1/*
2 * Copyright 1993-2018 NVIDIA Corporation. All rights reserved.
3 *
4 * NOTICE TO LICENSEE:
5 *
6 * This source code and/or documentation ("Licensed Deliverables") are
7 * subject to NVIDIA intellectual property rights under U.S. and
8 * international Copyright laws.
9 *
10 * These Licensed Deliverables contained herein is PROPRIETARY and
11 * CONFIDENTIAL to NVIDIA and is being provided under the terms and
12 * conditions of a form of NVIDIA software license agreement by and
13 * between NVIDIA and Licensee ("License Agreement") or electronically
14 * accepted by Licensee. Notwithstanding any terms or conditions to
15 * the contrary in the License Agreement, reproduction or disclosure
16 * of the Licensed Deliverables to any third party without the express
17 * written consent of NVIDIA is prohibited.
18 *
19 * NOTWITHSTANDING ANY TERMS OR CONDITIONS TO THE CONTRARY IN THE
20 * LICENSE AGREEMENT, NVIDIA MAKES NO REPRESENTATION ABOUT THE
21 * SUITABILITY OF THESE LICENSED DELIVERABLES FOR ANY PURPOSE. IT IS
22 * PROVIDED "AS IS" WITHOUT EXPRESS OR IMPLIED WARRANTY OF ANY KIND.
23 * NVIDIA DISCLAIMS ALL WARRANTIES WITH REGARD TO THESE LICENSED
24 * DELIVERABLES, INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY,
25 * NONINFRINGEMENT, AND FITNESS FOR A PARTICULAR PURPOSE.
26 * NOTWITHSTANDING ANY TERMS OR CONDITIONS TO THE CONTRARY IN THE
27 * LICENSE AGREEMENT, IN NO EVENT SHALL NVIDIA BE LIABLE FOR ANY
28 * SPECIAL, INDIRECT, INCIDENTAL, OR CONSEQUENTIAL DAMAGES, OR ANY
29 * DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS,
30 * WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS
31 * ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE
32 * OF THESE LICENSED DELIVERABLES.
33 *
34 * U.S. Government End Users. These Licensed Deliverables are a
35 * "commercial item" as that term is defined at 48 C.F.R. 2.101 (OCT
36 * 1995), consisting of "commercial computer software" and "commercial
37 * computer software documentation" as such terms are used in 48
38 * C.F.R. 12.212 (SEPT 1995) and is provided to the U.S. Government
39 * only as a commercial end item. Consistent with 48 C.F.R.12.212 and
40 * 48 C.F.R. 227.7202-1 through 227.7202-4 (JUNE 1995), all
41 * U.S. Government End Users acquire the Licensed Deliverables with
42 * only those rights set forth herein.
43 *
44 * Any use of the Licensed Deliverables in individual and commercial
45 * software must include, in the user documentation and internal
46 * comments to the code, the above Disclaimer and U.S. Government End
47 * Users Notice.
48 */
49
50#ifndef __cuda_cuda_h__
51#define __cuda_cuda_h__
52
53#include <stdlib.h>
54#ifdef _MSC_VER
55typedef unsigned __int32 cuuint32_t;
56typedef unsigned __int64 cuuint64_t;
57#else
58#include <stdint.h>
59typedef uint32_t cuuint32_t;
60typedef uint64_t cuuint64_t;
61#endif
62
63#if defined(__CUDA_API_VERSION_INTERNAL) || defined(__DOXYGEN_ONLY__) || defined(CUDA_ENABLE_DEPRECATED)
64#define __CUDA_DEPRECATED
65#elif defined(_MSC_VER)
66#define __CUDA_DEPRECATED __declspec(deprecated)
67#elif defined(__GNUC__)
68#define __CUDA_DEPRECATED __attribute__((deprecated))
69#else
70#define __CUDA_DEPRECATED
71#endif
72
73#if defined(CUDA_FORCE_API_VERSION)
74#error "CUDA_FORCE_API_VERSION is no longer supported."
75#endif
76
77#if defined(__CUDA_API_VERSION_INTERNAL) || defined(CUDA_API_PER_THREAD_DEFAULT_STREAM)
78 #define __CUDA_API_PER_THREAD_DEFAULT_STREAM
79 #define __CUDA_API_PTDS(api) api ## _ptds
80 #define __CUDA_API_PTSZ(api) api ## _ptsz
81#else
82 #define __CUDA_API_PTDS(api) api
83 #define __CUDA_API_PTSZ(api) api
84#endif
85
86#define cuDeviceTotalMem cuDeviceTotalMem_v2
87#define cuCtxCreate cuCtxCreate_v2
88#define cuCtxCreate_v3 cuCtxCreate_v3
89#define cuModuleGetGlobal cuModuleGetGlobal_v2
90#define cuMemGetInfo cuMemGetInfo_v2
91#define cuMemAlloc cuMemAlloc_v2
92#define cuMemAllocPitch cuMemAllocPitch_v2
93#define cuMemFree cuMemFree_v2
94#define cuMemGetAddressRange cuMemGetAddressRange_v2
95#define cuMemAllocHost cuMemAllocHost_v2
96#define cuMemHostGetDevicePointer cuMemHostGetDevicePointer_v2
97#define cuMemcpyHtoD __CUDA_API_PTDS(cuMemcpyHtoD_v2)
98#define cuMemcpyDtoH __CUDA_API_PTDS(cuMemcpyDtoH_v2)
99#define cuMemcpyDtoD __CUDA_API_PTDS(cuMemcpyDtoD_v2)
100#define cuMemcpyDtoA __CUDA_API_PTDS(cuMemcpyDtoA_v2)
101#define cuMemcpyAtoD __CUDA_API_PTDS(cuMemcpyAtoD_v2)
102#define cuMemcpyHtoA __CUDA_API_PTDS(cuMemcpyHtoA_v2)
103#define cuMemcpyAtoH __CUDA_API_PTDS(cuMemcpyAtoH_v2)
104#define cuMemcpyAtoA __CUDA_API_PTDS(cuMemcpyAtoA_v2)
105#define cuMemcpyHtoAAsync __CUDA_API_PTSZ(cuMemcpyHtoAAsync_v2)
106#define cuMemcpyAtoHAsync __CUDA_API_PTSZ(cuMemcpyAtoHAsync_v2)
107#define cuMemcpy2D __CUDA_API_PTDS(cuMemcpy2D_v2)
108#define cuMemcpy2DUnaligned __CUDA_API_PTDS(cuMemcpy2DUnaligned_v2)
109#define cuMemcpy3D __CUDA_API_PTDS(cuMemcpy3D_v2)
110#define cuMemcpyHtoDAsync __CUDA_API_PTSZ(cuMemcpyHtoDAsync_v2)
111#define cuMemcpyDtoHAsync __CUDA_API_PTSZ(cuMemcpyDtoHAsync_v2)
112#define cuMemcpyDtoDAsync __CUDA_API_PTSZ(cuMemcpyDtoDAsync_v2)
113#define cuMemcpy2DAsync __CUDA_API_PTSZ(cuMemcpy2DAsync_v2)
114#define cuMemcpy3DAsync __CUDA_API_PTSZ(cuMemcpy3DAsync_v2)
115#define cuMemsetD8 __CUDA_API_PTDS(cuMemsetD8_v2)
116#define cuMemsetD16 __CUDA_API_PTDS(cuMemsetD16_v2)
117#define cuMemsetD32 __CUDA_API_PTDS(cuMemsetD32_v2)
118#define cuMemsetD2D8 __CUDA_API_PTDS(cuMemsetD2D8_v2)
119#define cuMemsetD2D16 __CUDA_API_PTDS(cuMemsetD2D16_v2)
120#define cuMemsetD2D32 __CUDA_API_PTDS(cuMemsetD2D32_v2)
121#define cuArrayCreate cuArrayCreate_v2
122#define cuArrayGetDescriptor cuArrayGetDescriptor_v2
123#define cuArray3DCreate cuArray3DCreate_v2
124#define cuArray3DGetDescriptor cuArray3DGetDescriptor_v2
125#define cuTexRefSetAddress cuTexRefSetAddress_v2
126#define cuTexRefGetAddress cuTexRefGetAddress_v2
127#define cuGraphicsResourceGetMappedPointer cuGraphicsResourceGetMappedPointer_v2
128#define cuCtxDestroy cuCtxDestroy_v2
129#define cuCtxPopCurrent cuCtxPopCurrent_v2
130#define cuCtxPushCurrent cuCtxPushCurrent_v2
131#define cuStreamDestroy cuStreamDestroy_v2
132#define cuEventDestroy cuEventDestroy_v2
133#define cuTexRefSetAddress2D cuTexRefSetAddress2D_v3
134#define cuLinkCreate cuLinkCreate_v2
135#define cuLinkAddData cuLinkAddData_v2
136#define cuLinkAddFile cuLinkAddFile_v2
137#define cuMemHostRegister cuMemHostRegister_v2
138#define cuGraphicsResourceSetMapFlags cuGraphicsResourceSetMapFlags_v2
139#define cuStreamBeginCapture __CUDA_API_PTSZ(cuStreamBeginCapture_v2)
140#define cuDevicePrimaryCtxRelease cuDevicePrimaryCtxRelease_v2
141#define cuDevicePrimaryCtxReset cuDevicePrimaryCtxReset_v2
142#define cuDevicePrimaryCtxSetFlags cuDevicePrimaryCtxSetFlags_v2
143#define cuDeviceGetUuid_v2 cuDeviceGetUuid_v2
144#define cuIpcOpenMemHandle cuIpcOpenMemHandle_v2
145#define cuGraphInstantiate cuGraphInstantiate_v2
146
147#if defined(__CUDA_API_PER_THREAD_DEFAULT_STREAM)
148 #define cuMemcpy __CUDA_API_PTDS(cuMemcpy)
149 #define cuMemcpyAsync __CUDA_API_PTSZ(cuMemcpyAsync)
150 #define cuMemcpyPeer __CUDA_API_PTDS(cuMemcpyPeer)
151 #define cuMemcpyPeerAsync __CUDA_API_PTSZ(cuMemcpyPeerAsync)
152 #define cuMemcpy3DPeer __CUDA_API_PTDS(cuMemcpy3DPeer)
153 #define cuMemcpy3DPeerAsync __CUDA_API_PTSZ(cuMemcpy3DPeerAsync)
154 #define cuMemPrefetchAsync __CUDA_API_PTSZ(cuMemPrefetchAsync)
155
156 #define cuMemsetD8Async __CUDA_API_PTSZ(cuMemsetD8Async)
157 #define cuMemsetD16Async __CUDA_API_PTSZ(cuMemsetD16Async)
158 #define cuMemsetD32Async __CUDA_API_PTSZ(cuMemsetD32Async)
159 #define cuMemsetD2D8Async __CUDA_API_PTSZ(cuMemsetD2D8Async)
160 #define cuMemsetD2D16Async __CUDA_API_PTSZ(cuMemsetD2D16Async)
161 #define cuMemsetD2D32Async __CUDA_API_PTSZ(cuMemsetD2D32Async)
162
163 #define cuStreamGetPriority __CUDA_API_PTSZ(cuStreamGetPriority)
164 #define cuStreamGetFlags __CUDA_API_PTSZ(cuStreamGetFlags)
165 #define cuStreamGetCtx __CUDA_API_PTSZ(cuStreamGetCtx)
166 #define cuStreamWaitEvent __CUDA_API_PTSZ(cuStreamWaitEvent)
167 #define cuStreamEndCapture __CUDA_API_PTSZ(cuStreamEndCapture)
168 #define cuStreamIsCapturing __CUDA_API_PTSZ(cuStreamIsCapturing)
169 #define cuStreamGetCaptureInfo __CUDA_API_PTSZ(cuStreamGetCaptureInfo)
170 #define cuStreamGetCaptureInfo_v2 __CUDA_API_PTSZ(cuStreamGetCaptureInfo_v2)
171 #define cuStreamUpdateCaptureDependencies __CUDA_API_PTSZ(cuStreamUpdateCaptureDependencies)
172 #define cuStreamAddCallback __CUDA_API_PTSZ(cuStreamAddCallback)
173 #define cuStreamAttachMemAsync __CUDA_API_PTSZ(cuStreamAttachMemAsync)
174 #define cuStreamQuery __CUDA_API_PTSZ(cuStreamQuery)
175 #define cuStreamSynchronize __CUDA_API_PTSZ(cuStreamSynchronize)
176 #define cuEventRecord __CUDA_API_PTSZ(cuEventRecord)
177 #define cuEventRecordWithFlags __CUDA_API_PTSZ(cuEventRecordWithFlags)
178 #define cuLaunchKernel __CUDA_API_PTSZ(cuLaunchKernel)
179 #define cuLaunchHostFunc __CUDA_API_PTSZ(cuLaunchHostFunc)
180 #define cuGraphicsMapResources __CUDA_API_PTSZ(cuGraphicsMapResources)
181 #define cuGraphicsUnmapResources __CUDA_API_PTSZ(cuGraphicsUnmapResources)
182
183 #define cuStreamWriteValue32 __CUDA_API_PTSZ(cuStreamWriteValue32)
184 #define cuStreamWaitValue32 __CUDA_API_PTSZ(cuStreamWaitValue32)
185 #define cuStreamWriteValue64 __CUDA_API_PTSZ(cuStreamWriteValue64)
186 #define cuStreamWaitValue64 __CUDA_API_PTSZ(cuStreamWaitValue64)
187 #define cuStreamBatchMemOp __CUDA_API_PTSZ(cuStreamBatchMemOp)
188
189 #define cuLaunchCooperativeKernel __CUDA_API_PTSZ(cuLaunchCooperativeKernel)
190
191 #define cuSignalExternalSemaphoresAsync __CUDA_API_PTSZ(cuSignalExternalSemaphoresAsync)
192 #define cuWaitExternalSemaphoresAsync __CUDA_API_PTSZ(cuWaitExternalSemaphoresAsync)
193
194 #define cuGraphUpload __CUDA_API_PTSZ(cuGraphUpload)
195 #define cuGraphLaunch __CUDA_API_PTSZ(cuGraphLaunch)
196 #define cuStreamCopyAttributes __CUDA_API_PTSZ(cuStreamCopyAttributes)
197 #define cuStreamGetAttribute __CUDA_API_PTSZ(cuStreamGetAttribute)
198 #define cuStreamSetAttribute __CUDA_API_PTSZ(cuStreamSetAttribute)
199 #define cuMemMapArrayAsync __CUDA_API_PTSZ(cuMemMapArrayAsync)
200
201 #define cuMemFreeAsync __CUDA_API_PTSZ(cuMemFreeAsync)
202 #define cuMemAllocAsync __CUDA_API_PTSZ(cuMemAllocAsync)
203 #define cuMemAllocFromPoolAsync __CUDA_API_PTSZ(cuMemAllocFromPoolAsync)
204#endif
205
206/**
207 * \file cuda.h
208 * \brief Header file for the CUDA Toolkit application programming interface.
209 *
210 * \file cudaGL.h
211 * \brief Header file for the OpenGL interoperability functions of the
212 * low-level CUDA driver application programming interface.
213 *
214 * \file cudaD3D9.h
215 * \brief Header file for the Direct3D 9 interoperability functions of the
216 * low-level CUDA driver application programming interface.
217 */
218
219/**
220 * \defgroup CUDA_TYPES Data types used by CUDA driver
221 * @{
222 */
223
224/**
225 * CUDA API version number
226 */
227#define CUDA_VERSION 11040
228
229#ifdef __cplusplus
230extern "C" {
231#endif
232
233/**
234 * CUDA device pointer
235 * CUdeviceptr is defined as an unsigned integer type whose size matches the size of a pointer on the target platform.
236 */
237#if defined(_WIN64) || defined(__LP64__)
238typedef unsigned long long CUdeviceptr_v2;
239#else
240typedef unsigned int CUdeviceptr_v2;
241#endif
242typedef CUdeviceptr_v2 CUdeviceptr; /**< CUDA device pointer */
243
244typedef int CUdevice_v1; /**< CUDA device */
245typedef CUdevice_v1 CUdevice; /**< CUDA device */
246typedef struct CUctx_st *CUcontext; /**< CUDA context */
247typedef struct CUmod_st *CUmodule; /**< CUDA module */
248typedef struct CUfunc_st *CUfunction; /**< CUDA function */
249typedef struct CUarray_st *CUarray; /**< CUDA array */
250typedef struct CUmipmappedArray_st *CUmipmappedArray; /**< CUDA mipmapped array */
251typedef struct CUtexref_st *CUtexref; /**< CUDA texture reference */
252typedef struct CUsurfref_st *CUsurfref; /**< CUDA surface reference */
253typedef struct CUevent_st *CUevent; /**< CUDA event */
254typedef struct CUstream_st *CUstream; /**< CUDA stream */
255typedef struct CUgraphicsResource_st *CUgraphicsResource; /**< CUDA graphics interop resource */
256typedef unsigned long long CUtexObject_v1; /**< An opaque value that represents a CUDA texture object */
257typedef CUtexObject_v1 CUtexObject; /**< An opaque value that represents a CUDA texture object */
258typedef unsigned long long CUsurfObject_v1; /**< An opaque value that represents a CUDA surface object */
259typedef CUsurfObject_v1 CUsurfObject; /**< An opaque value that represents a CUDA surface object */
260typedef struct CUextMemory_st *CUexternalMemory; /**< CUDA external memory */
261typedef struct CUextSemaphore_st *CUexternalSemaphore; /**< CUDA external semaphore */
262typedef struct CUgraph_st *CUgraph; /**< CUDA graph */
263typedef struct CUgraphNode_st *CUgraphNode; /**< CUDA graph node */
264typedef struct CUgraphExec_st *CUgraphExec; /**< CUDA executable graph */
265typedef struct CUmemPoolHandle_st *CUmemoryPool; /**< CUDA memory pool */
266typedef struct CUuserObject_st *CUuserObject; /**< CUDA user object for graphs */
267
268#ifndef CU_UUID_HAS_BEEN_DEFINED
269#define CU_UUID_HAS_BEEN_DEFINED
270typedef struct CUuuid_st { /**< CUDA definition of UUID */
271 char bytes[16];
272} CUuuid;
273#endif
274
275/**
276 * CUDA IPC handle size
277 */
278#define CU_IPC_HANDLE_SIZE 64
279
280/**
281 * CUDA IPC event handle
282 */
283typedef struct CUipcEventHandle_st {
284 char reserved[CU_IPC_HANDLE_SIZE];
285} CUipcEventHandle_v1;
286typedef CUipcEventHandle_v1 CUipcEventHandle;
287
288/**
289 * CUDA IPC mem handle
290 */
291typedef struct CUipcMemHandle_st {
292 char reserved[CU_IPC_HANDLE_SIZE];
293} CUipcMemHandle_v1;
294typedef CUipcMemHandle_v1 CUipcMemHandle;
295
296/**
297 * CUDA Ipc Mem Flags
298 */
299typedef enum CUipcMem_flags_enum {
300 CU_IPC_MEM_LAZY_ENABLE_PEER_ACCESS = 0x1 /**< Automatically enable peer access between remote devices as needed */
301} CUipcMem_flags;
302
303
304/**
305 * CUDA Mem Attach Flags
306 */
307typedef enum CUmemAttach_flags_enum {
308 CU_MEM_ATTACH_GLOBAL = 0x1, /**< Memory can be accessed by any stream on any device */
309 CU_MEM_ATTACH_HOST = 0x2, /**< Memory cannot be accessed by any stream on any device */
310 CU_MEM_ATTACH_SINGLE = 0x4 /**< Memory can only be accessed by a single stream on the associated device */
311} CUmemAttach_flags;
312
313/**
314 * Context creation flags
315 */
316typedef enum CUctx_flags_enum {
317 CU_CTX_SCHED_AUTO = 0x00, /**< Automatic scheduling */
318 CU_CTX_SCHED_SPIN = 0x01, /**< Set spin as default scheduling */
319 CU_CTX_SCHED_YIELD = 0x02, /**< Set yield as default scheduling */
320 CU_CTX_SCHED_BLOCKING_SYNC = 0x04, /**< Set blocking synchronization as default scheduling */
321 CU_CTX_BLOCKING_SYNC = 0x04, /**< Set blocking synchronization as default scheduling
322 * \deprecated This flag was deprecated as of CUDA 4.0
323 * and was replaced with ::CU_CTX_SCHED_BLOCKING_SYNC. */
324 CU_CTX_SCHED_MASK = 0x07,
325 CU_CTX_MAP_HOST = 0x08, /**< \deprecated This flag was deprecated as of CUDA 11.0
326 * and it no longer has any effect. All contexts
327 * as of CUDA 3.2 behave as though the flag is enabled. */
328 CU_CTX_LMEM_RESIZE_TO_MAX = 0x10, /**< Keep local memory allocation after launch */
329 CU_CTX_FLAGS_MASK = 0x1f
330} CUctx_flags;
331
332/**
333 * Stream creation flags
334 */
335typedef enum CUstream_flags_enum {
336 CU_STREAM_DEFAULT = 0x0, /**< Default stream flag */
337 CU_STREAM_NON_BLOCKING = 0x1 /**< Stream does not synchronize with stream 0 (the NULL stream) */
338} CUstream_flags;
339
340/**
341 * Legacy stream handle
342 *
343 * Stream handle that can be passed as a CUstream to use an implicit stream
344 * with legacy synchronization behavior.
345 *
346 * See details of the \link_sync_behavior
347 */
348#define CU_STREAM_LEGACY ((CUstream)0x1)
349
350/**
351 * Per-thread stream handle
352 *
353 * Stream handle that can be passed as a CUstream to use an implicit stream
354 * with per-thread synchronization behavior.
355 *
356 * See details of the \link_sync_behavior
357 */
358#define CU_STREAM_PER_THREAD ((CUstream)0x2)
359
360/**
361 * Event creation flags
362 */
363typedef enum CUevent_flags_enum {
364 CU_EVENT_DEFAULT = 0x0, /**< Default event flag */
365 CU_EVENT_BLOCKING_SYNC = 0x1, /**< Event uses blocking synchronization */
366 CU_EVENT_DISABLE_TIMING = 0x2, /**< Event will not record timing data */
367 CU_EVENT_INTERPROCESS = 0x4 /**< Event is suitable for interprocess use. CU_EVENT_DISABLE_TIMING must be set */
368} CUevent_flags;
369
370/**
371 * Event record flags
372 */
373typedef enum CUevent_record_flags_enum {
374 CU_EVENT_RECORD_DEFAULT = 0x0, /**< Default event record flag */
375 CU_EVENT_RECORD_EXTERNAL = 0x1 /**< When using stream capture, create an event record node
376 * instead of the default behavior. This flag is invalid
377 * when used outside of capture. */
378} CUevent_record_flags;
379
380/**
381 * Event wait flags
382 */
383typedef enum CUevent_wait_flags_enum {
384 CU_EVENT_WAIT_DEFAULT = 0x0, /**< Default event wait flag */
385 CU_EVENT_WAIT_EXTERNAL = 0x1 /**< When using stream capture, create an event wait node
386 * instead of the default behavior. This flag is invalid
387 * when used outside of capture.*/
388} CUevent_wait_flags;
389
390/**
391 * Flags for ::cuStreamWaitValue32 and ::cuStreamWaitValue64
392 */
393typedef enum CUstreamWaitValue_flags_enum {
394 CU_STREAM_WAIT_VALUE_GEQ = 0x0, /**< Wait until (int32_t)(*addr - value) >= 0 (or int64_t for 64 bit
395 values). Note this is a cyclic comparison which ignores wraparound.
396 (Default behavior.) */
397 CU_STREAM_WAIT_VALUE_EQ = 0x1, /**< Wait until *addr == value. */
398 CU_STREAM_WAIT_VALUE_AND = 0x2, /**< Wait until (*addr & value) != 0. */
399 CU_STREAM_WAIT_VALUE_NOR = 0x3, /**< Wait until ~(*addr | value) != 0. Support for this operation can be
400 queried with ::cuDeviceGetAttribute() and
401 ::CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_WAIT_VALUE_NOR.*/
402 CU_STREAM_WAIT_VALUE_FLUSH = 1<<30 /**< Follow the wait operation with a flush of outstanding remote writes. This
403 means that, if a remote write operation is guaranteed to have reached the
404 device before the wait can be satisfied, that write is guaranteed to be
405 visible to downstream device work. The device is permitted to reorder
406 remote writes internally. For example, this flag would be required if
407 two remote writes arrive in a defined order, the wait is satisfied by the
408 second write, and downstream work needs to observe the first write.
409 Support for this operation is restricted to selected platforms and can be
410 queried with ::CU_DEVICE_ATTRIBUTE_CAN_USE_WAIT_VALUE_FLUSH.*/
411} CUstreamWaitValue_flags;
412
413/**
414 * Flags for ::cuStreamWriteValue32
415 */
416typedef enum CUstreamWriteValue_flags_enum {
417 CU_STREAM_WRITE_VALUE_DEFAULT = 0x0, /**< Default behavior */
418 CU_STREAM_WRITE_VALUE_NO_MEMORY_BARRIER = 0x1 /**< Permits the write to be reordered with writes which were issued
419 before it, as a performance optimization. Normally,
420 ::cuStreamWriteValue32 will provide a memory fence before the
421 write, which has similar semantics to
422 __threadfence_system() but is scoped to the stream
423 rather than a CUDA thread. */
424} CUstreamWriteValue_flags;
425
426/**
427 * Operations for ::cuStreamBatchMemOp
428 */
429typedef enum CUstreamBatchMemOpType_enum {
430 CU_STREAM_MEM_OP_WAIT_VALUE_32 = 1, /**< Represents a ::cuStreamWaitValue32 operation */
431 CU_STREAM_MEM_OP_WRITE_VALUE_32 = 2, /**< Represents a ::cuStreamWriteValue32 operation */
432 CU_STREAM_MEM_OP_WAIT_VALUE_64 = 4, /**< Represents a ::cuStreamWaitValue64 operation */
433 CU_STREAM_MEM_OP_WRITE_VALUE_64 = 5, /**< Represents a ::cuStreamWriteValue64 operation */
434 CU_STREAM_MEM_OP_FLUSH_REMOTE_WRITES = 3 /**< This has the same effect as ::CU_STREAM_WAIT_VALUE_FLUSH, but as a
435 standalone operation. */
436} CUstreamBatchMemOpType;
437
438/**
439 * Per-operation parameters for ::cuStreamBatchMemOp
440 */
441typedef union CUstreamBatchMemOpParams_union {
442 CUstreamBatchMemOpType operation;
443 struct CUstreamMemOpWaitValueParams_st {
444 CUstreamBatchMemOpType operation;
445 CUdeviceptr address;
446 union {
447 cuuint32_t value;
448 cuuint64_t value64;
449 };
450 unsigned int flags;
451 CUdeviceptr alias; /**< For driver internal use. Initial value is unimportant. */
452 } waitValue;
453 struct CUstreamMemOpWriteValueParams_st {
454 CUstreamBatchMemOpType operation;
455 CUdeviceptr address;
456 union {
457 cuuint32_t value;
458 cuuint64_t value64;
459 };
460 unsigned int flags;
461 CUdeviceptr alias; /**< For driver internal use. Initial value is unimportant. */
462 } writeValue;
463 struct CUstreamMemOpFlushRemoteWritesParams_st {
464 CUstreamBatchMemOpType operation;
465 unsigned int flags;
466 } flushRemoteWrites;
467 cuuint64_t pad[6];
468} CUstreamBatchMemOpParams_v1;
469typedef CUstreamBatchMemOpParams_v1 CUstreamBatchMemOpParams;
470
471/**
472 * Occupancy calculator flag
473 */
474typedef enum CUoccupancy_flags_enum {
475 CU_OCCUPANCY_DEFAULT = 0x0, /**< Default behavior */
476 CU_OCCUPANCY_DISABLE_CACHING_OVERRIDE = 0x1 /**< Assume global caching is enabled and cannot be automatically turned off */
477} CUoccupancy_flags;
478
479/**
480 * Flags for ::cuStreamUpdateCaptureDependencies
481 */
482typedef enum CUstreamUpdateCaptureDependencies_flags_enum {
483 CU_STREAM_ADD_CAPTURE_DEPENDENCIES = 0x0, /**< Add new nodes to the dependency set */
484 CU_STREAM_SET_CAPTURE_DEPENDENCIES = 0x1 /**< Replace the dependency set with the new nodes */
485} CUstreamUpdateCaptureDependencies_flags;
486
487/**
488 * Array formats
489 */
490typedef enum CUarray_format_enum {
491 CU_AD_FORMAT_UNSIGNED_INT8 = 0x01, /**< Unsigned 8-bit integers */
492 CU_AD_FORMAT_UNSIGNED_INT16 = 0x02, /**< Unsigned 16-bit integers */
493 CU_AD_FORMAT_UNSIGNED_INT32 = 0x03, /**< Unsigned 32-bit integers */
494 CU_AD_FORMAT_SIGNED_INT8 = 0x08, /**< Signed 8-bit integers */
495 CU_AD_FORMAT_SIGNED_INT16 = 0x09, /**< Signed 16-bit integers */
496 CU_AD_FORMAT_SIGNED_INT32 = 0x0a, /**< Signed 32-bit integers */
497 CU_AD_FORMAT_HALF = 0x10, /**< 16-bit floating point */
498 CU_AD_FORMAT_FLOAT = 0x20, /**< 32-bit floating point */
499 CU_AD_FORMAT_NV12 = 0xb0
500} CUarray_format;
501
502/**
503 * Texture reference addressing modes
504 */
505typedef enum CUaddress_mode_enum {
506 CU_TR_ADDRESS_MODE_WRAP = 0, /**< Wrapping address mode */
507 CU_TR_ADDRESS_MODE_CLAMP = 1, /**< Clamp to edge address mode */
508 CU_TR_ADDRESS_MODE_MIRROR = 2, /**< Mirror address mode */
509 CU_TR_ADDRESS_MODE_BORDER = 3 /**< Border address mode */
510} CUaddress_mode;
511
512/**
513 * Texture reference filtering modes
514 */
515typedef enum CUfilter_mode_enum {
516 CU_TR_FILTER_MODE_POINT = 0, /**< Point filter mode */
517 CU_TR_FILTER_MODE_LINEAR = 1 /**< Linear filter mode */
518} CUfilter_mode;
519
520/**
521 * Device properties
522 */
523typedef enum CUdevice_attribute_enum {
524 CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK = 1, /**< Maximum number of threads per block */
525 CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_X = 2, /**< Maximum block dimension X */
526 CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Y = 3, /**< Maximum block dimension Y */
527 CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Z = 4, /**< Maximum block dimension Z */
528 CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_X = 5, /**< Maximum grid dimension X */
529 CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Y = 6, /**< Maximum grid dimension Y */
530 CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Z = 7, /**< Maximum grid dimension Z */
531 CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK = 8, /**< Maximum shared memory available per block in bytes */
532 CU_DEVICE_ATTRIBUTE_SHARED_MEMORY_PER_BLOCK = 8, /**< Deprecated, use CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK */
533 CU_DEVICE_ATTRIBUTE_TOTAL_CONSTANT_MEMORY = 9, /**< Memory available on device for __constant__ variables in a CUDA C kernel in bytes */
534 CU_DEVICE_ATTRIBUTE_WARP_SIZE = 10, /**< Warp size in threads */
535 CU_DEVICE_ATTRIBUTE_MAX_PITCH = 11, /**< Maximum pitch in bytes allowed by memory copies */
536 CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_BLOCK = 12, /**< Maximum number of 32-bit registers available per block */
537 CU_DEVICE_ATTRIBUTE_REGISTERS_PER_BLOCK = 12, /**< Deprecated, use CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_BLOCK */
538 CU_DEVICE_ATTRIBUTE_CLOCK_RATE = 13, /**< Typical clock frequency in kilohertz */
539 CU_DEVICE_ATTRIBUTE_TEXTURE_ALIGNMENT = 14, /**< Alignment requirement for textures */
540 CU_DEVICE_ATTRIBUTE_GPU_OVERLAP = 15, /**< Device can possibly copy memory and execute a kernel concurrently. Deprecated. Use instead CU_DEVICE_ATTRIBUTE_ASYNC_ENGINE_COUNT. */
541 CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT = 16, /**< Number of multiprocessors on device */
542 CU_DEVICE_ATTRIBUTE_KERNEL_EXEC_TIMEOUT = 17, /**< Specifies whether there is a run time limit on kernels */
543 CU_DEVICE_ATTRIBUTE_INTEGRATED = 18, /**< Device is integrated with host memory */
544 CU_DEVICE_ATTRIBUTE_CAN_MAP_HOST_MEMORY = 19, /**< Device can map host memory into CUDA address space */
545 CU_DEVICE_ATTRIBUTE_COMPUTE_MODE = 20, /**< Compute mode (See ::CUcomputemode for details) */
546 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_WIDTH = 21, /**< Maximum 1D texture width */
547 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_WIDTH = 22, /**< Maximum 2D texture width */
548 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_HEIGHT = 23, /**< Maximum 2D texture height */
549 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_WIDTH = 24, /**< Maximum 3D texture width */
550 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_HEIGHT = 25, /**< Maximum 3D texture height */
551 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_DEPTH = 26, /**< Maximum 3D texture depth */
552 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_WIDTH = 27, /**< Maximum 2D layered texture width */
553 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_HEIGHT = 28, /**< Maximum 2D layered texture height */
554 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_LAYERS = 29, /**< Maximum layers in a 2D layered texture */
555 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_ARRAY_WIDTH = 27, /**< Deprecated, use CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_WIDTH */
556 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_ARRAY_HEIGHT = 28, /**< Deprecated, use CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_HEIGHT */
557 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_ARRAY_NUMSLICES = 29, /**< Deprecated, use CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_LAYERS */
558 CU_DEVICE_ATTRIBUTE_SURFACE_ALIGNMENT = 30, /**< Alignment requirement for surfaces */
559 CU_DEVICE_ATTRIBUTE_CONCURRENT_KERNELS = 31, /**< Device can possibly execute multiple kernels concurrently */
560 CU_DEVICE_ATTRIBUTE_ECC_ENABLED = 32, /**< Device has ECC support enabled */
561 CU_DEVICE_ATTRIBUTE_PCI_BUS_ID = 33, /**< PCI bus ID of the device */
562 CU_DEVICE_ATTRIBUTE_PCI_DEVICE_ID = 34, /**< PCI device ID of the device */
563 CU_DEVICE_ATTRIBUTE_TCC_DRIVER = 35, /**< Device is using TCC driver model */
564 CU_DEVICE_ATTRIBUTE_MEMORY_CLOCK_RATE = 36, /**< Peak memory clock frequency in kilohertz */
565 CU_DEVICE_ATTRIBUTE_GLOBAL_MEMORY_BUS_WIDTH = 37, /**< Global memory bus width in bits */
566 CU_DEVICE_ATTRIBUTE_L2_CACHE_SIZE = 38, /**< Size of L2 cache in bytes */
567 CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR = 39, /**< Maximum resident threads per multiprocessor */
568 CU_DEVICE_ATTRIBUTE_ASYNC_ENGINE_COUNT = 40, /**< Number of asynchronous engines */
569 CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING = 41, /**< Device shares a unified address space with the host */
570 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LAYERED_WIDTH = 42, /**< Maximum 1D layered texture width */
571 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LAYERED_LAYERS = 43, /**< Maximum layers in a 1D layered texture */
572 CU_DEVICE_ATTRIBUTE_CAN_TEX2D_GATHER = 44, /**< Deprecated, do not use. */
573 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_GATHER_WIDTH = 45, /**< Maximum 2D texture width if CUDA_ARRAY3D_TEXTURE_GATHER is set */
574 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_GATHER_HEIGHT = 46, /**< Maximum 2D texture height if CUDA_ARRAY3D_TEXTURE_GATHER is set */
575 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_WIDTH_ALTERNATE = 47, /**< Alternate maximum 3D texture width */
576 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_HEIGHT_ALTERNATE = 48, /**< Alternate maximum 3D texture height */
577 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_DEPTH_ALTERNATE = 49, /**< Alternate maximum 3D texture depth */
578 CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID = 50, /**< PCI domain ID of the device */
579 CU_DEVICE_ATTRIBUTE_TEXTURE_PITCH_ALIGNMENT = 51, /**< Pitch alignment requirement for textures */
580 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURECUBEMAP_WIDTH = 52, /**< Maximum cubemap texture width/height */
581 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURECUBEMAP_LAYERED_WIDTH = 53, /**< Maximum cubemap layered texture width/height */
582 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURECUBEMAP_LAYERED_LAYERS = 54, /**< Maximum layers in a cubemap layered texture */
583 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE1D_WIDTH = 55, /**< Maximum 1D surface width */
584 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE2D_WIDTH = 56, /**< Maximum 2D surface width */
585 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE2D_HEIGHT = 57, /**< Maximum 2D surface height */
586 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE3D_WIDTH = 58, /**< Maximum 3D surface width */
587 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE3D_HEIGHT = 59, /**< Maximum 3D surface height */
588 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE3D_DEPTH = 60, /**< Maximum 3D surface depth */
589 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE1D_LAYERED_WIDTH = 61, /**< Maximum 1D layered surface width */
590 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE1D_LAYERED_LAYERS = 62, /**< Maximum layers in a 1D layered surface */
591 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE2D_LAYERED_WIDTH = 63, /**< Maximum 2D layered surface width */
592 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE2D_LAYERED_HEIGHT = 64, /**< Maximum 2D layered surface height */
593 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE2D_LAYERED_LAYERS = 65, /**< Maximum layers in a 2D layered surface */
594 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACECUBEMAP_WIDTH = 66, /**< Maximum cubemap surface width */
595 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACECUBEMAP_LAYERED_WIDTH = 67, /**< Maximum cubemap layered surface width */
596 CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACECUBEMAP_LAYERED_LAYERS = 68, /**< Maximum layers in a cubemap layered surface */
597 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LINEAR_WIDTH = 69, /**< Deprecated, do not use. Use cudaDeviceGetTexture1DLinearMaxWidth() or cuDeviceGetTexture1DLinearMaxWidth() instead. */
598 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_WIDTH = 70, /**< Maximum 2D linear texture width */
599 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_HEIGHT = 71, /**< Maximum 2D linear texture height */
600 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_PITCH = 72, /**< Maximum 2D linear texture pitch in bytes */
601 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_MIPMAPPED_WIDTH = 73, /**< Maximum mipmapped 2D texture width */
602 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_MIPMAPPED_HEIGHT = 74, /**< Maximum mipmapped 2D texture height */
603 CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR = 75, /**< Major compute capability version number */
604 CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR = 76, /**< Minor compute capability version number */
605 CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_MIPMAPPED_WIDTH = 77, /**< Maximum mipmapped 1D texture width */
606 CU_DEVICE_ATTRIBUTE_STREAM_PRIORITIES_SUPPORTED = 78, /**< Device supports stream priorities */
607 CU_DEVICE_ATTRIBUTE_GLOBAL_L1_CACHE_SUPPORTED = 79, /**< Device supports caching globals in L1 */
608 CU_DEVICE_ATTRIBUTE_LOCAL_L1_CACHE_SUPPORTED = 80, /**< Device supports caching locals in L1 */
609 CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_MULTIPROCESSOR = 81, /**< Maximum shared memory available per multiprocessor in bytes */
610 CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_MULTIPROCESSOR = 82, /**< Maximum number of 32-bit registers available per multiprocessor */
611 CU_DEVICE_ATTRIBUTE_MANAGED_MEMORY = 83, /**< Device can allocate managed memory on this system */
612 CU_DEVICE_ATTRIBUTE_MULTI_GPU_BOARD = 84, /**< Device is on a multi-GPU board */
613 CU_DEVICE_ATTRIBUTE_MULTI_GPU_BOARD_GROUP_ID = 85, /**< Unique id for a group of devices on the same multi-GPU board */
614 CU_DEVICE_ATTRIBUTE_HOST_NATIVE_ATOMIC_SUPPORTED = 86, /**< Link between the device and the host supports native atomic operations (this is a placeholder attribute, and is not supported on any current hardware)*/
615 CU_DEVICE_ATTRIBUTE_SINGLE_TO_DOUBLE_PRECISION_PERF_RATIO = 87, /**< Ratio of single precision performance (in floating-point operations per second) to double precision performance */
616 CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS = 88, /**< Device supports coherently accessing pageable memory without calling cudaHostRegister on it */
617 CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS = 89, /**< Device can coherently access managed memory concurrently with the CPU */
618 CU_DEVICE_ATTRIBUTE_COMPUTE_PREEMPTION_SUPPORTED = 90, /**< Device supports compute preemption. */
619 CU_DEVICE_ATTRIBUTE_CAN_USE_HOST_POINTER_FOR_REGISTERED_MEM = 91, /**< Device can access host registered memory at the same virtual address as the CPU */
620 CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_MEM_OPS = 92, /**< ::cuStreamBatchMemOp and related APIs are supported. */
621 CU_DEVICE_ATTRIBUTE_CAN_USE_64_BIT_STREAM_MEM_OPS = 93, /**< 64-bit operations are supported in ::cuStreamBatchMemOp and related APIs. */
622 CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_WAIT_VALUE_NOR = 94, /**< ::CU_STREAM_WAIT_VALUE_NOR is supported. */
623 CU_DEVICE_ATTRIBUTE_COOPERATIVE_LAUNCH = 95, /**< Device supports launching cooperative kernels via ::cuLaunchCooperativeKernel */
624 CU_DEVICE_ATTRIBUTE_COOPERATIVE_MULTI_DEVICE_LAUNCH = 96, /**< Deprecated, ::cuLaunchCooperativeKernelMultiDevice is deprecated. */
625 CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN = 97, /**< Maximum optin shared memory per block */
626 CU_DEVICE_ATTRIBUTE_CAN_FLUSH_REMOTE_WRITES = 98, /**< The ::CU_STREAM_WAIT_VALUE_FLUSH flag and the ::CU_STREAM_MEM_OP_FLUSH_REMOTE_WRITES MemOp are supported on the device. See \ref CUDA_MEMOP for additional details. */
627 CU_DEVICE_ATTRIBUTE_HOST_REGISTER_SUPPORTED = 99, /**< Device supports host memory registration via ::cudaHostRegister. */
628 CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS_USES_HOST_PAGE_TABLES = 100, /**< Device accesses pageable memory via the host's page tables. */
629 CU_DEVICE_ATTRIBUTE_DIRECT_MANAGED_MEM_ACCESS_FROM_HOST = 101, /**< The host can directly access managed memory on the device without migration. */
630 CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED = 102, /**< Deprecated, Use CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED*/
631 CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED = 102, /**< Device supports virtual memory management APIs like ::cuMemAddressReserve, ::cuMemCreate, ::cuMemMap and related APIs */
632 CU_DEVICE_ATTRIBUTE_HANDLE_TYPE_POSIX_FILE_DESCRIPTOR_SUPPORTED = 103, /**< Device supports exporting memory to a posix file descriptor with ::cuMemExportToShareableHandle, if requested via ::cuMemCreate */
633 CU_DEVICE_ATTRIBUTE_HANDLE_TYPE_WIN32_HANDLE_SUPPORTED = 104, /**< Device supports exporting memory to a Win32 NT handle with ::cuMemExportToShareableHandle, if requested via ::cuMemCreate */
634 CU_DEVICE_ATTRIBUTE_HANDLE_TYPE_WIN32_KMT_HANDLE_SUPPORTED = 105, /**< Device supports exporting memory to a Win32 KMT handle with ::cuMemExportToShareableHandle, if requested ::cuMemCreate */
635 CU_DEVICE_ATTRIBUTE_MAX_BLOCKS_PER_MULTIPROCESSOR = 106, /**< Maximum number of blocks per multiprocessor */
636 CU_DEVICE_ATTRIBUTE_GENERIC_COMPRESSION_SUPPORTED = 107, /**< Device supports compression of memory */
637 CU_DEVICE_ATTRIBUTE_MAX_PERSISTING_L2_CACHE_SIZE = 108, /**< Maximum L2 persisting lines capacity setting in bytes. */
638 CU_DEVICE_ATTRIBUTE_MAX_ACCESS_POLICY_WINDOW_SIZE = 109, /**< Maximum value of CUaccessPolicyWindow::num_bytes. */
639 CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_WITH_CUDA_VMM_SUPPORTED = 110, /**< Device supports specifying the GPUDirect RDMA flag with ::cuMemCreate */
640 CU_DEVICE_ATTRIBUTE_RESERVED_SHARED_MEMORY_PER_BLOCK = 111, /**< Shared memory reserved by CUDA driver per block in bytes */
641 CU_DEVICE_ATTRIBUTE_SPARSE_CUDA_ARRAY_SUPPORTED = 112, /**< Device supports sparse CUDA arrays and sparse CUDA mipmapped arrays */
642 CU_DEVICE_ATTRIBUTE_READ_ONLY_HOST_REGISTER_SUPPORTED = 113, /**< Device supports using the ::cuMemHostRegister flag CU_MEMHOSTERGISTER_READ_ONLY to register memory that must be mapped as read-only to the GPU */
643 CU_DEVICE_ATTRIBUTE_TIMELINE_SEMAPHORE_INTEROP_SUPPORTED = 114, /**< External timeline semaphore interop is supported on the device */
644 CU_DEVICE_ATTRIBUTE_MEMORY_POOLS_SUPPORTED = 115, /**< Device supports using the ::cuMemAllocAsync and ::cuMemPool family of APIs */
645 CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_SUPPORTED = 116, /**< Device supports GPUDirect RDMA APIs, like nvidia_p2p_get_pages (see https://docs.nvidia.com/cuda/gpudirect-rdma for more information) */
646 CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_FLUSH_WRITES_OPTIONS = 117, /**< The returned attribute shall be interpreted as a bitmask, where the individual bits are described by the ::CUflushGPUDirectRDMAWritesOptions enum */
647 CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_WRITES_ORDERING = 118, /**< GPUDirect RDMA writes to the device do not need to be flushed for consumers within the scope indicated by the returned attribute. See ::CUGPUDirectRDMAWritesOrdering for the numerical values returned here. */
648 CU_DEVICE_ATTRIBUTE_MEMPOOL_SUPPORTED_HANDLE_TYPES = 119, /**< Handle types supported with mempool based IPC */
649 CU_DEVICE_ATTRIBUTE_MAX
650} CUdevice_attribute;
651
652/**
653 * Legacy device properties
654 */
655typedef struct CUdevprop_st {
656 int maxThreadsPerBlock; /**< Maximum number of threads per block */
657 int maxThreadsDim[3]; /**< Maximum size of each dimension of a block */
658 int maxGridSize[3]; /**< Maximum size of each dimension of a grid */
659 int sharedMemPerBlock; /**< Shared memory available per block in bytes */
660 int totalConstantMemory; /**< Constant memory available on device in bytes */
661 int SIMDWidth; /**< Warp size in threads */
662 int memPitch; /**< Maximum pitch in bytes allowed by memory copies */
663 int regsPerBlock; /**< 32-bit registers available per block */
664 int clockRate; /**< Clock frequency in kilohertz */
665 int textureAlign; /**< Alignment requirement for textures */
666} CUdevprop_v1;
667typedef CUdevprop_v1 CUdevprop;
668
669/**
670 * Pointer information
671 */
672typedef enum CUpointer_attribute_enum {
673 CU_POINTER_ATTRIBUTE_CONTEXT = 1, /**< The ::CUcontext on which a pointer was allocated or registered */
674 CU_POINTER_ATTRIBUTE_MEMORY_TYPE = 2, /**< The ::CUmemorytype describing the physical location of a pointer */
675 CU_POINTER_ATTRIBUTE_DEVICE_POINTER = 3, /**< The address at which a pointer's memory may be accessed on the device */
676 CU_POINTER_ATTRIBUTE_HOST_POINTER = 4, /**< The address at which a pointer's memory may be accessed on the host */
677 CU_POINTER_ATTRIBUTE_P2P_TOKENS = 5, /**< A pair of tokens for use with the nv-p2p.h Linux kernel interface */
678 CU_POINTER_ATTRIBUTE_SYNC_MEMOPS = 6, /**< Synchronize every synchronous memory operation initiated on this region */
679 CU_POINTER_ATTRIBUTE_BUFFER_ID = 7, /**< A process-wide unique ID for an allocated memory region*/
680 CU_POINTER_ATTRIBUTE_IS_MANAGED = 8, /**< Indicates if the pointer points to managed memory */
681 CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL = 9, /**< A device ordinal of a device on which a pointer was allocated or registered */
682 CU_POINTER_ATTRIBUTE_IS_LEGACY_CUDA_IPC_CAPABLE = 10, /**< 1 if this pointer maps to an allocation that is suitable for ::cudaIpcGetMemHandle, 0 otherwise **/
683 CU_POINTER_ATTRIBUTE_RANGE_START_ADDR = 11, /**< Starting address for this requested pointer */
684 CU_POINTER_ATTRIBUTE_RANGE_SIZE = 12, /**< Size of the address range for this requested pointer */
685 CU_POINTER_ATTRIBUTE_MAPPED = 13, /**< 1 if this pointer is in a valid address range that is mapped to a backing allocation, 0 otherwise **/
686 CU_POINTER_ATTRIBUTE_ALLOWED_HANDLE_TYPES = 14, /**< Bitmask of allowed ::CUmemAllocationHandleType for this allocation **/
687 CU_POINTER_ATTRIBUTE_IS_GPU_DIRECT_RDMA_CAPABLE = 15, /**< 1 if the memory this pointer is referencing can be used with the GPUDirect RDMA API **/
688 CU_POINTER_ATTRIBUTE_ACCESS_FLAGS = 16, /**< Returns the access flags the device associated with the current context has on the corresponding memory referenced by the pointer given */
689 CU_POINTER_ATTRIBUTE_MEMPOOL_HANDLE = 17 /**< Returns the mempool handle for the allocation if it was allocated from a mempool. Otherwise returns NULL. **/
690} CUpointer_attribute;
691
692/**
693 * Function properties
694 */
695typedef enum CUfunction_attribute_enum {
696 /**
697 * The maximum number of threads per block, beyond which a launch of the
698 * function would fail. This number depends on both the function and the
699 * device on which the function is currently loaded.
700 */
701 CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK = 0,
702
703 /**
704 * The size in bytes of statically-allocated shared memory required by
705 * this function. This does not include dynamically-allocated shared
706 * memory requested by the user at runtime.
707 */
708 CU_FUNC_ATTRIBUTE_SHARED_SIZE_BYTES = 1,
709
710 /**
711 * The size in bytes of user-allocated constant memory required by this
712 * function.
713 */
714 CU_FUNC_ATTRIBUTE_CONST_SIZE_BYTES = 2,
715
716 /**
717 * The size in bytes of local memory used by each thread of this function.
718 */
719 CU_FUNC_ATTRIBUTE_LOCAL_SIZE_BYTES = 3,
720
721 /**
722 * The number of registers used by each thread of this function.
723 */
724 CU_FUNC_ATTRIBUTE_NUM_REGS = 4,
725
726 /**
727 * The PTX virtual architecture version for which the function was
728 * compiled. This value is the major PTX version * 10 + the minor PTX
729 * version, so a PTX version 1.3 function would return the value 13.
730 * Note that this may return the undefined value of 0 for cubins
731 * compiled prior to CUDA 3.0.
732 */
733 CU_FUNC_ATTRIBUTE_PTX_VERSION = 5,
734
735 /**
736 * The binary architecture version for which the function was compiled.
737 * This value is the major binary version * 10 + the minor binary version,
738 * so a binary version 1.3 function would return the value 13. Note that
739 * this will return a value of 10 for legacy cubins that do not have a
740 * properly-encoded binary architecture version.
741 */
742 CU_FUNC_ATTRIBUTE_BINARY_VERSION = 6,
743
744 /**
745 * The attribute to indicate whether the function has been compiled with
746 * user specified option "-Xptxas --dlcm=ca" set .
747 */
748 CU_FUNC_ATTRIBUTE_CACHE_MODE_CA = 7,
749
750 /**
751 * The maximum size in bytes of dynamically-allocated shared memory that can be used by
752 * this function. If the user-specified dynamic shared memory size is larger than this
753 * value, the launch will fail.
754 * See ::cuFuncSetAttribute
755 */
756 CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES = 8,
757
758 /**
759 * On devices where the L1 cache and shared memory use the same hardware resources,
760 * this sets the shared memory carveout preference, in percent of the total shared memory.
761 * Refer to ::CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_MULTIPROCESSOR.
762 * This is only a hint, and the driver can choose a different ratio if required to execute the function.
763 * See ::cuFuncSetAttribute
764 */
765 CU_FUNC_ATTRIBUTE_PREFERRED_SHARED_MEMORY_CARVEOUT = 9,
766
767 CU_FUNC_ATTRIBUTE_MAX
768} CUfunction_attribute;
769
770/**
771 * Function cache configurations
772 */
773typedef enum CUfunc_cache_enum {
774 CU_FUNC_CACHE_PREFER_NONE = 0x00, /**< no preference for shared memory or L1 (default) */
775 CU_FUNC_CACHE_PREFER_SHARED = 0x01, /**< prefer larger shared memory and smaller L1 cache */
776 CU_FUNC_CACHE_PREFER_L1 = 0x02, /**< prefer larger L1 cache and smaller shared memory */
777 CU_FUNC_CACHE_PREFER_EQUAL = 0x03 /**< prefer equal sized L1 cache and shared memory */
778} CUfunc_cache;
779
780/**
781 * Shared memory configurations
782 */
783typedef enum CUsharedconfig_enum {
784 CU_SHARED_MEM_CONFIG_DEFAULT_BANK_SIZE = 0x00, /**< set default shared memory bank size */
785 CU_SHARED_MEM_CONFIG_FOUR_BYTE_BANK_SIZE = 0x01, /**< set shared memory bank width to four bytes */
786 CU_SHARED_MEM_CONFIG_EIGHT_BYTE_BANK_SIZE = 0x02 /**< set shared memory bank width to eight bytes */
787} CUsharedconfig;
788
789/**
790 * Shared memory carveout configurations. These may be passed to ::cuFuncSetAttribute
791 */
792typedef enum CUshared_carveout_enum {
793 CU_SHAREDMEM_CARVEOUT_DEFAULT = -1, /**< No preference for shared memory or L1 (default) */
794 CU_SHAREDMEM_CARVEOUT_MAX_SHARED = 100, /**< Prefer maximum available shared memory, minimum L1 cache */
795 CU_SHAREDMEM_CARVEOUT_MAX_L1 = 0 /**< Prefer maximum available L1 cache, minimum shared memory */
796} CUshared_carveout;
797
798/**
799 * Memory types
800 */
801typedef enum CUmemorytype_enum {
802 CU_MEMORYTYPE_HOST = 0x01, /**< Host memory */
803 CU_MEMORYTYPE_DEVICE = 0x02, /**< Device memory */
804 CU_MEMORYTYPE_ARRAY = 0x03, /**< Array memory */
805 CU_MEMORYTYPE_UNIFIED = 0x04 /**< Unified device or host memory */
806} CUmemorytype;
807
808/**
809 * Compute Modes
810 */
811typedef enum CUcomputemode_enum {
812 CU_COMPUTEMODE_DEFAULT = 0, /**< Default compute mode (Multiple contexts allowed per device) */
813 CU_COMPUTEMODE_PROHIBITED = 2, /**< Compute-prohibited mode (No contexts can be created on this device at this time) */
814 CU_COMPUTEMODE_EXCLUSIVE_PROCESS = 3 /**< Compute-exclusive-process mode (Only one context used by a single process can be present on this device at a time) */
815} CUcomputemode;
816
817/**
818 * Memory advise values
819 */
820typedef enum CUmem_advise_enum {
821 CU_MEM_ADVISE_SET_READ_MOSTLY = 1, /**< Data will mostly be read and only occasionally be written to */
822 CU_MEM_ADVISE_UNSET_READ_MOSTLY = 2, /**< Undo the effect of ::CU_MEM_ADVISE_SET_READ_MOSTLY */
823 CU_MEM_ADVISE_SET_PREFERRED_LOCATION = 3, /**< Set the preferred location for the data as the specified device */
824 CU_MEM_ADVISE_UNSET_PREFERRED_LOCATION = 4, /**< Clear the preferred location for the data */
825 CU_MEM_ADVISE_SET_ACCESSED_BY = 5, /**< Data will be accessed by the specified device, so prevent page faults as much as possible */
826 CU_MEM_ADVISE_UNSET_ACCESSED_BY = 6 /**< Let the Unified Memory subsystem decide on the page faulting policy for the specified device */
827} CUmem_advise;
828
829typedef enum CUmem_range_attribute_enum {
830 CU_MEM_RANGE_ATTRIBUTE_READ_MOSTLY = 1, /**< Whether the range will mostly be read and only occasionally be written to */
831 CU_MEM_RANGE_ATTRIBUTE_PREFERRED_LOCATION = 2, /**< The preferred location of the range */
832 CU_MEM_RANGE_ATTRIBUTE_ACCESSED_BY = 3, /**< Memory range has ::CU_MEM_ADVISE_SET_ACCESSED_BY set for specified device */
833 CU_MEM_RANGE_ATTRIBUTE_LAST_PREFETCH_LOCATION = 4 /**< The last location to which the range was prefetched */
834} CUmem_range_attribute;
835
836/**
837 * Online compiler and linker options
838 */
839typedef enum CUjit_option_enum
840{
841 /**
842 * Max number of registers that a thread may use.\n
843 * Option type: unsigned int\n
844 * Applies to: compiler only
845 */
846 CU_JIT_MAX_REGISTERS = 0,
847
848 /**
849 * IN: Specifies minimum number of threads per block to target compilation
850 * for\n
851 * OUT: Returns the number of threads the compiler actually targeted.
852 * This restricts the resource utilization of the compiler (e.g. max
853 * registers) such that a block with the given number of threads should be
854 * able to launch based on register limitations. Note, this option does not
855 * currently take into account any other resource limitations, such as
856 * shared memory utilization.\n
857 * Cannot be combined with ::CU_JIT_TARGET.\n
858 * Option type: unsigned int\n
859 * Applies to: compiler only
860 */
861 CU_JIT_THREADS_PER_BLOCK,
862
863 /**
864 * Overwrites the option value with the total wall clock time, in
865 * milliseconds, spent in the compiler and linker\n
866 * Option type: float\n
867 * Applies to: compiler and linker
868 */
869 CU_JIT_WALL_TIME,
870
871 /**
872 * Pointer to a buffer in which to print any log messages
873 * that are informational in nature (the buffer size is specified via
874 * option ::CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES)\n
875 * Option type: char *\n
876 * Applies to: compiler and linker
877 */
878 CU_JIT_INFO_LOG_BUFFER,
879
880 /**
881 * IN: Log buffer size in bytes. Log messages will be capped at this size
882 * (including null terminator)\n
883 * OUT: Amount of log buffer filled with messages\n
884 * Option type: unsigned int\n
885 * Applies to: compiler and linker
886 */
887 CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES,
888
889 /**
890 * Pointer to a buffer in which to print any log messages that
891 * reflect errors (the buffer size is specified via option
892 * ::CU_JIT_ERROR_LOG_BUFFER_SIZE_BYTES)\n
893 * Option type: char *\n
894 * Applies to: compiler and linker
895 */
896 CU_JIT_ERROR_LOG_BUFFER,
897
898 /**
899 * IN: Log buffer size in bytes. Log messages will be capped at this size
900 * (including null terminator)\n
901 * OUT: Amount of log buffer filled with messages\n
902 * Option type: unsigned int\n
903 * Applies to: compiler and linker
904 */
905 CU_JIT_ERROR_LOG_BUFFER_SIZE_BYTES,
906
907 /**
908 * Level of optimizations to apply to generated code (0 - 4), with 4
909 * being the default and highest level of optimizations.\n
910 * Option type: unsigned int\n
911 * Applies to: compiler only
912 */
913 CU_JIT_OPTIMIZATION_LEVEL,
914
915 /**
916 * No option value required. Determines the target based on the current
917 * attached context (default)\n
918 * Option type: No option value needed\n
919 * Applies to: compiler and linker
920 */
921 CU_JIT_TARGET_FROM_CUCONTEXT,
922
923 /**
924 * Target is chosen based on supplied ::CUjit_target. Cannot be
925 * combined with ::CU_JIT_THREADS_PER_BLOCK.\n
926 * Option type: unsigned int for enumerated type ::CUjit_target\n
927 * Applies to: compiler and linker
928 */
929 CU_JIT_TARGET,
930
931 /**
932 * Specifies choice of fallback strategy if matching cubin is not found.
933 * Choice is based on supplied ::CUjit_fallback. This option cannot be
934 * used with cuLink* APIs as the linker requires exact matches.\n
935 * Option type: unsigned int for enumerated type ::CUjit_fallback\n
936 * Applies to: compiler only
937 */
938 CU_JIT_FALLBACK_STRATEGY,
939
940 /**
941 * Specifies whether to create debug information in output (-g)
942 * (0: false, default)\n
943 * Option type: int\n
944 * Applies to: compiler and linker
945 */
946 CU_JIT_GENERATE_DEBUG_INFO,
947
948 /**
949 * Generate verbose log messages (0: false, default)\n
950 * Option type: int\n
951 * Applies to: compiler and linker
952 */
953 CU_JIT_LOG_VERBOSE,
954
955 /**
956 * Generate line number information (-lineinfo) (0: false, default)\n
957 * Option type: int\n
958 * Applies to: compiler only
959 */
960 CU_JIT_GENERATE_LINE_INFO,
961
962 /**
963 * Specifies whether to enable caching explicitly (-dlcm) \n
964 * Choice is based on supplied ::CUjit_cacheMode_enum.\n
965 * Option type: unsigned int for enumerated type ::CUjit_cacheMode_enum\n
966 * Applies to: compiler only
967 */
968 CU_JIT_CACHE_MODE,
969
970 /**
971 * The below jit options are used for internal purposes only, in this version of CUDA
972 */
973 CU_JIT_NEW_SM3X_OPT,
974 CU_JIT_FAST_COMPILE,
975
976 /**
977 * Array of device symbol names that will be relocated to the corresponding
978 * host addresses stored in ::CU_JIT_GLOBAL_SYMBOL_ADDRESSES.\n
979 * Must contain ::CU_JIT_GLOBAL_SYMBOL_COUNT entries.\n
980 * When loading a device module, driver will relocate all encountered
981 * unresolved symbols to the host addresses.\n
982 * It is only allowed to register symbols that correspond to unresolved
983 * global variables.\n
984 * It is illegal to register the same device symbol at multiple addresses.\n
985 * Option type: const char **\n
986 * Applies to: dynamic linker only
987 */
988 CU_JIT_GLOBAL_SYMBOL_NAMES,
989
990 /**
991 * Array of host addresses that will be used to relocate corresponding
992 * device symbols stored in ::CU_JIT_GLOBAL_SYMBOL_NAMES.\n
993 * Must contain ::CU_JIT_GLOBAL_SYMBOL_COUNT entries.\n
994 * Option type: void **\n
995 * Applies to: dynamic linker only
996 */
997 CU_JIT_GLOBAL_SYMBOL_ADDRESSES,
998
999 /**
1000 * Number of entries in ::CU_JIT_GLOBAL_SYMBOL_NAMES and
1001 * ::CU_JIT_GLOBAL_SYMBOL_ADDRESSES arrays.\n
1002 * Option type: unsigned int\n
1003 * Applies to: dynamic linker only
1004 */
1005 CU_JIT_GLOBAL_SYMBOL_COUNT,
1006
1007 /**
1008 * Enable link-time optimization (-dlto) for device code (0: false, default)\n
1009 * Option type: int\n
1010 * Applies to: compiler and linker
1011 */
1012 CU_JIT_LTO,
1013
1014 /**
1015 * Control single-precision denormals (-ftz) support (0: false, default).
1016 * 1 : flushes denormal values to zero
1017 * 0 : preserves denormal values
1018 * Option type: int\n
1019 * Applies to: link-time optimization specified with CU_JIT_LTO
1020 */
1021 CU_JIT_FTZ,
1022
1023 /**
1024 * Control single-precision floating-point division and reciprocals
1025 * (-prec-div) support (1: true, default).
1026 * 1 : Enables the IEEE round-to-nearest mode
1027 * 0 : Enables the fast approximation mode
1028 * Option type: int\n
1029 * Applies to: link-time optimization specified with CU_JIT_LTO
1030 */
1031 CU_JIT_PREC_DIV,
1032
1033 /**
1034 * Control single-precision floating-point square root
1035 * (-prec-sqrt) support (1: true, default).
1036 * 1 : Enables the IEEE round-to-nearest mode
1037 * 0 : Enables the fast approximation mode
1038 * Option type: int\n
1039 * Applies to: link-time optimization specified with CU_JIT_LTO
1040 */
1041 CU_JIT_PREC_SQRT,
1042
1043 /**
1044 * Enable/Disable the contraction of floating-point multiplies
1045 * and adds/subtracts into floating-point multiply-add (-fma)
1046 * operations (1: Enable, default; 0: Disable).
1047 * Option type: int\n
1048 * Applies to: link-time optimization specified with CU_JIT_LTO
1049 */
1050 CU_JIT_FMA,
1051
1052 CU_JIT_NUM_OPTIONS
1053
1054} CUjit_option;
1055
1056/**
1057 * Online compilation targets
1058 */
1059typedef enum CUjit_target_enum
1060{
1061 CU_TARGET_COMPUTE_20 = 20, /**< Compute device class 2.0 */
1062 CU_TARGET_COMPUTE_21 = 21, /**< Compute device class 2.1 */
1063 CU_TARGET_COMPUTE_30 = 30, /**< Compute device class 3.0 */
1064 CU_TARGET_COMPUTE_32 = 32, /**< Compute device class 3.2 */
1065 CU_TARGET_COMPUTE_35 = 35, /**< Compute device class 3.5 */
1066 CU_TARGET_COMPUTE_37 = 37, /**< Compute device class 3.7 */
1067 CU_TARGET_COMPUTE_50 = 50, /**< Compute device class 5.0 */
1068 CU_TARGET_COMPUTE_52 = 52, /**< Compute device class 5.2 */
1069 CU_TARGET_COMPUTE_53 = 53, /**< Compute device class 5.3 */
1070 CU_TARGET_COMPUTE_60 = 60, /**< Compute device class 6.0.*/
1071 CU_TARGET_COMPUTE_61 = 61, /**< Compute device class 6.1.*/
1072 CU_TARGET_COMPUTE_62 = 62, /**< Compute device class 6.2.*/
1073 CU_TARGET_COMPUTE_70 = 70, /**< Compute device class 7.0.*/
1074 CU_TARGET_COMPUTE_72 = 72, /**< Compute device class 7.2.*/
1075 CU_TARGET_COMPUTE_75 = 75, /**< Compute device class 7.5.*/
1076 CU_TARGET_COMPUTE_80 = 80, /**< Compute device class 8.0.*/
1077 CU_TARGET_COMPUTE_86 = 86 /**< Compute device class 8.6.*/
1078} CUjit_target;
1079
1080/**
1081 * Cubin matching fallback strategies
1082 */
1083typedef enum CUjit_fallback_enum
1084{
1085 CU_PREFER_PTX = 0, /**< Prefer to compile ptx if exact binary match not found */
1086
1087 CU_PREFER_BINARY /**< Prefer to fall back to compatible binary code if exact match not found */
1088
1089} CUjit_fallback;
1090
1091/**
1092 * Caching modes for dlcm
1093 */
1094typedef enum CUjit_cacheMode_enum
1095{
1096 CU_JIT_CACHE_OPTION_NONE = 0, /**< Compile with no -dlcm flag specified */
1097 CU_JIT_CACHE_OPTION_CG, /**< Compile with L1 cache disabled */
1098 CU_JIT_CACHE_OPTION_CA /**< Compile with L1 cache enabled */
1099} CUjit_cacheMode;
1100
1101/**
1102 * Device code formats
1103 */
1104typedef enum CUjitInputType_enum
1105{
1106 /**
1107 * Compiled device-class-specific device code\n
1108 * Applicable options: none
1109 */
1110 CU_JIT_INPUT_CUBIN = 0,
1111
1112 /**
1113 * PTX source code\n
1114 * Applicable options: PTX compiler options
1115 */
1116 CU_JIT_INPUT_PTX,
1117
1118 /**
1119 * Bundle of multiple cubins and/or PTX of some device code\n
1120 * Applicable options: PTX compiler options, ::CU_JIT_FALLBACK_STRATEGY
1121 */
1122 CU_JIT_INPUT_FATBINARY,
1123
1124 /**
1125 * Host object with embedded device code\n
1126 * Applicable options: PTX compiler options, ::CU_JIT_FALLBACK_STRATEGY
1127 */
1128 CU_JIT_INPUT_OBJECT,
1129
1130 /**
1131 * Archive of host objects with embedded device code\n
1132 * Applicable options: PTX compiler options, ::CU_JIT_FALLBACK_STRATEGY
1133 */
1134 CU_JIT_INPUT_LIBRARY,
1135
1136 /**
1137 * High-level intermediate code for link-time optimization\n
1138 * Applicable options: NVVM compiler options, PTX compiler options
1139 */
1140 CU_JIT_INPUT_NVVM,
1141
1142 CU_JIT_NUM_INPUT_TYPES
1143} CUjitInputType;
1144
1145typedef struct CUlinkState_st *CUlinkState;
1146
1147/**
1148 * Flags to register a graphics resource
1149 */
1150typedef enum CUgraphicsRegisterFlags_enum {
1151 CU_GRAPHICS_REGISTER_FLAGS_NONE = 0x00,
1152 CU_GRAPHICS_REGISTER_FLAGS_READ_ONLY = 0x01,
1153 CU_GRAPHICS_REGISTER_FLAGS_WRITE_DISCARD = 0x02,
1154 CU_GRAPHICS_REGISTER_FLAGS_SURFACE_LDST = 0x04,
1155 CU_GRAPHICS_REGISTER_FLAGS_TEXTURE_GATHER = 0x08
1156} CUgraphicsRegisterFlags;
1157
1158/**
1159 * Flags for mapping and unmapping interop resources
1160 */
1161typedef enum CUgraphicsMapResourceFlags_enum {
1162 CU_GRAPHICS_MAP_RESOURCE_FLAGS_NONE = 0x00,
1163 CU_GRAPHICS_MAP_RESOURCE_FLAGS_READ_ONLY = 0x01,
1164 CU_GRAPHICS_MAP_RESOURCE_FLAGS_WRITE_DISCARD = 0x02
1165} CUgraphicsMapResourceFlags;
1166
1167/**
1168 * Array indices for cube faces
1169 */
1170typedef enum CUarray_cubemap_face_enum {
1171 CU_CUBEMAP_FACE_POSITIVE_X = 0x00, /**< Positive X face of cubemap */
1172 CU_CUBEMAP_FACE_NEGATIVE_X = 0x01, /**< Negative X face of cubemap */
1173 CU_CUBEMAP_FACE_POSITIVE_Y = 0x02, /**< Positive Y face of cubemap */
1174 CU_CUBEMAP_FACE_NEGATIVE_Y = 0x03, /**< Negative Y face of cubemap */
1175 CU_CUBEMAP_FACE_POSITIVE_Z = 0x04, /**< Positive Z face of cubemap */
1176 CU_CUBEMAP_FACE_NEGATIVE_Z = 0x05 /**< Negative Z face of cubemap */
1177} CUarray_cubemap_face;
1178
1179/**
1180 * Limits
1181 */
1182typedef enum CUlimit_enum {
1183 CU_LIMIT_STACK_SIZE = 0x00, /**< GPU thread stack size */
1184 CU_LIMIT_PRINTF_FIFO_SIZE = 0x01, /**< GPU printf FIFO size */
1185 CU_LIMIT_MALLOC_HEAP_SIZE = 0x02, /**< GPU malloc heap size */
1186 CU_LIMIT_DEV_RUNTIME_SYNC_DEPTH = 0x03, /**< GPU device runtime launch synchronize depth */
1187 CU_LIMIT_DEV_RUNTIME_PENDING_LAUNCH_COUNT = 0x04, /**< GPU device runtime pending launch count */
1188 CU_LIMIT_MAX_L2_FETCH_GRANULARITY = 0x05, /**< A value between 0 and 128 that indicates the maximum fetch granularity of L2 (in Bytes). This is a hint */
1189 CU_LIMIT_PERSISTING_L2_CACHE_SIZE = 0x06, /**< A size in bytes for L2 persisting lines cache size */
1190 CU_LIMIT_MAX
1191} CUlimit;
1192
1193/**
1194 * Resource types
1195 */
1196typedef enum CUresourcetype_enum {
1197 CU_RESOURCE_TYPE_ARRAY = 0x00, /**< Array resource */
1198 CU_RESOURCE_TYPE_MIPMAPPED_ARRAY = 0x01, /**< Mipmapped array resource */
1199 CU_RESOURCE_TYPE_LINEAR = 0x02, /**< Linear resource */
1200 CU_RESOURCE_TYPE_PITCH2D = 0x03 /**< Pitch 2D resource */
1201} CUresourcetype;
1202
1203#ifdef _WIN32
1204#define CUDA_CB __stdcall
1205#else
1206#define CUDA_CB
1207#endif
1208
1209/**
1210 * CUDA host function
1211 * \param userData Argument value passed to the function
1212 */
1213typedef void (CUDA_CB *CUhostFn)(void *userData);
1214
1215/**
1216 * Specifies performance hint with ::CUaccessPolicyWindow for hitProp and missProp members.
1217 */
1218typedef enum CUaccessProperty_enum {
1219 CU_ACCESS_PROPERTY_NORMAL = 0, /**< Normal cache persistence. */
1220 CU_ACCESS_PROPERTY_STREAMING = 1, /**< Streaming access is less likely to persit from cache. */
1221 CU_ACCESS_PROPERTY_PERSISTING = 2 /**< Persisting access is more likely to persist in cache.*/
1222} CUaccessProperty;
1223
1224/**
1225 * Specifies an access policy for a window, a contiguous extent of memory
1226 * beginning at base_ptr and ending at base_ptr + num_bytes.
1227 * num_bytes is limited by CU_DEVICE_ATTRIBUTE_MAX_ACCESS_POLICY_WINDOW_SIZE.
1228 * Partition into many segments and assign segments such that:
1229 * sum of "hit segments" / window == approx. ratio.
1230 * sum of "miss segments" / window == approx 1-ratio.
1231 * Segments and ratio specifications are fitted to the capabilities of
1232 * the architecture.
1233 * Accesses in a hit segment apply the hitProp access policy.
1234 * Accesses in a miss segment apply the missProp access policy.
1235 */
1236typedef struct CUaccessPolicyWindow_st {
1237 void *base_ptr; /**< Starting address of the access policy window. CUDA driver may align it. */
1238 size_t num_bytes; /**< Size in bytes of the window policy. CUDA driver may restrict the maximum size and alignment. */
1239 float hitRatio; /**< hitRatio specifies percentage of lines assigned hitProp, rest are assigned missProp. */
1240 CUaccessProperty hitProp; /**< ::CUaccessProperty set for hit. */
1241 CUaccessProperty missProp; /**< ::CUaccessProperty set for miss. Must be either NORMAL or STREAMING */
1242} CUaccessPolicyWindow_v1;
1243typedef CUaccessPolicyWindow_v1 CUaccessPolicyWindow;
1244
1245/**
1246 * GPU kernel node parameters
1247 */
1248typedef struct CUDA_KERNEL_NODE_PARAMS_st {
1249 CUfunction func; /**< Kernel to launch */
1250 unsigned int gridDimX; /**< Width of grid in blocks */
1251 unsigned int gridDimY; /**< Height of grid in blocks */
1252 unsigned int gridDimZ; /**< Depth of grid in blocks */
1253 unsigned int blockDimX; /**< X dimension of each thread block */
1254 unsigned int blockDimY; /**< Y dimension of each thread block */
1255 unsigned int blockDimZ; /**< Z dimension of each thread block */
1256 unsigned int sharedMemBytes; /**< Dynamic shared-memory size per thread block in bytes */
1257 void **kernelParams; /**< Array of pointers to kernel parameters */
1258 void **extra; /**< Extra options */
1259} CUDA_KERNEL_NODE_PARAMS_v1;
1260typedef CUDA_KERNEL_NODE_PARAMS_v1 CUDA_KERNEL_NODE_PARAMS;
1261
1262/**
1263 * Memset node parameters
1264 */
1265typedef struct CUDA_MEMSET_NODE_PARAMS_st {
1266 CUdeviceptr dst; /**< Destination device pointer */
1267 size_t pitch; /**< Pitch of destination device pointer. Unused if height is 1 */
1268 unsigned int value; /**< Value to be set */
1269 unsigned int elementSize; /**< Size of each element in bytes. Must be 1, 2, or 4. */
1270 size_t width; /**< Width of the row in elements */
1271 size_t height; /**< Number of rows */
1272} CUDA_MEMSET_NODE_PARAMS_v1;
1273typedef CUDA_MEMSET_NODE_PARAMS_v1 CUDA_MEMSET_NODE_PARAMS;
1274
1275/**
1276 * Host node parameters
1277 */
1278typedef struct CUDA_HOST_NODE_PARAMS_st {
1279 CUhostFn fn; /**< The function to call when the node executes */
1280 void* userData; /**< Argument to pass to the function */
1281} CUDA_HOST_NODE_PARAMS_v1;
1282typedef CUDA_HOST_NODE_PARAMS_v1 CUDA_HOST_NODE_PARAMS;
1283
1284/**
1285 * Graph node types
1286 */
1287typedef enum CUgraphNodeType_enum {
1288 CU_GRAPH_NODE_TYPE_KERNEL = 0, /**< GPU kernel node */
1289 CU_GRAPH_NODE_TYPE_MEMCPY = 1, /**< Memcpy node */
1290 CU_GRAPH_NODE_TYPE_MEMSET = 2, /**< Memset node */
1291 CU_GRAPH_NODE_TYPE_HOST = 3, /**< Host (executable) node */
1292 CU_GRAPH_NODE_TYPE_GRAPH = 4, /**< Node which executes an embedded graph */
1293 CU_GRAPH_NODE_TYPE_EMPTY = 5, /**< Empty (no-op) node */
1294 CU_GRAPH_NODE_TYPE_WAIT_EVENT = 6, /**< External event wait node */
1295 CU_GRAPH_NODE_TYPE_EVENT_RECORD = 7, /**< External event record node */
1296 CU_GRAPH_NODE_TYPE_EXT_SEMAS_SIGNAL = 8, /**< External semaphore signal node */
1297 CU_GRAPH_NODE_TYPE_EXT_SEMAS_WAIT = 9, /**< External semaphore wait node */
1298 CU_GRAPH_NODE_TYPE_MEM_ALLOC = 10,/**< Memory Allocation Node */
1299 CU_GRAPH_NODE_TYPE_MEM_FREE = 11 /**< Memory Free Node */
1300} CUgraphNodeType;
1301
1302typedef enum CUsynchronizationPolicy_enum {
1303 CU_SYNC_POLICY_AUTO = 1,
1304 CU_SYNC_POLICY_SPIN = 2,
1305 CU_SYNC_POLICY_YIELD = 3,
1306 CU_SYNC_POLICY_BLOCKING_SYNC = 4
1307} CUsynchronizationPolicy;
1308
1309/**
1310 * Graph kernel node Attributes
1311 */
1312typedef enum CUkernelNodeAttrID_enum {
1313 CU_KERNEL_NODE_ATTRIBUTE_ACCESS_POLICY_WINDOW = 1, /**< Identifier for ::CUkernelNodeAttrValue::accessPolicyWindow. */
1314 CU_KERNEL_NODE_ATTRIBUTE_COOPERATIVE = 2 /**< Allows a kernel node to be cooperative (see ::cuLaunchCooperativeKernel). */
1315} CUkernelNodeAttrID;
1316
1317/**
1318 * Graph kernel node attributes union, used with ::cuKernelNodeSetAttribute/::cuKernelNodeGetAttribute
1319 */
1320typedef union CUkernelNodeAttrValue_union {
1321 CUaccessPolicyWindow accessPolicyWindow; /**< Attribute ::CUaccessPolicyWindow. */
1322 int cooperative; /**< Nonzero indicates a cooperative kernel (see ::cuLaunchCooperativeKernel). */
1323} CUkernelNodeAttrValue_v1;
1324typedef CUkernelNodeAttrValue_v1 CUkernelNodeAttrValue;
1325
1326/**
1327 * Possible stream capture statuses returned by ::cuStreamIsCapturing
1328 */
1329typedef enum CUstreamCaptureStatus_enum {
1330 CU_STREAM_CAPTURE_STATUS_NONE = 0, /**< Stream is not capturing */
1331 CU_STREAM_CAPTURE_STATUS_ACTIVE = 1, /**< Stream is actively capturing */
1332 CU_STREAM_CAPTURE_STATUS_INVALIDATED = 2 /**< Stream is part of a capture sequence that
1333 has been invalidated, but not terminated */
1334} CUstreamCaptureStatus;
1335
1336/**
1337 * Possible modes for stream capture thread interactions. For more details see
1338 * ::cuStreamBeginCapture and ::cuThreadExchangeStreamCaptureMode
1339 */
1340typedef enum CUstreamCaptureMode_enum {
1341 CU_STREAM_CAPTURE_MODE_GLOBAL = 0,
1342 CU_STREAM_CAPTURE_MODE_THREAD_LOCAL = 1,
1343 CU_STREAM_CAPTURE_MODE_RELAXED = 2
1344} CUstreamCaptureMode;
1345
1346/**
1347 * Stream Attributes
1348 */
1349typedef enum CUstreamAttrID_enum {
1350 CU_STREAM_ATTRIBUTE_ACCESS_POLICY_WINDOW = 1, /**< Identifier for ::CUstreamAttrValue::accessPolicyWindow. */
1351 CU_STREAM_ATTRIBUTE_SYNCHRONIZATION_POLICY = 3 /**< ::CUsynchronizationPolicy for work queued up in this stream */
1352} CUstreamAttrID;
1353
1354/**
1355 * Stream attributes union, used with ::cuStreamSetAttribute/::cuStreamGetAttribute
1356 */
1357typedef union CUstreamAttrValue_union {
1358 CUaccessPolicyWindow accessPolicyWindow; /**< Attribute ::CUaccessPolicyWindow. */
1359 CUsynchronizationPolicy syncPolicy; /**< Value for ::CU_STREAM_ATTRIBUTE_SYNCHRONIZATION_POLICY. */
1360} CUstreamAttrValue_v1;
1361typedef CUstreamAttrValue_v1 CUstreamAttrValue;
1362
1363/**
1364 * Flags to specify search options. For more details see ::cuGetProcAddress
1365 */
1366typedef enum CUdriverProcAddress_flags_enum {
1367 CU_GET_PROC_ADDRESS_DEFAULT = 0, /**< Default search mode for driver symbols. */
1368 CU_GET_PROC_ADDRESS_LEGACY_STREAM = 1 << 0, /**< Search for legacy versions of driver symbols. */
1369 CU_GET_PROC_ADDRESS_PER_THREAD_DEFAULT_STREAM = 1 << 1 /**< Search for per-thread versions of driver symbols. */
1370} CUdriverProcAddress_flags;
1371
1372/**
1373 * Execution Affinity Types
1374 */
1375typedef enum CUexecAffinityType_enum {
1376 CU_EXEC_AFFINITY_TYPE_SM_COUNT = 0, /**< Create a context with limited SMs. */
1377 CU_EXEC_AFFINITY_TYPE_MAX
1378} CUexecAffinityType;
1379
1380/**
1381 * Value for ::CU_EXEC_AFFINITY_TYPE_SM_COUNT
1382 */
1383typedef struct CUexecAffinitySmCount_st {
1384 unsigned int val; /**< The number of SMs the context is limited to use. */
1385} CUexecAffinitySmCount_v1;
1386typedef CUexecAffinitySmCount_v1 CUexecAffinitySmCount;
1387
1388/**
1389 * Execution Affinity Parameters
1390 */
1391typedef struct CUexecAffinityParam_st {
1392 CUexecAffinityType type;
1393 union {
1394 CUexecAffinitySmCount smCount; /** Value for ::CU_EXEC_AFFINITY_TYPE_SM_COUNT */
1395 } param;
1396} CUexecAffinityParam_v1;
1397typedef CUexecAffinityParam_v1 CUexecAffinityParam;
1398
1399/**
1400 * Error codes
1401 */
1402typedef enum cudaError_enum {
1403 /**
1404 * The API call returned with no errors. In the case of query calls, this
1405 * also means that the operation being queried is complete (see
1406 * ::cuEventQuery() and ::cuStreamQuery()).
1407 */
1408 CUDA_SUCCESS = 0,
1409
1410 /**
1411 * This indicates that one or more of the parameters passed to the API call
1412 * is not within an acceptable range of values.
1413 */
1414 CUDA_ERROR_INVALID_VALUE = 1,
1415
1416 /**
1417 * The API call failed because it was unable to allocate enough memory to
1418 * perform the requested operation.
1419 */
1420 CUDA_ERROR_OUT_OF_MEMORY = 2,
1421
1422 /**
1423 * This indicates that the CUDA driver has not been initialized with
1424 * ::cuInit() or that initialization has failed.
1425 */
1426 CUDA_ERROR_NOT_INITIALIZED = 3,
1427
1428 /**
1429 * This indicates that the CUDA driver is in the process of shutting down.
1430 */
1431 CUDA_ERROR_DEINITIALIZED = 4,
1432
1433 /**
1434 * This indicates profiler is not initialized for this run. This can
1435 * happen when the application is running with external profiling tools
1436 * like visual profiler.
1437 */
1438 CUDA_ERROR_PROFILER_DISABLED = 5,
1439
1440 /**
1441 * \deprecated
1442 * This error return is deprecated as of CUDA 5.0. It is no longer an error
1443 * to attempt to enable/disable the profiling via ::cuProfilerStart or
1444 * ::cuProfilerStop without initialization.
1445 */
1446 CUDA_ERROR_PROFILER_NOT_INITIALIZED = 6,
1447
1448 /**
1449 * \deprecated
1450 * This error return is deprecated as of CUDA 5.0. It is no longer an error
1451 * to call cuProfilerStart() when profiling is already enabled.
1452 */
1453 CUDA_ERROR_PROFILER_ALREADY_STARTED = 7,
1454
1455 /**
1456 * \deprecated
1457 * This error return is deprecated as of CUDA 5.0. It is no longer an error
1458 * to call cuProfilerStop() when profiling is already disabled.
1459 */
1460 CUDA_ERROR_PROFILER_ALREADY_STOPPED = 8,
1461
1462 /**
1463 * This indicates that the CUDA driver that the application has loaded is a
1464 * stub library. Applications that run with the stub rather than a real
1465 * driver loaded will result in CUDA API returning this error.
1466 */
1467 CUDA_ERROR_STUB_LIBRARY = 34,
1468
1469 /**
1470 * This indicates that no CUDA-capable devices were detected by the installed
1471 * CUDA driver.
1472 */
1473 CUDA_ERROR_NO_DEVICE = 100,
1474
1475 /**
1476 * This indicates that the device ordinal supplied by the user does not
1477 * correspond to a valid CUDA device or that the action requested is
1478 * invalid for the specified device.
1479 */
1480 CUDA_ERROR_INVALID_DEVICE = 101,
1481
1482 /**
1483 * This error indicates that the Grid license is not applied.
1484 */
1485 CUDA_ERROR_DEVICE_NOT_LICENSED = 102,
1486
1487 /**
1488 * This indicates that the device kernel image is invalid. This can also
1489 * indicate an invalid CUDA module.
1490 */
1491 CUDA_ERROR_INVALID_IMAGE = 200,
1492
1493 /**
1494 * This most frequently indicates that there is no context bound to the
1495 * current thread. This can also be returned if the context passed to an
1496 * API call is not a valid handle (such as a context that has had
1497 * ::cuCtxDestroy() invoked on it). This can also be returned if a user
1498 * mixes different API versions (i.e. 3010 context with 3020 API calls).
1499 * See ::cuCtxGetApiVersion() for more details.
1500 */
1501 CUDA_ERROR_INVALID_CONTEXT = 201,
1502
1503 /**
1504 * This indicated that the context being supplied as a parameter to the
1505 * API call was already the active context.
1506 * \deprecated
1507 * This error return is deprecated as of CUDA 3.2. It is no longer an
1508 * error to attempt to push the active context via ::cuCtxPushCurrent().
1509 */
1510 CUDA_ERROR_CONTEXT_ALREADY_CURRENT = 202,
1511
1512 /**
1513 * This indicates that a map or register operation has failed.
1514 */
1515 CUDA_ERROR_MAP_FAILED = 205,
1516
1517 /**
1518 * This indicates that an unmap or unregister operation has failed.
1519 */
1520 CUDA_ERROR_UNMAP_FAILED = 206,
1521
1522 /**
1523 * This indicates that the specified array is currently mapped and thus
1524 * cannot be destroyed.
1525 */
1526 CUDA_ERROR_ARRAY_IS_MAPPED = 207,
1527
1528 /**
1529 * This indicates that the resource is already mapped.
1530 */
1531 CUDA_ERROR_ALREADY_MAPPED = 208,
1532
1533 /**
1534 * This indicates that there is no kernel image available that is suitable
1535 * for the device. This can occur when a user specifies code generation
1536 * options for a particular CUDA source file that do not include the
1537 * corresponding device configuration.
1538 */
1539 CUDA_ERROR_NO_BINARY_FOR_GPU = 209,
1540
1541 /**
1542 * This indicates that a resource has already been acquired.
1543 */
1544 CUDA_ERROR_ALREADY_ACQUIRED = 210,
1545
1546 /**
1547 * This indicates that a resource is not mapped.
1548 */
1549 CUDA_ERROR_NOT_MAPPED = 211,
1550
1551 /**
1552 * This indicates that a mapped resource is not available for access as an
1553 * array.
1554 */
1555 CUDA_ERROR_NOT_MAPPED_AS_ARRAY = 212,
1556
1557 /**
1558 * This indicates that a mapped resource is not available for access as a
1559 * pointer.
1560 */
1561 CUDA_ERROR_NOT_MAPPED_AS_POINTER = 213,
1562
1563 /**
1564 * This indicates that an uncorrectable ECC error was detected during
1565 * execution.
1566 */
1567 CUDA_ERROR_ECC_UNCORRECTABLE = 214,
1568
1569 /**
1570 * This indicates that the ::CUlimit passed to the API call is not
1571 * supported by the active device.
1572 */
1573 CUDA_ERROR_UNSUPPORTED_LIMIT = 215,
1574
1575 /**
1576 * This indicates that the ::CUcontext passed to the API call can
1577 * only be bound to a single CPU thread at a time but is already
1578 * bound to a CPU thread.
1579 */
1580 CUDA_ERROR_CONTEXT_ALREADY_IN_USE = 216,
1581
1582 /**
1583 * This indicates that peer access is not supported across the given
1584 * devices.
1585 */
1586 CUDA_ERROR_PEER_ACCESS_UNSUPPORTED = 217,
1587
1588 /**
1589 * This indicates that a PTX JIT compilation failed.
1590 */
1591 CUDA_ERROR_INVALID_PTX = 218,
1592
1593 /**
1594 * This indicates an error with OpenGL or DirectX context.
1595 */
1596 CUDA_ERROR_INVALID_GRAPHICS_CONTEXT = 219,
1597
1598 /**
1599 * This indicates that an uncorrectable NVLink error was detected during the
1600 * execution.
1601 */
1602 CUDA_ERROR_NVLINK_UNCORRECTABLE = 220,
1603
1604 /**
1605 * This indicates that the PTX JIT compiler library was not found.
1606 */
1607 CUDA_ERROR_JIT_COMPILER_NOT_FOUND = 221,
1608
1609 /**
1610 * This indicates that the provided PTX was compiled with an unsupported toolchain.
1611 */
1612
1613 CUDA_ERROR_UNSUPPORTED_PTX_VERSION = 222,
1614
1615 /**
1616 * This indicates that the PTX JIT compilation was disabled.
1617 */
1618 CUDA_ERROR_JIT_COMPILATION_DISABLED = 223,
1619
1620 /**
1621 * This indicates that the ::CUexecAffinityType passed to the API call is not
1622 * supported by the active device.
1623 */
1624 CUDA_ERROR_UNSUPPORTED_EXEC_AFFINITY = 224,
1625
1626 /**
1627 * This indicates that the device kernel source is invalid.
1628 */
1629 CUDA_ERROR_INVALID_SOURCE = 300,
1630
1631 /**
1632 * This indicates that the file specified was not found.
1633 */
1634 CUDA_ERROR_FILE_NOT_FOUND = 301,
1635
1636 /**
1637 * This indicates that a link to a shared object failed to resolve.
1638 */
1639 CUDA_ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND = 302,
1640
1641 /**
1642 * This indicates that initialization of a shared object failed.
1643 */
1644 CUDA_ERROR_SHARED_OBJECT_INIT_FAILED = 303,
1645
1646 /**
1647 * This indicates that an OS call failed.
1648 */
1649 CUDA_ERROR_OPERATING_SYSTEM = 304,
1650
1651 /**
1652 * This indicates that a resource handle passed to the API call was not
1653 * valid. Resource handles are opaque types like ::CUstream and ::CUevent.
1654 */
1655 CUDA_ERROR_INVALID_HANDLE = 400,
1656
1657 /**
1658 * This indicates that a resource required by the API call is not in a
1659 * valid state to perform the requested operation.
1660 */
1661 CUDA_ERROR_ILLEGAL_STATE = 401,
1662
1663 /**
1664 * This indicates that a named symbol was not found. Examples of symbols
1665 * are global/constant variable names, driver function names, texture names,
1666 * and surface names.
1667 */
1668 CUDA_ERROR_NOT_FOUND = 500,
1669
1670 /**
1671 * This indicates that asynchronous operations issued previously have not
1672 * completed yet. This result is not actually an error, but must be indicated
1673 * differently than ::CUDA_SUCCESS (which indicates completion). Calls that
1674 * may return this value include ::cuEventQuery() and ::cuStreamQuery().
1675 */
1676 CUDA_ERROR_NOT_READY = 600,
1677
1678 /**
1679 * While executing a kernel, the device encountered a
1680 * load or store instruction on an invalid memory address.
1681 * This leaves the process in an inconsistent state and any further CUDA work
1682 * will return the same error. To continue using CUDA, the process must be terminated
1683 * and relaunched.
1684 */
1685 CUDA_ERROR_ILLEGAL_ADDRESS = 700,
1686
1687 /**
1688 * This indicates that a launch did not occur because it did not have
1689 * appropriate resources. This error usually indicates that the user has
1690 * attempted to pass too many arguments to the device kernel, or the
1691 * kernel launch specifies too many threads for the kernel's register
1692 * count. Passing arguments of the wrong size (i.e. a 64-bit pointer
1693 * when a 32-bit int is expected) is equivalent to passing too many
1694 * arguments and can also result in this error.
1695 */
1696 CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES = 701,
1697
1698 /**
1699 * This indicates that the device kernel took too long to execute. This can
1700 * only occur if timeouts are enabled - see the device attribute
1701 * ::CU_DEVICE_ATTRIBUTE_KERNEL_EXEC_TIMEOUT for more information.
1702 * This leaves the process in an inconsistent state and any further CUDA work
1703 * will return the same error. To continue using CUDA, the process must be terminated
1704 * and relaunched.
1705 */
1706 CUDA_ERROR_LAUNCH_TIMEOUT = 702,
1707
1708 /**
1709 * This error indicates a kernel launch that uses an incompatible texturing
1710 * mode.
1711 */
1712 CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING = 703,
1713
1714 /**
1715 * This error indicates that a call to ::cuCtxEnablePeerAccess() is
1716 * trying to re-enable peer access to a context which has already
1717 * had peer access to it enabled.
1718 */
1719 CUDA_ERROR_PEER_ACCESS_ALREADY_ENABLED = 704,
1720
1721 /**
1722 * This error indicates that ::cuCtxDisablePeerAccess() is
1723 * trying to disable peer access which has not been enabled yet
1724 * via ::cuCtxEnablePeerAccess().
1725 */
1726 CUDA_ERROR_PEER_ACCESS_NOT_ENABLED = 705,
1727
1728 /**
1729 * This error indicates that the primary context for the specified device
1730 * has already been initialized.
1731 */
1732 CUDA_ERROR_PRIMARY_CONTEXT_ACTIVE = 708,
1733
1734 /**
1735 * This error indicates that the context current to the calling thread
1736 * has been destroyed using ::cuCtxDestroy, or is a primary context which
1737 * has not yet been initialized.
1738 */
1739 CUDA_ERROR_CONTEXT_IS_DESTROYED = 709,
1740
1741 /**
1742 * A device-side assert triggered during kernel execution. The context
1743 * cannot be used anymore, and must be destroyed. All existing device
1744 * memory allocations from this context are invalid and must be
1745 * reconstructed if the program is to continue using CUDA.
1746 */
1747 CUDA_ERROR_ASSERT = 710,
1748
1749 /**
1750 * This error indicates that the hardware resources required to enable
1751 * peer access have been exhausted for one or more of the devices
1752 * passed to ::cuCtxEnablePeerAccess().
1753 */
1754 CUDA_ERROR_TOO_MANY_PEERS = 711,
1755
1756 /**
1757 * This error indicates that the memory range passed to ::cuMemHostRegister()
1758 * has already been registered.
1759 */
1760 CUDA_ERROR_HOST_MEMORY_ALREADY_REGISTERED = 712,
1761
1762 /**
1763 * This error indicates that the pointer passed to ::cuMemHostUnregister()
1764 * does not correspond to any currently registered memory region.
1765 */
1766 CUDA_ERROR_HOST_MEMORY_NOT_REGISTERED = 713,
1767
1768 /**
1769 * While executing a kernel, the device encountered a stack error.
1770 * This can be due to stack corruption or exceeding the stack size limit.
1771 * This leaves the process in an inconsistent state and any further CUDA work
1772 * will return the same error. To continue using CUDA, the process must be terminated
1773 * and relaunched.
1774 */
1775 CUDA_ERROR_HARDWARE_STACK_ERROR = 714,
1776
1777 /**
1778 * While executing a kernel, the device encountered an illegal instruction.
1779 * This leaves the process in an inconsistent state and any further CUDA work
1780 * will return the same error. To continue using CUDA, the process must be terminated
1781 * and relaunched.
1782 */
1783 CUDA_ERROR_ILLEGAL_INSTRUCTION = 715,
1784
1785 /**
1786 * While executing a kernel, the device encountered a load or store instruction
1787 * on a memory address which is not aligned.
1788 * This leaves the process in an inconsistent state and any further CUDA work
1789 * will return the same error. To continue using CUDA, the process must be terminated
1790 * and relaunched.
1791 */
1792 CUDA_ERROR_MISALIGNED_ADDRESS = 716,
1793
1794 /**
1795 * While executing a kernel, the device encountered an instruction
1796 * which can only operate on memory locations in certain address spaces
1797 * (global, shared, or local), but was supplied a memory address not
1798 * belonging to an allowed address space.
1799 * This leaves the process in an inconsistent state and any further CUDA work
1800 * will return the same error. To continue using CUDA, the process must be terminated
1801 * and relaunched.
1802 */
1803 CUDA_ERROR_INVALID_ADDRESS_SPACE = 717,
1804
1805 /**
1806 * While executing a kernel, the device program counter wrapped its address space.
1807 * This leaves the process in an inconsistent state and any further CUDA work
1808 * will return the same error. To continue using CUDA, the process must be terminated
1809 * and relaunched.
1810 */
1811 CUDA_ERROR_INVALID_PC = 718,
1812
1813 /**
1814 * An exception occurred on the device while executing a kernel. Common
1815 * causes include dereferencing an invalid device pointer and accessing
1816 * out of bounds shared memory. Less common cases can be system specific - more
1817 * information about these cases can be found in the system specific user guide.
1818 * This leaves the process in an inconsistent state and any further CUDA work
1819 * will return the same error. To continue using CUDA, the process must be terminated
1820 * and relaunched.
1821 */
1822 CUDA_ERROR_LAUNCH_FAILED = 719,
1823
1824 /**
1825 * This error indicates that the number of blocks launched per grid for a kernel that was
1826 * launched via either ::cuLaunchCooperativeKernel or ::cuLaunchCooperativeKernelMultiDevice
1827 * exceeds the maximum number of blocks as allowed by ::cuOccupancyMaxActiveBlocksPerMultiprocessor
1828 * or ::cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags times the number of multiprocessors
1829 * as specified by the device attribute ::CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT.
1830 */
1831 CUDA_ERROR_COOPERATIVE_LAUNCH_TOO_LARGE = 720,
1832
1833 /**
1834 * This error indicates that the attempted operation is not permitted.
1835 */
1836 CUDA_ERROR_NOT_PERMITTED = 800,
1837
1838 /**
1839 * This error indicates that the attempted operation is not supported
1840 * on the current system or device.
1841 */
1842 CUDA_ERROR_NOT_SUPPORTED = 801,
1843
1844 /**
1845 * This error indicates that the system is not yet ready to start any CUDA
1846 * work. To continue using CUDA, verify the system configuration is in a
1847 * valid state and all required driver daemons are actively running.
1848 * More information about this error can be found in the system specific
1849 * user guide.
1850 */
1851 CUDA_ERROR_SYSTEM_NOT_READY = 802,
1852
1853 /**
1854 * This error indicates that there is a mismatch between the versions of
1855 * the display driver and the CUDA driver. Refer to the compatibility documentation
1856 * for supported versions.
1857 */
1858 CUDA_ERROR_SYSTEM_DRIVER_MISMATCH = 803,
1859
1860 /**
1861 * This error indicates that the system was upgraded to run with forward compatibility
1862 * but the visible hardware detected by CUDA does not support this configuration.
1863 * Refer to the compatibility documentation for the supported hardware matrix or ensure
1864 * that only supported hardware is visible during initialization via the CUDA_VISIBLE_DEVICES
1865 * environment variable.
1866 */
1867 CUDA_ERROR_COMPAT_NOT_SUPPORTED_ON_DEVICE = 804,
1868
1869 /**
1870 * This error indicates that the MPS client failed to connect to the MPS control daemon or the MPS server.
1871 */
1872 CUDA_ERROR_MPS_CONNECTION_FAILED = 805,
1873
1874 /**
1875 * This error indicates that the remote procedural call between the MPS server and the MPS client failed.
1876 */
1877 CUDA_ERROR_MPS_RPC_FAILURE = 806,
1878
1879 /**
1880 * This error indicates that the MPS server is not ready to accept new MPS client requests.
1881 * This error can be returned when the MPS server is in the process of recovering from a fatal failure.
1882 */
1883 CUDA_ERROR_MPS_SERVER_NOT_READY = 807,
1884
1885 /**
1886 * This error indicates that the hardware resources required to create MPS client have been exhausted.
1887 */
1888 CUDA_ERROR_MPS_MAX_CLIENTS_REACHED = 808,
1889
1890 /**
1891 * This error indicates the the hardware resources required to support device connections have been exhausted.
1892 */
1893 CUDA_ERROR_MPS_MAX_CONNECTIONS_REACHED = 809,
1894
1895 /**
1896 * This error indicates that the operation is not permitted when
1897 * the stream is capturing.
1898 */
1899 CUDA_ERROR_STREAM_CAPTURE_UNSUPPORTED = 900,
1900
1901 /**
1902 * This error indicates that the current capture sequence on the stream
1903 * has been invalidated due to a previous error.
1904 */
1905 CUDA_ERROR_STREAM_CAPTURE_INVALIDATED = 901,
1906
1907 /**
1908 * This error indicates that the operation would have resulted in a merge
1909 * of two independent capture sequences.
1910 */
1911 CUDA_ERROR_STREAM_CAPTURE_MERGE = 902,
1912
1913 /**
1914 * This error indicates that the capture was not initiated in this stream.
1915 */
1916 CUDA_ERROR_STREAM_CAPTURE_UNMATCHED = 903,
1917
1918 /**
1919 * This error indicates that the capture sequence contains a fork that was
1920 * not joined to the primary stream.
1921 */
1922 CUDA_ERROR_STREAM_CAPTURE_UNJOINED = 904,
1923
1924 /**
1925 * This error indicates that a dependency would have been created which
1926 * crosses the capture sequence boundary. Only implicit in-stream ordering
1927 * dependencies are allowed to cross the boundary.
1928 */
1929 CUDA_ERROR_STREAM_CAPTURE_ISOLATION = 905,
1930
1931 /**
1932 * This error indicates a disallowed implicit dependency on a current capture
1933 * sequence from cudaStreamLegacy.
1934 */
1935 CUDA_ERROR_STREAM_CAPTURE_IMPLICIT = 906,
1936
1937 /**
1938 * This error indicates that the operation is not permitted on an event which
1939 * was last recorded in a capturing stream.
1940 */
1941 CUDA_ERROR_CAPTURED_EVENT = 907,
1942
1943 /**
1944 * A stream capture sequence not initiated with the ::CU_STREAM_CAPTURE_MODE_RELAXED
1945 * argument to ::cuStreamBeginCapture was passed to ::cuStreamEndCapture in a
1946 * different thread.
1947 */
1948 CUDA_ERROR_STREAM_CAPTURE_WRONG_THREAD = 908,
1949
1950 /**
1951 * This error indicates that the timeout specified for the wait operation has lapsed.
1952 */
1953 CUDA_ERROR_TIMEOUT = 909,
1954
1955 /**
1956 * This error indicates that the graph update was not performed because it included
1957 * changes which violated constraints specific to instantiated graph update.
1958 */
1959 CUDA_ERROR_GRAPH_EXEC_UPDATE_FAILURE = 910,
1960
1961 /**
1962 * This indicates that an async error has occurred in a device outside of CUDA.
1963 * If CUDA was waiting for an external device's signal before consuming shared data,
1964 * the external device signaled an error indicating that the data is not valid for
1965 * consumption. This leaves the process in an inconsistent state and any further CUDA
1966 * work will return the same error. To continue using CUDA, the process must be
1967 * terminated and relaunched.
1968 */
1969 CUDA_ERROR_EXTERNAL_DEVICE = 911,
1970
1971 /**
1972 * This indicates that an unknown internal error has occurred.
1973 */
1974 CUDA_ERROR_UNKNOWN = 999
1975} CUresult;
1976
1977/**
1978 * P2P Attributes
1979 */
1980typedef enum CUdevice_P2PAttribute_enum {
1981 CU_DEVICE_P2P_ATTRIBUTE_PERFORMANCE_RANK = 0x01, /**< A relative value indicating the performance of the link between two devices */
1982 CU_DEVICE_P2P_ATTRIBUTE_ACCESS_SUPPORTED = 0x02, /**< P2P Access is enable */
1983 CU_DEVICE_P2P_ATTRIBUTE_NATIVE_ATOMIC_SUPPORTED = 0x03, /**< Atomic operation over the link supported */
1984 CU_DEVICE_P2P_ATTRIBUTE_ACCESS_ACCESS_SUPPORTED = 0x04, /**< \deprecated use CU_DEVICE_P2P_ATTRIBUTE_CUDA_ARRAY_ACCESS_SUPPORTED instead */
1985 CU_DEVICE_P2P_ATTRIBUTE_CUDA_ARRAY_ACCESS_SUPPORTED = 0x04 /**< Accessing CUDA arrays over the link supported */
1986} CUdevice_P2PAttribute;
1987
1988/**
1989 * CUDA stream callback
1990 * \param hStream The stream the callback was added to, as passed to ::cuStreamAddCallback. May be NULL.
1991 * \param status ::CUDA_SUCCESS or any persistent error on the stream.
1992 * \param userData User parameter provided at registration.
1993 */
1994typedef void (CUDA_CB *CUstreamCallback)(CUstream hStream, CUresult status, void *userData);
1995
1996/**
1997 * Block size to per-block dynamic shared memory mapping for a certain
1998 * kernel \param blockSize Block size of the kernel.
1999 *
2000 * \return The dynamic shared memory needed by a block.
2001 */
2002typedef size_t (CUDA_CB *CUoccupancyB2DSize)(int blockSize);
2003
2004/**
2005 * If set, host memory is portable between CUDA contexts.
2006 * Flag for ::cuMemHostAlloc()
2007 */
2008#define CU_MEMHOSTALLOC_PORTABLE 0x01
2009
2010/**
2011 * If set, host memory is mapped into CUDA address space and
2012 * ::cuMemHostGetDevicePointer() may be called on the host pointer.
2013 * Flag for ::cuMemHostAlloc()
2014 */
2015#define CU_MEMHOSTALLOC_DEVICEMAP 0x02
2016
2017/**
2018 * If set, host memory is allocated as write-combined - fast to write,
2019 * faster to DMA, slow to read except via SSE4 streaming load instruction
2020 * (MOVNTDQA).
2021 * Flag for ::cuMemHostAlloc()
2022 */
2023#define CU_MEMHOSTALLOC_WRITECOMBINED 0x04
2024
2025/**
2026 * If set, host memory is portable between CUDA contexts.
2027 * Flag for ::cuMemHostRegister()
2028 */
2029#define CU_MEMHOSTREGISTER_PORTABLE 0x01
2030
2031/**
2032 * If set, host memory is mapped into CUDA address space and
2033 * ::cuMemHostGetDevicePointer() may be called on the host pointer.
2034 * Flag for ::cuMemHostRegister()
2035 */
2036#define CU_MEMHOSTREGISTER_DEVICEMAP 0x02
2037
2038/**
2039 * If set, the passed memory pointer is treated as pointing to some
2040 * memory-mapped I/O space, e.g. belonging to a third-party PCIe device.
2041 * On Windows the flag is a no-op.
2042 * On Linux that memory is marked as non cache-coherent for the GPU and
2043 * is expected to be physically contiguous. It may return
2044 * CUDA_ERROR_NOT_PERMITTED if run as an unprivileged user,
2045 * CUDA_ERROR_NOT_SUPPORTED on older Linux kernel versions.
2046 * On all other platforms, it is not supported and CUDA_ERROR_NOT_SUPPORTED
2047 * is returned.
2048 * Flag for ::cuMemHostRegister()
2049 */
2050#define CU_MEMHOSTREGISTER_IOMEMORY 0x04
2051
2052/**
2053* If set, the passed memory pointer is treated as pointing to memory that is
2054* considered read-only by the device. On platforms without
2055* CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS_USES_HOST_PAGE_TABLES, this flag is
2056* required in order to register memory mapped to the CPU as read-only. Support
2057* for the use of this flag can be queried from the device attribute
2058* CU_DEVICE_ATTRIBUTE_READ_ONLY_HOST_REGISTER_SUPPORTED. Using this flag with
2059* a current context associated with a device that does not have this attribute
2060* set will cause ::cuMemHostRegister to error with CUDA_ERROR_NOT_SUPPORTED.
2061*/
2062#define CU_MEMHOSTREGISTER_READ_ONLY 0x08
2063
2064/**
2065 * 2D memory copy parameters
2066 */
2067typedef struct CUDA_MEMCPY2D_st {
2068 size_t srcXInBytes; /**< Source X in bytes */
2069 size_t srcY; /**< Source Y */
2070
2071 CUmemorytype srcMemoryType; /**< Source memory type (host, device, array) */
2072 const void *srcHost; /**< Source host pointer */
2073 CUdeviceptr srcDevice; /**< Source device pointer */
2074 CUarray srcArray; /**< Source array reference */
2075 size_t srcPitch; /**< Source pitch (ignored when src is array) */
2076
2077 size_t dstXInBytes; /**< Destination X in bytes */
2078 size_t dstY; /**< Destination Y */
2079
2080 CUmemorytype dstMemoryType; /**< Destination memory type (host, device, array) */
2081 void *dstHost; /**< Destination host pointer */
2082 CUdeviceptr dstDevice; /**< Destination device pointer */
2083 CUarray dstArray; /**< Destination array reference */
2084 size_t dstPitch; /**< Destination pitch (ignored when dst is array) */
2085
2086 size_t WidthInBytes; /**< Width of 2D memory copy in bytes */
2087 size_t Height; /**< Height of 2D memory copy */
2088} CUDA_MEMCPY2D_v2;
2089typedef CUDA_MEMCPY2D_v2 CUDA_MEMCPY2D;
2090
2091/**
2092 * 3D memory copy parameters
2093 */
2094typedef struct CUDA_MEMCPY3D_st {
2095 size_t srcXInBytes; /**< Source X in bytes */
2096 size_t srcY; /**< Source Y */
2097 size_t srcZ; /**< Source Z */
2098 size_t srcLOD; /**< Source LOD */
2099 CUmemorytype srcMemoryType; /**< Source memory type (host, device, array) */
2100 const void *srcHost; /**< Source host pointer */
2101 CUdeviceptr srcDevice; /**< Source device pointer */
2102 CUarray srcArray; /**< Source array reference */
2103 void *reserved0; /**< Must be NULL */
2104 size_t srcPitch; /**< Source pitch (ignored when src is array) */
2105 size_t srcHeight; /**< Source height (ignored when src is array; may be 0 if Depth==1) */
2106
2107 size_t dstXInBytes; /**< Destination X in bytes */
2108 size_t dstY; /**< Destination Y */
2109 size_t dstZ; /**< Destination Z */
2110 size_t dstLOD; /**< Destination LOD */
2111 CUmemorytype dstMemoryType; /**< Destination memory type (host, device, array) */
2112 void *dstHost; /**< Destination host pointer */
2113 CUdeviceptr dstDevice; /**< Destination device pointer */
2114 CUarray dstArray; /**< Destination array reference */
2115 void *reserved1; /**< Must be NULL */
2116 size_t dstPitch; /**< Destination pitch (ignored when dst is array) */
2117 size_t dstHeight; /**< Destination height (ignored when dst is array; may be 0 if Depth==1) */
2118
2119 size_t WidthInBytes; /**< Width of 3D memory copy in bytes */
2120 size_t Height; /**< Height of 3D memory copy */
2121 size_t Depth; /**< Depth of 3D memory copy */
2122} CUDA_MEMCPY3D_v2;
2123typedef CUDA_MEMCPY3D_v2 CUDA_MEMCPY3D;
2124
2125/**
2126 * 3D memory cross-context copy parameters
2127 */
2128typedef struct CUDA_MEMCPY3D_PEER_st {
2129 size_t srcXInBytes; /**< Source X in bytes */
2130 size_t srcY; /**< Source Y */
2131 size_t srcZ; /**< Source Z */
2132 size_t srcLOD; /**< Source LOD */
2133 CUmemorytype srcMemoryType; /**< Source memory type (host, device, array) */
2134 const void *srcHost; /**< Source host pointer */
2135 CUdeviceptr srcDevice; /**< Source device pointer */
2136 CUarray srcArray; /**< Source array reference */
2137 CUcontext srcContext; /**< Source context (ignored with srcMemoryType is ::CU_MEMORYTYPE_ARRAY) */
2138 size_t srcPitch; /**< Source pitch (ignored when src is array) */
2139 size_t srcHeight; /**< Source height (ignored when src is array; may be 0 if Depth==1) */
2140
2141 size_t dstXInBytes; /**< Destination X in bytes */
2142 size_t dstY; /**< Destination Y */
2143 size_t dstZ; /**< Destination Z */
2144 size_t dstLOD; /**< Destination LOD */
2145 CUmemorytype dstMemoryType; /**< Destination memory type (host, device, array) */
2146 void *dstHost; /**< Destination host pointer */
2147 CUdeviceptr dstDevice; /**< Destination device pointer */
2148 CUarray dstArray; /**< Destination array reference */
2149 CUcontext dstContext; /**< Destination context (ignored with dstMemoryType is ::CU_MEMORYTYPE_ARRAY) */
2150 size_t dstPitch; /**< Destination pitch (ignored when dst is array) */
2151 size_t dstHeight; /**< Destination height (ignored when dst is array; may be 0 if Depth==1) */
2152
2153 size_t WidthInBytes; /**< Width of 3D memory copy in bytes */
2154 size_t Height; /**< Height of 3D memory copy */
2155 size_t Depth; /**< Depth of 3D memory copy */
2156} CUDA_MEMCPY3D_PEER_v1;
2157typedef CUDA_MEMCPY3D_PEER_v1 CUDA_MEMCPY3D_PEER;
2158
2159/**
2160 * Array descriptor
2161 */
2162typedef struct CUDA_ARRAY_DESCRIPTOR_st
2163{
2164 size_t Width; /**< Width of array */
2165 size_t Height; /**< Height of array */
2166
2167 CUarray_format Format; /**< Array format */
2168 unsigned int NumChannels; /**< Channels per array element */
2169} CUDA_ARRAY_DESCRIPTOR_v2;
2170typedef CUDA_ARRAY_DESCRIPTOR_v2 CUDA_ARRAY_DESCRIPTOR;
2171
2172/**
2173 * 3D array descriptor
2174 */
2175typedef struct CUDA_ARRAY3D_DESCRIPTOR_st
2176{
2177 size_t Width; /**< Width of 3D array */
2178 size_t Height; /**< Height of 3D array */
2179 size_t Depth; /**< Depth of 3D array */
2180
2181 CUarray_format Format; /**< Array format */
2182 unsigned int NumChannels; /**< Channels per array element */
2183 unsigned int Flags; /**< Flags */
2184} CUDA_ARRAY3D_DESCRIPTOR_v2;
2185typedef CUDA_ARRAY3D_DESCRIPTOR_v2 CUDA_ARRAY3D_DESCRIPTOR;
2186
2187/**
2188 * Indicates that the layered sparse CUDA array or CUDA mipmapped array has a single mip tail region for all layers
2189 */
2190#define CU_ARRAY_SPARSE_PROPERTIES_SINGLE_MIPTAIL 0x1
2191
2192/**
2193 * CUDA array sparse properties
2194 */
2195typedef struct CUDA_ARRAY_SPARSE_PROPERTIES_st {
2196 struct {
2197 unsigned int width; /**< Width of sparse tile in elements */
2198 unsigned int height; /**< Height of sparse tile in elements */
2199 unsigned int depth; /**< Depth of sparse tile in elements */
2200 } tileExtent;
2201
2202 /**
2203 * First mip level at which the mip tail begins.
2204 */
2205 unsigned int miptailFirstLevel;
2206 /**
2207 * Total size of the mip tail.
2208 */
2209 unsigned long long miptailSize;
2210 /**
2211 * Flags will either be zero or ::CU_ARRAY_SPARSE_PROPERTIES_SINGLE_MIPTAIL
2212 */
2213 unsigned int flags;
2214 unsigned int reserved[4];
2215} CUDA_ARRAY_SPARSE_PROPERTIES_v1;
2216typedef CUDA_ARRAY_SPARSE_PROPERTIES_v1 CUDA_ARRAY_SPARSE_PROPERTIES;
2217
2218/**
2219 * CUDA Resource descriptor
2220 */
2221typedef struct CUDA_RESOURCE_DESC_st
2222{
2223 CUresourcetype resType; /**< Resource type */
2224
2225 union {
2226 struct {
2227 CUarray hArray; /**< CUDA array */
2228 } array;
2229 struct {
2230 CUmipmappedArray hMipmappedArray; /**< CUDA mipmapped array */
2231 } mipmap;
2232 struct {
2233 CUdeviceptr devPtr; /**< Device pointer */
2234 CUarray_format format; /**< Array format */
2235 unsigned int numChannels; /**< Channels per array element */
2236 size_t sizeInBytes; /**< Size in bytes */
2237 } linear;
2238 struct {
2239 CUdeviceptr devPtr; /**< Device pointer */
2240 CUarray_format format; /**< Array format */
2241 unsigned int numChannels; /**< Channels per array element */
2242 size_t width; /**< Width of the array in elements */
2243 size_t height; /**< Height of the array in elements */
2244 size_t pitchInBytes; /**< Pitch between two rows in bytes */
2245 } pitch2D;
2246 struct {
2247 int reserved[32];
2248 } reserved;
2249 } res;
2250
2251 unsigned int flags; /**< Flags (must be zero) */
2252} CUDA_RESOURCE_DESC_v1;
2253typedef CUDA_RESOURCE_DESC_v1 CUDA_RESOURCE_DESC;
2254
2255/**
2256 * Texture descriptor
2257 */
2258typedef struct CUDA_TEXTURE_DESC_st {
2259 CUaddress_mode addressMode[3]; /**< Address modes */
2260 CUfilter_mode filterMode; /**< Filter mode */
2261 unsigned int flags; /**< Flags */
2262 unsigned int maxAnisotropy; /**< Maximum anisotropy ratio */
2263 CUfilter_mode mipmapFilterMode; /**< Mipmap filter mode */
2264 float mipmapLevelBias; /**< Mipmap level bias */
2265 float minMipmapLevelClamp; /**< Mipmap minimum level clamp */
2266 float maxMipmapLevelClamp; /**< Mipmap maximum level clamp */
2267 float borderColor[4]; /**< Border Color */
2268 int reserved[12];
2269} CUDA_TEXTURE_DESC_v1;
2270typedef CUDA_TEXTURE_DESC_v1 CUDA_TEXTURE_DESC;
2271
2272/**
2273 * Resource view format
2274 */
2275typedef enum CUresourceViewFormat_enum
2276{
2277 CU_RES_VIEW_FORMAT_NONE = 0x00, /**< No resource view format (use underlying resource format) */
2278 CU_RES_VIEW_FORMAT_UINT_1X8 = 0x01, /**< 1 channel unsigned 8-bit integers */
2279 CU_RES_VIEW_FORMAT_UINT_2X8 = 0x02, /**< 2 channel unsigned 8-bit integers */
2280 CU_RES_VIEW_FORMAT_UINT_4X8 = 0x03, /**< 4 channel unsigned 8-bit integers */
2281 CU_RES_VIEW_FORMAT_SINT_1X8 = 0x04, /**< 1 channel signed 8-bit integers */
2282 CU_RES_VIEW_FORMAT_SINT_2X8 = 0x05, /**< 2 channel signed 8-bit integers */
2283 CU_RES_VIEW_FORMAT_SINT_4X8 = 0x06, /**< 4 channel signed 8-bit integers */
2284 CU_RES_VIEW_FORMAT_UINT_1X16 = 0x07, /**< 1 channel unsigned 16-bit integers */
2285 CU_RES_VIEW_FORMAT_UINT_2X16 = 0x08, /**< 2 channel unsigned 16-bit integers */
2286 CU_RES_VIEW_FORMAT_UINT_4X16 = 0x09, /**< 4 channel unsigned 16-bit integers */
2287 CU_RES_VIEW_FORMAT_SINT_1X16 = 0x0a, /**< 1 channel signed 16-bit integers */
2288 CU_RES_VIEW_FORMAT_SINT_2X16 = 0x0b, /**< 2 channel signed 16-bit integers */
2289 CU_RES_VIEW_FORMAT_SINT_4X16 = 0x0c, /**< 4 channel signed 16-bit integers */
2290 CU_RES_VIEW_FORMAT_UINT_1X32 = 0x0d, /**< 1 channel unsigned 32-bit integers */
2291 CU_RES_VIEW_FORMAT_UINT_2X32 = 0x0e, /**< 2 channel unsigned 32-bit integers */
2292 CU_RES_VIEW_FORMAT_UINT_4X32 = 0x0f, /**< 4 channel unsigned 32-bit integers */
2293 CU_RES_VIEW_FORMAT_SINT_1X32 = 0x10, /**< 1 channel signed 32-bit integers */
2294 CU_RES_VIEW_FORMAT_SINT_2X32 = 0x11, /**< 2 channel signed 32-bit integers */
2295 CU_RES_VIEW_FORMAT_SINT_4X32 = 0x12, /**< 4 channel signed 32-bit integers */
2296 CU_RES_VIEW_FORMAT_FLOAT_1X16 = 0x13, /**< 1 channel 16-bit floating point */
2297 CU_RES_VIEW_FORMAT_FLOAT_2X16 = 0x14, /**< 2 channel 16-bit floating point */
2298 CU_RES_VIEW_FORMAT_FLOAT_4X16 = 0x15, /**< 4 channel 16-bit floating point */
2299 CU_RES_VIEW_FORMAT_FLOAT_1X32 = 0x16, /**< 1 channel 32-bit floating point */
2300 CU_RES_VIEW_FORMAT_FLOAT_2X32 = 0x17, /**< 2 channel 32-bit floating point */
2301 CU_RES_VIEW_FORMAT_FLOAT_4X32 = 0x18, /**< 4 channel 32-bit floating point */
2302 CU_RES_VIEW_FORMAT_UNSIGNED_BC1 = 0x19, /**< Block compressed 1 */
2303 CU_RES_VIEW_FORMAT_UNSIGNED_BC2 = 0x1a, /**< Block compressed 2 */
2304 CU_RES_VIEW_FORMAT_UNSIGNED_BC3 = 0x1b, /**< Block compressed 3 */
2305 CU_RES_VIEW_FORMAT_UNSIGNED_BC4 = 0x1c, /**< Block compressed 4 unsigned */
2306 CU_RES_VIEW_FORMAT_SIGNED_BC4 = 0x1d, /**< Block compressed 4 signed */
2307 CU_RES_VIEW_FORMAT_UNSIGNED_BC5 = 0x1e, /**< Block compressed 5 unsigned */
2308 CU_RES_VIEW_FORMAT_SIGNED_BC5 = 0x1f, /**< Block compressed 5 signed */
2309 CU_RES_VIEW_FORMAT_UNSIGNED_BC6H = 0x20, /**< Block compressed 6 unsigned half-float */
2310 CU_RES_VIEW_FORMAT_SIGNED_BC6H = 0x21, /**< Block compressed 6 signed half-float */
2311 CU_RES_VIEW_FORMAT_UNSIGNED_BC7 = 0x22 /**< Block compressed 7 */
2312} CUresourceViewFormat;
2313
2314/**
2315 * Resource view descriptor
2316 */
2317typedef struct CUDA_RESOURCE_VIEW_DESC_st
2318{
2319 CUresourceViewFormat format; /**< Resource view format */
2320 size_t width; /**< Width of the resource view */
2321 size_t height; /**< Height of the resource view */
2322 size_t depth; /**< Depth of the resource view */
2323 unsigned int firstMipmapLevel; /**< First defined mipmap level */
2324 unsigned int lastMipmapLevel; /**< Last defined mipmap level */
2325 unsigned int firstLayer; /**< First layer index */
2326 unsigned int lastLayer; /**< Last layer index */
2327 unsigned int reserved[16];
2328} CUDA_RESOURCE_VIEW_DESC_v1;
2329typedef CUDA_RESOURCE_VIEW_DESC_v1 CUDA_RESOURCE_VIEW_DESC;
2330
2331/**
2332 * GPU Direct v3 tokens
2333 */
2334typedef struct CUDA_POINTER_ATTRIBUTE_P2P_TOKENS_st {
2335 unsigned long long p2pToken;
2336 unsigned int vaSpaceToken;
2337} CUDA_POINTER_ATTRIBUTE_P2P_TOKENS_v1;
2338typedef CUDA_POINTER_ATTRIBUTE_P2P_TOKENS_v1 CUDA_POINTER_ATTRIBUTE_P2P_TOKENS;
2339
2340/**
2341* Access flags that specify the level of access the current context's device has
2342* on the memory referenced.
2343*/
2344typedef enum CUDA_POINTER_ATTRIBUTE_ACCESS_FLAGS_enum {
2345 CU_POINTER_ATTRIBUTE_ACCESS_FLAG_NONE = 0x0, /**< No access, meaning the device cannot access this memory at all, thus must be staged through accessible memory in order to complete certain operations */
2346 CU_POINTER_ATTRIBUTE_ACCESS_FLAG_READ = 0x1, /**< Read-only access, meaning writes to this memory are considered invalid accesses and thus return error in that case. */
2347 CU_POINTER_ATTRIBUTE_ACCESS_FLAG_READWRITE = 0x3 /**< Read-write access, the device has full read-write access to the memory */
2348} CUDA_POINTER_ATTRIBUTE_ACCESS_FLAGS;
2349
2350/**
2351 * Kernel launch parameters
2352 */
2353typedef struct CUDA_LAUNCH_PARAMS_st {
2354 CUfunction function; /**< Kernel to launch */
2355 unsigned int gridDimX; /**< Width of grid in blocks */
2356 unsigned int gridDimY; /**< Height of grid in blocks */
2357 unsigned int gridDimZ; /**< Depth of grid in blocks */
2358 unsigned int blockDimX; /**< X dimension of each thread block */
2359 unsigned int blockDimY; /**< Y dimension of each thread block */
2360 unsigned int blockDimZ; /**< Z dimension of each thread block */
2361 unsigned int sharedMemBytes; /**< Dynamic shared-memory size per thread block in bytes */
2362 CUstream hStream; /**< Stream identifier */
2363 void **kernelParams; /**< Array of pointers to kernel parameters */
2364} CUDA_LAUNCH_PARAMS_v1;
2365typedef CUDA_LAUNCH_PARAMS_v1 CUDA_LAUNCH_PARAMS;
2366
2367/**
2368 * External memory handle types
2369 */
2370typedef enum CUexternalMemoryHandleType_enum {
2371 /**
2372 * Handle is an opaque file descriptor
2373 */
2374 CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD = 1,
2375 /**
2376 * Handle is an opaque shared NT handle
2377 */
2378 CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32 = 2,
2379 /**
2380 * Handle is an opaque, globally shared handle
2381 */
2382 CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT = 3,
2383 /**
2384 * Handle is a D3D12 heap object
2385 */
2386 CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_HEAP = 4,
2387 /**
2388 * Handle is a D3D12 committed resource
2389 */
2390 CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_RESOURCE = 5,
2391 /**
2392 * Handle is a shared NT handle to a D3D11 resource
2393 */
2394 CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_RESOURCE = 6,
2395 /**
2396 * Handle is a globally shared handle to a D3D11 resource
2397 */
2398 CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_RESOURCE_KMT = 7,
2399 /**
2400 * Handle is an NvSciBuf object
2401 */
2402 CU_EXTERNAL_MEMORY_HANDLE_TYPE_NVSCIBUF = 8
2403} CUexternalMemoryHandleType;
2404
2405/**
2406 * Indicates that the external memory object is a dedicated resource
2407 */
2408#define CUDA_EXTERNAL_MEMORY_DEDICATED 0x1
2409
2410/** When the \p flags parameter of ::CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS
2411 * contains this flag, it indicates that signaling an external semaphore object
2412 * should skip performing appropriate memory synchronization operations over all
2413 * the external memory objects that are imported as ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_NVSCIBUF,
2414 * which otherwise are performed by default to ensure data coherency with other
2415 * importers of the same NvSciBuf memory objects.
2416 */
2417#define CUDA_EXTERNAL_SEMAPHORE_SIGNAL_SKIP_NVSCIBUF_MEMSYNC 0x01
2418
2419/** When the \p flags parameter of ::CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS
2420 * contains this flag, it indicates that waiting on an external semaphore object
2421 * should skip performing appropriate memory synchronization operations over all
2422 * the external memory objects that are imported as ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_NVSCIBUF,
2423 * which otherwise are performed by default to ensure data coherency with other
2424 * importers of the same NvSciBuf memory objects.
2425 */
2426#define CUDA_EXTERNAL_SEMAPHORE_WAIT_SKIP_NVSCIBUF_MEMSYNC 0x02
2427
2428/**
2429 * When \p flags of ::cuDeviceGetNvSciSyncAttributes is set to this,
2430 * it indicates that application needs signaler specific NvSciSyncAttr
2431 * to be filled by ::cuDeviceGetNvSciSyncAttributes.
2432 */
2433#define CUDA_NVSCISYNC_ATTR_SIGNAL 0x1
2434
2435/**
2436 * When \p flags of ::cuDeviceGetNvSciSyncAttributes is set to this,
2437 * it indicates that application needs waiter specific NvSciSyncAttr
2438 * to be filled by ::cuDeviceGetNvSciSyncAttributes.
2439 */
2440#define CUDA_NVSCISYNC_ATTR_WAIT 0x2
2441/**
2442 * External memory handle descriptor
2443 */
2444typedef struct CUDA_EXTERNAL_MEMORY_HANDLE_DESC_st {
2445 /**
2446 * Type of the handle
2447 */
2448 CUexternalMemoryHandleType type;
2449 union {
2450 /**
2451 * File descriptor referencing the memory object. Valid
2452 * when type is
2453 * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD
2454 */
2455 int fd;
2456 /**
2457 * Win32 handle referencing the semaphore object. Valid when
2458 * type is one of the following:
2459 * - ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32
2460 * - ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT
2461 * - ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_HEAP
2462 * - ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_RESOURCE
2463 * - ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_RESOURCE
2464 * - ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_RESOURCE_KMT
2465 * Exactly one of 'handle' and 'name' must be non-NULL. If
2466 * type is one of the following:
2467 * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT
2468 * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_RESOURCE_KMT
2469 * then 'name' must be NULL.
2470 */
2471 struct {
2472 /**
2473 * Valid NT handle. Must be NULL if 'name' is non-NULL
2474 */
2475 void *handle;
2476 /**
2477 * Name of a valid memory object.
2478 * Must be NULL if 'handle' is non-NULL.
2479 */
2480 const void *name;
2481 } win32;
2482 /**
2483 * A handle representing an NvSciBuf Object. Valid when type
2484 * is ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_NVSCIBUF
2485 */
2486 const void *nvSciBufObject;
2487 } handle;
2488 /**
2489 * Size of the memory allocation
2490 */
2491 unsigned long long size;
2492 /**
2493 * Flags must either be zero or ::CUDA_EXTERNAL_MEMORY_DEDICATED
2494 */
2495 unsigned int flags;
2496 unsigned int reserved[16];
2497} CUDA_EXTERNAL_MEMORY_HANDLE_DESC_v1;
2498typedef CUDA_EXTERNAL_MEMORY_HANDLE_DESC_v1 CUDA_EXTERNAL_MEMORY_HANDLE_DESC;
2499
2500/**
2501 * External memory buffer descriptor
2502 */
2503typedef struct CUDA_EXTERNAL_MEMORY_BUFFER_DESC_st {
2504 /**
2505 * Offset into the memory object where the buffer's base is
2506 */
2507 unsigned long long offset;
2508 /**
2509 * Size of the buffer
2510 */
2511 unsigned long long size;
2512 /**
2513 * Flags reserved for future use. Must be zero.
2514 */
2515 unsigned int flags;
2516 unsigned int reserved[16];
2517} CUDA_EXTERNAL_MEMORY_BUFFER_DESC_v1;
2518typedef CUDA_EXTERNAL_MEMORY_BUFFER_DESC_v1 CUDA_EXTERNAL_MEMORY_BUFFER_DESC;
2519
2520/**
2521 * External memory mipmap descriptor
2522 */
2523typedef struct CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC_st {
2524 /**
2525 * Offset into the memory object where the base level of the
2526 * mipmap chain is.
2527 */
2528 unsigned long long offset;
2529 /**
2530 * Format, dimension and type of base level of the mipmap chain
2531 */
2532 CUDA_ARRAY3D_DESCRIPTOR arrayDesc;
2533 /**
2534 * Total number of levels in the mipmap chain
2535 */
2536 unsigned int numLevels;
2537 unsigned int reserved[16];
2538} CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC_v1;
2539typedef CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC_v1 CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC;
2540
2541/**
2542 * External semaphore handle types
2543 */
2544typedef enum CUexternalSemaphoreHandleType_enum {
2545 /**
2546 * Handle is an opaque file descriptor
2547 */
2548 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD = 1,
2549 /**
2550 * Handle is an opaque shared NT handle
2551 */
2552 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32 = 2,
2553 /**
2554 * Handle is an opaque, globally shared handle
2555 */
2556 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT = 3,
2557 /**
2558 * Handle is a shared NT handle referencing a D3D12 fence object
2559 */
2560 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D12_FENCE = 4,
2561 /**
2562 * Handle is a shared NT handle referencing a D3D11 fence object
2563 */
2564 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D11_FENCE = 5,
2565 /**
2566 * Opaque handle to NvSciSync Object
2567 */
2568 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_NVSCISYNC = 6,
2569 /**
2570 * Handle is a shared NT handle referencing a D3D11 keyed mutex object
2571 */
2572 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D11_KEYED_MUTEX = 7,
2573 /**
2574 * Handle is a globally shared handle referencing a D3D11 keyed mutex object
2575 */
2576 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D11_KEYED_MUTEX_KMT = 8,
2577 /**
2578 * Handle is an opaque file descriptor referencing a timeline semaphore
2579 */
2580 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_TIMELINE_SEMAPHORE_FD = 9,
2581 /**
2582 * Handle is an opaque shared NT handle referencing a timeline semaphore
2583 */
2584 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_TIMELINE_SEMAPHORE_WIN32 = 10
2585} CUexternalSemaphoreHandleType;
2586
2587/**
2588 * External semaphore handle descriptor
2589 */
2590typedef struct CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC_st {
2591 /**
2592 * Type of the handle
2593 */
2594 CUexternalSemaphoreHandleType type;
2595 union {
2596 /**
2597 * File descriptor referencing the semaphore object. Valid
2598 * when type is one of the following:
2599 * - ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD
2600 * - ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_TIMELINE_SEMAPHORE_FD
2601 */
2602 int fd;
2603 /**
2604 * Win32 handle referencing the semaphore object. Valid when
2605 * type is one of the following:
2606 * - ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32
2607 * - ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT
2608 * - ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D12_FENCE
2609 * - ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D11_FENCE
2610 * - ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D11_KEYED_MUTEX
2611 * - ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_TIMELINE_SEMAPHORE_WIN32
2612 * Exactly one of 'handle' and 'name' must be non-NULL. If
2613 * type is one of the following:
2614 * - ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT
2615 * - ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D11_KEYED_MUTEX_KMT
2616 * then 'name' must be NULL.
2617 */
2618 struct {
2619 /**
2620 * Valid NT handle. Must be NULL if 'name' is non-NULL
2621 */
2622 void *handle;
2623 /**
2624 * Name of a valid synchronization primitive.
2625 * Must be NULL if 'handle' is non-NULL.
2626 */
2627 const void *name;
2628 } win32;
2629 /**
2630 * Valid NvSciSyncObj. Must be non NULL
2631 */
2632 const void* nvSciSyncObj;
2633 } handle;
2634 /**
2635 * Flags reserved for the future. Must be zero.
2636 */
2637 unsigned int flags;
2638 unsigned int reserved[16];
2639} CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC_v1;
2640typedef CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC_v1 CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC;
2641
2642/**
2643 * External semaphore signal parameters
2644 */
2645typedef struct CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS_st {
2646 struct {
2647 /**
2648 * Parameters for fence objects
2649 */
2650 struct {
2651 /**
2652 * Value of fence to be signaled
2653 */
2654 unsigned long long value;
2655 } fence;
2656 union {
2657 /**
2658 * Pointer to NvSciSyncFence. Valid if ::CUexternalSemaphoreHandleType
2659 * is of type ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_NVSCISYNC.
2660 */
2661 void *fence;
2662 unsigned long long reserved;
2663 } nvSciSync;
2664 /**
2665 * Parameters for keyed mutex objects
2666 */
2667 struct {
2668 /**
2669 * Value of key to release the mutex with
2670 */
2671 unsigned long long key;
2672 } keyedMutex;
2673 unsigned int reserved[12];
2674 } params;
2675 /**
2676 * Only when ::CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS is used to
2677 * signal a ::CUexternalSemaphore of type
2678 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_NVSCISYNC, the valid flag is
2679 * ::CUDA_EXTERNAL_SEMAPHORE_SIGNAL_SKIP_NVSCIBUF_MEMSYNC which indicates
2680 * that while signaling the ::CUexternalSemaphore, no memory synchronization
2681 * operations should be performed for any external memory object imported
2682 * as ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_NVSCIBUF.
2683 * For all other types of ::CUexternalSemaphore, flags must be zero.
2684 */
2685 unsigned int flags;
2686 unsigned int reserved[16];
2687} CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS_v1;
2688typedef CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS_v1 CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS;
2689
2690/**
2691 * External semaphore wait parameters
2692 */
2693typedef struct CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS_st {
2694 struct {
2695 /**
2696 * Parameters for fence objects
2697 */
2698 struct {
2699 /**
2700 * Value of fence to be waited on
2701 */
2702 unsigned long long value;
2703 } fence;
2704 /**
2705 * Pointer to NvSciSyncFence. Valid if CUexternalSemaphoreHandleType
2706 * is of type CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_NVSCISYNC.
2707 */
2708 union {
2709 void *fence;
2710 unsigned long long reserved;
2711 } nvSciSync;
2712 /**
2713 * Parameters for keyed mutex objects
2714 */
2715 struct {
2716 /**
2717 * Value of key to acquire the mutex with
2718 */
2719 unsigned long long key;
2720 /**
2721 * Timeout in milliseconds to wait to acquire the mutex
2722 */
2723 unsigned int timeoutMs;
2724 } keyedMutex;
2725 unsigned int reserved[10];
2726 } params;
2727 /**
2728 * Only when ::CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS is used to wait on
2729 * a ::CUexternalSemaphore of type ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_NVSCISYNC,
2730 * the valid flag is ::CUDA_EXTERNAL_SEMAPHORE_WAIT_SKIP_NVSCIBUF_MEMSYNC
2731 * which indicates that while waiting for the ::CUexternalSemaphore, no memory
2732 * synchronization operations should be performed for any external memory
2733 * object imported as ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_NVSCIBUF.
2734 * For all other types of ::CUexternalSemaphore, flags must be zero.
2735 */
2736 unsigned int flags;
2737 unsigned int reserved[16];
2738} CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS_v1;
2739typedef CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS_v1 CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS;
2740
2741/**
2742 * Semaphore signal node parameters
2743 */
2744typedef struct CUDA_EXT_SEM_SIGNAL_NODE_PARAMS_st {
2745 CUexternalSemaphore* extSemArray; /**< Array of external semaphore handles. */
2746 const CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS* paramsArray; /**< Array of external semaphore signal parameters. */
2747 unsigned int numExtSems; /**< Number of handles and parameters supplied in extSemArray and paramsArray. */
2748} CUDA_EXT_SEM_SIGNAL_NODE_PARAMS_v1;
2749typedef CUDA_EXT_SEM_SIGNAL_NODE_PARAMS_v1 CUDA_EXT_SEM_SIGNAL_NODE_PARAMS;
2750
2751/**
2752 * Semaphore wait node parameters
2753 */
2754typedef struct CUDA_EXT_SEM_WAIT_NODE_PARAMS_st {
2755 CUexternalSemaphore* extSemArray; /**< Array of external semaphore handles. */
2756 const CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS* paramsArray; /**< Array of external semaphore wait parameters. */
2757 unsigned int numExtSems; /**< Number of handles and parameters supplied in extSemArray and paramsArray. */
2758} CUDA_EXT_SEM_WAIT_NODE_PARAMS_v1;
2759typedef CUDA_EXT_SEM_WAIT_NODE_PARAMS_v1 CUDA_EXT_SEM_WAIT_NODE_PARAMS;
2760
2761typedef unsigned long long CUmemGenericAllocationHandle_v1;
2762typedef CUmemGenericAllocationHandle_v1 CUmemGenericAllocationHandle;
2763
2764/**
2765 * Flags for specifying particular handle types
2766 */
2767typedef enum CUmemAllocationHandleType_enum {
2768 CU_MEM_HANDLE_TYPE_NONE = 0x0, /**< Does not allow any export mechanism. > */
2769 CU_MEM_HANDLE_TYPE_POSIX_FILE_DESCRIPTOR = 0x1, /**< Allows a file descriptor to be used for exporting. Permitted only on POSIX systems. (int) */
2770 CU_MEM_HANDLE_TYPE_WIN32 = 0x2, /**< Allows a Win32 NT handle to be used for exporting. (HANDLE) */
2771 CU_MEM_HANDLE_TYPE_WIN32_KMT = 0x4, /**< Allows a Win32 KMT handle to be used for exporting. (D3DKMT_HANDLE) */
2772 CU_MEM_HANDLE_TYPE_MAX = 0x7FFFFFFF
2773} CUmemAllocationHandleType;
2774
2775/**
2776 * Specifies the memory protection flags for mapping.
2777 */
2778typedef enum CUmemAccess_flags_enum {
2779 CU_MEM_ACCESS_FLAGS_PROT_NONE = 0x0, /**< Default, make the address range not accessible */
2780 CU_MEM_ACCESS_FLAGS_PROT_READ = 0x1, /**< Make the address range read accessible */
2781 CU_MEM_ACCESS_FLAGS_PROT_READWRITE = 0x3, /**< Make the address range read-write accessible */
2782 CU_MEM_ACCESS_FLAGS_PROT_MAX = 0x7FFFFFFF
2783} CUmemAccess_flags;
2784
2785/**
2786 * Specifies the type of location
2787 */
2788typedef enum CUmemLocationType_enum {
2789 CU_MEM_LOCATION_TYPE_INVALID = 0x0,
2790 CU_MEM_LOCATION_TYPE_DEVICE = 0x1, /**< Location is a device location, thus id is a device ordinal */
2791 CU_MEM_LOCATION_TYPE_MAX = 0x7FFFFFFF
2792} CUmemLocationType;
2793
2794/**
2795* Defines the allocation types available
2796*/
2797typedef enum CUmemAllocationType_enum {
2798 CU_MEM_ALLOCATION_TYPE_INVALID = 0x0,
2799
2800 /** This allocation type is 'pinned', i.e. cannot migrate from its current
2801 * location while the application is actively using it
2802 */
2803 CU_MEM_ALLOCATION_TYPE_PINNED = 0x1,
2804 CU_MEM_ALLOCATION_TYPE_MAX = 0x7FFFFFFF
2805} CUmemAllocationType;
2806
2807/**
2808* Flag for requesting different optimal and required granularities for an allocation.
2809*/
2810typedef enum CUmemAllocationGranularity_flags_enum {
2811 CU_MEM_ALLOC_GRANULARITY_MINIMUM = 0x0, /**< Minimum required granularity for allocation */
2812 CU_MEM_ALLOC_GRANULARITY_RECOMMENDED = 0x1 /**< Recommended granularity for allocation for best performance */
2813} CUmemAllocationGranularity_flags;
2814
2815/**
2816 * Sparse subresource types
2817 */
2818typedef enum CUarraySparseSubresourceType_enum {
2819 CU_ARRAY_SPARSE_SUBRESOURCE_TYPE_SPARSE_LEVEL = 0,
2820 CU_ARRAY_SPARSE_SUBRESOURCE_TYPE_MIPTAIL = 1
2821} CUarraySparseSubresourceType;
2822
2823/**
2824 * Memory operation types
2825 */
2826typedef enum CUmemOperationType_enum {
2827 CU_MEM_OPERATION_TYPE_MAP = 1,
2828 CU_MEM_OPERATION_TYPE_UNMAP = 2
2829} CUmemOperationType;
2830
2831/**
2832 * Memory handle types
2833 */
2834typedef enum CUmemHandleType_enum {
2835 CU_MEM_HANDLE_TYPE_GENERIC = 0
2836} CUmemHandleType;
2837
2838/**
2839 * Specifies the CUDA array or CUDA mipmapped array memory mapping information
2840 */
2841typedef struct CUarrayMapInfo_st {
2842 CUresourcetype resourceType; /**< Resource type */
2843
2844 union {
2845 CUmipmappedArray mipmap;
2846 CUarray array;
2847 } resource;
2848
2849 CUarraySparseSubresourceType subresourceType; /**< Sparse subresource type */
2850
2851 union {
2852 struct {
2853 unsigned int level; /**< For CUDA mipmapped arrays must a valid mipmap level. For CUDA arrays must be zero */
2854 unsigned int layer; /**< For CUDA layered arrays must be a valid layer index. Otherwise, must be zero */
2855 unsigned int offsetX; /**< Starting X offset in elements */
2856 unsigned int offsetY; /**< Starting Y offset in elements */
2857 unsigned int offsetZ; /**< Starting Z offset in elements */
2858 unsigned int extentWidth; /**< Width in elements */
2859 unsigned int extentHeight; /**< Height in elements */
2860 unsigned int extentDepth; /**< Depth in elements */
2861 } sparseLevel;
2862 struct {
2863 unsigned int layer; /**< For CUDA layered arrays must be a valid layer index. Otherwise, must be zero */
2864 unsigned long long offset; /**< Offset within mip tail */
2865 unsigned long long size; /**< Extent in bytes */
2866 } miptail;
2867 } subresource;
2868
2869 CUmemOperationType memOperationType; /**< Memory operation type */
2870 CUmemHandleType memHandleType; /**< Memory handle type */
2871
2872 union {
2873 CUmemGenericAllocationHandle memHandle;
2874 } memHandle;
2875
2876 unsigned long long offset; /**< Offset within the memory */
2877 unsigned int deviceBitMask; /**< Device ordinal bit mask */
2878 unsigned int flags; /**< flags for future use, must be zero now. */
2879 unsigned int reserved[2]; /**< Reserved for future use, must be zero now. */
2880} CUarrayMapInfo_v1;
2881typedef CUarrayMapInfo_v1 CUarrayMapInfo;
2882
2883/**
2884 * Specifies a memory location.
2885 */
2886typedef struct CUmemLocation_st {
2887 CUmemLocationType type; /**< Specifies the location type, which modifies the meaning of id. */
2888 int id; /**< identifier for a given this location's ::CUmemLocationType. */
2889} CUmemLocation_v1;
2890typedef CUmemLocation_v1 CUmemLocation;
2891
2892/**
2893 * Specifies compression attribute for an allocation.
2894 */
2895typedef enum CUmemAllocationCompType_enum {
2896 CU_MEM_ALLOCATION_COMP_NONE = 0x0, /**< Allocating non-compressible memory */
2897 CU_MEM_ALLOCATION_COMP_GENERIC = 0x1 /**< Allocating compressible memory */
2898} CUmemAllocationCompType;
2899
2900/**
2901 * This flag if set indicates that the memory will be used as a tile pool.
2902 */
2903#define CU_MEM_CREATE_USAGE_TILE_POOL 0x1
2904
2905/**
2906* Specifies the allocation properties for a allocation.
2907*/
2908typedef struct CUmemAllocationProp_st {
2909 /** Allocation type */
2910 CUmemAllocationType type;
2911 /** requested ::CUmemAllocationHandleType */
2912 CUmemAllocationHandleType requestedHandleTypes;
2913 /** Location of allocation */
2914 CUmemLocation location;
2915 /**
2916 * Windows-specific POBJECT_ATTRIBUTES required when
2917 * ::CU_MEM_HANDLE_TYPE_WIN32 is specified. This object attributes structure
2918 * includes security attributes that define
2919 * the scope of which exported allocations may be transferred to other
2920 * processes. In all other cases, this field is required to be zero.
2921 */
2922 void *win32HandleMetaData;
2923 struct {
2924 /**
2925 * Allocation hint for requesting compressible memory.
2926 * On devices that support Compute Data Compression, compressible
2927 * memory can be used to accelerate accesses to data with unstructured
2928 * sparsity and other compressible data patterns. Applications are
2929 * expected to query allocation property of the handle obtained with
2930 * ::cuMemCreate using ::cuMemGetAllocationPropertiesFromHandle to
2931 * validate if the obtained allocation is compressible or not. Note that
2932 * compressed memory may not be mappable on all devices.
2933 */
2934 unsigned char compressionType;
2935 unsigned char gpuDirectRDMACapable;
2936 /** Bitmask indicating intended usage for this allocation */
2937 unsigned short usage;
2938 unsigned char reserved[4];
2939 } allocFlags;
2940} CUmemAllocationProp_v1;
2941typedef CUmemAllocationProp_v1 CUmemAllocationProp;
2942
2943/**
2944 * Memory access descriptor
2945 */
2946typedef struct CUmemAccessDesc_st {
2947 CUmemLocation location; /**< Location on which the request is to change it's accessibility */
2948 CUmemAccess_flags flags; /**< ::CUmemProt accessibility flags to set on the request */
2949} CUmemAccessDesc_v1;
2950typedef CUmemAccessDesc_v1 CUmemAccessDesc;
2951
2952typedef enum CUgraphExecUpdateResult_enum {
2953 CU_GRAPH_EXEC_UPDATE_SUCCESS = 0x0, /**< The update succeeded */
2954 CU_GRAPH_EXEC_UPDATE_ERROR = 0x1, /**< The update failed for an unexpected reason which is described in the return value of the function */
2955 CU_GRAPH_EXEC_UPDATE_ERROR_TOPOLOGY_CHANGED = 0x2, /**< The update failed because the topology changed */
2956 CU_GRAPH_EXEC_UPDATE_ERROR_NODE_TYPE_CHANGED = 0x3, /**< The update failed because a node type changed */
2957 CU_GRAPH_EXEC_UPDATE_ERROR_FUNCTION_CHANGED = 0x4, /**< The update failed because the function of a kernel node changed (CUDA driver < 11.2) */
2958 CU_GRAPH_EXEC_UPDATE_ERROR_PARAMETERS_CHANGED = 0x5, /**< The update failed because the parameters changed in a way that is not supported */
2959 CU_GRAPH_EXEC_UPDATE_ERROR_NOT_SUPPORTED = 0x6, /**< The update failed because something about the node is not supported */
2960 CU_GRAPH_EXEC_UPDATE_ERROR_UNSUPPORTED_FUNCTION_CHANGE = 0x7 /**< The update failed because the function of a kernel node changed in an unsupported way */
2961} CUgraphExecUpdateResult;
2962
2963/**
2964 * CUDA memory pool attributes
2965 */
2966typedef enum CUmemPool_attribute_enum {
2967 /**
2968 * (value type = int)
2969 * Allow cuMemAllocAsync to use memory asynchronously freed
2970 * in another streams as long as a stream ordering dependency
2971 * of the allocating stream on the free action exists.
2972 * Cuda events and null stream interactions can create the required
2973 * stream ordered dependencies. (default enabled)
2974 */
2975 CU_MEMPOOL_ATTR_REUSE_FOLLOW_EVENT_DEPENDENCIES = 1,
2976
2977 /**
2978 * (value type = int)
2979 * Allow reuse of already completed frees when there is no dependency
2980 * between the free and allocation. (default enabled)
2981 */
2982 CU_MEMPOOL_ATTR_REUSE_ALLOW_OPPORTUNISTIC,
2983
2984 /**
2985 * (value type = int)
2986 * Allow cuMemAllocAsync to insert new stream dependencies
2987 * in order to establish the stream ordering required to reuse
2988 * a piece of memory released by cuFreeAsync (default enabled).
2989 */
2990 CU_MEMPOOL_ATTR_REUSE_ALLOW_INTERNAL_DEPENDENCIES,
2991
2992 /**
2993 * (value type = cuuint64_t)
2994 * Amount of reserved memory in bytes to hold onto before trying
2995 * to release memory back to the OS. When more than the release
2996 * threshold bytes of memory are held by the memory pool, the
2997 * allocator will try to release memory back to the OS on the
2998 * next call to stream, event or context synchronize. (default 0)
2999 */
3000 CU_MEMPOOL_ATTR_RELEASE_THRESHOLD,
3001
3002 /**
3003 * (value type = cuuint64_t)
3004 * Amount of backing memory currently allocated for the mempool.
3005 */
3006 CU_MEMPOOL_ATTR_RESERVED_MEM_CURRENT,
3007
3008 /**
3009 * (value type = cuuint64_t)
3010 * High watermark of backing memory allocated for the mempool since the
3011 * last time it was reset. High watermark can only be reset to zero.
3012 */
3013 CU_MEMPOOL_ATTR_RESERVED_MEM_HIGH,
3014
3015 /**
3016 * (value type = cuuint64_t)
3017 * Amount of memory from the pool that is currently in use by the application.
3018 */
3019 CU_MEMPOOL_ATTR_USED_MEM_CURRENT,
3020
3021 /**
3022 * (value type = cuuint64_t)
3023 * High watermark of the amount of memory from the pool that was in use by the application since
3024 * the last time it was reset. High watermark can only be reset to zero.
3025 */
3026 CU_MEMPOOL_ATTR_USED_MEM_HIGH
3027} CUmemPool_attribute;
3028
3029/**
3030 * Specifies the properties of allocations made from the pool.
3031 */
3032typedef struct CUmemPoolProps_st {
3033 CUmemAllocationType allocType; /**< Allocation type. Currently must be specified as CU_MEM_ALLOCATION_TYPE_PINNED */
3034 CUmemAllocationHandleType handleTypes; /**< Handle types that will be supported by allocations from the pool. */
3035 CUmemLocation location; /**< Location where allocations should reside. */
3036 /**
3037 * Windows-specific LPSECURITYATTRIBUTES required when
3038 * ::CU_MEM_HANDLE_TYPE_WIN32 is specified. This security attribute defines
3039 * the scope of which exported allocations may be transferred to other
3040 * processes. In all other cases, this field is required to be zero.
3041 */
3042 void *win32SecurityAttributes;
3043 unsigned char reserved[64]; /**< reserved for future use, must be 0 */
3044} CUmemPoolProps_v1;
3045typedef CUmemPoolProps_v1 CUmemPoolProps;
3046
3047/**
3048 * Opaque data for exporting a pool allocation
3049 */
3050typedef struct CUmemPoolPtrExportData_st {
3051 unsigned char reserved[64];
3052} CUmemPoolPtrExportData_v1;
3053typedef CUmemPoolPtrExportData_v1 CUmemPoolPtrExportData;
3054
3055/**
3056 * Memory allocation node parameters
3057 */
3058typedef struct CUDA_MEM_ALLOC_NODE_PARAMS_st {
3059 /**
3060 * in: location where the allocation should reside (specified in ::location).
3061 * ::handleTypes must be ::CU_MEM_HANDLE_TYPE_NONE. IPC is not supported.
3062 */
3063 CUmemPoolProps poolProps;
3064 const CUmemAccessDesc *accessDescs; /**< in: array of memory access descriptors. Used to describe peer GPU access */
3065 size_t accessDescCount; /**< in: number of memory access descriptors. Must not exceed the number of GPUs. */
3066 size_t bytesize; /**< in: size in bytes of the requested allocation */
3067 CUdeviceptr dptr; /**< out: address of the allocation returned by CUDA */
3068} CUDA_MEM_ALLOC_NODE_PARAMS;
3069
3070typedef enum CUgraphMem_attribute_enum {
3071 /**
3072 * (value type = cuuint64_t)
3073 * Amount of memory, in bytes, currently associated with graphs
3074 */
3075 CU_GRAPH_MEM_ATTR_USED_MEM_CURRENT,
3076
3077 /**
3078 * (value type = cuuint64_t)
3079 * High watermark of memory, in bytes, associated with graphs since the
3080 * last time it was reset. High watermark can only be reset to zero.
3081 */
3082 CU_GRAPH_MEM_ATTR_USED_MEM_HIGH,
3083
3084 /**
3085 * (value type = cuuint64_t)
3086 * Amount of memory, in bytes, currently allocated for use by
3087 * the CUDA graphs asynchronous allocator.
3088 */
3089 CU_GRAPH_MEM_ATTR_RESERVED_MEM_CURRENT,
3090
3091 /**
3092 * (value type = cuuint64_t)
3093 * High watermark of memory, in bytes, currently allocated for use by
3094 * the CUDA graphs asynchronous allocator.
3095 */
3096 CU_GRAPH_MEM_ATTR_RESERVED_MEM_HIGH
3097} CUgraphMem_attribute;
3098
3099/**
3100 * If set, each kernel launched as part of ::cuLaunchCooperativeKernelMultiDevice only
3101 * waits for prior work in the stream corresponding to that GPU to complete before the
3102 * kernel begins execution.
3103 */
3104#define CUDA_COOPERATIVE_LAUNCH_MULTI_DEVICE_NO_PRE_LAUNCH_SYNC 0x01
3105
3106/**
3107 * If set, any subsequent work pushed in a stream that participated in a call to
3108 * ::cuLaunchCooperativeKernelMultiDevice will only wait for the kernel launched on
3109 * the GPU corresponding to that stream to complete before it begins execution.
3110 */
3111#define CUDA_COOPERATIVE_LAUNCH_MULTI_DEVICE_NO_POST_LAUNCH_SYNC 0x02
3112
3113/**
3114 * If set, the CUDA array is a collection of layers, where each layer is either a 1D
3115 * or a 2D array and the Depth member of CUDA_ARRAY3D_DESCRIPTOR specifies the number
3116 * of layers, not the depth of a 3D array.
3117 */
3118#define CUDA_ARRAY3D_LAYERED 0x01
3119
3120/**
3121 * Deprecated, use CUDA_ARRAY3D_LAYERED
3122 */
3123#define CUDA_ARRAY3D_2DARRAY 0x01
3124
3125/**
3126 * This flag must be set in order to bind a surface reference
3127 * to the CUDA array
3128 */
3129#define CUDA_ARRAY3D_SURFACE_LDST 0x02
3130
3131/**
3132 * If set, the CUDA array is a collection of six 2D arrays, representing faces of a cube. The
3133 * width of such a CUDA array must be equal to its height, and Depth must be six.
3134 * If ::CUDA_ARRAY3D_LAYERED flag is also set, then the CUDA array is a collection of cubemaps
3135 * and Depth must be a multiple of six.
3136 */
3137#define CUDA_ARRAY3D_CUBEMAP 0x04
3138
3139/**
3140 * This flag must be set in order to perform texture gather operations
3141 * on a CUDA array.
3142 */
3143#define CUDA_ARRAY3D_TEXTURE_GATHER 0x08
3144
3145/**
3146 * This flag if set indicates that the CUDA
3147 * array is a DEPTH_TEXTURE.
3148 */
3149#define CUDA_ARRAY3D_DEPTH_TEXTURE 0x10
3150
3151/**
3152 * This flag indicates that the CUDA array may be bound as a color target
3153 * in an external graphics API
3154 */
3155#define CUDA_ARRAY3D_COLOR_ATTACHMENT 0x20
3156
3157/**
3158 * This flag if set indicates that the CUDA array or CUDA mipmapped array
3159 * is a sparse CUDA array or CUDA mipmapped array respectively
3160 */
3161#define CUDA_ARRAY3D_SPARSE 0x40
3162
3163/**
3164 * Override the texref format with a format inferred from the array.
3165 * Flag for ::cuTexRefSetArray()
3166 */
3167#define CU_TRSA_OVERRIDE_FORMAT 0x01
3168
3169/**
3170 * Read the texture as integers rather than promoting the values to floats
3171 * in the range [0,1].
3172 * Flag for ::cuTexRefSetFlags() and ::cuTexObjectCreate()
3173 */
3174#define CU_TRSF_READ_AS_INTEGER 0x01
3175
3176/**
3177 * Use normalized texture coordinates in the range [0,1) instead of [0,dim).
3178 * Flag for ::cuTexRefSetFlags() and ::cuTexObjectCreate()
3179 */
3180#define CU_TRSF_NORMALIZED_COORDINATES 0x02
3181
3182/**
3183 * Perform sRGB->linear conversion during texture read.
3184 * Flag for ::cuTexRefSetFlags() and ::cuTexObjectCreate()
3185 */
3186#define CU_TRSF_SRGB 0x10
3187
3188 /**
3189 * Disable any trilinear filtering optimizations.
3190 * Flag for ::cuTexRefSetFlags() and ::cuTexObjectCreate()
3191 */
3192#define CU_TRSF_DISABLE_TRILINEAR_OPTIMIZATION 0x20
3193
3194/**
3195 * End of array terminator for the \p extra parameter to
3196 * ::cuLaunchKernel
3197 */
3198#define CU_LAUNCH_PARAM_END ((void*)0x00)
3199
3200/**
3201 * Indicator that the next value in the \p extra parameter to
3202 * ::cuLaunchKernel will be a pointer to a buffer containing all kernel
3203 * parameters used for launching kernel \p f. This buffer needs to
3204 * honor all alignment/padding requirements of the individual parameters.
3205 * If ::CU_LAUNCH_PARAM_BUFFER_SIZE is not also specified in the
3206 * \p extra array, then ::CU_LAUNCH_PARAM_BUFFER_POINTER will have no
3207 * effect.
3208 */
3209#define CU_LAUNCH_PARAM_BUFFER_POINTER ((void*)0x01)
3210
3211/**
3212 * Indicator that the next value in the \p extra parameter to
3213 * ::cuLaunchKernel will be a pointer to a size_t which contains the
3214 * size of the buffer specified with ::CU_LAUNCH_PARAM_BUFFER_POINTER.
3215 * It is required that ::CU_LAUNCH_PARAM_BUFFER_POINTER also be specified
3216 * in the \p extra array if the value associated with
3217 * ::CU_LAUNCH_PARAM_BUFFER_SIZE is not zero.
3218 */
3219#define CU_LAUNCH_PARAM_BUFFER_SIZE ((void*)0x02)
3220
3221/**
3222 * For texture references loaded into the module, use default texunit from
3223 * texture reference.
3224 */
3225#define CU_PARAM_TR_DEFAULT -1
3226
3227/**
3228 * Device that represents the CPU
3229 */
3230#define CU_DEVICE_CPU ((CUdevice)-1)
3231
3232/**
3233 * Device that represents an invalid device
3234 */
3235#define CU_DEVICE_INVALID ((CUdevice)-2)
3236
3237/**
3238 * Bitmasks for ::CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_FLUSH_WRITES_OPTIONS
3239 */
3240typedef enum CUflushGPUDirectRDMAWritesOptions_enum {
3241 CU_FLUSH_GPU_DIRECT_RDMA_WRITES_OPTION_HOST = 1<<0, /**< ::cuFlushGPUDirectRDMAWrites() and its CUDA Runtime API counterpart are supported on the device. */
3242 CU_FLUSH_GPU_DIRECT_RDMA_WRITES_OPTION_MEMOPS = 1<<1 /**< The ::CU_STREAM_WAIT_VALUE_FLUSH flag and the ::CU_STREAM_MEM_OP_FLUSH_REMOTE_WRITES MemOp are supported on the device. */
3243} CUflushGPUDirectRDMAWritesOptions;
3244
3245/**
3246 * Platform native ordering for GPUDirect RDMA writes
3247 */
3248typedef enum CUGPUDirectRDMAWritesOrdering_enum {
3249 CU_GPU_DIRECT_RDMA_WRITES_ORDERING_NONE = 0, /**< The device does not natively support ordering of remote writes. ::cuFlushGPUDirectRDMAWrites() can be leveraged if supported. */
3250 CU_GPU_DIRECT_RDMA_WRITES_ORDERING_OWNER = 100, /**< Natively, the device can consistently consume remote writes, although other CUDA devices may not. */
3251 CU_GPU_DIRECT_RDMA_WRITES_ORDERING_ALL_DEVICES = 200 /**< Any CUDA device in the system can consistently consume remote writes to this device. */
3252} CUGPUDirectRDMAWritesOrdering;
3253
3254/**
3255 * The scopes for ::cuFlushGPUDirectRDMAWrites
3256 */
3257typedef enum CUflushGPUDirectRDMAWritesScope_enum {
3258 CU_FLUSH_GPU_DIRECT_RDMA_WRITES_TO_OWNER = 100, /**< Blocks until remote writes are visible to the CUDA device context owning the data. */
3259 CU_FLUSH_GPU_DIRECT_RDMA_WRITES_TO_ALL_DEVICES = 200 /**< Blocks until remote writes are visible to all CUDA device contexts. */
3260} CUflushGPUDirectRDMAWritesScope;
3261
3262/**
3263 * The targets for ::cuFlushGPUDirectRDMAWrites
3264 */
3265typedef enum CUflushGPUDirectRDMAWritesTarget_enum {
3266 CU_FLUSH_GPU_DIRECT_RDMA_WRITES_TARGET_CURRENT_CTX = 0 /**< Sets the target for ::cuFlushGPUDirectRDMAWrites() to the currently active CUDA device context. */
3267} CUflushGPUDirectRDMAWritesTarget;
3268
3269/**
3270 * The additional write options for ::cuGraphDebugDotPrint
3271 */
3272typedef enum CUgraphDebugDot_flags_enum {
3273 CU_GRAPH_DEBUG_DOT_FLAGS_VERBOSE = 1<<0, /** Output all debug data as if every debug flag is enabled */
3274 CU_GRAPH_DEBUG_DOT_FLAGS_RUNTIME_TYPES = 1<<1, /** Use CUDA Runtime structures for output */
3275 CU_GRAPH_DEBUG_DOT_FLAGS_KERNEL_NODE_PARAMS = 1<<2, /** Adds CUDA_KERNEL_NODE_PARAMS values to output */
3276 CU_GRAPH_DEBUG_DOT_FLAGS_MEMCPY_NODE_PARAMS = 1<<3, /** Adds CUDA_MEMCPY3D values to output */
3277 CU_GRAPH_DEBUG_DOT_FLAGS_MEMSET_NODE_PARAMS = 1<<4, /** Adds CUDA_MEMSET_NODE_PARAMS values to output */
3278 CU_GRAPH_DEBUG_DOT_FLAGS_HOST_NODE_PARAMS = 1<<5, /** Adds CUDA_HOST_NODE_PARAMS values to output */
3279 CU_GRAPH_DEBUG_DOT_FLAGS_EVENT_NODE_PARAMS = 1<<6, /** Adds CUevent handle from record and wait nodes to output */
3280 CU_GRAPH_DEBUG_DOT_FLAGS_EXT_SEMAS_SIGNAL_NODE_PARAMS = 1<<7, /** Adds CUDA_EXT_SEM_SIGNAL_NODE_PARAMS values to output */
3281 CU_GRAPH_DEBUG_DOT_FLAGS_EXT_SEMAS_WAIT_NODE_PARAMS = 1<<8, /** Adds CUDA_EXT_SEM_WAIT_NODE_PARAMS values to output */
3282 CU_GRAPH_DEBUG_DOT_FLAGS_KERNEL_NODE_ATTRIBUTES = 1<<9, /** Adds CUkernelNodeAttrValue values to output */
3283 CU_GRAPH_DEBUG_DOT_FLAGS_HANDLES = 1<<10, /** Adds node handles and every kernel function handle to output */
3284 CU_GRAPH_DEBUG_DOT_FLAGS_MEM_ALLOC_NODE_PARAMS = 1<<11, /** Adds memory alloc node parameters to output */
3285 CU_GRAPH_DEBUG_DOT_FLAGS_MEM_FREE_NODE_PARAMS = 1<<12 /** Adds memory free node parameters to output */
3286} CUgraphDebugDot_flags;
3287
3288/**
3289 * Flags for user objects for graphs
3290 */
3291typedef enum CUuserObject_flags_enum {
3292 CU_USER_OBJECT_NO_DESTRUCTOR_SYNC = 1 /**< Indicates the destructor execution is not synchronized by any CUDA handle. */
3293} CUuserObject_flags;
3294
3295/**
3296 * Flags for retaining user object references for graphs
3297 */
3298typedef enum CUuserObjectRetain_flags_enum {
3299 CU_GRAPH_USER_OBJECT_MOVE = 1 /**< Transfer references from the caller rather than creating new references. */
3300} CUuserObjectRetain_flags;
3301
3302/**
3303 * Flags for instantiating a graph
3304 */
3305typedef enum CUgraphInstantiate_flags_enum {
3306 CUDA_GRAPH_INSTANTIATE_FLAG_AUTO_FREE_ON_LAUNCH = 1 /**< Automatically free memory allocated in a graph before relaunching. */
3307} CUgraphInstantiate_flags;
3308
3309/** @} */ /* END CUDA_TYPES */
3310
3311#if defined(__GNUC__)
3312 #if defined(__CUDA_API_PUSH_VISIBILITY_DEFAULT)
3313 #pragma GCC visibility push(default)
3314 #endif
3315#endif
3316
3317#ifdef _WIN32
3318#define CUDAAPI __stdcall
3319#else
3320#define CUDAAPI
3321#endif
3322
3323/**
3324 * \defgroup CUDA_ERROR Error Handling
3325 *
3326 * ___MANBRIEF___ error handling functions of the low-level CUDA driver API
3327 * (___CURRENT_FILE___) ___ENDMANBRIEF___
3328 *
3329 * This section describes the error handling functions of the low-level CUDA
3330 * driver application programming interface.
3331 *
3332 * @{
3333 */
3334
3335/**
3336 * \brief Gets the string description of an error code
3337 *
3338 * Sets \p *pStr to the address of a NULL-terminated string description
3339 * of the error code \p error.
3340 * If the error code is not recognized, ::CUDA_ERROR_INVALID_VALUE
3341 * will be returned and \p *pStr will be set to the NULL address.
3342 *
3343 * \param error - Error code to convert to string
3344 * \param pStr - Address of the string pointer.
3345 *
3346 * \return
3347 * ::CUDA_SUCCESS,
3348 * ::CUDA_ERROR_INVALID_VALUE
3349 *
3350 * \sa
3351 * ::CUresult,
3352 * ::cudaGetErrorString
3353 */
3354CUresult CUDAAPI cuGetErrorString(CUresult error, const char **pStr);
3355
3356/**
3357 * \brief Gets the string representation of an error code enum name
3358 *
3359 * Sets \p *pStr to the address of a NULL-terminated string representation
3360 * of the name of the enum error code \p error.
3361 * If the error code is not recognized, ::CUDA_ERROR_INVALID_VALUE
3362 * will be returned and \p *pStr will be set to the NULL address.
3363 *
3364 * \param error - Error code to convert to string
3365 * \param pStr - Address of the string pointer.
3366 *
3367 * \return
3368 * ::CUDA_SUCCESS,
3369 * ::CUDA_ERROR_INVALID_VALUE
3370 *
3371 * \sa
3372 * ::CUresult,
3373 * ::cudaGetErrorName
3374 */
3375CUresult CUDAAPI cuGetErrorName(CUresult error, const char **pStr);
3376
3377/** @} */ /* END CUDA_ERROR */
3378
3379/**
3380 * \defgroup CUDA_INITIALIZE Initialization
3381 *
3382 * ___MANBRIEF___ initialization functions of the low-level CUDA driver API
3383 * (___CURRENT_FILE___) ___ENDMANBRIEF___
3384 *
3385 * This section describes the initialization functions of the low-level CUDA
3386 * driver application programming interface.
3387 *
3388 * @{
3389 */
3390
3391/**
3392 * \brief Initialize the CUDA driver API
3393 *
3394 * Initializes the driver API and must be called before any other function from
3395 * the driver API. Currently, the \p Flags parameter must be 0. If ::cuInit()
3396 * has not been called, any function from the driver API will return
3397 * ::CUDA_ERROR_NOT_INITIALIZED.
3398 *
3399 * \param Flags - Initialization flag for CUDA.
3400 *
3401 * \return
3402 * ::CUDA_SUCCESS,
3403 * ::CUDA_ERROR_INVALID_VALUE,
3404 * ::CUDA_ERROR_INVALID_DEVICE,
3405 * ::CUDA_ERROR_SYSTEM_DRIVER_MISMATCH,
3406 * ::CUDA_ERROR_COMPAT_NOT_SUPPORTED_ON_DEVICE
3407 * \notefnerr
3408 */
3409CUresult CUDAAPI cuInit(unsigned int Flags);
3410
3411/** @} */ /* END CUDA_INITIALIZE */
3412
3413/**
3414 * \defgroup CUDA_VERSION Version Management
3415 *
3416 * ___MANBRIEF___ version management functions of the low-level CUDA driver
3417 * API (___CURRENT_FILE___) ___ENDMANBRIEF___
3418 *
3419 * This section describes the version management functions of the low-level
3420 * CUDA driver application programming interface.
3421 *
3422 * @{
3423 */
3424
3425/**
3426 * \brief Returns the latest CUDA version supported by driver
3427 *
3428 * Returns in \p *driverVersion the version of CUDA supported by
3429 * the driver. The version is returned as
3430 * (1000 &times; major + 10 &times; minor). For example, CUDA 9.2
3431 * would be represented by 9020.
3432 *
3433 * This function automatically returns ::CUDA_ERROR_INVALID_VALUE if
3434 * \p driverVersion is NULL.
3435 *
3436 * \param driverVersion - Returns the CUDA driver version
3437 *
3438 * \return
3439 * ::CUDA_SUCCESS,
3440 * ::CUDA_ERROR_INVALID_VALUE
3441 * \notefnerr
3442 *
3443 * \sa
3444 * ::cudaDriverGetVersion,
3445 * ::cudaRuntimeGetVersion
3446 */
3447CUresult CUDAAPI cuDriverGetVersion(int *driverVersion);
3448
3449/** @} */ /* END CUDA_VERSION */
3450
3451/**
3452 * \defgroup CUDA_DEVICE Device Management
3453 *
3454 * ___MANBRIEF___ device management functions of the low-level CUDA driver API
3455 * (___CURRENT_FILE___) ___ENDMANBRIEF___
3456 *
3457 * This section describes the device management functions of the low-level
3458 * CUDA driver application programming interface.
3459 *
3460 * @{
3461 */
3462
3463/**
3464 * \brief Returns a handle to a compute device
3465 *
3466 * Returns in \p *device a device handle given an ordinal in the range <b>[0,
3467 * ::cuDeviceGetCount()-1]</b>.
3468 *
3469 * \param device - Returned device handle
3470 * \param ordinal - Device number to get handle for
3471 *
3472 * \return
3473 * ::CUDA_SUCCESS,
3474 * ::CUDA_ERROR_DEINITIALIZED,
3475 * ::CUDA_ERROR_NOT_INITIALIZED,
3476 * ::CUDA_ERROR_INVALID_CONTEXT,
3477 * ::CUDA_ERROR_INVALID_VALUE,
3478 * ::CUDA_ERROR_INVALID_DEVICE
3479 * \notefnerr
3480 *
3481 * \sa
3482 * ::cuDeviceGetAttribute,
3483 * ::cuDeviceGetCount,
3484 * ::cuDeviceGetName,
3485 * ::cuDeviceGetUuid,
3486 * ::cuDeviceGetLuid,
3487 * ::cuDeviceTotalMem,
3488 * ::cuDeviceGetExecAffinitySupport
3489 */
3490CUresult CUDAAPI cuDeviceGet(CUdevice *device, int ordinal);
3491
3492/**
3493 * \brief Returns the number of compute-capable devices
3494 *
3495 * Returns in \p *count the number of devices with compute capability greater
3496 * than or equal to 2.0 that are available for execution. If there is no such
3497 * device, ::cuDeviceGetCount() returns 0.
3498 *
3499 * \param count - Returned number of compute-capable devices
3500 *
3501 * \return
3502 * ::CUDA_SUCCESS,
3503 * ::CUDA_ERROR_DEINITIALIZED,
3504 * ::CUDA_ERROR_NOT_INITIALIZED,
3505 * ::CUDA_ERROR_INVALID_CONTEXT,
3506 * ::CUDA_ERROR_INVALID_VALUE
3507 * \notefnerr
3508 *
3509 * \sa
3510 * ::cuDeviceGetAttribute,
3511 * ::cuDeviceGetName,
3512 * ::cuDeviceGetUuid,
3513 * ::cuDeviceGetLuid,
3514 * ::cuDeviceGet,
3515 * ::cuDeviceTotalMem,
3516 * ::cuDeviceGetExecAffinitySupport,
3517 * ::cudaGetDeviceCount
3518 */
3519CUresult CUDAAPI cuDeviceGetCount(int *count);
3520
3521/**
3522 * \brief Returns an identifier string for the device
3523 *
3524 * Returns an ASCII string identifying the device \p dev in the NULL-terminated
3525 * string pointed to by \p name. \p len specifies the maximum length of the
3526 * string that may be returned.
3527 *
3528 * \param name - Returned identifier string for the device
3529 * \param len - Maximum length of string to store in \p name
3530 * \param dev - Device to get identifier string for
3531 *
3532 * \return
3533 * ::CUDA_SUCCESS,
3534 * ::CUDA_ERROR_DEINITIALIZED,
3535 * ::CUDA_ERROR_NOT_INITIALIZED,
3536 * ::CUDA_ERROR_INVALID_CONTEXT,
3537 * ::CUDA_ERROR_INVALID_VALUE,
3538 * ::CUDA_ERROR_INVALID_DEVICE
3539 * \notefnerr
3540 *
3541 * \sa
3542 * ::cuDeviceGetAttribute,
3543 * ::cuDeviceGetUuid,
3544 * ::cuDeviceGetLuid,
3545 * ::cuDeviceGetCount,
3546 * ::cuDeviceGet,
3547 * ::cuDeviceTotalMem,
3548 * ::cuDeviceGetExecAffinitySupport,
3549 * ::cudaGetDeviceProperties
3550 */
3551CUresult CUDAAPI cuDeviceGetName(char *name, int len, CUdevice dev);
3552
3553/**
3554 * \brief Return an UUID for the device
3555 *
3556 * Note there is a later version of this API, ::cuDeviceGetUuid_v2. It will
3557 * supplant this version in 12.0, which is retained for minor version compatibility.
3558 *
3559 * Returns 16-octets identifying the device \p dev in the structure
3560 * pointed by the \p uuid.
3561 *
3562 * \param uuid - Returned UUID
3563 * \param dev - Device to get identifier string for
3564 *
3565 * \return
3566 * ::CUDA_SUCCESS,
3567 * ::CUDA_ERROR_DEINITIALIZED,
3568 * ::CUDA_ERROR_NOT_INITIALIZED,
3569 * ::CUDA_ERROR_INVALID_VALUE,
3570 * ::CUDA_ERROR_INVALID_DEVICE
3571 * \notefnerr
3572 *
3573 * \sa
3574 * ::cuDeviceGetUuid_v2
3575 * ::cuDeviceGetAttribute,
3576 * ::cuDeviceGetCount,
3577 * ::cuDeviceGetName,
3578 * ::cuDeviceGetLuid,
3579 * ::cuDeviceGet,
3580 * ::cuDeviceTotalMem,
3581 * ::cuDeviceGetExecAffinitySupport,
3582 * ::cudaGetDeviceProperties
3583 */
3584CUresult CUDAAPI cuDeviceGetUuid(CUuuid *uuid, CUdevice dev);
3585
3586/**
3587 * \brief Return an UUID for the device (11.4+)
3588 *
3589 * Returns 16-octets identifying the device \p dev in the structure
3590 * pointed by the \p uuid. If the device is in MIG mode, returns its
3591 * MIG UUID which uniquely identifies the subscribed MIG compute instance.
3592 *
3593 * \param uuid - Returned UUID
3594 * \param dev - Device to get identifier string for
3595 *
3596 * \return
3597 * ::CUDA_SUCCESS,
3598 * ::CUDA_ERROR_DEINITIALIZED,
3599 * ::CUDA_ERROR_NOT_INITIALIZED,
3600 * ::CUDA_ERROR_INVALID_VALUE,
3601 * ::CUDA_ERROR_INVALID_DEVICE
3602 * \notefnerr
3603 *
3604 * \sa
3605 * ::cuDeviceGetAttribute,
3606 * ::cuDeviceGetCount,
3607 * ::cuDeviceGetName,
3608 * ::cuDeviceGetLuid,
3609 * ::cuDeviceGet,
3610 * ::cuDeviceTotalMem,
3611 * ::cudaGetDeviceProperties
3612 */
3613CUresult CUDAAPI cuDeviceGetUuid_v2(CUuuid *uuid, CUdevice dev);
3614
3615/**
3616 * \brief Return an LUID and device node mask for the device
3617 *
3618 * Return identifying information (\p luid and \p deviceNodeMask) to allow
3619 * matching device with graphics APIs.
3620 *
3621 * \param luid - Returned LUID
3622 * \param deviceNodeMask - Returned device node mask
3623 * \param dev - Device to get identifier string for
3624 *
3625 * \return
3626 * ::CUDA_SUCCESS,
3627 * ::CUDA_ERROR_DEINITIALIZED,
3628 * ::CUDA_ERROR_NOT_INITIALIZED,
3629 * ::CUDA_ERROR_INVALID_VALUE,
3630 * ::CUDA_ERROR_INVALID_DEVICE
3631 * \notefnerr
3632 *
3633 * \sa
3634 * ::cuDeviceGetAttribute,
3635 * ::cuDeviceGetCount,
3636 * ::cuDeviceGetName,
3637 * ::cuDeviceGet,
3638 * ::cuDeviceTotalMem,
3639 * ::cuDeviceGetExecAffinitySupport,
3640 * ::cudaGetDeviceProperties
3641 */
3642CUresult CUDAAPI cuDeviceGetLuid(char *luid, unsigned int *deviceNodeMask, CUdevice dev);
3643
3644/**
3645 * \brief Returns the total amount of memory on the device
3646 *
3647 * Returns in \p *bytes the total amount of memory available on the device
3648 * \p dev in bytes.
3649 *
3650 * \param bytes - Returned memory available on device in bytes
3651 * \param dev - Device handle
3652 *
3653 * \return
3654 * ::CUDA_SUCCESS,
3655 * ::CUDA_ERROR_DEINITIALIZED,
3656 * ::CUDA_ERROR_NOT_INITIALIZED,
3657 * ::CUDA_ERROR_INVALID_CONTEXT,
3658 * ::CUDA_ERROR_INVALID_VALUE,
3659 * ::CUDA_ERROR_INVALID_DEVICE
3660 * \notefnerr
3661 *
3662 * \sa
3663 * ::cuDeviceGetAttribute,
3664 * ::cuDeviceGetCount,
3665 * ::cuDeviceGetName,
3666 * ::cuDeviceGetUuid,
3667 * ::cuDeviceGet,
3668 * ::cuDeviceGetExecAffinitySupport,
3669 * ::cudaMemGetInfo
3670 */
3671CUresult CUDAAPI cuDeviceTotalMem(size_t *bytes, CUdevice dev);
3672
3673/**
3674 * \brief Returns the maximum number of elements allocatable in a 1D linear texture for a given texture element size.
3675 *
3676 * Returns in \p maxWidthInElements the maximum number of texture elements allocatable in a 1D linear texture
3677 * for given \p format and \p numChannels.
3678 *
3679 * \param maxWidthInElements - Returned maximum number of texture elements allocatable for given \p format and \p numChannels.
3680 * \param format - Texture format.
3681 * \param numChannels - Number of channels per texture element.
3682 * \param dev - Device handle.
3683 *
3684 * \return
3685 * ::CUDA_SUCCESS,
3686 * ::CUDA_ERROR_DEINITIALIZED,
3687 * ::CUDA_ERROR_NOT_INITIALIZED,
3688 * ::CUDA_ERROR_INVALID_CONTEXT,
3689 * ::CUDA_ERROR_INVALID_VALUE,
3690 * ::CUDA_ERROR_INVALID_DEVICE
3691 * \notefnerr
3692 *
3693 * \sa
3694 * ::cuDeviceGetAttribute,
3695 * ::cuDeviceGetCount,
3696 * ::cuDeviceGetName,
3697 * ::cuDeviceGetUuid,
3698 * ::cuDeviceGet,
3699 * ::cudaMemGetInfo,
3700 * ::cuDeviceTotalMem
3701 */
3702CUresult CUDAAPI cuDeviceGetTexture1DLinearMaxWidth(size_t *maxWidthInElements, CUarray_format format, unsigned numChannels, CUdevice dev);
3703
3704/**
3705 * \brief Returns information about the device
3706 *
3707 * Returns in \p *pi the integer value of the attribute \p attrib on device
3708 * \p dev. The supported attributes are:
3709 * - ::CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK: Maximum number of threads per
3710 * block;
3711 * - ::CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_X: Maximum x-dimension of a block;
3712 * - ::CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Y: Maximum y-dimension of a block;
3713 * - ::CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Z: Maximum z-dimension of a block;
3714 * - ::CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_X: Maximum x-dimension of a grid;
3715 * - ::CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Y: Maximum y-dimension of a grid;
3716 * - ::CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Z: Maximum z-dimension of a grid;
3717 * - ::CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK: Maximum amount of
3718 * shared memory available to a thread block in bytes;
3719 * - ::CU_DEVICE_ATTRIBUTE_TOTAL_CONSTANT_MEMORY: Memory available on device for
3720 * __constant__ variables in a CUDA C kernel in bytes;
3721 * - ::CU_DEVICE_ATTRIBUTE_WARP_SIZE: Warp size in threads;
3722 * - ::CU_DEVICE_ATTRIBUTE_MAX_PITCH: Maximum pitch in bytes allowed by the
3723 * memory copy functions that involve memory regions allocated through
3724 * ::cuMemAllocPitch();
3725 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_WIDTH: Maximum 1D
3726 * texture width;
3727 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LINEAR_WIDTH: Maximum width
3728 * for a 1D texture bound to linear memory;
3729 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_MIPMAPPED_WIDTH: Maximum
3730 * mipmapped 1D texture width;
3731 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_WIDTH: Maximum 2D
3732 * texture width;
3733 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_HEIGHT: Maximum 2D
3734 * texture height;
3735 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_WIDTH: Maximum width
3736 * for a 2D texture bound to linear memory;
3737 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_HEIGHT: Maximum height
3738 * for a 2D texture bound to linear memory;
3739 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_PITCH: Maximum pitch
3740 * in bytes for a 2D texture bound to linear memory;
3741 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_MIPMAPPED_WIDTH: Maximum
3742 * mipmapped 2D texture width;
3743 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_MIPMAPPED_HEIGHT: Maximum
3744 * mipmapped 2D texture height;
3745 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_WIDTH: Maximum 3D
3746 * texture width;
3747 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_HEIGHT: Maximum 3D
3748 * texture height;
3749 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_DEPTH: Maximum 3D
3750 * texture depth;
3751 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_WIDTH_ALTERNATE:
3752 * Alternate maximum 3D texture width, 0 if no alternate
3753 * maximum 3D texture size is supported;
3754 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_HEIGHT_ALTERNATE:
3755 * Alternate maximum 3D texture height, 0 if no alternate
3756 * maximum 3D texture size is supported;
3757 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE3D_DEPTH_ALTERNATE:
3758 * Alternate maximum 3D texture depth, 0 if no alternate
3759 * maximum 3D texture size is supported;
3760 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURECUBEMAP_WIDTH:
3761 * Maximum cubemap texture width or height;
3762 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LAYERED_WIDTH:
3763 * Maximum 1D layered texture width;
3764 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LAYERED_LAYERS:
3765 * Maximum layers in a 1D layered texture;
3766 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_WIDTH:
3767 * Maximum 2D layered texture width;
3768 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_HEIGHT:
3769 * Maximum 2D layered texture height;
3770 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LAYERED_LAYERS:
3771 * Maximum layers in a 2D layered texture;
3772 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURECUBEMAP_LAYERED_WIDTH:
3773 * Maximum cubemap layered texture width or height;
3774 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURECUBEMAP_LAYERED_LAYERS:
3775 * Maximum layers in a cubemap layered texture;
3776 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE1D_WIDTH:
3777 * Maximum 1D surface width;
3778 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE2D_WIDTH:
3779 * Maximum 2D surface width;
3780 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE2D_HEIGHT:
3781 * Maximum 2D surface height;
3782 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE3D_WIDTH:
3783 * Maximum 3D surface width;
3784 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE3D_HEIGHT:
3785 * Maximum 3D surface height;
3786 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE3D_DEPTH:
3787 * Maximum 3D surface depth;
3788 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE1D_LAYERED_WIDTH:
3789 * Maximum 1D layered surface width;
3790 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE1D_LAYERED_LAYERS:
3791 * Maximum layers in a 1D layered surface;
3792 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE2D_LAYERED_WIDTH:
3793 * Maximum 2D layered surface width;
3794 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE2D_LAYERED_HEIGHT:
3795 * Maximum 2D layered surface height;
3796 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACE2D_LAYERED_LAYERS:
3797 * Maximum layers in a 2D layered surface;
3798 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACECUBEMAP_WIDTH:
3799 * Maximum cubemap surface width;
3800 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACECUBEMAP_LAYERED_WIDTH:
3801 * Maximum cubemap layered surface width;
3802 * - ::CU_DEVICE_ATTRIBUTE_MAXIMUM_SURFACECUBEMAP_LAYERED_LAYERS:
3803 * Maximum layers in a cubemap layered surface;
3804 * - ::CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_BLOCK: Maximum number of 32-bit
3805 * registers available to a thread block;
3806 * - ::CU_DEVICE_ATTRIBUTE_CLOCK_RATE: The typical clock frequency in kilohertz;
3807 * - ::CU_DEVICE_ATTRIBUTE_TEXTURE_ALIGNMENT: Alignment requirement; texture
3808 * base addresses aligned to ::textureAlign bytes do not need an offset
3809 * applied to texture fetches;
3810 * - ::CU_DEVICE_ATTRIBUTE_TEXTURE_PITCH_ALIGNMENT: Pitch alignment requirement
3811 * for 2D texture references bound to pitched memory;
3812 * - ::CU_DEVICE_ATTRIBUTE_GPU_OVERLAP: 1 if the device can concurrently copy
3813 * memory between host and device while executing a kernel, or 0 if not;
3814 * - ::CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT: Number of multiprocessors on
3815 * the device;
3816 * - ::CU_DEVICE_ATTRIBUTE_KERNEL_EXEC_TIMEOUT: 1 if there is a run time limit
3817 * for kernels executed on the device, or 0 if not;
3818 * - ::CU_DEVICE_ATTRIBUTE_INTEGRATED: 1 if the device is integrated with the
3819 * memory subsystem, or 0 if not;
3820 * - ::CU_DEVICE_ATTRIBUTE_CAN_MAP_HOST_MEMORY: 1 if the device can map host
3821 * memory into the CUDA address space, or 0 if not;
3822 * - ::CU_DEVICE_ATTRIBUTE_COMPUTE_MODE: Compute mode that device is currently
3823 * in. Available modes are as follows:
3824 * - ::CU_COMPUTEMODE_DEFAULT: Default mode - Device is not restricted and
3825 * can have multiple CUDA contexts present at a single time.
3826 * - ::CU_COMPUTEMODE_PROHIBITED: Compute-prohibited mode - Device is
3827 * prohibited from creating new CUDA contexts.
3828 * - ::CU_COMPUTEMODE_EXCLUSIVE_PROCESS: Compute-exclusive-process mode - Device
3829 * can have only one context used by a single process at a time.
3830 * - ::CU_DEVICE_ATTRIBUTE_CONCURRENT_KERNELS: 1 if the device supports
3831 * executing multiple kernels within the same context simultaneously, or 0 if
3832 * not. It is not guaranteed that multiple kernels will be resident
3833 * on the device concurrently so this feature should not be relied upon for
3834 * correctness;
3835 * - ::CU_DEVICE_ATTRIBUTE_ECC_ENABLED: 1 if error correction is enabled on the
3836 * device, 0 if error correction is disabled or not supported by the device;
3837 * - ::CU_DEVICE_ATTRIBUTE_PCI_BUS_ID: PCI bus identifier of the device;
3838 * - ::CU_DEVICE_ATTRIBUTE_PCI_DEVICE_ID: PCI device (also known as slot) identifier
3839 * of the device;
3840 * - ::CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID: PCI domain identifier of the device
3841 * - ::CU_DEVICE_ATTRIBUTE_TCC_DRIVER: 1 if the device is using a TCC driver. TCC
3842 * is only available on Tesla hardware running Windows Vista or later;
3843 * - ::CU_DEVICE_ATTRIBUTE_MEMORY_CLOCK_RATE: Peak memory clock frequency in kilohertz;
3844 * - ::CU_DEVICE_ATTRIBUTE_GLOBAL_MEMORY_BUS_WIDTH: Global memory bus width in bits;
3845 * - ::CU_DEVICE_ATTRIBUTE_L2_CACHE_SIZE: Size of L2 cache in bytes. 0 if the device doesn't have L2 cache;
3846 * - ::CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR: Maximum resident threads per multiprocessor;
3847 * - ::CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING: 1 if the device shares a unified address space with
3848 * the host, or 0 if not;
3849 * - ::CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR: Major compute capability version number;
3850 * - ::CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR: Minor compute capability version number;
3851 * - ::CU_DEVICE_ATTRIBUTE_GLOBAL_L1_CACHE_SUPPORTED: 1 if device supports caching globals
3852 * in L1 cache, 0 if caching globals in L1 cache is not supported by the device;
3853 * - ::CU_DEVICE_ATTRIBUTE_LOCAL_L1_CACHE_SUPPORTED: 1 if device supports caching locals
3854 * in L1 cache, 0 if caching locals in L1 cache is not supported by the device;
3855 * - ::CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_MULTIPROCESSOR: Maximum amount of
3856 * shared memory available to a multiprocessor in bytes; this amount is shared
3857 * by all thread blocks simultaneously resident on a multiprocessor;
3858 * - ::CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_MULTIPROCESSOR: Maximum number of 32-bit
3859 * registers available to a multiprocessor; this number is shared by all thread
3860 * blocks simultaneously resident on a multiprocessor;
3861 * - ::CU_DEVICE_ATTRIBUTE_MANAGED_MEMORY: 1 if device supports allocating managed memory
3862 * on this system, 0 if allocating managed memory is not supported by the device on this system.
3863 * - ::CU_DEVICE_ATTRIBUTE_MULTI_GPU_BOARD: 1 if device is on a multi-GPU board, 0 if not.
3864 * - ::CU_DEVICE_ATTRIBUTE_MULTI_GPU_BOARD_GROUP_ID: Unique identifier for a group of devices
3865 * associated with the same board. Devices on the same multi-GPU board will share the same identifier.
3866 * - ::CU_DEVICE_ATTRIBUTE_HOST_NATIVE_ATOMIC_SUPPORTED: 1 if Link between the device and the host
3867 * supports native atomic operations.
3868 * - ::CU_DEVICE_ATTRIBUTE_SINGLE_TO_DOUBLE_PRECISION_PERF_RATIO: Ratio of single precision performance
3869 * (in floating-point operations per second) to double precision performance.
3870 * - ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS: Device supports coherently accessing
3871 * pageable memory without calling cudaHostRegister on it.
3872 * - ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS: Device can coherently access managed memory
3873 * concurrently with the CPU.
3874 * - ::CU_DEVICE_ATTRIBUTE_COMPUTE_PREEMPTION_SUPPORTED: Device supports Compute Preemption.
3875 * - ::CU_DEVICE_ATTRIBUTE_CAN_USE_HOST_POINTER_FOR_REGISTERED_MEM: Device can access host registered
3876 * memory at the same virtual address as the CPU.
3877 * - ::CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN: The maximum per block shared memory size
3878 * supported on this device. This is the maximum value that can be opted into when using the cuFuncSetAttribute() call.
3879 * For more details see ::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES
3880 * - ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS_USES_HOST_PAGE_TABLES: Device accesses pageable memory via the host's
3881 * page tables.
3882 * - ::CU_DEVICE_ATTRIBUTE_DIRECT_MANAGED_MEM_ACCESS_FROM_HOST: The host can directly access managed memory on the device without migration.
3883 * - ::CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED: Device supports virtual memory management APIs like ::cuMemAddressReserve, ::cuMemCreate, ::cuMemMap and related APIs
3884 * - ::CU_DEVICE_ATTRIBUTE_HANDLE_TYPE_POSIX_FILE_DESCRIPTOR_SUPPORTED: Device supports exporting memory to a posix file descriptor with ::cuMemExportToShareableHandle, if requested via ::cuMemCreate
3885 * - ::CU_DEVICE_ATTRIBUTE_HANDLE_TYPE_WIN32_HANDLE_SUPPORTED: Device supports exporting memory to a Win32 NT handle with ::cuMemExportToShareableHandle, if requested via ::cuMemCreate
3886 * - ::CU_DEVICE_ATTRIBUTE_HANDLE_TYPE_WIN32_KMT_HANDLE_SUPPORTED: Device supports exporting memory to a Win32 KMT handle with ::cuMemExportToShareableHandle, if requested ::cuMemCreate
3887 * - ::CU_DEVICE_ATTRIBUTE_MAX_PERSISTING_L2_CACHE_SIZE: Maximum L2 persisting lines capacity setting in bytes.
3888 * - ::CU_DEVICE_ATTRIBUTE_MAX_ACCESS_POLICY_WINDOW_SIZE: Maximum value of CUaccessPolicyWindow::num_bytes.
3889 * - ::CU_DEVICE_ATTRIBUTE_MAX_BLOCKS_PER_MULTIPROCESSOR: Maximum number of thread blocks that can reside on a multiprocessor.
3890 * - ::CU_DEVICE_ATTRIBUTE_GENERIC_COMPRESSION_SUPPORTED: Device supports compressible memory allocation via ::cuMemCreate
3891 * - ::CU_DEVICE_ATTRIBUTE_RESERVED_SHARED_MEMORY_PER_BLOCK: Amount of shared memory per block reserved by CUDA driver in bytes.
3892 * - ::CU_DEVICE_ATTRIBUTE_READ_ONLY_HOST_REGISTER_SUPPORTED: Device supports using the ::cuMemHostRegister flag CU_MEMHOSTERGISTER_READ_ONLY to register memory that must be mapped as read-only to the GPU
3893 * - ::CU_DEVICE_ATTRIBUTE_MEMORY_POOLS_SUPPORTED: Device supports using the ::cuMemAllocAsync and ::cuMemPool family of APIs
3894 *
3895 * \param pi - Returned device attribute value
3896 * \param attrib - Device attribute to query
3897 * \param dev - Device handle
3898 *
3899 * \return
3900 * ::CUDA_SUCCESS,
3901 * ::CUDA_ERROR_DEINITIALIZED,
3902 * ::CUDA_ERROR_NOT_INITIALIZED,
3903 * ::CUDA_ERROR_INVALID_CONTEXT,
3904 * ::CUDA_ERROR_INVALID_VALUE,
3905 * ::CUDA_ERROR_INVALID_DEVICE
3906 * \notefnerr
3907 *
3908 * \sa
3909 * ::cuDeviceGetCount,
3910 * ::cuDeviceGetName,
3911 * ::cuDeviceGetUuid,
3912 * ::cuDeviceGet,
3913 * ::cuDeviceTotalMem,
3914 * ::cuDeviceGetExecAffinitySupport,
3915 * ::cudaDeviceGetAttribute,
3916 * ::cudaGetDeviceProperties
3917 */
3918CUresult CUDAAPI cuDeviceGetAttribute(int *pi, CUdevice_attribute attrib, CUdevice dev);
3919
3920/**
3921 * \brief Return NvSciSync attributes that this device can support.
3922 *
3923 * Returns in \p nvSciSyncAttrList, the properties of NvSciSync that
3924 * this CUDA device, \p dev can support. The returned \p nvSciSyncAttrList
3925 * can be used to create an NvSciSync object that matches this device's capabilities.
3926 *
3927 * If NvSciSyncAttrKey_RequiredPerm field in \p nvSciSyncAttrList is
3928 * already set this API will return ::CUDA_ERROR_INVALID_VALUE.
3929 *
3930 * The applications should set \p nvSciSyncAttrList to a valid
3931 * NvSciSyncAttrList failing which this API will return
3932 * ::CUDA_ERROR_INVALID_HANDLE.
3933 *
3934 * The \p flags controls how applications intends to use
3935 * the NvSciSync created from the \p nvSciSyncAttrList. The valid flags are:
3936 * - ::CUDA_NVSCISYNC_ATTR_SIGNAL, specifies that the applications intends to
3937 * signal an NvSciSync on this CUDA device.
3938 * - ::CUDA_NVSCISYNC_ATTR_WAIT, specifies that the applications intends to
3939 * wait on an NvSciSync on this CUDA device.
3940 *
3941 * At least one of these flags must be set, failing which the API
3942 * returns ::CUDA_ERROR_INVALID_VALUE. Both the flags are orthogonal
3943 * to one another: a developer may set both these flags that allows to
3944 * set both wait and signal specific attributes in the same \p nvSciSyncAttrList.
3945 *
3946 * \param nvSciSyncAttrList - Return NvSciSync attributes supported.
3947 * \param dev - Valid Cuda Device to get NvSciSync attributes for.
3948 * \param flags - flags describing NvSciSync usage.
3949 *
3950 * \return
3951 *
3952 * ::CUDA_SUCCESS,
3953 * ::CUDA_ERROR_DEINITIALIZED,
3954 * ::CUDA_ERROR_NOT_INITIALIZED,
3955 * ::CUDA_ERROR_INVALID_VALUE,
3956 * ::CUDA_ERROR_INVALID_HANDLE,
3957 * ::CUDA_ERROR_INVALID_DEVICE,
3958 * ::CUDA_ERROR_NOT_SUPPORTED,
3959 * ::CUDA_ERROR_OUT_OF_MEMORY
3960 *
3961 * \sa
3962 * ::cuImportExternalSemaphore,
3963 * ::cuDestroyExternalSemaphore,
3964 * ::cuSignalExternalSemaphoresAsync,
3965 * ::cuWaitExternalSemaphoresAsync
3966 */
3967CUresult CUDAAPI cuDeviceGetNvSciSyncAttributes(void *nvSciSyncAttrList, CUdevice dev, int flags);
3968
3969/**
3970 * \brief Sets the current memory pool of a device
3971 *
3972 * The memory pool must be local to the specified device.
3973 * ::cuMemAllocAsync allocates from the current mempool of the provided stream's device.
3974 * By default, a device's current memory pool is its default memory pool.
3975 *
3976 * \note Use ::cuMemAllocFromPoolAsync to specify asynchronous allocations from a device different
3977 * than the one the stream runs on.
3978 *
3979 * \returns
3980 * ::CUDA_SUCCESS,
3981 * ::CUDA_ERROR_INVALID_VALUE
3982 *
3983 * \sa ::cuDeviceGetDefaultMemPool, ::cuDeviceGetMemPool, ::cuMemPoolCreate, ::cuMemPoolDestroy, ::cuMemAllocFromPoolAsync
3984 */
3985CUresult CUDAAPI cuDeviceSetMemPool(CUdevice dev, CUmemoryPool pool);
3986
3987/**
3988 * \brief Gets the current mempool for a device
3989 *
3990 * Returns the last pool provided to ::cuDeviceSetMemPool for this device
3991 * or the device's default memory pool if ::cuDeviceSetMemPool has never been called.
3992 * By default the current mempool is the default mempool for a device.
3993 * Otherwise the returned pool must have been set with ::cuDeviceSetMemPool.
3994 *
3995 * \returns
3996 * ::CUDA_SUCCESS,
3997 * ::CUDA_ERROR_INVALID_VALUE
3998 *
3999 * \sa ::cuDeviceGetDefaultMemPool, ::cuMemPoolCreate, ::cuDeviceSetMemPool
4000 */
4001CUresult CUDAAPI cuDeviceGetMemPool(CUmemoryPool *pool, CUdevice dev);
4002
4003/**
4004 * \brief Returns the default mempool of a device
4005 *
4006 * The default mempool of a device contains device memory from that device.
4007 *
4008 * \return
4009 * ::CUDA_SUCCESS,
4010 * ::CUDA_ERROR_DEINITIALIZED
4011 * ::CUDA_ERROR_NOT_INITIALIZED,
4012 * ::CUDA_ERROR_INVALID_VALUE,
4013 * ::CUDA_ERROR_INVALID_DEVICE,
4014 * ::CUDA_ERROR_NOT_SUPPORTED
4015 * \notefnerr
4016 *
4017 * \sa ::cuMemAllocAsync, ::cuMemPoolTrimTo, ::cuMemPoolGetAttribute, ::cuMemPoolSetAttribute, cuMemPoolSetAccess, ::cuDeviceGetMemPool, ::cuMemPoolCreate
4018 */
4019CUresult CUDAAPI cuDeviceGetDefaultMemPool(CUmemoryPool *pool_out, CUdevice dev);
4020
4021/**
4022 * \brief Blocks until remote writes are visible to the specified scope
4023 *
4024 * Blocks until GPUDirect RDMA writes to the target context via mappings
4025 * created through APIs like nvidia_p2p_get_pages (see
4026 * https://docs.nvidia.com/cuda/gpudirect-rdma for more information), are
4027 * visible to the specified scope.
4028 *
4029 * If the scope equals or lies within the scope indicated by
4030 * ::CU_DEVICE_ATTRIBUTE_GPU_DIRECT_RDMA_WRITES_ORDERING, the call
4031 * will be a no-op and can be safely omitted for performance. This can be
4032 * determined by comparing the numerical values between the two enums, with
4033 * smaller scopes having smaller values.
4034 *
4035 * Users may query support for this API via
4036 * ::CU_DEVICE_ATTRIBUTE_FLUSH_FLUSH_GPU_DIRECT_RDMA_OPTIONS.
4037 *
4038 * \param target - The target of the operation, see ::CUflushGPUDirectRDMAWritesTarget
4039 * \param scope - The scope of the operation, see ::CUflushGPUDirectRDMAWritesScope
4040 *
4041 * \return
4042 * ::CUDA_SUCCESS,
4043 * ::CUDA_ERROR_DEINITIALIZED,
4044 * ::CUDA_ERROR_NOT_INITIALIZED,
4045 * ::CUDA_ERROR_INVALID_CONTEXT,
4046 * ::CUDA_ERROR_INVALID_VALUE,
4047 * \notefnerr
4048 *
4049 */
4050CUresult CUDAAPI cuFlushGPUDirectRDMAWrites(CUflushGPUDirectRDMAWritesTarget target, CUflushGPUDirectRDMAWritesScope scope);
4051
4052/** @} */ /* END CUDA_DEVICE */
4053
4054/**
4055 * \defgroup CUDA_DEVICE_DEPRECATED Device Management [DEPRECATED]
4056 *
4057 * ___MANBRIEF___ deprecated device management functions of the low-level CUDA
4058 * driver API (___CURRENT_FILE___) ___ENDMANBRIEF___
4059 *
4060 * This section describes the device management functions of the low-level
4061 * CUDA driver application programming interface.
4062 *
4063 * @{
4064 */
4065
4066/**
4067 * \brief Returns properties for a selected device
4068 *
4069 * \deprecated
4070 *
4071 * This function was deprecated as of CUDA 5.0 and replaced by ::cuDeviceGetAttribute().
4072 *
4073 * Returns in \p *prop the properties of device \p dev. The ::CUdevprop
4074 * structure is defined as:
4075 *
4076 * \code
4077 typedef struct CUdevprop_st {
4078 int maxThreadsPerBlock;
4079 int maxThreadsDim[3];
4080 int maxGridSize[3];
4081 int sharedMemPerBlock;
4082 int totalConstantMemory;
4083 int SIMDWidth;
4084 int memPitch;
4085 int regsPerBlock;
4086 int clockRate;
4087 int textureAlign
4088 } CUdevprop;
4089 * \endcode
4090 * where:
4091 *
4092 * - ::maxThreadsPerBlock is the maximum number of threads per block;
4093 * - ::maxThreadsDim[3] is the maximum sizes of each dimension of a block;
4094 * - ::maxGridSize[3] is the maximum sizes of each dimension of a grid;
4095 * - ::sharedMemPerBlock is the total amount of shared memory available per
4096 * block in bytes;
4097 * - ::totalConstantMemory is the total amount of constant memory available on
4098 * the device in bytes;
4099 * - ::SIMDWidth is the warp size;
4100 * - ::memPitch is the maximum pitch allowed by the memory copy functions that
4101 * involve memory regions allocated through ::cuMemAllocPitch();
4102 * - ::regsPerBlock is the total number of registers available per block;
4103 * - ::clockRate is the clock frequency in kilohertz;
4104 * - ::textureAlign is the alignment requirement; texture base addresses that
4105 * are aligned to ::textureAlign bytes do not need an offset applied to
4106 * texture fetches.
4107 *
4108 * \param prop - Returned properties of device
4109 * \param dev - Device to get properties for
4110 *
4111 * \return
4112 * ::CUDA_SUCCESS,
4113 * ::CUDA_ERROR_DEINITIALIZED,
4114 * ::CUDA_ERROR_NOT_INITIALIZED,
4115 * ::CUDA_ERROR_INVALID_CONTEXT,
4116 * ::CUDA_ERROR_INVALID_VALUE,
4117 * ::CUDA_ERROR_INVALID_DEVICE
4118 * \notefnerr
4119 *
4120 * \sa
4121 * ::cuDeviceGetAttribute,
4122 * ::cuDeviceGetCount,
4123 * ::cuDeviceGetName,
4124 * ::cuDeviceGetUuid,
4125 * ::cuDeviceGet,
4126 * ::cuDeviceTotalMem
4127 */
4128__CUDA_DEPRECATED CUresult CUDAAPI cuDeviceGetProperties(CUdevprop *prop, CUdevice dev);
4129
4130/**
4131 * \brief Returns the compute capability of the device
4132 *
4133 * \deprecated
4134 *
4135 * This function was deprecated as of CUDA 5.0 and its functionality superseded
4136 * by ::cuDeviceGetAttribute().
4137 *
4138 * Returns in \p *major and \p *minor the major and minor revision numbers that
4139 * define the compute capability of the device \p dev.
4140 *
4141 * \param major - Major revision number
4142 * \param minor - Minor revision number
4143 * \param dev - Device handle
4144 *
4145 * \return
4146 * ::CUDA_SUCCESS,
4147 * ::CUDA_ERROR_DEINITIALIZED,
4148 * ::CUDA_ERROR_NOT_INITIALIZED,
4149 * ::CUDA_ERROR_INVALID_CONTEXT,
4150 * ::CUDA_ERROR_INVALID_VALUE,
4151 * ::CUDA_ERROR_INVALID_DEVICE
4152 * \notefnerr
4153 *
4154 * \sa
4155 * ::cuDeviceGetAttribute,
4156 * ::cuDeviceGetCount,
4157 * ::cuDeviceGetName,
4158 * ::cuDeviceGetUuid,
4159 * ::cuDeviceGet,
4160 * ::cuDeviceTotalMem
4161 */
4162__CUDA_DEPRECATED CUresult CUDAAPI cuDeviceComputeCapability(int *major, int *minor, CUdevice dev);
4163
4164/** @} */ /* END CUDA_DEVICE_DEPRECATED */
4165
4166/**
4167 * \defgroup CUDA_PRIMARY_CTX Primary Context Management
4168 *
4169 * ___MANBRIEF___ primary context management functions of the low-level CUDA driver
4170 * API (___CURRENT_FILE___) ___ENDMANBRIEF___
4171 *
4172 * This section describes the primary context management functions of the low-level
4173 * CUDA driver application programming interface.
4174 *
4175 * The primary context is unique per device and shared with the CUDA runtime API.
4176 * These functions allow integration with other libraries using CUDA.
4177 *
4178 * @{
4179 */
4180
4181/**
4182 * \brief Retain the primary context on the GPU
4183 *
4184 * Retains the primary context on the device.
4185 * Once the user successfully retains the primary context, the primary context
4186 * will be active and available to the user until the user releases it
4187 * with ::cuDevicePrimaryCtxRelease() or resets it with ::cuDevicePrimaryCtxReset().
4188 * Unlike ::cuCtxCreate() the newly retained context is not pushed onto the stack.
4189 *
4190 * Retaining the primary context for the first time will fail with ::CUDA_ERROR_UNKNOWN
4191 * if the compute mode of the device is ::CU_COMPUTEMODE_PROHIBITED. The function
4192 * ::cuDeviceGetAttribute() can be used with ::CU_DEVICE_ATTRIBUTE_COMPUTE_MODE to
4193 * determine the compute mode of the device.
4194 * The <i>nvidia-smi</i> tool can be used to set the compute mode for
4195 * devices. Documentation for <i>nvidia-smi</i> can be obtained by passing a
4196 * -h option to it.
4197 *
4198 * Please note that the primary context always supports pinned allocations. Other
4199 * flags can be specified by ::cuDevicePrimaryCtxSetFlags().
4200 *
4201 * \param pctx - Returned context handle of the new context
4202 * \param dev - Device for which primary context is requested
4203 *
4204 * \return
4205 * ::CUDA_SUCCESS,
4206 * ::CUDA_ERROR_DEINITIALIZED,
4207 * ::CUDA_ERROR_NOT_INITIALIZED,
4208 * ::CUDA_ERROR_INVALID_CONTEXT,
4209 * ::CUDA_ERROR_INVALID_DEVICE,
4210 * ::CUDA_ERROR_INVALID_VALUE,
4211 * ::CUDA_ERROR_OUT_OF_MEMORY,
4212 * ::CUDA_ERROR_UNKNOWN
4213 * \notefnerr
4214 *
4215 * \sa ::cuDevicePrimaryCtxRelease,
4216 * ::cuDevicePrimaryCtxSetFlags,
4217 * ::cuCtxCreate,
4218 * ::cuCtxGetApiVersion,
4219 * ::cuCtxGetCacheConfig,
4220 * ::cuCtxGetDevice,
4221 * ::cuCtxGetFlags,
4222 * ::cuCtxGetLimit,
4223 * ::cuCtxPopCurrent,
4224 * ::cuCtxPushCurrent,
4225 * ::cuCtxSetCacheConfig,
4226 * ::cuCtxSetLimit,
4227 * ::cuCtxSynchronize
4228 */
4229CUresult CUDAAPI cuDevicePrimaryCtxRetain(CUcontext *pctx, CUdevice dev);
4230
4231/**
4232 * \brief Release the primary context on the GPU
4233 *
4234 * Releases the primary context interop on the device.
4235 * A retained context should always be released once the user is done using
4236 * it. The context is automatically reset once the last reference to it is
4237 * released. This behavior is different when the primary context was retained
4238 * by the CUDA runtime from CUDA 4.0 and earlier. In this case, the primary
4239 * context remains always active.
4240 *
4241 * Releasing a primary context that has not been previously retained will
4242 * fail with ::CUDA_ERROR_INVALID_CONTEXT.
4243 *
4244 * Please note that unlike ::cuCtxDestroy() this method does not pop the context
4245 * from stack in any circumstances.
4246 *
4247 * \param dev - Device which primary context is released
4248 *
4249 * \return
4250 * ::CUDA_SUCCESS,
4251 * ::CUDA_ERROR_DEINITIALIZED,
4252 * ::CUDA_ERROR_NOT_INITIALIZED,
4253 * ::CUDA_ERROR_INVALID_DEVICE,
4254 * ::CUDA_ERROR_INVALID_CONTEXT
4255 * \notefnerr
4256 *
4257 * \sa ::cuDevicePrimaryCtxRetain,
4258 * ::cuCtxDestroy,
4259 * ::cuCtxGetApiVersion,
4260 * ::cuCtxGetCacheConfig,
4261 * ::cuCtxGetDevice,
4262 * ::cuCtxGetFlags,
4263 * ::cuCtxGetLimit,
4264 * ::cuCtxPopCurrent,
4265 * ::cuCtxPushCurrent,
4266 * ::cuCtxSetCacheConfig,
4267 * ::cuCtxSetLimit,
4268 * ::cuCtxSynchronize
4269 */
4270CUresult CUDAAPI cuDevicePrimaryCtxRelease(CUdevice dev);
4271
4272/**
4273 * \brief Set flags for the primary context
4274 *
4275 * Sets the flags for the primary context on the device overwriting perviously
4276 * set ones.
4277 *
4278 * The three LSBs of the \p flags parameter can be used to control how the OS
4279 * thread, which owns the CUDA context at the time of an API call, interacts
4280 * with the OS scheduler when waiting for results from the GPU. Only one of
4281 * the scheduling flags can be set when creating a context.
4282 *
4283 * - ::CU_CTX_SCHED_SPIN: Instruct CUDA to actively spin when waiting for
4284 * results from the GPU. This can decrease latency when waiting for the GPU,
4285 * but may lower the performance of CPU threads if they are performing work in
4286 * parallel with the CUDA thread.
4287 *
4288 * - ::CU_CTX_SCHED_YIELD: Instruct CUDA to yield its thread when waiting for
4289 * results from the GPU. This can increase latency when waiting for the GPU,
4290 * but can increase the performance of CPU threads performing work in parallel
4291 * with the GPU.
4292 *
4293 * - ::CU_CTX_SCHED_BLOCKING_SYNC: Instruct CUDA to block the CPU thread on a
4294 * synchronization primitive when waiting for the GPU to finish work.
4295 *
4296 * - ::CU_CTX_BLOCKING_SYNC: Instruct CUDA to block the CPU thread on a
4297 * synchronization primitive when waiting for the GPU to finish work. <br>
4298 * <b>Deprecated:</b> This flag was deprecated as of CUDA 4.0 and was
4299 * replaced with ::CU_CTX_SCHED_BLOCKING_SYNC.
4300 *
4301 * - ::CU_CTX_SCHED_AUTO: The default value if the \p flags parameter is zero,
4302 * uses a heuristic based on the number of active CUDA contexts in the
4303 * process \e C and the number of logical processors in the system \e P. If
4304 * \e C > \e P, then CUDA will yield to other OS threads when waiting for
4305 * the GPU (::CU_CTX_SCHED_YIELD), otherwise CUDA will not yield while
4306 * waiting for results and actively spin on the processor (::CU_CTX_SCHED_SPIN).
4307 * Additionally, on Tegra devices, ::CU_CTX_SCHED_AUTO uses a heuristic based on
4308 * the power profile of the platform and may choose ::CU_CTX_SCHED_BLOCKING_SYNC
4309 * for low-powered devices.
4310 *
4311 * - ::CU_CTX_LMEM_RESIZE_TO_MAX: Instruct CUDA to not reduce local memory
4312 * after resizing local memory for a kernel. This can prevent thrashing by
4313 * local memory allocations when launching many kernels with high local
4314 * memory usage at the cost of potentially increased memory usage. <br>
4315 * <b>Deprecated:</b> This flag is deprecated and the behavior enabled
4316 * by this flag is now the default and cannot be disabled.
4317 *
4318 * \param dev - Device for which the primary context flags are set
4319 * \param flags - New flags for the device
4320 *
4321 * \return
4322 * ::CUDA_SUCCESS,
4323 * ::CUDA_ERROR_DEINITIALIZED,
4324 * ::CUDA_ERROR_NOT_INITIALIZED,
4325 * ::CUDA_ERROR_INVALID_DEVICE,
4326 * ::CUDA_ERROR_INVALID_VALUE,
4327 * \notefnerr
4328 *
4329 * \sa ::cuDevicePrimaryCtxRetain,
4330 * ::cuDevicePrimaryCtxGetState,
4331 * ::cuCtxCreate,
4332 * ::cuCtxGetFlags,
4333 * ::cudaSetDeviceFlags
4334 */
4335CUresult CUDAAPI cuDevicePrimaryCtxSetFlags(CUdevice dev, unsigned int flags);
4336
4337/**
4338 * \brief Get the state of the primary context
4339 *
4340 * Returns in \p *flags the flags for the primary context of \p dev, and in
4341 * \p *active whether it is active. See ::cuDevicePrimaryCtxSetFlags for flag
4342 * values.
4343 *
4344 * \param dev - Device to get primary context flags for
4345 * \param flags - Pointer to store flags
4346 * \param active - Pointer to store context state; 0 = inactive, 1 = active
4347 *
4348 * \return
4349 * ::CUDA_SUCCESS,
4350 * ::CUDA_ERROR_DEINITIALIZED,
4351 * ::CUDA_ERROR_NOT_INITIALIZED,
4352 * ::CUDA_ERROR_INVALID_DEVICE,
4353 * ::CUDA_ERROR_INVALID_VALUE,
4354 * \notefnerr
4355 *
4356 * \sa
4357 * ::cuDevicePrimaryCtxSetFlags,
4358 * ::cuCtxGetFlags,
4359 * ::cudaGetDeviceFlags
4360 */
4361CUresult CUDAAPI cuDevicePrimaryCtxGetState(CUdevice dev, unsigned int *flags, int *active);
4362
4363/**
4364 * \brief Destroy all allocations and reset all state on the primary context
4365 *
4366 * Explicitly destroys and cleans up all resources associated with the current
4367 * device in the current process.
4368 *
4369 * Note that it is responsibility of the calling function to ensure that no
4370 * other module in the process is using the device any more. For that reason
4371 * it is recommended to use ::cuDevicePrimaryCtxRelease() in most cases.
4372 * However it is safe for other modules to call ::cuDevicePrimaryCtxRelease()
4373 * even after resetting the device.
4374 * Resetting the primary context does not release it, an application that has
4375 * retained the primary context should explicitly release its usage.
4376 *
4377 * \param dev - Device for which primary context is destroyed
4378 *
4379 * \return
4380 * ::CUDA_SUCCESS,
4381 * ::CUDA_ERROR_DEINITIALIZED,
4382 * ::CUDA_ERROR_NOT_INITIALIZED,
4383 * ::CUDA_ERROR_INVALID_DEVICE,
4384 * ::CUDA_ERROR_PRIMARY_CONTEXT_ACTIVE
4385 * \notefnerr
4386 *
4387 * \sa ::cuDevicePrimaryCtxRetain,
4388 * ::cuDevicePrimaryCtxRelease,
4389 * ::cuCtxGetApiVersion,
4390 * ::cuCtxGetCacheConfig,
4391 * ::cuCtxGetDevice,
4392 * ::cuCtxGetFlags,
4393 * ::cuCtxGetLimit,
4394 * ::cuCtxPopCurrent,
4395 * ::cuCtxPushCurrent,
4396 * ::cuCtxSetCacheConfig,
4397 * ::cuCtxSetLimit,
4398 * ::cuCtxSynchronize,
4399 * ::cudaDeviceReset
4400 */
4401CUresult CUDAAPI cuDevicePrimaryCtxReset(CUdevice dev);
4402
4403/** @} */ /* END CUDA_PRIMARY_CTX */
4404
4405/**
4406 * \brief Returns information about the execution affinity support of the device.
4407 *
4408 * Returns in \p *pi whether execution affinity type \p type is supported by device \p dev.
4409 * The supported types are:
4410 * - ::CU_EXEC_AFFINITY_TYPE_SM_COUNT: 1 if context with limited SMs is supported by the device,
4411 * or 0 if not;
4412 *
4413 * \param pi - 1 if the execution affinity type \p type is supported by the device, or 0 if not
4414 * \param type - Execution affinity type to query
4415 * \param dev - Device handle
4416 *
4417 * \return
4418 * ::CUDA_SUCCESS,
4419 * ::CUDA_ERROR_DEINITIALIZED,
4420 * ::CUDA_ERROR_NOT_INITIALIZED,
4421 * ::CUDA_ERROR_INVALID_CONTEXT,
4422 * ::CUDA_ERROR_INVALID_VALUE,
4423 * ::CUDA_ERROR_INVALID_DEVICE
4424 * \notefnerr
4425 *
4426 * \sa
4427 * ::cuDeviceGetAttribute,
4428 * ::cuDeviceGetCount,
4429 * ::cuDeviceGetName,
4430 * ::cuDeviceGetUuid,
4431 * ::cuDeviceGet,
4432 * ::cuDeviceTotalMem
4433 */
4434CUresult CUDAAPI cuDeviceGetExecAffinitySupport(int *pi, CUexecAffinityType type, CUdevice dev);
4435
4436/**
4437 * \defgroup CUDA_CTX Context Management
4438 *
4439 * ___MANBRIEF___ context management functions of the low-level CUDA driver
4440 * API (___CURRENT_FILE___) ___ENDMANBRIEF___
4441 *
4442 * This section describes the context management functions of the low-level
4443 * CUDA driver application programming interface.
4444 *
4445 * Please note that some functions are described in
4446 * \ref CUDA_PRIMARY_CTX "Primary Context Management" section.
4447 *
4448 * @{
4449 */
4450
4451/**
4452 * \brief Create a CUDA context
4453 *
4454 * \note In most cases it is recommended to use ::cuDevicePrimaryCtxRetain.
4455 *
4456 * Creates a new CUDA context and associates it with the calling thread. The
4457 * \p flags parameter is described below. The context is created with a usage
4458 * count of 1 and the caller of ::cuCtxCreate() must call ::cuCtxDestroy() or
4459 * when done using the context. If a context is already current to the thread,
4460 * it is supplanted by the newly created context and may be restored by a subsequent
4461 * call to ::cuCtxPopCurrent().
4462 *
4463 * The three LSBs of the \p flags parameter can be used to control how the OS
4464 * thread, which owns the CUDA context at the time of an API call, interacts
4465 * with the OS scheduler when waiting for results from the GPU. Only one of
4466 * the scheduling flags can be set when creating a context.
4467 *
4468 * - ::CU_CTX_SCHED_SPIN: Instruct CUDA to actively spin when waiting for
4469 * results from the GPU. This can decrease latency when waiting for the GPU,
4470 * but may lower the performance of CPU threads if they are performing work in
4471 * parallel with the CUDA thread.
4472 *
4473 * - ::CU_CTX_SCHED_YIELD: Instruct CUDA to yield its thread when waiting for
4474 * results from the GPU. This can increase latency when waiting for the GPU,
4475 * but can increase the performance of CPU threads performing work in parallel
4476 * with the GPU.
4477 *
4478 * - ::CU_CTX_SCHED_BLOCKING_SYNC: Instruct CUDA to block the CPU thread on a
4479 * synchronization primitive when waiting for the GPU to finish work.
4480 *
4481 * - ::CU_CTX_BLOCKING_SYNC: Instruct CUDA to block the CPU thread on a
4482 * synchronization primitive when waiting for the GPU to finish work. <br>
4483 * <b>Deprecated:</b> This flag was deprecated as of CUDA 4.0 and was
4484 * replaced with ::CU_CTX_SCHED_BLOCKING_SYNC.
4485 *
4486 * - ::CU_CTX_SCHED_AUTO: The default value if the \p flags parameter is zero,
4487 * uses a heuristic based on the number of active CUDA contexts in the
4488 * process \e C and the number of logical processors in the system \e P. If
4489 * \e C > \e P, then CUDA will yield to other OS threads when waiting for
4490 * the GPU (::CU_CTX_SCHED_YIELD), otherwise CUDA will not yield while
4491 * waiting for results and actively spin on the processor (::CU_CTX_SCHED_SPIN).
4492 * Additionally, on Tegra devices, ::CU_CTX_SCHED_AUTO uses a heuristic based on
4493 * the power profile of the platform and may choose ::CU_CTX_SCHED_BLOCKING_SYNC
4494 * for low-powered devices.
4495 *
4496 * - ::CU_CTX_MAP_HOST: Instruct CUDA to support mapped pinned allocations.
4497 * This flag must be set in order to allocate pinned host memory that is
4498 * accessible to the GPU.
4499 *
4500 * - ::CU_CTX_LMEM_RESIZE_TO_MAX: Instruct CUDA to not reduce local memory
4501 * after resizing local memory for a kernel. This can prevent thrashing by
4502 * local memory allocations when launching many kernels with high local
4503 * memory usage at the cost of potentially increased memory usage. <br>
4504 * <b>Deprecated:</b> This flag is deprecated and the behavior enabled
4505 * by this flag is now the default and cannot be disabled.
4506 * Instead, the per-thread stack size can be controlled with ::cuCtxSetLimit().
4507 *
4508 * Context creation will fail with ::CUDA_ERROR_UNKNOWN if the compute mode of
4509 * the device is ::CU_COMPUTEMODE_PROHIBITED. The function ::cuDeviceGetAttribute()
4510 * can be used with ::CU_DEVICE_ATTRIBUTE_COMPUTE_MODE to determine the
4511 * compute mode of the device. The <i>nvidia-smi</i> tool can be used to set
4512 * the compute mode for * devices.
4513 * Documentation for <i>nvidia-smi</i> can be obtained by passing a
4514 * -h option to it.
4515 *
4516 * \param pctx - Returned context handle of the new context
4517 * \param flags - Context creation flags
4518 * \param dev - Device to create context on
4519 *
4520 * \return
4521 * ::CUDA_SUCCESS,
4522 * ::CUDA_ERROR_DEINITIALIZED,
4523 * ::CUDA_ERROR_NOT_INITIALIZED,
4524 * ::CUDA_ERROR_INVALID_CONTEXT,
4525 * ::CUDA_ERROR_INVALID_DEVICE,
4526 * ::CUDA_ERROR_INVALID_VALUE,
4527 * ::CUDA_ERROR_OUT_OF_MEMORY,
4528 * ::CUDA_ERROR_UNKNOWN
4529 * \notefnerr
4530 *
4531 * \sa ::cuCtxDestroy,
4532 * ::cuCtxGetApiVersion,
4533 * ::cuCtxGetCacheConfig,
4534 * ::cuCtxGetDevice,
4535 * ::cuCtxGetFlags,
4536 * ::cuCtxGetLimit,
4537 * ::cuCtxPopCurrent,
4538 * ::cuCtxPushCurrent,
4539 * ::cuCtxSetCacheConfig,
4540 * ::cuCtxSetLimit,
4541 * ::cuCtxSynchronize
4542 */
4543CUresult CUDAAPI cuCtxCreate(CUcontext *pctx, unsigned int flags, CUdevice dev);
4544
4545/**
4546 * \brief Create a CUDA context with execution affinity
4547 *
4548 * Creates a new CUDA context with execution affinity and associates it with
4549 * the calling thread. The \p paramsArray and \p flags parameter are described below.
4550 * The context is created with a usage count of 1 and the caller of ::cuCtxCreate() must
4551 * call ::cuCtxDestroy() or when done using the context. If a context is already
4552 * current to the thread, it is supplanted by the newly created context and may
4553 * be restored by a subsequent call to ::cuCtxPopCurrent().
4554 *
4555 * The type and the amount of execution resource the context can use is limited by \p paramsArray
4556 * and \p numParams. The \p paramsArray is an array of \p CUexecAffinityParam and the \p numParams
4557 * describes the size of the array. If two \p CUexecAffinityParam in the array have the same type,
4558 * the latter execution affinity parameter overrides the former execution affinity parameter.
4559 * The supported execution affinity types are:
4560 * - ::CU_EXEC_AFFINITY_TYPE_SM_COUNT limits the portion of SMs that the context can use. The portion
4561 * of SMs is specified as the number of SMs via \p CUexecAffinitySmCount. This limit will be internally
4562 * rounded up to the next hardware-supported amount. Hence, it is imperative to query the actual execution
4563 * affinity of the context via \p cuCtxGetExecAffinity after context creation. Currently, this attribute
4564 * is only supported under Volta+ MPS.
4565 *
4566 * The three LSBs of the \p flags parameter can be used to control how the OS
4567 * thread, which owns the CUDA context at the time of an API call, interacts
4568 * with the OS scheduler when waiting for results from the GPU. Only one of
4569 * the scheduling flags can be set when creating a context.
4570 *
4571 * - ::CU_CTX_SCHED_SPIN: Instruct CUDA to actively spin when waiting for
4572 * results from the GPU. This can decrease latency when waiting for the GPU,
4573 * but may lower the performance of CPU threads if they are performing work in
4574 * parallel with the CUDA thread.
4575 *
4576 * - ::CU_CTX_SCHED_YIELD: Instruct CUDA to yield its thread when waiting for
4577 * results from the GPU. This can increase latency when waiting for the GPU,
4578 * but can increase the performance of CPU threads performing work in parallel
4579 * with the GPU.
4580 *
4581 * - ::CU_CTX_SCHED_BLOCKING_SYNC: Instruct CUDA to block the CPU thread on a
4582 * synchronization primitive when waiting for the GPU to finish work.
4583 *
4584 * - ::CU_CTX_BLOCKING_SYNC: Instruct CUDA to block the CPU thread on a
4585 * synchronization primitive when waiting for the GPU to finish work. <br>
4586 * <b>Deprecated:</b> This flag was deprecated as of CUDA 4.0 and was
4587 * replaced with ::CU_CTX_SCHED_BLOCKING_SYNC.
4588 *
4589 * - ::CU_CTX_SCHED_AUTO: The default value if the \p flags parameter is zero,
4590 * uses a heuristic based on the number of active CUDA contexts in the
4591 * process \e C and the number of logical processors in the system \e P. If
4592 * \e C > \e P, then CUDA will yield to other OS threads when waiting for
4593 * the GPU (::CU_CTX_SCHED_YIELD), otherwise CUDA will not yield while
4594 * waiting for results and actively spin on the processor (::CU_CTX_SCHED_SPIN).
4595 * Additionally, on Tegra devices, ::CU_CTX_SCHED_AUTO uses a heuristic based on
4596 * the power profile of the platform and may choose ::CU_CTX_SCHED_BLOCKING_SYNC
4597 * for low-powered devices.
4598 *
4599 * - ::CU_CTX_MAP_HOST: Instruct CUDA to support mapped pinned allocations.
4600 * This flag must be set in order to allocate pinned host memory that is
4601 * accessible to the GPU.
4602 *
4603 * - ::CU_CTX_LMEM_RESIZE_TO_MAX: Instruct CUDA to not reduce local memory
4604 * after resizing local memory for a kernel. This can prevent thrashing by
4605 * local memory allocations when launching many kernels with high local
4606 * memory usage at the cost of potentially increased memory usage. <br>
4607 * <b>Deprecated:</b> This flag is deprecated and the behavior enabled
4608 * by this flag is now the default and cannot be disabled.
4609 * Instead, the per-thread stack size can be controlled with ::cuCtxSetLimit().
4610 *
4611 * Context creation will fail with ::CUDA_ERROR_UNKNOWN if the compute mode of
4612 * the device is ::CU_COMPUTEMODE_PROHIBITED. The function ::cuDeviceGetAttribute()
4613 * can be used with ::CU_DEVICE_ATTRIBUTE_COMPUTE_MODE to determine the
4614 * compute mode of the device. The <i>nvidia-smi</i> tool can be used to set
4615 * the compute mode for * devices.
4616 * Documentation for <i>nvidia-smi</i> can be obtained by passing a
4617 * -h option to it.
4618 *
4619 * \param pctx - Returned context handle of the new context
4620 * \param paramsArray - Execution affinity parameters
4621 * \param numParams - Number of execution affinity parameters
4622 * \param flags - Context creation flags
4623 * \param dev - Device to create context on
4624 *
4625 * \return
4626 * ::CUDA_SUCCESS,
4627 * ::CUDA_ERROR_DEINITIALIZED,
4628 * ::CUDA_ERROR_NOT_INITIALIZED,
4629 * ::CUDA_ERROR_INVALID_CONTEXT,
4630 * ::CUDA_ERROR_INVALID_DEVICE,
4631 * ::CUDA_ERROR_INVALID_VALUE,
4632 * ::CUDA_ERROR_OUT_OF_MEMORY,
4633 * ::CUDA_ERROR_UNSUPPORTED_EXEC_AFFINITY,
4634 * ::CUDA_ERROR_UNKNOWN
4635 * \notefnerr
4636 *
4637 * \sa ::cuCtxDestroy,
4638 * ::cuCtxGetApiVersion,
4639 * ::cuCtxGetCacheConfig,
4640 * ::cuCtxGetDevice,
4641 * ::cuCtxGetFlags,
4642 * ::cuCtxGetLimit,
4643 * ::cuCtxPopCurrent,
4644 * ::cuCtxPushCurrent,
4645 * ::cuCtxSetCacheConfig,
4646 * ::cuCtxSetLimit,
4647 * ::cuCtxSynchronize,
4648 * ::CUexecAffinityParam
4649 */
4650CUresult CUDAAPI cuCtxCreate_v3(CUcontext *pctx, CUexecAffinityParam *paramsArray, int numParams, unsigned int flags, CUdevice dev);
4651
4652/**
4653 * \brief Destroy a CUDA context
4654 *
4655 * Destroys the CUDA context specified by \p ctx. The context \p ctx will be
4656 * destroyed regardless of how many threads it is current to.
4657 * It is the responsibility of the calling function to ensure that no API
4658 * call issues using \p ctx while ::cuCtxDestroy() is executing.
4659 *
4660 * If \p ctx is current to the calling thread then \p ctx will also be
4661 * popped from the current thread's context stack (as though ::cuCtxPopCurrent()
4662 * were called). If \p ctx is current to other threads, then \p ctx will
4663 * remain current to those threads, and attempting to access \p ctx from
4664 * those threads will result in the error ::CUDA_ERROR_CONTEXT_IS_DESTROYED.
4665 *
4666 * \param ctx - Context to destroy
4667 *
4668 * \return
4669 * ::CUDA_SUCCESS,
4670 * ::CUDA_ERROR_DEINITIALIZED,
4671 * ::CUDA_ERROR_NOT_INITIALIZED,
4672 * ::CUDA_ERROR_INVALID_CONTEXT,
4673 * ::CUDA_ERROR_INVALID_VALUE
4674 * \notefnerr
4675 *
4676 * \sa ::cuCtxCreate,
4677 * ::cuCtxGetApiVersion,
4678 * ::cuCtxGetCacheConfig,
4679 * ::cuCtxGetDevice,
4680 * ::cuCtxGetFlags,
4681 * ::cuCtxGetLimit,
4682 * ::cuCtxPopCurrent,
4683 * ::cuCtxPushCurrent,
4684 * ::cuCtxSetCacheConfig,
4685 * ::cuCtxSetLimit,
4686 * ::cuCtxSynchronize
4687 */
4688CUresult CUDAAPI cuCtxDestroy(CUcontext ctx);
4689
4690/**
4691 * \brief Pushes a context on the current CPU thread
4692 *
4693 * Pushes the given context \p ctx onto the CPU thread's stack of current
4694 * contexts. The specified context becomes the CPU thread's current context, so
4695 * all CUDA functions that operate on the current context are affected.
4696 *
4697 * The previous current context may be made current again by calling
4698 * ::cuCtxDestroy() or ::cuCtxPopCurrent().
4699 *
4700 * \param ctx - Context to push
4701 *
4702 * \return
4703 * ::CUDA_SUCCESS,
4704 * ::CUDA_ERROR_DEINITIALIZED,
4705 * ::CUDA_ERROR_NOT_INITIALIZED,
4706 * ::CUDA_ERROR_INVALID_CONTEXT,
4707 * ::CUDA_ERROR_INVALID_VALUE
4708 * \notefnerr
4709 *
4710 * \sa ::cuCtxCreate,
4711 * ::cuCtxDestroy,
4712 * ::cuCtxGetApiVersion,
4713 * ::cuCtxGetCacheConfig,
4714 * ::cuCtxGetDevice,
4715 * ::cuCtxGetFlags,
4716 * ::cuCtxGetLimit,
4717 * ::cuCtxPopCurrent,
4718 * ::cuCtxSetCacheConfig,
4719 * ::cuCtxSetLimit,
4720 * ::cuCtxSynchronize
4721 */
4722CUresult CUDAAPI cuCtxPushCurrent(CUcontext ctx);
4723
4724/**
4725 * \brief Pops the current CUDA context from the current CPU thread.
4726 *
4727 * Pops the current CUDA context from the CPU thread and passes back the
4728 * old context handle in \p *pctx. That context may then be made current
4729 * to a different CPU thread by calling ::cuCtxPushCurrent().
4730 *
4731 * If a context was current to the CPU thread before ::cuCtxCreate() or
4732 * ::cuCtxPushCurrent() was called, this function makes that context current to
4733 * the CPU thread again.
4734 *
4735 * \param pctx - Returned new context handle
4736 *
4737 * \return
4738 * ::CUDA_SUCCESS,
4739 * ::CUDA_ERROR_DEINITIALIZED,
4740 * ::CUDA_ERROR_NOT_INITIALIZED,
4741 * ::CUDA_ERROR_INVALID_CONTEXT
4742 * \notefnerr
4743 *
4744 * \sa ::cuCtxCreate,
4745 * ::cuCtxDestroy,
4746 * ::cuCtxGetApiVersion,
4747 * ::cuCtxGetCacheConfig,
4748 * ::cuCtxGetDevice,
4749 * ::cuCtxGetFlags,
4750 * ::cuCtxGetLimit,
4751 * ::cuCtxPushCurrent,
4752 * ::cuCtxSetCacheConfig,
4753 * ::cuCtxSetLimit,
4754 * ::cuCtxSynchronize
4755 */
4756CUresult CUDAAPI cuCtxPopCurrent(CUcontext *pctx);
4757
4758/**
4759 * \brief Binds the specified CUDA context to the calling CPU thread
4760 *
4761 * Binds the specified CUDA context to the calling CPU thread.
4762 * If \p ctx is NULL then the CUDA context previously bound to the
4763 * calling CPU thread is unbound and ::CUDA_SUCCESS is returned.
4764 *
4765 * If there exists a CUDA context stack on the calling CPU thread, this
4766 * will replace the top of that stack with \p ctx.
4767 * If \p ctx is NULL then this will be equivalent to popping the top
4768 * of the calling CPU thread's CUDA context stack (or a no-op if the
4769 * calling CPU thread's CUDA context stack is empty).
4770 *
4771 * \param ctx - Context to bind to the calling CPU thread
4772 *
4773 * \return
4774 * ::CUDA_SUCCESS,
4775 * ::CUDA_ERROR_DEINITIALIZED,
4776 * ::CUDA_ERROR_NOT_INITIALIZED,
4777 * ::CUDA_ERROR_INVALID_CONTEXT
4778 * \notefnerr
4779 *
4780 * \sa
4781 * ::cuCtxGetCurrent,
4782 * ::cuCtxCreate,
4783 * ::cuCtxDestroy,
4784 * ::cudaSetDevice
4785 */
4786CUresult CUDAAPI cuCtxSetCurrent(CUcontext ctx);
4787
4788/**
4789 * \brief Returns the CUDA context bound to the calling CPU thread.
4790 *
4791 * Returns in \p *pctx the CUDA context bound to the calling CPU thread.
4792 * If no context is bound to the calling CPU thread then \p *pctx is
4793 * set to NULL and ::CUDA_SUCCESS is returned.
4794 *
4795 * \param pctx - Returned context handle
4796 *
4797 * \return
4798 * ::CUDA_SUCCESS,
4799 * ::CUDA_ERROR_DEINITIALIZED,
4800 * ::CUDA_ERROR_NOT_INITIALIZED,
4801 * \notefnerr
4802 *
4803 * \sa
4804 * ::cuCtxSetCurrent,
4805 * ::cuCtxCreate,
4806 * ::cuCtxDestroy,
4807 * ::cudaGetDevice
4808 */
4809CUresult CUDAAPI cuCtxGetCurrent(CUcontext *pctx);
4810
4811/**
4812 * \brief Returns the device ID for the current context
4813 *
4814 * Returns in \p *device the ordinal of the current context's device.
4815 *
4816 * \param device - Returned device ID for the current context
4817 *
4818 * \return
4819 * ::CUDA_SUCCESS,
4820 * ::CUDA_ERROR_DEINITIALIZED,
4821 * ::CUDA_ERROR_NOT_INITIALIZED,
4822 * ::CUDA_ERROR_INVALID_CONTEXT,
4823 * ::CUDA_ERROR_INVALID_VALUE,
4824 * \notefnerr
4825 *
4826 * \sa ::cuCtxCreate,
4827 * ::cuCtxDestroy,
4828 * ::cuCtxGetApiVersion,
4829 * ::cuCtxGetCacheConfig,
4830 * ::cuCtxGetFlags,
4831 * ::cuCtxGetLimit,
4832 * ::cuCtxPopCurrent,
4833 * ::cuCtxPushCurrent,
4834 * ::cuCtxSetCacheConfig,
4835 * ::cuCtxSetLimit,
4836 * ::cuCtxSynchronize,
4837 * ::cudaGetDevice
4838 */
4839CUresult CUDAAPI cuCtxGetDevice(CUdevice *device);
4840
4841/**
4842 * \brief Returns the flags for the current context
4843 *
4844 * Returns in \p *flags the flags of the current context. See ::cuCtxCreate
4845 * for flag values.
4846 *
4847 * \param flags - Pointer to store flags of current context
4848 *
4849 * \return
4850 * ::CUDA_SUCCESS,
4851 * ::CUDA_ERROR_DEINITIALIZED,
4852 * ::CUDA_ERROR_NOT_INITIALIZED,
4853 * ::CUDA_ERROR_INVALID_CONTEXT,
4854 * ::CUDA_ERROR_INVALID_VALUE,
4855 * \notefnerr
4856 *
4857 * \sa ::cuCtxCreate,
4858 * ::cuCtxGetApiVersion,
4859 * ::cuCtxGetCacheConfig,
4860 * ::cuCtxGetCurrent,
4861 * ::cuCtxGetDevice,
4862 * ::cuCtxGetLimit,
4863 * ::cuCtxGetSharedMemConfig,
4864 * ::cuCtxGetStreamPriorityRange,
4865 * ::cudaGetDeviceFlags
4866 */
4867CUresult CUDAAPI cuCtxGetFlags(unsigned int *flags);
4868
4869/**
4870 * \brief Block for a context's tasks to complete
4871 *
4872 * Blocks until the device has completed all preceding requested tasks.
4873 * ::cuCtxSynchronize() returns an error if one of the preceding tasks failed.
4874 * If the context was created with the ::CU_CTX_SCHED_BLOCKING_SYNC flag, the
4875 * CPU thread will block until the GPU context has finished its work.
4876 *
4877 * \return
4878 * ::CUDA_SUCCESS,
4879 * ::CUDA_ERROR_DEINITIALIZED,
4880 * ::CUDA_ERROR_NOT_INITIALIZED,
4881 * ::CUDA_ERROR_INVALID_CONTEXT
4882 * \notefnerr
4883 *
4884 * \sa ::cuCtxCreate,
4885 * ::cuCtxDestroy,
4886 * ::cuCtxGetApiVersion,
4887 * ::cuCtxGetCacheConfig,
4888 * ::cuCtxGetDevice,
4889 * ::cuCtxGetFlags,
4890 * ::cuCtxGetLimit,
4891 * ::cuCtxPopCurrent,
4892 * ::cuCtxPushCurrent,
4893 * ::cuCtxSetCacheConfig,
4894 * ::cuCtxSetLimit,
4895 * ::cudaDeviceSynchronize
4896 */
4897CUresult CUDAAPI cuCtxSynchronize(void);
4898
4899/**
4900 * \brief Set resource limits
4901 *
4902 * Setting \p limit to \p value is a request by the application to update
4903 * the current limit maintained by the context. The driver is free to
4904 * modify the requested value to meet h/w requirements (this could be
4905 * clamping to minimum or maximum values, rounding up to nearest element
4906 * size, etc). The application can use ::cuCtxGetLimit() to find out exactly
4907 * what the limit has been set to.
4908 *
4909 * Setting each ::CUlimit has its own specific restrictions, so each is
4910 * discussed here.
4911 *
4912 * - ::CU_LIMIT_STACK_SIZE controls the stack size in bytes of each GPU thread.
4913 * The driver automatically increases the per-thread stack size
4914 * for each kernel launch as needed. This size isn't reset back to the
4915 * original value after each launch. Setting this value will take effect
4916 * immediately, and if necessary, the device will block until all preceding
4917 * requested tasks are complete.
4918 *
4919 * - ::CU_LIMIT_PRINTF_FIFO_SIZE controls the size in bytes of the FIFO used
4920 * by the ::printf() device system call. Setting ::CU_LIMIT_PRINTF_FIFO_SIZE
4921 * must be performed before launching any kernel that uses the ::printf()
4922 * device system call, otherwise ::CUDA_ERROR_INVALID_VALUE will be returned.
4923 *
4924 * - ::CU_LIMIT_MALLOC_HEAP_SIZE controls the size in bytes of the heap used
4925 * by the ::malloc() and ::free() device system calls. Setting
4926 * ::CU_LIMIT_MALLOC_HEAP_SIZE must be performed before launching any kernel
4927 * that uses the ::malloc() or ::free() device system calls, otherwise
4928 * ::CUDA_ERROR_INVALID_VALUE will be returned.
4929 *
4930 * - ::CU_LIMIT_DEV_RUNTIME_SYNC_DEPTH controls the maximum nesting depth of
4931 * a grid at which a thread can safely call ::cudaDeviceSynchronize(). Setting
4932 * this limit must be performed before any launch of a kernel that uses the
4933 * device runtime and calls ::cudaDeviceSynchronize() above the default sync
4934 * depth, two levels of grids. Calls to ::cudaDeviceSynchronize() will fail
4935 * with error code ::cudaErrorSyncDepthExceeded if the limitation is
4936 * violated. This limit can be set smaller than the default or up the maximum
4937 * launch depth of 24. When setting this limit, keep in mind that additional
4938 * levels of sync depth require the driver to reserve large amounts of device
4939 * memory which can no longer be used for user allocations. If these
4940 * reservations of device memory fail, ::cuCtxSetLimit() will return
4941 * ::CUDA_ERROR_OUT_OF_MEMORY, and the limit can be reset to a lower value.
4942 * This limit is only applicable to devices of compute capability 3.5 and
4943 * higher. Attempting to set this limit on devices of compute capability less
4944 * than 3.5 will result in the error ::CUDA_ERROR_UNSUPPORTED_LIMIT being
4945 * returned.
4946 *
4947 * - ::CU_LIMIT_DEV_RUNTIME_PENDING_LAUNCH_COUNT controls the maximum number of
4948 * outstanding device runtime launches that can be made from the current
4949 * context. A grid is outstanding from the point of launch up until the grid
4950 * is known to have been completed. Device runtime launches which violate
4951 * this limitation fail and return ::cudaErrorLaunchPendingCountExceeded when
4952 * ::cudaGetLastError() is called after launch. If more pending launches than
4953 * the default (2048 launches) are needed for a module using the device
4954 * runtime, this limit can be increased. Keep in mind that being able to
4955 * sustain additional pending launches will require the driver to reserve
4956 * larger amounts of device memory upfront which can no longer be used for
4957 * allocations. If these reservations fail, ::cuCtxSetLimit() will return
4958 * ::CUDA_ERROR_OUT_OF_MEMORY, and the limit can be reset to a lower value.
4959 * This limit is only applicable to devices of compute capability 3.5 and
4960 * higher. Attempting to set this limit on devices of compute capability less
4961 * than 3.5 will result in the error ::CUDA_ERROR_UNSUPPORTED_LIMIT being
4962 * returned.
4963 *
4964 * - ::CU_LIMIT_MAX_L2_FETCH_GRANULARITY controls the L2 cache fetch granularity.
4965 * Values can range from 0B to 128B. This is purely a performance hint and
4966 * it can be ignored or clamped depending on the platform.
4967 *
4968 * - ::CU_LIMIT_PERSISTING_L2_CACHE_SIZE controls size in bytes available for
4969 * persisting L2 cache. This is purely a performance hint and it can be
4970 * ignored or clamped depending on the platform.
4971 *
4972 * \param limit - Limit to set
4973 * \param value - Size of limit
4974 *
4975 * \return
4976 * ::CUDA_SUCCESS,
4977 * ::CUDA_ERROR_INVALID_VALUE,
4978 * ::CUDA_ERROR_UNSUPPORTED_LIMIT,
4979 * ::CUDA_ERROR_OUT_OF_MEMORY,
4980 * ::CUDA_ERROR_INVALID_CONTEXT
4981 * \notefnerr
4982 *
4983 * \sa ::cuCtxCreate,
4984 * ::cuCtxDestroy,
4985 * ::cuCtxGetApiVersion,
4986 * ::cuCtxGetCacheConfig,
4987 * ::cuCtxGetDevice,
4988 * ::cuCtxGetFlags,
4989 * ::cuCtxGetLimit,
4990 * ::cuCtxPopCurrent,
4991 * ::cuCtxPushCurrent,
4992 * ::cuCtxSetCacheConfig,
4993 * ::cuCtxSynchronize,
4994 * ::cudaDeviceSetLimit
4995 */
4996CUresult CUDAAPI cuCtxSetLimit(CUlimit limit, size_t value);
4997
4998/**
4999 * \brief Returns resource limits
5000 *
5001 * Returns in \p *pvalue the current size of \p limit. The supported
5002 * ::CUlimit values are:
5003 * - ::CU_LIMIT_STACK_SIZE: stack size in bytes of each GPU thread.
5004 * - ::CU_LIMIT_PRINTF_FIFO_SIZE: size in bytes of the FIFO used by the
5005 * ::printf() device system call.
5006 * - ::CU_LIMIT_MALLOC_HEAP_SIZE: size in bytes of the heap used by the
5007 * ::malloc() and ::free() device system calls.
5008 * - ::CU_LIMIT_DEV_RUNTIME_SYNC_DEPTH: maximum grid depth at which a thread
5009 * can issue the device runtime call ::cudaDeviceSynchronize() to wait on
5010 * child grid launches to complete.
5011 * - ::CU_LIMIT_DEV_RUNTIME_PENDING_LAUNCH_COUNT: maximum number of outstanding
5012 * device runtime launches that can be made from this context.
5013 * - ::CU_LIMIT_MAX_L2_FETCH_GRANULARITY: L2 cache fetch granularity.
5014 * - ::CU_LIMIT_PERSISTING_L2_CACHE_SIZE: Persisting L2 cache size in bytes
5015 *
5016 * \param limit - Limit to query
5017 * \param pvalue - Returned size of limit
5018 *
5019 * \return
5020 * ::CUDA_SUCCESS,
5021 * ::CUDA_ERROR_INVALID_VALUE,
5022 * ::CUDA_ERROR_UNSUPPORTED_LIMIT
5023 * \notefnerr
5024 *
5025 * \sa ::cuCtxCreate,
5026 * ::cuCtxDestroy,
5027 * ::cuCtxGetApiVersion,
5028 * ::cuCtxGetCacheConfig,
5029 * ::cuCtxGetDevice,
5030 * ::cuCtxGetFlags,
5031 * ::cuCtxPopCurrent,
5032 * ::cuCtxPushCurrent,
5033 * ::cuCtxSetCacheConfig,
5034 * ::cuCtxSetLimit,
5035 * ::cuCtxSynchronize,
5036 * ::cudaDeviceGetLimit
5037 */
5038CUresult CUDAAPI cuCtxGetLimit(size_t *pvalue, CUlimit limit);
5039
5040/**
5041 * \brief Returns the preferred cache configuration for the current context.
5042 *
5043 * On devices where the L1 cache and shared memory use the same hardware
5044 * resources, this function returns through \p pconfig the preferred cache configuration
5045 * for the current context. This is only a preference. The driver will use
5046 * the requested configuration if possible, but it is free to choose a different
5047 * configuration if required to execute functions.
5048 *
5049 * This will return a \p pconfig of ::CU_FUNC_CACHE_PREFER_NONE on devices
5050 * where the size of the L1 cache and shared memory are fixed.
5051 *
5052 * The supported cache configurations are:
5053 * - ::CU_FUNC_CACHE_PREFER_NONE: no preference for shared memory or L1 (default)
5054 * - ::CU_FUNC_CACHE_PREFER_SHARED: prefer larger shared memory and smaller L1 cache
5055 * - ::CU_FUNC_CACHE_PREFER_L1: prefer larger L1 cache and smaller shared memory
5056 * - ::CU_FUNC_CACHE_PREFER_EQUAL: prefer equal sized L1 cache and shared memory
5057 *
5058 * \param pconfig - Returned cache configuration
5059 *
5060 * \return
5061 * ::CUDA_SUCCESS,
5062 * ::CUDA_ERROR_DEINITIALIZED,
5063 * ::CUDA_ERROR_NOT_INITIALIZED,
5064 * ::CUDA_ERROR_INVALID_CONTEXT,
5065 * ::CUDA_ERROR_INVALID_VALUE
5066 * \notefnerr
5067 *
5068 * \sa ::cuCtxCreate,
5069 * ::cuCtxDestroy,
5070 * ::cuCtxGetApiVersion,
5071 * ::cuCtxGetDevice,
5072 * ::cuCtxGetFlags,
5073 * ::cuCtxGetLimit,
5074 * ::cuCtxPopCurrent,
5075 * ::cuCtxPushCurrent,
5076 * ::cuCtxSetCacheConfig,
5077 * ::cuCtxSetLimit,
5078 * ::cuCtxSynchronize,
5079 * ::cuFuncSetCacheConfig,
5080 * ::cudaDeviceGetCacheConfig
5081 */
5082CUresult CUDAAPI cuCtxGetCacheConfig(CUfunc_cache *pconfig);
5083
5084/**
5085 * \brief Sets the preferred cache configuration for the current context.
5086 *
5087 * On devices where the L1 cache and shared memory use the same hardware
5088 * resources, this sets through \p config the preferred cache configuration for
5089 * the current context. This is only a preference. The driver will use
5090 * the requested configuration if possible, but it is free to choose a different
5091 * configuration if required to execute the function. Any function preference
5092 * set via ::cuFuncSetCacheConfig() will be preferred over this context-wide
5093 * setting. Setting the context-wide cache configuration to
5094 * ::CU_FUNC_CACHE_PREFER_NONE will cause subsequent kernel launches to prefer
5095 * to not change the cache configuration unless required to launch the kernel.
5096 *
5097 * This setting does nothing on devices where the size of the L1 cache and
5098 * shared memory are fixed.
5099 *
5100 * Launching a kernel with a different preference than the most recent
5101 * preference setting may insert a device-side synchronization point.
5102 *
5103 * The supported cache configurations are:
5104 * - ::CU_FUNC_CACHE_PREFER_NONE: no preference for shared memory or L1 (default)
5105 * - ::CU_FUNC_CACHE_PREFER_SHARED: prefer larger shared memory and smaller L1 cache
5106 * - ::CU_FUNC_CACHE_PREFER_L1: prefer larger L1 cache and smaller shared memory
5107 * - ::CU_FUNC_CACHE_PREFER_EQUAL: prefer equal sized L1 cache and shared memory
5108 *
5109 * \param config - Requested cache configuration
5110 *
5111 * \return
5112 * ::CUDA_SUCCESS,
5113 * ::CUDA_ERROR_DEINITIALIZED,
5114 * ::CUDA_ERROR_NOT_INITIALIZED,
5115 * ::CUDA_ERROR_INVALID_CONTEXT,
5116 * ::CUDA_ERROR_INVALID_VALUE
5117 * \notefnerr
5118 *
5119 * \sa ::cuCtxCreate,
5120 * ::cuCtxDestroy,
5121 * ::cuCtxGetApiVersion,
5122 * ::cuCtxGetCacheConfig,
5123 * ::cuCtxGetDevice,
5124 * ::cuCtxGetFlags,
5125 * ::cuCtxGetLimit,
5126 * ::cuCtxPopCurrent,
5127 * ::cuCtxPushCurrent,
5128 * ::cuCtxSetLimit,
5129 * ::cuCtxSynchronize,
5130 * ::cuFuncSetCacheConfig,
5131 * ::cudaDeviceSetCacheConfig
5132 */
5133CUresult CUDAAPI cuCtxSetCacheConfig(CUfunc_cache config);
5134
5135/**
5136 * \brief Returns the current shared memory configuration for the current context.
5137 *
5138 * This function will return in \p pConfig the current size of shared memory banks
5139 * in the current context. On devices with configurable shared memory banks,
5140 * ::cuCtxSetSharedMemConfig can be used to change this setting, so that all
5141 * subsequent kernel launches will by default use the new bank size. When
5142 * ::cuCtxGetSharedMemConfig is called on devices without configurable shared
5143 * memory, it will return the fixed bank size of the hardware.
5144 *
5145 * The returned bank configurations can be either:
5146 * - ::CU_SHARED_MEM_CONFIG_FOUR_BYTE_BANK_SIZE: shared memory bank width is
5147 * four bytes.
5148 * - ::CU_SHARED_MEM_CONFIG_EIGHT_BYTE_BANK_SIZE: shared memory bank width will
5149 * eight bytes.
5150 *
5151 * \param pConfig - returned shared memory configuration
5152 * \return
5153 * ::CUDA_SUCCESS,
5154 * ::CUDA_ERROR_DEINITIALIZED,
5155 * ::CUDA_ERROR_NOT_INITIALIZED,
5156 * ::CUDA_ERROR_INVALID_CONTEXT,
5157 * ::CUDA_ERROR_INVALID_VALUE
5158 * \notefnerr
5159 *
5160 * \sa ::cuCtxCreate,
5161 * ::cuCtxDestroy,
5162 * ::cuCtxGetApiVersion,
5163 * ::cuCtxGetCacheConfig,
5164 * ::cuCtxGetDevice,
5165 * ::cuCtxGetFlags,
5166 * ::cuCtxGetLimit,
5167 * ::cuCtxPopCurrent,
5168 * ::cuCtxPushCurrent,
5169 * ::cuCtxSetLimit,
5170 * ::cuCtxSynchronize,
5171 * ::cuCtxGetSharedMemConfig,
5172 * ::cuFuncSetCacheConfig,
5173 * ::cudaDeviceGetSharedMemConfig
5174 */
5175CUresult CUDAAPI cuCtxGetSharedMemConfig(CUsharedconfig *pConfig);
5176
5177/**
5178 * \brief Sets the shared memory configuration for the current context.
5179 *
5180 * On devices with configurable shared memory banks, this function will set
5181 * the context's shared memory bank size which is used for subsequent kernel
5182 * launches.
5183 *
5184 * Changed the shared memory configuration between launches may insert a device
5185 * side synchronization point between those launches.
5186 *
5187 * Changing the shared memory bank size will not increase shared memory usage
5188 * or affect occupancy of kernels, but may have major effects on performance.
5189 * Larger bank sizes will allow for greater potential bandwidth to shared memory,
5190 * but will change what kinds of accesses to shared memory will result in bank
5191 * conflicts.
5192 *
5193 * This function will do nothing on devices with fixed shared memory bank size.
5194 *
5195 * The supported bank configurations are:
5196 * - ::CU_SHARED_MEM_CONFIG_DEFAULT_BANK_SIZE: set bank width to the default initial
5197 * setting (currently, four bytes).
5198 * - ::CU_SHARED_MEM_CONFIG_FOUR_BYTE_BANK_SIZE: set shared memory bank width to
5199 * be natively four bytes.
5200 * - ::CU_SHARED_MEM_CONFIG_EIGHT_BYTE_BANK_SIZE: set shared memory bank width to
5201 * be natively eight bytes.
5202 *
5203 * \param config - requested shared memory configuration
5204 *
5205 * \return
5206 * ::CUDA_SUCCESS,
5207 * ::CUDA_ERROR_DEINITIALIZED,
5208 * ::CUDA_ERROR_NOT_INITIALIZED,
5209 * ::CUDA_ERROR_INVALID_CONTEXT,
5210 * ::CUDA_ERROR_INVALID_VALUE
5211 * \notefnerr
5212 *
5213 * \sa ::cuCtxCreate,
5214 * ::cuCtxDestroy,
5215 * ::cuCtxGetApiVersion,
5216 * ::cuCtxGetCacheConfig,
5217 * ::cuCtxGetDevice,
5218 * ::cuCtxGetFlags,
5219 * ::cuCtxGetLimit,
5220 * ::cuCtxPopCurrent,
5221 * ::cuCtxPushCurrent,
5222 * ::cuCtxSetLimit,
5223 * ::cuCtxSynchronize,
5224 * ::cuCtxGetSharedMemConfig,
5225 * ::cuFuncSetCacheConfig,
5226 * ::cudaDeviceSetSharedMemConfig
5227 */
5228CUresult CUDAAPI cuCtxSetSharedMemConfig(CUsharedconfig config);
5229
5230/**
5231 * \brief Gets the context's API version.
5232 *
5233 * Returns a version number in \p version corresponding to the capabilities of
5234 * the context (e.g. 3010 or 3020), which library developers can use to direct
5235 * callers to a specific API version. If \p ctx is NULL, returns the API version
5236 * used to create the currently bound context.
5237 *
5238 * Note that new API versions are only introduced when context capabilities are
5239 * changed that break binary compatibility, so the API version and driver version
5240 * may be different. For example, it is valid for the API version to be 3020 while
5241 * the driver version is 4020.
5242 *
5243 * \param ctx - Context to check
5244 * \param version - Pointer to version
5245 *
5246 * \return
5247 * ::CUDA_SUCCESS,
5248 * ::CUDA_ERROR_DEINITIALIZED,
5249 * ::CUDA_ERROR_NOT_INITIALIZED,
5250 * ::CUDA_ERROR_INVALID_CONTEXT,
5251 * ::CUDA_ERROR_INVALID_VALUE,
5252 * ::CUDA_ERROR_UNKNOWN
5253 * \notefnerr
5254 *
5255 * \sa ::cuCtxCreate,
5256 * ::cuCtxDestroy,
5257 * ::cuCtxGetDevice,
5258 * ::cuCtxGetFlags,
5259 * ::cuCtxGetLimit,
5260 * ::cuCtxPopCurrent,
5261 * ::cuCtxPushCurrent,
5262 * ::cuCtxSetCacheConfig,
5263 * ::cuCtxSetLimit,
5264 * ::cuCtxSynchronize
5265 */
5266CUresult CUDAAPI cuCtxGetApiVersion(CUcontext ctx, unsigned int *version);
5267
5268/**
5269 * \brief Returns numerical values that correspond to the least and
5270 * greatest stream priorities.
5271 *
5272 * Returns in \p *leastPriority and \p *greatestPriority the numerical values that correspond
5273 * to the least and greatest stream priorities respectively. Stream priorities
5274 * follow a convention where lower numbers imply greater priorities. The range of
5275 * meaningful stream priorities is given by [\p *greatestPriority, \p *leastPriority].
5276 * If the user attempts to create a stream with a priority value that is
5277 * outside the meaningful range as specified by this API, the priority is
5278 * automatically clamped down or up to either \p *leastPriority or \p *greatestPriority
5279 * respectively. See ::cuStreamCreateWithPriority for details on creating a
5280 * priority stream.
5281 * A NULL may be passed in for \p *leastPriority or \p *greatestPriority if the value
5282 * is not desired.
5283 *
5284 * This function will return '0' in both \p *leastPriority and \p *greatestPriority if
5285 * the current context's device does not support stream priorities
5286 * (see ::cuDeviceGetAttribute).
5287 *
5288 * \param leastPriority - Pointer to an int in which the numerical value for least
5289 * stream priority is returned
5290 * \param greatestPriority - Pointer to an int in which the numerical value for greatest
5291 * stream priority is returned
5292 *
5293 * \return
5294 * ::CUDA_SUCCESS,
5295 * ::CUDA_ERROR_INVALID_VALUE,
5296 * \notefnerr
5297 *
5298 * \sa ::cuStreamCreateWithPriority,
5299 * ::cuStreamGetPriority,
5300 * ::cuCtxGetDevice,
5301 * ::cuCtxGetFlags,
5302 * ::cuCtxSetLimit,
5303 * ::cuCtxSynchronize,
5304 * ::cudaDeviceGetStreamPriorityRange
5305 */
5306CUresult CUDAAPI cuCtxGetStreamPriorityRange(int *leastPriority, int *greatestPriority);
5307
5308/**
5309 * \brief Resets all persisting lines in cache to normal status.
5310 *
5311 * ::cuCtxResetPersistingL2Cache Resets all persisting lines in cache to normal
5312 * status. Takes effect on function return.
5313 *
5314 * \return
5315 * ::CUDA_SUCCESS,
5316 * ::CUDA_ERROR_NOT_SUPPORTED
5317 * \notefnerr
5318 *
5319 * \sa
5320 * ::CUaccessPolicyWindow
5321 */
5322CUresult CUDAAPI cuCtxResetPersistingL2Cache(void);
5323
5324/**
5325 * \brief Returns the execution affinity setting for the current context.
5326 *
5327 * Returns in \p *pExecAffinity the current value of \p type. The supported
5328 * ::CUexecAffinityType values are:
5329 * - ::CU_EXEC_AFFINITY_TYPE_SM_COUNT: number of SMs the context is limited to use.
5330 *
5331 * \param type - Execution affinity type to query
5332 * \param pExecAffinity - Returned execution affinity
5333 *
5334 * \return
5335 * ::CUDA_SUCCESS,
5336 * ::CUDA_ERROR_DEINITIALIZED,
5337 * ::CUDA_ERROR_NOT_INITIALIZED,
5338 * ::CUDA_ERROR_INVALID_CONTEXT,
5339 * ::CUDA_ERROR_INVALID_VALUE,
5340 * ::CUDA_ERROR_UNSUPPORTED_EXEC_AFFINITY
5341 * \notefnerr
5342 *
5343 * \sa
5344 * ::CUexecAffinityParam
5345 */
5346CUresult CUDAAPI cuCtxGetExecAffinity(CUexecAffinityParam *pExecAffinity, CUexecAffinityType type);
5347
5348
5349/** @} */ /* END CUDA_CTX */
5350
5351/**
5352 * \defgroup CUDA_CTX_DEPRECATED Context Management [DEPRECATED]
5353 *
5354 * ___MANBRIEF___ deprecated context management functions of the low-level CUDA
5355 * driver API (___CURRENT_FILE___) ___ENDMANBRIEF___
5356 *
5357 * This section describes the deprecated context management functions of the low-level
5358 * CUDA driver application programming interface.
5359 *
5360 * @{
5361 */
5362
5363/**
5364 * \brief Increment a context's usage-count
5365 *
5366 * \deprecated
5367 *
5368 * Note that this function is deprecated and should not be used.
5369 *
5370 * Increments the usage count of the context and passes back a context handle
5371 * in \p *pctx that must be passed to ::cuCtxDetach() when the application is
5372 * done with the context. ::cuCtxAttach() fails if there is no context current
5373 * to the thread.
5374 *
5375 * Currently, the \p flags parameter must be 0.
5376 *
5377 * \param pctx - Returned context handle of the current context
5378 * \param flags - Context attach flags (must be 0)
5379 *
5380 * \return
5381 * ::CUDA_SUCCESS,
5382 * ::CUDA_ERROR_DEINITIALIZED,
5383 * ::CUDA_ERROR_NOT_INITIALIZED,
5384 * ::CUDA_ERROR_INVALID_CONTEXT,
5385 * ::CUDA_ERROR_INVALID_VALUE
5386 * \notefnerr
5387 *
5388 * \sa ::cuCtxCreate,
5389 * ::cuCtxDestroy,
5390 * ::cuCtxDetach,
5391 * ::cuCtxGetApiVersion,
5392 * ::cuCtxGetCacheConfig,
5393 * ::cuCtxGetDevice,
5394 * ::cuCtxGetFlags,
5395 * ::cuCtxGetLimit,
5396 * ::cuCtxPopCurrent,
5397 * ::cuCtxPushCurrent,
5398 * ::cuCtxSetCacheConfig,
5399 * ::cuCtxSetLimit,
5400 * ::cuCtxSynchronize
5401 */
5402__CUDA_DEPRECATED CUresult CUDAAPI cuCtxAttach(CUcontext *pctx, unsigned int flags);
5403
5404/**
5405 * \brief Decrement a context's usage-count
5406 *
5407 * \deprecated
5408 *
5409 * Note that this function is deprecated and should not be used.
5410 *
5411 * Decrements the usage count of the context \p ctx, and destroys the context
5412 * if the usage count goes to 0. The context must be a handle that was passed
5413 * back by ::cuCtxCreate() or ::cuCtxAttach(), and must be current to the
5414 * calling thread.
5415 *
5416 * \param ctx - Context to destroy
5417 *
5418 * \return
5419 * ::CUDA_SUCCESS,
5420 * ::CUDA_ERROR_DEINITIALIZED,
5421 * ::CUDA_ERROR_NOT_INITIALIZED,
5422 * ::CUDA_ERROR_INVALID_CONTEXT
5423 * \notefnerr
5424 *
5425 * \sa ::cuCtxCreate,
5426 * ::cuCtxDestroy,
5427 * ::cuCtxGetApiVersion,
5428 * ::cuCtxGetCacheConfig,
5429 * ::cuCtxGetDevice,
5430 * ::cuCtxGetFlags,
5431 * ::cuCtxGetLimit,
5432 * ::cuCtxPopCurrent,
5433 * ::cuCtxPushCurrent,
5434 * ::cuCtxSetCacheConfig,
5435 * ::cuCtxSetLimit,
5436 * ::cuCtxSynchronize
5437 */
5438__CUDA_DEPRECATED CUresult CUDAAPI cuCtxDetach(CUcontext ctx);
5439
5440/** @} */ /* END CUDA_CTX_DEPRECATED */
5441
5442
5443/**
5444 * \defgroup CUDA_MODULE Module Management
5445 *
5446 * ___MANBRIEF___ module management functions of the low-level CUDA driver API
5447 * (___CURRENT_FILE___) ___ENDMANBRIEF___
5448 *
5449 * This section describes the module management functions of the low-level CUDA
5450 * driver application programming interface.
5451 *
5452 * @{
5453 */
5454
5455/**
5456 * \brief Loads a compute module
5457 *
5458 * Takes a filename \p fname and loads the corresponding module \p module into
5459 * the current context. The CUDA driver API does not attempt to lazily
5460 * allocate the resources needed by a module; if the memory for functions and
5461 * data (constant and global) needed by the module cannot be allocated,
5462 * ::cuModuleLoad() fails. The file should be a \e cubin file as output by
5463 * \b nvcc, or a \e PTX file either as output by \b nvcc or handwritten, or
5464 * a \e fatbin file as output by \b nvcc from toolchain 4.0 or later.
5465 *
5466 * \param module - Returned module
5467 * \param fname - Filename of module to load
5468 *
5469 * \return
5470 * ::CUDA_SUCCESS,
5471 * ::CUDA_ERROR_DEINITIALIZED,
5472 * ::CUDA_ERROR_NOT_INITIALIZED,
5473 * ::CUDA_ERROR_INVALID_CONTEXT,
5474 * ::CUDA_ERROR_INVALID_VALUE,
5475 * ::CUDA_ERROR_INVALID_PTX,
5476 * ::CUDA_ERROR_UNSUPPORTED_PTX_VERSION,
5477 * ::CUDA_ERROR_NOT_FOUND,
5478 * ::CUDA_ERROR_OUT_OF_MEMORY,
5479 * ::CUDA_ERROR_FILE_NOT_FOUND,
5480 * ::CUDA_ERROR_NO_BINARY_FOR_GPU,
5481 * ::CUDA_ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND,
5482 * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED,
5483 * ::CUDA_ERROR_JIT_COMPILER_NOT_FOUND
5484 * \notefnerr
5485 *
5486 * \sa ::cuModuleGetFunction,
5487 * ::cuModuleGetGlobal,
5488 * ::cuModuleGetTexRef,
5489 * ::cuModuleLoadData,
5490 * ::cuModuleLoadDataEx,
5491 * ::cuModuleLoadFatBinary,
5492 * ::cuModuleUnload
5493 */
5494CUresult CUDAAPI cuModuleLoad(CUmodule *module, const char *fname);
5495
5496/**
5497 * \brief Load a module's data
5498 *
5499 * Takes a pointer \p image and loads the corresponding module \p module into
5500 * the current context. The pointer may be obtained by mapping a \e cubin or
5501 * \e PTX or \e fatbin file, passing a \e cubin or \e PTX or \e fatbin file
5502 * as a NULL-terminated text string, or incorporating a \e cubin or \e fatbin
5503 * object into the executable resources and using operating system calls such
5504 * as Windows \c FindResource() to obtain the pointer.
5505 *
5506 * \param module - Returned module
5507 * \param image - Module data to load
5508 *
5509 * \return
5510 * ::CUDA_SUCCESS,
5511 * ::CUDA_ERROR_DEINITIALIZED,
5512 * ::CUDA_ERROR_NOT_INITIALIZED,
5513 * ::CUDA_ERROR_INVALID_CONTEXT,
5514 * ::CUDA_ERROR_INVALID_VALUE,
5515 * ::CUDA_ERROR_INVALID_PTX,
5516 * ::CUDA_ERROR_UNSUPPORTED_PTX_VERSION,
5517 * ::CUDA_ERROR_OUT_OF_MEMORY,
5518 * ::CUDA_ERROR_NO_BINARY_FOR_GPU,
5519 * ::CUDA_ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND,
5520 * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED,
5521 * ::CUDA_ERROR_JIT_COMPILER_NOT_FOUND
5522 * \notefnerr
5523 *
5524 * \sa ::cuModuleGetFunction,
5525 * ::cuModuleGetGlobal,
5526 * ::cuModuleGetTexRef,
5527 * ::cuModuleLoad,
5528 * ::cuModuleLoadDataEx,
5529 * ::cuModuleLoadFatBinary,
5530 * ::cuModuleUnload
5531 */
5532CUresult CUDAAPI cuModuleLoadData(CUmodule *module, const void *image);
5533
5534/**
5535 * \brief Load a module's data with options
5536 *
5537 * Takes a pointer \p image and loads the corresponding module \p module into
5538 * the current context. The pointer may be obtained by mapping a \e cubin or
5539 * \e PTX or \e fatbin file, passing a \e cubin or \e PTX or \e fatbin file
5540 * as a NULL-terminated text string, or incorporating a \e cubin or \e fatbin
5541 * object into the executable resources and using operating system calls such
5542 * as Windows \c FindResource() to obtain the pointer. Options are passed as
5543 * an array via \p options and any corresponding parameters are passed in
5544 * \p optionValues. The number of total options is supplied via \p numOptions.
5545 * Any outputs will be returned via \p optionValues.
5546 *
5547 * \param module - Returned module
5548 * \param image - Module data to load
5549 * \param numOptions - Number of options
5550 * \param options - Options for JIT
5551 * \param optionValues - Option values for JIT
5552 *
5553 * \return
5554 * ::CUDA_SUCCESS,
5555 * ::CUDA_ERROR_DEINITIALIZED,
5556 * ::CUDA_ERROR_NOT_INITIALIZED,
5557 * ::CUDA_ERROR_INVALID_CONTEXT,
5558 * ::CUDA_ERROR_INVALID_VALUE,
5559 * ::CUDA_ERROR_INVALID_PTX,
5560 * ::CUDA_ERROR_UNSUPPORTED_PTX_VERSION,
5561 * ::CUDA_ERROR_OUT_OF_MEMORY,
5562 * ::CUDA_ERROR_NO_BINARY_FOR_GPU,
5563 * ::CUDA_ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND,
5564 * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED,
5565 * ::CUDA_ERROR_JIT_COMPILER_NOT_FOUND
5566 * \notefnerr
5567 *
5568 * \sa ::cuModuleGetFunction,
5569 * ::cuModuleGetGlobal,
5570 * ::cuModuleGetTexRef,
5571 * ::cuModuleLoad,
5572 * ::cuModuleLoadData,
5573 * ::cuModuleLoadFatBinary,
5574 * ::cuModuleUnload
5575 */
5576CUresult CUDAAPI cuModuleLoadDataEx(CUmodule *module, const void *image, unsigned int numOptions, CUjit_option *options, void **optionValues);
5577
5578/**
5579 * \brief Load a module's data
5580 *
5581 * Takes a pointer \p fatCubin and loads the corresponding module \p module
5582 * into the current context. The pointer represents a <i>fat binary</i> object,
5583 * which is a collection of different \e cubin and/or \e PTX files, all
5584 * representing the same device code, but compiled and optimized for different
5585 * architectures.
5586 *
5587 * Prior to CUDA 4.0, there was no documented API for constructing and using
5588 * fat binary objects by programmers. Starting with CUDA 4.0, fat binary
5589 * objects can be constructed by providing the <i>-fatbin option</i> to \b nvcc.
5590 * More information can be found in the \b nvcc document.
5591 *
5592 * \param module - Returned module
5593 * \param fatCubin - Fat binary to load
5594 *
5595 * \return
5596 * ::CUDA_SUCCESS,
5597 * ::CUDA_ERROR_DEINITIALIZED,
5598 * ::CUDA_ERROR_NOT_INITIALIZED,
5599 * ::CUDA_ERROR_INVALID_CONTEXT,
5600 * ::CUDA_ERROR_INVALID_VALUE,
5601 * ::CUDA_ERROR_INVALID_PTX,
5602 * ::CUDA_ERROR_UNSUPPORTED_PTX_VERSION,
5603 * ::CUDA_ERROR_NOT_FOUND,
5604 * ::CUDA_ERROR_OUT_OF_MEMORY,
5605 * ::CUDA_ERROR_NO_BINARY_FOR_GPU,
5606 * ::CUDA_ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND,
5607 * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED,
5608 * ::CUDA_ERROR_JIT_COMPILER_NOT_FOUND
5609 * \notefnerr
5610 *
5611 * \sa ::cuModuleGetFunction,
5612 * ::cuModuleGetGlobal,
5613 * ::cuModuleGetTexRef,
5614 * ::cuModuleLoad,
5615 * ::cuModuleLoadData,
5616 * ::cuModuleLoadDataEx,
5617 * ::cuModuleUnload
5618 */
5619CUresult CUDAAPI cuModuleLoadFatBinary(CUmodule *module, const void *fatCubin);
5620
5621/**
5622 * \brief Unloads a module
5623 *
5624 * Unloads a module \p hmod from the current context.
5625 *
5626 * \param hmod - Module to unload
5627 *
5628 * \return
5629 * ::CUDA_SUCCESS,
5630 * ::CUDA_ERROR_DEINITIALIZED,
5631 * ::CUDA_ERROR_NOT_INITIALIZED,
5632 * ::CUDA_ERROR_INVALID_CONTEXT,
5633 * ::CUDA_ERROR_INVALID_VALUE
5634 * \notefnerr
5635 *
5636 * \sa ::cuModuleGetFunction,
5637 * ::cuModuleGetGlobal,
5638 * ::cuModuleGetTexRef,
5639 * ::cuModuleLoad,
5640 * ::cuModuleLoadData,
5641 * ::cuModuleLoadDataEx,
5642 * ::cuModuleLoadFatBinary
5643 */
5644CUresult CUDAAPI cuModuleUnload(CUmodule hmod);
5645
5646/**
5647 * \brief Returns a function handle
5648 *
5649 * Returns in \p *hfunc the handle of the function of name \p name located in
5650 * module \p hmod. If no function of that name exists, ::cuModuleGetFunction()
5651 * returns ::CUDA_ERROR_NOT_FOUND.
5652 *
5653 * \param hfunc - Returned function handle
5654 * \param hmod - Module to retrieve function from
5655 * \param name - Name of function to retrieve
5656 *
5657 * \return
5658 * ::CUDA_SUCCESS,
5659 * ::CUDA_ERROR_DEINITIALIZED,
5660 * ::CUDA_ERROR_NOT_INITIALIZED,
5661 * ::CUDA_ERROR_INVALID_CONTEXT,
5662 * ::CUDA_ERROR_INVALID_VALUE,
5663 * ::CUDA_ERROR_NOT_FOUND
5664 * \notefnerr
5665 *
5666 * \sa ::cuModuleGetGlobal,
5667 * ::cuModuleGetTexRef,
5668 * ::cuModuleLoad,
5669 * ::cuModuleLoadData,
5670 * ::cuModuleLoadDataEx,
5671 * ::cuModuleLoadFatBinary,
5672 * ::cuModuleUnload
5673 */
5674CUresult CUDAAPI cuModuleGetFunction(CUfunction *hfunc, CUmodule hmod, const char *name);
5675
5676/**
5677 * \brief Returns a global pointer from a module
5678 *
5679 * Returns in \p *dptr and \p *bytes the base pointer and size of the
5680 * global of name \p name located in module \p hmod. If no variable of that name
5681 * exists, ::cuModuleGetGlobal() returns ::CUDA_ERROR_NOT_FOUND. Both
5682 * parameters \p dptr and \p bytes are optional. If one of them is
5683 * NULL, it is ignored.
5684 *
5685 * \param dptr - Returned global device pointer
5686 * \param bytes - Returned global size in bytes
5687 * \param hmod - Module to retrieve global from
5688 * \param name - Name of global to retrieve
5689 *
5690 * \return
5691 * ::CUDA_SUCCESS,
5692 * ::CUDA_ERROR_DEINITIALIZED,
5693 * ::CUDA_ERROR_NOT_INITIALIZED,
5694 * ::CUDA_ERROR_INVALID_CONTEXT,
5695 * ::CUDA_ERROR_INVALID_VALUE,
5696 * ::CUDA_ERROR_NOT_FOUND
5697 * \notefnerr
5698 *
5699 * \sa ::cuModuleGetFunction,
5700 * ::cuModuleGetTexRef,
5701 * ::cuModuleLoad,
5702 * ::cuModuleLoadData,
5703 * ::cuModuleLoadDataEx,
5704 * ::cuModuleLoadFatBinary,
5705 * ::cuModuleUnload,
5706 * ::cudaGetSymbolAddress,
5707 * ::cudaGetSymbolSize
5708 */
5709CUresult CUDAAPI cuModuleGetGlobal(CUdeviceptr *dptr, size_t *bytes, CUmodule hmod, const char *name);
5710
5711/**
5712 * \brief Returns a handle to a texture reference
5713 *
5714 * Returns in \p *pTexRef the handle of the texture reference of name \p name
5715 * in the module \p hmod. If no texture reference of that name exists,
5716 * ::cuModuleGetTexRef() returns ::CUDA_ERROR_NOT_FOUND. This texture reference
5717 * handle should not be destroyed, since it will be destroyed when the module
5718 * is unloaded.
5719 *
5720 * \param pTexRef - Returned texture reference
5721 * \param hmod - Module to retrieve texture reference from
5722 * \param name - Name of texture reference to retrieve
5723 *
5724 * \return
5725 * ::CUDA_SUCCESS,
5726 * ::CUDA_ERROR_DEINITIALIZED,
5727 * ::CUDA_ERROR_NOT_INITIALIZED,
5728 * ::CUDA_ERROR_INVALID_CONTEXT,
5729 * ::CUDA_ERROR_INVALID_VALUE,
5730 * ::CUDA_ERROR_NOT_FOUND
5731 * \notefnerr
5732 *
5733 * \sa ::cuModuleGetFunction,
5734 * ::cuModuleGetGlobal,
5735 * ::cuModuleGetSurfRef,
5736 * ::cuModuleLoad,
5737 * ::cuModuleLoadData,
5738 * ::cuModuleLoadDataEx,
5739 * ::cuModuleLoadFatBinary,
5740 * ::cuModuleUnload,
5741 * ::cudaGetTextureReference
5742 */
5743CUresult CUDAAPI cuModuleGetTexRef(CUtexref *pTexRef, CUmodule hmod, const char *name);
5744
5745/**
5746 * \brief Returns a handle to a surface reference
5747 *
5748 * Returns in \p *pSurfRef the handle of the surface reference of name \p name
5749 * in the module \p hmod. If no surface reference of that name exists,
5750 * ::cuModuleGetSurfRef() returns ::CUDA_ERROR_NOT_FOUND.
5751 *
5752 * \param pSurfRef - Returned surface reference
5753 * \param hmod - Module to retrieve surface reference from
5754 * \param name - Name of surface reference to retrieve
5755 *
5756 * \return
5757 * ::CUDA_SUCCESS,
5758 * ::CUDA_ERROR_DEINITIALIZED,
5759 * ::CUDA_ERROR_NOT_INITIALIZED,
5760 * ::CUDA_ERROR_INVALID_CONTEXT,
5761 * ::CUDA_ERROR_INVALID_VALUE,
5762 * ::CUDA_ERROR_NOT_FOUND
5763 * \notefnerr
5764 *
5765 * \sa ::cuModuleGetFunction,
5766 * ::cuModuleGetGlobal,
5767 * ::cuModuleGetTexRef,
5768 * ::cuModuleLoad,
5769 * ::cuModuleLoadData,
5770 * ::cuModuleLoadDataEx,
5771 * ::cuModuleLoadFatBinary,
5772 * ::cuModuleUnload,
5773 * ::cudaGetSurfaceReference
5774 */
5775CUresult CUDAAPI cuModuleGetSurfRef(CUsurfref *pSurfRef, CUmodule hmod, const char *name);
5776
5777/**
5778 * \brief Creates a pending JIT linker invocation.
5779 *
5780 * If the call is successful, the caller owns the returned CUlinkState, which
5781 * should eventually be destroyed with ::cuLinkDestroy. The
5782 * device code machine size (32 or 64 bit) will match the calling application.
5783 *
5784 * Both linker and compiler options may be specified. Compiler options will
5785 * be applied to inputs to this linker action which must be compiled from PTX.
5786 * The options ::CU_JIT_WALL_TIME,
5787 * ::CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES, and ::CU_JIT_ERROR_LOG_BUFFER_SIZE_BYTES
5788 * will accumulate data until the CUlinkState is destroyed.
5789 *
5790 * \p optionValues must remain valid for the life of the CUlinkState if output
5791 * options are used. No other references to inputs are maintained after this
5792 * call returns.
5793 *
5794 * \param numOptions Size of options arrays
5795 * \param options Array of linker and compiler options
5796 * \param optionValues Array of option values, each cast to void *
5797 * \param stateOut On success, this will contain a CUlinkState to specify
5798 * and complete this action
5799 *
5800 * \return
5801 * ::CUDA_SUCCESS,
5802 * ::CUDA_ERROR_DEINITIALIZED,
5803 * ::CUDA_ERROR_NOT_INITIALIZED,
5804 * ::CUDA_ERROR_INVALID_CONTEXT,
5805 * ::CUDA_ERROR_INVALID_VALUE,
5806 * ::CUDA_ERROR_OUT_OF_MEMORY,
5807 * ::CUDA_ERROR_JIT_COMPILER_NOT_FOUND
5808 * \notefnerr
5809 *
5810 * \sa ::cuLinkAddData,
5811 * ::cuLinkAddFile,
5812 * ::cuLinkComplete,
5813 * ::cuLinkDestroy
5814 */
5815CUresult CUDAAPI
5816cuLinkCreate(unsigned int numOptions, CUjit_option *options, void **optionValues, CUlinkState *stateOut);
5817
5818/**
5819 * \brief Add an input to a pending linker invocation
5820 *
5821 * Ownership of \p data is retained by the caller. No reference is retained to any
5822 * inputs after this call returns.
5823 *
5824 * This method accepts only compiler options, which are used if the data must
5825 * be compiled from PTX, and does not accept any of
5826 * ::CU_JIT_WALL_TIME, ::CU_JIT_INFO_LOG_BUFFER, ::CU_JIT_ERROR_LOG_BUFFER,
5827 * ::CU_JIT_TARGET_FROM_CUCONTEXT, or ::CU_JIT_TARGET.
5828 *
5829 * \param state A pending linker action.
5830 * \param type The type of the input data.
5831 * \param data The input data. PTX must be NULL-terminated.
5832 * \param size The length of the input data.
5833 * \param name An optional name for this input in log messages.
5834 * \param numOptions Size of options.
5835 * \param options Options to be applied only for this input (overrides options from ::cuLinkCreate).
5836 * \param optionValues Array of option values, each cast to void *.
5837 *
5838 * \return
5839 * ::CUDA_SUCCESS,
5840 * ::CUDA_ERROR_INVALID_HANDLE,
5841 * ::CUDA_ERROR_INVALID_VALUE,
5842 * ::CUDA_ERROR_INVALID_IMAGE,
5843 * ::CUDA_ERROR_INVALID_PTX,
5844 * ::CUDA_ERROR_UNSUPPORTED_PTX_VERSION,
5845 * ::CUDA_ERROR_OUT_OF_MEMORY,
5846 * ::CUDA_ERROR_NO_BINARY_FOR_GPU
5847 *
5848 * \sa ::cuLinkCreate,
5849 * ::cuLinkAddFile,
5850 * ::cuLinkComplete,
5851 * ::cuLinkDestroy
5852 */
5853CUresult CUDAAPI
5854cuLinkAddData(CUlinkState state, CUjitInputType type, void *data, size_t size, const char *name,
5855 unsigned int numOptions, CUjit_option *options, void **optionValues);
5856
5857/**
5858 * \brief Add a file input to a pending linker invocation
5859 *
5860 * No reference is retained to any inputs after this call returns.
5861 *
5862 * This method accepts only compiler options, which are used if the input
5863 * must be compiled from PTX, and does not accept any of
5864 * ::CU_JIT_WALL_TIME, ::CU_JIT_INFO_LOG_BUFFER, ::CU_JIT_ERROR_LOG_BUFFER,
5865 * ::CU_JIT_TARGET_FROM_CUCONTEXT, or ::CU_JIT_TARGET.
5866 *
5867 * This method is equivalent to invoking ::cuLinkAddData on the contents
5868 * of the file.
5869 *
5870 * \param state A pending linker action
5871 * \param type The type of the input data
5872 * \param path Path to the input file
5873 * \param numOptions Size of options
5874 * \param options Options to be applied only for this input (overrides options from ::cuLinkCreate)
5875 * \param optionValues Array of option values, each cast to void *
5876 *
5877 * \return
5878 * ::CUDA_SUCCESS,
5879 * ::CUDA_ERROR_FILE_NOT_FOUND
5880 * ::CUDA_ERROR_INVALID_HANDLE,
5881 * ::CUDA_ERROR_INVALID_VALUE,
5882 * ::CUDA_ERROR_INVALID_IMAGE,
5883 * ::CUDA_ERROR_INVALID_PTX,
5884 * ::CUDA_ERROR_UNSUPPORTED_PTX_VERSION,
5885 * ::CUDA_ERROR_OUT_OF_MEMORY,
5886 * ::CUDA_ERROR_NO_BINARY_FOR_GPU
5887 *
5888 * \sa ::cuLinkCreate,
5889 * ::cuLinkAddData,
5890 * ::cuLinkComplete,
5891 * ::cuLinkDestroy
5892 */
5893CUresult CUDAAPI
5894cuLinkAddFile(CUlinkState state, CUjitInputType type, const char *path,
5895 unsigned int numOptions, CUjit_option *options, void **optionValues);
5896
5897/**
5898 * \brief Complete a pending linker invocation
5899 *
5900 * Completes the pending linker action and returns the cubin image for the linked
5901 * device code, which can be used with ::cuModuleLoadData. The cubin is owned by
5902 * \p state, so it should be loaded before \p state is destroyed via ::cuLinkDestroy.
5903 * This call does not destroy \p state.
5904 *
5905 * \param state A pending linker invocation
5906 * \param cubinOut On success, this will point to the output image
5907 * \param sizeOut Optional parameter to receive the size of the generated image
5908 *
5909 * \return
5910 * ::CUDA_SUCCESS,
5911 * ::CUDA_ERROR_INVALID_HANDLE,
5912 * ::CUDA_ERROR_OUT_OF_MEMORY
5913 *
5914 * \sa ::cuLinkCreate,
5915 * ::cuLinkAddData,
5916 * ::cuLinkAddFile,
5917 * ::cuLinkDestroy,
5918 * ::cuModuleLoadData
5919 */
5920CUresult CUDAAPI
5921cuLinkComplete(CUlinkState state, void **cubinOut, size_t *sizeOut);
5922
5923/**
5924 * \brief Destroys state for a JIT linker invocation.
5925 *
5926 * \param state State object for the linker invocation
5927 *
5928 * \return
5929 * ::CUDA_SUCCESS,
5930 * ::CUDA_ERROR_INVALID_HANDLE
5931 *
5932 * \sa ::cuLinkCreate
5933 */
5934CUresult CUDAAPI
5935cuLinkDestroy(CUlinkState state);
5936
5937/** @} */ /* END CUDA_MODULE */
5938
5939
5940/**
5941 * \defgroup CUDA_MEM Memory Management
5942 *
5943 * ___MANBRIEF___ memory management functions of the low-level CUDA driver API
5944 * (___CURRENT_FILE___) ___ENDMANBRIEF___
5945 *
5946 * This section describes the memory management functions of the low-level CUDA
5947 * driver application programming interface.
5948 *
5949 * @{
5950 */
5951
5952/**
5953 * \brief Gets free and total memory
5954 *
5955 * Returns in \p *free and \p *total respectively, the free and total amount of
5956 * memory available for allocation by the CUDA context, in bytes.
5957 *
5958 * \param free - Returned free memory in bytes
5959 * \param total - Returned total memory in bytes
5960 *
5961 * \return
5962 * ::CUDA_SUCCESS,
5963 * ::CUDA_ERROR_DEINITIALIZED,
5964 * ::CUDA_ERROR_NOT_INITIALIZED,
5965 * ::CUDA_ERROR_INVALID_CONTEXT,
5966 * ::CUDA_ERROR_INVALID_VALUE
5967 * \notefnerr
5968 *
5969 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
5970 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
5971 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
5972 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
5973 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
5974 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
5975 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
5976 * ::cuMemGetAddressRange, ::cuMemHostAlloc,
5977 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
5978 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
5979 * ::cudaMemGetInfo
5980 */
5981CUresult CUDAAPI cuMemGetInfo(size_t *free, size_t *total);
5982
5983/**
5984 * \brief Allocates device memory
5985 *
5986 * Allocates \p bytesize bytes of linear memory on the device and returns in
5987 * \p *dptr a pointer to the allocated memory. The allocated memory is suitably
5988 * aligned for any kind of variable. The memory is not cleared. If \p bytesize
5989 * is 0, ::cuMemAlloc() returns ::CUDA_ERROR_INVALID_VALUE.
5990 *
5991 * \param dptr - Returned device pointer
5992 * \param bytesize - Requested allocation size in bytes
5993 *
5994 * \return
5995 * ::CUDA_SUCCESS,
5996 * ::CUDA_ERROR_DEINITIALIZED,
5997 * ::CUDA_ERROR_NOT_INITIALIZED,
5998 * ::CUDA_ERROR_INVALID_CONTEXT,
5999 * ::CUDA_ERROR_INVALID_VALUE,
6000 * ::CUDA_ERROR_OUT_OF_MEMORY
6001 * \notefnerr
6002 *
6003 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
6004 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAllocHost,
6005 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
6006 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
6007 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
6008 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
6009 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
6010 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
6011 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
6012 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
6013 * ::cudaMalloc
6014 */
6015CUresult CUDAAPI cuMemAlloc(CUdeviceptr *dptr, size_t bytesize);
6016
6017/**
6018 * \brief Allocates pitched device memory
6019 *
6020 * Allocates at least \p WidthInBytes * \p Height bytes of linear memory on
6021 * the device and returns in \p *dptr a pointer to the allocated memory. The
6022 * function may pad the allocation to ensure that corresponding pointers in
6023 * any given row will continue to meet the alignment requirements for
6024 * coalescing as the address is updated from row to row. \p ElementSizeBytes
6025 * specifies the size of the largest reads and writes that will be performed
6026 * on the memory range. \p ElementSizeBytes may be 4, 8 or 16 (since coalesced
6027 * memory transactions are not possible on other data sizes). If
6028 * \p ElementSizeBytes is smaller than the actual read/write size of a kernel,
6029 * the kernel will run correctly, but possibly at reduced speed. The pitch
6030 * returned in \p *pPitch by ::cuMemAllocPitch() is the width in bytes of the
6031 * allocation. The intended usage of pitch is as a separate parameter of the
6032 * allocation, used to compute addresses within the 2D array. Given the row
6033 * and column of an array element of type \b T, the address is computed as:
6034 * \code
6035 T* pElement = (T*)((char*)BaseAddress + Row * Pitch) + Column;
6036 * \endcode
6037 *
6038 * The pitch returned by ::cuMemAllocPitch() is guaranteed to work with
6039 * ::cuMemcpy2D() under all circumstances. For allocations of 2D arrays, it is
6040 * recommended that programmers consider performing pitch allocations using
6041 * ::cuMemAllocPitch(). Due to alignment restrictions in the hardware, this is
6042 * especially true if the application will be performing 2D memory copies
6043 * between different regions of device memory (whether linear memory or CUDA
6044 * arrays).
6045 *
6046 * The byte alignment of the pitch returned by ::cuMemAllocPitch() is guaranteed
6047 * to match or exceed the alignment requirement for texture binding with
6048 * ::cuTexRefSetAddress2D().
6049 *
6050 * \param dptr - Returned device pointer
6051 * \param pPitch - Returned pitch of allocation in bytes
6052 * \param WidthInBytes - Requested allocation width in bytes
6053 * \param Height - Requested allocation height in rows
6054 * \param ElementSizeBytes - Size of largest reads/writes for range
6055 *
6056 * \return
6057 * ::CUDA_SUCCESS,
6058 * ::CUDA_ERROR_DEINITIALIZED,
6059 * ::CUDA_ERROR_NOT_INITIALIZED,
6060 * ::CUDA_ERROR_INVALID_CONTEXT,
6061 * ::CUDA_ERROR_INVALID_VALUE,
6062 * ::CUDA_ERROR_OUT_OF_MEMORY
6063 * \notefnerr
6064 *
6065 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
6066 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
6067 * ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
6068 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
6069 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
6070 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
6071 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
6072 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
6073 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
6074 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
6075 * ::cudaMallocPitch
6076 */
6077CUresult CUDAAPI cuMemAllocPitch(CUdeviceptr *dptr, size_t *pPitch, size_t WidthInBytes, size_t Height, unsigned int ElementSizeBytes);
6078
6079/**
6080 * \brief Frees device memory
6081 *
6082 * Frees the memory space pointed to by \p dptr, which must have been returned
6083 * by a previous call to ::cuMemAlloc() or ::cuMemAllocPitch().
6084 *
6085 * \param dptr - Pointer to memory to free
6086 *
6087 * \return
6088 * ::CUDA_SUCCESS,
6089 * ::CUDA_ERROR_DEINITIALIZED,
6090 * ::CUDA_ERROR_NOT_INITIALIZED,
6091 * ::CUDA_ERROR_INVALID_CONTEXT,
6092 * ::CUDA_ERROR_INVALID_VALUE
6093 * \notefnerr
6094 *
6095 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
6096 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
6097 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
6098 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
6099 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
6100 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
6101 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFreeHost,
6102 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
6103 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
6104 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
6105 * ::cudaFree
6106 */
6107CUresult CUDAAPI cuMemFree(CUdeviceptr dptr);
6108
6109/**
6110 * \brief Get information on memory allocations
6111 *
6112 * Returns the base address in \p *pbase and size in \p *psize of the
6113 * allocation by ::cuMemAlloc() or ::cuMemAllocPitch() that contains the input
6114 * pointer \p dptr. Both parameters \p pbase and \p psize are optional. If one
6115 * of them is NULL, it is ignored.
6116 *
6117 * \param pbase - Returned base address
6118 * \param psize - Returned size of device memory allocation
6119 * \param dptr - Device pointer to query
6120 *
6121 * \return
6122 * ::CUDA_SUCCESS,
6123 * ::CUDA_ERROR_DEINITIALIZED,
6124 * ::CUDA_ERROR_NOT_INITIALIZED,
6125 * ::CUDA_ERROR_INVALID_CONTEXT,
6126 * ::CUDA_ERROR_NOT_FOUND,
6127 * ::CUDA_ERROR_INVALID_VALUE
6128 * \notefnerr
6129 *
6130 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
6131 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
6132 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
6133 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
6134 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
6135 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
6136 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
6137 * ::cuMemGetInfo, ::cuMemHostAlloc,
6138 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
6139 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32
6140 */
6141CUresult CUDAAPI cuMemGetAddressRange(CUdeviceptr *pbase, size_t *psize, CUdeviceptr dptr);
6142
6143/**
6144 * \brief Allocates page-locked host memory
6145 *
6146 * Allocates \p bytesize bytes of host memory that is page-locked and
6147 * accessible to the device. The driver tracks the virtual memory ranges
6148 * allocated with this function and automatically accelerates calls to
6149 * functions such as ::cuMemcpy(). Since the memory can be accessed directly by
6150 * the device, it can be read or written with much higher bandwidth than
6151 * pageable memory obtained with functions such as ::malloc(). Allocating
6152 * excessive amounts of memory with ::cuMemAllocHost() may degrade system
6153 * performance, since it reduces the amount of memory available to the system
6154 * for paging. As a result, this function is best used sparingly to allocate
6155 * staging areas for data exchange between host and device.
6156 *
6157 * Note all host memory allocated using ::cuMemHostAlloc() will automatically
6158 * be immediately accessible to all contexts on all devices which support unified
6159 * addressing (as may be queried using ::CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING).
6160 * The device pointer that may be used to access this host memory from those
6161 * contexts is always equal to the returned host pointer \p *pp.
6162 * See \ref CUDA_UNIFIED for additional details.
6163 *
6164 * \param pp - Returned host pointer to page-locked memory
6165 * \param bytesize - Requested allocation size in bytes
6166 *
6167 * \return
6168 * ::CUDA_SUCCESS,
6169 * ::CUDA_ERROR_DEINITIALIZED,
6170 * ::CUDA_ERROR_NOT_INITIALIZED,
6171 * ::CUDA_ERROR_INVALID_CONTEXT,
6172 * ::CUDA_ERROR_INVALID_VALUE,
6173 * ::CUDA_ERROR_OUT_OF_MEMORY
6174 * \notefnerr
6175 *
6176 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
6177 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc,
6178 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
6179 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
6180 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
6181 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
6182 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
6183 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
6184 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
6185 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
6186 * ::cudaMallocHost
6187 */
6188CUresult CUDAAPI cuMemAllocHost(void **pp, size_t bytesize);
6189
6190/**
6191 * \brief Frees page-locked host memory
6192 *
6193 * Frees the memory space pointed to by \p p, which must have been returned by
6194 * a previous call to ::cuMemAllocHost().
6195 *
6196 * \param p - Pointer to memory to free
6197 *
6198 * \return
6199 * ::CUDA_SUCCESS,
6200 * ::CUDA_ERROR_DEINITIALIZED,
6201 * ::CUDA_ERROR_NOT_INITIALIZED,
6202 * ::CUDA_ERROR_INVALID_CONTEXT,
6203 * ::CUDA_ERROR_INVALID_VALUE
6204 * \notefnerr
6205 *
6206 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
6207 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
6208 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
6209 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
6210 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
6211 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
6212 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree,
6213 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
6214 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
6215 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
6216 * ::cudaFreeHost
6217 */
6218CUresult CUDAAPI cuMemFreeHost(void *p);
6219
6220/**
6221 * \brief Allocates page-locked host memory
6222 *
6223 * Allocates \p bytesize bytes of host memory that is page-locked and accessible
6224 * to the device. The driver tracks the virtual memory ranges allocated with
6225 * this function and automatically accelerates calls to functions such as
6226 * ::cuMemcpyHtoD(). Since the memory can be accessed directly by the device,
6227 * it can be read or written with much higher bandwidth than pageable memory
6228 * obtained with functions such as ::malloc(). Allocating excessive amounts of
6229 * pinned memory may degrade system performance, since it reduces the amount
6230 * of memory available to the system for paging. As a result, this function is
6231 * best used sparingly to allocate staging areas for data exchange between
6232 * host and device.
6233 *
6234 * The \p Flags parameter enables different options to be specified that
6235 * affect the allocation, as follows.
6236 *
6237 * - ::CU_MEMHOSTALLOC_PORTABLE: The memory returned by this call will be
6238 * considered as pinned memory by all CUDA contexts, not just the one that
6239 * performed the allocation.
6240 *
6241 * - ::CU_MEMHOSTALLOC_DEVICEMAP: Maps the allocation into the CUDA address
6242 * space. The device pointer to the memory may be obtained by calling
6243 * ::cuMemHostGetDevicePointer().
6244 *
6245 * - ::CU_MEMHOSTALLOC_WRITECOMBINED: Allocates the memory as write-combined
6246 * (WC). WC memory can be transferred across the PCI Express bus more
6247 * quickly on some system configurations, but cannot be read efficiently by
6248 * most CPUs. WC memory is a good option for buffers that will be written by
6249 * the CPU and read by the GPU via mapped pinned memory or host->device
6250 * transfers.
6251 *
6252 * All of these flags are orthogonal to one another: a developer may allocate
6253 * memory that is portable, mapped and/or write-combined with no restrictions.
6254 *
6255 * The ::CU_MEMHOSTALLOC_DEVICEMAP flag may be specified on CUDA contexts for
6256 * devices that do not support mapped pinned memory. The failure is deferred
6257 * to ::cuMemHostGetDevicePointer() because the memory may be mapped into
6258 * other CUDA contexts via the ::CU_MEMHOSTALLOC_PORTABLE flag.
6259 *
6260 * The memory allocated by this function must be freed with ::cuMemFreeHost().
6261 *
6262 * Note all host memory allocated using ::cuMemHostAlloc() will automatically
6263 * be immediately accessible to all contexts on all devices which support unified
6264 * addressing (as may be queried using ::CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING).
6265 * Unless the flag ::CU_MEMHOSTALLOC_WRITECOMBINED is specified, the device pointer
6266 * that may be used to access this host memory from those contexts is always equal
6267 * to the returned host pointer \p *pp. If the flag ::CU_MEMHOSTALLOC_WRITECOMBINED
6268 * is specified, then the function ::cuMemHostGetDevicePointer() must be used
6269 * to query the device pointer, even if the context supports unified addressing.
6270 * See \ref CUDA_UNIFIED for additional details.
6271 *
6272 * \param pp - Returned host pointer to page-locked memory
6273 * \param bytesize - Requested allocation size in bytes
6274 * \param Flags - Flags for allocation request
6275 *
6276 * \return
6277 * ::CUDA_SUCCESS,
6278 * ::CUDA_ERROR_DEINITIALIZED,
6279 * ::CUDA_ERROR_NOT_INITIALIZED,
6280 * ::CUDA_ERROR_INVALID_CONTEXT,
6281 * ::CUDA_ERROR_INVALID_VALUE,
6282 * ::CUDA_ERROR_OUT_OF_MEMORY
6283 * \notefnerr
6284 *
6285 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
6286 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
6287 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
6288 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
6289 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
6290 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
6291 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
6292 * ::cuMemGetAddressRange, ::cuMemGetInfo,
6293 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
6294 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
6295 * ::cudaHostAlloc
6296 */
6297CUresult CUDAAPI cuMemHostAlloc(void **pp, size_t bytesize, unsigned int Flags);
6298
6299/**
6300 * \brief Passes back device pointer of mapped pinned memory
6301 *
6302 * Passes back the device pointer \p pdptr corresponding to the mapped, pinned
6303 * host buffer \p p allocated by ::cuMemHostAlloc.
6304 *
6305 * ::cuMemHostGetDevicePointer() will fail if the ::CU_MEMHOSTALLOC_DEVICEMAP
6306 * flag was not specified at the time the memory was allocated, or if the
6307 * function is called on a GPU that does not support mapped pinned memory.
6308 *
6309 * For devices that have a non-zero value for the device attribute
6310 * ::CU_DEVICE_ATTRIBUTE_CAN_USE_HOST_POINTER_FOR_REGISTERED_MEM, the memory
6311 * can also be accessed from the device using the host pointer \p p.
6312 * The device pointer returned by ::cuMemHostGetDevicePointer() may or may not
6313 * match the original host pointer \p p and depends on the devices visible to the
6314 * application. If all devices visible to the application have a non-zero value for the
6315 * device attribute, the device pointer returned by ::cuMemHostGetDevicePointer()
6316 * will match the original pointer \p p. If any device visible to the application
6317 * has a zero value for the device attribute, the device pointer returned by
6318 * ::cuMemHostGetDevicePointer() will not match the original host pointer \p p,
6319 * but it will be suitable for use on all devices provided Unified Virtual Addressing
6320 * is enabled. In such systems, it is valid to access the memory using either pointer
6321 * on devices that have a non-zero value for the device attribute. Note however that
6322 * such devices should access the memory using only of the two pointers and not both.
6323 *
6324 * \p Flags provides for future releases. For now, it must be set to 0.
6325 *
6326 * \param pdptr - Returned device pointer
6327 * \param p - Host pointer
6328 * \param Flags - Options (must be 0)
6329 *
6330 * \return
6331 * ::CUDA_SUCCESS,
6332 * ::CUDA_ERROR_DEINITIALIZED,
6333 * ::CUDA_ERROR_NOT_INITIALIZED,
6334 * ::CUDA_ERROR_INVALID_CONTEXT,
6335 * ::CUDA_ERROR_INVALID_VALUE
6336 * \notefnerr
6337 *
6338 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
6339 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
6340 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
6341 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
6342 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
6343 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
6344 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
6345 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
6346 * ::cuMemsetD2D8, ::cuMemsetD2D16,
6347 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
6348 * ::cudaHostGetDevicePointer
6349 */
6350CUresult CUDAAPI cuMemHostGetDevicePointer(CUdeviceptr *pdptr, void *p, unsigned int Flags);
6351
6352/**
6353 * \brief Passes back flags that were used for a pinned allocation
6354 *
6355 * Passes back the flags \p pFlags that were specified when allocating
6356 * the pinned host buffer \p p allocated by ::cuMemHostAlloc.
6357 *
6358 * ::cuMemHostGetFlags() will fail if the pointer does not reside in
6359 * an allocation performed by ::cuMemAllocHost() or ::cuMemHostAlloc().
6360 *
6361 * \param pFlags - Returned flags word
6362 * \param p - Host pointer
6363 *
6364 * \return
6365 * ::CUDA_SUCCESS,
6366 * ::CUDA_ERROR_DEINITIALIZED,
6367 * ::CUDA_ERROR_NOT_INITIALIZED,
6368 * ::CUDA_ERROR_INVALID_CONTEXT,
6369 * ::CUDA_ERROR_INVALID_VALUE
6370 * \notefnerr
6371 *
6372 * \sa
6373 * ::cuMemAllocHost,
6374 * ::cuMemHostAlloc,
6375 * ::cudaHostGetFlags
6376 */
6377CUresult CUDAAPI cuMemHostGetFlags(unsigned int *pFlags, void *p);
6378
6379/**
6380 * \brief Allocates memory that will be automatically managed by the Unified Memory system
6381 *
6382 * Allocates \p bytesize bytes of managed memory on the device and returns in
6383 * \p *dptr a pointer to the allocated memory. If the device doesn't support
6384 * allocating managed memory, ::CUDA_ERROR_NOT_SUPPORTED is returned. Support
6385 * for managed memory can be queried using the device attribute
6386 * ::CU_DEVICE_ATTRIBUTE_MANAGED_MEMORY. The allocated memory is suitably
6387 * aligned for any kind of variable. The memory is not cleared. If \p bytesize
6388 * is 0, ::cuMemAllocManaged returns ::CUDA_ERROR_INVALID_VALUE. The pointer
6389 * is valid on the CPU and on all GPUs in the system that support managed memory.
6390 * All accesses to this pointer must obey the Unified Memory programming model.
6391 *
6392 * \p flags specifies the default stream association for this allocation.
6393 * \p flags must be one of ::CU_MEM_ATTACH_GLOBAL or ::CU_MEM_ATTACH_HOST. If
6394 * ::CU_MEM_ATTACH_GLOBAL is specified, then this memory is accessible from
6395 * any stream on any device. If ::CU_MEM_ATTACH_HOST is specified, then the
6396 * allocation should not be accessed from devices that have a zero value for the
6397 * device attribute ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS; an explicit call to
6398 * ::cuStreamAttachMemAsync will be required to enable access on such devices.
6399 *
6400 * If the association is later changed via ::cuStreamAttachMemAsync to
6401 * a single stream, the default association as specified during ::cuMemAllocManaged
6402 * is restored when that stream is destroyed. For __managed__ variables, the
6403 * default association is always ::CU_MEM_ATTACH_GLOBAL. Note that destroying a
6404 * stream is an asynchronous operation, and as a result, the change to default
6405 * association won't happen until all work in the stream has completed.
6406 *
6407 * Memory allocated with ::cuMemAllocManaged should be released with ::cuMemFree.
6408 *
6409 * Device memory oversubscription is possible for GPUs that have a non-zero value for the
6410 * device attribute ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS. Managed memory on
6411 * such GPUs may be evicted from device memory to host memory at any time by the Unified
6412 * Memory driver in order to make room for other allocations.
6413 *
6414 * In a multi-GPU system where all GPUs have a non-zero value for the device attribute
6415 * ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS, managed memory may not be populated when this
6416 * API returns and instead may be populated on access. In such systems, managed memory can
6417 * migrate to any processor's memory at any time. The Unified Memory driver will employ heuristics to
6418 * maintain data locality and prevent excessive page faults to the extent possible. The application
6419 * can also guide the driver about memory usage patterns via ::cuMemAdvise. The application
6420 * can also explicitly migrate memory to a desired processor's memory via
6421 * ::cuMemPrefetchAsync.
6422 *
6423 * In a multi-GPU system where all of the GPUs have a zero value for the device attribute
6424 * ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS and all the GPUs have peer-to-peer support
6425 * with each other, the physical storage for managed memory is created on the GPU which is active
6426 * at the time ::cuMemAllocManaged is called. All other GPUs will reference the data at reduced
6427 * bandwidth via peer mappings over the PCIe bus. The Unified Memory driver does not migrate
6428 * memory among such GPUs.
6429 *
6430 * In a multi-GPU system where not all GPUs have peer-to-peer support with each other and
6431 * where the value of the device attribute ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS
6432 * is zero for at least one of those GPUs, the location chosen for physical storage of managed
6433 * memory is system-dependent.
6434 * - On Linux, the location chosen will be device memory as long as the current set of active
6435 * contexts are on devices that either have peer-to-peer support with each other or have a
6436 * non-zero value for the device attribute ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS.
6437 * If there is an active context on a GPU that does not have a non-zero value for that device
6438 * attribute and it does not have peer-to-peer support with the other devices that have active
6439 * contexts on them, then the location for physical storage will be 'zero-copy' or host memory.
6440 * Note that this means that managed memory that is located in device memory is migrated to
6441 * host memory if a new context is created on a GPU that doesn't have a non-zero value for
6442 * the device attribute and does not support peer-to-peer with at least one of the other devices
6443 * that has an active context. This in turn implies that context creation may fail if there is
6444 * insufficient host memory to migrate all managed allocations.
6445 * - On Windows, the physical storage is always created in 'zero-copy' or host memory.
6446 * All GPUs will reference the data at reduced bandwidth over the PCIe bus. In these
6447 * circumstances, use of the environment variable CUDA_VISIBLE_DEVICES is recommended to
6448 * restrict CUDA to only use those GPUs that have peer-to-peer support.
6449 * Alternatively, users can also set CUDA_MANAGED_FORCE_DEVICE_ALLOC to a
6450 * non-zero value to force the driver to always use device memory for physical storage.
6451 * When this environment variable is set to a non-zero value, all contexts created in
6452 * that process on devices that support managed memory have to be peer-to-peer compatible
6453 * with each other. Context creation will fail if a context is created on a device that
6454 * supports managed memory and is not peer-to-peer compatible with any of the other
6455 * managed memory supporting devices on which contexts were previously created, even if
6456 * those contexts have been destroyed. These environment variables are described
6457 * in the CUDA programming guide under the "CUDA environment variables" section.
6458 * - On ARM, managed memory is not available on discrete gpu with Drive PX-2.
6459 *
6460 * \param dptr - Returned device pointer
6461 * \param bytesize - Requested allocation size in bytes
6462 * \param flags - Must be one of ::CU_MEM_ATTACH_GLOBAL or ::CU_MEM_ATTACH_HOST
6463 *
6464 * \return
6465 * ::CUDA_SUCCESS,
6466 * ::CUDA_ERROR_DEINITIALIZED,
6467 * ::CUDA_ERROR_NOT_INITIALIZED,
6468 * ::CUDA_ERROR_INVALID_CONTEXT,
6469 * ::CUDA_ERROR_NOT_SUPPORTED,
6470 * ::CUDA_ERROR_INVALID_VALUE,
6471 * ::CUDA_ERROR_OUT_OF_MEMORY
6472 * \notefnerr
6473 *
6474 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
6475 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAllocHost,
6476 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
6477 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
6478 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
6479 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
6480 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
6481 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
6482 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
6483 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
6484 * ::cuDeviceGetAttribute, ::cuStreamAttachMemAsync,
6485 * ::cudaMallocManaged
6486 */
6487CUresult CUDAAPI cuMemAllocManaged(CUdeviceptr *dptr, size_t bytesize, unsigned int flags);
6488
6489/**
6490 * \brief Returns a handle to a compute device
6491 *
6492 * Returns in \p *device a device handle given a PCI bus ID string.
6493 *
6494 * \param dev - Returned device handle
6495 *
6496 * \param pciBusId - String in one of the following forms:
6497 * [domain]:[bus]:[device].[function]
6498 * [domain]:[bus]:[device]
6499 * [bus]:[device].[function]
6500 * where \p domain, \p bus, \p device, and \p function are all hexadecimal values
6501 *
6502 * \return
6503 * ::CUDA_SUCCESS,
6504 * ::CUDA_ERROR_DEINITIALIZED,
6505 * ::CUDA_ERROR_NOT_INITIALIZED,
6506 * ::CUDA_ERROR_INVALID_VALUE,
6507 * ::CUDA_ERROR_INVALID_DEVICE
6508 * \notefnerr
6509 *
6510 * \sa
6511 * ::cuDeviceGet,
6512 * ::cuDeviceGetAttribute,
6513 * ::cuDeviceGetPCIBusId,
6514 * ::cudaDeviceGetByPCIBusId
6515 */
6516CUresult CUDAAPI cuDeviceGetByPCIBusId(CUdevice *dev, const char *pciBusId);
6517
6518/**
6519 * \brief Returns a PCI Bus Id string for the device
6520 *
6521 * Returns an ASCII string identifying the device \p dev in the NULL-terminated
6522 * string pointed to by \p pciBusId. \p len specifies the maximum length of the
6523 * string that may be returned.
6524 *
6525 * \param pciBusId - Returned identifier string for the device in the following format
6526 * [domain]:[bus]:[device].[function]
6527 * where \p domain, \p bus, \p device, and \p function are all hexadecimal values.
6528 * pciBusId should be large enough to store 13 characters including the NULL-terminator.
6529 *
6530 * \param len - Maximum length of string to store in \p name
6531 *
6532 * \param dev - Device to get identifier string for
6533 *
6534 * \return
6535 * ::CUDA_SUCCESS,
6536 * ::CUDA_ERROR_DEINITIALIZED,
6537 * ::CUDA_ERROR_NOT_INITIALIZED,
6538 * ::CUDA_ERROR_INVALID_VALUE,
6539 * ::CUDA_ERROR_INVALID_DEVICE
6540 * \notefnerr
6541 *
6542 * \sa
6543 * ::cuDeviceGet,
6544 * ::cuDeviceGetAttribute,
6545 * ::cuDeviceGetByPCIBusId,
6546 * ::cudaDeviceGetPCIBusId
6547 */
6548CUresult CUDAAPI cuDeviceGetPCIBusId(char *pciBusId, int len, CUdevice dev);
6549
6550/**
6551 * \brief Gets an interprocess handle for a previously allocated event
6552 *
6553 * Takes as input a previously allocated event. This event must have been
6554 * created with the ::CU_EVENT_INTERPROCESS and ::CU_EVENT_DISABLE_TIMING
6555 * flags set. This opaque handle may be copied into other processes and
6556 * opened with ::cuIpcOpenEventHandle to allow efficient hardware
6557 * synchronization between GPU work in different processes.
6558 *
6559 * After the event has been opened in the importing process,
6560 * ::cuEventRecord, ::cuEventSynchronize, ::cuStreamWaitEvent and
6561 * ::cuEventQuery may be used in either process. Performing operations
6562 * on the imported event after the exported event has been freed
6563 * with ::cuEventDestroy will result in undefined behavior.
6564 *
6565 * IPC functionality is restricted to devices with support for unified
6566 * addressing on Linux and Windows operating systems.
6567 * IPC functionality on Windows is restricted to GPUs in TCC mode
6568 *
6569 * \param pHandle - Pointer to a user allocated CUipcEventHandle
6570 * in which to return the opaque event handle
6571 * \param event - Event allocated with ::CU_EVENT_INTERPROCESS and
6572 * ::CU_EVENT_DISABLE_TIMING flags.
6573 *
6574 * \return
6575 * ::CUDA_SUCCESS,
6576 * ::CUDA_ERROR_INVALID_HANDLE,
6577 * ::CUDA_ERROR_OUT_OF_MEMORY,
6578 * ::CUDA_ERROR_MAP_FAILED,
6579 * ::CUDA_ERROR_INVALID_VALUE
6580 *
6581 * \sa
6582 * ::cuEventCreate,
6583 * ::cuEventDestroy,
6584 * ::cuEventSynchronize,
6585 * ::cuEventQuery,
6586 * ::cuStreamWaitEvent,
6587 * ::cuIpcOpenEventHandle,
6588 * ::cuIpcGetMemHandle,
6589 * ::cuIpcOpenMemHandle,
6590 * ::cuIpcCloseMemHandle,
6591 * ::cudaIpcGetEventHandle
6592 */
6593CUresult CUDAAPI cuIpcGetEventHandle(CUipcEventHandle *pHandle, CUevent event);
6594
6595/**
6596 * \brief Opens an interprocess event handle for use in the current process
6597 *
6598 * Opens an interprocess event handle exported from another process with
6599 * ::cuIpcGetEventHandle. This function returns a ::CUevent that behaves like
6600 * a locally created event with the ::CU_EVENT_DISABLE_TIMING flag specified.
6601 * This event must be freed with ::cuEventDestroy.
6602 *
6603 * Performing operations on the imported event after the exported event has
6604 * been freed with ::cuEventDestroy will result in undefined behavior.
6605 *
6606 * IPC functionality is restricted to devices with support for unified
6607 * addressing on Linux and Windows operating systems.
6608 * IPC functionality on Windows is restricted to GPUs in TCC mode
6609 *
6610 * \param phEvent - Returns the imported event
6611 * \param handle - Interprocess handle to open
6612 *
6613 * \returns
6614 * ::CUDA_SUCCESS,
6615 * ::CUDA_ERROR_INVALID_CONTEXT,
6616 * ::CUDA_ERROR_MAP_FAILED,
6617 * ::CUDA_ERROR_PEER_ACCESS_UNSUPPORTED,
6618 * ::CUDA_ERROR_INVALID_HANDLE,
6619 * ::CUDA_ERROR_INVALID_VALUE
6620 *
6621 * \sa
6622 * ::cuEventCreate,
6623 * ::cuEventDestroy,
6624 * ::cuEventSynchronize,
6625 * ::cuEventQuery,
6626 * ::cuStreamWaitEvent,
6627 * ::cuIpcGetEventHandle,
6628 * ::cuIpcGetMemHandle,
6629 * ::cuIpcOpenMemHandle,
6630 * ::cuIpcCloseMemHandle,
6631 * ::cudaIpcOpenEventHandle
6632 */
6633CUresult CUDAAPI cuIpcOpenEventHandle(CUevent *phEvent, CUipcEventHandle handle);
6634
6635/**
6636 * \brief Gets an interprocess memory handle for an existing device memory
6637 * allocation
6638 *
6639 * Takes a pointer to the base of an existing device memory allocation created
6640 * with ::cuMemAlloc and exports it for use in another process. This is a
6641 * lightweight operation and may be called multiple times on an allocation
6642 * without adverse effects.
6643 *
6644 * If a region of memory is freed with ::cuMemFree and a subsequent call
6645 * to ::cuMemAlloc returns memory with the same device address,
6646 * ::cuIpcGetMemHandle will return a unique handle for the
6647 * new memory.
6648 *
6649 * IPC functionality is restricted to devices with support for unified
6650 * addressing on Linux and Windows operating systems.
6651 * IPC functionality on Windows is restricted to GPUs in TCC mode
6652 *
6653 * \param pHandle - Pointer to user allocated ::CUipcMemHandle to return
6654 * the handle in.
6655 * \param dptr - Base pointer to previously allocated device memory
6656 *
6657 * \returns
6658 * ::CUDA_SUCCESS,
6659 * ::CUDA_ERROR_INVALID_HANDLE,
6660 * ::CUDA_ERROR_INVALID_CONTEXT,
6661 * ::CUDA_ERROR_OUT_OF_MEMORY,
6662 * ::CUDA_ERROR_MAP_FAILED,
6663 * ::CUDA_ERROR_INVALID_VALUE
6664 *
6665 * \sa
6666 * ::cuMemAlloc,
6667 * ::cuMemFree,
6668 * ::cuIpcGetEventHandle,
6669 * ::cuIpcOpenEventHandle,
6670 * ::cuIpcOpenMemHandle,
6671 * ::cuIpcCloseMemHandle,
6672 * ::cudaIpcGetMemHandle
6673 */
6674CUresult CUDAAPI cuIpcGetMemHandle(CUipcMemHandle *pHandle, CUdeviceptr dptr);
6675
6676/**
6677 * \brief Opens an interprocess memory handle exported from another process
6678 * and returns a device pointer usable in the local process.
6679 *
6680 * Maps memory exported from another process with ::cuIpcGetMemHandle into
6681 * the current device address space. For contexts on different devices
6682 * ::cuIpcOpenMemHandle can attempt to enable peer access between the
6683 * devices as if the user called ::cuCtxEnablePeerAccess. This behavior is
6684 * controlled by the ::CU_IPC_MEM_LAZY_ENABLE_PEER_ACCESS flag.
6685 * ::cuDeviceCanAccessPeer can determine if a mapping is possible.
6686 *
6687 * Contexts that may open ::CUipcMemHandles are restricted in the following way.
6688 * ::CUipcMemHandles from each ::CUdevice in a given process may only be opened
6689 * by one ::CUcontext per ::CUdevice per other process.
6690 *
6691 * If the memory handle has already been opened by the current context, the
6692 * reference count on the handle is incremented by 1 and the existing device pointer
6693 * is returned.
6694 *
6695 * Memory returned from ::cuIpcOpenMemHandle must be freed with
6696 * ::cuIpcCloseMemHandle.
6697 *
6698 * Calling ::cuMemFree on an exported memory region before calling
6699 * ::cuIpcCloseMemHandle in the importing context will result in undefined
6700 * behavior.
6701 *
6702 * IPC functionality is restricted to devices with support for unified
6703 * addressing on Linux and Windows operating systems.
6704 * IPC functionality on Windows is restricted to GPUs in TCC mode
6705 *
6706 * \param pdptr - Returned device pointer
6707 * \param handle - ::CUipcMemHandle to open
6708 * \param Flags - Flags for this operation. Must be specified as ::CU_IPC_MEM_LAZY_ENABLE_PEER_ACCESS
6709 *
6710 * \returns
6711 * ::CUDA_SUCCESS,
6712 * ::CUDA_ERROR_INVALID_CONTEXT,
6713 * ::CUDA_ERROR_MAP_FAILED,
6714 * ::CUDA_ERROR_INVALID_HANDLE,
6715 * ::CUDA_ERROR_TOO_MANY_PEERS,
6716 * ::CUDA_ERROR_INVALID_VALUE
6717 *
6718 * \note No guarantees are made about the address returned in \p *pdptr.
6719 * In particular, multiple processes may not receive the same address for the same \p handle.
6720 *
6721 * \sa
6722 * ::cuMemAlloc,
6723 * ::cuMemFree,
6724 * ::cuIpcGetEventHandle,
6725 * ::cuIpcOpenEventHandle,
6726 * ::cuIpcGetMemHandle,
6727 * ::cuIpcCloseMemHandle,
6728 * ::cuCtxEnablePeerAccess,
6729 * ::cuDeviceCanAccessPeer,
6730 * ::cudaIpcOpenMemHandle
6731 */
6732CUresult CUDAAPI cuIpcOpenMemHandle(CUdeviceptr *pdptr, CUipcMemHandle handle, unsigned int Flags);
6733
6734/**
6735 * \brief Attempts to close memory mapped with ::cuIpcOpenMemHandle
6736 *
6737 * Decrements the reference count of the memory returned by ::cuIpcOpenMemHandle by 1.
6738 * When the reference count reaches 0, this API unmaps the memory. The original allocation
6739 * in the exporting process as well as imported mappings in other processes
6740 * will be unaffected.
6741 *
6742 * Any resources used to enable peer access will be freed if this is the
6743 * last mapping using them.
6744 *
6745 * IPC functionality is restricted to devices with support for unified
6746 * addressing on Linux and Windows operating systems.
6747 * IPC functionality on Windows is restricted to GPUs in TCC mode
6748 *
6749 * \param dptr - Device pointer returned by ::cuIpcOpenMemHandle
6750 *
6751 * \returns
6752 * ::CUDA_SUCCESS,
6753 * ::CUDA_ERROR_INVALID_CONTEXT,
6754 * ::CUDA_ERROR_MAP_FAILED,
6755 * ::CUDA_ERROR_INVALID_HANDLE,
6756 * ::CUDA_ERROR_INVALID_VALUE
6757 * \sa
6758 * ::cuMemAlloc,
6759 * ::cuMemFree,
6760 * ::cuIpcGetEventHandle,
6761 * ::cuIpcOpenEventHandle,
6762 * ::cuIpcGetMemHandle,
6763 * ::cuIpcOpenMemHandle,
6764 * ::cudaIpcCloseMemHandle
6765 */
6766CUresult CUDAAPI cuIpcCloseMemHandle(CUdeviceptr dptr);
6767
6768/**
6769 * \brief Registers an existing host memory range for use by CUDA
6770 *
6771 * Page-locks the memory range specified by \p p and \p bytesize and maps it
6772 * for the device(s) as specified by \p Flags. This memory range also is added
6773 * to the same tracking mechanism as ::cuMemHostAlloc to automatically accelerate
6774 * calls to functions such as ::cuMemcpyHtoD(). Since the memory can be accessed
6775 * directly by the device, it can be read or written with much higher bandwidth
6776 * than pageable memory that has not been registered. Page-locking excessive
6777 * amounts of memory may degrade system performance, since it reduces the amount
6778 * of memory available to the system for paging. As a result, this function is
6779 * best used sparingly to register staging areas for data exchange between
6780 * host and device.
6781 *
6782 * This function has limited support on Mac OS X. OS 10.7 or higher is required.
6783 *
6784 * The \p Flags parameter enables different options to be specified that
6785 * affect the allocation, as follows.
6786 *
6787 * - ::CU_MEMHOSTREGISTER_PORTABLE: The memory returned by this call will be
6788 * considered as pinned memory by all CUDA contexts, not just the one that
6789 * performed the allocation.
6790 *
6791 * - ::CU_MEMHOSTREGISTER_DEVICEMAP: Maps the allocation into the CUDA address
6792 * space. The device pointer to the memory may be obtained by calling
6793 * ::cuMemHostGetDevicePointer().
6794 *
6795 * - ::CU_MEMHOSTREGISTER_IOMEMORY: The pointer is treated as pointing to some
6796 * I/O memory space, e.g. the PCI Express resource of a 3rd party device.
6797 *
6798 * - ::CU_MEMHOSTREGISTER_READ_ONLY: The pointer is treated as pointing to memory
6799 * that is considered read-only by the device. On platforms without
6800 * CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS_USES_HOST_PAGE_TABLES, this flag is
6801 * required in order to register memory mapped to the CPU as read-only. Support
6802 * for the use of this flag can be queried from the device attribute
6803 * CU_DEVICE_ATTRIBUTE_READ_ONLY_HOST_REGISTER_SUPPORTED. Using this flag with
6804 * a current context associated with a device that does not have this attribute
6805 * set will cause ::cuMemHostRegister to error with CUDA_ERROR_NOT_SUPPORTED.
6806 *
6807 * All of these flags are orthogonal to one another: a developer may page-lock
6808 * memory that is portable or mapped with no restrictions.
6809 *
6810 * The ::CU_MEMHOSTREGISTER_DEVICEMAP flag may be specified on CUDA contexts for
6811 * devices that do not support mapped pinned memory. The failure is deferred
6812 * to ::cuMemHostGetDevicePointer() because the memory may be mapped into
6813 * other CUDA contexts via the ::CU_MEMHOSTREGISTER_PORTABLE flag.
6814 *
6815 * For devices that have a non-zero value for the device attribute
6816 * ::CU_DEVICE_ATTRIBUTE_CAN_USE_HOST_POINTER_FOR_REGISTERED_MEM, the memory
6817 * can also be accessed from the device using the host pointer \p p.
6818 * The device pointer returned by ::cuMemHostGetDevicePointer() may or may not
6819 * match the original host pointer \p ptr and depends on the devices visible to the
6820 * application. If all devices visible to the application have a non-zero value for the
6821 * device attribute, the device pointer returned by ::cuMemHostGetDevicePointer()
6822 * will match the original pointer \p ptr. If any device visible to the application
6823 * has a zero value for the device attribute, the device pointer returned by
6824 * ::cuMemHostGetDevicePointer() will not match the original host pointer \p ptr,
6825 * but it will be suitable for use on all devices provided Unified Virtual Addressing
6826 * is enabled. In such systems, it is valid to access the memory using either pointer
6827 * on devices that have a non-zero value for the device attribute. Note however that
6828 * such devices should access the memory using only of the two pointers and not both.
6829 *
6830 * The memory page-locked by this function must be unregistered with
6831 * ::cuMemHostUnregister().
6832 *
6833 * \param p - Host pointer to memory to page-lock
6834 * \param bytesize - Size in bytes of the address range to page-lock
6835 * \param Flags - Flags for allocation request
6836 *
6837 * \return
6838 * ::CUDA_SUCCESS,
6839 * ::CUDA_ERROR_DEINITIALIZED,
6840 * ::CUDA_ERROR_NOT_INITIALIZED,
6841 * ::CUDA_ERROR_INVALID_CONTEXT,
6842 * ::CUDA_ERROR_INVALID_VALUE,
6843 * ::CUDA_ERROR_OUT_OF_MEMORY,
6844 * ::CUDA_ERROR_HOST_MEMORY_ALREADY_REGISTERED,
6845 * ::CUDA_ERROR_NOT_PERMITTED,
6846 * ::CUDA_ERROR_NOT_SUPPORTED
6847 * \notefnerr
6848 *
6849 * \sa
6850 * ::cuMemHostUnregister,
6851 * ::cuMemHostGetFlags,
6852 * ::cuMemHostGetDevicePointer,
6853 * ::cudaHostRegister
6854 */
6855CUresult CUDAAPI cuMemHostRegister(void *p, size_t bytesize, unsigned int Flags);
6856
6857/**
6858 * \brief Unregisters a memory range that was registered with cuMemHostRegister.
6859 *
6860 * Unmaps the memory range whose base address is specified by \p p, and makes
6861 * it pageable again.
6862 *
6863 * The base address must be the same one specified to ::cuMemHostRegister().
6864 *
6865 * \param p - Host pointer to memory to unregister
6866 *
6867 * \return
6868 * ::CUDA_SUCCESS,
6869 * ::CUDA_ERROR_DEINITIALIZED,
6870 * ::CUDA_ERROR_NOT_INITIALIZED,
6871 * ::CUDA_ERROR_INVALID_CONTEXT,
6872 * ::CUDA_ERROR_INVALID_VALUE,
6873 * ::CUDA_ERROR_OUT_OF_MEMORY,
6874 * ::CUDA_ERROR_HOST_MEMORY_NOT_REGISTERED,
6875 * \notefnerr
6876 *
6877 * \sa
6878 * ::cuMemHostRegister,
6879 * ::cudaHostUnregister
6880 */
6881CUresult CUDAAPI cuMemHostUnregister(void *p);
6882
6883/**
6884 * \brief Copies memory
6885 *
6886 * Copies data between two pointers.
6887 * \p dst and \p src are base pointers of the destination and source, respectively.
6888 * \p ByteCount specifies the number of bytes to copy.
6889 * Note that this function infers the type of the transfer (host to host, host to
6890 * device, device to device, or device to host) from the pointer values. This
6891 * function is only allowed in contexts which support unified addressing.
6892 *
6893 * \param dst - Destination unified virtual address space pointer
6894 * \param src - Source unified virtual address space pointer
6895 * \param ByteCount - Size of memory copy in bytes
6896 *
6897 * \return
6898 * ::CUDA_SUCCESS,
6899 * ::CUDA_ERROR_DEINITIALIZED,
6900 * ::CUDA_ERROR_NOT_INITIALIZED,
6901 * ::CUDA_ERROR_INVALID_CONTEXT,
6902 * ::CUDA_ERROR_INVALID_VALUE
6903 * \notefnerr
6904 * \note_sync
6905 * \note_memcpy
6906 *
6907 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
6908 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
6909 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
6910 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
6911 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA,
6912 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
6913 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
6914 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
6915 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
6916 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
6917 * ::cudaMemcpy,
6918 * ::cudaMemcpyToSymbol,
6919 * ::cudaMemcpyFromSymbol
6920 */
6921CUresult CUDAAPI cuMemcpy(CUdeviceptr dst, CUdeviceptr src, size_t ByteCount);
6922
6923/**
6924 * \brief Copies device memory between two contexts
6925 *
6926 * Copies from device memory in one context to device memory in another
6927 * context. \p dstDevice is the base device pointer of the destination memory
6928 * and \p dstContext is the destination context. \p srcDevice is the base
6929 * device pointer of the source memory and \p srcContext is the source pointer.
6930 * \p ByteCount specifies the number of bytes to copy.
6931 *
6932 * \param dstDevice - Destination device pointer
6933 * \param dstContext - Destination context
6934 * \param srcDevice - Source device pointer
6935 * \param srcContext - Source context
6936 * \param ByteCount - Size of memory copy in bytes
6937 *
6938 * \return
6939 * ::CUDA_SUCCESS,
6940 * ::CUDA_ERROR_DEINITIALIZED,
6941 * ::CUDA_ERROR_NOT_INITIALIZED,
6942 * ::CUDA_ERROR_INVALID_CONTEXT,
6943 * ::CUDA_ERROR_INVALID_VALUE
6944 * \notefnerr
6945 * \note_sync
6946 *
6947 * \sa ::cuMemcpyDtoD, ::cuMemcpy3DPeer, ::cuMemcpyDtoDAsync, ::cuMemcpyPeerAsync,
6948 * ::cuMemcpy3DPeerAsync,
6949 * ::cudaMemcpyPeer
6950 */
6951CUresult CUDAAPI cuMemcpyPeer(CUdeviceptr dstDevice, CUcontext dstContext, CUdeviceptr srcDevice, CUcontext srcContext, size_t ByteCount);
6952
6953/**
6954 * \brief Copies memory from Host to Device
6955 *
6956 * Copies from host memory to device memory. \p dstDevice and \p srcHost are
6957 * the base addresses of the destination and source, respectively. \p ByteCount
6958 * specifies the number of bytes to copy.
6959 *
6960 * \param dstDevice - Destination device pointer
6961 * \param srcHost - Source host pointer
6962 * \param ByteCount - Size of memory copy in bytes
6963 *
6964 * \return
6965 * ::CUDA_SUCCESS,
6966 * ::CUDA_ERROR_DEINITIALIZED,
6967 * ::CUDA_ERROR_NOT_INITIALIZED,
6968 * ::CUDA_ERROR_INVALID_CONTEXT,
6969 * ::CUDA_ERROR_INVALID_VALUE
6970 * \notefnerr
6971 * \note_sync
6972 * \note_memcpy
6973 *
6974 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
6975 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
6976 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
6977 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
6978 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
6979 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
6980 * ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
6981 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
6982 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
6983 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
6984 * ::cudaMemcpy,
6985 * ::cudaMemcpyToSymbol
6986 */
6987CUresult CUDAAPI cuMemcpyHtoD(CUdeviceptr dstDevice, const void *srcHost, size_t ByteCount);
6988
6989/**
6990 * \brief Copies memory from Device to Host
6991 *
6992 * Copies from device to host memory. \p dstHost and \p srcDevice specify the
6993 * base pointers of the destination and source, respectively. \p ByteCount
6994 * specifies the number of bytes to copy.
6995 *
6996 * \param dstHost - Destination host pointer
6997 * \param srcDevice - Source device pointer
6998 * \param ByteCount - Size of memory copy in bytes
6999 *
7000 * \return
7001 * ::CUDA_SUCCESS,
7002 * ::CUDA_ERROR_DEINITIALIZED,
7003 * ::CUDA_ERROR_NOT_INITIALIZED,
7004 * ::CUDA_ERROR_INVALID_CONTEXT,
7005 * ::CUDA_ERROR_INVALID_VALUE
7006 * \notefnerr
7007 * \note_sync
7008 * \note_memcpy
7009 *
7010 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
7011 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
7012 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
7013 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
7014 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
7015 * ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
7016 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
7017 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
7018 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
7019 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
7020 * ::cudaMemcpy,
7021 * ::cudaMemcpyFromSymbol
7022 */
7023CUresult CUDAAPI cuMemcpyDtoH(void *dstHost, CUdeviceptr srcDevice, size_t ByteCount);
7024
7025/**
7026 * \brief Copies memory from Device to Device
7027 *
7028 * Copies from device memory to device memory. \p dstDevice and \p srcDevice
7029 * are the base pointers of the destination and source, respectively.
7030 * \p ByteCount specifies the number of bytes to copy.
7031 *
7032 * \param dstDevice - Destination device pointer
7033 * \param srcDevice - Source device pointer
7034 * \param ByteCount - Size of memory copy in bytes
7035 *
7036 * \return
7037 * ::CUDA_SUCCESS,
7038 * ::CUDA_ERROR_DEINITIALIZED,
7039 * ::CUDA_ERROR_NOT_INITIALIZED,
7040 * ::CUDA_ERROR_INVALID_CONTEXT,
7041 * ::CUDA_ERROR_INVALID_VALUE
7042 * \notefnerr
7043 * \note_sync
7044 *
7045 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
7046 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
7047 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
7048 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
7049 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA,
7050 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
7051 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
7052 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
7053 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
7054 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
7055 * ::cudaMemcpy,
7056 * ::cudaMemcpyToSymbol,
7057 * ::cudaMemcpyFromSymbol
7058 */
7059CUresult CUDAAPI cuMemcpyDtoD(CUdeviceptr dstDevice, CUdeviceptr srcDevice, size_t ByteCount);
7060
7061/**
7062 * \brief Copies memory from Device to Array
7063 *
7064 * Copies from device memory to a 1D CUDA array. \p dstArray and \p dstOffset
7065 * specify the CUDA array handle and starting index of the destination data.
7066 * \p srcDevice specifies the base pointer of the source. \p ByteCount
7067 * specifies the number of bytes to copy.
7068 *
7069 * \param dstArray - Destination array
7070 * \param dstOffset - Offset in bytes of destination array
7071 * \param srcDevice - Source device pointer
7072 * \param ByteCount - Size of memory copy in bytes
7073 *
7074 * \return
7075 * ::CUDA_SUCCESS,
7076 * ::CUDA_ERROR_DEINITIALIZED,
7077 * ::CUDA_ERROR_NOT_INITIALIZED,
7078 * ::CUDA_ERROR_INVALID_CONTEXT,
7079 * ::CUDA_ERROR_INVALID_VALUE
7080 * \notefnerr
7081 * \note_sync
7082 *
7083 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
7084 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
7085 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
7086 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
7087 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
7088 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
7089 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
7090 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
7091 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
7092 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
7093 * ::cudaMemcpyToArray
7094 */
7095CUresult CUDAAPI cuMemcpyDtoA(CUarray dstArray, size_t dstOffset, CUdeviceptr srcDevice, size_t ByteCount);
7096
7097/**
7098 * \brief Copies memory from Array to Device
7099 *
7100 * Copies from one 1D CUDA array to device memory. \p dstDevice specifies the
7101 * base pointer of the destination and must be naturally aligned with the CUDA
7102 * array elements. \p srcArray and \p srcOffset specify the CUDA array handle
7103 * and the offset in bytes into the array where the copy is to begin.
7104 * \p ByteCount specifies the number of bytes to copy and must be evenly
7105 * divisible by the array element size.
7106 *
7107 * \param dstDevice - Destination device pointer
7108 * \param srcArray - Source array
7109 * \param srcOffset - Offset in bytes of source array
7110 * \param ByteCount - Size of memory copy in bytes
7111 *
7112 * \return
7113 * ::CUDA_SUCCESS,
7114 * ::CUDA_ERROR_DEINITIALIZED,
7115 * ::CUDA_ERROR_NOT_INITIALIZED,
7116 * ::CUDA_ERROR_INVALID_CONTEXT,
7117 * ::CUDA_ERROR_INVALID_VALUE
7118 * \notefnerr
7119 * \note_sync
7120 *
7121 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
7122 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
7123 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
7124 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA,
7125 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
7126 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
7127 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
7128 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
7129 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
7130 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
7131 * ::cudaMemcpyFromArray
7132 */
7133CUresult CUDAAPI cuMemcpyAtoD(CUdeviceptr dstDevice, CUarray srcArray, size_t srcOffset, size_t ByteCount);
7134
7135/**
7136 * \brief Copies memory from Host to Array
7137 *
7138 * Copies from host memory to a 1D CUDA array. \p dstArray and \p dstOffset
7139 * specify the CUDA array handle and starting offset in bytes of the destination
7140 * data. \p pSrc specifies the base address of the source. \p ByteCount specifies
7141 * the number of bytes to copy.
7142 *
7143 * \param dstArray - Destination array
7144 * \param dstOffset - Offset in bytes of destination array
7145 * \param srcHost - Source host pointer
7146 * \param ByteCount - Size of memory copy in bytes
7147 *
7148 * \return
7149 * ::CUDA_SUCCESS,
7150 * ::CUDA_ERROR_DEINITIALIZED,
7151 * ::CUDA_ERROR_NOT_INITIALIZED,
7152 * ::CUDA_ERROR_INVALID_CONTEXT,
7153 * ::CUDA_ERROR_INVALID_VALUE
7154 * \notefnerr
7155 * \note_sync
7156 * \note_memcpy
7157 *
7158 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
7159 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
7160 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
7161 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
7162 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
7163 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoAAsync,
7164 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
7165 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
7166 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
7167 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
7168 * ::cudaMemcpyToArray
7169 */
7170CUresult CUDAAPI cuMemcpyHtoA(CUarray dstArray, size_t dstOffset, const void *srcHost, size_t ByteCount);
7171
7172/**
7173 * \brief Copies memory from Array to Host
7174 *
7175 * Copies from one 1D CUDA array to host memory. \p dstHost specifies the base
7176 * pointer of the destination. \p srcArray and \p srcOffset specify the CUDA
7177 * array handle and starting offset in bytes of the source data.
7178 * \p ByteCount specifies the number of bytes to copy.
7179 *
7180 * \param dstHost - Destination device pointer
7181 * \param srcArray - Source array
7182 * \param srcOffset - Offset in bytes of source array
7183 * \param ByteCount - Size of memory copy in bytes
7184 *
7185 * \return
7186 * ::CUDA_SUCCESS,
7187 * ::CUDA_ERROR_DEINITIALIZED,
7188 * ::CUDA_ERROR_NOT_INITIALIZED,
7189 * ::CUDA_ERROR_INVALID_CONTEXT,
7190 * ::CUDA_ERROR_INVALID_VALUE
7191 * \notefnerr
7192 * \note_sync
7193 * \note_memcpy
7194 *
7195 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
7196 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
7197 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
7198 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
7199 * ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
7200 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
7201 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
7202 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
7203 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
7204 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
7205 * ::cudaMemcpyFromArray
7206 */
7207CUresult CUDAAPI cuMemcpyAtoH(void *dstHost, CUarray srcArray, size_t srcOffset, size_t ByteCount);
7208
7209/**
7210 * \brief Copies memory from Array to Array
7211 *
7212 * Copies from one 1D CUDA array to another. \p dstArray and \p srcArray
7213 * specify the handles of the destination and source CUDA arrays for the copy,
7214 * respectively. \p dstOffset and \p srcOffset specify the destination and
7215 * source offsets in bytes into the CUDA arrays. \p ByteCount is the number of
7216 * bytes to be copied. The size of the elements in the CUDA arrays need not be
7217 * the same format, but the elements must be the same size; and count must be
7218 * evenly divisible by that size.
7219 *
7220 * \param dstArray - Destination array
7221 * \param dstOffset - Offset in bytes of destination array
7222 * \param srcArray - Source array
7223 * \param srcOffset - Offset in bytes of source array
7224 * \param ByteCount - Size of memory copy in bytes
7225 *
7226 * \return
7227 * ::CUDA_SUCCESS,
7228 * ::CUDA_ERROR_DEINITIALIZED,
7229 * ::CUDA_ERROR_NOT_INITIALIZED,
7230 * ::CUDA_ERROR_INVALID_CONTEXT,
7231 * ::CUDA_ERROR_INVALID_VALUE
7232 * \notefnerr
7233 * \note_sync
7234 *
7235 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
7236 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
7237 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
7238 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoD,
7239 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
7240 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
7241 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
7242 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
7243 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
7244 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
7245 * ::cudaMemcpyArrayToArray
7246 */
7247CUresult CUDAAPI cuMemcpyAtoA(CUarray dstArray, size_t dstOffset, CUarray srcArray, size_t srcOffset, size_t ByteCount);
7248
7249/**
7250 * \brief Copies memory for 2D arrays
7251 *
7252 * Perform a 2D memory copy according to the parameters specified in \p pCopy.
7253 * The ::CUDA_MEMCPY2D structure is defined as:
7254 *
7255 * \code
7256 typedef struct CUDA_MEMCPY2D_st {
7257 unsigned int srcXInBytes, srcY;
7258 CUmemorytype srcMemoryType;
7259 const void *srcHost;
7260 CUdeviceptr srcDevice;
7261 CUarray srcArray;
7262 unsigned int srcPitch;
7263
7264 unsigned int dstXInBytes, dstY;
7265 CUmemorytype dstMemoryType;
7266 void *dstHost;
7267 CUdeviceptr dstDevice;
7268 CUarray dstArray;
7269 unsigned int dstPitch;
7270
7271 unsigned int WidthInBytes;
7272 unsigned int Height;
7273 } CUDA_MEMCPY2D;
7274 * \endcode
7275 * where:
7276 * - ::srcMemoryType and ::dstMemoryType specify the type of memory of the
7277 * source and destination, respectively; ::CUmemorytype_enum is defined as:
7278 *
7279 * \code
7280 typedef enum CUmemorytype_enum {
7281 CU_MEMORYTYPE_HOST = 0x01,
7282 CU_MEMORYTYPE_DEVICE = 0x02,
7283 CU_MEMORYTYPE_ARRAY = 0x03,
7284 CU_MEMORYTYPE_UNIFIED = 0x04
7285 } CUmemorytype;
7286 * \endcode
7287 *
7288 * \par
7289 * If ::srcMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::srcDevice and ::srcPitch
7290 * specify the (unified virtual address space) base address of the source data
7291 * and the bytes per row to apply. ::srcArray is ignored.
7292 * This value may be used only if unified addressing is supported in the calling
7293 * context.
7294 *
7295 * \par
7296 * If ::srcMemoryType is ::CU_MEMORYTYPE_HOST, ::srcHost and ::srcPitch
7297 * specify the (host) base address of the source data and the bytes per row to
7298 * apply. ::srcArray is ignored.
7299 *
7300 * \par
7301 * If ::srcMemoryType is ::CU_MEMORYTYPE_DEVICE, ::srcDevice and ::srcPitch
7302 * specify the (device) base address of the source data and the bytes per row
7303 * to apply. ::srcArray is ignored.
7304 *
7305 * \par
7306 * If ::srcMemoryType is ::CU_MEMORYTYPE_ARRAY, ::srcArray specifies the
7307 * handle of the source data. ::srcHost, ::srcDevice and ::srcPitch are
7308 * ignored.
7309 *
7310 * \par
7311 * If ::dstMemoryType is ::CU_MEMORYTYPE_HOST, ::dstHost and ::dstPitch
7312 * specify the (host) base address of the destination data and the bytes per
7313 * row to apply. ::dstArray is ignored.
7314 *
7315 * \par
7316 * If ::dstMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::dstDevice and ::dstPitch
7317 * specify the (unified virtual address space) base address of the source data
7318 * and the bytes per row to apply. ::dstArray is ignored.
7319 * This value may be used only if unified addressing is supported in the calling
7320 * context.
7321 *
7322 * \par
7323 * If ::dstMemoryType is ::CU_MEMORYTYPE_DEVICE, ::dstDevice and ::dstPitch
7324 * specify the (device) base address of the destination data and the bytes per
7325 * row to apply. ::dstArray is ignored.
7326 *
7327 * \par
7328 * If ::dstMemoryType is ::CU_MEMORYTYPE_ARRAY, ::dstArray specifies the
7329 * handle of the destination data. ::dstHost, ::dstDevice and ::dstPitch are
7330 * ignored.
7331 *
7332 * - ::srcXInBytes and ::srcY specify the base address of the source data for
7333 * the copy.
7334 *
7335 * \par
7336 * For host pointers, the starting address is
7337 * \code
7338 void* Start = (void*)((char*)srcHost+srcY*srcPitch + srcXInBytes);
7339 * \endcode
7340 *
7341 * \par
7342 * For device pointers, the starting address is
7343 * \code
7344 CUdeviceptr Start = srcDevice+srcY*srcPitch+srcXInBytes;
7345 * \endcode
7346 *
7347 * \par
7348 * For CUDA arrays, ::srcXInBytes must be evenly divisible by the array
7349 * element size.
7350 *
7351 * - ::dstXInBytes and ::dstY specify the base address of the destination data
7352 * for the copy.
7353 *
7354 * \par
7355 * For host pointers, the base address is
7356 * \code
7357 void* dstStart = (void*)((char*)dstHost+dstY*dstPitch + dstXInBytes);
7358 * \endcode
7359 *
7360 * \par
7361 * For device pointers, the starting address is
7362 * \code
7363 CUdeviceptr dstStart = dstDevice+dstY*dstPitch+dstXInBytes;
7364 * \endcode
7365 *
7366 * \par
7367 * For CUDA arrays, ::dstXInBytes must be evenly divisible by the array
7368 * element size.
7369 *
7370 * - ::WidthInBytes and ::Height specify the width (in bytes) and height of
7371 * the 2D copy being performed.
7372 * - If specified, ::srcPitch must be greater than or equal to ::WidthInBytes +
7373 * ::srcXInBytes, and ::dstPitch must be greater than or equal to
7374 * ::WidthInBytes + dstXInBytes.
7375 *
7376 * \par
7377 * ::cuMemcpy2D() returns an error if any pitch is greater than the maximum
7378 * allowed (::CU_DEVICE_ATTRIBUTE_MAX_PITCH). ::cuMemAllocPitch() passes back
7379 * pitches that always work with ::cuMemcpy2D(). On intra-device memory copies
7380 * (device to device, CUDA array to device, CUDA array to CUDA array),
7381 * ::cuMemcpy2D() may fail for pitches not computed by ::cuMemAllocPitch().
7382 * ::cuMemcpy2DUnaligned() does not have this restriction, but may run
7383 * significantly slower in the cases where ::cuMemcpy2D() would have returned
7384 * an error code.
7385 *
7386 * \param pCopy - Parameters for the memory copy
7387 *
7388 * \return
7389 * ::CUDA_SUCCESS,
7390 * ::CUDA_ERROR_DEINITIALIZED,
7391 * ::CUDA_ERROR_NOT_INITIALIZED,
7392 * ::CUDA_ERROR_INVALID_CONTEXT,
7393 * ::CUDA_ERROR_INVALID_VALUE
7394 * \notefnerr
7395 * \note_sync
7396 *
7397 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
7398 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
7399 * ::cuMemAllocPitch, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
7400 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
7401 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
7402 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
7403 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
7404 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
7405 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
7406 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
7407 * ::cudaMemcpy2D,
7408 * ::cudaMemcpy2DToArray,
7409 * ::cudaMemcpy2DFromArray
7410 */
7411CUresult CUDAAPI cuMemcpy2D(const CUDA_MEMCPY2D *pCopy);
7412
7413/**
7414 * \brief Copies memory for 2D arrays
7415 *
7416 * Perform a 2D memory copy according to the parameters specified in \p pCopy.
7417 * The ::CUDA_MEMCPY2D structure is defined as:
7418 *
7419 * \code
7420 typedef struct CUDA_MEMCPY2D_st {
7421 unsigned int srcXInBytes, srcY;
7422 CUmemorytype srcMemoryType;
7423 const void *srcHost;
7424 CUdeviceptr srcDevice;
7425 CUarray srcArray;
7426 unsigned int srcPitch;
7427 unsigned int dstXInBytes, dstY;
7428 CUmemorytype dstMemoryType;
7429 void *dstHost;
7430 CUdeviceptr dstDevice;
7431 CUarray dstArray;
7432 unsigned int dstPitch;
7433 unsigned int WidthInBytes;
7434 unsigned int Height;
7435 } CUDA_MEMCPY2D;
7436 * \endcode
7437 * where:
7438 * - ::srcMemoryType and ::dstMemoryType specify the type of memory of the
7439 * source and destination, respectively; ::CUmemorytype_enum is defined as:
7440 *
7441 * \code
7442 typedef enum CUmemorytype_enum {
7443 CU_MEMORYTYPE_HOST = 0x01,
7444 CU_MEMORYTYPE_DEVICE = 0x02,
7445 CU_MEMORYTYPE_ARRAY = 0x03,
7446 CU_MEMORYTYPE_UNIFIED = 0x04
7447 } CUmemorytype;
7448 * \endcode
7449 *
7450 * \par
7451 * If ::srcMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::srcDevice and ::srcPitch
7452 * specify the (unified virtual address space) base address of the source data
7453 * and the bytes per row to apply. ::srcArray is ignored.
7454 * This value may be used only if unified addressing is supported in the calling
7455 * context.
7456 *
7457 * \par
7458 * If ::srcMemoryType is ::CU_MEMORYTYPE_HOST, ::srcHost and ::srcPitch
7459 * specify the (host) base address of the source data and the bytes per row to
7460 * apply. ::srcArray is ignored.
7461 *
7462 * \par
7463 * If ::srcMemoryType is ::CU_MEMORYTYPE_DEVICE, ::srcDevice and ::srcPitch
7464 * specify the (device) base address of the source data and the bytes per row
7465 * to apply. ::srcArray is ignored.
7466 *
7467 * \par
7468 * If ::srcMemoryType is ::CU_MEMORYTYPE_ARRAY, ::srcArray specifies the
7469 * handle of the source data. ::srcHost, ::srcDevice and ::srcPitch are
7470 * ignored.
7471 *
7472 * \par
7473 * If ::dstMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::dstDevice and ::dstPitch
7474 * specify the (unified virtual address space) base address of the source data
7475 * and the bytes per row to apply. ::dstArray is ignored.
7476 * This value may be used only if unified addressing is supported in the calling
7477 * context.
7478 *
7479 * \par
7480 * If ::dstMemoryType is ::CU_MEMORYTYPE_HOST, ::dstHost and ::dstPitch
7481 * specify the (host) base address of the destination data and the bytes per
7482 * row to apply. ::dstArray is ignored.
7483 *
7484 * \par
7485 * If ::dstMemoryType is ::CU_MEMORYTYPE_DEVICE, ::dstDevice and ::dstPitch
7486 * specify the (device) base address of the destination data and the bytes per
7487 * row to apply. ::dstArray is ignored.
7488 *
7489 * \par
7490 * If ::dstMemoryType is ::CU_MEMORYTYPE_ARRAY, ::dstArray specifies the
7491 * handle of the destination data. ::dstHost, ::dstDevice and ::dstPitch are
7492 * ignored.
7493 *
7494 * - ::srcXInBytes and ::srcY specify the base address of the source data for
7495 * the copy.
7496 *
7497 * \par
7498 * For host pointers, the starting address is
7499 * \code
7500 void* Start = (void*)((char*)srcHost+srcY*srcPitch + srcXInBytes);
7501 * \endcode
7502 *
7503 * \par
7504 * For device pointers, the starting address is
7505 * \code
7506 CUdeviceptr Start = srcDevice+srcY*srcPitch+srcXInBytes;
7507 * \endcode
7508 *
7509 * \par
7510 * For CUDA arrays, ::srcXInBytes must be evenly divisible by the array
7511 * element size.
7512 *
7513 * - ::dstXInBytes and ::dstY specify the base address of the destination data
7514 * for the copy.
7515 *
7516 * \par
7517 * For host pointers, the base address is
7518 * \code
7519 void* dstStart = (void*)((char*)dstHost+dstY*dstPitch + dstXInBytes);
7520 * \endcode
7521 *
7522 * \par
7523 * For device pointers, the starting address is
7524 * \code
7525 CUdeviceptr dstStart = dstDevice+dstY*dstPitch+dstXInBytes;
7526 * \endcode
7527 *
7528 * \par
7529 * For CUDA arrays, ::dstXInBytes must be evenly divisible by the array
7530 * element size.
7531 *
7532 * - ::WidthInBytes and ::Height specify the width (in bytes) and height of
7533 * the 2D copy being performed.
7534 * - If specified, ::srcPitch must be greater than or equal to ::WidthInBytes +
7535 * ::srcXInBytes, and ::dstPitch must be greater than or equal to
7536 * ::WidthInBytes + dstXInBytes.
7537 *
7538 * \par
7539 * ::cuMemcpy2D() returns an error if any pitch is greater than the maximum
7540 * allowed (::CU_DEVICE_ATTRIBUTE_MAX_PITCH). ::cuMemAllocPitch() passes back
7541 * pitches that always work with ::cuMemcpy2D(). On intra-device memory copies
7542 * (device to device, CUDA array to device, CUDA array to CUDA array),
7543 * ::cuMemcpy2D() may fail for pitches not computed by ::cuMemAllocPitch().
7544 * ::cuMemcpy2DUnaligned() does not have this restriction, but may run
7545 * significantly slower in the cases where ::cuMemcpy2D() would have returned
7546 * an error code.
7547 *
7548 * \param pCopy - Parameters for the memory copy
7549 *
7550 * \return
7551 * ::CUDA_SUCCESS,
7552 * ::CUDA_ERROR_DEINITIALIZED,
7553 * ::CUDA_ERROR_NOT_INITIALIZED,
7554 * ::CUDA_ERROR_INVALID_CONTEXT,
7555 * ::CUDA_ERROR_INVALID_VALUE
7556 * \notefnerr
7557 * \note_sync
7558 *
7559 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
7560 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
7561 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync,
7562 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
7563 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
7564 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
7565 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
7566 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
7567 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
7568 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
7569 * ::cudaMemcpy2D,
7570 * ::cudaMemcpy2DToArray,
7571 * ::cudaMemcpy2DFromArray
7572 */
7573CUresult CUDAAPI cuMemcpy2DUnaligned(const CUDA_MEMCPY2D *pCopy);
7574
7575/**
7576 * \brief Copies memory for 3D arrays
7577 *
7578 * Perform a 3D memory copy according to the parameters specified in
7579 * \p pCopy. The ::CUDA_MEMCPY3D structure is defined as:
7580 *
7581 * \code
7582 typedef struct CUDA_MEMCPY3D_st {
7583
7584 unsigned int srcXInBytes, srcY, srcZ;
7585 unsigned int srcLOD;
7586 CUmemorytype srcMemoryType;
7587 const void *srcHost;
7588 CUdeviceptr srcDevice;
7589 CUarray srcArray;
7590 unsigned int srcPitch; // ignored when src is array
7591 unsigned int srcHeight; // ignored when src is array; may be 0 if Depth==1
7592
7593 unsigned int dstXInBytes, dstY, dstZ;
7594 unsigned int dstLOD;
7595 CUmemorytype dstMemoryType;
7596 void *dstHost;
7597 CUdeviceptr dstDevice;
7598 CUarray dstArray;
7599 unsigned int dstPitch; // ignored when dst is array
7600 unsigned int dstHeight; // ignored when dst is array; may be 0 if Depth==1
7601
7602 unsigned int WidthInBytes;
7603 unsigned int Height;
7604 unsigned int Depth;
7605 } CUDA_MEMCPY3D;
7606 * \endcode
7607 * where:
7608 * - ::srcMemoryType and ::dstMemoryType specify the type of memory of the
7609 * source and destination, respectively; ::CUmemorytype_enum is defined as:
7610 *
7611 * \code
7612 typedef enum CUmemorytype_enum {
7613 CU_MEMORYTYPE_HOST = 0x01,
7614 CU_MEMORYTYPE_DEVICE = 0x02,
7615 CU_MEMORYTYPE_ARRAY = 0x03,
7616 CU_MEMORYTYPE_UNIFIED = 0x04
7617 } CUmemorytype;
7618 * \endcode
7619 *
7620 * \par
7621 * If ::srcMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::srcDevice and ::srcPitch
7622 * specify the (unified virtual address space) base address of the source data
7623 * and the bytes per row to apply. ::srcArray is ignored.
7624 * This value may be used only if unified addressing is supported in the calling
7625 * context.
7626 *
7627 * \par
7628 * If ::srcMemoryType is ::CU_MEMORYTYPE_HOST, ::srcHost, ::srcPitch and
7629 * ::srcHeight specify the (host) base address of the source data, the bytes
7630 * per row, and the height of each 2D slice of the 3D array. ::srcArray is
7631 * ignored.
7632 *
7633 * \par
7634 * If ::srcMemoryType is ::CU_MEMORYTYPE_DEVICE, ::srcDevice, ::srcPitch and
7635 * ::srcHeight specify the (device) base address of the source data, the bytes
7636 * per row, and the height of each 2D slice of the 3D array. ::srcArray is
7637 * ignored.
7638 *
7639 * \par
7640 * If ::srcMemoryType is ::CU_MEMORYTYPE_ARRAY, ::srcArray specifies the
7641 * handle of the source data. ::srcHost, ::srcDevice, ::srcPitch and
7642 * ::srcHeight are ignored.
7643 *
7644 * \par
7645 * If ::dstMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::dstDevice and ::dstPitch
7646 * specify the (unified virtual address space) base address of the source data
7647 * and the bytes per row to apply. ::dstArray is ignored.
7648 * This value may be used only if unified addressing is supported in the calling
7649 * context.
7650 *
7651 * \par
7652 * If ::dstMemoryType is ::CU_MEMORYTYPE_HOST, ::dstHost and ::dstPitch
7653 * specify the (host) base address of the destination data, the bytes per row,
7654 * and the height of each 2D slice of the 3D array. ::dstArray is ignored.
7655 *
7656 * \par
7657 * If ::dstMemoryType is ::CU_MEMORYTYPE_DEVICE, ::dstDevice and ::dstPitch
7658 * specify the (device) base address of the destination data, the bytes per
7659 * row, and the height of each 2D slice of the 3D array. ::dstArray is ignored.
7660 *
7661 * \par
7662 * If ::dstMemoryType is ::CU_MEMORYTYPE_ARRAY, ::dstArray specifies the
7663 * handle of the destination data. ::dstHost, ::dstDevice, ::dstPitch and
7664 * ::dstHeight are ignored.
7665 *
7666 * - ::srcXInBytes, ::srcY and ::srcZ specify the base address of the source
7667 * data for the copy.
7668 *
7669 * \par
7670 * For host pointers, the starting address is
7671 * \code
7672 void* Start = (void*)((char*)srcHost+(srcZ*srcHeight+srcY)*srcPitch + srcXInBytes);
7673 * \endcode
7674 *
7675 * \par
7676 * For device pointers, the starting address is
7677 * \code
7678 CUdeviceptr Start = srcDevice+(srcZ*srcHeight+srcY)*srcPitch+srcXInBytes;
7679 * \endcode
7680 *
7681 * \par
7682 * For CUDA arrays, ::srcXInBytes must be evenly divisible by the array
7683 * element size.
7684 *
7685 * - dstXInBytes, ::dstY and ::dstZ specify the base address of the
7686 * destination data for the copy.
7687 *
7688 * \par
7689 * For host pointers, the base address is
7690 * \code
7691 void* dstStart = (void*)((char*)dstHost+(dstZ*dstHeight+dstY)*dstPitch + dstXInBytes);
7692 * \endcode
7693 *
7694 * \par
7695 * For device pointers, the starting address is
7696 * \code
7697 CUdeviceptr dstStart = dstDevice+(dstZ*dstHeight+dstY)*dstPitch+dstXInBytes;
7698 * \endcode
7699 *
7700 * \par
7701 * For CUDA arrays, ::dstXInBytes must be evenly divisible by the array
7702 * element size.
7703 *
7704 * - ::WidthInBytes, ::Height and ::Depth specify the width (in bytes), height
7705 * and depth of the 3D copy being performed.
7706 * - If specified, ::srcPitch must be greater than or equal to ::WidthInBytes +
7707 * ::srcXInBytes, and ::dstPitch must be greater than or equal to
7708 * ::WidthInBytes + dstXInBytes.
7709 * - If specified, ::srcHeight must be greater than or equal to ::Height +
7710 * ::srcY, and ::dstHeight must be greater than or equal to ::Height + ::dstY.
7711 *
7712 * \par
7713 * ::cuMemcpy3D() returns an error if any pitch is greater than the maximum
7714 * allowed (::CU_DEVICE_ATTRIBUTE_MAX_PITCH).
7715 *
7716 * The ::srcLOD and ::dstLOD members of the ::CUDA_MEMCPY3D structure must be
7717 * set to 0.
7718 *
7719 * \param pCopy - Parameters for the memory copy
7720 *
7721 * \return
7722 * ::CUDA_SUCCESS,
7723 * ::CUDA_ERROR_DEINITIALIZED,
7724 * ::CUDA_ERROR_NOT_INITIALIZED,
7725 * ::CUDA_ERROR_INVALID_CONTEXT,
7726 * ::CUDA_ERROR_INVALID_VALUE
7727 * \notefnerr
7728 * \note_sync
7729 *
7730 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
7731 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
7732 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
7733 * ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
7734 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
7735 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
7736 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
7737 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
7738 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
7739 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
7740 * ::cudaMemcpy3D
7741 */
7742CUresult CUDAAPI cuMemcpy3D(const CUDA_MEMCPY3D *pCopy);
7743
7744/**
7745 * \brief Copies memory between contexts
7746 *
7747 * Perform a 3D memory copy according to the parameters specified in
7748 * \p pCopy. See the definition of the ::CUDA_MEMCPY3D_PEER structure
7749 * for documentation of its parameters.
7750 *
7751 * \param pCopy - Parameters for the memory copy
7752 *
7753 * \return
7754 * ::CUDA_SUCCESS,
7755 * ::CUDA_ERROR_DEINITIALIZED,
7756 * ::CUDA_ERROR_NOT_INITIALIZED,
7757 * ::CUDA_ERROR_INVALID_CONTEXT,
7758 * ::CUDA_ERROR_INVALID_VALUE
7759 * \notefnerr
7760 * \note_sync
7761 *
7762 * \sa ::cuMemcpyDtoD, ::cuMemcpyPeer, ::cuMemcpyDtoDAsync, ::cuMemcpyPeerAsync,
7763 * ::cuMemcpy3DPeerAsync,
7764 * ::cudaMemcpy3DPeer
7765 */
7766CUresult CUDAAPI cuMemcpy3DPeer(const CUDA_MEMCPY3D_PEER *pCopy);
7767
7768/**
7769 * \brief Copies memory asynchronously
7770 *
7771 * Copies data between two pointers.
7772 * \p dst and \p src are base pointers of the destination and source, respectively.
7773 * \p ByteCount specifies the number of bytes to copy.
7774 * Note that this function infers the type of the transfer (host to host, host to
7775 * device, device to device, or device to host) from the pointer values. This
7776 * function is only allowed in contexts which support unified addressing.
7777 *
7778 * \param dst - Destination unified virtual address space pointer
7779 * \param src - Source unified virtual address space pointer
7780 * \param ByteCount - Size of memory copy in bytes
7781 * \param hStream - Stream identifier
7782 *
7783 * \return
7784 * ::CUDA_SUCCESS,
7785 * ::CUDA_ERROR_DEINITIALIZED,
7786 * ::CUDA_ERROR_NOT_INITIALIZED,
7787 * ::CUDA_ERROR_INVALID_CONTEXT,
7788 * ::CUDA_ERROR_INVALID_VALUE,
7789 * ::CUDA_ERROR_INVALID_HANDLE
7790 * \notefnerr
7791 * \note_async
7792 * \note_null_stream
7793 * \note_memcpy
7794 *
7795 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
7796 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
7797 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
7798 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
7799 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD,
7800 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
7801 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
7802 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
7803 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async,
7804 * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async,
7805 * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async,
7806 * ::cuMemsetD32, ::cuMemsetD32Async,
7807 * ::cudaMemcpyAsync,
7808 * ::cudaMemcpyToSymbolAsync,
7809 * ::cudaMemcpyFromSymbolAsync
7810 */
7811CUresult CUDAAPI cuMemcpyAsync(CUdeviceptr dst, CUdeviceptr src, size_t ByteCount, CUstream hStream);
7812
7813/**
7814 * \brief Copies device memory between two contexts asynchronously.
7815 *
7816 * Copies from device memory in one context to device memory in another
7817 * context. \p dstDevice is the base device pointer of the destination memory
7818 * and \p dstContext is the destination context. \p srcDevice is the base
7819 * device pointer of the source memory and \p srcContext is the source pointer.
7820 * \p ByteCount specifies the number of bytes to copy.
7821 *
7822 * \param dstDevice - Destination device pointer
7823 * \param dstContext - Destination context
7824 * \param srcDevice - Source device pointer
7825 * \param srcContext - Source context
7826 * \param ByteCount - Size of memory copy in bytes
7827 * \param hStream - Stream identifier
7828 *
7829 * \return
7830 * ::CUDA_SUCCESS,
7831 * ::CUDA_ERROR_DEINITIALIZED,
7832 * ::CUDA_ERROR_NOT_INITIALIZED,
7833 * ::CUDA_ERROR_INVALID_CONTEXT,
7834 * ::CUDA_ERROR_INVALID_VALUE,
7835 * ::CUDA_ERROR_INVALID_HANDLE
7836 * \notefnerr
7837 * \note_async
7838 * \note_null_stream
7839 *
7840 * \sa ::cuMemcpyDtoD, ::cuMemcpyPeer, ::cuMemcpy3DPeer, ::cuMemcpyDtoDAsync,
7841 * ::cuMemcpy3DPeerAsync,
7842 * ::cudaMemcpyPeerAsync
7843 */
7844CUresult CUDAAPI cuMemcpyPeerAsync(CUdeviceptr dstDevice, CUcontext dstContext, CUdeviceptr srcDevice, CUcontext srcContext, size_t ByteCount, CUstream hStream);
7845
7846/**
7847 * \brief Copies memory from Host to Device
7848 *
7849 * Copies from host memory to device memory. \p dstDevice and \p srcHost are
7850 * the base addresses of the destination and source, respectively. \p ByteCount
7851 * specifies the number of bytes to copy.
7852 *
7853 * \param dstDevice - Destination device pointer
7854 * \param srcHost - Source host pointer
7855 * \param ByteCount - Size of memory copy in bytes
7856 * \param hStream - Stream identifier
7857 *
7858 * \return
7859 * ::CUDA_SUCCESS,
7860 * ::CUDA_ERROR_DEINITIALIZED,
7861 * ::CUDA_ERROR_NOT_INITIALIZED,
7862 * ::CUDA_ERROR_INVALID_CONTEXT,
7863 * ::CUDA_ERROR_INVALID_VALUE,
7864 * ::CUDA_ERROR_INVALID_HANDLE
7865 * \notefnerr
7866 * \note_async
7867 * \note_null_stream
7868 * \note_memcpy
7869 *
7870 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
7871 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
7872 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
7873 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
7874 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
7875 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
7876 * ::cuMemcpyHtoD, ::cuMemFree, ::cuMemFreeHost,
7877 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
7878 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async,
7879 * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async,
7880 * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async,
7881 * ::cuMemsetD32, ::cuMemsetD32Async,
7882 * ::cudaMemcpyAsync,
7883 * ::cudaMemcpyToSymbolAsync
7884 */
7885CUresult CUDAAPI cuMemcpyHtoDAsync(CUdeviceptr dstDevice, const void *srcHost, size_t ByteCount, CUstream hStream);
7886
7887/**
7888 * \brief Copies memory from Device to Host
7889 *
7890 * Copies from device to host memory. \p dstHost and \p srcDevice specify the
7891 * base pointers of the destination and source, respectively. \p ByteCount
7892 * specifies the number of bytes to copy.
7893 *
7894 * \param dstHost - Destination host pointer
7895 * \param srcDevice - Source device pointer
7896 * \param ByteCount - Size of memory copy in bytes
7897 * \param hStream - Stream identifier
7898 *
7899 * \return
7900 * ::CUDA_SUCCESS,
7901 * ::CUDA_ERROR_DEINITIALIZED,
7902 * ::CUDA_ERROR_NOT_INITIALIZED,
7903 * ::CUDA_ERROR_INVALID_CONTEXT,
7904 * ::CUDA_ERROR_INVALID_VALUE,
7905 * ::CUDA_ERROR_INVALID_HANDLE
7906 * \notefnerr
7907 * \note_async
7908 * \note_null_stream
7909 * \note_memcpy
7910 *
7911 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
7912 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
7913 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
7914 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
7915 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
7916 * ::cuMemcpyDtoH, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
7917 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
7918 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
7919 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async,
7920 * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async,
7921 * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async,
7922 * ::cuMemsetD32, ::cuMemsetD32Async,
7923 * ::cudaMemcpyAsync,
7924 * ::cudaMemcpyFromSymbolAsync
7925 */
7926CUresult CUDAAPI cuMemcpyDtoHAsync(void *dstHost, CUdeviceptr srcDevice, size_t ByteCount, CUstream hStream);
7927
7928/**
7929 * \brief Copies memory from Device to Device
7930 *
7931 * Copies from device memory to device memory. \p dstDevice and \p srcDevice
7932 * are the base pointers of the destination and source, respectively.
7933 * \p ByteCount specifies the number of bytes to copy.
7934 *
7935 * \param dstDevice - Destination device pointer
7936 * \param srcDevice - Source device pointer
7937 * \param ByteCount - Size of memory copy in bytes
7938 * \param hStream - Stream identifier
7939 *
7940 * \return
7941 * ::CUDA_SUCCESS,
7942 * ::CUDA_ERROR_DEINITIALIZED,
7943 * ::CUDA_ERROR_NOT_INITIALIZED,
7944 * ::CUDA_ERROR_INVALID_CONTEXT,
7945 * ::CUDA_ERROR_INVALID_VALUE,
7946 * ::CUDA_ERROR_INVALID_HANDLE
7947 * \notefnerr
7948 * \note_async
7949 * \note_null_stream
7950 *
7951 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
7952 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
7953 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
7954 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
7955 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD,
7956 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
7957 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
7958 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
7959 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async,
7960 * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async,
7961 * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async,
7962 * ::cuMemsetD32, ::cuMemsetD32Async,
7963 * ::cudaMemcpyAsync,
7964 * ::cudaMemcpyToSymbolAsync,
7965 * ::cudaMemcpyFromSymbolAsync
7966 */
7967CUresult CUDAAPI cuMemcpyDtoDAsync(CUdeviceptr dstDevice, CUdeviceptr srcDevice, size_t ByteCount, CUstream hStream);
7968
7969/**
7970 * \brief Copies memory from Host to Array
7971 *
7972 * Copies from host memory to a 1D CUDA array. \p dstArray and \p dstOffset
7973 * specify the CUDA array handle and starting offset in bytes of the
7974 * destination data. \p srcHost specifies the base address of the source.
7975 * \p ByteCount specifies the number of bytes to copy.
7976 *
7977 * \param dstArray - Destination array
7978 * \param dstOffset - Offset in bytes of destination array
7979 * \param srcHost - Source host pointer
7980 * \param ByteCount - Size of memory copy in bytes
7981 * \param hStream - Stream identifier
7982 *
7983 * \return
7984 * ::CUDA_SUCCESS,
7985 * ::CUDA_ERROR_DEINITIALIZED,
7986 * ::CUDA_ERROR_NOT_INITIALIZED,
7987 * ::CUDA_ERROR_INVALID_CONTEXT,
7988 * ::CUDA_ERROR_INVALID_VALUE,
7989 * ::CUDA_ERROR_INVALID_HANDLE
7990 * \notefnerr
7991 * \note_async
7992 * \note_null_stream
7993 * \note_memcpy
7994 *
7995 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
7996 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
7997 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
7998 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
7999 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
8000 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA,
8001 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
8002 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
8003 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async,
8004 * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async,
8005 * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async,
8006 * ::cuMemsetD32, ::cuMemsetD32Async,
8007 * ::cudaMemcpyToArrayAsync
8008 */
8009CUresult CUDAAPI cuMemcpyHtoAAsync(CUarray dstArray, size_t dstOffset, const void *srcHost, size_t ByteCount, CUstream hStream);
8010
8011/**
8012 * \brief Copies memory from Array to Host
8013 *
8014 * Copies from one 1D CUDA array to host memory. \p dstHost specifies the base
8015 * pointer of the destination. \p srcArray and \p srcOffset specify the CUDA
8016 * array handle and starting offset in bytes of the source data.
8017 * \p ByteCount specifies the number of bytes to copy.
8018 *
8019 * \param dstHost - Destination pointer
8020 * \param srcArray - Source array
8021 * \param srcOffset - Offset in bytes of source array
8022 * \param ByteCount - Size of memory copy in bytes
8023 * \param hStream - Stream identifier
8024 *
8025 * \return
8026 * ::CUDA_SUCCESS,
8027 * ::CUDA_ERROR_DEINITIALIZED,
8028 * ::CUDA_ERROR_NOT_INITIALIZED,
8029 * ::CUDA_ERROR_INVALID_CONTEXT,
8030 * ::CUDA_ERROR_INVALID_VALUE,
8031 * ::CUDA_ERROR_INVALID_HANDLE
8032 * \notefnerr
8033 * \note_async
8034 * \note_null_stream
8035 * \note_memcpy
8036 *
8037 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
8038 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
8039 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
8040 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
8041 * ::cuMemcpyAtoH, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
8042 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
8043 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
8044 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
8045 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async,
8046 * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async,
8047 * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async,
8048 * ::cuMemsetD32, ::cuMemsetD32Async,
8049 * ::cudaMemcpyFromArrayAsync
8050 */
8051CUresult CUDAAPI cuMemcpyAtoHAsync(void *dstHost, CUarray srcArray, size_t srcOffset, size_t ByteCount, CUstream hStream);
8052
8053/**
8054 * \brief Copies memory for 2D arrays
8055 *
8056 * Perform a 2D memory copy according to the parameters specified in \p pCopy.
8057 * The ::CUDA_MEMCPY2D structure is defined as:
8058 *
8059 * \code
8060 typedef struct CUDA_MEMCPY2D_st {
8061 unsigned int srcXInBytes, srcY;
8062 CUmemorytype srcMemoryType;
8063 const void *srcHost;
8064 CUdeviceptr srcDevice;
8065 CUarray srcArray;
8066 unsigned int srcPitch;
8067 unsigned int dstXInBytes, dstY;
8068 CUmemorytype dstMemoryType;
8069 void *dstHost;
8070 CUdeviceptr dstDevice;
8071 CUarray dstArray;
8072 unsigned int dstPitch;
8073 unsigned int WidthInBytes;
8074 unsigned int Height;
8075 } CUDA_MEMCPY2D;
8076 * \endcode
8077 * where:
8078 * - ::srcMemoryType and ::dstMemoryType specify the type of memory of the
8079 * source and destination, respectively; ::CUmemorytype_enum is defined as:
8080 *
8081 * \code
8082 typedef enum CUmemorytype_enum {
8083 CU_MEMORYTYPE_HOST = 0x01,
8084 CU_MEMORYTYPE_DEVICE = 0x02,
8085 CU_MEMORYTYPE_ARRAY = 0x03,
8086 CU_MEMORYTYPE_UNIFIED = 0x04
8087 } CUmemorytype;
8088 * \endcode
8089 *
8090 * \par
8091 * If ::srcMemoryType is ::CU_MEMORYTYPE_HOST, ::srcHost and ::srcPitch
8092 * specify the (host) base address of the source data and the bytes per row to
8093 * apply. ::srcArray is ignored.
8094 *
8095 * \par
8096 * If ::srcMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::srcDevice and ::srcPitch
8097 * specify the (unified virtual address space) base address of the source data
8098 * and the bytes per row to apply. ::srcArray is ignored.
8099 * This value may be used only if unified addressing is supported in the calling
8100 * context.
8101 *
8102 * \par
8103 * If ::srcMemoryType is ::CU_MEMORYTYPE_DEVICE, ::srcDevice and ::srcPitch
8104 * specify the (device) base address of the source data and the bytes per row
8105 * to apply. ::srcArray is ignored.
8106 *
8107 * \par
8108 * If ::srcMemoryType is ::CU_MEMORYTYPE_ARRAY, ::srcArray specifies the
8109 * handle of the source data. ::srcHost, ::srcDevice and ::srcPitch are
8110 * ignored.
8111 *
8112 * \par
8113 * If ::dstMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::dstDevice and ::dstPitch
8114 * specify the (unified virtual address space) base address of the source data
8115 * and the bytes per row to apply. ::dstArray is ignored.
8116 * This value may be used only if unified addressing is supported in the calling
8117 * context.
8118 *
8119 * \par
8120 * If ::dstMemoryType is ::CU_MEMORYTYPE_HOST, ::dstHost and ::dstPitch
8121 * specify the (host) base address of the destination data and the bytes per
8122 * row to apply. ::dstArray is ignored.
8123 *
8124 * \par
8125 * If ::dstMemoryType is ::CU_MEMORYTYPE_DEVICE, ::dstDevice and ::dstPitch
8126 * specify the (device) base address of the destination data and the bytes per
8127 * row to apply. ::dstArray is ignored.
8128 *
8129 * \par
8130 * If ::dstMemoryType is ::CU_MEMORYTYPE_ARRAY, ::dstArray specifies the
8131 * handle of the destination data. ::dstHost, ::dstDevice and ::dstPitch are
8132 * ignored.
8133 *
8134 * - ::srcXInBytes and ::srcY specify the base address of the source data for
8135 * the copy.
8136 *
8137 * \par
8138 * For host pointers, the starting address is
8139 * \code
8140 void* Start = (void*)((char*)srcHost+srcY*srcPitch + srcXInBytes);
8141 * \endcode
8142 *
8143 * \par
8144 * For device pointers, the starting address is
8145 * \code
8146 CUdeviceptr Start = srcDevice+srcY*srcPitch+srcXInBytes;
8147 * \endcode
8148 *
8149 * \par
8150 * For CUDA arrays, ::srcXInBytes must be evenly divisible by the array
8151 * element size.
8152 *
8153 * - ::dstXInBytes and ::dstY specify the base address of the destination data
8154 * for the copy.
8155 *
8156 * \par
8157 * For host pointers, the base address is
8158 * \code
8159 void* dstStart = (void*)((char*)dstHost+dstY*dstPitch + dstXInBytes);
8160 * \endcode
8161 *
8162 * \par
8163 * For device pointers, the starting address is
8164 * \code
8165 CUdeviceptr dstStart = dstDevice+dstY*dstPitch+dstXInBytes;
8166 * \endcode
8167 *
8168 * \par
8169 * For CUDA arrays, ::dstXInBytes must be evenly divisible by the array
8170 * element size.
8171 *
8172 * - ::WidthInBytes and ::Height specify the width (in bytes) and height of
8173 * the 2D copy being performed.
8174 * - If specified, ::srcPitch must be greater than or equal to ::WidthInBytes +
8175 * ::srcXInBytes, and ::dstPitch must be greater than or equal to
8176 * ::WidthInBytes + dstXInBytes.
8177 * - If specified, ::srcPitch must be greater than or equal to ::WidthInBytes +
8178 * ::srcXInBytes, and ::dstPitch must be greater than or equal to
8179 * ::WidthInBytes + dstXInBytes.
8180 * - If specified, ::srcHeight must be greater than or equal to ::Height +
8181 * ::srcY, and ::dstHeight must be greater than or equal to ::Height + ::dstY.
8182 *
8183 * \par
8184 * ::cuMemcpy2DAsync() returns an error if any pitch is greater than the maximum
8185 * allowed (::CU_DEVICE_ATTRIBUTE_MAX_PITCH). ::cuMemAllocPitch() passes back
8186 * pitches that always work with ::cuMemcpy2D(). On intra-device memory copies
8187 * (device to device, CUDA array to device, CUDA array to CUDA array),
8188 * ::cuMemcpy2DAsync() may fail for pitches not computed by ::cuMemAllocPitch().
8189 *
8190 * \param pCopy - Parameters for the memory copy
8191 * \param hStream - Stream identifier
8192 *
8193 * \return
8194 * ::CUDA_SUCCESS,
8195 * ::CUDA_ERROR_DEINITIALIZED,
8196 * ::CUDA_ERROR_NOT_INITIALIZED,
8197 * ::CUDA_ERROR_INVALID_CONTEXT,
8198 * ::CUDA_ERROR_INVALID_VALUE,
8199 * ::CUDA_ERROR_INVALID_HANDLE
8200 * \notefnerr
8201 * \note_async
8202 * \note_null_stream
8203 *
8204 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
8205 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
8206 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DUnaligned,
8207 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
8208 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
8209 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
8210 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
8211 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
8212 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async,
8213 * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async,
8214 * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async,
8215 * ::cuMemsetD32, ::cuMemsetD32Async,
8216 * ::cudaMemcpy2DAsync,
8217 * ::cudaMemcpy2DToArrayAsync,
8218 * ::cudaMemcpy2DFromArrayAsync
8219 */
8220CUresult CUDAAPI cuMemcpy2DAsync(const CUDA_MEMCPY2D *pCopy, CUstream hStream);
8221
8222/**
8223 * \brief Copies memory for 3D arrays
8224 *
8225 * Perform a 3D memory copy according to the parameters specified in
8226 * \p pCopy. The ::CUDA_MEMCPY3D structure is defined as:
8227 *
8228 * \code
8229 typedef struct CUDA_MEMCPY3D_st {
8230
8231 unsigned int srcXInBytes, srcY, srcZ;
8232 unsigned int srcLOD;
8233 CUmemorytype srcMemoryType;
8234 const void *srcHost;
8235 CUdeviceptr srcDevice;
8236 CUarray srcArray;
8237 unsigned int srcPitch; // ignored when src is array
8238 unsigned int srcHeight; // ignored when src is array; may be 0 if Depth==1
8239
8240 unsigned int dstXInBytes, dstY, dstZ;
8241 unsigned int dstLOD;
8242 CUmemorytype dstMemoryType;
8243 void *dstHost;
8244 CUdeviceptr dstDevice;
8245 CUarray dstArray;
8246 unsigned int dstPitch; // ignored when dst is array
8247 unsigned int dstHeight; // ignored when dst is array; may be 0 if Depth==1
8248
8249 unsigned int WidthInBytes;
8250 unsigned int Height;
8251 unsigned int Depth;
8252 } CUDA_MEMCPY3D;
8253 * \endcode
8254 * where:
8255 * - ::srcMemoryType and ::dstMemoryType specify the type of memory of the
8256 * source and destination, respectively; ::CUmemorytype_enum is defined as:
8257 *
8258 * \code
8259 typedef enum CUmemorytype_enum {
8260 CU_MEMORYTYPE_HOST = 0x01,
8261 CU_MEMORYTYPE_DEVICE = 0x02,
8262 CU_MEMORYTYPE_ARRAY = 0x03,
8263 CU_MEMORYTYPE_UNIFIED = 0x04
8264 } CUmemorytype;
8265 * \endcode
8266 *
8267 * \par
8268 * If ::srcMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::srcDevice and ::srcPitch
8269 * specify the (unified virtual address space) base address of the source data
8270 * and the bytes per row to apply. ::srcArray is ignored.
8271 * This value may be used only if unified addressing is supported in the calling
8272 * context.
8273 *
8274 * \par
8275 * If ::srcMemoryType is ::CU_MEMORYTYPE_HOST, ::srcHost, ::srcPitch and
8276 * ::srcHeight specify the (host) base address of the source data, the bytes
8277 * per row, and the height of each 2D slice of the 3D array. ::srcArray is
8278 * ignored.
8279 *
8280 * \par
8281 * If ::srcMemoryType is ::CU_MEMORYTYPE_DEVICE, ::srcDevice, ::srcPitch and
8282 * ::srcHeight specify the (device) base address of the source data, the bytes
8283 * per row, and the height of each 2D slice of the 3D array. ::srcArray is
8284 * ignored.
8285 *
8286 * \par
8287 * If ::srcMemoryType is ::CU_MEMORYTYPE_ARRAY, ::srcArray specifies the
8288 * handle of the source data. ::srcHost, ::srcDevice, ::srcPitch and
8289 * ::srcHeight are ignored.
8290 *
8291 * \par
8292 * If ::dstMemoryType is ::CU_MEMORYTYPE_UNIFIED, ::dstDevice and ::dstPitch
8293 * specify the (unified virtual address space) base address of the source data
8294 * and the bytes per row to apply. ::dstArray is ignored.
8295 * This value may be used only if unified addressing is supported in the calling
8296 * context.
8297 *
8298 * \par
8299 * If ::dstMemoryType is ::CU_MEMORYTYPE_HOST, ::dstHost and ::dstPitch
8300 * specify the (host) base address of the destination data, the bytes per row,
8301 * and the height of each 2D slice of the 3D array. ::dstArray is ignored.
8302 *
8303 * \par
8304 * If ::dstMemoryType is ::CU_MEMORYTYPE_DEVICE, ::dstDevice and ::dstPitch
8305 * specify the (device) base address of the destination data, the bytes per
8306 * row, and the height of each 2D slice of the 3D array. ::dstArray is ignored.
8307 *
8308 * \par
8309 * If ::dstMemoryType is ::CU_MEMORYTYPE_ARRAY, ::dstArray specifies the
8310 * handle of the destination data. ::dstHost, ::dstDevice, ::dstPitch and
8311 * ::dstHeight are ignored.
8312 *
8313 * - ::srcXInBytes, ::srcY and ::srcZ specify the base address of the source
8314 * data for the copy.
8315 *
8316 * \par
8317 * For host pointers, the starting address is
8318 * \code
8319 void* Start = (void*)((char*)srcHost+(srcZ*srcHeight+srcY)*srcPitch + srcXInBytes);
8320 * \endcode
8321 *
8322 * \par
8323 * For device pointers, the starting address is
8324 * \code
8325 CUdeviceptr Start = srcDevice+(srcZ*srcHeight+srcY)*srcPitch+srcXInBytes;
8326 * \endcode
8327 *
8328 * \par
8329 * For CUDA arrays, ::srcXInBytes must be evenly divisible by the array
8330 * element size.
8331 *
8332 * - dstXInBytes, ::dstY and ::dstZ specify the base address of the
8333 * destination data for the copy.
8334 *
8335 * \par
8336 * For host pointers, the base address is
8337 * \code
8338 void* dstStart = (void*)((char*)dstHost+(dstZ*dstHeight+dstY)*dstPitch + dstXInBytes);
8339 * \endcode
8340 *
8341 * \par
8342 * For device pointers, the starting address is
8343 * \code
8344 CUdeviceptr dstStart = dstDevice+(dstZ*dstHeight+dstY)*dstPitch+dstXInBytes;
8345 * \endcode
8346 *
8347 * \par
8348 * For CUDA arrays, ::dstXInBytes must be evenly divisible by the array
8349 * element size.
8350 *
8351 * - ::WidthInBytes, ::Height and ::Depth specify the width (in bytes), height
8352 * and depth of the 3D copy being performed.
8353 * - If specified, ::srcPitch must be greater than or equal to ::WidthInBytes +
8354 * ::srcXInBytes, and ::dstPitch must be greater than or equal to
8355 * ::WidthInBytes + dstXInBytes.
8356 * - If specified, ::srcHeight must be greater than or equal to ::Height +
8357 * ::srcY, and ::dstHeight must be greater than or equal to ::Height + ::dstY.
8358 *
8359 * \par
8360 * ::cuMemcpy3DAsync() returns an error if any pitch is greater than the maximum
8361 * allowed (::CU_DEVICE_ATTRIBUTE_MAX_PITCH).
8362 *
8363 * The ::srcLOD and ::dstLOD members of the ::CUDA_MEMCPY3D structure must be
8364 * set to 0.
8365 *
8366 * \param pCopy - Parameters for the memory copy
8367 * \param hStream - Stream identifier
8368 *
8369 * \return
8370 * ::CUDA_SUCCESS,
8371 * ::CUDA_ERROR_DEINITIALIZED,
8372 * ::CUDA_ERROR_NOT_INITIALIZED,
8373 * ::CUDA_ERROR_INVALID_CONTEXT,
8374 * ::CUDA_ERROR_INVALID_VALUE,
8375 * ::CUDA_ERROR_INVALID_HANDLE
8376 * \notefnerr
8377 * \note_async
8378 * \note_null_stream
8379 *
8380 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
8381 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
8382 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
8383 * ::cuMemcpy3D, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
8384 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
8385 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
8386 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
8387 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
8388 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async,
8389 * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async,
8390 * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async,
8391 * ::cuMemsetD32, ::cuMemsetD32Async,
8392 * ::cudaMemcpy3DAsync
8393 */
8394CUresult CUDAAPI cuMemcpy3DAsync(const CUDA_MEMCPY3D *pCopy, CUstream hStream);
8395
8396/**
8397 * \brief Copies memory between contexts asynchronously.
8398 *
8399 * Perform a 3D memory copy according to the parameters specified in
8400 * \p pCopy. See the definition of the ::CUDA_MEMCPY3D_PEER structure
8401 * for documentation of its parameters.
8402 *
8403 * \param pCopy - Parameters for the memory copy
8404 * \param hStream - Stream identifier
8405 *
8406 * \return
8407 * ::CUDA_SUCCESS,
8408 * ::CUDA_ERROR_DEINITIALIZED,
8409 * ::CUDA_ERROR_NOT_INITIALIZED,
8410 * ::CUDA_ERROR_INVALID_CONTEXT,
8411 * ::CUDA_ERROR_INVALID_VALUE
8412 * \notefnerr
8413 * \note_async
8414 * \note_null_stream
8415 *
8416 * \sa ::cuMemcpyDtoD, ::cuMemcpyPeer, ::cuMemcpyDtoDAsync, ::cuMemcpyPeerAsync,
8417 * ::cuMemcpy3DPeerAsync,
8418 * ::cudaMemcpy3DPeerAsync
8419 */
8420CUresult CUDAAPI cuMemcpy3DPeerAsync(const CUDA_MEMCPY3D_PEER *pCopy, CUstream hStream);
8421
8422/**
8423 * \brief Initializes device memory
8424 *
8425 * Sets the memory range of \p N 8-bit values to the specified value
8426 * \p uc.
8427 *
8428 * \param dstDevice - Destination device pointer
8429 * \param uc - Value to set
8430 * \param N - Number of elements
8431 *
8432 * \return
8433 * ::CUDA_SUCCESS,
8434 * ::CUDA_ERROR_DEINITIALIZED,
8435 * ::CUDA_ERROR_NOT_INITIALIZED,
8436 * ::CUDA_ERROR_INVALID_CONTEXT,
8437 * ::CUDA_ERROR_INVALID_VALUE
8438 * \notefnerr
8439 * \note_memset
8440 *
8441 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
8442 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
8443 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
8444 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
8445 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
8446 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
8447 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
8448 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
8449 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async,
8450 * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async,
8451 * ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async,
8452 * ::cuMemsetD32, ::cuMemsetD32Async,
8453 * ::cudaMemset
8454 */
8455CUresult CUDAAPI cuMemsetD8(CUdeviceptr dstDevice, unsigned char uc, size_t N);
8456
8457/**
8458 * \brief Initializes device memory
8459 *
8460 * Sets the memory range of \p N 16-bit values to the specified value
8461 * \p us. The \p dstDevice pointer must be two byte aligned.
8462 *
8463 * \param dstDevice - Destination device pointer
8464 * \param us - Value to set
8465 * \param N - Number of elements
8466 *
8467 * \return
8468 * ::CUDA_SUCCESS,
8469 * ::CUDA_ERROR_DEINITIALIZED,
8470 * ::CUDA_ERROR_NOT_INITIALIZED,
8471 * ::CUDA_ERROR_INVALID_CONTEXT,
8472 * ::CUDA_ERROR_INVALID_VALUE
8473 * \notefnerr
8474 * \note_memset
8475 *
8476 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
8477 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
8478 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
8479 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
8480 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
8481 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
8482 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
8483 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
8484 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async,
8485 * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async,
8486 * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16Async,
8487 * ::cuMemsetD32, ::cuMemsetD32Async,
8488 * ::cudaMemset
8489 */
8490CUresult CUDAAPI cuMemsetD16(CUdeviceptr dstDevice, unsigned short us, size_t N);
8491
8492/**
8493 * \brief Initializes device memory
8494 *
8495 * Sets the memory range of \p N 32-bit values to the specified value
8496 * \p ui. The \p dstDevice pointer must be four byte aligned.
8497 *
8498 * \param dstDevice - Destination device pointer
8499 * \param ui - Value to set
8500 * \param N - Number of elements
8501 *
8502 * \return
8503 * ::CUDA_SUCCESS,
8504 * ::CUDA_ERROR_DEINITIALIZED,
8505 * ::CUDA_ERROR_NOT_INITIALIZED,
8506 * ::CUDA_ERROR_INVALID_CONTEXT,
8507 * ::CUDA_ERROR_INVALID_VALUE
8508 * \notefnerr
8509 * \note_memset
8510 *
8511 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
8512 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
8513 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
8514 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
8515 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
8516 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
8517 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
8518 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
8519 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async,
8520 * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async,
8521 * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async,
8522 * ::cuMemsetD32Async,
8523 * ::cudaMemset
8524 */
8525CUresult CUDAAPI cuMemsetD32(CUdeviceptr dstDevice, unsigned int ui, size_t N);
8526
8527/**
8528 * \brief Initializes device memory
8529 *
8530 * Sets the 2D memory range of \p Width 8-bit values to the specified value
8531 * \p uc. \p Height specifies the number of rows to set, and \p dstPitch
8532 * specifies the number of bytes between each row. This function performs
8533 * fastest when the pitch is one that has been passed back by
8534 * ::cuMemAllocPitch().
8535 *
8536 * \param dstDevice - Destination device pointer
8537 * \param dstPitch - Pitch of destination device pointer(Unused if \p Height is 1)
8538 * \param uc - Value to set
8539 * \param Width - Width of row
8540 * \param Height - Number of rows
8541 *
8542 * \return
8543 * ::CUDA_SUCCESS,
8544 * ::CUDA_ERROR_DEINITIALIZED,
8545 * ::CUDA_ERROR_NOT_INITIALIZED,
8546 * ::CUDA_ERROR_INVALID_CONTEXT,
8547 * ::CUDA_ERROR_INVALID_VALUE
8548 * \notefnerr
8549 * \note_memset
8550 *
8551 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
8552 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
8553 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
8554 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
8555 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
8556 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
8557 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
8558 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
8559 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8Async,
8560 * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async,
8561 * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async,
8562 * ::cuMemsetD32, ::cuMemsetD32Async,
8563 * ::cudaMemset2D
8564 */
8565CUresult CUDAAPI cuMemsetD2D8(CUdeviceptr dstDevice, size_t dstPitch, unsigned char uc, size_t Width, size_t Height);
8566
8567/**
8568 * \brief Initializes device memory
8569 *
8570 * Sets the 2D memory range of \p Width 16-bit values to the specified value
8571 * \p us. \p Height specifies the number of rows to set, and \p dstPitch
8572 * specifies the number of bytes between each row. The \p dstDevice pointer
8573 * and \p dstPitch offset must be two byte aligned. This function performs
8574 * fastest when the pitch is one that has been passed back by
8575 * ::cuMemAllocPitch().
8576 *
8577 * \param dstDevice - Destination device pointer
8578 * \param dstPitch - Pitch of destination device pointer(Unused if \p Height is 1)
8579 * \param us - Value to set
8580 * \param Width - Width of row
8581 * \param Height - Number of rows
8582 *
8583 * \return
8584 * ::CUDA_SUCCESS,
8585 * ::CUDA_ERROR_DEINITIALIZED,
8586 * ::CUDA_ERROR_NOT_INITIALIZED,
8587 * ::CUDA_ERROR_INVALID_CONTEXT,
8588 * ::CUDA_ERROR_INVALID_VALUE
8589 * \notefnerr
8590 * \note_memset
8591 *
8592 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
8593 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
8594 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
8595 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
8596 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
8597 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
8598 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
8599 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
8600 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async,
8601 * ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async,
8602 * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async,
8603 * ::cuMemsetD32, ::cuMemsetD32Async,
8604 * ::cudaMemset2D
8605 */
8606CUresult CUDAAPI cuMemsetD2D16(CUdeviceptr dstDevice, size_t dstPitch, unsigned short us, size_t Width, size_t Height);
8607
8608/**
8609 * \brief Initializes device memory
8610 *
8611 * Sets the 2D memory range of \p Width 32-bit values to the specified value
8612 * \p ui. \p Height specifies the number of rows to set, and \p dstPitch
8613 * specifies the number of bytes between each row. The \p dstDevice pointer
8614 * and \p dstPitch offset must be four byte aligned. This function performs
8615 * fastest when the pitch is one that has been passed back by
8616 * ::cuMemAllocPitch().
8617 *
8618 * \param dstDevice - Destination device pointer
8619 * \param dstPitch - Pitch of destination device pointer(Unused if \p Height is 1)
8620 * \param ui - Value to set
8621 * \param Width - Width of row
8622 * \param Height - Number of rows
8623 *
8624 * \return
8625 * ::CUDA_SUCCESS,
8626 * ::CUDA_ERROR_DEINITIALIZED,
8627 * ::CUDA_ERROR_NOT_INITIALIZED,
8628 * ::CUDA_ERROR_INVALID_CONTEXT,
8629 * ::CUDA_ERROR_INVALID_VALUE
8630 * \notefnerr
8631 * \note_memset
8632 *
8633 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
8634 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
8635 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
8636 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
8637 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
8638 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
8639 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
8640 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
8641 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async,
8642 * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32Async,
8643 * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async,
8644 * ::cuMemsetD32, ::cuMemsetD32Async,
8645 * ::cudaMemset2D
8646 */
8647CUresult CUDAAPI cuMemsetD2D32(CUdeviceptr dstDevice, size_t dstPitch, unsigned int ui, size_t Width, size_t Height);
8648
8649/**
8650 * \brief Sets device memory
8651 *
8652 * Sets the memory range of \p N 8-bit values to the specified value
8653 * \p uc.
8654 *
8655 * \param dstDevice - Destination device pointer
8656 * \param uc - Value to set
8657 * \param N - Number of elements
8658 * \param hStream - Stream identifier
8659 *
8660 * \return
8661 * ::CUDA_SUCCESS,
8662 * ::CUDA_ERROR_DEINITIALIZED,
8663 * ::CUDA_ERROR_NOT_INITIALIZED,
8664 * ::CUDA_ERROR_INVALID_CONTEXT,
8665 * ::CUDA_ERROR_INVALID_VALUE
8666 * \notefnerr
8667 * \note_memset
8668 * \note_null_stream
8669 *
8670 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
8671 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
8672 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
8673 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
8674 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
8675 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
8676 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
8677 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
8678 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async,
8679 * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async,
8680 * ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD16Async,
8681 * ::cuMemsetD32, ::cuMemsetD32Async,
8682 * ::cudaMemsetAsync
8683 */
8684CUresult CUDAAPI cuMemsetD8Async(CUdeviceptr dstDevice, unsigned char uc, size_t N, CUstream hStream);
8685
8686/**
8687 * \brief Sets device memory
8688 *
8689 * Sets the memory range of \p N 16-bit values to the specified value
8690 * \p us. The \p dstDevice pointer must be two byte aligned.
8691 *
8692 * \param dstDevice - Destination device pointer
8693 * \param us - Value to set
8694 * \param N - Number of elements
8695 * \param hStream - Stream identifier
8696 *
8697 * \return
8698 * ::CUDA_SUCCESS,
8699 * ::CUDA_ERROR_DEINITIALIZED,
8700 * ::CUDA_ERROR_NOT_INITIALIZED,
8701 * ::CUDA_ERROR_INVALID_CONTEXT,
8702 * ::CUDA_ERROR_INVALID_VALUE
8703 * \notefnerr
8704 * \note_memset
8705 * \note_null_stream
8706 *
8707 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
8708 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
8709 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
8710 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
8711 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
8712 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
8713 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
8714 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
8715 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async,
8716 * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async,
8717 * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16,
8718 * ::cuMemsetD32, ::cuMemsetD32Async,
8719 * ::cudaMemsetAsync
8720 */
8721CUresult CUDAAPI cuMemsetD16Async(CUdeviceptr dstDevice, unsigned short us, size_t N, CUstream hStream);
8722
8723/**
8724 * \brief Sets device memory
8725 *
8726 * Sets the memory range of \p N 32-bit values to the specified value
8727 * \p ui. The \p dstDevice pointer must be four byte aligned.
8728 *
8729 * \param dstDevice - Destination device pointer
8730 * \param ui - Value to set
8731 * \param N - Number of elements
8732 * \param hStream - Stream identifier
8733 *
8734 * \return
8735 * ::CUDA_SUCCESS,
8736 * ::CUDA_ERROR_DEINITIALIZED,
8737 * ::CUDA_ERROR_NOT_INITIALIZED,
8738 * ::CUDA_ERROR_INVALID_CONTEXT,
8739 * ::CUDA_ERROR_INVALID_VALUE
8740 * \notefnerr
8741 * \note_memset
8742 * \note_null_stream
8743 *
8744 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
8745 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
8746 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
8747 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
8748 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
8749 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
8750 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
8751 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
8752 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async,
8753 * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async,
8754 * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async, ::cuMemsetD32,
8755 * ::cudaMemsetAsync
8756 */
8757CUresult CUDAAPI cuMemsetD32Async(CUdeviceptr dstDevice, unsigned int ui, size_t N, CUstream hStream);
8758
8759/**
8760 * \brief Sets device memory
8761 *
8762 * Sets the 2D memory range of \p Width 8-bit values to the specified value
8763 * \p uc. \p Height specifies the number of rows to set, and \p dstPitch
8764 * specifies the number of bytes between each row. This function performs
8765 * fastest when the pitch is one that has been passed back by
8766 * ::cuMemAllocPitch().
8767 *
8768 * \param dstDevice - Destination device pointer
8769 * \param dstPitch - Pitch of destination device pointer(Unused if \p Height is 1)
8770 * \param uc - Value to set
8771 * \param Width - Width of row
8772 * \param Height - Number of rows
8773 * \param hStream - Stream identifier
8774 *
8775 * \return
8776 * ::CUDA_SUCCESS,
8777 * ::CUDA_ERROR_DEINITIALIZED,
8778 * ::CUDA_ERROR_NOT_INITIALIZED,
8779 * ::CUDA_ERROR_INVALID_CONTEXT,
8780 * ::CUDA_ERROR_INVALID_VALUE
8781 * \notefnerr
8782 * \note_memset
8783 * \note_null_stream
8784 *
8785 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
8786 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
8787 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
8788 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
8789 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
8790 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
8791 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
8792 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
8793 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8,
8794 * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32, ::cuMemsetD2D32Async,
8795 * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async,
8796 * ::cuMemsetD32, ::cuMemsetD32Async,
8797 * ::cudaMemset2DAsync
8798 */
8799CUresult CUDAAPI cuMemsetD2D8Async(CUdeviceptr dstDevice, size_t dstPitch, unsigned char uc, size_t Width, size_t Height, CUstream hStream);
8800
8801/**
8802 * \brief Sets device memory
8803 *
8804 * Sets the 2D memory range of \p Width 16-bit values to the specified value
8805 * \p us. \p Height specifies the number of rows to set, and \p dstPitch
8806 * specifies the number of bytes between each row. The \p dstDevice pointer
8807 * and \p dstPitch offset must be two byte aligned. This function performs
8808 * fastest when the pitch is one that has been passed back by
8809 * ::cuMemAllocPitch().
8810 *
8811 * \param dstDevice - Destination device pointer
8812 * \param dstPitch - Pitch of destination device pointer(Unused if \p Height is 1)
8813 * \param us - Value to set
8814 * \param Width - Width of row
8815 * \param Height - Number of rows
8816 * \param hStream - Stream identifier
8817 *
8818 * \return
8819 * ::CUDA_SUCCESS,
8820 * ::CUDA_ERROR_DEINITIALIZED,
8821 * ::CUDA_ERROR_NOT_INITIALIZED,
8822 * ::CUDA_ERROR_INVALID_CONTEXT,
8823 * ::CUDA_ERROR_INVALID_VALUE
8824 * \notefnerr
8825 * \note_memset
8826 * \note_null_stream
8827 *
8828 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
8829 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
8830 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
8831 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
8832 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
8833 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
8834 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
8835 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
8836 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async,
8837 * ::cuMemsetD2D16, ::cuMemsetD2D32, ::cuMemsetD2D32Async,
8838 * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async,
8839 * ::cuMemsetD32, ::cuMemsetD32Async,
8840 * ::cudaMemset2DAsync
8841 */
8842CUresult CUDAAPI cuMemsetD2D16Async(CUdeviceptr dstDevice, size_t dstPitch, unsigned short us, size_t Width, size_t Height, CUstream hStream);
8843
8844/**
8845 * \brief Sets device memory
8846 *
8847 * Sets the 2D memory range of \p Width 32-bit values to the specified value
8848 * \p ui. \p Height specifies the number of rows to set, and \p dstPitch
8849 * specifies the number of bytes between each row. The \p dstDevice pointer
8850 * and \p dstPitch offset must be four byte aligned. This function performs
8851 * fastest when the pitch is one that has been passed back by
8852 * ::cuMemAllocPitch().
8853 *
8854 * \param dstDevice - Destination device pointer
8855 * \param dstPitch - Pitch of destination device pointer(Unused if \p Height is 1)
8856 * \param ui - Value to set
8857 * \param Width - Width of row
8858 * \param Height - Number of rows
8859 * \param hStream - Stream identifier
8860 *
8861 * \return
8862 * ::CUDA_SUCCESS,
8863 * ::CUDA_ERROR_DEINITIALIZED,
8864 * ::CUDA_ERROR_NOT_INITIALIZED,
8865 * ::CUDA_ERROR_INVALID_CONTEXT,
8866 * ::CUDA_ERROR_INVALID_VALUE
8867 * \notefnerr
8868 * \note_memset
8869 * \note_null_stream
8870 *
8871 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
8872 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
8873 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
8874 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
8875 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
8876 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
8877 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
8878 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
8879 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D8Async,
8880 * ::cuMemsetD2D16, ::cuMemsetD2D16Async, ::cuMemsetD2D32,
8881 * ::cuMemsetD8, ::cuMemsetD8Async, ::cuMemsetD16, ::cuMemsetD16Async,
8882 * ::cuMemsetD32, ::cuMemsetD32Async,
8883 * ::cudaMemset2DAsync
8884 */
8885CUresult CUDAAPI cuMemsetD2D32Async(CUdeviceptr dstDevice, size_t dstPitch, unsigned int ui, size_t Width, size_t Height, CUstream hStream);
8886
8887/**
8888 * \brief Creates a 1D or 2D CUDA array
8889 *
8890 * Creates a CUDA array according to the ::CUDA_ARRAY_DESCRIPTOR structure
8891 * \p pAllocateArray and returns a handle to the new CUDA array in \p *pHandle.
8892 * The ::CUDA_ARRAY_DESCRIPTOR is defined as:
8893 *
8894 * \code
8895 typedef struct {
8896 unsigned int Width;
8897 unsigned int Height;
8898 CUarray_format Format;
8899 unsigned int NumChannels;
8900 } CUDA_ARRAY_DESCRIPTOR;
8901 * \endcode
8902 * where:
8903 *
8904 * - \p Width, and \p Height are the width, and height of the CUDA array (in
8905 * elements); the CUDA array is one-dimensional if height is 0, two-dimensional
8906 * otherwise;
8907 * - ::Format specifies the format of the elements; ::CUarray_format is
8908 * defined as:
8909 * \code
8910 typedef enum CUarray_format_enum {
8911 CU_AD_FORMAT_UNSIGNED_INT8 = 0x01,
8912 CU_AD_FORMAT_UNSIGNED_INT16 = 0x02,
8913 CU_AD_FORMAT_UNSIGNED_INT32 = 0x03,
8914 CU_AD_FORMAT_SIGNED_INT8 = 0x08,
8915 CU_AD_FORMAT_SIGNED_INT16 = 0x09,
8916 CU_AD_FORMAT_SIGNED_INT32 = 0x0a,
8917 CU_AD_FORMAT_HALF = 0x10,
8918 CU_AD_FORMAT_FLOAT = 0x20
8919 } CUarray_format;
8920 * \endcode
8921 * - \p NumChannels specifies the number of packed components per CUDA array
8922 * element; it may be 1, 2, or 4;
8923 *
8924 * Here are examples of CUDA array descriptions:
8925 *
8926 * Description for a CUDA array of 2048 floats:
8927 * \code
8928 CUDA_ARRAY_DESCRIPTOR desc;
8929 desc.Format = CU_AD_FORMAT_FLOAT;
8930 desc.NumChannels = 1;
8931 desc.Width = 2048;
8932 desc.Height = 1;
8933 * \endcode
8934 *
8935 * Description for a 64 x 64 CUDA array of floats:
8936 * \code
8937 CUDA_ARRAY_DESCRIPTOR desc;
8938 desc.Format = CU_AD_FORMAT_FLOAT;
8939 desc.NumChannels = 1;
8940 desc.Width = 64;
8941 desc.Height = 64;
8942 * \endcode
8943 *
8944 * Description for a \p width x \p height CUDA array of 64-bit, 4x16-bit
8945 * float16's:
8946 * \code
8947 CUDA_ARRAY_DESCRIPTOR desc;
8948 desc.FormatFlags = CU_AD_FORMAT_HALF;
8949 desc.NumChannels = 4;
8950 desc.Width = width;
8951 desc.Height = height;
8952 * \endcode
8953 *
8954 * Description for a \p width x \p height CUDA array of 16-bit elements, each
8955 * of which is two 8-bit unsigned chars:
8956 * \code
8957 CUDA_ARRAY_DESCRIPTOR arrayDesc;
8958 desc.FormatFlags = CU_AD_FORMAT_UNSIGNED_INT8;
8959 desc.NumChannels = 2;
8960 desc.Width = width;
8961 desc.Height = height;
8962 * \endcode
8963 *
8964 * \param pHandle - Returned array
8965 * \param pAllocateArray - Array descriptor
8966 *
8967 * \return
8968 * ::CUDA_SUCCESS,
8969 * ::CUDA_ERROR_DEINITIALIZED,
8970 * ::CUDA_ERROR_NOT_INITIALIZED,
8971 * ::CUDA_ERROR_INVALID_CONTEXT,
8972 * ::CUDA_ERROR_INVALID_VALUE,
8973 * ::CUDA_ERROR_OUT_OF_MEMORY,
8974 * ::CUDA_ERROR_UNKNOWN
8975 * \notefnerr
8976 *
8977 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor,
8978 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
8979 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
8980 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
8981 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
8982 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
8983 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
8984 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
8985 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
8986 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
8987 * ::cudaMallocArray
8988 */
8989CUresult CUDAAPI cuArrayCreate(CUarray *pHandle, const CUDA_ARRAY_DESCRIPTOR *pAllocateArray);
8990
8991/**
8992 * \brief Get a 1D or 2D CUDA array descriptor
8993 *
8994 * Returns in \p *pArrayDescriptor a descriptor containing information on the
8995 * format and dimensions of the CUDA array \p hArray. It is useful for
8996 * subroutines that have been passed a CUDA array, but need to know the CUDA
8997 * array parameters for validation or other purposes.
8998 *
8999 * \param pArrayDescriptor - Returned array descriptor
9000 * \param hArray - Array to get descriptor of
9001 *
9002 * \return
9003 * ::CUDA_SUCCESS,
9004 * ::CUDA_ERROR_DEINITIALIZED,
9005 * ::CUDA_ERROR_NOT_INITIALIZED,
9006 * ::CUDA_ERROR_INVALID_CONTEXT,
9007 * ::CUDA_ERROR_INVALID_VALUE,
9008 * ::CUDA_ERROR_INVALID_HANDLE
9009 * \notefnerr
9010 *
9011 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
9012 * ::cuArrayDestroy, ::cuMemAlloc, ::cuMemAllocHost,
9013 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
9014 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
9015 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
9016 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
9017 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
9018 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
9019 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
9020 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
9021 * ::cudaArrayGetInfo
9022 */
9023CUresult CUDAAPI cuArrayGetDescriptor(CUDA_ARRAY_DESCRIPTOR *pArrayDescriptor, CUarray hArray);
9024
9025/**
9026 * \brief Returns the layout properties of a sparse CUDA array
9027 *
9028 * Returns the layout properties of a sparse CUDA array in \p sparseProperties
9029 * If the CUDA array is not allocated with flag ::CUDA_ARRAY3D_SPARSE
9030 * ::CUDA_ERROR_INVALID_VALUE will be returned.
9031 *
9032 * If the returned value in ::CUDA_ARRAY_SPARSE_PROPERTIES::flags contains ::CU_ARRAY_SPARSE_PROPERTIES_SINGLE_MIPTAIL,
9033 * then ::CUDA_ARRAY_SPARSE_PROPERTIES::miptailSize represents the total size of the array. Otherwise, it will be zero.
9034 * Also, the returned value in ::CUDA_ARRAY_SPARSE_PROPERTIES::miptailFirstLevel is always zero.
9035 * Note that the \p array must have been allocated using ::cuArrayCreate or ::cuArray3DCreate. For CUDA arrays obtained
9036 * using ::cuMipmappedArrayGetLevel, ::CUDA_ERROR_INVALID_VALUE will be returned. Instead, ::cuMipmappedArrayGetSparseProperties
9037 * must be used to obtain the sparse properties of the entire CUDA mipmapped array to which \p array belongs to.
9038 *
9039 * \return
9040 * ::CUDA_SUCCESS
9041 * ::CUDA_ERROR_INVALID_VALUE
9042 *
9043 * \param[out] sparseProperties - Pointer to ::CUDA_ARRAY_SPARSE_PROPERTIES
9044 * \param[in] array - CUDA array to get the sparse properties of
9045 * \sa ::cuMipmappedArrayGetSparseProperties, ::cuMemMapArrayAsync
9046 */
9047CUresult CUDAAPI cuArrayGetSparseProperties(CUDA_ARRAY_SPARSE_PROPERTIES *sparseProperties, CUarray array);
9048
9049/**
9050 * \brief Returns the layout properties of a sparse CUDA mipmapped array
9051 *
9052 * Returns the sparse array layout properties in \p sparseProperties
9053 * If the CUDA mipmapped array is not allocated with flag ::CUDA_ARRAY3D_SPARSE
9054 * ::CUDA_ERROR_INVALID_VALUE will be returned.
9055 *
9056 * For non-layered CUDA mipmapped arrays, ::CUDA_ARRAY_SPARSE_PROPERTIES::miptailSize returns the
9057 * size of the mip tail region. The mip tail region includes all mip levels whose width, height or depth
9058 * is less than that of the tile.
9059 * For layered CUDA mipmapped arrays, if ::CUDA_ARRAY_SPARSE_PROPERTIES::flags contains ::CU_ARRAY_SPARSE_PROPERTIES_SINGLE_MIPTAIL,
9060 * then ::CUDA_ARRAY_SPARSE_PROPERTIES::miptailSize specifies the size of the mip tail of all layers combined.
9061 * Otherwise, ::CUDA_ARRAY_SPARSE_PROPERTIES::miptailSize specifies mip tail size per layer.
9062 * The returned value of ::CUDA_ARRAY_SPARSE_PROPERTIES::miptailFirstLevel is valid only if ::CUDA_ARRAY_SPARSE_PROPERTIES::miptailSize is non-zero.
9063 *
9064 * \return
9065 * ::CUDA_SUCCESS
9066 * ::CUDA_ERROR_INVALID_VALUE
9067 *
9068 * \param[out] sparseProperties - Pointer to ::CUDA_ARRAY_SPARSE_PROPERTIES
9069 * \param[in] mipmap - CUDA mipmapped array to get the sparse properties of
9070 * \sa ::cuArrayGetSparseProperties, ::cuMemMapArrayAsync
9071 */
9072CUresult CUDAAPI cuMipmappedArrayGetSparseProperties(CUDA_ARRAY_SPARSE_PROPERTIES *sparseProperties, CUmipmappedArray mipmap);
9073
9074/**
9075 * \brief Gets a CUDA array plane from a CUDA array
9076 *
9077 * Returns in \p pPlaneArray a CUDA array that represents a single format plane
9078 * of the CUDA array \p hArray.
9079 *
9080 * If \p planeIdx is greater than the maximum number of planes in this array or if the array does
9081 * not have a multi-planar format e.g: ::CU_AD_FORMAT_NV12, then ::CUDA_ERROR_INVALID_VALUE is returned.
9082 *
9083 * Note that if the \p hArray has format ::CU_AD_FORMAT_NV12, then passing in 0 for \p planeIdx returns
9084 * a CUDA array of the same size as \p hArray but with one channel and ::CU_AD_FORMAT_UNSIGNED_INT8 as its format.
9085 * If 1 is passed for \p planeIdx, then the returned CUDA array has half the height and width
9086 * of \p hArray with two channels and ::CU_AD_FORMAT_UNSIGNED_INT8 as its format.
9087 *
9088 * \param pPlaneArray - Returned CUDA array referenced by the \p planeIdx
9089 * \param hArray - Multiplanar CUDA array
9090 * \param planeIdx - Plane index
9091 *
9092 * \return
9093 * ::CUDA_SUCCESS,
9094 * ::CUDA_ERROR_DEINITIALIZED,
9095 * ::CUDA_ERROR_NOT_INITIALIZED,
9096 * ::CUDA_ERROR_INVALID_CONTEXT,
9097 * ::CUDA_ERROR_INVALID_VALUE,
9098 * ::CUDA_ERROR_INVALID_HANDLE
9099 * \notefnerr
9100 *
9101 * \sa
9102 * ::cuArrayCreate,
9103 * ::cudaGetArrayPlane
9104 */
9105CUresult CUDAAPI cuArrayGetPlane(CUarray *pPlaneArray, CUarray hArray, unsigned int planeIdx);
9106
9107/**
9108 * \brief Destroys a CUDA array
9109 *
9110 * Destroys the CUDA array \p hArray.
9111 *
9112 * \param hArray - Array to destroy
9113 *
9114 * \return
9115 * ::CUDA_SUCCESS,
9116 * ::CUDA_ERROR_DEINITIALIZED,
9117 * ::CUDA_ERROR_NOT_INITIALIZED,
9118 * ::CUDA_ERROR_INVALID_CONTEXT,
9119 * ::CUDA_ERROR_INVALID_HANDLE,
9120 * ::CUDA_ERROR_ARRAY_IS_MAPPED,
9121 * ::CUDA_ERROR_CONTEXT_IS_DESTROYED
9122 * \notefnerr
9123 *
9124 * \sa ::cuArray3DCreate, ::cuArray3DGetDescriptor, ::cuArrayCreate,
9125 * ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
9126 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
9127 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
9128 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
9129 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
9130 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
9131 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
9132 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
9133 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
9134 * ::cudaFreeArray
9135 */
9136CUresult CUDAAPI cuArrayDestroy(CUarray hArray);
9137
9138/**
9139 * \brief Creates a 3D CUDA array
9140 *
9141 * Creates a CUDA array according to the ::CUDA_ARRAY3D_DESCRIPTOR structure
9142 * \p pAllocateArray and returns a handle to the new CUDA array in \p *pHandle.
9143 * The ::CUDA_ARRAY3D_DESCRIPTOR is defined as:
9144 *
9145 * \code
9146 typedef struct {
9147 unsigned int Width;
9148 unsigned int Height;
9149 unsigned int Depth;
9150 CUarray_format Format;
9151 unsigned int NumChannels;
9152 unsigned int Flags;
9153 } CUDA_ARRAY3D_DESCRIPTOR;
9154 * \endcode
9155 * where:
9156 *
9157 * - \p Width, \p Height, and \p Depth are the width, height, and depth of the
9158 * CUDA array (in elements); the following types of CUDA arrays can be allocated:
9159 * - A 1D array is allocated if \p Height and \p Depth extents are both zero.
9160 * - A 2D array is allocated if only \p Depth extent is zero.
9161 * - A 3D array is allocated if all three extents are non-zero.
9162 * - A 1D layered CUDA array is allocated if only \p Height is zero and the
9163 * ::CUDA_ARRAY3D_LAYERED flag is set. Each layer is a 1D array. The number
9164 * of layers is determined by the depth extent.
9165 * - A 2D layered CUDA array is allocated if all three extents are non-zero and
9166 * the ::CUDA_ARRAY3D_LAYERED flag is set. Each layer is a 2D array. The number
9167 * of layers is determined by the depth extent.
9168 * - A cubemap CUDA array is allocated if all three extents are non-zero and the
9169 * ::CUDA_ARRAY3D_CUBEMAP flag is set. \p Width must be equal to \p Height, and
9170 * \p Depth must be six. A cubemap is a special type of 2D layered CUDA array,
9171 * where the six layers represent the six faces of a cube. The order of the six
9172 * layers in memory is the same as that listed in ::CUarray_cubemap_face.
9173 * - A cubemap layered CUDA array is allocated if all three extents are non-zero,
9174 * and both, ::CUDA_ARRAY3D_CUBEMAP and ::CUDA_ARRAY3D_LAYERED flags are set.
9175 * \p Width must be equal to \p Height, and \p Depth must be a multiple of six.
9176 * A cubemap layered CUDA array is a special type of 2D layered CUDA array that
9177 * consists of a collection of cubemaps. The first six layers represent the first
9178 * cubemap, the next six layers form the second cubemap, and so on.
9179 *
9180 * - ::Format specifies the format of the elements; ::CUarray_format is
9181 * defined as:
9182 * \code
9183 typedef enum CUarray_format_enum {
9184 CU_AD_FORMAT_UNSIGNED_INT8 = 0x01,
9185 CU_AD_FORMAT_UNSIGNED_INT16 = 0x02,
9186 CU_AD_FORMAT_UNSIGNED_INT32 = 0x03,
9187 CU_AD_FORMAT_SIGNED_INT8 = 0x08,
9188 CU_AD_FORMAT_SIGNED_INT16 = 0x09,
9189 CU_AD_FORMAT_SIGNED_INT32 = 0x0a,
9190 CU_AD_FORMAT_HALF = 0x10,
9191 CU_AD_FORMAT_FLOAT = 0x20
9192 } CUarray_format;
9193 * \endcode
9194 *
9195 * - \p NumChannels specifies the number of packed components per CUDA array
9196 * element; it may be 1, 2, or 4;
9197 *
9198 * - ::Flags may be set to
9199 * - ::CUDA_ARRAY3D_LAYERED to enable creation of layered CUDA arrays. If this flag is set,
9200 * \p Depth specifies the number of layers, not the depth of a 3D array.
9201 * - ::CUDA_ARRAY3D_SURFACE_LDST to enable surface references to be bound to the CUDA array.
9202 * If this flag is not set, ::cuSurfRefSetArray will fail when attempting to bind the CUDA array
9203 * to a surface reference.
9204 * - ::CUDA_ARRAY3D_CUBEMAP to enable creation of cubemaps. If this flag is set, \p Width must be
9205 * equal to \p Height, and \p Depth must be six. If the ::CUDA_ARRAY3D_LAYERED flag is also set,
9206 * then \p Depth must be a multiple of six.
9207 * - ::CUDA_ARRAY3D_TEXTURE_GATHER to indicate that the CUDA array will be used for texture gather.
9208 * Texture gather can only be performed on 2D CUDA arrays.
9209 *
9210 * \p Width, \p Height and \p Depth must meet certain size requirements as listed in the following table.
9211 * All values are specified in elements. Note that for brevity's sake, the full name of the device attribute
9212 * is not specified. For ex., TEXTURE1D_WIDTH refers to the device attribute
9213 * ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_WIDTH.
9214 *
9215 * Note that 2D CUDA arrays have different size requirements if the ::CUDA_ARRAY3D_TEXTURE_GATHER flag
9216 * is set. \p Width and \p Height must not be greater than ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_GATHER_WIDTH
9217 * and ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_GATHER_HEIGHT respectively, in that case.
9218 *
9219 * <table>
9220 * <tr><td><b>CUDA array type</b></td>
9221 * <td><b>Valid extents that must always be met<br>{(width range in elements), (height range),
9222 * (depth range)}</b></td>
9223 * <td><b>Valid extents with CUDA_ARRAY3D_SURFACE_LDST set<br>
9224 * {(width range in elements), (height range), (depth range)}</b></td></tr>
9225 * <tr><td>1D</td>
9226 * <td><small>{ (1,TEXTURE1D_WIDTH), 0, 0 }</small></td>
9227 * <td><small>{ (1,SURFACE1D_WIDTH), 0, 0 }</small></td></tr>
9228 * <tr><td>2D</td>
9229 * <td><small>{ (1,TEXTURE2D_WIDTH), (1,TEXTURE2D_HEIGHT), 0 }</small></td>
9230 * <td><small>{ (1,SURFACE2D_WIDTH), (1,SURFACE2D_HEIGHT), 0 }</small></td></tr>
9231 * <tr><td>3D</td>
9232 * <td><small>{ (1,TEXTURE3D_WIDTH), (1,TEXTURE3D_HEIGHT), (1,TEXTURE3D_DEPTH) }
9233 * <br>OR<br>{ (1,TEXTURE3D_WIDTH_ALTERNATE), (1,TEXTURE3D_HEIGHT_ALTERNATE),
9234 * (1,TEXTURE3D_DEPTH_ALTERNATE) }</small></td>
9235 * <td><small>{ (1,SURFACE3D_WIDTH), (1,SURFACE3D_HEIGHT),
9236 * (1,SURFACE3D_DEPTH) }</small></td></tr>
9237 * <tr><td>1D Layered</td>
9238 * <td><small>{ (1,TEXTURE1D_LAYERED_WIDTH), 0,
9239 * (1,TEXTURE1D_LAYERED_LAYERS) }</small></td>
9240 * <td><small>{ (1,SURFACE1D_LAYERED_WIDTH), 0,
9241 * (1,SURFACE1D_LAYERED_LAYERS) }</small></td></tr>
9242 * <tr><td>2D Layered</td>
9243 * <td><small>{ (1,TEXTURE2D_LAYERED_WIDTH), (1,TEXTURE2D_LAYERED_HEIGHT),
9244 * (1,TEXTURE2D_LAYERED_LAYERS) }</small></td>
9245 * <td><small>{ (1,SURFACE2D_LAYERED_WIDTH), (1,SURFACE2D_LAYERED_HEIGHT),
9246 * (1,SURFACE2D_LAYERED_LAYERS) }</small></td></tr>
9247 * <tr><td>Cubemap</td>
9248 * <td><small>{ (1,TEXTURECUBEMAP_WIDTH), (1,TEXTURECUBEMAP_WIDTH), 6 }</small></td>
9249 * <td><small>{ (1,SURFACECUBEMAP_WIDTH),
9250 * (1,SURFACECUBEMAP_WIDTH), 6 }</small></td></tr>
9251 * <tr><td>Cubemap Layered</td>
9252 * <td><small>{ (1,TEXTURECUBEMAP_LAYERED_WIDTH), (1,TEXTURECUBEMAP_LAYERED_WIDTH),
9253 * (1,TEXTURECUBEMAP_LAYERED_LAYERS) }</small></td>
9254 * <td><small>{ (1,SURFACECUBEMAP_LAYERED_WIDTH), (1,SURFACECUBEMAP_LAYERED_WIDTH),
9255 * (1,SURFACECUBEMAP_LAYERED_LAYERS) }</small></td></tr>
9256 * </table>
9257 *
9258 * Here are examples of CUDA array descriptions:
9259 *
9260 * Description for a CUDA array of 2048 floats:
9261 * \code
9262 CUDA_ARRAY3D_DESCRIPTOR desc;
9263 desc.Format = CU_AD_FORMAT_FLOAT;
9264 desc.NumChannels = 1;
9265 desc.Width = 2048;
9266 desc.Height = 0;
9267 desc.Depth = 0;
9268 * \endcode
9269 *
9270 * Description for a 64 x 64 CUDA array of floats:
9271 * \code
9272 CUDA_ARRAY3D_DESCRIPTOR desc;
9273 desc.Format = CU_AD_FORMAT_FLOAT;
9274 desc.NumChannels = 1;
9275 desc.Width = 64;
9276 desc.Height = 64;
9277 desc.Depth = 0;
9278 * \endcode
9279 *
9280 * Description for a \p width x \p height x \p depth CUDA array of 64-bit,
9281 * 4x16-bit float16's:
9282 * \code
9283 CUDA_ARRAY3D_DESCRIPTOR desc;
9284 desc.FormatFlags = CU_AD_FORMAT_HALF;
9285 desc.NumChannels = 4;
9286 desc.Width = width;
9287 desc.Height = height;
9288 desc.Depth = depth;
9289 * \endcode
9290 *
9291 * \param pHandle - Returned array
9292 * \param pAllocateArray - 3D array descriptor
9293 *
9294 * \return
9295 * ::CUDA_SUCCESS,
9296 * ::CUDA_ERROR_DEINITIALIZED,
9297 * ::CUDA_ERROR_NOT_INITIALIZED,
9298 * ::CUDA_ERROR_INVALID_CONTEXT,
9299 * ::CUDA_ERROR_INVALID_VALUE,
9300 * ::CUDA_ERROR_OUT_OF_MEMORY,
9301 * ::CUDA_ERROR_UNKNOWN
9302 * \notefnerr
9303 *
9304 * \sa ::cuArray3DGetDescriptor, ::cuArrayCreate,
9305 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
9306 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
9307 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
9308 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
9309 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
9310 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
9311 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
9312 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
9313 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
9314 * ::cudaMalloc3DArray
9315 */
9316CUresult CUDAAPI cuArray3DCreate(CUarray *pHandle, const CUDA_ARRAY3D_DESCRIPTOR *pAllocateArray);
9317
9318/**
9319 * \brief Get a 3D CUDA array descriptor
9320 *
9321 * Returns in \p *pArrayDescriptor a descriptor containing information on the
9322 * format and dimensions of the CUDA array \p hArray. It is useful for
9323 * subroutines that have been passed a CUDA array, but need to know the CUDA
9324 * array parameters for validation or other purposes.
9325 *
9326 * This function may be called on 1D and 2D arrays, in which case the \p Height
9327 * and/or \p Depth members of the descriptor struct will be set to 0.
9328 *
9329 * \param pArrayDescriptor - Returned 3D array descriptor
9330 * \param hArray - 3D array to get descriptor of
9331 *
9332 * \return
9333 * ::CUDA_SUCCESS,
9334 * ::CUDA_ERROR_DEINITIALIZED,
9335 * ::CUDA_ERROR_NOT_INITIALIZED,
9336 * ::CUDA_ERROR_INVALID_CONTEXT,
9337 * ::CUDA_ERROR_INVALID_VALUE,
9338 * ::CUDA_ERROR_INVALID_HANDLE,
9339 * ::CUDA_ERROR_CONTEXT_IS_DESTROYED
9340 * \notefnerr
9341 *
9342 * \sa ::cuArray3DCreate, ::cuArrayCreate,
9343 * ::cuArrayDestroy, ::cuArrayGetDescriptor, ::cuMemAlloc, ::cuMemAllocHost,
9344 * ::cuMemAllocPitch, ::cuMemcpy2D, ::cuMemcpy2DAsync, ::cuMemcpy2DUnaligned,
9345 * ::cuMemcpy3D, ::cuMemcpy3DAsync, ::cuMemcpyAtoA, ::cuMemcpyAtoD,
9346 * ::cuMemcpyAtoH, ::cuMemcpyAtoHAsync, ::cuMemcpyDtoA, ::cuMemcpyDtoD, ::cuMemcpyDtoDAsync,
9347 * ::cuMemcpyDtoH, ::cuMemcpyDtoHAsync, ::cuMemcpyHtoA, ::cuMemcpyHtoAAsync,
9348 * ::cuMemcpyHtoD, ::cuMemcpyHtoDAsync, ::cuMemFree, ::cuMemFreeHost,
9349 * ::cuMemGetAddressRange, ::cuMemGetInfo, ::cuMemHostAlloc,
9350 * ::cuMemHostGetDevicePointer, ::cuMemsetD2D8, ::cuMemsetD2D16,
9351 * ::cuMemsetD2D32, ::cuMemsetD8, ::cuMemsetD16, ::cuMemsetD32,
9352 * ::cudaArrayGetInfo
9353 */
9354CUresult CUDAAPI cuArray3DGetDescriptor(CUDA_ARRAY3D_DESCRIPTOR *pArrayDescriptor, CUarray hArray);
9355
9356/**
9357 * \brief Creates a CUDA mipmapped array
9358 *
9359 * Creates a CUDA mipmapped array according to the ::CUDA_ARRAY3D_DESCRIPTOR structure
9360 * \p pMipmappedArrayDesc and returns a handle to the new CUDA mipmapped array in \p *pHandle.
9361 * \p numMipmapLevels specifies the number of mipmap levels to be allocated. This value is
9362 * clamped to the range [1, 1 + floor(log2(max(width, height, depth)))].
9363 *
9364 * The ::CUDA_ARRAY3D_DESCRIPTOR is defined as:
9365 *
9366 * \code
9367 typedef struct {
9368 unsigned int Width;
9369 unsigned int Height;
9370 unsigned int Depth;
9371 CUarray_format Format;
9372 unsigned int NumChannels;
9373 unsigned int Flags;
9374 } CUDA_ARRAY3D_DESCRIPTOR;
9375 * \endcode
9376 * where:
9377 *
9378 * - \p Width, \p Height, and \p Depth are the width, height, and depth of the
9379 * CUDA array (in elements); the following types of CUDA arrays can be allocated:
9380 * - A 1D mipmapped array is allocated if \p Height and \p Depth extents are both zero.
9381 * - A 2D mipmapped array is allocated if only \p Depth extent is zero.
9382 * - A 3D mipmapped array is allocated if all three extents are non-zero.
9383 * - A 1D layered CUDA mipmapped array is allocated if only \p Height is zero and the
9384 * ::CUDA_ARRAY3D_LAYERED flag is set. Each layer is a 1D array. The number
9385 * of layers is determined by the depth extent.
9386 * - A 2D layered CUDA mipmapped array is allocated if all three extents are non-zero and
9387 * the ::CUDA_ARRAY3D_LAYERED flag is set. Each layer is a 2D array. The number
9388 * of layers is determined by the depth extent.
9389 * - A cubemap CUDA mipmapped array is allocated if all three extents are non-zero and the
9390 * ::CUDA_ARRAY3D_CUBEMAP flag is set. \p Width must be equal to \p Height, and
9391 * \p Depth must be six. A cubemap is a special type of 2D layered CUDA array,
9392 * where the six layers represent the six faces of a cube. The order of the six
9393 * layers in memory is the same as that listed in ::CUarray_cubemap_face.
9394 * - A cubemap layered CUDA mipmapped array is allocated if all three extents are non-zero,
9395 * and both, ::CUDA_ARRAY3D_CUBEMAP and ::CUDA_ARRAY3D_LAYERED flags are set.
9396 * \p Width must be equal to \p Height, and \p Depth must be a multiple of six.
9397 * A cubemap layered CUDA array is a special type of 2D layered CUDA array that
9398 * consists of a collection of cubemaps. The first six layers represent the first
9399 * cubemap, the next six layers form the second cubemap, and so on.
9400 *
9401 * - ::Format specifies the format of the elements; ::CUarray_format is
9402 * defined as:
9403 * \code
9404 typedef enum CUarray_format_enum {
9405 CU_AD_FORMAT_UNSIGNED_INT8 = 0x01,
9406 CU_AD_FORMAT_UNSIGNED_INT16 = 0x02,
9407 CU_AD_FORMAT_UNSIGNED_INT32 = 0x03,
9408 CU_AD_FORMAT_SIGNED_INT8 = 0x08,
9409 CU_AD_FORMAT_SIGNED_INT16 = 0x09,
9410 CU_AD_FORMAT_SIGNED_INT32 = 0x0a,
9411 CU_AD_FORMAT_HALF = 0x10,
9412 CU_AD_FORMAT_FLOAT = 0x20
9413 } CUarray_format;
9414 * \endcode
9415 *
9416 * - \p NumChannels specifies the number of packed components per CUDA array
9417 * element; it may be 1, 2, or 4;
9418 *
9419 * - ::Flags may be set to
9420 * - ::CUDA_ARRAY3D_LAYERED to enable creation of layered CUDA mipmapped arrays. If this flag is set,
9421 * \p Depth specifies the number of layers, not the depth of a 3D array.
9422 * - ::CUDA_ARRAY3D_SURFACE_LDST to enable surface references to be bound to individual mipmap levels of
9423 * the CUDA mipmapped array. If this flag is not set, ::cuSurfRefSetArray will fail when attempting to
9424 * bind a mipmap level of the CUDA mipmapped array to a surface reference.
9425 * - ::CUDA_ARRAY3D_CUBEMAP to enable creation of mipmapped cubemaps. If this flag is set, \p Width must be
9426 * equal to \p Height, and \p Depth must be six. If the ::CUDA_ARRAY3D_LAYERED flag is also set,
9427 * then \p Depth must be a multiple of six.
9428 * - ::CUDA_ARRAY3D_TEXTURE_GATHER to indicate that the CUDA mipmapped array will be used for texture gather.
9429 * Texture gather can only be performed on 2D CUDA mipmapped arrays.
9430 *
9431 * \p Width, \p Height and \p Depth must meet certain size requirements as listed in the following table.
9432 * All values are specified in elements. Note that for brevity's sake, the full name of the device attribute
9433 * is not specified. For ex., TEXTURE1D_MIPMAPPED_WIDTH refers to the device attribute
9434 * ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_MIPMAPPED_WIDTH.
9435 *
9436 * <table>
9437 * <tr><td><b>CUDA array type</b></td>
9438 * <td><b>Valid extents that must always be met<br>{(width range in elements), (height range),
9439 * (depth range)}</b></td>
9440 * <td><b>Valid extents with CUDA_ARRAY3D_SURFACE_LDST set<br>
9441 * {(width range in elements), (height range), (depth range)}</b></td></tr>
9442 * <tr><td>1D</td>
9443 * <td><small>{ (1,TEXTURE1D_MIPMAPPED_WIDTH), 0, 0 }</small></td>
9444 * <td><small>{ (1,SURFACE1D_WIDTH), 0, 0 }</small></td></tr>
9445 * <tr><td>2D</td>
9446 * <td><small>{ (1,TEXTURE2D_MIPMAPPED_WIDTH), (1,TEXTURE2D_MIPMAPPED_HEIGHT), 0 }</small></td>
9447 * <td><small>{ (1,SURFACE2D_WIDTH), (1,SURFACE2D_HEIGHT), 0 }</small></td></tr>
9448 * <tr><td>3D</td>
9449 * <td><small>{ (1,TEXTURE3D_WIDTH), (1,TEXTURE3D_HEIGHT), (1,TEXTURE3D_DEPTH) }
9450 * <br>OR<br>{ (1,TEXTURE3D_WIDTH_ALTERNATE), (1,TEXTURE3D_HEIGHT_ALTERNATE),
9451 * (1,TEXTURE3D_DEPTH_ALTERNATE) }</small></td>
9452 * <td><small>{ (1,SURFACE3D_WIDTH), (1,SURFACE3D_HEIGHT),
9453 * (1,SURFACE3D_DEPTH) }</small></td></tr>
9454 * <tr><td>1D Layered</td>
9455 * <td><small>{ (1,TEXTURE1D_LAYERED_WIDTH), 0,
9456 * (1,TEXTURE1D_LAYERED_LAYERS) }</small></td>
9457 * <td><small>{ (1,SURFACE1D_LAYERED_WIDTH), 0,
9458 * (1,SURFACE1D_LAYERED_LAYERS) }</small></td></tr>
9459 * <tr><td>2D Layered</td>
9460 * <td><small>{ (1,TEXTURE2D_LAYERED_WIDTH), (1,TEXTURE2D_LAYERED_HEIGHT),
9461 * (1,TEXTURE2D_LAYERED_LAYERS) }</small></td>
9462 * <td><small>{ (1,SURFACE2D_LAYERED_WIDTH), (1,SURFACE2D_LAYERED_HEIGHT),
9463 * (1,SURFACE2D_LAYERED_LAYERS) }</small></td></tr>
9464 * <tr><td>Cubemap</td>
9465 * <td><small>{ (1,TEXTURECUBEMAP_WIDTH), (1,TEXTURECUBEMAP_WIDTH), 6 }</small></td>
9466 * <td><small>{ (1,SURFACECUBEMAP_WIDTH),
9467 * (1,SURFACECUBEMAP_WIDTH), 6 }</small></td></tr>
9468 * <tr><td>Cubemap Layered</td>
9469 * <td><small>{ (1,TEXTURECUBEMAP_LAYERED_WIDTH), (1,TEXTURECUBEMAP_LAYERED_WIDTH),
9470 * (1,TEXTURECUBEMAP_LAYERED_LAYERS) }</small></td>
9471 * <td><small>{ (1,SURFACECUBEMAP_LAYERED_WIDTH), (1,SURFACECUBEMAP_LAYERED_WIDTH),
9472 * (1,SURFACECUBEMAP_LAYERED_LAYERS) }</small></td></tr>
9473 * </table>
9474 *
9475 *
9476 * \param pHandle - Returned mipmapped array
9477 * \param pMipmappedArrayDesc - mipmapped array descriptor
9478 * \param numMipmapLevels - Number of mipmap levels
9479 *
9480 * \return
9481 * ::CUDA_SUCCESS,
9482 * ::CUDA_ERROR_DEINITIALIZED,
9483 * ::CUDA_ERROR_NOT_INITIALIZED,
9484 * ::CUDA_ERROR_INVALID_CONTEXT,
9485 * ::CUDA_ERROR_INVALID_VALUE,
9486 * ::CUDA_ERROR_OUT_OF_MEMORY,
9487 * ::CUDA_ERROR_UNKNOWN
9488 * \notefnerr
9489 *
9490 * \sa
9491 * ::cuMipmappedArrayDestroy,
9492 * ::cuMipmappedArrayGetLevel,
9493 * ::cuArrayCreate,
9494 * ::cudaMallocMipmappedArray
9495 */
9496CUresult CUDAAPI cuMipmappedArrayCreate(CUmipmappedArray *pHandle, const CUDA_ARRAY3D_DESCRIPTOR *pMipmappedArrayDesc, unsigned int numMipmapLevels);
9497
9498/**
9499 * \brief Gets a mipmap level of a CUDA mipmapped array
9500 *
9501 * Returns in \p *pLevelArray a CUDA array that represents a single mipmap level
9502 * of the CUDA mipmapped array \p hMipmappedArray.
9503 *
9504 * If \p level is greater than the maximum number of levels in this mipmapped array,
9505 * ::CUDA_ERROR_INVALID_VALUE is returned.
9506 *
9507 * \param pLevelArray - Returned mipmap level CUDA array
9508 * \param hMipmappedArray - CUDA mipmapped array
9509 * \param level - Mipmap level
9510 *
9511 * \return
9512 * ::CUDA_SUCCESS,
9513 * ::CUDA_ERROR_DEINITIALIZED,
9514 * ::CUDA_ERROR_NOT_INITIALIZED,
9515 * ::CUDA_ERROR_INVALID_CONTEXT,
9516 * ::CUDA_ERROR_INVALID_VALUE,
9517 * ::CUDA_ERROR_INVALID_HANDLE
9518 * \notefnerr
9519 *
9520 * \sa
9521 * ::cuMipmappedArrayCreate,
9522 * ::cuMipmappedArrayDestroy,
9523 * ::cuArrayCreate,
9524 * ::cudaGetMipmappedArrayLevel
9525 */
9526CUresult CUDAAPI cuMipmappedArrayGetLevel(CUarray *pLevelArray, CUmipmappedArray hMipmappedArray, unsigned int level);
9527
9528/**
9529 * \brief Destroys a CUDA mipmapped array
9530 *
9531 * Destroys the CUDA mipmapped array \p hMipmappedArray.
9532 *
9533 * \param hMipmappedArray - Mipmapped array to destroy
9534 *
9535 * \return
9536 * ::CUDA_SUCCESS,
9537 * ::CUDA_ERROR_DEINITIALIZED,
9538 * ::CUDA_ERROR_NOT_INITIALIZED,
9539 * ::CUDA_ERROR_INVALID_CONTEXT,
9540 * ::CUDA_ERROR_INVALID_HANDLE,
9541 * ::CUDA_ERROR_ARRAY_IS_MAPPED,
9542 * ::CUDA_ERROR_CONTEXT_IS_DESTROYED
9543 * \notefnerr
9544 *
9545 * \sa
9546 * ::cuMipmappedArrayCreate,
9547 * ::cuMipmappedArrayGetLevel,
9548 * ::cuArrayCreate,
9549 * ::cudaFreeMipmappedArray
9550 */
9551CUresult CUDAAPI cuMipmappedArrayDestroy(CUmipmappedArray hMipmappedArray);
9552
9553/** @} */ /* END CUDA_MEM */
9554
9555/**
9556 * \defgroup CUDA_VA Virtual Memory Management
9557 *
9558 * ___MANBRIEF___ virtual memory management functions of the low-level CUDA driver API
9559 * (___CURRENT_FILE___) ___ENDMANBRIEF___
9560 *
9561 * This section describes the virtual memory management functions of the low-level CUDA
9562 * driver application programming interface.
9563 *
9564 * @{
9565 */
9566
9567/**
9568* \brief Allocate an address range reservation.
9569*
9570* Reserves a virtual address range based on the given parameters, giving
9571* the starting address of the range in \p ptr. This API requires a system that
9572* supports UVA. The size and address parameters must be a multiple of the
9573* host page size and the alignment must be a power of two or zero for default
9574* alignment.
9575*
9576* \param[out] ptr - Resulting pointer to start of virtual address range allocated
9577* \param[in] size - Size of the reserved virtual address range requested
9578* \param[in] alignment - Alignment of the reserved virtual address range requested
9579* \param[in] addr - Fixed starting address range requested
9580* \param[in] flags - Currently unused, must be zero
9581* \return
9582* ::CUDA_SUCCESS,
9583* ::CUDA_ERROR_INVALID_VALUE,
9584* ::CUDA_ERROR_OUT_OF_MEMORY,
9585* ::CUDA_ERROR_NOT_INITIALIZED,
9586* ::CUDA_ERROR_DEINITIALIZED,
9587* ::CUDA_ERROR_NOT_PERMITTED,
9588* ::CUDA_ERROR_NOT_SUPPORTED
9589*
9590* \sa ::cuMemAddressFree
9591*/
9592CUresult CUDAAPI cuMemAddressReserve(CUdeviceptr *ptr, size_t size, size_t alignment, CUdeviceptr addr, unsigned long long flags);
9593
9594/**
9595* \brief Free an address range reservation.
9596*
9597* Frees a virtual address range reserved by cuMemAddressReserve. The size
9598* must match what was given to memAddressReserve and the ptr given must
9599* match what was returned from memAddressReserve.
9600*
9601* \param[in] ptr - Starting address of the virtual address range to free
9602* \param[in] size - Size of the virtual address region to free
9603* \return
9604* ::CUDA_SUCCESS,
9605* ::CUDA_ERROR_INVALID_VALUE,
9606* ::CUDA_ERROR_NOT_INITIALIZED,
9607* ::CUDA_ERROR_DEINITIALIZED,
9608* ::CUDA_ERROR_NOT_PERMITTED,
9609* ::CUDA_ERROR_NOT_SUPPORTED
9610*
9611* \sa ::cuMemAddressReserve
9612*/
9613CUresult CUDAAPI cuMemAddressFree(CUdeviceptr ptr, size_t size);
9614
9615/**
9616* \brief Create a CUDA memory handle representing a memory allocation of a given size described by the given properties
9617*
9618* This creates a memory allocation on the target device specified through the
9619* \p prop structure. The created allocation will not have any device or host
9620* mappings. The generic memory \p handle for the allocation can be
9621* mapped to the address space of calling process via ::cuMemMap. This handle
9622* cannot be transmitted directly to other processes (see
9623* ::cuMemExportToShareableHandle). On Windows, the caller must also pass
9624* an LPSECURITYATTRIBUTE in \p prop to be associated with this handle which
9625* limits or allows access to this handle for a recipient process (see
9626* ::CUmemAllocationProp::win32HandleMetaData for more). The \p size of this
9627* allocation must be a multiple of the the value given via
9628* ::cuMemGetAllocationGranularity with the ::CU_MEM_ALLOC_GRANULARITY_MINIMUM
9629* flag.
9630* If ::CUmemAllocationProp::allocFlags::usage contains ::CU_MEM_CREATE_USAGE_TILE_POOL flag then
9631* the memory allocation is intended only to be used as backing tile pool for sparse CUDA arrays
9632* and sparse CUDA mipmapped arrays.
9633* (see ::cuMemMapArrayAsync).
9634*
9635* \param[out] handle - Value of handle returned. All operations on this allocation are to be performed using this handle.
9636* \param[in] size - Size of the allocation requested
9637* \param[in] prop - Properties of the allocation to create.
9638* \param[in] flags - flags for future use, must be zero now.
9639* \return
9640* ::CUDA_SUCCESS,
9641* ::CUDA_ERROR_INVALID_VALUE,
9642* ::CUDA_ERROR_OUT_OF_MEMORY,
9643* ::CUDA_ERROR_INVALID_DEVICE,
9644* ::CUDA_ERROR_NOT_INITIALIZED,
9645* ::CUDA_ERROR_DEINITIALIZED,
9646* ::CUDA_ERROR_NOT_PERMITTED,
9647* ::CUDA_ERROR_NOT_SUPPORTED
9648* \notefnerr
9649*
9650* \sa ::cuMemRelease, ::cuMemExportToShareableHandle, ::cuMemImportFromShareableHandle
9651*/
9652CUresult CUDAAPI cuMemCreate(CUmemGenericAllocationHandle *handle, size_t size, const CUmemAllocationProp *prop, unsigned long long flags);
9653
9654/**
9655* \brief Release a memory handle representing a memory allocation which was previously allocated through cuMemCreate.
9656*
9657* Frees the memory that was allocated on a device through cuMemCreate.
9658*
9659* The memory allocation will be freed when all outstanding mappings to the memory
9660* are unmapped and when all outstanding references to the handle (including it's
9661* shareable counterparts) are also released. The generic memory handle can be
9662* freed when there are still outstanding mappings made with this handle. Each
9663* time a recipient process imports a shareable handle, it needs to pair it with
9664* ::cuMemRelease for the handle to be freed. If \p handle is not a valid handle
9665* the behavior is undefined.
9666*
9667* \param[in] handle Value of handle which was returned previously by cuMemCreate.
9668* \return
9669* ::CUDA_SUCCESS,
9670* ::CUDA_ERROR_INVALID_VALUE,
9671* ::CUDA_ERROR_NOT_INITIALIZED,
9672* ::CUDA_ERROR_DEINITIALIZED,
9673* ::CUDA_ERROR_NOT_PERMITTED,
9674* ::CUDA_ERROR_NOT_SUPPORTED
9675* \notefnerr
9676*
9677* \sa ::cuMemCreate
9678*/
9679CUresult CUDAAPI cuMemRelease(CUmemGenericAllocationHandle handle);
9680
9681/**
9682* \brief Maps an allocation handle to a reserved virtual address range.
9683*
9684* Maps bytes of memory represented by \p handle starting from byte \p offset to
9685* \p size to address range [\p addr, \p addr + \p size]. This range must be an
9686* address reservation previously reserved with ::cuMemAddressReserve, and
9687* \p offset + \p size must be less than the size of the memory allocation.
9688* Both \p ptr, \p size, and \p offset must be a multiple of the value given via
9689* ::cuMemGetAllocationGranularity with the ::CU_MEM_ALLOC_GRANULARITY_MINIMUM flag.
9690*
9691* Please note calling ::cuMemMap does not make the address accessible,
9692* the caller needs to update accessibility of a contiguous mapped VA
9693* range by calling ::cuMemSetAccess.
9694*
9695* Once a recipient process obtains a shareable memory handle
9696* from ::cuMemImportFromShareableHandle, the process must
9697* use ::cuMemMap to map the memory into its address ranges before
9698* setting accessibility with ::cuMemSetAccess.
9699*
9700* ::cuMemMap can only create mappings on VA range reservations
9701* that are not currently mapped.
9702*
9703* \param[in] ptr - Address where memory will be mapped.
9704* \param[in] size - Size of the memory mapping.
9705* \param[in] offset - Offset into the memory represented by
9706* - \p handle from which to start mapping
9707* - Note: currently must be zero.
9708* \param[in] handle - Handle to a shareable memory
9709* \param[in] flags - flags for future use, must be zero now.
9710* \return
9711* ::CUDA_SUCCESS,
9712* ::CUDA_ERROR_INVALID_VALUE,
9713* ::CUDA_ERROR_INVALID_DEVICE,
9714* ::CUDA_ERROR_OUT_OF_MEMORY,
9715* ::CUDA_ERROR_NOT_INITIALIZED,
9716* ::CUDA_ERROR_DEINITIALIZED,
9717* ::CUDA_ERROR_NOT_PERMITTED,
9718* ::CUDA_ERROR_NOT_SUPPORTED
9719* \notefnerr
9720*
9721* \sa ::cuMemUnmap, ::cuMemSetAccess, ::cuMemCreate, ::cuMemAddressReserve, ::cuMemImportFromShareableHandle
9722*/
9723CUresult CUDAAPI cuMemMap(CUdeviceptr ptr, size_t size, size_t offset, CUmemGenericAllocationHandle handle, unsigned long long flags);
9724
9725/**
9726 * \brief Maps or unmaps subregions of sparse CUDA arrays and sparse CUDA mipmapped arrays
9727 *
9728 * Performs map or unmap operations on subregions of sparse CUDA arrays and sparse CUDA mipmapped arrays.
9729 * Each operation is specified by a ::CUarrayMapInfo entry in the \p mapInfoList array of size \p count.
9730 * The structure ::CUarrayMapInfo is defined as follow:
9731 \code
9732 typedef struct CUarrayMapInfo_st {
9733 CUresourcetype resourceType;
9734 union {
9735 CUmipmappedArray mipmap;
9736 CUarray array;
9737 } resource;
9738
9739 CUarraySparseSubresourceType subresourceType;
9740 union {
9741 struct {
9742 unsigned int level;
9743 unsigned int layer;
9744 unsigned int offsetX;
9745 unsigned int offsetY;
9746 unsigned int offsetZ;
9747 unsigned int extentWidth;
9748 unsigned int extentHeight;
9749 unsigned int extentDepth;
9750 } sparseLevel;
9751 struct {
9752 unsigned int layer;
9753 unsigned long long offset;
9754 unsigned long long size;
9755 } miptail;
9756 } subresource;
9757
9758 CUmemOperationType memOperationType;
9759
9760 CUmemHandleType memHandleType;
9761 union {
9762 CUmemGenericAllocationHandle memHandle;
9763 } memHandle;
9764
9765 unsigned long long offset;
9766 unsigned int deviceBitMask;
9767 unsigned int flags;
9768 unsigned int reserved[2];
9769 } CUarrayMapInfo;
9770 \endcode
9771 *
9772 * where ::CUarrayMapInfo::resourceType specifies the type of resource to be operated on.
9773 * If ::CUarrayMapInfo::resourceType is set to ::CUresourcetype::CU_RESOURCE_TYPE_ARRAY then
9774 * ::CUarrayMapInfo::resource::array must be set to a valid sparse CUDA array handle.
9775 * The CUDA array must be either a 2D, 2D layered or 3D CUDA array and must have been allocated using
9776 * ::cuArrayCreate or ::cuArray3DCreate with the flag ::CUDA_ARRAY3D_SPARSE.
9777 * For CUDA arrays obtained using ::cuMipmappedArrayGetLevel, ::CUDA_ERROR_INVALID_VALUE will be returned.
9778 * If ::CUarrayMapInfo::resourceType is set to ::CUresourcetype::CU_RESOURCE_TYPE_MIPMAPPED_ARRAY
9779 * then ::CUarrayMapInfo::resource::mipmap must be set to a valid sparse CUDA mipmapped array handle.
9780 * The CUDA mipmapped array must be either a 2D, 2D layered or 3D CUDA mipmapped array and must have been
9781 * allocated using ::cuMipmappedArrayCreate with the flag ::CUDA_ARRAY3D_SPARSE.
9782 *
9783 * ::CUarrayMapInfo::subresourceType specifies the type of subresource within the resource.
9784 * ::CUarraySparseSubresourceType_enum is defined as:
9785 \code
9786 typedef enum CUarraySparseSubresourceType_enum {
9787 CU_ARRAY_SPARSE_SUBRESOURCE_TYPE_SPARSE_LEVEL = 0,
9788 CU_ARRAY_SPARSE_SUBRESOURCE_TYPE_MIPTAIL = 1
9789 } CUarraySparseSubresourceType;
9790 \endcode
9791 *
9792 * where ::CUarraySparseSubresourceType::CU_ARRAY_SPARSE_SUBRESOURCE_TYPE_SPARSE_LEVEL indicates a
9793 * sparse-miplevel which spans at least one tile in every dimension. The remaining miplevels which
9794 * are too small to span at least one tile in any dimension constitute the mip tail region as indicated by
9795 * ::CUarraySparseSubresourceType::CU_ARRAY_SPARSE_SUBRESOURCE_TYPE_MIPTAIL subresource type.
9796 *
9797 * If ::CUarrayMapInfo::subresourceType is set to ::CUarraySparseSubresourceType::CU_ARRAY_SPARSE_SUBRESOURCE_TYPE_SPARSE_LEVEL
9798 * then ::CUarrayMapInfo::subresource::sparseLevel struct must contain valid array subregion offsets and extents.
9799 * The ::CUarrayMapInfo::subresource::sparseLevel::offsetX, ::CUarrayMapInfo::subresource::sparseLevel::offsetY
9800 * and ::CUarrayMapInfo::subresource::sparseLevel::offsetZ must specify valid X, Y and Z offsets respectively.
9801 * The ::CUarrayMapInfo::subresource::sparseLevel::extentWidth, ::CUarrayMapInfo::subresource::sparseLevel::extentHeight
9802 * and ::CUarrayMapInfo::subresource::sparseLevel::extentDepth must specify valid width, height and depth extents respectively.
9803 * These offsets and extents must be aligned to the corresponding tile dimension.
9804 * For CUDA mipmapped arrays ::CUarrayMapInfo::subresource::sparseLevel::level must specify a valid mip level index. Otherwise,
9805 * must be zero.
9806 * For layered CUDA arrays and layered CUDA mipmapped arrays ::CUarrayMapInfo::subresource::sparseLevel::layer must specify a valid layer index. Otherwise,
9807 * must be zero.
9808 * ::CUarrayMapInfo::subresource::sparseLevel::offsetZ must be zero and ::CUarrayMapInfo::subresource::sparseLevel::extentDepth
9809 * must be set to 1 for 2D and 2D layered CUDA arrays and CUDA mipmapped arrays.
9810 * Tile extents can be obtained by calling ::cuArrayGetSparseProperties and ::cuMipmappedArrayGetSparseProperties
9811 *
9812 * If ::CUarrayMapInfo::subresourceType is set to ::CUarraySparseSubresourceType::CU_ARRAY_SPARSE_SUBRESOURCE_TYPE_MIPTAIL
9813 * then ::CUarrayMapInfo::subresource::miptail struct must contain valid mip tail offset in
9814 * ::CUarrayMapInfo::subresource::miptail::offset and size in ::CUarrayMapInfo::subresource::miptail::size.
9815 * Both, mip tail offset and mip tail size must be aligned to the tile size.
9816 * For layered CUDA mipmapped arrays which don't have the flag ::CU_ARRAY_SPARSE_PROPERTIES_SINGLE_MIPTAIL set in ::CUDA_ARRAY_SPARSE_PROPERTIES::flags
9817 * as returned by ::cuMipmappedArrayGetSparseProperties, ::CUarrayMapInfo::subresource::miptail::layer must specify a valid layer index.
9818 * Otherwise, must be zero.
9819 *
9820 * ::CUarrayMapInfo::memOperationType specifies the type of operation. ::CUmemOperationType is defined as:
9821 \code
9822 typedef enum CUmemOperationType_enum {
9823 CU_MEM_OPERATION_TYPE_MAP = 1,
9824 CU_MEM_OPERATION_TYPE_UNMAP = 2
9825 } CUmemOperationType;
9826 \endcode
9827 * If ::CUarrayMapInfo::memOperationType is set to ::CUmemOperationType::CU_MEM_OPERATION_TYPE_MAP then the subresource
9828 * will be mapped onto the tile pool memory specified by ::CUarrayMapInfo::memHandle at offset ::CUarrayMapInfo::offset.
9829 * The tile pool allocation has to be created by specifying the ::CU_MEM_CREATE_USAGE_TILE_POOL flag when calling ::cuMemCreate. Also,
9830 * ::CUarrayMapInfo::memHandleType must be set to ::CUmemHandleType::CU_MEM_HANDLE_TYPE_GENERIC.
9831 *
9832 * If ::CUarrayMapInfo::memOperationType is set to ::CUmemOperationType::CU_MEM_OPERATION_TYPE_UNMAP then an unmapping operation
9833 * is performed. ::CUarrayMapInfo::memHandle must be NULL.
9834 *
9835 * ::CUarrayMapInfo::deviceBitMask specifies the list of devices that must map or unmap physical memory.
9836 * Currently, this mask must have exactly one bit set, and the corresponding device must match the device associated with the stream.
9837 * If ::CUarrayMapInfo::memOperationType is set to ::CUmemOperationType::CU_MEM_OPERATION_TYPE_MAP, the device must also match
9838 * the device associated with the tile pool memory allocation as specified by ::CUarrayMapInfo::memHandle.
9839 *
9840 * ::CUarrayMapInfo::flags and ::CUarrayMapInfo::reserved[] are unused and must be set to zero.
9841 *
9842 * \return
9843 * ::CUDA_SUCCESS,
9844 * ::CUDA_ERROR_INVALID_VALUE,
9845 * ::CUDA_ERROR_INVALID_HANDLE
9846 *
9847 * \param[in] mapInfoList - List of ::CUarrayMapInfo
9848 * \param[in] count - Count of ::CUarrayMapInfo in \p mapInfoList
9849 * \param[in] hStream - Stream identifier for the stream to use for map or unmap operations
9850 *
9851 * \sa ::cuMipmappedArrayCreate, ::cuArrayCreate, ::cuArray3DCreate, ::cuMemCreate, ::cuArrayGetSparseProperties, ::cuMipmappedArrayGetSparseProperties
9852 */
9853CUresult CUDAAPI cuMemMapArrayAsync(CUarrayMapInfo *mapInfoList, unsigned int count, CUstream hStream);
9854
9855/**
9856* \brief Unmap the backing memory of a given address range.
9857*
9858* The range must be the entire contiguous address range that was mapped to. In
9859* other words, ::cuMemUnmap cannot unmap a sub-range of an address range mapped
9860* by ::cuMemCreate / ::cuMemMap. Any backing memory allocations will be freed
9861* if there are no existing mappings and there are no unreleased memory handles.
9862*
9863* When ::cuMemUnmap returns successfully the address range is converted to an
9864* address reservation and can be used for a future calls to ::cuMemMap. Any new
9865* mapping to this virtual address will need to have access granted through
9866* ::cuMemSetAccess, as all mappings start with no accessibility setup.
9867*
9868* \param[in] ptr - Starting address for the virtual address range to unmap
9869* \param[in] size - Size of the virtual address range to unmap
9870* \returns
9871* ::CUDA_SUCCESS,
9872* ::CUDA_ERROR_INVALID_VALUE,
9873* ::CUDA_ERROR_NOT_INITIALIZED,
9874* ::CUDA_ERROR_DEINITIALIZED,
9875* ::CUDA_ERROR_NOT_PERMITTED,
9876* ::CUDA_ERROR_NOT_SUPPORTED
9877* \notefnerr
9878* \note_sync
9879*
9880* \sa ::cuMemCreate, ::cuMemAddressReserve
9881*/
9882CUresult CUDAAPI cuMemUnmap(CUdeviceptr ptr, size_t size);
9883
9884/**
9885* \brief Set the access flags for each location specified in \p desc for the given virtual address range
9886*
9887* Given the virtual address range via \p ptr and \p size, and the locations
9888* in the array given by \p desc and \p count, set the access flags for the
9889* target locations. The range must be a fully mapped address range
9890* containing all allocations created by ::cuMemMap / ::cuMemCreate.
9891*
9892* \param[in] ptr - Starting address for the virtual address range
9893* \param[in] size - Length of the virtual address range
9894* \param[in] desc - Array of ::CUmemAccessDesc that describe how to change the
9895* - mapping for each location specified
9896* \param[in] count - Number of ::CUmemAccessDesc in \p desc
9897* \returns
9898* ::CUDA_SUCCESS,
9899* ::CUDA_ERROR_INVALID_VALUE,
9900* ::CUDA_ERROR_INVALID_DEVICE,
9901* ::CUDA_ERROR_NOT_SUPPORTED
9902* \notefnerr
9903* \note_sync
9904*
9905* \sa ::cuMemSetAccess, ::cuMemCreate, :cuMemMap
9906*/
9907CUresult CUDAAPI cuMemSetAccess(CUdeviceptr ptr, size_t size, const CUmemAccessDesc *desc, size_t count);
9908
9909/**
9910* \brief Get the access \p flags set for the given \p location and \p ptr
9911*
9912* \param[out] flags - Flags set for this location
9913* \param[in] location - Location in which to check the flags for
9914* \param[in] ptr - Address in which to check the access flags for
9915* \returns
9916* ::CUDA_SUCCESS,
9917* ::CUDA_ERROR_INVALID_VALUE,
9918* ::CUDA_ERROR_INVALID_DEVICE,
9919* ::CUDA_ERROR_NOT_INITIALIZED,
9920* ::CUDA_ERROR_DEINITIALIZED,
9921* ::CUDA_ERROR_NOT_PERMITTED,
9922* ::CUDA_ERROR_NOT_SUPPORTED
9923*
9924* \sa ::cuMemSetAccess
9925*/
9926CUresult CUDAAPI cuMemGetAccess(unsigned long long *flags, const CUmemLocation *location, CUdeviceptr ptr);
9927
9928/**
9929* \brief Exports an allocation to a requested shareable handle type
9930*
9931* Given a CUDA memory handle, create a shareable memory
9932* allocation handle that can be used to share the memory with other
9933* processes. The recipient process can convert the shareable handle back into a
9934* CUDA memory handle using ::cuMemImportFromShareableHandle and map
9935* it with ::cuMemMap. The implementation of what this handle is and how it
9936* can be transferred is defined by the requested handle type in \p handleType
9937*
9938* Once all shareable handles are closed and the allocation is released, the allocated
9939* memory referenced will be released back to the OS and uses of the CUDA handle afterward
9940* will lead to undefined behavior.
9941*
9942* This API can also be used in conjunction with other APIs (e.g. Vulkan, OpenGL)
9943* that support importing memory from the shareable type
9944*
9945* \param[out] shareableHandle - Pointer to the location in which to store the requested handle type
9946* \param[in] handle - CUDA handle for the memory allocation
9947* \param[in] handleType - Type of shareable handle requested (defines type and size of the \p shareableHandle output parameter)
9948* \param[in] flags - Reserved, must be zero
9949* \returns
9950* ::CUDA_SUCCESS,
9951* ::CUDA_ERROR_INVALID_VALUE,
9952* ::CUDA_ERROR_NOT_INITIALIZED,
9953* ::CUDA_ERROR_DEINITIALIZED,
9954* ::CUDA_ERROR_NOT_PERMITTED,
9955* ::CUDA_ERROR_NOT_SUPPORTED
9956*
9957* \sa ::cuMemImportFromShareableHandle
9958*/
9959CUresult CUDAAPI cuMemExportToShareableHandle(void *shareableHandle, CUmemGenericAllocationHandle handle, CUmemAllocationHandleType handleType, unsigned long long flags);
9960
9961/**
9962* \brief Imports an allocation from a requested shareable handle type.
9963*
9964* If the current process cannot support the memory described by this shareable
9965* handle, this API will error as CUDA_ERROR_NOT_SUPPORTED.
9966*
9967* \note Importing shareable handles exported from some graphics APIs(VUlkan, OpenGL, etc)
9968* created on devices under an SLI group may not be supported, and thus this API will
9969* return CUDA_ERROR_NOT_SUPPORTED.
9970* There is no guarantee that the contents of \p handle will be the same CUDA memory handle
9971* for the same given OS shareable handle, or the same underlying allocation.
9972*
9973* \param[out] handle - CUDA Memory handle for the memory allocation.
9974* \param[in] osHandle - Shareable Handle representing the memory allocation that is to be imported.
9975* \param[in] shHandleType - handle type of the exported handle ::CUmemAllocationHandleType.
9976* \returns
9977* ::CUDA_SUCCESS,
9978* ::CUDA_ERROR_INVALID_VALUE,
9979* ::CUDA_ERROR_NOT_INITIALIZED,
9980* ::CUDA_ERROR_DEINITIALIZED,
9981* ::CUDA_ERROR_NOT_PERMITTED,
9982* ::CUDA_ERROR_NOT_SUPPORTED
9983*
9984* \sa ::cuMemExportToShareableHandle, ::cuMemMap, ::cuMemRelease
9985*/
9986CUresult CUDAAPI cuMemImportFromShareableHandle(CUmemGenericAllocationHandle *handle, void *osHandle, CUmemAllocationHandleType shHandleType);
9987
9988/**
9989* \brief Calculates either the minimal or recommended granularity
9990*
9991* Calculates either the minimal or recommended granularity
9992* for a given allocation specification and returns it in granularity. This
9993* granularity can be used as a multiple for alignment, size, or address mapping.
9994*
9995* \param[out] granularity Returned granularity.
9996* \param[in] prop Property for which to determine the granularity for
9997* \param[in] option Determines which granularity to return
9998* \returns
9999* ::CUDA_SUCCESS,
10000* ::CUDA_ERROR_INVALID_VALUE,
10001* ::CUDA_ERROR_NOT_INITIALIZED,
10002* ::CUDA_ERROR_DEINITIALIZED,
10003* ::CUDA_ERROR_NOT_PERMITTED,
10004* ::CUDA_ERROR_NOT_SUPPORTED
10005*
10006* \sa ::cuMemCreate, ::cuMemMap
10007*/
10008CUresult CUDAAPI cuMemGetAllocationGranularity(size_t *granularity, const CUmemAllocationProp *prop, CUmemAllocationGranularity_flags option);
10009
10010/**
10011* \brief Retrieve the contents of the property structure defining properties for this handle
10012*
10013* \param[out] prop - Pointer to a properties structure which will hold the information about this handle
10014* \param[in] handle - Handle which to perform the query on
10015* \returns
10016* ::CUDA_SUCCESS,
10017* ::CUDA_ERROR_INVALID_VALUE,
10018* ::CUDA_ERROR_NOT_INITIALIZED,
10019* ::CUDA_ERROR_DEINITIALIZED,
10020* ::CUDA_ERROR_NOT_PERMITTED,
10021* ::CUDA_ERROR_NOT_SUPPORTED
10022*
10023* \sa ::cuMemCreate, ::cuMemImportFromShareableHandle
10024*/
10025CUresult CUDAAPI cuMemGetAllocationPropertiesFromHandle(CUmemAllocationProp *prop, CUmemGenericAllocationHandle handle);
10026
10027/**
10028* \brief Given an address \p addr, returns the allocation handle of the backing memory allocation.
10029*
10030* The handle is guaranteed to be the same handle value used to map the memory. If the address
10031* requested is not mapped, the function will fail. The returned handle must be released with
10032* corresponding number of calls to ::cuMemRelease.
10033*
10034* \note The address \p addr, can be any address in a range previously mapped
10035* by ::cuMemMap, and not necessarily the start address.
10036*
10037* \param[out] handle CUDA Memory handle for the backing memory allocation.
10038* \param[in] addr Memory address to query, that has been mapped previously.
10039* \returns
10040* ::CUDA_SUCCESS,
10041* ::CUDA_ERROR_INVALID_VALUE,
10042* ::CUDA_ERROR_NOT_INITIALIZED,
10043* ::CUDA_ERROR_DEINITIALIZED,
10044* ::CUDA_ERROR_NOT_PERMITTED,
10045* ::CUDA_ERROR_NOT_SUPPORTED
10046*
10047* \sa ::cuMemCreate, ::cuMemRelease, ::cuMemMap
10048*/
10049CUresult CUDAAPI cuMemRetainAllocationHandle(CUmemGenericAllocationHandle *handle, void *addr);
10050
10051/** @} */ /* END CUDA_VA */
10052
10053/**
10054 * \defgroup CUDA_MALLOC_ASYNC Stream Ordered Memory Allocator
10055 *
10056 * ___MANBRIEF___ Functions for performing allocation and free operations in stream order.
10057 * Functions for controlling the behavior of the underlying allocator.
10058 * (___CURRENT_FILE___) ___ENDMANBRIEF___
10059 *
10060 * This section describes the stream ordered memory allocator exposed by the
10061 * low-level CUDA driver application programming interface.
10062 *
10063 * @{
10064 *
10065 * \section CUDA_MALLOC_ASYNC_overview overview
10066 *
10067 * The asynchronous allocator allows the user to allocate and free in stream order.
10068 * All asynchronous accesses of the allocation must happen between
10069 * the stream executions of the allocation and the free. If the memory is accessed
10070 * outside of the promised stream order, a use before allocation / use after free error
10071 * will cause undefined behavior.
10072 *
10073 * The allocator is free to reallocate the memory as long as it can guarantee
10074 * that compliant memory accesses will not overlap temporally.
10075 * The allocator may refer to internal stream ordering as well as inter-stream dependencies
10076 * (such as CUDA events and null stream dependencies) when establishing the temporal guarantee.
10077 * The allocator may also insert inter-stream dependencies to establish the temporal guarantee.
10078 *
10079 * \section CUDA_MALLOC_ASYNC_support Supported Platforms
10080 *
10081 * Whether or not a device supports the integrated stream ordered memory allocator
10082 * may be queried by calling ::cuDeviceGetAttribute() with the device attribute
10083 * ::CU_DEVICE_ATTRIBUTE_MEMORY_POOLS_SUPPORTED
10084 */
10085
10086/**
10087 * \brief Frees memory with stream ordered semantics
10088 *
10089 * Inserts a free operation into \p hStream.
10090 * The allocation must not be accessed after stream execution reaches the free.
10091 * After this API returns, accessing the memory from any subsequent work launched on the GPU
10092 * or querying its pointer attributes results in undefined behavior.
10093 *
10094 * \note During stream capture, this function results in the creation of a free node and
10095 * must therefore be passed the address of a graph allocation.
10096 *
10097 * \param dptr - memory to free
10098 * \param hStream - The stream establishing the stream ordering contract.
10099 * \returns
10100 * ::CUDA_SUCCESS,
10101 * ::CUDA_ERROR_INVALID_VALUE,
10102 * ::CUDA_ERROR_NOT_INITIALIZED,
10103 * ::CUDA_ERROR_INVALID_CONTEXT (default stream specified with no current context),
10104 * ::CUDA_ERROR_NOT_SUPPORTED
10105 */
10106CUresult CUDAAPI cuMemFreeAsync(CUdeviceptr dptr, CUstream hStream);
10107
10108/**
10109 * \brief Allocates memory with stream ordered semantics
10110 *
10111 * Inserts an allocation operation into \p hStream.
10112 * A pointer to the allocated memory is returned immediately in *dptr.
10113 * The allocation must not be accessed until the the allocation operation completes.
10114 * The allocation comes from the memory pool current to the stream's device.
10115 *
10116 * \note The default memory pool of a device contains device memory from that device.
10117 * \note Basic stream ordering allows future work submitted into the same stream to use the allocation.
10118 * Stream query, stream synchronize, and CUDA events can be used to guarantee that the allocation
10119 * operation completes before work submitted in a separate stream runs.
10120 * \note During stream capture, this function results in the creation of an allocation node. In this case,
10121 * the allocation is owned by the graph instead of the memory pool. The memory pool's properties
10122 * are used to set the node's creation parameters.
10123 *
10124 * \param[out] dptr - Returned device pointer
10125 * \param[in] bytesize - Number of bytes to allocate
10126 * \param[in] hStream - The stream establishing the stream ordering contract and the memory pool to allocate from
10127 * \returns
10128 * ::CUDA_SUCCESS,
10129 * ::CUDA_ERROR_INVALID_VALUE,
10130 * ::CUDA_ERROR_NOT_INITIALIZED,
10131 * ::CUDA_ERROR_INVALID_CONTEXT (default stream specified with no current context),
10132 * ::CUDA_ERROR_NOT_SUPPORTED,
10133 * ::CUDA_ERROR_OUT_OF_MEMORY
10134 *
10135 * \sa ::cuMemAllocFromPoolAsync, ::cuMemFreeAsync, ::cuDeviceSetMemPool,
10136 * ::cuDeviceGetDefaultMemPool, ::cuDeviceGetMemPool, ::cuMemPoolCreate,
10137 * ::cuMemPoolSetAccess, ::cuMemPoolSetAttribute
10138 */
10139CUresult CUDAAPI cuMemAllocAsync(CUdeviceptr *dptr, size_t bytesize, CUstream hStream);
10140
10141/**
10142 * \brief Tries to release memory back to the OS
10143 *
10144 * Releases memory back to the OS until the pool contains fewer than minBytesToKeep
10145 * reserved bytes, or there is no more memory that the allocator can safely release.
10146 * The allocator cannot release OS allocations that back outstanding asynchronous allocations.
10147 * The OS allocations may happen at different granularity from the user allocations.
10148 *
10149 * \note: Allocations that have not been freed count as outstanding.
10150 * \note: Allocations that have been asynchronously freed but whose completion has
10151 * not been observed on the host (eg. by a synchronize) can count as outstanding.
10152 *
10153 * \param[in] pool - The memory pool to trim
10154 * \param[in] minBytesToKeep - If the pool has less than minBytesToKeep reserved,
10155 * the TrimTo operation is a no-op. Otherwise the pool will be guaranteed to have
10156 * at least minBytesToKeep bytes reserved after the operation.
10157 * \returns
10158 * ::CUDA_SUCCESS,
10159 * ::CUDA_ERROR_NOT_INITIALIZED,
10160 * ::CUDA_ERROR_INVALID_VALUE
10161 *
10162 * \sa ::cuMemAllocAsync, ::cuMemFreeAsync, ::cuDeviceGetDefaultMemPool,
10163 * ::cuDeviceGetMemPool, ::cuMemPoolCreate
10164 */
10165CUresult CUDAAPI cuMemPoolTrimTo(CUmemoryPool pool, size_t minBytesToKeep);
10166
10167/**
10168 * \brief Sets attributes of a memory pool
10169 *
10170 * Supported attributes are:
10171 * - ::CU_MEMPOOL_ATTR_RELEASE_THRESHOLD: (value type = cuuint64_t)
10172 * Amount of reserved memory in bytes to hold onto before trying
10173 * to release memory back to the OS. When more than the release
10174 * threshold bytes of memory are held by the memory pool, the
10175 * allocator will try to release memory back to the OS on the
10176 * next call to stream, event or context synchronize. (default 0)
10177 * - ::CU_MEMPOOL_ATTR_REUSE_FOLLOW_EVENT_DEPENDENCIES: (value type = int)
10178 * Allow ::cuMemAllocAsync to use memory asynchronously freed
10179 * in another stream as long as a stream ordering dependency
10180 * of the allocating stream on the free action exists.
10181 * Cuda events and null stream interactions can create the required
10182 * stream ordered dependencies. (default enabled)
10183 * - ::CU_MEMPOOL_ATTR_REUSE_ALLOW_OPPORTUNISTIC: (value type = int)
10184 * Allow reuse of already completed frees when there is no dependency
10185 * between the free and allocation. (default enabled)
10186 * - ::CU_MEMPOOL_ATTR_REUSE_ALLOW_INTERNAL_DEPENDENCIES: (value type = int)
10187 * Allow ::cuMemAllocAsync to insert new stream dependencies
10188 * in order to establish the stream ordering required to reuse
10189 * a piece of memory released by ::cuMemFreeAsync (default enabled).
10190 * - ::CU_MEMPOOL_ATTR_RESERVED_MEM_HIGH: (value type = cuuint64_t)
10191 * Reset the high watermark that tracks the amount of backing memory that was
10192 * allocated for the memory pool. It is illegal to set this attribute to a non-zero value.
10193 * - ::CU_MEMPOOL_ATTR_USED_MEM_HIGH: (value type = cuuint64_t)
10194 * Reset the high watermark that tracks the amount of used memory that was
10195 * allocated for the memory pool.
10196 *
10197 * \param[in] pool - The memory pool to modify
10198 * \param[in] attr - The attribute to modify
10199 * \param[in] value - Pointer to the value to assign
10200 *
10201 * \returns
10202 * ::CUDA_SUCCESS,
10203 * ::CUDA_ERROR_INVALID_VALUE
10204 *
10205 * \sa ::cuMemAllocAsync, ::cuMemFreeAsync, ::cuDeviceGetDefaultMemPool,
10206 * ::cuDeviceGetMemPool, ::cuMemPoolCreate
10207 */
10208CUresult CUDAAPI cuMemPoolSetAttribute(CUmemoryPool pool, CUmemPool_attribute attr, void *value);
10209
10210/**
10211 * \brief Gets attributes of a memory pool
10212 *
10213 * Supported attributes are:
10214 * - ::CU_MEMPOOL_ATTR_RELEASE_THRESHOLD: (value type = cuuint64_t)
10215 * Amount of reserved memory in bytes to hold onto before trying
10216 * to release memory back to the OS. When more than the release
10217 * threshold bytes of memory are held by the memory pool, the
10218 * allocator will try to release memory back to the OS on the
10219 * next call to stream, event or context synchronize. (default 0)
10220 * - ::CU_MEMPOOL_ATTR_REUSE_FOLLOW_EVENT_DEPENDENCIES: (value type = int)
10221 * Allow ::cuMemAllocAsync to use memory asynchronously freed
10222 * in another stream as long as a stream ordering dependency
10223 * of the allocating stream on the free action exists.
10224 * Cuda events and null stream interactions can create the required
10225 * stream ordered dependencies. (default enabled)
10226 * - ::CU_MEMPOOL_ATTR_REUSE_ALLOW_OPPORTUNISTIC: (value type = int)
10227 * Allow reuse of already completed frees when there is no dependency
10228 * between the free and allocation. (default enabled)
10229 * - ::CU_MEMPOOL_ATTR_REUSE_ALLOW_INTERNAL_DEPENDENCIES: (value type = int)
10230 * Allow ::cuMemAllocAsync to insert new stream dependencies
10231 * in order to establish the stream ordering required to reuse
10232 * a piece of memory released by ::cuMemFreeAsync (default enabled).
10233 * - ::CU_MEMPOOL_ATTR_RESERVED_MEM_CURRENT: (value type = cuuint64_t)
10234 * Amount of backing memory currently allocated for the mempool
10235 * - ::CU_MEMPOOL_ATTR_RESERVED_MEM_HIGH: (value type = cuuint64_t)
10236 * High watermark of backing memory allocated for the mempool since the
10237 * last time it was reset.
10238 * - ::CU_MEMPOOL_ATTR_USED_MEM_CURRENT: (value type = cuuint64_t)
10239 * Amount of memory from the pool that is currently in use by the application.
10240 * - ::CU_MEMPOOL_ATTR_USED_MEM_HIGH: (value type = cuuint64_t)
10241 * High watermark of the amount of memory from the pool that was in use by the application.
10242 *
10243 * \param[in] pool - The memory pool to get attributes of
10244 * \param[in] attr - The attribute to get
10245 * \param[out] value - Retrieved value
10246 *
10247 * \returns
10248 * ::CUDA_SUCCESS,
10249 * ::CUDA_ERROR_NOT_INITIALIZED,
10250 * ::CUDA_ERROR_INVALID_VALUE
10251 *
10252 * \sa ::cuMemAllocAsync, ::cuMemFreeAsync, ::cuDeviceGetDefaultMemPool,
10253 * ::cuDeviceGetMemPool, ::cuMemPoolCreate
10254 */
10255CUresult CUDAAPI cuMemPoolGetAttribute(CUmemoryPool pool, CUmemPool_attribute attr, void *value);
10256
10257/**
10258 * \brief Controls visibility of pools between devices
10259 *
10260 * \param[in] pool - The pool being modified
10261 * \param[in] map - Array of access descriptors. Each descriptor instructs the access to enable for a single gpu.
10262 * \param[in] count - Number of descriptors in the map array.
10263 *
10264 * \returns
10265 * ::CUDA_SUCCESS,
10266 * ::CUDA_ERROR_NOT_INITIALIZED,
10267 * ::CUDA_ERROR_INVALID_VALUE
10268 *
10269 * \sa ::cuMemAllocAsync, ::cuMemFreeAsync, ::cuDeviceGetDefaultMemPool,
10270 * ::cuDeviceGetMemPool, ::cuMemPoolCreate
10271 */
10272CUresult CUDAAPI cuMemPoolSetAccess(CUmemoryPool pool, const CUmemAccessDesc *map, size_t count);
10273
10274/**
10275 * \brief Returns the accessibility of a pool from a device
10276 *
10277 * Returns the accessibility of the pool's memory from the specified location.
10278 *
10279 * \param[out] flags - the accessibility of the pool from the specified location
10280 * \param[in] memPool - the pool being queried
10281 * \param[in] location - the location accessing the pool
10282 *
10283 * \sa ::cuMemAllocAsync, ::cuMemFreeAsync, ::cuDeviceGetDefaultMemPool,
10284 * ::cuDeviceGetMemPool, ::cuMemPoolCreate
10285 */
10286CUresult CUDAAPI cuMemPoolGetAccess(CUmemAccess_flags *flags, CUmemoryPool memPool, CUmemLocation *location);
10287
10288/**
10289 * \brief Creates a memory pool
10290 *
10291 * Creates a CUDA memory pool and returns the handle in \p pool. The \p poolProps determines
10292 * the properties of the pool such as the backing device and IPC capabilities.
10293 *
10294 * By default, the pool's memory will be accessible from the device it is allocated on.
10295 *
10296 * \note Specifying CU_MEM_HANDLE_TYPE_NONE creates a memory pool that will not support IPC.
10297 *
10298 * \returns
10299 * ::CUDA_SUCCESS,
10300 * ::CUDA_ERROR_NOT_INITIALIZED,
10301 * ::CUDA_ERROR_INVALID_VALUE,
10302 * ::CUDA_ERROR_OUT_OF_MEMORY,
10303 * ::CUDA_ERROR_NOT_SUPPORTED
10304 *
10305 * \sa ::cuDeviceSetMemPool, ::cuDeviceGetMemPool, ::cuDeviceGetDefaultMemPool,
10306 * ::cuMemAllocFromPoolAsync, ::cuMemPoolExportToShareableHandle
10307 */
10308CUresult CUDAAPI cuMemPoolCreate(CUmemoryPool *pool, const CUmemPoolProps *poolProps);
10309
10310/**
10311 * \brief Destroys the specified memory pool
10312 *
10313 * If any pointers obtained from this pool haven't been freed or
10314 * the pool has free operations that haven't completed
10315 * when ::cuMemPoolDestroy is invoked, the function will return immediately and the
10316 * resources associated with the pool will be released automatically
10317 * once there are no more outstanding allocations.
10318 *
10319 * Destroying the current mempool of a device sets the default mempool of
10320 * that device as the current mempool for that device.
10321 *
10322 * \note A device's default memory pool cannot be destroyed.
10323 *
10324 * \returns
10325 * ::CUDA_SUCCESS,
10326 * ::CUDA_ERROR_INVALID_VALUE
10327 *
10328 * \sa ::cuMemFreeAsync, ::cuDeviceSetMemPool, ::cuDeviceGetMemPool,
10329 * ::cuDeviceGetDefaultMemPool, ::cuMemPoolCreate
10330 */
10331CUresult CUDAAPI cuMemPoolDestroy(CUmemoryPool pool);
10332
10333/**
10334 * \brief Allocates memory from a specified pool with stream ordered semantics.
10335 *
10336 * Inserts an allocation operation into \p hStream.
10337 * A pointer to the allocated memory is returned immediately in *dptr.
10338 * The allocation must not be accessed until the the allocation operation completes.
10339 * The allocation comes from the specified memory pool.
10340 *
10341 * \note
10342 * - The specified memory pool may be from a device different than that of the specified \p hStream.
10343 *
10344 * - Basic stream ordering allows future work submitted into the same stream to use the allocation.
10345 * Stream query, stream synchronize, and CUDA events can be used to guarantee that the allocation
10346 * operation completes before work submitted in a separate stream runs.
10347 *
10348 * \note During stream capture, this function results in the creation of an allocation node. In this case,
10349 * the allocation is owned by the graph instead of the memory pool. The memory pool's properties
10350 * are used to set the node's creation parameters.
10351 *
10352 * \param[out] dptr - Returned device pointer
10353 * \param[in] bytesize - Number of bytes to allocate
10354 * \param[in] pool - The pool to allocate from
10355 * \param[in] hStream - The stream establishing the stream ordering semantic
10356 *
10357 * \returns
10358 * ::CUDA_SUCCESS,
10359 * ::CUDA_ERROR_INVALID_VALUE,
10360 * ::CUDA_ERROR_NOT_INITIALIZED,
10361 * ::CUDA_ERROR_INVALID_CONTEXT (default stream specified with no current context),
10362 * ::CUDA_ERROR_NOT_SUPPORTED,
10363 * ::CUDA_ERROR_OUT_OF_MEMORY
10364 *
10365 * \sa ::cuMemAllocAsync, ::cuMemFreeAsync, ::cuDeviceGetDefaultMemPool,
10366 * ::cuDeviceGetMemPool, ::cuMemPoolCreate, ::cuMemPoolSetAccess,
10367 * ::cuMemPoolSetAttribute
10368 */
10369CUresult CUDAAPI cuMemAllocFromPoolAsync(CUdeviceptr *dptr, size_t bytesize, CUmemoryPool pool, CUstream hStream);
10370
10371/**
10372 * \brief Exports a memory pool to the requested handle type.
10373 *
10374 * Given an IPC capable mempool, create an OS handle to share the pool with another process.
10375 * A recipient process can convert the shareable handle into a mempool with ::cuMemPoolImportFromShareableHandle.
10376 * Individual pointers can then be shared with the ::cuMemPoolExportPointer and ::cuMemPoolImportPointer APIs.
10377 * The implementation of what the shareable handle is and how it can be transferred is defined by the requested
10378 * handle type.
10379 *
10380 * \note: To create an IPC capable mempool, create a mempool with a CUmemAllocationHandleType other than CU_MEM_HANDLE_TYPE_NONE.
10381 *
10382 * \param[out] handle_out - Returned OS handle
10383 * \param[in] pool - pool to export
10384 * \param[in] handleType - the type of handle to create
10385 * \param[in] flags - must be 0
10386 *
10387 * \returns
10388 * ::CUDA_SUCCESS,
10389 * ::CUDA_ERROR_INVALID_VALUE,
10390 * ::CUDA_ERROR_NOT_INITIALIZED,
10391 * ::CUDA_ERROR_OUT_OF_MEMORY
10392 *
10393 * \sa ::cuMemPoolImportFromShareableHandle, ::cuMemPoolExportPointer,
10394 * ::cuMemPoolImportPointer, ::cuMemAllocAsync, ::cuMemFreeAsync,
10395 * ::cuDeviceGetDefaultMemPool, ::cuDeviceGetMemPool, ::cuMemPoolCreate,
10396 * ::cuMemPoolSetAccess, ::cuMemPoolSetAttribute
10397 */
10398CUresult CUDAAPI cuMemPoolExportToShareableHandle(void *handle_out, CUmemoryPool pool, CUmemAllocationHandleType handleType, unsigned long long flags);
10399
10400/**
10401 * \brief imports a memory pool from a shared handle.
10402 *
10403 * Specific allocations can be imported from the imported pool with cuMemPoolImportPointer.
10404 *
10405 * \note Imported memory pools do not support creating new allocations.
10406 * As such imported memory pools may not be used in cuDeviceSetMemPool
10407 * or ::cuMemAllocFromPoolAsync calls.
10408 *
10409 * \param[out] pool_out - Returned memory pool
10410 * \param[in] handle - OS handle of the pool to open
10411 * \param[in] handleType - The type of handle being imported
10412 * \param[in] flags - must be 0
10413 *
10414 * \returns
10415 * ::CUDA_SUCCESS,
10416 * ::CUDA_ERROR_INVALID_VALUE,
10417 * ::CUDA_ERROR_NOT_INITIALIZED,
10418 * ::CUDA_ERROR_OUT_OF_MEMORY
10419 *
10420 * \sa ::cuMemPoolExportToShareableHandle, ::cuMemPoolExportPointer, ::cuMemPoolImportPointer
10421 */
10422CUresult CUDAAPI cuMemPoolImportFromShareableHandle(
10423 CUmemoryPool *pool_out,
10424 void *handle,
10425 CUmemAllocationHandleType handleType,
10426 unsigned long long flags);
10427
10428/**
10429 * \brief Export data to share a memory pool allocation between processes.
10430 *
10431 * Constructs \p shareData_out for sharing a specific allocation from an already shared memory pool.
10432 * The recipient process can import the allocation with the ::cuMemPoolImportPointer api.
10433 * The data is not a handle and may be shared through any IPC mechanism.
10434 *
10435 * \param[out] shareData_out - Returned export data
10436 * \param[in] ptr - pointer to memory being exported
10437 *
10438 * \returns
10439 * ::CUDA_SUCCESS,
10440 * ::CUDA_ERROR_INVALID_VALUE,
10441 * ::CUDA_ERROR_NOT_INITIALIZED,
10442 * ::CUDA_ERROR_OUT_OF_MEMORY
10443 *
10444 * \sa ::cuMemPoolExportToShareableHandle, ::cuMemPoolImportFromShareableHandle, ::cuMemPoolImportPointer
10445 */
10446CUresult CUDAAPI cuMemPoolExportPointer(CUmemPoolPtrExportData *shareData_out, CUdeviceptr ptr);
10447
10448/**
10449 * \brief Import a memory pool allocation from another process.
10450 *
10451 * Returns in \p ptr_out a pointer to the imported memory.
10452 * The imported memory must not be accessed before the allocation operation completes
10453 * in the exporting process. The imported memory must be freed from all importing processes before
10454 * being freed in the exporting process. The pointer may be freed with cuMemFree
10455 * or cuMemFreeAsync. If cuMemFreeAsync is used, the free must be completed
10456 * on the importing process before the free operation on the exporting process.
10457 *
10458 * \note The cuMemFreeAsync api may be used in the exporting process before
10459 * the cuMemFreeAsync operation completes in its stream as long as the
10460 * cuMemFreeAsync in the exporting process specifies a stream with
10461 * a stream dependency on the importing process's cuMemFreeAsync.
10462 *
10463 * \param[out] ptr_out - pointer to imported memory
10464 * \param[in] pool - pool from which to import
10465 * \param[in] shareData - data specifying the memory to import
10466 *
10467 * \returns
10468 * ::CUDA_SUCCESS,
10469 * ::CUDA_ERROR_INVALID_VALUE,
10470 * ::CUDA_ERROR_NOT_INITIALIZED,
10471 * ::CUDA_ERROR_OUT_OF_MEMORY
10472 *
10473 * \sa ::cuMemPoolExportToShareableHandle, ::cuMemPoolImportFromShareableHandle, ::cuMemPoolExportPointer
10474 */
10475CUresult CUDAAPI cuMemPoolImportPointer(CUdeviceptr *ptr_out, CUmemoryPool pool, CUmemPoolPtrExportData *shareData);
10476
10477/** @} */ /* END CUDA_MALLOC_ASYNC */
10478
10479/**
10480 * \defgroup CUDA_UNIFIED Unified Addressing
10481 *
10482 * ___MANBRIEF___ unified addressing functions of the low-level CUDA driver
10483 * API (___CURRENT_FILE___) ___ENDMANBRIEF___
10484 *
10485 * This section describes the unified addressing functions of the
10486 * low-level CUDA driver application programming interface.
10487 *
10488 * @{
10489 *
10490 * \section CUDA_UNIFIED_overview Overview
10491 *
10492 * CUDA devices can share a unified address space with the host.
10493 * For these devices there is no distinction between a device
10494 * pointer and a host pointer -- the same pointer value may be
10495 * used to access memory from the host program and from a kernel
10496 * running on the device (with exceptions enumerated below).
10497 *
10498 * \section CUDA_UNIFIED_support Supported Platforms
10499 *
10500 * Whether or not a device supports unified addressing may be
10501 * queried by calling ::cuDeviceGetAttribute() with the device
10502 * attribute ::CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING.
10503 *
10504 * Unified addressing is automatically enabled in 64-bit processes
10505 *
10506 * \section CUDA_UNIFIED_lookup Looking Up Information from Pointer Values
10507 *
10508 * It is possible to look up information about the memory which backs a
10509 * pointer value. For instance, one may want to know if a pointer points
10510 * to host or device memory. As another example, in the case of device
10511 * memory, one may want to know on which CUDA device the memory
10512 * resides. These properties may be queried using the function
10513 * ::cuPointerGetAttribute()
10514 *
10515 * Since pointers are unique, it is not necessary to specify information
10516 * about the pointers specified to the various copy functions in the
10517 * CUDA API. The function ::cuMemcpy() may be used to perform a copy
10518 * between two pointers, ignoring whether they point to host or device
10519 * memory (making ::cuMemcpyHtoD(), ::cuMemcpyDtoD(), and ::cuMemcpyDtoH()
10520 * unnecessary for devices supporting unified addressing). For
10521 * multidimensional copies, the memory type ::CU_MEMORYTYPE_UNIFIED may be
10522 * used to specify that the CUDA driver should infer the location of the
10523 * pointer from its value.
10524 *
10525 * \section CUDA_UNIFIED_automaphost Automatic Mapping of Host Allocated Host Memory
10526 *
10527 * All host memory allocated in all contexts using ::cuMemAllocHost() and
10528 * ::cuMemHostAlloc() is always directly accessible from all contexts on
10529 * all devices that support unified addressing. This is the case regardless
10530 * of whether or not the flags ::CU_MEMHOSTALLOC_PORTABLE and
10531 * ::CU_MEMHOSTALLOC_DEVICEMAP are specified.
10532 *
10533 * The pointer value through which allocated host memory may be accessed
10534 * in kernels on all devices that support unified addressing is the same
10535 * as the pointer value through which that memory is accessed on the host,
10536 * so it is not necessary to call ::cuMemHostGetDevicePointer() to get the device
10537 * pointer for these allocations.
10538 *
10539 * Note that this is not the case for memory allocated using the flag
10540 * ::CU_MEMHOSTALLOC_WRITECOMBINED, as discussed below.
10541 *
10542 * \section CUDA_UNIFIED_autopeerregister Automatic Registration of Peer Memory
10543 *
10544 * Upon enabling direct access from a context that supports unified addressing
10545 * to another peer context that supports unified addressing using
10546 * ::cuCtxEnablePeerAccess() all memory allocated in the peer context using
10547 * ::cuMemAlloc() and ::cuMemAllocPitch() will immediately be accessible
10548 * by the current context. The device pointer value through
10549 * which any peer memory may be accessed in the current context
10550 * is the same pointer value through which that memory may be
10551 * accessed in the peer context.
10552 *
10553 * \section CUDA_UNIFIED_exceptions Exceptions, Disjoint Addressing
10554 *
10555 * Not all memory may be accessed on devices through the same pointer
10556 * value through which they are accessed on the host. These exceptions
10557 * are host memory registered using ::cuMemHostRegister() and host memory
10558 * allocated using the flag ::CU_MEMHOSTALLOC_WRITECOMBINED. For these
10559 * exceptions, there exists a distinct host and device address for the
10560 * memory. The device address is guaranteed to not overlap any valid host
10561 * pointer range and is guaranteed to have the same value across all
10562 * contexts that support unified addressing.
10563 *
10564 * This device address may be queried using ::cuMemHostGetDevicePointer()
10565 * when a context using unified addressing is current. Either the host
10566 * or the unified device pointer value may be used to refer to this memory
10567 * through ::cuMemcpy() and similar functions using the
10568 * ::CU_MEMORYTYPE_UNIFIED memory type.
10569 *
10570 */
10571
10572/**
10573 * \brief Returns information about a pointer
10574 *
10575 * The supported attributes are:
10576 *
10577 * - ::CU_POINTER_ATTRIBUTE_CONTEXT:
10578 *
10579 * Returns in \p *data the ::CUcontext in which \p ptr was allocated or
10580 * registered.
10581 * The type of \p data must be ::CUcontext *.
10582 *
10583 * If \p ptr was not allocated by, mapped by, or registered with
10584 * a ::CUcontext which uses unified virtual addressing then
10585 * ::CUDA_ERROR_INVALID_VALUE is returned.
10586 *
10587 * - ::CU_POINTER_ATTRIBUTE_MEMORY_TYPE:
10588 *
10589 * Returns in \p *data the physical memory type of the memory that
10590 * \p ptr addresses as a ::CUmemorytype enumerated value.
10591 * The type of \p data must be unsigned int.
10592 *
10593 * If \p ptr addresses device memory then \p *data is set to
10594 * ::CU_MEMORYTYPE_DEVICE. The particular ::CUdevice on which the
10595 * memory resides is the ::CUdevice of the ::CUcontext returned by the
10596 * ::CU_POINTER_ATTRIBUTE_CONTEXT attribute of \p ptr.
10597 *
10598 * If \p ptr addresses host memory then \p *data is set to
10599 * ::CU_MEMORYTYPE_HOST.
10600 *
10601 * If \p ptr was not allocated by, mapped by, or registered with
10602 * a ::CUcontext which uses unified virtual addressing then
10603 * ::CUDA_ERROR_INVALID_VALUE is returned.
10604 *
10605 * If the current ::CUcontext does not support unified virtual
10606 * addressing then ::CUDA_ERROR_INVALID_CONTEXT is returned.
10607 *
10608 * - ::CU_POINTER_ATTRIBUTE_DEVICE_POINTER:
10609 *
10610 * Returns in \p *data the device pointer value through which
10611 * \p ptr may be accessed by kernels running in the current
10612 * ::CUcontext.
10613 * The type of \p data must be CUdeviceptr *.
10614 *
10615 * If there exists no device pointer value through which
10616 * kernels running in the current ::CUcontext may access
10617 * \p ptr then ::CUDA_ERROR_INVALID_VALUE is returned.
10618 *
10619 * If there is no current ::CUcontext then
10620 * ::CUDA_ERROR_INVALID_CONTEXT is returned.
10621 *
10622 * Except in the exceptional disjoint addressing cases discussed
10623 * below, the value returned in \p *data will equal the input
10624 * value \p ptr.
10625 *
10626 * - ::CU_POINTER_ATTRIBUTE_HOST_POINTER:
10627 *
10628 * Returns in \p *data the host pointer value through which
10629 * \p ptr may be accessed by by the host program.
10630 * The type of \p data must be void **.
10631 * If there exists no host pointer value through which
10632 * the host program may directly access \p ptr then
10633 * ::CUDA_ERROR_INVALID_VALUE is returned.
10634 *
10635 * Except in the exceptional disjoint addressing cases discussed
10636 * below, the value returned in \p *data will equal the input
10637 * value \p ptr.
10638 *
10639 * - ::CU_POINTER_ATTRIBUTE_P2P_TOKENS:
10640 *
10641 * Returns in \p *data two tokens for use with the nv-p2p.h Linux
10642 * kernel interface. \p data must be a struct of type
10643 * CUDA_POINTER_ATTRIBUTE_P2P_TOKENS.
10644 *
10645 * \p ptr must be a pointer to memory obtained from :cuMemAlloc().
10646 * Note that p2pToken and vaSpaceToken are only valid for the
10647 * lifetime of the source allocation. A subsequent allocation at
10648 * the same address may return completely different tokens.
10649 * Querying this attribute has a side effect of setting the attribute
10650 * ::CU_POINTER_ATTRIBUTE_SYNC_MEMOPS for the region of memory that
10651 * \p ptr points to.
10652 *
10653 * - ::CU_POINTER_ATTRIBUTE_SYNC_MEMOPS:
10654 *
10655 * A boolean attribute which when set, ensures that synchronous memory operations
10656 * initiated on the region of memory that \p ptr points to will always synchronize.
10657 * See further documentation in the section titled "API synchronization behavior"
10658 * to learn more about cases when synchronous memory operations can
10659 * exhibit asynchronous behavior.
10660 *
10661 * - ::CU_POINTER_ATTRIBUTE_BUFFER_ID:
10662 *
10663 * Returns in \p *data a buffer ID which is guaranteed to be unique within the process.
10664 * \p data must point to an unsigned long long.
10665 *
10666 * \p ptr must be a pointer to memory obtained from a CUDA memory allocation API.
10667 * Every memory allocation from any of the CUDA memory allocation APIs will
10668 * have a unique ID over a process lifetime. Subsequent allocations do not reuse IDs
10669 * from previous freed allocations. IDs are only unique within a single process.
10670 *
10671 *
10672 * - ::CU_POINTER_ATTRIBUTE_IS_MANAGED:
10673 *
10674 * Returns in \p *data a boolean that indicates whether the pointer points to
10675 * managed memory or not.
10676 *
10677 * If \p ptr is not a valid CUDA pointer then ::CUDA_ERROR_INVALID_VALUE is returned.
10678 *
10679 * - ::CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL:
10680 *
10681 * Returns in \p *data an integer representing a device ordinal of a device against
10682 * which the memory was allocated or registered.
10683 *
10684 * - ::CU_POINTER_ATTRIBUTE_IS_LEGACY_CUDA_IPC_CAPABLE:
10685 *
10686 * Returns in \p *data a boolean that indicates if this pointer maps to
10687 * an allocation that is suitable for ::cudaIpcGetMemHandle.
10688 *
10689 * - ::CU_POINTER_ATTRIBUTE_RANGE_START_ADDR:
10690 *
10691 * Returns in \p *data the starting address for the allocation referenced
10692 * by the device pointer \p ptr. Note that this is not necessarily the
10693 * address of the mapped region, but the address of the mappable address
10694 * range \p ptr references (e.g. from ::cuMemAddressReserve).
10695 *
10696 * - ::CU_POINTER_ATTRIBUTE_RANGE_SIZE:
10697 *
10698 * Returns in \p *data the size for the allocation referenced by the device
10699 * pointer \p ptr. Note that this is not necessarily the size of the mapped
10700 * region, but the size of the mappable address range \p ptr references
10701 * (e.g. from ::cuMemAddressReserve). To retrieve the size of the mapped
10702 * region, see ::cuMemGetAddressRange
10703 *
10704 * - ::CU_POINTER_ATTRIBUTE_MAPPED:
10705 *
10706 * Returns in \p *data a boolean that indicates if this pointer is in a
10707 * valid address range that is mapped to a backing allocation.
10708 *
10709 * - ::CU_POINTER_ATTRIBUTE_ALLOWED_HANDLE_TYPES:
10710 *
10711 * Returns a bitmask of the allowed handle types for an allocation that may
10712 * be passed to ::cuMemExportToShareableHandle.
10713 *
10714 * - ::CU_POINTER_ATTRIBUTE_MEMPOOL_HANDLE:
10715 *
10716 * Returns in \p *data the handle to the mempool that the allocation was obtained from.
10717 *
10718 * \par
10719 *
10720 * Note that for most allocations in the unified virtual address space
10721 * the host and device pointer for accessing the allocation will be the
10722 * same. The exceptions to this are
10723 * - user memory registered using ::cuMemHostRegister
10724 * - host memory allocated using ::cuMemHostAlloc with the
10725 * ::CU_MEMHOSTALLOC_WRITECOMBINED flag
10726 * For these types of allocation there will exist separate, disjoint host
10727 * and device addresses for accessing the allocation. In particular
10728 * - The host address will correspond to an invalid unmapped device address
10729 * (which will result in an exception if accessed from the device)
10730 * - The device address will correspond to an invalid unmapped host address
10731 * (which will result in an exception if accessed from the host).
10732 * For these types of allocations, querying ::CU_POINTER_ATTRIBUTE_HOST_POINTER
10733 * and ::CU_POINTER_ATTRIBUTE_DEVICE_POINTER may be used to retrieve the host
10734 * and device addresses from either address.
10735 *
10736 * \param data - Returned pointer attribute value
10737 * \param attribute - Pointer attribute to query
10738 * \param ptr - Pointer
10739 *
10740 * \return
10741 * ::CUDA_SUCCESS,
10742 * ::CUDA_ERROR_DEINITIALIZED,
10743 * ::CUDA_ERROR_NOT_INITIALIZED,
10744 * ::CUDA_ERROR_INVALID_CONTEXT,
10745 * ::CUDA_ERROR_INVALID_VALUE,
10746 * ::CUDA_ERROR_INVALID_DEVICE
10747 * \notefnerr
10748 *
10749 * \sa
10750 * ::cuPointerSetAttribute,
10751 * ::cuMemAlloc,
10752 * ::cuMemFree,
10753 * ::cuMemAllocHost,
10754 * ::cuMemFreeHost,
10755 * ::cuMemHostAlloc,
10756 * ::cuMemHostRegister,
10757 * ::cuMemHostUnregister,
10758 * ::cudaPointerGetAttributes
10759 */
10760CUresult CUDAAPI cuPointerGetAttribute(void *data, CUpointer_attribute attribute, CUdeviceptr ptr);
10761
10762/**
10763 * \brief Prefetches memory to the specified destination device
10764 *
10765 * Prefetches memory to the specified destination device. \p devPtr is the
10766 * base device pointer of the memory to be prefetched and \p dstDevice is the
10767 * destination device. \p count specifies the number of bytes to copy. \p hStream
10768 * is the stream in which the operation is enqueued. The memory range must refer
10769 * to managed memory allocated via ::cuMemAllocManaged or declared via __managed__ variables.
10770 *
10771 * Passing in CU_DEVICE_CPU for \p dstDevice will prefetch the data to host memory. If
10772 * \p dstDevice is a GPU, then the device attribute ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS
10773 * must be non-zero. Additionally, \p hStream must be associated with a device that has a
10774 * non-zero value for the device attribute ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS.
10775 *
10776 * The start address and end address of the memory range will be rounded down and rounded up
10777 * respectively to be aligned to CPU page size before the prefetch operation is enqueued
10778 * in the stream.
10779 *
10780 * If no physical memory has been allocated for this region, then this memory region
10781 * will be populated and mapped on the destination device. If there's insufficient
10782 * memory to prefetch the desired region, the Unified Memory driver may evict pages from other
10783 * ::cuMemAllocManaged allocations to host memory in order to make room. Device memory
10784 * allocated using ::cuMemAlloc or ::cuArrayCreate will not be evicted.
10785 *
10786 * By default, any mappings to the previous location of the migrated pages are removed and
10787 * mappings for the new location are only setup on \p dstDevice. The exact behavior however
10788 * also depends on the settings applied to this memory range via ::cuMemAdvise as described
10789 * below:
10790 *
10791 * If ::CU_MEM_ADVISE_SET_READ_MOSTLY was set on any subset of this memory range,
10792 * then that subset will create a read-only copy of the pages on \p dstDevice.
10793 *
10794 * If ::CU_MEM_ADVISE_SET_PREFERRED_LOCATION was called on any subset of this memory
10795 * range, then the pages will be migrated to \p dstDevice even if \p dstDevice is not the
10796 * preferred location of any pages in the memory range.
10797 *
10798 * If ::CU_MEM_ADVISE_SET_ACCESSED_BY was called on any subset of this memory range,
10799 * then mappings to those pages from all the appropriate processors are updated to
10800 * refer to the new location if establishing such a mapping is possible. Otherwise,
10801 * those mappings are cleared.
10802 *
10803 * Note that this API is not required for functionality and only serves to improve performance
10804 * by allowing the application to migrate data to a suitable location before it is accessed.
10805 * Memory accesses to this range are always coherent and are allowed even when the data is
10806 * actively being migrated.
10807 *
10808 * Note that this function is asynchronous with respect to the host and all work
10809 * on other devices.
10810 *
10811 * \param devPtr - Pointer to be prefetched
10812 * \param count - Size in bytes
10813 * \param dstDevice - Destination device to prefetch to
10814 * \param hStream - Stream to enqueue prefetch operation
10815 *
10816 * \return
10817 * ::CUDA_SUCCESS,
10818 * ::CUDA_ERROR_INVALID_VALUE,
10819 * ::CUDA_ERROR_INVALID_DEVICE
10820 * \notefnerr
10821 * \note_async
10822 * \note_null_stream
10823 *
10824 * \sa ::cuMemcpy, ::cuMemcpyPeer, ::cuMemcpyAsync,
10825 * ::cuMemcpy3DPeerAsync, ::cuMemAdvise,
10826 * ::cudaMemPrefetchAsync
10827 */
10828CUresult CUDAAPI cuMemPrefetchAsync(CUdeviceptr devPtr, size_t count, CUdevice dstDevice, CUstream hStream);
10829
10830/**
10831 * \brief Advise about the usage of a given memory range
10832 *
10833 * Advise the Unified Memory subsystem about the usage pattern for the memory range
10834 * starting at \p devPtr with a size of \p count bytes. The start address and end address of the memory
10835 * range will be rounded down and rounded up respectively to be aligned to CPU page size before the
10836 * advice is applied. The memory range must refer to managed memory allocated via ::cuMemAllocManaged
10837 * or declared via __managed__ variables. The memory range could also refer to system-allocated pageable
10838 * memory provided it represents a valid, host-accessible region of memory and all additional constraints
10839 * imposed by \p advice as outlined below are also satisfied. Specifying an invalid system-allocated pageable
10840 * memory range results in an error being returned.
10841 *
10842 * The \p advice parameter can take the following values:
10843 * - ::CU_MEM_ADVISE_SET_READ_MOSTLY: This implies that the data is mostly going to be read
10844 * from and only occasionally written to. Any read accesses from any processor to this region will create a
10845 * read-only copy of at least the accessed pages in that processor's memory. Additionally, if ::cuMemPrefetchAsync
10846 * is called on this region, it will create a read-only copy of the data on the destination processor.
10847 * If any processor writes to this region, all copies of the corresponding page will be invalidated
10848 * except for the one where the write occurred. The \p device argument is ignored for this advice.
10849 * Note that for a page to be read-duplicated, the accessing processor must either be the CPU or a GPU
10850 * that has a non-zero value for the device attribute ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS.
10851 * Also, if a context is created on a device that does not have the device attribute
10852 * ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS set, then read-duplication will not occur until
10853 * all such contexts are destroyed.
10854 * If the memory region refers to valid system-allocated pageable memory, then the accessing device must
10855 * have a non-zero value for the device attribute ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS for a read-only
10856 * copy to be created on that device. Note however that if the accessing device also has a non-zero value for the
10857 * device attribute ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS_USES_HOST_PAGE_TABLES, then setting this advice
10858 * will not create a read-only copy when that device accesses this memory region.
10859 *
10860 * - ::CU_MEM_ADVISE_UNSET_READ_MOSTLY: Undoes the effect of ::CU_MEM_ADVISE_SET_READ_MOSTLY and also prevents the
10861 * Unified Memory driver from attempting heuristic read-duplication on the memory range. Any read-duplicated
10862 * copies of the data will be collapsed into a single copy. The location for the collapsed
10863 * copy will be the preferred location if the page has a preferred location and one of the read-duplicated
10864 * copies was resident at that location. Otherwise, the location chosen is arbitrary.
10865 *
10866 * - ::CU_MEM_ADVISE_SET_PREFERRED_LOCATION: This advice sets the preferred location for the
10867 * data to be the memory belonging to \p device. Passing in CU_DEVICE_CPU for \p device sets the
10868 * preferred location as host memory. If \p device is a GPU, then it must have a non-zero value for the
10869 * device attribute ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS. Setting the preferred location
10870 * does not cause data to migrate to that location immediately. Instead, it guides the migration policy
10871 * when a fault occurs on that memory region. If the data is already in its preferred location and the
10872 * faulting processor can establish a mapping without requiring the data to be migrated, then
10873 * data migration will be avoided. On the other hand, if the data is not in its preferred location
10874 * or if a direct mapping cannot be established, then it will be migrated to the processor accessing
10875 * it. It is important to note that setting the preferred location does not prevent data prefetching
10876 * done using ::cuMemPrefetchAsync.
10877 * Having a preferred location can override the page thrash detection and resolution logic in the Unified
10878 * Memory driver. Normally, if a page is detected to be constantly thrashing between for example host and device
10879 * memory, the page may eventually be pinned to host memory by the Unified Memory driver. But
10880 * if the preferred location is set as device memory, then the page will continue to thrash indefinitely.
10881 * If ::CU_MEM_ADVISE_SET_READ_MOSTLY is also set on this memory region or any subset of it, then the
10882 * policies associated with that advice will override the policies of this advice, unless read accesses from
10883 * \p device will not result in a read-only copy being created on that device as outlined in description for
10884 * the advice ::CU_MEM_ADVISE_SET_READ_MOSTLY.
10885 * If the memory region refers to valid system-allocated pageable memory, then \p device must have a non-zero
10886 * value for the device attribute ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS. Additionally, if \p device has
10887 * a non-zero value for the device attribute ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS_USES_HOST_PAGE_TABLES,
10888 * then this call has no effect. Note however that this behavior may change in the future.
10889 *
10890 * - ::CU_MEM_ADVISE_UNSET_PREFERRED_LOCATION: Undoes the effect of ::CU_MEM_ADVISE_SET_PREFERRED_LOCATION
10891 * and changes the preferred location to none.
10892 *
10893 * - ::CU_MEM_ADVISE_SET_ACCESSED_BY: This advice implies that the data will be accessed by \p device.
10894 * Passing in ::CU_DEVICE_CPU for \p device will set the advice for the CPU. If \p device is a GPU, then
10895 * the device attribute ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS must be non-zero.
10896 * This advice does not cause data migration and has no impact on the location of the data per se. Instead,
10897 * it causes the data to always be mapped in the specified processor's page tables, as long as the
10898 * location of the data permits a mapping to be established. If the data gets migrated for any reason,
10899 * the mappings are updated accordingly.
10900 * This advice is recommended in scenarios where data locality is not important, but avoiding faults is.
10901 * Consider for example a system containing multiple GPUs with peer-to-peer access enabled, where the
10902 * data located on one GPU is occasionally accessed by peer GPUs. In such scenarios, migrating data
10903 * over to the other GPUs is not as important because the accesses are infrequent and the overhead of
10904 * migration may be too high. But preventing faults can still help improve performance, and so having
10905 * a mapping set up in advance is useful. Note that on CPU access of this data, the data may be migrated
10906 * to host memory because the CPU typically cannot access device memory directly. Any GPU that had the
10907 * ::CU_MEM_ADVISE_SET_ACCESSED_BY flag set for this data will now have its mapping updated to point to the
10908 * page in host memory.
10909 * If ::CU_MEM_ADVISE_SET_READ_MOSTLY is also set on this memory region or any subset of it, then the
10910 * policies associated with that advice will override the policies of this advice. Additionally, if the
10911 * preferred location of this memory region or any subset of it is also \p device, then the policies
10912 * associated with ::CU_MEM_ADVISE_SET_PREFERRED_LOCATION will override the policies of this advice.
10913 * If the memory region refers to valid system-allocated pageable memory, then \p device must have a non-zero
10914 * value for the device attribute ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS. Additionally, if \p device has
10915 * a non-zero value for the device attribute ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS_USES_HOST_PAGE_TABLES,
10916 * then this call has no effect.
10917 *
10918 * - ::CU_MEM_ADVISE_UNSET_ACCESSED_BY: Undoes the effect of ::CU_MEM_ADVISE_SET_ACCESSED_BY. Any mappings to
10919 * the data from \p device may be removed at any time causing accesses to result in non-fatal page faults.
10920 * If the memory region refers to valid system-allocated pageable memory, then \p device must have a non-zero
10921 * value for the device attribute ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS. Additionally, if \p device has
10922 * a non-zero value for the device attribute ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS_USES_HOST_PAGE_TABLES,
10923 * then this call has no effect.
10924 *
10925 * \param devPtr - Pointer to memory to set the advice for
10926 * \param count - Size in bytes of the memory range
10927 * \param advice - Advice to be applied for the specified memory range
10928 * \param device - Device to apply the advice for
10929 *
10930 * \return
10931 * ::CUDA_SUCCESS,
10932 * ::CUDA_ERROR_INVALID_VALUE,
10933 * ::CUDA_ERROR_INVALID_DEVICE
10934 * \notefnerr
10935 * \note_async
10936 * \note_null_stream
10937 *
10938 * \sa ::cuMemcpy, ::cuMemcpyPeer, ::cuMemcpyAsync,
10939 * ::cuMemcpy3DPeerAsync, ::cuMemPrefetchAsync,
10940 * ::cudaMemAdvise
10941 */
10942CUresult CUDAAPI cuMemAdvise(CUdeviceptr devPtr, size_t count, CUmem_advise advice, CUdevice device);
10943
10944/**
10945 * \brief Query an attribute of a given memory range
10946 *
10947 * Query an attribute about the memory range starting at \p devPtr with a size of \p count bytes. The
10948 * memory range must refer to managed memory allocated via ::cuMemAllocManaged or declared via
10949 * __managed__ variables.
10950 *
10951 * The \p attribute parameter can take the following values:
10952 * - ::CU_MEM_RANGE_ATTRIBUTE_READ_MOSTLY: If this attribute is specified, \p data will be interpreted
10953 * as a 32-bit integer, and \p dataSize must be 4. The result returned will be 1 if all pages in the given
10954 * memory range have read-duplication enabled, or 0 otherwise.
10955 * - ::CU_MEM_RANGE_ATTRIBUTE_PREFERRED_LOCATION: If this attribute is specified, \p data will be
10956 * interpreted as a 32-bit integer, and \p dataSize must be 4. The result returned will be a GPU device
10957 * id if all pages in the memory range have that GPU as their preferred location, or it will be CU_DEVICE_CPU
10958 * if all pages in the memory range have the CPU as their preferred location, or it will be CU_DEVICE_INVALID
10959 * if either all the pages don't have the same preferred location or some of the pages don't have a
10960 * preferred location at all. Note that the actual location of the pages in the memory range at the time of
10961 * the query may be different from the preferred location.
10962 * - ::CU_MEM_RANGE_ATTRIBUTE_ACCESSED_BY: If this attribute is specified, \p data will be interpreted
10963 * as an array of 32-bit integers, and \p dataSize must be a non-zero multiple of 4. The result returned
10964 * will be a list of device ids that had ::CU_MEM_ADVISE_SET_ACCESSED_BY set for that entire memory range.
10965 * If any device does not have that advice set for the entire memory range, that device will not be included.
10966 * If \p data is larger than the number of devices that have that advice set for that memory range,
10967 * CU_DEVICE_INVALID will be returned in all the extra space provided. For ex., if \p dataSize is 12
10968 * (i.e. \p data has 3 elements) and only device 0 has the advice set, then the result returned will be
10969 * { 0, CU_DEVICE_INVALID, CU_DEVICE_INVALID }. If \p data is smaller than the number of devices that have
10970 * that advice set, then only as many devices will be returned as can fit in the array. There is no
10971 * guarantee on which specific devices will be returned, however.
10972 * - ::CU_MEM_RANGE_ATTRIBUTE_LAST_PREFETCH_LOCATION: If this attribute is specified, \p data will be
10973 * interpreted as a 32-bit integer, and \p dataSize must be 4. The result returned will be the last location
10974 * to which all pages in the memory range were prefetched explicitly via ::cuMemPrefetchAsync. This will either be
10975 * a GPU id or CU_DEVICE_CPU depending on whether the last location for prefetch was a GPU or the CPU
10976 * respectively. If any page in the memory range was never explicitly prefetched or if all pages were not
10977 * prefetched to the same location, CU_DEVICE_INVALID will be returned. Note that this simply returns the
10978 * last location that the application requested to prefetch the memory range to. It gives no indication as to
10979 * whether the prefetch operation to that location has completed or even begun.
10980 *
10981 * \param data - A pointers to a memory location where the result
10982 * of each attribute query will be written to.
10983 * \param dataSize - Array containing the size of data
10984 * \param attribute - The attribute to query
10985 * \param devPtr - Start of the range to query
10986 * \param count - Size of the range to query
10987 *
10988 * \return
10989 * ::CUDA_SUCCESS,
10990 * ::CUDA_ERROR_INVALID_VALUE,
10991 * ::CUDA_ERROR_INVALID_DEVICE
10992 * \notefnerr
10993 * \note_async
10994 * \note_null_stream
10995 *
10996 * \sa ::cuMemRangeGetAttributes, ::cuMemPrefetchAsync,
10997 * ::cuMemAdvise,
10998 * ::cudaMemRangeGetAttribute
10999 */
11000CUresult CUDAAPI cuMemRangeGetAttribute(void *data, size_t dataSize, CUmem_range_attribute attribute, CUdeviceptr devPtr, size_t count);
11001
11002/**
11003 * \brief Query attributes of a given memory range.
11004 *
11005 * Query attributes of the memory range starting at \p devPtr with a size of \p count bytes. The
11006 * memory range must refer to managed memory allocated via ::cuMemAllocManaged or declared via
11007 * __managed__ variables. The \p attributes array will be interpreted to have \p numAttributes
11008 * entries. The \p dataSizes array will also be interpreted to have \p numAttributes entries.
11009 * The results of the query will be stored in \p data.
11010 *
11011 * The list of supported attributes are given below. Please refer to ::cuMemRangeGetAttribute for
11012 * attribute descriptions and restrictions.
11013 *
11014 * - ::CU_MEM_RANGE_ATTRIBUTE_READ_MOSTLY
11015 * - ::CU_MEM_RANGE_ATTRIBUTE_PREFERRED_LOCATION
11016 * - ::CU_MEM_RANGE_ATTRIBUTE_ACCESSED_BY
11017 * - ::CU_MEM_RANGE_ATTRIBUTE_LAST_PREFETCH_LOCATION
11018 *
11019 * \param data - A two-dimensional array containing pointers to memory
11020 * locations where the result of each attribute query will be written to.
11021 * \param dataSizes - Array containing the sizes of each result
11022 * \param attributes - An array of attributes to query
11023 * (numAttributes and the number of attributes in this array should match)
11024 * \param numAttributes - Number of attributes to query
11025 * \param devPtr - Start of the range to query
11026 * \param count - Size of the range to query
11027 *
11028 * \return
11029 * ::CUDA_SUCCESS,
11030 * ::CUDA_ERROR_DEINITIALIZED,
11031 * ::CUDA_ERROR_INVALID_CONTEXT,
11032 * ::CUDA_ERROR_INVALID_VALUE,
11033 * ::CUDA_ERROR_INVALID_DEVICE
11034 * \notefnerr
11035 *
11036 * \sa ::cuMemRangeGetAttribute, ::cuMemAdvise,
11037 * ::cuMemPrefetchAsync,
11038 * ::cudaMemRangeGetAttributes
11039 */
11040CUresult CUDAAPI cuMemRangeGetAttributes(void **data, size_t *dataSizes, CUmem_range_attribute *attributes, size_t numAttributes, CUdeviceptr devPtr, size_t count);
11041
11042/**
11043 * \brief Set attributes on a previously allocated memory region
11044 *
11045 * The supported attributes are:
11046 *
11047 * - ::CU_POINTER_ATTRIBUTE_SYNC_MEMOPS:
11048 *
11049 * A boolean attribute that can either be set (1) or unset (0). When set,
11050 * the region of memory that \p ptr points to is guaranteed to always synchronize
11051 * memory operations that are synchronous. If there are some previously initiated
11052 * synchronous memory operations that are pending when this attribute is set, the
11053 * function does not return until those memory operations are complete.
11054 * See further documentation in the section titled "API synchronization behavior"
11055 * to learn more about cases when synchronous memory operations can
11056 * exhibit asynchronous behavior.
11057 * \p value will be considered as a pointer to an unsigned integer to which this attribute is to be set.
11058 *
11059 * \param value - Pointer to memory containing the value to be set
11060 * \param attribute - Pointer attribute to set
11061 * \param ptr - Pointer to a memory region allocated using CUDA memory allocation APIs
11062 *
11063 * \return
11064 * ::CUDA_SUCCESS,
11065 * ::CUDA_ERROR_DEINITIALIZED,
11066 * ::CUDA_ERROR_NOT_INITIALIZED,
11067 * ::CUDA_ERROR_INVALID_CONTEXT,
11068 * ::CUDA_ERROR_INVALID_VALUE,
11069 * ::CUDA_ERROR_INVALID_DEVICE
11070 * \notefnerr
11071 *
11072 * \sa ::cuPointerGetAttribute,
11073 * ::cuPointerGetAttributes,
11074 * ::cuMemAlloc,
11075 * ::cuMemFree,
11076 * ::cuMemAllocHost,
11077 * ::cuMemFreeHost,
11078 * ::cuMemHostAlloc,
11079 * ::cuMemHostRegister,
11080 * ::cuMemHostUnregister
11081 */
11082CUresult CUDAAPI cuPointerSetAttribute(const void *value, CUpointer_attribute attribute, CUdeviceptr ptr);
11083
11084/**
11085 * \brief Returns information about a pointer.
11086 *
11087 * The supported attributes are (refer to ::cuPointerGetAttribute for attribute descriptions and restrictions):
11088 *
11089 * - ::CU_POINTER_ATTRIBUTE_CONTEXT
11090 * - ::CU_POINTER_ATTRIBUTE_MEMORY_TYPE
11091 * - ::CU_POINTER_ATTRIBUTE_DEVICE_POINTER
11092 * - ::CU_POINTER_ATTRIBUTE_HOST_POINTER
11093 * - ::CU_POINTER_ATTRIBUTE_SYNC_MEMOPS
11094 * - ::CU_POINTER_ATTRIBUTE_BUFFER_ID
11095 * - ::CU_POINTER_ATTRIBUTE_IS_MANAGED
11096 * - ::CU_POINTER_ATTRIBUTE_DEVICE_ORDINAL
11097 * - ::CU_POINTER_ATTRIBUTE_RANGE_START_ADDR
11098 * - ::CU_POINTER_ATTRIBUTE_RANGE_SIZE
11099 * - ::CU_POINTER_ATTRIBUTE_MAPPED
11100 * - ::CU_POINTER_ATTRIBUTE_IS_LEGACY_CUDA_IPC_CAPABLE
11101 * - ::CU_POINTER_ATTRIBUTE_ALLOWED_HANDLE_TYPES
11102 * - ::CU_POINTER_ATTRIBUTE_MEMPOOL_HANDLE
11103 *
11104 * \param numAttributes - Number of attributes to query
11105 * \param attributes - An array of attributes to query
11106 * (numAttributes and the number of attributes in this array should match)
11107 * \param data - A two-dimensional array containing pointers to memory
11108 * locations where the result of each attribute query will be written to.
11109 * \param ptr - Pointer to query
11110 *
11111 * Unlike ::cuPointerGetAttribute, this function will not return an error when the \p ptr
11112 * encountered is not a valid CUDA pointer. Instead, the attributes are assigned default NULL values
11113 * and CUDA_SUCCESS is returned.
11114 *
11115 * If \p ptr was not allocated by, mapped by, or registered with a ::CUcontext which uses UVA
11116 * (Unified Virtual Addressing), ::CUDA_ERROR_INVALID_CONTEXT is returned.
11117 *
11118 * \return
11119 * ::CUDA_SUCCESS,
11120 * ::CUDA_ERROR_DEINITIALIZED,
11121 * ::CUDA_ERROR_INVALID_CONTEXT,
11122 * ::CUDA_ERROR_INVALID_VALUE,
11123 * ::CUDA_ERROR_INVALID_DEVICE
11124 * \notefnerr
11125 *
11126 * \sa
11127 * ::cuPointerGetAttribute,
11128 * ::cuPointerSetAttribute,
11129 * ::cudaPointerGetAttributes
11130 */
11131CUresult CUDAAPI cuPointerGetAttributes(unsigned int numAttributes, CUpointer_attribute *attributes, void **data, CUdeviceptr ptr);
11132
11133/** @} */ /* END CUDA_UNIFIED */
11134
11135/**
11136 * \defgroup CUDA_STREAM Stream Management
11137 *
11138 * ___MANBRIEF___ stream management functions of the low-level CUDA driver API
11139 * (___CURRENT_FILE___) ___ENDMANBRIEF___
11140 *
11141 * This section describes the stream management functions of the low-level CUDA
11142 * driver application programming interface.
11143 *
11144 * @{
11145 */
11146
11147/**
11148 * \brief Create a stream
11149 *
11150 * Creates a stream and returns a handle in \p phStream. The \p Flags argument
11151 * determines behaviors of the stream.
11152 *
11153 * Valid values for \p Flags are:
11154 * - ::CU_STREAM_DEFAULT: Default stream creation flag.
11155 * - ::CU_STREAM_NON_BLOCKING: Specifies that work running in the created
11156 * stream may run concurrently with work in stream 0 (the NULL stream), and that
11157 * the created stream should perform no implicit synchronization with stream 0.
11158 *
11159 * \param phStream - Returned newly created stream
11160 * \param Flags - Parameters for stream creation
11161 *
11162 * \return
11163 * ::CUDA_SUCCESS,
11164 * ::CUDA_ERROR_DEINITIALIZED,
11165 * ::CUDA_ERROR_NOT_INITIALIZED,
11166 * ::CUDA_ERROR_INVALID_CONTEXT,
11167 * ::CUDA_ERROR_INVALID_VALUE,
11168 * ::CUDA_ERROR_OUT_OF_MEMORY
11169 * \notefnerr
11170 *
11171 * \sa ::cuStreamDestroy,
11172 * ::cuStreamCreateWithPriority,
11173 * ::cuStreamGetPriority,
11174 * ::cuStreamGetFlags,
11175 * ::cuStreamWaitEvent,
11176 * ::cuStreamQuery,
11177 * ::cuStreamSynchronize,
11178 * ::cuStreamAddCallback,
11179 * ::cudaStreamCreate,
11180 * ::cudaStreamCreateWithFlags
11181 */
11182CUresult CUDAAPI cuStreamCreate(CUstream *phStream, unsigned int Flags);
11183
11184/**
11185 * \brief Create a stream with the given priority
11186 *
11187 * Creates a stream with the specified priority and returns a handle in \p phStream.
11188 * This API alters the scheduler priority of work in the stream. Work in a higher
11189 * priority stream may preempt work already executing in a low priority stream.
11190 *
11191 * \p priority follows a convention where lower numbers represent higher priorities.
11192 * '0' represents default priority. The range of meaningful numerical priorities can
11193 * be queried using ::cuCtxGetStreamPriorityRange. If the specified priority is
11194 * outside the numerical range returned by ::cuCtxGetStreamPriorityRange,
11195 * it will automatically be clamped to the lowest or the highest number in the range.
11196 *
11197 * \param phStream - Returned newly created stream
11198 * \param flags - Flags for stream creation. See ::cuStreamCreate for a list of
11199 * valid flags
11200 * \param priority - Stream priority. Lower numbers represent higher priorities.
11201 * See ::cuCtxGetStreamPriorityRange for more information about
11202 * meaningful stream priorities that can be passed.
11203 *
11204 * \return
11205 * ::CUDA_SUCCESS,
11206 * ::CUDA_ERROR_DEINITIALIZED,
11207 * ::CUDA_ERROR_NOT_INITIALIZED,
11208 * ::CUDA_ERROR_INVALID_CONTEXT,
11209 * ::CUDA_ERROR_INVALID_VALUE,
11210 * ::CUDA_ERROR_OUT_OF_MEMORY
11211 * \notefnerr
11212 *
11213 * \note Stream priorities are supported only on GPUs
11214 * with compute capability 3.5 or higher.
11215 *
11216 * \note In the current implementation, only compute kernels launched in
11217 * priority streams are affected by the stream's priority. Stream priorities have
11218 * no effect on host-to-device and device-to-host memory operations.
11219 *
11220 * \sa ::cuStreamDestroy,
11221 * ::cuStreamCreate,
11222 * ::cuStreamGetPriority,
11223 * ::cuCtxGetStreamPriorityRange,
11224 * ::cuStreamGetFlags,
11225 * ::cuStreamWaitEvent,
11226 * ::cuStreamQuery,
11227 * ::cuStreamSynchronize,
11228 * ::cuStreamAddCallback,
11229 * ::cudaStreamCreateWithPriority
11230 */
11231CUresult CUDAAPI cuStreamCreateWithPriority(CUstream *phStream, unsigned int flags, int priority);
11232
11233
11234/**
11235 * \brief Query the priority of a given stream
11236 *
11237 * Query the priority of a stream created using ::cuStreamCreate or ::cuStreamCreateWithPriority
11238 * and return the priority in \p priority. Note that if the stream was created with a
11239 * priority outside the numerical range returned by ::cuCtxGetStreamPriorityRange,
11240 * this function returns the clamped priority.
11241 * See ::cuStreamCreateWithPriority for details about priority clamping.
11242 *
11243 * \param hStream - Handle to the stream to be queried
11244 * \param priority - Pointer to a signed integer in which the stream's priority is returned
11245 * \return
11246 * ::CUDA_SUCCESS,
11247 * ::CUDA_ERROR_DEINITIALIZED,
11248 * ::CUDA_ERROR_NOT_INITIALIZED,
11249 * ::CUDA_ERROR_INVALID_CONTEXT,
11250 * ::CUDA_ERROR_INVALID_VALUE,
11251 * ::CUDA_ERROR_INVALID_HANDLE,
11252 * ::CUDA_ERROR_OUT_OF_MEMORY
11253 * \notefnerr
11254 *
11255 * \sa ::cuStreamDestroy,
11256 * ::cuStreamCreate,
11257 * ::cuStreamCreateWithPriority,
11258 * ::cuCtxGetStreamPriorityRange,
11259 * ::cuStreamGetFlags,
11260 * ::cudaStreamGetPriority
11261 */
11262CUresult CUDAAPI cuStreamGetPriority(CUstream hStream, int *priority);
11263
11264/**
11265 * \brief Query the flags of a given stream
11266 *
11267 * Query the flags of a stream created using ::cuStreamCreate or ::cuStreamCreateWithPriority
11268 * and return the flags in \p flags.
11269 *
11270 * \param hStream - Handle to the stream to be queried
11271 * \param flags - Pointer to an unsigned integer in which the stream's flags are returned
11272 * The value returned in \p flags is a logical 'OR' of all flags that
11273 * were used while creating this stream. See ::cuStreamCreate for the list
11274 * of valid flags
11275 * \return
11276 * ::CUDA_SUCCESS,
11277 * ::CUDA_ERROR_DEINITIALIZED,
11278 * ::CUDA_ERROR_NOT_INITIALIZED,
11279 * ::CUDA_ERROR_INVALID_CONTEXT,
11280 * ::CUDA_ERROR_INVALID_VALUE,
11281 * ::CUDA_ERROR_INVALID_HANDLE,
11282 * ::CUDA_ERROR_OUT_OF_MEMORY
11283 * \notefnerr
11284 *
11285 * \sa ::cuStreamDestroy,
11286 * ::cuStreamCreate,
11287 * ::cuStreamGetPriority,
11288 * ::cudaStreamGetFlags
11289 */
11290CUresult CUDAAPI cuStreamGetFlags(CUstream hStream, unsigned int *flags);
11291
11292/**
11293 * \brief Query the context associated with a stream
11294 *
11295 * Returns the CUDA context that the stream is associated with.
11296 *
11297 * The stream handle \p hStream can refer to any of the following:
11298 * <ul>
11299 * <li>a stream created via any of the CUDA driver APIs such as ::cuStreamCreate
11300 * and ::cuStreamCreateWithPriority, or their runtime API equivalents such as
11301 * ::cudaStreamCreate, ::cudaStreamCreateWithFlags and ::cudaStreamCreateWithPriority.
11302 * The returned context is the context that was active in the calling thread when the
11303 * stream was created. Passing an invalid handle will result in undefined behavior.</li>
11304 * <li>any of the special streams such as the NULL stream, ::CU_STREAM_LEGACY and
11305 * ::CU_STREAM_PER_THREAD. The runtime API equivalents of these are also accepted,
11306 * which are NULL, ::cudaStreamLegacy and ::cudaStreamPerThread respectively.
11307 * Specifying any of the special handles will return the context current to the
11308 * calling thread. If no context is current to the calling thread,
11309 * ::CUDA_ERROR_INVALID_CONTEXT is returned.</li>
11310 * </ul>
11311 *
11312 * \param hStream - Handle to the stream to be queried
11313 * \param pctx - Returned context associated with the stream
11314 *
11315 * \return
11316 * ::CUDA_SUCCESS,
11317 * ::CUDA_ERROR_DEINITIALIZED,
11318 * ::CUDA_ERROR_NOT_INITIALIZED,
11319 * ::CUDA_ERROR_INVALID_CONTEXT,
11320 * ::CUDA_ERROR_INVALID_HANDLE,
11321 * \notefnerr
11322 *
11323 * \sa ::cuStreamDestroy,
11324 * ::cuStreamCreateWithPriority,
11325 * ::cuStreamGetPriority,
11326 * ::cuStreamGetFlags,
11327 * ::cuStreamWaitEvent,
11328 * ::cuStreamQuery,
11329 * ::cuStreamSynchronize,
11330 * ::cuStreamAddCallback,
11331 * ::cudaStreamCreate,
11332 * ::cudaStreamCreateWithFlags
11333 */
11334CUresult CUDAAPI cuStreamGetCtx(CUstream hStream, CUcontext *pctx);
11335
11336/**
11337 * \brief Make a compute stream wait on an event
11338 *
11339 * Makes all future work submitted to \p hStream wait for all work captured in
11340 * \p hEvent. See ::cuEventRecord() for details on what is captured by an event.
11341 * The synchronization will be performed efficiently on the device when applicable.
11342 * \p hEvent may be from a different context or device than \p hStream.
11343 *
11344 * flags include:
11345 * - ::CU_EVENT_WAIT_DEFAULT: Default event creation flag.
11346 * - ::CU_EVENT_WAIT_EXTERNAL: Event is captured in the graph as an external
11347 * event node when performing stream capture. This flag is invalid outside
11348 * of stream capture.
11349 *
11350 * \param hStream - Stream to wait
11351 * \param hEvent - Event to wait on (may not be NULL)
11352 * \param Flags - See ::CUevent_capture_flags
11353 *
11354 * \return
11355 * ::CUDA_SUCCESS,
11356 * ::CUDA_ERROR_DEINITIALIZED,
11357 * ::CUDA_ERROR_NOT_INITIALIZED,
11358 * ::CUDA_ERROR_INVALID_CONTEXT,
11359 * ::CUDA_ERROR_INVALID_HANDLE,
11360 * \note_null_stream
11361 * \notefnerr
11362 *
11363 * \sa ::cuStreamCreate,
11364 * ::cuEventRecord,
11365 * ::cuStreamQuery,
11366 * ::cuStreamSynchronize,
11367 * ::cuStreamAddCallback,
11368 * ::cuStreamDestroy,
11369 * ::cudaStreamWaitEvent
11370 */
11371CUresult CUDAAPI cuStreamWaitEvent(CUstream hStream, CUevent hEvent, unsigned int Flags);
11372
11373/**
11374 * \brief Add a callback to a compute stream
11375 *
11376 * \note This function is slated for eventual deprecation and removal. If
11377 * you do not require the callback to execute in case of a device error,
11378 * consider using ::cuLaunchHostFunc. Additionally, this function is not
11379 * supported with ::cuStreamBeginCapture and ::cuStreamEndCapture, unlike
11380 * ::cuLaunchHostFunc.
11381 *
11382 * Adds a callback to be called on the host after all currently enqueued
11383 * items in the stream have completed. For each
11384 * cuStreamAddCallback call, the callback will be executed exactly once.
11385 * The callback will block later work in the stream until it is finished.
11386 *
11387 * The callback may be passed ::CUDA_SUCCESS or an error code. In the event
11388 * of a device error, all subsequently executed callbacks will receive an
11389 * appropriate ::CUresult.
11390 *
11391 * Callbacks must not make any CUDA API calls. Attempting to use a CUDA API
11392 * will result in ::CUDA_ERROR_NOT_PERMITTED. Callbacks must not perform any
11393 * synchronization that may depend on outstanding device work or other callbacks
11394 * that are not mandated to run earlier. Callbacks without a mandated order
11395 * (in independent streams) execute in undefined order and may be serialized.
11396 *
11397 * For the purposes of Unified Memory, callback execution makes a number of
11398 * guarantees:
11399 * <ul>
11400 * <li>The callback stream is considered idle for the duration of the
11401 * callback. Thus, for example, a callback may always use memory attached
11402 * to the callback stream.</li>
11403 * <li>The start of execution of a callback has the same effect as
11404 * synchronizing an event recorded in the same stream immediately prior to
11405 * the callback. It thus synchronizes streams which have been "joined"
11406 * prior to the callback.</li>
11407 * <li>Adding device work to any stream does not have the effect of making
11408 * the stream active until all preceding host functions and stream callbacks
11409 * have executed. Thus, for
11410 * example, a callback might use global attached memory even if work has
11411 * been added to another stream, if the work has been ordered behind the
11412 * callback with an event.</li>
11413 * <li>Completion of a callback does not cause a stream to become
11414 * active except as described above. The callback stream will remain idle
11415 * if no device work follows the callback, and will remain idle across
11416 * consecutive callbacks without device work in between. Thus, for example,
11417 * stream synchronization can be done by signaling from a callback at the
11418 * end of the stream.</li>
11419 * </ul>
11420 *
11421 * \param hStream - Stream to add callback to
11422 * \param callback - The function to call once preceding stream operations are complete
11423 * \param userData - User specified data to be passed to the callback function
11424 * \param flags - Reserved for future use, must be 0
11425 *
11426 * \return
11427 * ::CUDA_SUCCESS,
11428 * ::CUDA_ERROR_DEINITIALIZED,
11429 * ::CUDA_ERROR_NOT_INITIALIZED,
11430 * ::CUDA_ERROR_INVALID_CONTEXT,
11431 * ::CUDA_ERROR_INVALID_HANDLE,
11432 * ::CUDA_ERROR_NOT_SUPPORTED
11433 * \note_null_stream
11434 * \notefnerr
11435 *
11436 * \sa ::cuStreamCreate,
11437 * ::cuStreamQuery,
11438 * ::cuStreamSynchronize,
11439 * ::cuStreamWaitEvent,
11440 * ::cuStreamDestroy,
11441 * ::cuMemAllocManaged,
11442 * ::cuStreamAttachMemAsync,
11443 * ::cuStreamLaunchHostFunc,
11444 * ::cudaStreamAddCallback
11445 */
11446CUresult CUDAAPI cuStreamAddCallback(CUstream hStream, CUstreamCallback callback, void *userData, unsigned int flags);
11447
11448/**
11449 * \brief Begins graph capture on a stream
11450 *
11451 * Begin graph capture on \p hStream. When a stream is in capture mode, all operations
11452 * pushed into the stream will not be executed, but will instead be captured into
11453 * a graph, which will be returned via ::cuStreamEndCapture. Capture may not be initiated
11454 * if \p stream is CU_STREAM_LEGACY. Capture must be ended on the same stream in which
11455 * it was initiated, and it may only be initiated if the stream is not already in capture
11456 * mode. The capture mode may be queried via ::cuStreamIsCapturing. A unique id
11457 * representing the capture sequence may be queried via ::cuStreamGetCaptureInfo.
11458 *
11459 * If \p mode is not ::CU_STREAM_CAPTURE_MODE_RELAXED, ::cuStreamEndCapture must be
11460 * called on this stream from the same thread.
11461 *
11462 * \param hStream - Stream in which to initiate capture
11463 * \param mode - Controls the interaction of this capture sequence with other API
11464 * calls that are potentially unsafe. For more details see
11465 * ::cuThreadExchangeStreamCaptureMode.
11466 *
11467 * \note Kernels captured using this API must not use texture and surface references.
11468 * Reading or writing through any texture or surface reference is undefined
11469 * behavior. This restriction does not apply to texture and surface objects.
11470 *
11471 * \return
11472 * ::CUDA_SUCCESS,
11473 * ::CUDA_ERROR_DEINITIALIZED,
11474 * ::CUDA_ERROR_NOT_INITIALIZED,
11475 * ::CUDA_ERROR_INVALID_VALUE
11476 * \notefnerr
11477 *
11478 * \sa
11479 * ::cuStreamCreate,
11480 * ::cuStreamIsCapturing,
11481 * ::cuStreamEndCapture,
11482 * ::cuThreadExchangeStreamCaptureMode
11483 */
11484CUresult CUDAAPI cuStreamBeginCapture(CUstream hStream, CUstreamCaptureMode mode);
11485
11486/**
11487 * \brief Swaps the stream capture interaction mode for a thread
11488 *
11489 * Sets the calling thread's stream capture interaction mode to the value contained
11490 * in \p *mode, and overwrites \p *mode with the previous mode for the thread. To
11491 * facilitate deterministic behavior across function or module boundaries, callers
11492 * are encouraged to use this API in a push-pop fashion: \code
11493 CUstreamCaptureMode mode = desiredMode;
11494 cuThreadExchangeStreamCaptureMode(&mode);
11495 ...
11496 cuThreadExchangeStreamCaptureMode(&mode); // restore previous mode
11497 * \endcode
11498 *
11499 * During stream capture (see ::cuStreamBeginCapture), some actions, such as a call
11500 * to ::cudaMalloc, may be unsafe. In the case of ::cudaMalloc, the operation is
11501 * not enqueued asynchronously to a stream, and is not observed by stream capture.
11502 * Therefore, if the sequence of operations captured via ::cuStreamBeginCapture
11503 * depended on the allocation being replayed whenever the graph is launched, the
11504 * captured graph would be invalid.
11505 *
11506 * Therefore, stream capture places restrictions on API calls that can be made within
11507 * or concurrently to a ::cuStreamBeginCapture-::cuStreamEndCapture sequence. This
11508 * behavior can be controlled via this API and flags to ::cuStreamBeginCapture.
11509 *
11510 * A thread's mode is one of the following:
11511 * - \p CU_STREAM_CAPTURE_MODE_GLOBAL: This is the default mode. If the local thread has
11512 * an ongoing capture sequence that was not initiated with
11513 * \p CU_STREAM_CAPTURE_MODE_RELAXED at \p cuStreamBeginCapture, or if any other thread
11514 * has a concurrent capture sequence initiated with \p CU_STREAM_CAPTURE_MODE_GLOBAL,
11515 * this thread is prohibited from potentially unsafe API calls.
11516 * - \p CU_STREAM_CAPTURE_MODE_THREAD_LOCAL: If the local thread has an ongoing capture
11517 * sequence not initiated with \p CU_STREAM_CAPTURE_MODE_RELAXED, it is prohibited
11518 * from potentially unsafe API calls. Concurrent capture sequences in other threads
11519 * are ignored.
11520 * - \p CU_STREAM_CAPTURE_MODE_RELAXED: The local thread is not prohibited from potentially
11521 * unsafe API calls. Note that the thread is still prohibited from API calls which
11522 * necessarily conflict with stream capture, for example, attempting ::cuEventQuery
11523 * on an event that was last recorded inside a capture sequence.
11524 *
11525 * \param mode - Pointer to mode value to swap with the current mode
11526 *
11527 * \return
11528 * ::CUDA_SUCCESS,
11529 * ::CUDA_ERROR_DEINITIALIZED,
11530 * ::CUDA_ERROR_NOT_INITIALIZED,
11531 * ::CUDA_ERROR_INVALID_VALUE
11532 * \notefnerr
11533 *
11534 * \sa
11535 * ::cuStreamBeginCapture
11536 */
11537CUresult CUDAAPI cuThreadExchangeStreamCaptureMode(CUstreamCaptureMode *mode);
11538
11539/**
11540 * \brief Ends capture on a stream, returning the captured graph
11541 *
11542 * End capture on \p hStream, returning the captured graph via \p phGraph.
11543 * Capture must have been initiated on \p hStream via a call to ::cuStreamBeginCapture.
11544 * If capture was invalidated, due to a violation of the rules of stream capture, then
11545 * a NULL graph will be returned.
11546 *
11547 * If the \p mode argument to ::cuStreamBeginCapture was not
11548 * ::CU_STREAM_CAPTURE_MODE_RELAXED, this call must be from the same thread as
11549 * ::cuStreamBeginCapture.
11550 *
11551 * \param hStream - Stream to query
11552 * \param phGraph - The captured graph
11553 *
11554 * \return
11555 * ::CUDA_SUCCESS,
11556 * ::CUDA_ERROR_DEINITIALIZED,
11557 * ::CUDA_ERROR_NOT_INITIALIZED,
11558 * ::CUDA_ERROR_INVALID_VALUE,
11559 * ::CUDA_ERROR_STREAM_CAPTURE_WRONG_THREAD
11560 * \notefnerr
11561 *
11562 * \sa
11563 * ::cuStreamCreate,
11564 * ::cuStreamBeginCapture,
11565 * ::cuStreamIsCapturing
11566 */
11567CUresult CUDAAPI cuStreamEndCapture(CUstream hStream, CUgraph *phGraph);
11568
11569/**
11570 * \brief Returns a stream's capture status
11571 *
11572 * Return the capture status of \p hStream via \p captureStatus. After a successful
11573 * call, \p *captureStatus will contain one of the following:
11574 * - ::CU_STREAM_CAPTURE_STATUS_NONE: The stream is not capturing.
11575 * - ::CU_STREAM_CAPTURE_STATUS_ACTIVE: The stream is capturing.
11576 * - ::CU_STREAM_CAPTURE_STATUS_INVALIDATED: The stream was capturing but an error
11577 * has invalidated the capture sequence. The capture sequence must be terminated
11578 * with ::cuStreamEndCapture on the stream where it was initiated in order to
11579 * continue using \p hStream.
11580 *
11581 * Note that, if this is called on ::CU_STREAM_LEGACY (the "null stream") while
11582 * a blocking stream in the same context is capturing, it will return
11583 * ::CUDA_ERROR_STREAM_CAPTURE_IMPLICIT and \p *captureStatus is unspecified
11584 * after the call. The blocking stream capture is not invalidated.
11585 *
11586 * When a blocking stream is capturing, the legacy stream is in an
11587 * unusable state until the blocking stream capture is terminated. The legacy
11588 * stream is not supported for stream capture, but attempted use would have an
11589 * implicit dependency on the capturing stream(s).
11590 *
11591 * \param hStream - Stream to query
11592 * \param captureStatus - Returns the stream's capture status
11593 *
11594 * \return
11595 * ::CUDA_SUCCESS,
11596 * ::CUDA_ERROR_DEINITIALIZED,
11597 * ::CUDA_ERROR_NOT_INITIALIZED,
11598 * ::CUDA_ERROR_INVALID_VALUE,
11599 * ::CUDA_ERROR_STREAM_CAPTURE_IMPLICIT
11600 * \notefnerr
11601 *
11602 * \sa
11603 * ::cuStreamCreate,
11604 * ::cuStreamBeginCapture,
11605 * ::cuStreamEndCapture
11606 */
11607CUresult CUDAAPI cuStreamIsCapturing(CUstream hStream, CUstreamCaptureStatus *captureStatus);
11608
11609/**
11610 * \brief Query capture status of a stream
11611 *
11612 * Note there is a later version of this API, ::cuStreamGetCaptureInfo_v2. It will
11613 * supplant this version in 12.0, which is retained for minor version compatibility.
11614 *
11615 * Query the capture status of a stream and and get an id for
11616 * the capture sequence, which is unique over the lifetime of the process.
11617 *
11618 * If called on ::CU_STREAM_LEGACY (the "null stream") while a stream not created
11619 * with ::CU_STREAM_NON_BLOCKING is capturing, returns ::CUDA_ERROR_STREAM_CAPTURE_IMPLICIT.
11620 *
11621 * A valid id is returned only if both of the following are true:
11622 * - the call returns CUDA_SUCCESS
11623 * - captureStatus is set to ::CU_STREAM_CAPTURE_STATUS_ACTIVE
11624 *
11625 * \return
11626 * ::CUDA_SUCCESS,
11627 * ::CUDA_ERROR_STREAM_CAPTURE_IMPLICIT
11628 * \notefnerr
11629 *
11630 * \sa
11631 * ::cuStreamGetCaptureInfo_v2,
11632 * ::cuStreamBeginCapture,
11633 * ::cuStreamIsCapturing
11634 */
11635CUresult CUDAAPI cuStreamGetCaptureInfo(CUstream hStream, CUstreamCaptureStatus *captureStatus_out, cuuint64_t *id_out);
11636
11637/**
11638 * \brief Query a stream's capture state (11.3+)
11639 *
11640 * Query stream state related to stream capture.
11641 *
11642 * If called on ::CU_STREAM_LEGACY (the "null stream") while a stream not created
11643 * with ::CU_STREAM_NON_BLOCKING is capturing, returns ::CUDA_ERROR_STREAM_CAPTURE_IMPLICIT.
11644 *
11645 * Valid data (other than capture status) is returned only if both of the following are true:
11646 * - the call returns CUDA_SUCCESS
11647 * - the returned capture status is ::CU_STREAM_CAPTURE_STATUS_ACTIVE
11648 *
11649 * This version of cuStreamGetCaptureInfo is introduced in CUDA 11.3 and will supplant the
11650 * previous version in 12.0. Developers requiring compatibility across minor versions to
11651 * CUDA 11.0 (driver version 445) should use ::cuStreamGetCaptureInfo or include a fallback
11652 * path.
11653 *
11654 * \param hStream - The stream to query
11655 * \param captureStatus_out - Location to return the capture status of the stream; required
11656 * \param id_out - Optional location to return an id for the capture sequence, which is
11657 * unique over the lifetime of the process
11658 * \param graph_out - Optional location to return the graph being captured into. All
11659 * operations other than destroy and node removal are permitted on the graph
11660 * while the capture sequence is in progress. This API does not transfer
11661 * ownership of the graph, which is transferred or destroyed at
11662 * ::cuStreamEndCapture. Note that the graph handle may be invalidated before
11663 * end of capture for certain errors. Nodes that are or become
11664 * unreachable from the original stream at ::cuStreamEndCapture due to direct
11665 * actions on the graph do not trigger ::CUDA_ERROR_STREAM_CAPTURE_UNJOINED.
11666 * \param dependencies_out - Optional location to store a pointer to an array of nodes.
11667 * The next node to be captured in the stream will depend on this set of nodes,
11668 * absent operations such as event wait which modify this set. The array pointer
11669 * is valid until the next API call which operates on the stream or until end of
11670 * capture. The node handles may be copied out and are valid until they or the
11671 * graph is destroyed. The driver-owned array may also be passed directly to
11672 * APIs that operate on the graph (not the stream) without copying.
11673 * \param numDependencies_out - Optional location to store the size of the array
11674 * returned in dependencies_out.
11675 *
11676 * \return
11677 * ::CUDA_SUCCESS,
11678 * ::CUDA_ERROR_INVALID_VALUE,
11679 * ::CUDA_ERROR_STREAM_CAPTURE_IMPLICIT
11680 * \note_graph_thread_safety
11681 * \notefnerr
11682 *
11683 * \sa
11684 * ::cuStreamGetCaptureInfo,
11685 * ::cuStreamBeginCapture,
11686 * ::cuStreamIsCapturing,
11687 * ::cuStreamUpdateCaptureDependencies
11688 */
11689CUresult CUDAAPI cuStreamGetCaptureInfo_v2(CUstream hStream, CUstreamCaptureStatus *captureStatus_out,
11690 cuuint64_t *id_out, CUgraph *graph_out, const CUgraphNode **dependencies_out, size_t *numDependencies_out);
11691
11692/**
11693 * \brief Update the set of dependencies in a capturing stream (11.3+)
11694 *
11695 * Modifies the dependency set of a capturing stream. The dependency set is the set
11696 * of nodes that the next captured node in the stream will depend on.
11697 *
11698 * Valid flags are ::CU_STREAM_ADD_CAPTURE_DEPENDENCIES and
11699 * ::CU_STREAM_SET_CAPTURE_DEPENDENCIES. These control whether the set passed to
11700 * the API is added to the existing set or replaces it. A flags value of 0 defaults
11701 * to ::CU_STREAM_ADD_CAPTURE_DEPENDENCIES.
11702 *
11703 * Nodes that are removed from the dependency set via this API do not result in
11704 * ::CUDA_ERROR_STREAM_CAPTURE_UNJOINED if they are unreachable from the stream at
11705 * ::cuStreamEndCapture.
11706 *
11707 * Returns ::CUDA_ERROR_ILLEGAL_STATE if the stream is not capturing.
11708 *
11709 * This API is new in CUDA 11.3. Developers requiring compatibility across minor
11710 * versions to CUDA 11.0 should not use this API or provide a fallback.
11711 *
11712 * \return
11713 * ::CUDA_SUCCESS,
11714 * ::CUDA_ERROR_INVALID_VALUE,
11715 * ::CUDA_ERROR_ILLEGAL_STATE
11716 *
11717 * \sa
11718 * ::cuStreamBeginCapture,
11719 * ::cuStreamGetCaptureInfo,
11720 * ::cuStreamGetCaptureInfo_v2
11721 */
11722CUresult CUDAAPI cuStreamUpdateCaptureDependencies(CUstream hStream, CUgraphNode *dependencies, size_t numDependencies, unsigned int flags);
11723
11724/**
11725 * \brief Attach memory to a stream asynchronously
11726 *
11727 * Enqueues an operation in \p hStream to specify stream association of
11728 * \p length bytes of memory starting from \p dptr. This function is a
11729 * stream-ordered operation, meaning that it is dependent on, and will
11730 * only take effect when, previous work in stream has completed. Any
11731 * previous association is automatically replaced.
11732 *
11733 * \p dptr must point to one of the following types of memories:
11734 * - managed memory declared using the __managed__ keyword or allocated with
11735 * ::cuMemAllocManaged.
11736 * - a valid host-accessible region of system-allocated pageable memory. This
11737 * type of memory may only be specified if the device associated with the
11738 * stream reports a non-zero value for the device attribute
11739 * ::CU_DEVICE_ATTRIBUTE_PAGEABLE_MEMORY_ACCESS.
11740 *
11741 * For managed allocations, \p length must be either zero or the entire
11742 * allocation's size. Both indicate that the entire allocation's stream
11743 * association is being changed. Currently, it is not possible to change stream
11744 * association for a portion of a managed allocation.
11745 *
11746 * For pageable host allocations, \p length must be non-zero.
11747 *
11748 * The stream association is specified using \p flags which must be
11749 * one of ::CUmemAttach_flags.
11750 * If the ::CU_MEM_ATTACH_GLOBAL flag is specified, the memory can be accessed
11751 * by any stream on any device.
11752 * If the ::CU_MEM_ATTACH_HOST flag is specified, the program makes a guarantee
11753 * that it won't access the memory on the device from any stream on a device that
11754 * has a zero value for the device attribute ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS.
11755 * If the ::CU_MEM_ATTACH_SINGLE flag is specified and \p hStream is associated with
11756 * a device that has a zero value for the device attribute ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS,
11757 * the program makes a guarantee that it will only access the memory on the device
11758 * from \p hStream. It is illegal to attach singly to the NULL stream, because the
11759 * NULL stream is a virtual global stream and not a specific stream. An error will
11760 * be returned in this case.
11761 *
11762 * When memory is associated with a single stream, the Unified Memory system will
11763 * allow CPU access to this memory region so long as all operations in \p hStream
11764 * have completed, regardless of whether other streams are active. In effect,
11765 * this constrains exclusive ownership of the managed memory region by
11766 * an active GPU to per-stream activity instead of whole-GPU activity.
11767 *
11768 * Accessing memory on the device from streams that are not associated with
11769 * it will produce undefined results. No error checking is performed by the
11770 * Unified Memory system to ensure that kernels launched into other streams
11771 * do not access this region.
11772 *
11773 * It is a program's responsibility to order calls to ::cuStreamAttachMemAsync
11774 * via events, synchronization or other means to ensure legal access to memory
11775 * at all times. Data visibility and coherency will be changed appropriately
11776 * for all kernels which follow a stream-association change.
11777 *
11778 * If \p hStream is destroyed while data is associated with it, the association is
11779 * removed and the association reverts to the default visibility of the allocation
11780 * as specified at ::cuMemAllocManaged. For __managed__ variables, the default
11781 * association is always ::CU_MEM_ATTACH_GLOBAL. Note that destroying a stream is an
11782 * asynchronous operation, and as a result, the change to default association won't
11783 * happen until all work in the stream has completed.
11784 *
11785 * \param hStream - Stream in which to enqueue the attach operation
11786 * \param dptr - Pointer to memory (must be a pointer to managed memory or
11787 * to a valid host-accessible region of system-allocated
11788 * pageable memory)
11789 * \param length - Length of memory
11790 * \param flags - Must be one of ::CUmemAttach_flags
11791 *
11792 * \return
11793 * ::CUDA_SUCCESS,
11794 * ::CUDA_ERROR_DEINITIALIZED,
11795 * ::CUDA_ERROR_NOT_INITIALIZED,
11796 * ::CUDA_ERROR_INVALID_CONTEXT,
11797 * ::CUDA_ERROR_INVALID_HANDLE,
11798 * ::CUDA_ERROR_NOT_SUPPORTED
11799 * \note_null_stream
11800 * \notefnerr
11801 *
11802 * \sa ::cuStreamCreate,
11803 * ::cuStreamQuery,
11804 * ::cuStreamSynchronize,
11805 * ::cuStreamWaitEvent,
11806 * ::cuStreamDestroy,
11807 * ::cuMemAllocManaged,
11808 * ::cudaStreamAttachMemAsync
11809 */
11810CUresult CUDAAPI cuStreamAttachMemAsync(CUstream hStream, CUdeviceptr dptr, size_t length, unsigned int flags);
11811
11812/**
11813 * \brief Determine status of a compute stream
11814 *
11815 * Returns ::CUDA_SUCCESS if all operations in the stream specified by
11816 * \p hStream have completed, or ::CUDA_ERROR_NOT_READY if not.
11817 *
11818 * For the purposes of Unified Memory, a return value of ::CUDA_SUCCESS
11819 * is equivalent to having called ::cuStreamSynchronize().
11820 *
11821 * \param hStream - Stream to query status of
11822 *
11823 * \return
11824 * ::CUDA_SUCCESS,
11825 * ::CUDA_ERROR_DEINITIALIZED,
11826 * ::CUDA_ERROR_NOT_INITIALIZED,
11827 * ::CUDA_ERROR_INVALID_CONTEXT,
11828 * ::CUDA_ERROR_INVALID_HANDLE,
11829 * ::CUDA_ERROR_NOT_READY
11830 * \note_null_stream
11831 * \notefnerr
11832 *
11833 * \sa ::cuStreamCreate,
11834 * ::cuStreamWaitEvent,
11835 * ::cuStreamDestroy,
11836 * ::cuStreamSynchronize,
11837 * ::cuStreamAddCallback,
11838 * ::cudaStreamQuery
11839 */
11840CUresult CUDAAPI cuStreamQuery(CUstream hStream);
11841
11842/**
11843 * \brief Wait until a stream's tasks are completed
11844 *
11845 * Waits until the device has completed all operations in the stream specified
11846 * by \p hStream. If the context was created with the
11847 * ::CU_CTX_SCHED_BLOCKING_SYNC flag, the CPU thread will block until the
11848 * stream is finished with all of its tasks.
11849 *
11850 * \param hStream - Stream to wait for
11851 *
11852 * \return
11853 * ::CUDA_SUCCESS,
11854 * ::CUDA_ERROR_DEINITIALIZED,
11855 * ::CUDA_ERROR_NOT_INITIALIZED,
11856 * ::CUDA_ERROR_INVALID_CONTEXT,
11857 * ::CUDA_ERROR_INVALID_HANDLE
11858
11859 * \note_null_stream
11860 * \notefnerr
11861 *
11862 * \sa ::cuStreamCreate,
11863 * ::cuStreamDestroy,
11864 * ::cuStreamWaitEvent,
11865 * ::cuStreamQuery,
11866 * ::cuStreamAddCallback,
11867 * ::cudaStreamSynchronize
11868 */
11869CUresult CUDAAPI cuStreamSynchronize(CUstream hStream);
11870
11871/**
11872 * \brief Destroys a stream
11873 *
11874 * Destroys the stream specified by \p hStream.
11875 *
11876 * In case the device is still doing work in the stream \p hStream
11877 * when ::cuStreamDestroy() is called, the function will return immediately
11878 * and the resources associated with \p hStream will be released automatically
11879 * once the device has completed all work in \p hStream.
11880 *
11881 * \param hStream - Stream to destroy
11882 *
11883 * \return
11884 * ::CUDA_SUCCESS,
11885 * ::CUDA_ERROR_DEINITIALIZED,
11886 * ::CUDA_ERROR_NOT_INITIALIZED,
11887 * ::CUDA_ERROR_INVALID_CONTEXT,
11888 * ::CUDA_ERROR_INVALID_VALUE,
11889 * ::CUDA_ERROR_INVALID_HANDLE
11890 * \notefnerr
11891 *
11892 * \sa ::cuStreamCreate,
11893 * ::cuStreamWaitEvent,
11894 * ::cuStreamQuery,
11895 * ::cuStreamSynchronize,
11896 * ::cuStreamAddCallback,
11897 * ::cudaStreamDestroy
11898 */
11899CUresult CUDAAPI cuStreamDestroy(CUstream hStream);
11900
11901/**
11902 * \brief Copies attributes from source stream to destination stream.
11903 *
11904 * Copies attributes from source stream \p src to destination stream \p dst.
11905 * Both streams must have the same context.
11906 *
11907 * \param[out] dst Destination stream
11908 * \param[in] src Source stream
11909 * For list of attributes see ::CUstreamAttrID
11910 *
11911 * \return
11912 * ::CUDA_SUCCESS,
11913 * ::CUDA_ERROR_INVALID_VALUE
11914 * \notefnerr
11915 *
11916 * \sa
11917 * ::CUaccessPolicyWindow
11918 */
11919CUresult CUDAAPI cuStreamCopyAttributes(CUstream dst, CUstream src);
11920
11921/**
11922 * \brief Queries stream attribute.
11923 *
11924 * Queries attribute \p attr from \p hStream and stores it in corresponding
11925 * member of \p value_out.
11926 *
11927 * \param[in] hStream
11928 * \param[in] attr
11929 * \param[out] value_out
11930 *
11931 * \return
11932 * ::CUDA_SUCCESS,
11933 * ::CUDA_ERROR_INVALID_VALUE,
11934 * ::CUDA_ERROR_INVALID_HANDLE
11935 * \notefnerr
11936 *
11937 * \sa
11938 * ::CUaccessPolicyWindow
11939 */
11940CUresult CUDAAPI cuStreamGetAttribute(CUstream hStream, CUstreamAttrID attr,
11941 CUstreamAttrValue *value_out);
11942
11943/**
11944 * \brief Sets stream attribute.
11945 *
11946 * Sets attribute \p attr on \p hStream from corresponding attribute of
11947 * \p value. The updated attribute will be applied to subsequent work
11948 * submitted to the stream. It will not affect previously submitted work.
11949 *
11950 * \param[out] hStream
11951 * \param[in] attr
11952 * \param[in] value
11953 *
11954 * \return
11955 * ::CUDA_SUCCESS,
11956 * ::CUDA_ERROR_INVALID_VALUE,
11957 * ::CUDA_ERROR_INVALID_HANDLE
11958 * \notefnerr
11959 *
11960 * \sa
11961 * ::CUaccessPolicyWindow
11962 */
11963CUresult CUDAAPI cuStreamSetAttribute(CUstream hStream, CUstreamAttrID attr,
11964 const CUstreamAttrValue *value);
11965
11966/** @} */ /* END CUDA_STREAM */
11967
11968
11969/**
11970 * \defgroup CUDA_EVENT Event Management
11971 *
11972 * ___MANBRIEF___ event management functions of the low-level CUDA driver API
11973 * (___CURRENT_FILE___) ___ENDMANBRIEF___
11974 *
11975 * This section describes the event management functions of the low-level CUDA
11976 * driver application programming interface.
11977 *
11978 * @{
11979 */
11980
11981/**
11982 * \brief Creates an event
11983 *
11984 * Creates an event *phEvent for the current context with the flags specified via
11985 * \p Flags. Valid flags include:
11986 * - ::CU_EVENT_DEFAULT: Default event creation flag.
11987 * - ::CU_EVENT_BLOCKING_SYNC: Specifies that the created event should use blocking
11988 * synchronization. A CPU thread that uses ::cuEventSynchronize() to wait on
11989 * an event created with this flag will block until the event has actually
11990 * been recorded.
11991 * - ::CU_EVENT_DISABLE_TIMING: Specifies that the created event does not need
11992 * to record timing data. Events created with this flag specified and
11993 * the ::CU_EVENT_BLOCKING_SYNC flag not specified will provide the best
11994 * performance when used with ::cuStreamWaitEvent() and ::cuEventQuery().
11995 * - ::CU_EVENT_INTERPROCESS: Specifies that the created event may be used as an
11996 * interprocess event by ::cuIpcGetEventHandle(). ::CU_EVENT_INTERPROCESS must
11997 * be specified along with ::CU_EVENT_DISABLE_TIMING.
11998 *
11999 * \param phEvent - Returns newly created event
12000 * \param Flags - Event creation flags
12001 *
12002 * \return
12003 * ::CUDA_SUCCESS,
12004 * ::CUDA_ERROR_DEINITIALIZED,
12005 * ::CUDA_ERROR_NOT_INITIALIZED,
12006 * ::CUDA_ERROR_INVALID_CONTEXT,
12007 * ::CUDA_ERROR_INVALID_VALUE,
12008 * ::CUDA_ERROR_OUT_OF_MEMORY
12009 * \notefnerr
12010 *
12011 * \sa
12012 * ::cuEventRecord,
12013 * ::cuEventQuery,
12014 * ::cuEventSynchronize,
12015 * ::cuEventDestroy,
12016 * ::cuEventElapsedTime,
12017 * ::cudaEventCreate,
12018 * ::cudaEventCreateWithFlags
12019 */
12020CUresult CUDAAPI cuEventCreate(CUevent *phEvent, unsigned int Flags);
12021
12022/**
12023 * \brief Records an event
12024 *
12025 * Captures in \p hEvent the contents of \p hStream at the time of this call.
12026 * \p hEvent and \p hStream must be from the same context.
12027 * Calls such as ::cuEventQuery() or ::cuStreamWaitEvent() will then
12028 * examine or wait for completion of the work that was captured. Uses of
12029 * \p hStream after this call do not modify \p hEvent. See note on default
12030 * stream behavior for what is captured in the default case.
12031 *
12032 * ::cuEventRecord() can be called multiple times on the same event and
12033 * will overwrite the previously captured state. Other APIs such as
12034 * ::cuStreamWaitEvent() use the most recently captured state at the time
12035 * of the API call, and are not affected by later calls to
12036 * ::cuEventRecord(). Before the first call to ::cuEventRecord(), an
12037 * event represents an empty set of work, so for example ::cuEventQuery()
12038 * would return ::CUDA_SUCCESS.
12039 *
12040 * \param hEvent - Event to record
12041 * \param hStream - Stream to record event for
12042 *
12043 * \return
12044 * ::CUDA_SUCCESS,
12045 * ::CUDA_ERROR_DEINITIALIZED,
12046 * ::CUDA_ERROR_NOT_INITIALIZED,
12047 * ::CUDA_ERROR_INVALID_CONTEXT,
12048 * ::CUDA_ERROR_INVALID_HANDLE,
12049 * ::CUDA_ERROR_INVALID_VALUE
12050 * \note_null_stream
12051 * \notefnerr
12052 *
12053 * \sa ::cuEventCreate,
12054 * ::cuEventQuery,
12055 * ::cuEventSynchronize,
12056 * ::cuStreamWaitEvent,
12057 * ::cuEventDestroy,
12058 * ::cuEventElapsedTime,
12059 * ::cudaEventRecord,
12060 * ::cuEventRecordWithFlags
12061 */
12062CUresult CUDAAPI cuEventRecord(CUevent hEvent, CUstream hStream);
12063
12064/**
12065 * \brief Records an event
12066 *
12067 * Captures in \p hEvent the contents of \p hStream at the time of this call.
12068 * \p hEvent and \p hStream must be from the same context.
12069 * Calls such as ::cuEventQuery() or ::cuStreamWaitEvent() will then
12070 * examine or wait for completion of the work that was captured. Uses of
12071 * \p hStream after this call do not modify \p hEvent. See note on default
12072 * stream behavior for what is captured in the default case.
12073 *
12074 * ::cuEventRecordWithFlags() can be called multiple times on the same event and
12075 * will overwrite the previously captured state. Other APIs such as
12076 * ::cuStreamWaitEvent() use the most recently captured state at the time
12077 * of the API call, and are not affected by later calls to
12078 * ::cuEventRecordWithFlags(). Before the first call to ::cuEventRecordWithFlags(), an
12079 * event represents an empty set of work, so for example ::cuEventQuery()
12080 * would return ::CUDA_SUCCESS.
12081 *
12082 * flags include:
12083 * - ::CU_EVENT_RECORD_DEFAULT: Default event creation flag.
12084 * - ::CU_EVENT_RECORD_EXTERNAL: Event is captured in the graph as an external
12085 * event node when performing stream capture. This flag is invalid outside
12086 * of stream capture.
12087 *
12088 * \param hEvent - Event to record
12089 * \param hStream - Stream to record event for
12090 * \param flags - See ::CUevent_capture_flags
12091 *
12092 * \return
12093 * ::CUDA_SUCCESS,
12094 * ::CUDA_ERROR_DEINITIALIZED,
12095 * ::CUDA_ERROR_NOT_INITIALIZED,
12096 * ::CUDA_ERROR_INVALID_CONTEXT,
12097 * ::CUDA_ERROR_INVALID_HANDLE,
12098 * ::CUDA_ERROR_INVALID_VALUE
12099 * \note_null_stream
12100 * \notefnerr
12101 *
12102 * \sa ::cuEventCreate,
12103 * ::cuEventQuery,
12104 * ::cuEventSynchronize,
12105 * ::cuStreamWaitEvent,
12106 * ::cuEventDestroy,
12107 * ::cuEventElapsedTime,
12108 * ::cuEventRecord,
12109 * ::cudaEventRecord
12110 */
12111CUresult CUDAAPI cuEventRecordWithFlags(CUevent hEvent, CUstream hStream, unsigned int flags);
12112
12113/**
12114 * \brief Queries an event's status
12115 *
12116 * Queries the status of all work currently captured by \p hEvent. See
12117 * ::cuEventRecord() for details on what is captured by an event.
12118 *
12119 * Returns ::CUDA_SUCCESS if all captured work has been completed, or
12120 * ::CUDA_ERROR_NOT_READY if any captured work is incomplete.
12121 *
12122 * For the purposes of Unified Memory, a return value of ::CUDA_SUCCESS
12123 * is equivalent to having called ::cuEventSynchronize().
12124 *
12125 * \param hEvent - Event to query
12126 *
12127 * \return
12128 * ::CUDA_SUCCESS,
12129 * ::CUDA_ERROR_DEINITIALIZED,
12130 * ::CUDA_ERROR_NOT_INITIALIZED,
12131 * ::CUDA_ERROR_INVALID_HANDLE,
12132 * ::CUDA_ERROR_INVALID_VALUE,
12133 * ::CUDA_ERROR_NOT_READY
12134 * \notefnerr
12135 *
12136 * \sa ::cuEventCreate,
12137 * ::cuEventRecord,
12138 * ::cuEventSynchronize,
12139 * ::cuEventDestroy,
12140 * ::cuEventElapsedTime,
12141 * ::cudaEventQuery
12142 */
12143CUresult CUDAAPI cuEventQuery(CUevent hEvent);
12144
12145/**
12146 * \brief Waits for an event to complete
12147 *
12148 * Waits until the completion of all work currently captured in \p hEvent.
12149 * See ::cuEventRecord() for details on what is captured by an event.
12150 *
12151 * Waiting for an event that was created with the ::CU_EVENT_BLOCKING_SYNC
12152 * flag will cause the calling CPU thread to block until the event has
12153 * been completed by the device. If the ::CU_EVENT_BLOCKING_SYNC flag has
12154 * not been set, then the CPU thread will busy-wait until the event has
12155 * been completed by the device.
12156 *
12157 * \param hEvent - Event to wait for
12158 *
12159 * \return
12160 * ::CUDA_SUCCESS,
12161 * ::CUDA_ERROR_DEINITIALIZED,
12162 * ::CUDA_ERROR_NOT_INITIALIZED,
12163 * ::CUDA_ERROR_INVALID_CONTEXT,
12164 * ::CUDA_ERROR_INVALID_HANDLE
12165 * \notefnerr
12166 *
12167 * \sa ::cuEventCreate,
12168 * ::cuEventRecord,
12169 * ::cuEventQuery,
12170 * ::cuEventDestroy,
12171 * ::cuEventElapsedTime,
12172 * ::cudaEventSynchronize
12173 */
12174CUresult CUDAAPI cuEventSynchronize(CUevent hEvent);
12175
12176/**
12177 * \brief Destroys an event
12178 *
12179 * Destroys the event specified by \p hEvent.
12180 *
12181 * An event may be destroyed before it is complete (i.e., while
12182 * ::cuEventQuery() would return ::CUDA_ERROR_NOT_READY). In this case, the
12183 * call does not block on completion of the event, and any associated
12184 * resources will automatically be released asynchronously at completion.
12185 *
12186 * \param hEvent - Event to destroy
12187 *
12188 * \return
12189 * ::CUDA_SUCCESS,
12190 * ::CUDA_ERROR_DEINITIALIZED,
12191 * ::CUDA_ERROR_NOT_INITIALIZED,
12192 * ::CUDA_ERROR_INVALID_CONTEXT,
12193 * ::CUDA_ERROR_INVALID_HANDLE
12194 * \notefnerr
12195 *
12196 * \sa ::cuEventCreate,
12197 * ::cuEventRecord,
12198 * ::cuEventQuery,
12199 * ::cuEventSynchronize,
12200 * ::cuEventElapsedTime,
12201 * ::cudaEventDestroy
12202 */
12203CUresult CUDAAPI cuEventDestroy(CUevent hEvent);
12204
12205/**
12206 * \brief Computes the elapsed time between two events
12207 *
12208 * Computes the elapsed time between two events (in milliseconds with a
12209 * resolution of around 0.5 microseconds).
12210 *
12211 * If either event was last recorded in a non-NULL stream, the resulting time
12212 * may be greater than expected (even if both used the same stream handle). This
12213 * happens because the ::cuEventRecord() operation takes place asynchronously
12214 * and there is no guarantee that the measured latency is actually just between
12215 * the two events. Any number of other different stream operations could execute
12216 * in between the two measured events, thus altering the timing in a significant
12217 * way.
12218 *
12219 * If ::cuEventRecord() has not been called on either event then
12220 * ::CUDA_ERROR_INVALID_HANDLE is returned. If ::cuEventRecord() has been called
12221 * on both events but one or both of them has not yet been completed (that is,
12222 * ::cuEventQuery() would return ::CUDA_ERROR_NOT_READY on at least one of the
12223 * events), ::CUDA_ERROR_NOT_READY is returned. If either event was created with
12224 * the ::CU_EVENT_DISABLE_TIMING flag, then this function will return
12225 * ::CUDA_ERROR_INVALID_HANDLE.
12226 *
12227 * \param pMilliseconds - Time between \p hStart and \p hEnd in ms
12228 * \param hStart - Starting event
12229 * \param hEnd - Ending event
12230 *
12231 * \return
12232 * ::CUDA_SUCCESS,
12233 * ::CUDA_ERROR_DEINITIALIZED,
12234 * ::CUDA_ERROR_NOT_INITIALIZED,
12235 * ::CUDA_ERROR_INVALID_CONTEXT,
12236 * ::CUDA_ERROR_INVALID_HANDLE,
12237 * ::CUDA_ERROR_NOT_READY
12238 * \notefnerr
12239 *
12240 * \sa ::cuEventCreate,
12241 * ::cuEventRecord,
12242 * ::cuEventQuery,
12243 * ::cuEventSynchronize,
12244 * ::cuEventDestroy,
12245 * ::cudaEventElapsedTime
12246 */
12247CUresult CUDAAPI cuEventElapsedTime(float *pMilliseconds, CUevent hStart, CUevent hEnd);
12248
12249/** @} */ /* END CUDA_EVENT */
12250
12251/**
12252 * \defgroup CUDA_EXTRES_INTEROP External Resource Interoperability
12253 *
12254 * ___MANBRIEF___ External resource interoperability functions of the low-level CUDA driver API
12255 * (___CURRENT_FILE___) ___ENDMANBRIEF___
12256 *
12257 * This section describes the external resource interoperability functions of the low-level CUDA
12258 * driver application programming interface.
12259 *
12260 * @{
12261 */
12262
12263 /**
12264 * \brief Imports an external memory object
12265 *
12266 * Imports an externally allocated memory object and returns
12267 * a handle to that in \p extMem_out.
12268 *
12269 * The properties of the handle being imported must be described in
12270 * \p memHandleDesc. The ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC structure
12271 * is defined as follows:
12272 *
12273 * \code
12274 typedef struct CUDA_EXTERNAL_MEMORY_HANDLE_DESC_st {
12275 CUexternalMemoryHandleType type;
12276 union {
12277 int fd;
12278 struct {
12279 void *handle;
12280 const void *name;
12281 } win32;
12282 const void *nvSciBufObject;
12283 } handle;
12284 unsigned long long size;
12285 unsigned int flags;
12286 } CUDA_EXTERNAL_MEMORY_HANDLE_DESC;
12287 * \endcode
12288 *
12289 * where ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::type specifies the type
12290 * of handle being imported. ::CUexternalMemoryHandleType is
12291 * defined as:
12292 *
12293 * \code
12294 typedef enum CUexternalMemoryHandleType_enum {
12295 CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD = 1,
12296 CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32 = 2,
12297 CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT = 3,
12298 CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_HEAP = 4,
12299 CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_RESOURCE = 5,
12300 CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_RESOURCE = 6,
12301 CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_RESOURCE_KMT = 7,
12302 CU_EXTERNAL_MEMORY_HANDLE_TYPE_NVSCIBUF = 8
12303 } CUexternalMemoryHandleType;
12304 * \endcode
12305 *
12306 * If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::type is
12307 * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD, then
12308 * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::fd must be a valid
12309 * file descriptor referencing a memory object. Ownership of
12310 * the file descriptor is transferred to the CUDA driver when the
12311 * handle is imported successfully. Performing any operations on the
12312 * file descriptor after it is imported results in undefined behavior.
12313 *
12314 * If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::type is
12315 * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32, then exactly one
12316 * of ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::handle and
12317 * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::name must not be
12318 * NULL. If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::handle
12319 * is not NULL, then it must represent a valid shared NT handle that
12320 * references a memory object. Ownership of this handle is
12321 * not transferred to CUDA after the import operation, so the
12322 * application must release the handle using the appropriate system
12323 * call. If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::name
12324 * is not NULL, then it must point to a NULL-terminated array of
12325 * UTF-16 characters that refers to a memory object.
12326 *
12327 * If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::type is
12328 * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_WIN32_KMT, then
12329 * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::handle must
12330 * be non-NULL and
12331 * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::name
12332 * must be NULL. The handle specified must be a globally shared KMT
12333 * handle. This handle does not hold a reference to the underlying
12334 * object, and thus will be invalid when all references to the
12335 * memory object are destroyed.
12336 *
12337 * If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::type is
12338 * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_HEAP, then exactly one
12339 * of ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::handle and
12340 * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::name must not be
12341 * NULL. If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::handle
12342 * is not NULL, then it must represent a valid shared NT handle that
12343 * is returned by ID3D12Device::CreateSharedHandle when referring to a
12344 * ID3D12Heap object. This handle holds a reference to the underlying
12345 * object. If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::name
12346 * is not NULL, then it must point to a NULL-terminated array of
12347 * UTF-16 characters that refers to a ID3D12Heap object.
12348 *
12349 * If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::type is
12350 * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_RESOURCE, then exactly one
12351 * of ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::handle and
12352 * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::name must not be
12353 * NULL. If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::handle
12354 * is not NULL, then it must represent a valid shared NT handle that
12355 * is returned by ID3D12Device::CreateSharedHandle when referring to a
12356 * ID3D12Resource object. This handle holds a reference to the
12357 * underlying object. If
12358 * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::name
12359 * is not NULL, then it must point to a NULL-terminated array of
12360 * UTF-16 characters that refers to a ID3D12Resource object.
12361 *
12362 * If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::type is
12363 * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_RESOURCE, then
12364 * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::handle must
12365 * represent a valid shared NT handle that is returned by
12366 * IDXGIResource1::CreateSharedHandle when referring to a
12367 * ID3D11Resource object. If
12368 * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::name
12369 * is not NULL, then it must point to a NULL-terminated array of
12370 * UTF-16 characters that refers to a ID3D11Resource object.
12371 *
12372 * If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::type is
12373 * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_RESOURCE_KMT, then
12374 * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::handle must
12375 * represent a valid shared KMT handle that is returned by
12376 * IDXGIResource::GetSharedHandle when referring to a
12377 * ID3D11Resource object and
12378 * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::win32::name
12379 * must be NULL.
12380 *
12381 * If ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::type is
12382 * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_NVSCIBUF, then
12383 * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::handle::nvSciBufObject must be non-NULL
12384 * and reference a valid NvSciBuf object.
12385 * If the NvSciBuf object imported into CUDA is also mapped by other drivers, then the
12386 * application must use ::cuWaitExternalSemaphoresAsync or ::cuSignalExternalSemaphoresAsync
12387 * as appropriate barriers to maintain coherence between CUDA and the other drivers.
12388 * See ::CUDA_EXTERNAL_SEMAPHORE_SIGNAL_SKIP_NVSCIBUF_MEMSYNC and ::CUDA_EXTERNAL_SEMAPHORE_WAIT_SKIP_NVSCIBUF_MEMSYNC
12389 * for memory synchronization.
12390 *
12391 *
12392 * The size of the memory object must be specified in
12393 * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::size.
12394 *
12395 * Specifying the flag ::CUDA_EXTERNAL_MEMORY_DEDICATED in
12396 * ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::flags indicates that the
12397 * resource is a dedicated resource. The definition of what a
12398 * dedicated resource is outside the scope of this extension.
12399 * This flag must be set if ::CUDA_EXTERNAL_MEMORY_HANDLE_DESC::type
12400 * is one of the following:
12401 * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D12_RESOURCE
12402 * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_RESOURCE
12403 * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_D3D11_RESOURCE_KMT
12404 *
12405 * \param extMem_out - Returned handle to an external memory object
12406 * \param memHandleDesc - Memory import handle descriptor
12407 *
12408 * \return
12409 * ::CUDA_SUCCESS,
12410 * ::CUDA_ERROR_NOT_INITIALIZED,
12411 * ::CUDA_ERROR_INVALID_VALUE,
12412 * ::CUDA_ERROR_INVALID_HANDLE
12413 * \notefnerr
12414 *
12415 * \note If the Vulkan memory imported into CUDA is mapped on the CPU then the
12416 * application must use vkInvalidateMappedMemoryRanges/vkFlushMappedMemoryRanges
12417 * as well as appropriate Vulkan pipeline barriers to maintain coherence between
12418 * CPU and GPU. For more information on these APIs, please refer to "Synchronization
12419 * and Cache Control" chapter from Vulkan specification.
12420 *
12421 * \sa ::cuDestroyExternalMemory,
12422 * ::cuExternalMemoryGetMappedBuffer,
12423 * ::cuExternalMemoryGetMappedMipmappedArray
12424 */
12425CUresult CUDAAPI cuImportExternalMemory(CUexternalMemory *extMem_out, const CUDA_EXTERNAL_MEMORY_HANDLE_DESC *memHandleDesc);
12426
12427/**
12428 * \brief Maps a buffer onto an imported memory object
12429 *
12430 * Maps a buffer onto an imported memory object and returns a device
12431 * pointer in \p devPtr.
12432 *
12433 * The properties of the buffer being mapped must be described in
12434 * \p bufferDesc. The ::CUDA_EXTERNAL_MEMORY_BUFFER_DESC structure is
12435 * defined as follows:
12436 *
12437 * \code
12438 typedef struct CUDA_EXTERNAL_MEMORY_BUFFER_DESC_st {
12439 unsigned long long offset;
12440 unsigned long long size;
12441 unsigned int flags;
12442 } CUDA_EXTERNAL_MEMORY_BUFFER_DESC;
12443 * \endcode
12444 *
12445 * where ::CUDA_EXTERNAL_MEMORY_BUFFER_DESC::offset is the offset in
12446 * the memory object where the buffer's base address is.
12447 * ::CUDA_EXTERNAL_MEMORY_BUFFER_DESC::size is the size of the buffer.
12448 * ::CUDA_EXTERNAL_MEMORY_BUFFER_DESC::flags must be zero.
12449 *
12450 * The offset and size have to be suitably aligned to match the
12451 * requirements of the external API. Mapping two buffers whose ranges
12452 * overlap may or may not result in the same virtual address being
12453 * returned for the overlapped portion. In such cases, the application
12454 * must ensure that all accesses to that region from the GPU are
12455 * volatile. Otherwise writes made via one address are not guaranteed
12456 * to be visible via the other address, even if they're issued by the
12457 * same thread. It is recommended that applications map the combined
12458 * range instead of mapping separate buffers and then apply the
12459 * appropriate offsets to the returned pointer to derive the
12460 * individual buffers.
12461 *
12462 * The returned pointer \p devPtr must be freed using ::cuMemFree.
12463 *
12464 * \param devPtr - Returned device pointer to buffer
12465 * \param extMem - Handle to external memory object
12466 * \param bufferDesc - Buffer descriptor
12467 *
12468 * \return
12469 * ::CUDA_SUCCESS,
12470 * ::CUDA_ERROR_NOT_INITIALIZED,
12471 * ::CUDA_ERROR_INVALID_VALUE,
12472 * ::CUDA_ERROR_INVALID_HANDLE
12473 * \notefnerr
12474 *
12475 * \sa ::cuImportExternalMemory,
12476 * ::cuDestroyExternalMemory,
12477 * ::cuExternalMemoryGetMappedMipmappedArray
12478 */
12479CUresult CUDAAPI cuExternalMemoryGetMappedBuffer(CUdeviceptr *devPtr, CUexternalMemory extMem, const CUDA_EXTERNAL_MEMORY_BUFFER_DESC *bufferDesc);
12480
12481/**
12482 * \brief Maps a CUDA mipmapped array onto an external memory object
12483 *
12484 * Maps a CUDA mipmapped array onto an external object and returns a
12485 * handle to it in \p mipmap.
12486 *
12487 * The properties of the CUDA mipmapped array being mapped must be
12488 * described in \p mipmapDesc. The structure
12489 * ::CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC is defined as follows:
12490 *
12491 * \code
12492 typedef struct CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC_st {
12493 unsigned long long offset;
12494 CUDA_ARRAY3D_DESCRIPTOR arrayDesc;
12495 unsigned int numLevels;
12496 } CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC;
12497 * \endcode
12498 *
12499 * where ::CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC::offset is the
12500 * offset in the memory object where the base level of the mipmap
12501 * chain is.
12502 * ::CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC::arrayDesc describes
12503 * the format, dimensions and type of the base level of the mipmap
12504 * chain. For further details on these parameters, please refer to the
12505 * documentation for ::cuMipmappedArrayCreate. Note that if the mipmapped
12506 * array is bound as a color target in the graphics API, then the flag
12507 * ::CUDA_ARRAY3D_COLOR_ATTACHMENT must be specified in
12508 * ::CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC::arrayDesc::Flags.
12509 * ::CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC::numLevels specifies
12510 * the total number of levels in the mipmap chain.
12511 *
12512 * If \p extMem was imported from a handle of type ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_NVSCIBUF, then
12513 * ::CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC::numLevels must be equal to 1.
12514 *
12515 * The returned CUDA mipmapped array must be freed using ::cuMipmappedArrayDestroy.
12516 *
12517 * \param mipmap - Returned CUDA mipmapped array
12518 * \param extMem - Handle to external memory object
12519 * \param mipmapDesc - CUDA array descriptor
12520 *
12521 * \return
12522 * ::CUDA_SUCCESS,
12523 * ::CUDA_ERROR_NOT_INITIALIZED,
12524 * ::CUDA_ERROR_INVALID_VALUE,
12525 * ::CUDA_ERROR_INVALID_HANDLE
12526 * \notefnerr
12527 *
12528 * \sa ::cuImportExternalMemory,
12529 * ::cuDestroyExternalMemory,
12530 * ::cuExternalMemoryGetMappedBuffer
12531 */
12532CUresult CUDAAPI cuExternalMemoryGetMappedMipmappedArray(CUmipmappedArray *mipmap, CUexternalMemory extMem, const CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC *mipmapDesc);
12533
12534/**
12535 * \brief Destroys an external memory object.
12536 *
12537 * Destroys the specified external memory object. Any existing buffers
12538 * and CUDA mipmapped arrays mapped onto this object must no longer be
12539 * used and must be explicitly freed using ::cuMemFree and
12540 * ::cuMipmappedArrayDestroy respectively.
12541 *
12542 * \param extMem - External memory object to be destroyed
12543 *
12544 * \return
12545 * ::CUDA_SUCCESS,
12546 * ::CUDA_ERROR_NOT_INITIALIZED,
12547 * ::CUDA_ERROR_INVALID_HANDLE
12548 * \notefnerr
12549 *
12550 * \sa ::cuImportExternalMemory,
12551 * ::cuExternalMemoryGetMappedBuffer,
12552 * ::cuExternalMemoryGetMappedMipmappedArray
12553 */
12554CUresult CUDAAPI cuDestroyExternalMemory(CUexternalMemory extMem);
12555
12556/**
12557 * \brief Imports an external semaphore
12558 *
12559 * Imports an externally allocated synchronization object and returns
12560 * a handle to that in \p extSem_out.
12561 *
12562 * The properties of the handle being imported must be described in
12563 * \p semHandleDesc. The ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC is
12564 * defined as follows:
12565 *
12566 * \code
12567 typedef struct CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC_st {
12568 CUexternalSemaphoreHandleType type;
12569 union {
12570 int fd;
12571 struct {
12572 void *handle;
12573 const void *name;
12574 } win32;
12575 const void* NvSciSyncObj;
12576 } handle;
12577 unsigned int flags;
12578 } CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC;
12579 * \endcode
12580 *
12581 * where ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::type specifies the type of
12582 * handle being imported. ::CUexternalSemaphoreHandleType is defined
12583 * as:
12584 *
12585 * \code
12586 typedef enum CUexternalSemaphoreHandleType_enum {
12587 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD = 1,
12588 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32 = 2,
12589 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT = 3,
12590 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D12_FENCE = 4,
12591 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D11_FENCE = 5,
12592 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_NVSCISYNC = 6,
12593 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D11_KEYED_MUTEX = 7,
12594 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D11_KEYED_MUTEX_KMT = 8,
12595 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_TIMELINE_SEMAPHORE_FD = 9,
12596 CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_TIMELINE_SEMAPHORE_WIN32 = 10
12597 } CUexternalSemaphoreHandleType;
12598 * \endcode
12599 *
12600 * If ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::type is
12601 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD, then
12602 * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::fd must be a valid
12603 * file descriptor referencing a synchronization object. Ownership of
12604 * the file descriptor is transferred to the CUDA driver when the
12605 * handle is imported successfully. Performing any operations on the
12606 * file descriptor after it is imported results in undefined behavior.
12607 *
12608 * If ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::type is
12609 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32, then exactly one
12610 * of ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::handle and
12611 * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::name must not be
12612 * NULL. If
12613 * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::handle
12614 * is not NULL, then it must represent a valid shared NT handle that
12615 * references a synchronization object. Ownership of this handle is
12616 * not transferred to CUDA after the import operation, so the
12617 * application must release the handle using the appropriate system
12618 * call. If ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::name
12619 * is not NULL, then it must name a valid synchronization object.
12620 *
12621 * If ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::type is
12622 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT, then
12623 * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::handle must
12624 * be non-NULL and
12625 * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::name
12626 * must be NULL. The handle specified must be a globally shared KMT
12627 * handle. This handle does not hold a reference to the underlying
12628 * object, and thus will be invalid when all references to the
12629 * synchronization object are destroyed.
12630 *
12631 * If ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::type is
12632 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D12_FENCE, then exactly one
12633 * of ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::handle and
12634 * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::name must not be
12635 * NULL. If
12636 * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::handle
12637 * is not NULL, then it must represent a valid shared NT handle that
12638 * is returned by ID3D12Device::CreateSharedHandle when referring to a
12639 * ID3D12Fence object. This handle holds a reference to the underlying
12640 * object. If
12641 * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::name
12642 * is not NULL, then it must name a valid synchronization object that
12643 * refers to a valid ID3D12Fence object.
12644 *
12645 * If ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::type is
12646 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D11_FENCE, then
12647 * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::handle
12648 * represents a valid shared NT handle that is returned by
12649 * ID3D11Fence::CreateSharedHandle. If
12650 * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::name
12651 * is not NULL, then it must name a valid synchronization object that
12652 * refers to a valid ID3D11Fence object.
12653 *
12654 * If ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::type is
12655 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_NVSCISYNC, then
12656 * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::nvSciSyncObj
12657 * represents a valid NvSciSyncObj.
12658 *
12659 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D11_KEYED_MUTEX, then
12660 * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::handle
12661 * represents a valid shared NT handle that
12662 * is returned by IDXGIResource1::CreateSharedHandle when referring to
12663 * a IDXGIKeyedMutex object. If
12664 * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::name
12665 * is not NULL, then it must name a valid synchronization object that
12666 * refers to a valid IDXGIKeyedMutex object.
12667 *
12668 * If ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::type is
12669 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D11_KEYED_MUTEX_KMT, then
12670 * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::handle
12671 * represents a valid shared KMT handle that
12672 * is returned by IDXGIResource::GetSharedHandle when referring to
12673 * a IDXGIKeyedMutex object and
12674 * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::name must be NULL.
12675 *
12676 * If ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::type is
12677 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_TIMELINE_SEMAPHORE_FD, then
12678 * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::fd must be a valid
12679 * file descriptor referencing a synchronization object. Ownership of
12680 * the file descriptor is transferred to the CUDA driver when the
12681 * handle is imported successfully. Performing any operations on the
12682 * file descriptor after it is imported results in undefined behavior.
12683 *
12684 * If ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::type is
12685 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_TIMELINE_SEMAPHORE_WIN32, then exactly one
12686 * of ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::handle and
12687 * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::name must not be
12688 * NULL. If
12689 * ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::handle
12690 * is not NULL, then it must represent a valid shared NT handle that
12691 * references a synchronization object. Ownership of this handle is
12692 * not transferred to CUDA after the import operation, so the
12693 * application must release the handle using the appropriate system
12694 * call. If ::CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC::handle::win32::name
12695 * is not NULL, then it must name a valid synchronization object.
12696 *
12697 * \param extSem_out - Returned handle to an external semaphore
12698 * \param semHandleDesc - Semaphore import handle descriptor
12699 *
12700 * \return
12701 * ::CUDA_SUCCESS,
12702 * ::CUDA_ERROR_NOT_INITIALIZED,
12703 * ::CUDA_ERROR_NOT_SUPPORTED,
12704 * ::CUDA_ERROR_INVALID_HANDLE
12705 * \notefnerr
12706 *
12707 * \sa ::cuDestroyExternalSemaphore,
12708 * ::cuSignalExternalSemaphoresAsync,
12709 * ::cuWaitExternalSemaphoresAsync
12710 */
12711CUresult CUDAAPI cuImportExternalSemaphore(CUexternalSemaphore *extSem_out, const CUDA_EXTERNAL_SEMAPHORE_HANDLE_DESC *semHandleDesc);
12712
12713/**
12714 * \brief Signals a set of external semaphore objects
12715 *
12716 * Enqueues a signal operation on a set of externally allocated
12717 * semaphore object in the specified stream. The operations will be
12718 * executed when all prior operations in the stream complete.
12719 *
12720 * The exact semantics of signaling a semaphore depends on the type of
12721 * the object.
12722 *
12723 * If the semaphore object is any one of the following types:
12724 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD,
12725 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32,
12726 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT
12727 * then signaling the semaphore will set it to the signaled state.
12728 *
12729 * If the semaphore object is any one of the following types:
12730 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D12_FENCE,
12731 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D11_FENCE,
12732 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_TIMELINE_SEMAPHORE_FD,
12733 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_TIMELINE_SEMAPHORE_WIN32
12734 * then the semaphore will be set to the value specified in
12735 * ::CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS::params::fence::value.
12736 *
12737 * If the semaphore object is of the type ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_NVSCISYNC
12738 * this API sets ::CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS::params::nvSciSync::fence
12739 * to a value that can be used by subsequent waiters of the same NvSciSync object
12740 * to order operations with those currently submitted in \p stream. Such an update
12741 * will overwrite previous contents of
12742 * ::CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS::params::nvSciSync::fence. By default,
12743 * signaling such an external semaphore object causes appropriate memory synchronization
12744 * operations to be performed over all external memory objects that are imported as
12745 * ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_NVSCIBUF. This ensures that any subsequent accesses
12746 * made by other importers of the same set of NvSciBuf memory object(s) are coherent.
12747 * These operations can be skipped by specifying the flag
12748 * ::CUDA_EXTERNAL_SEMAPHORE_SIGNAL_SKIP_NVSCIBUF_MEMSYNC, which can be used as a
12749 * performance optimization when data coherency is not required. But specifying this
12750 * flag in scenarios where data coherency is required results in undefined behavior.
12751 * Also, for semaphore object of the type ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_NVSCISYNC,
12752 * if the NvSciSyncAttrList used to create the NvSciSyncObj had not set the flags in
12753 * ::cuDeviceGetNvSciSyncAttributes to CUDA_NVSCISYNC_ATTR_SIGNAL, this API will return
12754 * CUDA_ERROR_NOT_SUPPORTED.
12755 *
12756 * If the semaphore object is any one of the following types:
12757 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D11_KEYED_MUTEX,
12758 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D11_KEYED_MUTEX_KMT
12759 * then the keyed mutex will be released with the key specified in
12760 * ::CUDA_EXTERNAL_SEMAPHORE_PARAMS::params::keyedmutex::key.
12761 *
12762 * \param extSemArray - Set of external semaphores to be signaled
12763 * \param paramsArray - Array of semaphore parameters
12764 * \param numExtSems - Number of semaphores to signal
12765 * \param stream - Stream to enqueue the signal operations in
12766 *
12767 * \return
12768 * ::CUDA_SUCCESS,
12769 * ::CUDA_ERROR_NOT_INITIALIZED,
12770 * ::CUDA_ERROR_INVALID_HANDLE,
12771 * ::CUDA_ERROR_NOT_SUPPORTED
12772 * \notefnerr
12773 *
12774 * \sa ::cuImportExternalSemaphore,
12775 * ::cuDestroyExternalSemaphore,
12776 * ::cuWaitExternalSemaphoresAsync
12777 */
12778CUresult CUDAAPI cuSignalExternalSemaphoresAsync(const CUexternalSemaphore *extSemArray, const CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS *paramsArray, unsigned int numExtSems, CUstream stream);
12779
12780/**
12781 * \brief Waits on a set of external semaphore objects
12782 *
12783 * Enqueues a wait operation on a set of externally allocated
12784 * semaphore object in the specified stream. The operations will be
12785 * executed when all prior operations in the stream complete.
12786 *
12787 * The exact semantics of waiting on a semaphore depends on the type
12788 * of the object.
12789 *
12790 * If the semaphore object is any one of the following types:
12791 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD,
12792 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32,
12793 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_WIN32_KMT
12794 * then waiting on the semaphore will wait until the semaphore reaches
12795 * the signaled state. The semaphore will then be reset to the
12796 * unsignaled state. Therefore for every signal operation, there can
12797 * only be one wait operation.
12798 *
12799 * If the semaphore object is any one of the following types:
12800 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D12_FENCE,
12801 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D11_FENCE,
12802 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_TIMELINE_SEMAPHORE_FD,
12803 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_TIMELINE_SEMAPHORE_WIN32
12804 * then waiting on the semaphore will wait until the value of the
12805 * semaphore is greater than or equal to
12806 * ::CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS::params::fence::value.
12807 *
12808 * If the semaphore object is of the type ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_NVSCISYNC
12809 * then, waiting on the semaphore will wait until the
12810 * ::CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS::params::nvSciSync::fence is signaled by the
12811 * signaler of the NvSciSyncObj that was associated with this semaphore object.
12812 * By default, waiting on such an external semaphore object causes appropriate
12813 * memory synchronization operations to be performed over all external memory objects
12814 * that are imported as ::CU_EXTERNAL_MEMORY_HANDLE_TYPE_NVSCIBUF. This ensures that
12815 * any subsequent accesses made by other importers of the same set of NvSciBuf memory
12816 * object(s) are coherent. These operations can be skipped by specifying the flag
12817 * ::CUDA_EXTERNAL_SEMAPHORE_WAIT_SKIP_NVSCIBUF_MEMSYNC, which can be used as a
12818 * performance optimization when data coherency is not required. But specifying this
12819 * flag in scenarios where data coherency is required results in undefined behavior.
12820 * Also, for semaphore object of the type ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_NVSCISYNC,
12821 * if the NvSciSyncAttrList used to create the NvSciSyncObj had not set the flags in
12822 * ::cuDeviceGetNvSciSyncAttributes to CUDA_NVSCISYNC_ATTR_WAIT, this API will return
12823 * CUDA_ERROR_NOT_SUPPORTED.
12824 *
12825 * If the semaphore object is any one of the following types:
12826 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D11_KEYED_MUTEX,
12827 * ::CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D11_KEYED_MUTEX_KMT
12828 * then the keyed mutex will be acquired when it is released with the key
12829 * specified in ::CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS::params::keyedmutex::key
12830 * or until the timeout specified by
12831 * ::CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS::params::keyedmutex::timeoutMs
12832 * has lapsed. The timeout interval can either be a finite value
12833 * specified in milliseconds or an infinite value. In case an infinite
12834 * value is specified the timeout never elapses. The windows INFINITE
12835 * macro must be used to specify infinite timeout.
12836 *
12837 * \param extSemArray - External semaphores to be waited on
12838 * \param paramsArray - Array of semaphore parameters
12839 * \param numExtSems - Number of semaphores to wait on
12840 * \param stream - Stream to enqueue the wait operations in
12841 *
12842 * \return
12843 * ::CUDA_SUCCESS,
12844 * ::CUDA_ERROR_NOT_INITIALIZED,
12845 * ::CUDA_ERROR_INVALID_HANDLE,
12846 * ::CUDA_ERROR_NOT_SUPPORTED,
12847 * ::CUDA_ERROR_TIMEOUT
12848 * \notefnerr
12849 *
12850 * \sa ::cuImportExternalSemaphore,
12851 * ::cuDestroyExternalSemaphore,
12852 * ::cuSignalExternalSemaphoresAsync
12853 */
12854CUresult CUDAAPI cuWaitExternalSemaphoresAsync(const CUexternalSemaphore *extSemArray, const CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS *paramsArray, unsigned int numExtSems, CUstream stream);
12855
12856/**
12857 * \brief Destroys an external semaphore
12858 *
12859 * Destroys an external semaphore object and releases any references
12860 * to the underlying resource. Any outstanding signals or waits must
12861 * have completed before the semaphore is destroyed.
12862 *
12863 * \param extSem - External semaphore to be destroyed
12864 *
12865 * \return
12866 * ::CUDA_SUCCESS,
12867 * ::CUDA_ERROR_NOT_INITIALIZED,
12868 * ::CUDA_ERROR_INVALID_HANDLE
12869 * \notefnerr
12870 *
12871 * \sa ::cuImportExternalSemaphore,
12872 * ::cuSignalExternalSemaphoresAsync,
12873 * ::cuWaitExternalSemaphoresAsync
12874 */
12875CUresult CUDAAPI cuDestroyExternalSemaphore(CUexternalSemaphore extSem);
12876
12877/** @} */ /* END CUDA_EXTRES_INTEROP */
12878
12879/**
12880 * \defgroup CUDA_MEMOP Stream memory operations
12881 *
12882 * ___MANBRIEF___ Stream memory operations of the low-level CUDA driver API
12883 * (___CURRENT_FILE___) ___ENDMANBRIEF___
12884 *
12885 * This section describes the stream memory operations of the low-level CUDA
12886 * driver application programming interface.
12887 *
12888 * The whole set of operations is disabled by default. Users are required
12889 * to explicitly enable them, e.g. on Linux by passing the kernel module
12890 * parameter shown below:
12891 * modprobe nvidia NVreg_EnableStreamMemOPs=1
12892 * There is currently no way to enable these operations on other operating
12893 * systems.
12894 *
12895 * Users can programmatically query whether the device supports these
12896 * operations with ::cuDeviceGetAttribute() and
12897 * ::CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_MEM_OPS.
12898 *
12899 * Support for the ::CU_STREAM_WAIT_VALUE_NOR flag can be queried with
12900 * ::CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_WAIT_VALUE_NOR.
12901 *
12902 * Support for the ::cuStreamWriteValue64() and ::cuStreamWaitValue64()
12903 * functions, as well as for the ::CU_STREAM_MEM_OP_WAIT_VALUE_64 and
12904 * ::CU_STREAM_MEM_OP_WRITE_VALUE_64 flags, can be queried with
12905 * ::CU_DEVICE_ATTRIBUTE_CAN_USE_64_BIT_STREAM_MEM_OPS.
12906 *
12907 * Support for both ::CU_STREAM_WAIT_VALUE_FLUSH and
12908 * ::CU_STREAM_MEM_OP_FLUSH_REMOTE_WRITES requires dedicated platform
12909 * hardware features and can be queried with ::cuDeviceGetAttribute() and
12910 * ::CU_DEVICE_ATTRIBUTE_CAN_FLUSH_REMOTE_WRITES.
12911 *
12912 * Note that all memory pointers passed as parameters to these operations
12913 * are device pointers. Where necessary a device pointer should be
12914 * obtained, for example with ::cuMemHostGetDevicePointer().
12915 *
12916 * None of the operations accepts pointers to managed memory buffers
12917 * (::cuMemAllocManaged).
12918 *
12919 * @{
12920 */
12921
12922/**
12923 * \brief Wait on a memory location
12924 *
12925 * Enqueues a synchronization of the stream on the given memory location. Work
12926 * ordered after the operation will block until the given condition on the
12927 * memory is satisfied. By default, the condition is to wait for
12928 * (int32_t)(*addr - value) >= 0, a cyclic greater-or-equal.
12929 * Other condition types can be specified via \p flags.
12930 *
12931 * If the memory was registered via ::cuMemHostRegister(), the device pointer
12932 * should be obtained with ::cuMemHostGetDevicePointer(). This function cannot
12933 * be used with managed memory (::cuMemAllocManaged).
12934 *
12935 * Support for this can be queried with ::cuDeviceGetAttribute() and
12936 * ::CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_MEM_OPS.
12937 *
12938 * Support for CU_STREAM_WAIT_VALUE_NOR can be queried with ::cuDeviceGetAttribute() and
12939 * ::CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_WAIT_VALUE_NOR.
12940 *
12941 * \param stream The stream to synchronize on the memory location.
12942 * \param addr The memory location to wait on.
12943 * \param value The value to compare with the memory location.
12944 * \param flags See ::CUstreamWaitValue_flags.
12945 *
12946 * \return
12947 * ::CUDA_SUCCESS,
12948 * ::CUDA_ERROR_INVALID_VALUE,
12949 * ::CUDA_ERROR_NOT_SUPPORTED
12950 * \notefnerr
12951 *
12952 * \sa ::cuStreamWaitValue64,
12953 * ::cuStreamWriteValue32,
12954 * ::cuStreamWriteValue64,
12955 * ::cuStreamBatchMemOp,
12956 * ::cuMemHostRegister,
12957 * ::cuStreamWaitEvent
12958 */
12959CUresult CUDAAPI cuStreamWaitValue32(CUstream stream, CUdeviceptr addr, cuuint32_t value, unsigned int flags);
12960
12961/**
12962 * \brief Wait on a memory location
12963 *
12964 * Enqueues a synchronization of the stream on the given memory location. Work
12965 * ordered after the operation will block until the given condition on the
12966 * memory is satisfied. By default, the condition is to wait for
12967 * (int64_t)(*addr - value) >= 0, a cyclic greater-or-equal.
12968 * Other condition types can be specified via \p flags.
12969 *
12970 * If the memory was registered via ::cuMemHostRegister(), the device pointer
12971 * should be obtained with ::cuMemHostGetDevicePointer().
12972 *
12973 * Support for this can be queried with ::cuDeviceGetAttribute() and
12974 * ::CU_DEVICE_ATTRIBUTE_CAN_USE_64_BIT_STREAM_MEM_OPS.
12975 *
12976 * \param stream The stream to synchronize on the memory location.
12977 * \param addr The memory location to wait on.
12978 * \param value The value to compare with the memory location.
12979 * \param flags See ::CUstreamWaitValue_flags.
12980 *
12981 * \return
12982 * ::CUDA_SUCCESS,
12983 * ::CUDA_ERROR_INVALID_VALUE,
12984 * ::CUDA_ERROR_NOT_SUPPORTED
12985 * \notefnerr
12986 *
12987 * \sa ::cuStreamWaitValue32,
12988 * ::cuStreamWriteValue32,
12989 * ::cuStreamWriteValue64,
12990 * ::cuStreamBatchMemOp,
12991 * ::cuMemHostRegister,
12992 * ::cuStreamWaitEvent
12993 */
12994CUresult CUDAAPI cuStreamWaitValue64(CUstream stream, CUdeviceptr addr, cuuint64_t value, unsigned int flags);
12995
12996/**
12997 * \brief Write a value to memory
12998 *
12999 * Write a value to memory. Unless the ::CU_STREAM_WRITE_VALUE_NO_MEMORY_BARRIER
13000 * flag is passed, the write is preceded by a system-wide memory fence,
13001 * equivalent to a __threadfence_system() but scoped to the stream
13002 * rather than a CUDA thread.
13003 *
13004 * If the memory was registered via ::cuMemHostRegister(), the device pointer
13005 * should be obtained with ::cuMemHostGetDevicePointer(). This function cannot
13006 * be used with managed memory (::cuMemAllocManaged).
13007 *
13008 * Support for this can be queried with ::cuDeviceGetAttribute() and
13009 * ::CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_MEM_OPS.
13010 *
13011 * \param stream The stream to do the write in.
13012 * \param addr The device address to write to.
13013 * \param value The value to write.
13014 * \param flags See ::CUstreamWriteValue_flags.
13015 *
13016 * \return
13017 * ::CUDA_SUCCESS,
13018 * ::CUDA_ERROR_INVALID_VALUE,
13019 * ::CUDA_ERROR_NOT_SUPPORTED
13020 * \notefnerr
13021 *
13022 * \sa ::cuStreamWriteValue64,
13023 * ::cuStreamWaitValue32,
13024 * ::cuStreamWaitValue64,
13025 * ::cuStreamBatchMemOp,
13026 * ::cuMemHostRegister,
13027 * ::cuEventRecord
13028 */
13029CUresult CUDAAPI cuStreamWriteValue32(CUstream stream, CUdeviceptr addr, cuuint32_t value, unsigned int flags);
13030
13031/**
13032 * \brief Write a value to memory
13033 *
13034 * Write a value to memory. Unless the ::CU_STREAM_WRITE_VALUE_NO_MEMORY_BARRIER
13035 * flag is passed, the write is preceded by a system-wide memory fence,
13036 * equivalent to a __threadfence_system() but scoped to the stream
13037 * rather than a CUDA thread.
13038 *
13039 * If the memory was registered via ::cuMemHostRegister(), the device pointer
13040 * should be obtained with ::cuMemHostGetDevicePointer().
13041 *
13042 * Support for this can be queried with ::cuDeviceGetAttribute() and
13043 * ::CU_DEVICE_ATTRIBUTE_CAN_USE_64_BIT_STREAM_MEM_OPS.
13044 *
13045 * \param stream The stream to do the write in.
13046 * \param addr The device address to write to.
13047 * \param value The value to write.
13048 * \param flags See ::CUstreamWriteValue_flags.
13049 *
13050 * \return
13051 * ::CUDA_SUCCESS,
13052 * ::CUDA_ERROR_INVALID_VALUE,
13053 * ::CUDA_ERROR_NOT_SUPPORTED
13054 * \notefnerr
13055 *
13056 * \sa ::cuStreamWriteValue32,
13057 * ::cuStreamWaitValue32,
13058 * ::cuStreamWaitValue64,
13059 * ::cuStreamBatchMemOp,
13060 * ::cuMemHostRegister,
13061 * ::cuEventRecord
13062 */
13063CUresult CUDAAPI cuStreamWriteValue64(CUstream stream, CUdeviceptr addr, cuuint64_t value, unsigned int flags);
13064
13065/**
13066 * \brief Batch operations to synchronize the stream via memory operations
13067 *
13068 * This is a batch version of ::cuStreamWaitValue32() and ::cuStreamWriteValue32().
13069 * Batching operations may avoid some performance overhead in both the API call
13070 * and the device execution versus adding them to the stream in separate API
13071 * calls. The operations are enqueued in the order they appear in the array.
13072 *
13073 * See ::CUstreamBatchMemOpType for the full set of supported operations, and
13074 * ::cuStreamWaitValue32(), ::cuStreamWaitValue64(), ::cuStreamWriteValue32(),
13075 * and ::cuStreamWriteValue64() for details of specific operations.
13076 *
13077 * Basic support for this can be queried with ::cuDeviceGetAttribute() and
13078 * ::CU_DEVICE_ATTRIBUTE_CAN_USE_STREAM_MEM_OPS. See related APIs for details
13079 * on querying support for specific operations.
13080 *
13081 * \param stream The stream to enqueue the operations in.
13082 * \param count The number of operations in the array. Must be less than 256.
13083 * \param paramArray The types and parameters of the individual operations.
13084 * \param flags Reserved for future expansion; must be 0.
13085 *
13086 * \return
13087 * ::CUDA_SUCCESS,
13088 * ::CUDA_ERROR_INVALID_VALUE,
13089 * ::CUDA_ERROR_NOT_SUPPORTED
13090 * \notefnerr
13091 *
13092 * \sa ::cuStreamWaitValue32,
13093 * ::cuStreamWaitValue64,
13094 * ::cuStreamWriteValue32,
13095 * ::cuStreamWriteValue64,
13096 * ::cuMemHostRegister
13097 */
13098CUresult CUDAAPI cuStreamBatchMemOp(CUstream stream, unsigned int count, CUstreamBatchMemOpParams *paramArray, unsigned int flags);
13099
13100/** @} */ /* END CUDA_MEMOP */
13101
13102/**
13103 * \defgroup CUDA_EXEC Execution Control
13104 *
13105 * ___MANBRIEF___ execution control functions of the low-level CUDA driver API
13106 * (___CURRENT_FILE___) ___ENDMANBRIEF___
13107 *
13108 * This section describes the execution control functions of the low-level CUDA
13109 * driver application programming interface.
13110 *
13111 * @{
13112 */
13113
13114/**
13115 * \brief Returns information about a function
13116 *
13117 * Returns in \p *pi the integer value of the attribute \p attrib on the kernel
13118 * given by \p hfunc. The supported attributes are:
13119 * - ::CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK: The maximum number of threads
13120 * per block, beyond which a launch of the function would fail. This number
13121 * depends on both the function and the device on which the function is
13122 * currently loaded.
13123 * - ::CU_FUNC_ATTRIBUTE_SHARED_SIZE_BYTES: The size in bytes of
13124 * statically-allocated shared memory per block required by this function.
13125 * This does not include dynamically-allocated shared memory requested by
13126 * the user at runtime.
13127 * - ::CU_FUNC_ATTRIBUTE_CONST_SIZE_BYTES: The size in bytes of user-allocated
13128 * constant memory required by this function.
13129 * - ::CU_FUNC_ATTRIBUTE_LOCAL_SIZE_BYTES: The size in bytes of local memory
13130 * used by each thread of this function.
13131 * - ::CU_FUNC_ATTRIBUTE_NUM_REGS: The number of registers used by each thread
13132 * of this function.
13133 * - ::CU_FUNC_ATTRIBUTE_PTX_VERSION: The PTX virtual architecture version for
13134 * which the function was compiled. This value is the major PTX version * 10
13135 * + the minor PTX version, so a PTX version 1.3 function would return the
13136 * value 13. Note that this may return the undefined value of 0 for cubins
13137 * compiled prior to CUDA 3.0.
13138 * - ::CU_FUNC_ATTRIBUTE_BINARY_VERSION: The binary architecture version for
13139 * which the function was compiled. This value is the major binary
13140 * version * 10 + the minor binary version, so a binary version 1.3 function
13141 * would return the value 13. Note that this will return a value of 10 for
13142 * legacy cubins that do not have a properly-encoded binary architecture
13143 * version.
13144 * - ::CU_FUNC_CACHE_MODE_CA: The attribute to indicate whether the function has
13145 * been compiled with user specified option "-Xptxas --dlcm=ca" set .
13146 * - ::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES: The maximum size in bytes of
13147 * dynamically-allocated shared memory.
13148 * - ::CU_FUNC_ATTRIBUTE_PREFERRED_SHARED_MEMORY_CARVEOUT: Preferred shared memory-L1
13149 * cache split ratio in percent of total shared memory.
13150 *
13151 * \param pi - Returned attribute value
13152 * \param attrib - Attribute requested
13153 * \param hfunc - Function to query attribute of
13154 *
13155 * \return
13156 * ::CUDA_SUCCESS,
13157 * ::CUDA_ERROR_DEINITIALIZED,
13158 * ::CUDA_ERROR_NOT_INITIALIZED,
13159 * ::CUDA_ERROR_INVALID_CONTEXT,
13160 * ::CUDA_ERROR_INVALID_HANDLE,
13161 * ::CUDA_ERROR_INVALID_VALUE
13162 * \notefnerr
13163 *
13164 * \sa ::cuCtxGetCacheConfig,
13165 * ::cuCtxSetCacheConfig,
13166 * ::cuFuncSetCacheConfig,
13167 * ::cuLaunchKernel,
13168 * ::cudaFuncGetAttributes,
13169 * ::cudaFuncSetAttribute
13170 */
13171CUresult CUDAAPI cuFuncGetAttribute(int *pi, CUfunction_attribute attrib, CUfunction hfunc);
13172
13173/**
13174 * \brief Sets information about a function
13175 *
13176 * This call sets the value of a specified attribute \p attrib on the kernel given
13177 * by \p hfunc to an integer value specified by \p val
13178 * This function returns CUDA_SUCCESS if the new value of the attribute could be
13179 * successfully set. If the set fails, this call will return an error.
13180 * Not all attributes can have values set. Attempting to set a value on a read-only
13181 * attribute will result in an error (CUDA_ERROR_INVALID_VALUE)
13182 *
13183 * Supported attributes for the cuFuncSetAttribute call are:
13184 * - ::CU_FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES: This maximum size in bytes of
13185 * dynamically-allocated shared memory. The value should contain the requested
13186 * maximum size of dynamically-allocated shared memory. The sum of this value and
13187 * the function attribute ::CU_FUNC_ATTRIBUTE_SHARED_SIZE_BYTES cannot exceed the
13188 * device attribute ::CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_BLOCK_OPTIN.
13189 * The maximal size of requestable dynamic shared memory may differ by GPU
13190 * architecture.
13191 * - ::CU_FUNC_ATTRIBUTE_PREFERRED_SHARED_MEMORY_CARVEOUT: On devices where the L1
13192 * cache and shared memory use the same hardware resources, this sets the shared memory
13193 * carveout preference, in percent of the total shared memory.
13194 * See ::CU_DEVICE_ATTRIBUTE_MAX_SHARED_MEMORY_PER_MULTIPROCESSOR
13195 * This is only a hint, and the driver can choose a different ratio if required to execute the function.
13196 *
13197 * \param hfunc - Function to query attribute of
13198 * \param attrib - Attribute requested
13199 * \param value - The value to set
13200 *
13201 * \return
13202 * ::CUDA_SUCCESS,
13203 * ::CUDA_ERROR_DEINITIALIZED,
13204 * ::CUDA_ERROR_NOT_INITIALIZED,
13205 * ::CUDA_ERROR_INVALID_CONTEXT,
13206 * ::CUDA_ERROR_INVALID_HANDLE,
13207 * ::CUDA_ERROR_INVALID_VALUE
13208 * \notefnerr
13209 *
13210 * \sa ::cuCtxGetCacheConfig,
13211 * ::cuCtxSetCacheConfig,
13212 * ::cuFuncSetCacheConfig,
13213 * ::cuLaunchKernel,
13214 * ::cudaFuncGetAttributes,
13215 * ::cudaFuncSetAttribute
13216 */
13217CUresult CUDAAPI cuFuncSetAttribute(CUfunction hfunc, CUfunction_attribute attrib, int value);
13218
13219/**
13220 * \brief Sets the preferred cache configuration for a device function
13221 *
13222 * On devices where the L1 cache and shared memory use the same hardware
13223 * resources, this sets through \p config the preferred cache configuration for
13224 * the device function \p hfunc. This is only a preference. The driver will use
13225 * the requested configuration if possible, but it is free to choose a different
13226 * configuration if required to execute \p hfunc. Any context-wide preference
13227 * set via ::cuCtxSetCacheConfig() will be overridden by this per-function
13228 * setting unless the per-function setting is ::CU_FUNC_CACHE_PREFER_NONE. In
13229 * that case, the current context-wide setting will be used.
13230 *
13231 * This setting does nothing on devices where the size of the L1 cache and
13232 * shared memory are fixed.
13233 *
13234 * Launching a kernel with a different preference than the most recent
13235 * preference setting may insert a device-side synchronization point.
13236 *
13237 *
13238 * The supported cache configurations are:
13239 * - ::CU_FUNC_CACHE_PREFER_NONE: no preference for shared memory or L1 (default)
13240 * - ::CU_FUNC_CACHE_PREFER_SHARED: prefer larger shared memory and smaller L1 cache
13241 * - ::CU_FUNC_CACHE_PREFER_L1: prefer larger L1 cache and smaller shared memory
13242 * - ::CU_FUNC_CACHE_PREFER_EQUAL: prefer equal sized L1 cache and shared memory
13243 *
13244 * \param hfunc - Kernel to configure cache for
13245 * \param config - Requested cache configuration
13246 *
13247 * \return
13248 * ::CUDA_SUCCESS,
13249 * ::CUDA_ERROR_INVALID_VALUE,
13250 * ::CUDA_ERROR_DEINITIALIZED,
13251 * ::CUDA_ERROR_NOT_INITIALIZED,
13252 * ::CUDA_ERROR_INVALID_CONTEXT
13253 * \notefnerr
13254 *
13255 * \sa ::cuCtxGetCacheConfig,
13256 * ::cuCtxSetCacheConfig,
13257 * ::cuFuncGetAttribute,
13258 * ::cuLaunchKernel,
13259 * ::cudaFuncSetCacheConfig
13260 */
13261CUresult CUDAAPI cuFuncSetCacheConfig(CUfunction hfunc, CUfunc_cache config);
13262
13263/**
13264 * \brief Sets the shared memory configuration for a device function.
13265 *
13266 * On devices with configurable shared memory banks, this function will
13267 * force all subsequent launches of the specified device function to have
13268 * the given shared memory bank size configuration. On any given launch of the
13269 * function, the shared memory configuration of the device will be temporarily
13270 * changed if needed to suit the function's preferred configuration. Changes in
13271 * shared memory configuration between subsequent launches of functions,
13272 * may introduce a device side synchronization point.
13273 *
13274 * Any per-function setting of shared memory bank size set via
13275 * ::cuFuncSetSharedMemConfig will override the context wide setting set with
13276 * ::cuCtxSetSharedMemConfig.
13277 *
13278 * Changing the shared memory bank size will not increase shared memory usage
13279 * or affect occupancy of kernels, but may have major effects on performance.
13280 * Larger bank sizes will allow for greater potential bandwidth to shared memory,
13281 * but will change what kinds of accesses to shared memory will result in bank
13282 * conflicts.
13283 *
13284 * This function will do nothing on devices with fixed shared memory bank size.
13285 *
13286 * The supported bank configurations are:
13287 * - ::CU_SHARED_MEM_CONFIG_DEFAULT_BANK_SIZE: use the context's shared memory
13288 * configuration when launching this function.
13289 * - ::CU_SHARED_MEM_CONFIG_FOUR_BYTE_BANK_SIZE: set shared memory bank width to
13290 * be natively four bytes when launching this function.
13291 * - ::CU_SHARED_MEM_CONFIG_EIGHT_BYTE_BANK_SIZE: set shared memory bank width to
13292 * be natively eight bytes when launching this function.
13293 *
13294 * \param hfunc - kernel to be given a shared memory config
13295 * \param config - requested shared memory configuration
13296 *
13297 * \return
13298 * ::CUDA_SUCCESS,
13299 * ::CUDA_ERROR_INVALID_VALUE,
13300 * ::CUDA_ERROR_DEINITIALIZED,
13301 * ::CUDA_ERROR_NOT_INITIALIZED,
13302 * ::CUDA_ERROR_INVALID_CONTEXT
13303 * \notefnerr
13304 *
13305 * \sa ::cuCtxGetCacheConfig,
13306 * ::cuCtxSetCacheConfig,
13307 * ::cuCtxGetSharedMemConfig,
13308 * ::cuCtxSetSharedMemConfig,
13309 * ::cuFuncGetAttribute,
13310 * ::cuLaunchKernel,
13311 * ::cudaFuncSetSharedMemConfig
13312 */
13313CUresult CUDAAPI cuFuncSetSharedMemConfig(CUfunction hfunc, CUsharedconfig config);
13314
13315/**
13316 * \brief Returns a module handle
13317 *
13318 * Returns in \p *hmod the handle of the module that function \p hfunc
13319 * is located in. The lifetime of the module corresponds to the lifetime of
13320 * the context it was loaded in or until the module is explicitly unloaded.
13321 *
13322 * The CUDA runtime manages its own modules loaded into the primary context.
13323 * If the handle returned by this API refers to a module loaded by the CUDA runtime,
13324 * calling ::cuModuleUnload() on that module will result in undefined behavior.
13325 *
13326 * \param hmod - Returned module handle
13327 * \param hfunc - Function to retrieve module for
13328 *
13329 * \return
13330 * ::CUDA_SUCCESS,
13331 * ::CUDA_ERROR_DEINITIALIZED,
13332 * ::CUDA_ERROR_NOT_INITIALIZED,
13333 * ::CUDA_ERROR_INVALID_CONTEXT,
13334 * ::CUDA_ERROR_INVALID_VALUE,
13335 * ::CUDA_ERROR_NOT_FOUND
13336 * \notefnerr
13337 *
13338 */
13339CUresult CUDAAPI cuFuncGetModule(CUmodule *hmod, CUfunction hfunc);
13340
13341/**
13342 * \brief Launches a CUDA function
13343 *
13344 * Invokes the kernel \p f on a \p gridDimX x \p gridDimY x \p gridDimZ
13345 * grid of blocks. Each block contains \p blockDimX x \p blockDimY x
13346 * \p blockDimZ threads.
13347 *
13348 * \p sharedMemBytes sets the amount of dynamic shared memory that will be
13349 * available to each thread block.
13350 *
13351 * Kernel parameters to \p f can be specified in one of two ways:
13352 *
13353 * 1) Kernel parameters can be specified via \p kernelParams. If \p f
13354 * has N parameters, then \p kernelParams needs to be an array of N
13355 * pointers. Each of \p kernelParams[0] through \p kernelParams[N-1]
13356 * must point to a region of memory from which the actual kernel
13357 * parameter will be copied. The number of kernel parameters and their
13358 * offsets and sizes do not need to be specified as that information is
13359 * retrieved directly from the kernel's image.
13360 *
13361 * 2) Kernel parameters can also be packaged by the application into
13362 * a single buffer that is passed in via the \p extra parameter.
13363 * This places the burden on the application of knowing each kernel
13364 * parameter's size and alignment/padding within the buffer. Here is
13365 * an example of using the \p extra parameter in this manner:
13366 * \code
13367 size_t argBufferSize;
13368 char argBuffer[256];
13369
13370 // populate argBuffer and argBufferSize
13371
13372 void *config[] = {
13373 CU_LAUNCH_PARAM_BUFFER_POINTER, argBuffer,
13374 CU_LAUNCH_PARAM_BUFFER_SIZE, &argBufferSize,
13375 CU_LAUNCH_PARAM_END
13376 };
13377 status = cuLaunchKernel(f, gx, gy, gz, bx, by, bz, sh, s, NULL, config);
13378 * \endcode
13379 *
13380 * The \p extra parameter exists to allow ::cuLaunchKernel to take
13381 * additional less commonly used arguments. \p extra specifies a list of
13382 * names of extra settings and their corresponding values. Each extra
13383 * setting name is immediately followed by the corresponding value. The
13384 * list must be terminated with either NULL or ::CU_LAUNCH_PARAM_END.
13385 *
13386 * - ::CU_LAUNCH_PARAM_END, which indicates the end of the \p extra
13387 * array;
13388 * - ::CU_LAUNCH_PARAM_BUFFER_POINTER, which specifies that the next
13389 * value in \p extra will be a pointer to a buffer containing all
13390 * the kernel parameters for launching kernel \p f;
13391 * - ::CU_LAUNCH_PARAM_BUFFER_SIZE, which specifies that the next
13392 * value in \p extra will be a pointer to a size_t containing the
13393 * size of the buffer specified with ::CU_LAUNCH_PARAM_BUFFER_POINTER;
13394 *
13395 * The error ::CUDA_ERROR_INVALID_VALUE will be returned if kernel
13396 * parameters are specified with both \p kernelParams and \p extra
13397 * (i.e. both \p kernelParams and \p extra are non-NULL).
13398 *
13399 * Calling ::cuLaunchKernel() invalidates the persistent function state
13400 * set through the following deprecated APIs:
13401 * ::cuFuncSetBlockShape(),
13402 * ::cuFuncSetSharedSize(),
13403 * ::cuParamSetSize(),
13404 * ::cuParamSeti(),
13405 * ::cuParamSetf(),
13406 * ::cuParamSetv().
13407 *
13408 * Note that to use ::cuLaunchKernel(), the kernel \p f must either have
13409 * been compiled with toolchain version 3.2 or later so that it will
13410 * contain kernel parameter information, or have no kernel parameters.
13411 * If either of these conditions is not met, then ::cuLaunchKernel() will
13412 * return ::CUDA_ERROR_INVALID_IMAGE.
13413 *
13414 * \param f - Kernel to launch
13415 * \param gridDimX - Width of grid in blocks
13416 * \param gridDimY - Height of grid in blocks
13417 * \param gridDimZ - Depth of grid in blocks
13418 * \param blockDimX - X dimension of each thread block
13419 * \param blockDimY - Y dimension of each thread block
13420 * \param blockDimZ - Z dimension of each thread block
13421 * \param sharedMemBytes - Dynamic shared-memory size per thread block in bytes
13422 * \param hStream - Stream identifier
13423 * \param kernelParams - Array of pointers to kernel parameters
13424 * \param extra - Extra options
13425 *
13426 * \return
13427 * ::CUDA_SUCCESS,
13428 * ::CUDA_ERROR_DEINITIALIZED,
13429 * ::CUDA_ERROR_NOT_INITIALIZED,
13430 * ::CUDA_ERROR_INVALID_CONTEXT,
13431 * ::CUDA_ERROR_INVALID_HANDLE,
13432 * ::CUDA_ERROR_INVALID_IMAGE,
13433 * ::CUDA_ERROR_INVALID_VALUE,
13434 * ::CUDA_ERROR_LAUNCH_FAILED,
13435 * ::CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES,
13436 * ::CUDA_ERROR_LAUNCH_TIMEOUT,
13437 * ::CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING,
13438 * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED
13439 * \note_null_stream
13440 * \notefnerr
13441 *
13442 * \sa ::cuCtxGetCacheConfig,
13443 * ::cuCtxSetCacheConfig,
13444 * ::cuFuncSetCacheConfig,
13445 * ::cuFuncGetAttribute,
13446 * ::cudaLaunchKernel
13447 */
13448CUresult CUDAAPI cuLaunchKernel(CUfunction f,
13449 unsigned int gridDimX,
13450 unsigned int gridDimY,
13451 unsigned int gridDimZ,
13452 unsigned int blockDimX,
13453 unsigned int blockDimY,
13454 unsigned int blockDimZ,
13455 unsigned int sharedMemBytes,
13456 CUstream hStream,
13457 void **kernelParams,
13458 void **extra);
13459
13460/**
13461 * \brief Launches a CUDA function where thread blocks can cooperate and synchronize as they execute
13462 *
13463 * Invokes the kernel \p f on a \p gridDimX x \p gridDimY x \p gridDimZ
13464 * grid of blocks. Each block contains \p blockDimX x \p blockDimY x
13465 * \p blockDimZ threads.
13466 *
13467 * \p sharedMemBytes sets the amount of dynamic shared memory that will be
13468 * available to each thread block.
13469 *
13470 * The device on which this kernel is invoked must have a non-zero value for
13471 * the device attribute ::CU_DEVICE_ATTRIBUTE_COOPERATIVE_LAUNCH.
13472 *
13473 * The total number of blocks launched cannot exceed the maximum number of blocks per
13474 * multiprocessor as returned by ::cuOccupancyMaxActiveBlocksPerMultiprocessor (or
13475 * ::cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags) times the number of multiprocessors
13476 * as specified by the device attribute ::CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT.
13477 *
13478 * The kernel cannot make use of CUDA dynamic parallelism.
13479 *
13480 * Kernel parameters must be specified via \p kernelParams. If \p f
13481 * has N parameters, then \p kernelParams needs to be an array of N
13482 * pointers. Each of \p kernelParams[0] through \p kernelParams[N-1]
13483 * must point to a region of memory from which the actual kernel
13484 * parameter will be copied. The number of kernel parameters and their
13485 * offsets and sizes do not need to be specified as that information is
13486 * retrieved directly from the kernel's image.
13487 *
13488 * Calling ::cuLaunchCooperativeKernel() sets persistent function state that is
13489 * the same as function state set through ::cuLaunchKernel API
13490 *
13491 * When the kernel \p f is launched via ::cuLaunchCooperativeKernel(), the previous
13492 * block shape, shared size and parameter info associated with \p f
13493 * is overwritten.
13494 *
13495 * Note that to use ::cuLaunchCooperativeKernel(), the kernel \p f must either have
13496 * been compiled with toolchain version 3.2 or later so that it will
13497 * contain kernel parameter information, or have no kernel parameters.
13498 * If either of these conditions is not met, then ::cuLaunchCooperativeKernel() will
13499 * return ::CUDA_ERROR_INVALID_IMAGE.
13500 *
13501 * \param f - Kernel to launch
13502 * \param gridDimX - Width of grid in blocks
13503 * \param gridDimY - Height of grid in blocks
13504 * \param gridDimZ - Depth of grid in blocks
13505 * \param blockDimX - X dimension of each thread block
13506 * \param blockDimY - Y dimension of each thread block
13507 * \param blockDimZ - Z dimension of each thread block
13508 * \param sharedMemBytes - Dynamic shared-memory size per thread block in bytes
13509 * \param hStream - Stream identifier
13510 * \param kernelParams - Array of pointers to kernel parameters
13511 *
13512 * \return
13513 * ::CUDA_SUCCESS,
13514 * ::CUDA_ERROR_DEINITIALIZED,
13515 * ::CUDA_ERROR_NOT_INITIALIZED,
13516 * ::CUDA_ERROR_INVALID_CONTEXT,
13517 * ::CUDA_ERROR_INVALID_HANDLE,
13518 * ::CUDA_ERROR_INVALID_IMAGE,
13519 * ::CUDA_ERROR_INVALID_VALUE,
13520 * ::CUDA_ERROR_LAUNCH_FAILED,
13521 * ::CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES,
13522 * ::CUDA_ERROR_LAUNCH_TIMEOUT,
13523 * ::CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING,
13524 * ::CUDA_ERROR_COOPERATIVE_LAUNCH_TOO_LARGE,
13525 * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED
13526 * \note_null_stream
13527 * \notefnerr
13528 *
13529 * \sa ::cuCtxGetCacheConfig,
13530 * ::cuCtxSetCacheConfig,
13531 * ::cuFuncSetCacheConfig,
13532 * ::cuFuncGetAttribute,
13533 * ::cuLaunchCooperativeKernelMultiDevice,
13534 * ::cudaLaunchCooperativeKernel
13535 */
13536CUresult CUDAAPI cuLaunchCooperativeKernel(CUfunction f,
13537 unsigned int gridDimX,
13538 unsigned int gridDimY,
13539 unsigned int gridDimZ,
13540 unsigned int blockDimX,
13541 unsigned int blockDimY,
13542 unsigned int blockDimZ,
13543 unsigned int sharedMemBytes,
13544 CUstream hStream,
13545 void **kernelParams);
13546
13547/**
13548 * \brief Launches CUDA functions on multiple devices where thread blocks can cooperate and synchronize as they execute
13549 *
13550 * \deprecated This function is deprecated as of CUDA 11.3.
13551 *
13552 * Invokes kernels as specified in the \p launchParamsList array where each element
13553 * of the array specifies all the parameters required to perform a single kernel launch.
13554 * These kernels can cooperate and synchronize as they execute. The size of the array is
13555 * specified by \p numDevices.
13556 *
13557 * No two kernels can be launched on the same device. All the devices targeted by this
13558 * multi-device launch must be identical. All devices must have a non-zero value for the
13559 * device attribute ::CU_DEVICE_ATTRIBUTE_COOPERATIVE_MULTI_DEVICE_LAUNCH.
13560 *
13561 * All kernels launched must be identical with respect to the compiled code. Note that
13562 * any __device__, __constant__ or __managed__ variables present in the module that owns
13563 * the kernel launched on each device, are independently instantiated on every device.
13564 * It is the application's responsibility to ensure these variables are initialized and
13565 * used appropriately.
13566 *
13567 * The size of the grids as specified in blocks, the size of the blocks themselves
13568 * and the amount of shared memory used by each thread block must also match across
13569 * all launched kernels.
13570 *
13571 * The streams used to launch these kernels must have been created via either ::cuStreamCreate
13572 * or ::cuStreamCreateWithPriority. The NULL stream or ::CU_STREAM_LEGACY or ::CU_STREAM_PER_THREAD
13573 * cannot be used.
13574 *
13575 * The total number of blocks launched per kernel cannot exceed the maximum number of blocks
13576 * per multiprocessor as returned by ::cuOccupancyMaxActiveBlocksPerMultiprocessor (or
13577 * ::cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags) times the number of multiprocessors
13578 * as specified by the device attribute ::CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT. Since the
13579 * total number of blocks launched per device has to match across all devices, the maximum
13580 * number of blocks that can be launched per device will be limited by the device with the
13581 * least number of multiprocessors.
13582 *
13583 * The kernels cannot make use of CUDA dynamic parallelism.
13584 *
13585 * The ::CUDA_LAUNCH_PARAMS structure is defined as:
13586 * \code
13587 typedef struct CUDA_LAUNCH_PARAMS_st
13588 {
13589 CUfunction function;
13590 unsigned int gridDimX;
13591 unsigned int gridDimY;
13592 unsigned int gridDimZ;
13593 unsigned int blockDimX;
13594 unsigned int blockDimY;
13595 unsigned int blockDimZ;
13596 unsigned int sharedMemBytes;
13597 CUstream hStream;
13598 void **kernelParams;
13599 } CUDA_LAUNCH_PARAMS;
13600 * \endcode
13601 * where:
13602 * - ::CUDA_LAUNCH_PARAMS::function specifies the kernel to be launched. All functions must
13603 * be identical with respect to the compiled code.
13604 * - ::CUDA_LAUNCH_PARAMS::gridDimX is the width of the grid in blocks. This must match across
13605 * all kernels launched.
13606 * - ::CUDA_LAUNCH_PARAMS::gridDimY is the height of the grid in blocks. This must match across
13607 * all kernels launched.
13608 * - ::CUDA_LAUNCH_PARAMS::gridDimZ is the depth of the grid in blocks. This must match across
13609 * all kernels launched.
13610 * - ::CUDA_LAUNCH_PARAMS::blockDimX is the X dimension of each thread block. This must match across
13611 * all kernels launched.
13612 * - ::CUDA_LAUNCH_PARAMS::blockDimX is the Y dimension of each thread block. This must match across
13613 * all kernels launched.
13614 * - ::CUDA_LAUNCH_PARAMS::blockDimZ is the Z dimension of each thread block. This must match across
13615 * all kernels launched.
13616 * - ::CUDA_LAUNCH_PARAMS::sharedMemBytes is the dynamic shared-memory size per thread block in bytes.
13617 * This must match across all kernels launched.
13618 * - ::CUDA_LAUNCH_PARAMS::hStream is the handle to the stream to perform the launch in. This cannot
13619 * be the NULL stream or ::CU_STREAM_LEGACY or ::CU_STREAM_PER_THREAD. The CUDA context associated
13620 * with this stream must match that associated with ::CUDA_LAUNCH_PARAMS::function.
13621 * - ::CUDA_LAUNCH_PARAMS::kernelParams is an array of pointers to kernel parameters. If
13622 * ::CUDA_LAUNCH_PARAMS::function has N parameters, then ::CUDA_LAUNCH_PARAMS::kernelParams
13623 * needs to be an array of N pointers. Each of ::CUDA_LAUNCH_PARAMS::kernelParams[0] through
13624 * ::CUDA_LAUNCH_PARAMS::kernelParams[N-1] must point to a region of memory from which the actual
13625 * kernel parameter will be copied. The number of kernel parameters and their offsets and sizes
13626 * do not need to be specified as that information is retrieved directly from the kernel's image.
13627 *
13628 * By default, the kernel won't begin execution on any GPU until all prior work in all the specified
13629 * streams has completed. This behavior can be overridden by specifying the flag
13630 * ::CUDA_COOPERATIVE_LAUNCH_MULTI_DEVICE_NO_PRE_LAUNCH_SYNC. When this flag is specified, each kernel
13631 * will only wait for prior work in the stream corresponding to that GPU to complete before it begins
13632 * execution.
13633 *
13634 * Similarly, by default, any subsequent work pushed in any of the specified streams will not begin
13635 * execution until the kernels on all GPUs have completed. This behavior can be overridden by specifying
13636 * the flag ::CUDA_COOPERATIVE_LAUNCH_MULTI_DEVICE_NO_POST_LAUNCH_SYNC. When this flag is specified,
13637 * any subsequent work pushed in any of the specified streams will only wait for the kernel launched
13638 * on the GPU corresponding to that stream to complete before it begins execution.
13639 *
13640 * Calling ::cuLaunchCooperativeKernelMultiDevice() sets persistent function state that is
13641 * the same as function state set through ::cuLaunchKernel API when called individually for each
13642 * element in \p launchParamsList.
13643 *
13644 * When kernels are launched via ::cuLaunchCooperativeKernelMultiDevice(), the previous
13645 * block shape, shared size and parameter info associated with each ::CUDA_LAUNCH_PARAMS::function
13646 * in \p launchParamsList is overwritten.
13647 *
13648 * Note that to use ::cuLaunchCooperativeKernelMultiDevice(), the kernels must either have
13649 * been compiled with toolchain version 3.2 or later so that it will
13650 * contain kernel parameter information, or have no kernel parameters.
13651 * If either of these conditions is not met, then ::cuLaunchCooperativeKernelMultiDevice() will
13652 * return ::CUDA_ERROR_INVALID_IMAGE.
13653 *
13654 * \param launchParamsList - List of launch parameters, one per device
13655 * \param numDevices - Size of the \p launchParamsList array
13656 * \param flags - Flags to control launch behavior
13657 *
13658 * \return
13659 * ::CUDA_SUCCESS,
13660 * ::CUDA_ERROR_DEINITIALIZED,
13661 * ::CUDA_ERROR_NOT_INITIALIZED,
13662 * ::CUDA_ERROR_INVALID_CONTEXT,
13663 * ::CUDA_ERROR_INVALID_HANDLE,
13664 * ::CUDA_ERROR_INVALID_IMAGE,
13665 * ::CUDA_ERROR_INVALID_VALUE,
13666 * ::CUDA_ERROR_LAUNCH_FAILED,
13667 * ::CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES,
13668 * ::CUDA_ERROR_LAUNCH_TIMEOUT,
13669 * ::CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING,
13670 * ::CUDA_ERROR_COOPERATIVE_LAUNCH_TOO_LARGE,
13671 * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED
13672 * \note_null_stream
13673 * \notefnerr
13674 *
13675 * \sa ::cuCtxGetCacheConfig,
13676 * ::cuCtxSetCacheConfig,
13677 * ::cuFuncSetCacheConfig,
13678 * ::cuFuncGetAttribute,
13679 * ::cuLaunchCooperativeKernel,
13680 * ::cudaLaunchCooperativeKernelMultiDevice
13681 */
13682__CUDA_DEPRECATED CUresult CUDAAPI cuLaunchCooperativeKernelMultiDevice(CUDA_LAUNCH_PARAMS *launchParamsList, unsigned int numDevices, unsigned int flags);
13683
13684/**
13685 * \brief Enqueues a host function call in a stream
13686 *
13687 * Enqueues a host function to run in a stream. The function will be called
13688 * after currently enqueued work and will block work added after it.
13689 *
13690 * The host function must not make any CUDA API calls. Attempting to use a
13691 * CUDA API may result in ::CUDA_ERROR_NOT_PERMITTED, but this is not required.
13692 * The host function must not perform any synchronization that may depend on
13693 * outstanding CUDA work not mandated to run earlier. Host functions without a
13694 * mandated order (such as in independent streams) execute in undefined order
13695 * and may be serialized.
13696 *
13697 * For the purposes of Unified Memory, execution makes a number of guarantees:
13698 * <ul>
13699 * <li>The stream is considered idle for the duration of the function's
13700 * execution. Thus, for example, the function may always use memory attached
13701 * to the stream it was enqueued in.</li>
13702 * <li>The start of execution of the function has the same effect as
13703 * synchronizing an event recorded in the same stream immediately prior to
13704 * the function. It thus synchronizes streams which have been "joined"
13705 * prior to the function.</li>
13706 * <li>Adding device work to any stream does not have the effect of making
13707 * the stream active until all preceding host functions and stream callbacks
13708 * have executed. Thus, for
13709 * example, a function might use global attached memory even if work has
13710 * been added to another stream, if the work has been ordered behind the
13711 * function call with an event.</li>
13712 * <li>Completion of the function does not cause a stream to become
13713 * active except as described above. The stream will remain idle
13714 * if no device work follows the function, and will remain idle across
13715 * consecutive host functions or stream callbacks without device work in
13716 * between. Thus, for example,
13717 * stream synchronization can be done by signaling from a host function at the
13718 * end of the stream.</li>
13719 * </ul>
13720 *
13721 * Note that, in contrast to ::cuStreamAddCallback, the function will not be
13722 * called in the event of an error in the CUDA context.
13723 *
13724 * \param hStream - Stream to enqueue function call in
13725 * \param fn - The function to call once preceding stream operations are complete
13726 * \param userData - User-specified data to be passed to the function
13727 *
13728 * \return
13729 * ::CUDA_SUCCESS,
13730 * ::CUDA_ERROR_DEINITIALIZED,
13731 * ::CUDA_ERROR_NOT_INITIALIZED,
13732 * ::CUDA_ERROR_INVALID_CONTEXT,
13733 * ::CUDA_ERROR_INVALID_HANDLE,
13734 * ::CUDA_ERROR_NOT_SUPPORTED
13735 * \note_null_stream
13736 * \notefnerr
13737 *
13738 * \sa ::cuStreamCreate,
13739 * ::cuStreamQuery,
13740 * ::cuStreamSynchronize,
13741 * ::cuStreamWaitEvent,
13742 * ::cuStreamDestroy,
13743 * ::cuMemAllocManaged,
13744 * ::cuStreamAttachMemAsync,
13745 * ::cuStreamAddCallback
13746 */
13747CUresult CUDAAPI cuLaunchHostFunc(CUstream hStream, CUhostFn fn, void *userData);
13748
13749/** @} */ /* END CUDA_EXEC */
13750
13751/**
13752 * \defgroup CUDA_EXEC_DEPRECATED Execution Control [DEPRECATED]
13753 *
13754 * ___MANBRIEF___ deprecated execution control functions of the low-level CUDA
13755 * driver API (___CURRENT_FILE___) ___ENDMANBRIEF___
13756 *
13757 * This section describes the deprecated execution control functions of the
13758 * low-level CUDA driver application programming interface.
13759 *
13760 * @{
13761 */
13762
13763/**
13764 * \brief Sets the block-dimensions for the function
13765 *
13766 * \deprecated
13767 *
13768 * Specifies the \p x, \p y, and \p z dimensions of the thread blocks that are
13769 * created when the kernel given by \p hfunc is launched.
13770 *
13771 * \param hfunc - Kernel to specify dimensions of
13772 * \param x - X dimension
13773 * \param y - Y dimension
13774 * \param z - Z dimension
13775 *
13776 * \return
13777 * ::CUDA_SUCCESS,
13778 * ::CUDA_ERROR_DEINITIALIZED,
13779 * ::CUDA_ERROR_NOT_INITIALIZED,
13780 * ::CUDA_ERROR_INVALID_CONTEXT,
13781 * ::CUDA_ERROR_INVALID_HANDLE,
13782 * ::CUDA_ERROR_INVALID_VALUE
13783 * \notefnerr
13784 *
13785 * \sa ::cuFuncSetSharedSize,
13786 * ::cuFuncSetCacheConfig,
13787 * ::cuFuncGetAttribute,
13788 * ::cuParamSetSize,
13789 * ::cuParamSeti,
13790 * ::cuParamSetf,
13791 * ::cuParamSetv,
13792 * ::cuLaunch,
13793 * ::cuLaunchGrid,
13794 * ::cuLaunchGridAsync,
13795 * ::cuLaunchKernel
13796 */
13797__CUDA_DEPRECATED CUresult CUDAAPI cuFuncSetBlockShape(CUfunction hfunc, int x, int y, int z);
13798
13799/**
13800 * \brief Sets the dynamic shared-memory size for the function
13801 *
13802 * \deprecated
13803 *
13804 * Sets through \p bytes the amount of dynamic shared memory that will be
13805 * available to each thread block when the kernel given by \p hfunc is launched.
13806 *
13807 * \param hfunc - Kernel to specify dynamic shared-memory size for
13808 * \param bytes - Dynamic shared-memory size per thread in bytes
13809 *
13810 * \return
13811 * ::CUDA_SUCCESS,
13812 * ::CUDA_ERROR_DEINITIALIZED,
13813 * ::CUDA_ERROR_NOT_INITIALIZED,
13814 * ::CUDA_ERROR_INVALID_CONTEXT,
13815 * ::CUDA_ERROR_INVALID_HANDLE,
13816 * ::CUDA_ERROR_INVALID_VALUE
13817 * \notefnerr
13818 *
13819 * \sa ::cuFuncSetBlockShape,
13820 * ::cuFuncSetCacheConfig,
13821 * ::cuFuncGetAttribute,
13822 * ::cuParamSetSize,
13823 * ::cuParamSeti,
13824 * ::cuParamSetf,
13825 * ::cuParamSetv,
13826 * ::cuLaunch,
13827 * ::cuLaunchGrid,
13828 * ::cuLaunchGridAsync,
13829 * ::cuLaunchKernel
13830 */
13831__CUDA_DEPRECATED CUresult CUDAAPI cuFuncSetSharedSize(CUfunction hfunc, unsigned int bytes);
13832
13833/**
13834 * \brief Sets the parameter size for the function
13835 *
13836 * \deprecated
13837 *
13838 * Sets through \p numbytes the total size in bytes needed by the function
13839 * parameters of the kernel corresponding to \p hfunc.
13840 *
13841 * \param hfunc - Kernel to set parameter size for
13842 * \param numbytes - Size of parameter list in bytes
13843 *
13844 * \return
13845 * ::CUDA_SUCCESS,
13846 * ::CUDA_ERROR_DEINITIALIZED,
13847 * ::CUDA_ERROR_NOT_INITIALIZED,
13848 * ::CUDA_ERROR_INVALID_CONTEXT,
13849 * ::CUDA_ERROR_INVALID_VALUE
13850 * \notefnerr
13851 *
13852 * \sa ::cuFuncSetBlockShape,
13853 * ::cuFuncSetSharedSize,
13854 * ::cuFuncGetAttribute,
13855 * ::cuParamSetf,
13856 * ::cuParamSeti,
13857 * ::cuParamSetv,
13858 * ::cuLaunch,
13859 * ::cuLaunchGrid,
13860 * ::cuLaunchGridAsync,
13861 * ::cuLaunchKernel
13862 */
13863__CUDA_DEPRECATED CUresult CUDAAPI cuParamSetSize(CUfunction hfunc, unsigned int numbytes);
13864
13865/**
13866 * \brief Adds an integer parameter to the function's argument list
13867 *
13868 * \deprecated
13869 *
13870 * Sets an integer parameter that will be specified the next time the
13871 * kernel corresponding to \p hfunc will be invoked. \p offset is a byte offset.
13872 *
13873 * \param hfunc - Kernel to add parameter to
13874 * \param offset - Offset to add parameter to argument list
13875 * \param value - Value of parameter
13876 *
13877 * \return
13878 * ::CUDA_SUCCESS,
13879 * ::CUDA_ERROR_DEINITIALIZED,
13880 * ::CUDA_ERROR_NOT_INITIALIZED,
13881 * ::CUDA_ERROR_INVALID_CONTEXT,
13882 * ::CUDA_ERROR_INVALID_VALUE
13883 * \notefnerr
13884 *
13885 * \sa ::cuFuncSetBlockShape,
13886 * ::cuFuncSetSharedSize,
13887 * ::cuFuncGetAttribute,
13888 * ::cuParamSetSize,
13889 * ::cuParamSetf,
13890 * ::cuParamSetv,
13891 * ::cuLaunch,
13892 * ::cuLaunchGrid,
13893 * ::cuLaunchGridAsync,
13894 * ::cuLaunchKernel
13895 */
13896__CUDA_DEPRECATED CUresult CUDAAPI cuParamSeti(CUfunction hfunc, int offset, unsigned int value);
13897
13898/**
13899 * \brief Adds a floating-point parameter to the function's argument list
13900 *
13901 * \deprecated
13902 *
13903 * Sets a floating-point parameter that will be specified the next time the
13904 * kernel corresponding to \p hfunc will be invoked. \p offset is a byte offset.
13905 *
13906 * \param hfunc - Kernel to add parameter to
13907 * \param offset - Offset to add parameter to argument list
13908 * \param value - Value of parameter
13909 *
13910 * \return
13911 * ::CUDA_SUCCESS,
13912 * ::CUDA_ERROR_DEINITIALIZED,
13913 * ::CUDA_ERROR_NOT_INITIALIZED,
13914 * ::CUDA_ERROR_INVALID_CONTEXT,
13915 * ::CUDA_ERROR_INVALID_VALUE
13916 * \notefnerr
13917 *
13918 * \sa ::cuFuncSetBlockShape,
13919 * ::cuFuncSetSharedSize,
13920 * ::cuFuncGetAttribute,
13921 * ::cuParamSetSize,
13922 * ::cuParamSeti,
13923 * ::cuParamSetv,
13924 * ::cuLaunch,
13925 * ::cuLaunchGrid,
13926 * ::cuLaunchGridAsync,
13927 * ::cuLaunchKernel
13928 */
13929__CUDA_DEPRECATED CUresult CUDAAPI cuParamSetf(CUfunction hfunc, int offset, float value);
13930
13931/**
13932 * \brief Adds arbitrary data to the function's argument list
13933 *
13934 * \deprecated
13935 *
13936 * Copies an arbitrary amount of data (specified in \p numbytes) from \p ptr
13937 * into the parameter space of the kernel corresponding to \p hfunc. \p offset
13938 * is a byte offset.
13939 *
13940 * \param hfunc - Kernel to add data to
13941 * \param offset - Offset to add data to argument list
13942 * \param ptr - Pointer to arbitrary data
13943 * \param numbytes - Size of data to copy in bytes
13944 *
13945 * \return
13946 * ::CUDA_SUCCESS,
13947 * ::CUDA_ERROR_DEINITIALIZED,
13948 * ::CUDA_ERROR_NOT_INITIALIZED,
13949 * ::CUDA_ERROR_INVALID_CONTEXT,
13950 * ::CUDA_ERROR_INVALID_VALUE
13951 * \notefnerr
13952 *
13953 * \sa ::cuFuncSetBlockShape,
13954 * ::cuFuncSetSharedSize,
13955 * ::cuFuncGetAttribute,
13956 * ::cuParamSetSize,
13957 * ::cuParamSetf,
13958 * ::cuParamSeti,
13959 * ::cuLaunch,
13960 * ::cuLaunchGrid,
13961 * ::cuLaunchGridAsync,
13962 * ::cuLaunchKernel
13963 */
13964__CUDA_DEPRECATED CUresult CUDAAPI cuParamSetv(CUfunction hfunc, int offset, void *ptr, unsigned int numbytes);
13965
13966/**
13967 * \brief Launches a CUDA function
13968 *
13969 * \deprecated
13970 *
13971 * Invokes the kernel \p f on a 1 x 1 x 1 grid of blocks. The block
13972 * contains the number of threads specified by a previous call to
13973 * ::cuFuncSetBlockShape().
13974 *
13975 * The block shape, dynamic shared memory size, and parameter information
13976 * must be set using
13977 * ::cuFuncSetBlockShape(),
13978 * ::cuFuncSetSharedSize(),
13979 * ::cuParamSetSize(),
13980 * ::cuParamSeti(),
13981 * ::cuParamSetf(), and
13982 * ::cuParamSetv()
13983 * prior to calling this function.
13984 *
13985 * Launching a function via ::cuLaunchKernel() invalidates the function's
13986 * block shape, dynamic shared memory size, and parameter information. After
13987 * launching via cuLaunchKernel, this state must be re-initialized prior to
13988 * calling this function. Failure to do so results in undefined behavior.
13989 *
13990 * \param f - Kernel to launch
13991 *
13992 * \return
13993 * ::CUDA_SUCCESS,
13994 * ::CUDA_ERROR_DEINITIALIZED,
13995 * ::CUDA_ERROR_NOT_INITIALIZED,
13996 * ::CUDA_ERROR_INVALID_CONTEXT,
13997 * ::CUDA_ERROR_INVALID_VALUE,
13998 * ::CUDA_ERROR_LAUNCH_FAILED,
13999 * ::CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES,
14000 * ::CUDA_ERROR_LAUNCH_TIMEOUT,
14001 * ::CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING,
14002 * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED
14003 * \notefnerr
14004 *
14005 * \sa ::cuFuncSetBlockShape,
14006 * ::cuFuncSetSharedSize,
14007 * ::cuFuncGetAttribute,
14008 * ::cuParamSetSize,
14009 * ::cuParamSetf,
14010 * ::cuParamSeti,
14011 * ::cuParamSetv,
14012 * ::cuLaunchGrid,
14013 * ::cuLaunchGridAsync,
14014 * ::cuLaunchKernel
14015 */
14016__CUDA_DEPRECATED CUresult CUDAAPI cuLaunch(CUfunction f);
14017
14018/**
14019 * \brief Launches a CUDA function
14020 *
14021 * \deprecated
14022 *
14023 * Invokes the kernel \p f on a \p grid_width x \p grid_height grid of
14024 * blocks. Each block contains the number of threads specified by a previous
14025 * call to ::cuFuncSetBlockShape().
14026 *
14027 * The block shape, dynamic shared memory size, and parameter information
14028 * must be set using
14029 * ::cuFuncSetBlockShape(),
14030 * ::cuFuncSetSharedSize(),
14031 * ::cuParamSetSize(),
14032 * ::cuParamSeti(),
14033 * ::cuParamSetf(), and
14034 * ::cuParamSetv()
14035 * prior to calling this function.
14036 *
14037 * Launching a function via ::cuLaunchKernel() invalidates the function's
14038 * block shape, dynamic shared memory size, and parameter information. After
14039 * launching via cuLaunchKernel, this state must be re-initialized prior to
14040 * calling this function. Failure to do so results in undefined behavior.
14041 *
14042 * \param f - Kernel to launch
14043 * \param grid_width - Width of grid in blocks
14044 * \param grid_height - Height of grid in blocks
14045 *
14046 * \return
14047 * ::CUDA_SUCCESS,
14048 * ::CUDA_ERROR_DEINITIALIZED,
14049 * ::CUDA_ERROR_NOT_INITIALIZED,
14050 * ::CUDA_ERROR_INVALID_CONTEXT,
14051 * ::CUDA_ERROR_INVALID_VALUE,
14052 * ::CUDA_ERROR_LAUNCH_FAILED,
14053 * ::CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES,
14054 * ::CUDA_ERROR_LAUNCH_TIMEOUT,
14055 * ::CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING,
14056 * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED
14057 * \notefnerr
14058 *
14059 * \sa ::cuFuncSetBlockShape,
14060 * ::cuFuncSetSharedSize,
14061 * ::cuFuncGetAttribute,
14062 * ::cuParamSetSize,
14063 * ::cuParamSetf,
14064 * ::cuParamSeti,
14065 * ::cuParamSetv,
14066 * ::cuLaunch,
14067 * ::cuLaunchGridAsync,
14068 * ::cuLaunchKernel
14069 */
14070__CUDA_DEPRECATED CUresult CUDAAPI cuLaunchGrid(CUfunction f, int grid_width, int grid_height);
14071
14072/**
14073 * \brief Launches a CUDA function
14074 *
14075 * \deprecated
14076 *
14077 * Invokes the kernel \p f on a \p grid_width x \p grid_height grid of
14078 * blocks. Each block contains the number of threads specified by a previous
14079 * call to ::cuFuncSetBlockShape().
14080 *
14081 * The block shape, dynamic shared memory size, and parameter information
14082 * must be set using
14083 * ::cuFuncSetBlockShape(),
14084 * ::cuFuncSetSharedSize(),
14085 * ::cuParamSetSize(),
14086 * ::cuParamSeti(),
14087 * ::cuParamSetf(), and
14088 * ::cuParamSetv()
14089 * prior to calling this function.
14090 *
14091 * Launching a function via ::cuLaunchKernel() invalidates the function's
14092 * block shape, dynamic shared memory size, and parameter information. After
14093 * launching via cuLaunchKernel, this state must be re-initialized prior to
14094 * calling this function. Failure to do so results in undefined behavior.
14095 *
14096 * \param f - Kernel to launch
14097 * \param grid_width - Width of grid in blocks
14098 * \param grid_height - Height of grid in blocks
14099 * \param hStream - Stream identifier
14100 *
14101 * \return
14102 * ::CUDA_SUCCESS,
14103 * ::CUDA_ERROR_DEINITIALIZED,
14104 * ::CUDA_ERROR_NOT_INITIALIZED,
14105 * ::CUDA_ERROR_INVALID_CONTEXT,
14106 * ::CUDA_ERROR_INVALID_HANDLE,
14107 * ::CUDA_ERROR_INVALID_VALUE,
14108 * ::CUDA_ERROR_LAUNCH_FAILED,
14109 * ::CUDA_ERROR_LAUNCH_OUT_OF_RESOURCES,
14110 * ::CUDA_ERROR_LAUNCH_TIMEOUT,
14111 * ::CUDA_ERROR_LAUNCH_INCOMPATIBLE_TEXTURING,
14112 * ::CUDA_ERROR_SHARED_OBJECT_INIT_FAILED
14113 *
14114 * \note In certain cases where cubins are created with no ABI (i.e., using \p ptxas \p --abi-compile \p no),
14115 * this function may serialize kernel launches. The CUDA driver retains asynchronous behavior by
14116 * growing the per-thread stack as needed per launch and not shrinking it afterwards.
14117 *
14118 * \note_null_stream
14119 * \notefnerr
14120 *
14121 * \sa ::cuFuncSetBlockShape,
14122 * ::cuFuncSetSharedSize,
14123 * ::cuFuncGetAttribute,
14124 * ::cuParamSetSize,
14125 * ::cuParamSetf,
14126 * ::cuParamSeti,
14127 * ::cuParamSetv,
14128 * ::cuLaunch,
14129 * ::cuLaunchGrid,
14130 * ::cuLaunchKernel
14131 */
14132__CUDA_DEPRECATED CUresult CUDAAPI cuLaunchGridAsync(CUfunction f, int grid_width, int grid_height, CUstream hStream);
14133
14134
14135/**
14136 * \brief Adds a texture-reference to the function's argument list
14137 *
14138 * \deprecated
14139 *
14140 * Makes the CUDA array or linear memory bound to the texture reference
14141 * \p hTexRef available to a device program as a texture. In this version of
14142 * CUDA, the texture-reference must be obtained via ::cuModuleGetTexRef() and
14143 * the \p texunit parameter must be set to ::CU_PARAM_TR_DEFAULT.
14144 *
14145 * \param hfunc - Kernel to add texture-reference to
14146 * \param texunit - Texture unit (must be ::CU_PARAM_TR_DEFAULT)
14147 * \param hTexRef - Texture-reference to add to argument list
14148 *
14149 * \return
14150 * ::CUDA_SUCCESS,
14151 * ::CUDA_ERROR_DEINITIALIZED,
14152 * ::CUDA_ERROR_NOT_INITIALIZED,
14153 * ::CUDA_ERROR_INVALID_CONTEXT,
14154 * ::CUDA_ERROR_INVALID_VALUE
14155 * \notefnerr
14156 */
14157__CUDA_DEPRECATED CUresult CUDAAPI cuParamSetTexRef(CUfunction hfunc, int texunit, CUtexref hTexRef);
14158/** @} */ /* END CUDA_EXEC_DEPRECATED */
14159
14160/**
14161 * \defgroup CUDA_GRAPH Graph Management
14162 *
14163 * ___MANBRIEF___ graph management functions of the low-level CUDA driver API
14164 * (___CURRENT_FILE___) ___ENDMANBRIEF___
14165 *
14166 * This section describes the graph management functions of the low-level CUDA
14167 * driver application programming interface.
14168 *
14169 * @{
14170 */
14171
14172/**
14173 * \brief Creates a graph
14174 *
14175 * Creates an empty graph, which is returned via \p phGraph.
14176 *
14177 * \param phGraph - Returns newly created graph
14178 * \param flags - Graph creation flags, must be 0
14179 *
14180 * \return
14181 * ::CUDA_SUCCESS,
14182 * ::CUDA_ERROR_DEINITIALIZED,
14183 * ::CUDA_ERROR_NOT_INITIALIZED,
14184 * ::CUDA_ERROR_INVALID_VALUE,
14185 * ::CUDA_ERROR_OUT_OF_MEMORY
14186 * \note_graph_thread_safety
14187 * \notefnerr
14188 *
14189 * \sa
14190 * ::cuGraphAddChildGraphNode,
14191 * ::cuGraphAddEmptyNode,
14192 * ::cuGraphAddKernelNode,
14193 * ::cuGraphAddHostNode,
14194 * ::cuGraphAddMemcpyNode,
14195 * ::cuGraphAddMemsetNode,
14196 * ::cuGraphInstantiate,
14197 * ::cuGraphDestroy,
14198 * ::cuGraphGetNodes,
14199 * ::cuGraphGetRootNodes,
14200 * ::cuGraphGetEdges,
14201 * ::cuGraphClone
14202 */
14203CUresult CUDAAPI cuGraphCreate(CUgraph *phGraph, unsigned int flags);
14204
14205/**
14206 * \brief Creates a kernel execution node and adds it to a graph
14207 *
14208 * Creates a new kernel execution node and adds it to \p hGraph with \p numDependencies
14209 * dependencies specified via \p dependencies and arguments specified in \p nodeParams.
14210 * It is possible for \p numDependencies to be 0, in which case the node will be placed
14211 * at the root of the graph. \p dependencies may not have any duplicate entries.
14212 * A handle to the new node will be returned in \p phGraphNode.
14213 *
14214 * The CUDA_KERNEL_NODE_PARAMS structure is defined as:
14215 *
14216 * \code
14217 * typedef struct CUDA_KERNEL_NODE_PARAMS_st {
14218 * CUfunction func;
14219 * unsigned int gridDimX;
14220 * unsigned int gridDimY;
14221 * unsigned int gridDimZ;
14222 * unsigned int blockDimX;
14223 * unsigned int blockDimY;
14224 * unsigned int blockDimZ;
14225 * unsigned int sharedMemBytes;
14226 * void **kernelParams;
14227 * void **extra;
14228 * } CUDA_KERNEL_NODE_PARAMS;
14229 * \endcode
14230 *
14231 * When the graph is launched, the node will invoke kernel \p func on a (\p gridDimX x
14232 * \p gridDimY x \p gridDimZ) grid of blocks. Each block contains
14233 * (\p blockDimX x \p blockDimY x \p blockDimZ) threads.
14234 *
14235 * \p sharedMemBytes sets the amount of dynamic shared memory that will be
14236 * available to each thread block.
14237 *
14238 * Kernel parameters to \p func can be specified in one of two ways:
14239 *
14240 * 1) Kernel parameters can be specified via \p kernelParams. If the kernel has N
14241 * parameters, then \p kernelParams needs to be an array of N pointers. Each pointer,
14242 * from \p kernelParams[0] to \p kernelParams[N-1], points to the region of memory from which the actual
14243 * parameter will be copied. The number of kernel parameters and their offsets and sizes do not need
14244 * to be specified as that information is retrieved directly from the kernel's image.
14245 *
14246 * 2) Kernel parameters for non-cooperative kernels can also be packaged by the application into a single
14247 * buffer that is passed in via \p extra. This places the burden on the application of knowing each
14248 * kernel parameter's size and alignment/padding within the buffer. The \p extra parameter exists
14249 * to allow this function to take additional less commonly used arguments. \p extra specifies
14250 * a list of names of extra settings and their corresponding values. Each extra setting name is
14251 * immediately followed by the corresponding value. The list must be terminated with either NULL or
14252 * CU_LAUNCH_PARAM_END.
14253 *
14254 * - ::CU_LAUNCH_PARAM_END, which indicates the end of the \p extra
14255 * array;
14256 * - ::CU_LAUNCH_PARAM_BUFFER_POINTER, which specifies that the next
14257 * value in \p extra will be a pointer to a buffer
14258 * containing all the kernel parameters for launching kernel
14259 * \p func;
14260 * - ::CU_LAUNCH_PARAM_BUFFER_SIZE, which specifies that the next
14261 * value in \p extra will be a pointer to a size_t
14262 * containing the size of the buffer specified with
14263 * ::CU_LAUNCH_PARAM_BUFFER_POINTER;
14264 *
14265 * The error ::CUDA_ERROR_INVALID_VALUE will be returned if kernel parameters are specified with both
14266 * \p kernelParams and \p extra (i.e. both \p kernelParams and \p extra are non-NULL).
14267 * ::CUDA_ERROR_INVALID_VALUE will be returned if \p extra is used for a cooperative kernel.
14268 *
14269 * The \p kernelParams or \p extra array, as well as the argument values it points to,
14270 * are copied during this call.
14271 *
14272 * \note Kernels launched using graphs must not use texture and surface references. Reading or
14273 * writing through any texture or surface reference is undefined behavior.
14274 * This restriction does not apply to texture and surface objects.
14275 *
14276 * \param phGraphNode - Returns newly created node
14277 * \param hGraph - Graph to which to add the node
14278 * \param dependencies - Dependencies of the node
14279 * \param numDependencies - Number of dependencies
14280 * \param nodeParams - Parameters for the GPU execution node
14281 *
14282 * \return
14283 * ::CUDA_SUCCESS,
14284 * ::CUDA_ERROR_DEINITIALIZED,
14285 * ::CUDA_ERROR_NOT_INITIALIZED,
14286 * ::CUDA_ERROR_INVALID_VALUE
14287 * \note_graph_thread_safety
14288 * \notefnerr
14289 *
14290 * \sa
14291 * ::cuLaunchKernel,
14292 * ::cuLaunchCooperativeKernel,
14293 * ::cuGraphKernelNodeGetParams,
14294 * ::cuGraphKernelNodeSetParams,
14295 * ::cuGraphCreate,
14296 * ::cuGraphDestroyNode,
14297 * ::cuGraphAddChildGraphNode,
14298 * ::cuGraphAddEmptyNode,
14299 * ::cuGraphAddHostNode,
14300 * ::cuGraphAddMemcpyNode,
14301 * ::cuGraphAddMemsetNode
14302 */
14303CUresult CUDAAPI cuGraphAddKernelNode(CUgraphNode *phGraphNode, CUgraph hGraph, const CUgraphNode *dependencies, size_t numDependencies, const CUDA_KERNEL_NODE_PARAMS *nodeParams);
14304
14305/**
14306 * \brief Returns a kernel node's parameters
14307 *
14308 * Returns the parameters of kernel node \p hNode in \p nodeParams.
14309 * The \p kernelParams or \p extra array returned in \p nodeParams,
14310 * as well as the argument values it points to, are owned by the node.
14311 * This memory remains valid until the node is destroyed or its
14312 * parameters are modified, and should not be modified
14313 * directly. Use ::cuGraphKernelNodeSetParams to update the
14314 * parameters of this node.
14315 *
14316 * The params will contain either \p kernelParams or \p extra,
14317 * according to which of these was most recently set on the node.
14318 *
14319 * \param hNode - Node to get the parameters for
14320 * \param nodeParams - Pointer to return the parameters
14321 *
14322 * \return
14323 * ::CUDA_SUCCESS,
14324 * ::CUDA_ERROR_DEINITIALIZED,
14325 * ::CUDA_ERROR_NOT_INITIALIZED,
14326 * ::CUDA_ERROR_INVALID_VALUE
14327 * \note_graph_thread_safety
14328 * \notefnerr
14329 *
14330 * \sa
14331 * ::cuLaunchKernel,
14332 * ::cuGraphAddKernelNode,
14333 * ::cuGraphKernelNodeSetParams
14334 */
14335CUresult CUDAAPI cuGraphKernelNodeGetParams(CUgraphNode hNode, CUDA_KERNEL_NODE_PARAMS *nodeParams);
14336
14337/**
14338 * \brief Sets a kernel node's parameters
14339 *
14340 * Sets the parameters of kernel node \p hNode to \p nodeParams.
14341 *
14342 * \param hNode - Node to set the parameters for
14343 * \param nodeParams - Parameters to copy
14344 *
14345 * \return
14346 * ::CUDA_SUCCESS,
14347 * ::CUDA_ERROR_INVALID_VALUE,
14348 * ::CUDA_ERROR_INVALID_HANDLE,
14349 * ::CUDA_ERROR_OUT_OF_MEMORY
14350 * \note_graph_thread_safety
14351 * \notefnerr
14352 *
14353 * \sa
14354 * ::cuLaunchKernel,
14355 * ::cuGraphAddKernelNode,
14356 * ::cuGraphKernelNodeGetParams
14357 */
14358CUresult CUDAAPI cuGraphKernelNodeSetParams(CUgraphNode hNode, const CUDA_KERNEL_NODE_PARAMS *nodeParams);
14359
14360/**
14361 * \brief Creates a memcpy node and adds it to a graph
14362 *
14363 * Creates a new memcpy node and adds it to \p hGraph with \p numDependencies
14364 * dependencies specified via \p dependencies.
14365 * It is possible for \p numDependencies to be 0, in which case the node will be placed
14366 * at the root of the graph. \p dependencies may not have any duplicate entries.
14367 * A handle to the new node will be returned in \p phGraphNode.
14368 *
14369 * When the graph is launched, the node will perform the memcpy described by \p copyParams.
14370 * See ::cuMemcpy3D() for a description of the structure and its restrictions.
14371 *
14372 * Memcpy nodes have some additional restrictions with regards to managed memory, if the
14373 * system contains at least one device which has a zero value for the device attribute
14374 * ::CU_DEVICE_ATTRIBUTE_CONCURRENT_MANAGED_ACCESS. If one or more of the operands refer
14375 * to managed memory, then using the memory type ::CU_MEMORYTYPE_UNIFIED is disallowed
14376 * for those operand(s). The managed memory will be treated as residing on either the
14377 * host or the device, depending on which memory type is specified.
14378 *
14379 * \param phGraphNode - Returns newly created node
14380 * \param hGraph - Graph to which to add the node
14381 * \param dependencies - Dependencies of the node
14382 * \param numDependencies - Number of dependencies
14383 * \param copyParams - Parameters for the memory copy
14384 * \param ctx - Context on which to run the node
14385 *
14386 * \return
14387 * ::CUDA_SUCCESS,
14388 * ::CUDA_ERROR_DEINITIALIZED,
14389 * ::CUDA_ERROR_NOT_INITIALIZED,
14390 * ::CUDA_ERROR_INVALID_VALUE
14391 * \note_graph_thread_safety
14392 * \notefnerr
14393 *
14394 * \sa
14395 * ::cuMemcpy3D,
14396 * ::cuGraphMemcpyNodeGetParams,
14397 * ::cuGraphMemcpyNodeSetParams,
14398 * ::cuGraphCreate,
14399 * ::cuGraphDestroyNode,
14400 * ::cuGraphAddChildGraphNode,
14401 * ::cuGraphAddEmptyNode,
14402 * ::cuGraphAddKernelNode,
14403 * ::cuGraphAddHostNode,
14404 * ::cuGraphAddMemsetNode
14405 */
14406CUresult CUDAAPI cuGraphAddMemcpyNode(CUgraphNode *phGraphNode, CUgraph hGraph, const CUgraphNode *dependencies, size_t numDependencies, const CUDA_MEMCPY3D *copyParams, CUcontext ctx);
14407
14408/**
14409 * \brief Returns a memcpy node's parameters
14410 *
14411 * Returns the parameters of memcpy node \p hNode in \p nodeParams.
14412 *
14413 * \param hNode - Node to get the parameters for
14414 * \param nodeParams - Pointer to return the parameters
14415 *
14416 * \return
14417 * ::CUDA_SUCCESS,
14418 * ::CUDA_ERROR_DEINITIALIZED,
14419 * ::CUDA_ERROR_NOT_INITIALIZED,
14420 * ::CUDA_ERROR_INVALID_VALUE
14421 * \note_graph_thread_safety
14422 * \notefnerr
14423 *
14424 * \sa
14425 * ::cuMemcpy3D,
14426 * ::cuGraphAddMemcpyNode,
14427 * ::cuGraphMemcpyNodeSetParams
14428 */
14429CUresult CUDAAPI cuGraphMemcpyNodeGetParams(CUgraphNode hNode, CUDA_MEMCPY3D *nodeParams);
14430
14431/**
14432 * \brief Sets a memcpy node's parameters
14433 *
14434 * Sets the parameters of memcpy node \p hNode to \p nodeParams.
14435 *
14436 * \param hNode - Node to set the parameters for
14437 * \param nodeParams - Parameters to copy
14438 *
14439 * \return
14440 * ::CUDA_SUCCESS,
14441 * ::CUDA_ERROR_DEINITIALIZED,
14442 * ::CUDA_ERROR_NOT_INITIALIZED,
14443 * ::CUDA_ERROR_INVALID_VALUE,
14444 * \note_graph_thread_safety
14445 * \notefnerr
14446 *
14447 * \sa
14448 * ::cuMemcpy3D,
14449 * ::cuGraphAddMemcpyNode,
14450 * ::cuGraphMemcpyNodeGetParams
14451 */
14452CUresult CUDAAPI cuGraphMemcpyNodeSetParams(CUgraphNode hNode, const CUDA_MEMCPY3D *nodeParams);
14453
14454/**
14455 * \brief Creates a memset node and adds it to a graph
14456 *
14457 * Creates a new memset node and adds it to \p hGraph with \p numDependencies
14458 * dependencies specified via \p dependencies.
14459 * It is possible for \p numDependencies to be 0, in which case the node will be placed
14460 * at the root of the graph. \p dependencies may not have any duplicate entries.
14461 * A handle to the new node will be returned in \p phGraphNode.
14462 *
14463 * The element size must be 1, 2, or 4 bytes.
14464 * When the graph is launched, the node will perform the memset described by \p memsetParams.
14465 *
14466 * \param phGraphNode - Returns newly created node
14467 * \param hGraph - Graph to which to add the node
14468 * \param dependencies - Dependencies of the node
14469 * \param numDependencies - Number of dependencies
14470 * \param memsetParams - Parameters for the memory set
14471 * \param ctx - Context on which to run the node
14472 *
14473 * \return
14474 * ::CUDA_SUCCESS,
14475 * ::CUDA_ERROR_DEINITIALIZED,
14476 * ::CUDA_ERROR_NOT_INITIALIZED,
14477 * ::CUDA_ERROR_INVALID_VALUE,
14478 * ::CUDA_ERROR_INVALID_CONTEXT
14479 * \note_graph_thread_safety
14480 * \notefnerr
14481 *
14482 * \sa
14483 * ::cuMemsetD2D32,
14484 * ::cuGraphMemsetNodeGetParams,
14485 * ::cuGraphMemsetNodeSetParams,
14486 * ::cuGraphCreate,
14487 * ::cuGraphDestroyNode,
14488 * ::cuGraphAddChildGraphNode,
14489 * ::cuGraphAddEmptyNode,
14490 * ::cuGraphAddKernelNode,
14491 * ::cuGraphAddHostNode,
14492 * ::cuGraphAddMemcpyNode
14493 */
14494CUresult CUDAAPI cuGraphAddMemsetNode(CUgraphNode *phGraphNode, CUgraph hGraph, const CUgraphNode *dependencies, size_t numDependencies, const CUDA_MEMSET_NODE_PARAMS *memsetParams, CUcontext ctx);
14495
14496/**
14497 * \brief Returns a memset node's parameters
14498 *
14499 * Returns the parameters of memset node \p hNode in \p nodeParams.
14500 *
14501 * \param hNode - Node to get the parameters for
14502 * \param nodeParams - Pointer to return the parameters
14503 *
14504 * \return
14505 * ::CUDA_SUCCESS,
14506 * ::CUDA_ERROR_DEINITIALIZED,
14507 * ::CUDA_ERROR_NOT_INITIALIZED,
14508 * ::CUDA_ERROR_INVALID_VALUE
14509 * \note_graph_thread_safety
14510 * \notefnerr
14511 *
14512 * \sa
14513 * ::cuMemsetD2D32,
14514 * ::cuGraphAddMemsetNode,
14515 * ::cuGraphMemsetNodeSetParams
14516 */
14517CUresult CUDAAPI cuGraphMemsetNodeGetParams(CUgraphNode hNode, CUDA_MEMSET_NODE_PARAMS *nodeParams);
14518
14519/**
14520 * \brief Sets a memset node's parameters
14521 *
14522 * Sets the parameters of memset node \p hNode to \p nodeParams.
14523 *
14524 * \param hNode - Node to set the parameters for
14525 * \param nodeParams - Parameters to copy
14526 *
14527 * \return
14528 * ::CUDA_SUCCESS,
14529 * ::CUDA_ERROR_DEINITIALIZED,
14530 * ::CUDA_ERROR_NOT_INITIALIZED,
14531 * ::CUDA_ERROR_INVALID_VALUE
14532 * \note_graph_thread_safety
14533 * \notefnerr
14534 *
14535 * \sa
14536 * ::cuMemsetD2D32,
14537 * ::cuGraphAddMemsetNode,
14538 * ::cuGraphMemsetNodeGetParams
14539 */
14540CUresult CUDAAPI cuGraphMemsetNodeSetParams(CUgraphNode hNode, const CUDA_MEMSET_NODE_PARAMS *nodeParams);
14541
14542/**
14543 * \brief Creates a host execution node and adds it to a graph
14544 *
14545 * Creates a new CPU execution node and adds it to \p hGraph with \p numDependencies
14546 * dependencies specified via \p dependencies and arguments specified in \p nodeParams.
14547 * It is possible for \p numDependencies to be 0, in which case the node will be placed
14548 * at the root of the graph. \p dependencies may not have any duplicate entries.
14549 * A handle to the new node will be returned in \p phGraphNode.
14550 *
14551 * When the graph is launched, the node will invoke the specified CPU function.
14552 * Host nodes are not supported under MPS with pre-Volta GPUs.
14553 *
14554 * \param phGraphNode - Returns newly created node
14555 * \param hGraph - Graph to which to add the node
14556 * \param dependencies - Dependencies of the node
14557 * \param numDependencies - Number of dependencies
14558 * \param nodeParams - Parameters for the host node
14559 *
14560 * \return
14561 * ::CUDA_SUCCESS,
14562 * ::CUDA_ERROR_DEINITIALIZED,
14563 * ::CUDA_ERROR_NOT_INITIALIZED,
14564 * ::CUDA_ERROR_NOT_SUPPORTED,
14565 * ::CUDA_ERROR_INVALID_VALUE
14566 * \note_graph_thread_safety
14567 * \notefnerr
14568 *
14569 * \sa
14570 * ::cuLaunchHostFunc,
14571 * ::cuGraphHostNodeGetParams,
14572 * ::cuGraphHostNodeSetParams,
14573 * ::cuGraphCreate,
14574 * ::cuGraphDestroyNode,
14575 * ::cuGraphAddChildGraphNode,
14576 * ::cuGraphAddEmptyNode,
14577 * ::cuGraphAddKernelNode,
14578 * ::cuGraphAddMemcpyNode,
14579 * ::cuGraphAddMemsetNode
14580 */
14581CUresult CUDAAPI cuGraphAddHostNode(CUgraphNode *phGraphNode, CUgraph hGraph, const CUgraphNode *dependencies, size_t numDependencies, const CUDA_HOST_NODE_PARAMS *nodeParams);
14582
14583/**
14584 * \brief Returns a host node's parameters
14585 *
14586 * Returns the parameters of host node \p hNode in \p nodeParams.
14587 *
14588 * \param hNode - Node to get the parameters for
14589 * \param nodeParams - Pointer to return the parameters
14590 *
14591 * \return
14592 * ::CUDA_SUCCESS,
14593 * ::CUDA_ERROR_DEINITIALIZED,
14594 * ::CUDA_ERROR_NOT_INITIALIZED,
14595 * ::CUDA_ERROR_INVALID_VALUE
14596 * \note_graph_thread_safety
14597 * \notefnerr
14598 *
14599 * \sa
14600 * ::cuLaunchHostFunc,
14601 * ::cuGraphAddHostNode,
14602 * ::cuGraphHostNodeSetParams
14603 */
14604CUresult CUDAAPI cuGraphHostNodeGetParams(CUgraphNode hNode, CUDA_HOST_NODE_PARAMS *nodeParams);
14605
14606/**
14607 * \brief Sets a host node's parameters
14608 *
14609 * Sets the parameters of host node \p hNode to \p nodeParams.
14610 *
14611 * \param hNode - Node to set the parameters for
14612 * \param nodeParams - Parameters to copy
14613 *
14614 * \return
14615 * ::CUDA_SUCCESS,
14616 * ::CUDA_ERROR_DEINITIALIZED,
14617 * ::CUDA_ERROR_NOT_INITIALIZED,
14618 * ::CUDA_ERROR_INVALID_VALUE
14619 * \note_graph_thread_safety
14620 * \notefnerr
14621 *
14622 * \sa
14623 * ::cuLaunchHostFunc,
14624 * ::cuGraphAddHostNode,
14625 * ::cuGraphHostNodeGetParams
14626 */
14627CUresult CUDAAPI cuGraphHostNodeSetParams(CUgraphNode hNode, const CUDA_HOST_NODE_PARAMS *nodeParams);
14628
14629/**
14630 * \brief Creates a child graph node and adds it to a graph
14631 *
14632 * Creates a new node which executes an embedded graph, and adds it to \p hGraph with
14633 * \p numDependencies dependencies specified via \p dependencies.
14634 * It is possible for \p numDependencies to be 0, in which case the node will be placed
14635 * at the root of the graph. \p dependencies may not have any duplicate entries.
14636 * A handle to the new node will be returned in \p phGraphNode.
14637 *
14638 * If \p hGraph contains allocation or free nodes, this call will return an error.
14639 *
14640 * The node executes an embedded child graph. The child graph is cloned in this call.
14641 *
14642 * \param phGraphNode - Returns newly created node
14643 * \param hGraph - Graph to which to add the node
14644 * \param dependencies - Dependencies of the node
14645 * \param numDependencies - Number of dependencies
14646 * \param childGraph - The graph to clone into this node
14647 *
14648 * \return
14649 * ::CUDA_SUCCESS,
14650 * ::CUDA_ERROR_DEINITIALIZED,
14651 * ::CUDA_ERROR_NOT_INITIALIZED,
14652 * ::CUDA_ERROR_INVALID_VALUE,
14653 * \note_graph_thread_safety
14654 * \notefnerr
14655 *
14656 * \sa
14657 * ::cuGraphChildGraphNodeGetGraph,
14658 * ::cuGraphCreate,
14659 * ::cuGraphDestroyNode,
14660 * ::cuGraphAddEmptyNode,
14661 * ::cuGraphAddKernelNode,
14662 * ::cuGraphAddHostNode,
14663 * ::cuGraphAddMemcpyNode,
14664 * ::cuGraphAddMemsetNode,
14665 * ::cuGraphClone
14666 */
14667CUresult CUDAAPI cuGraphAddChildGraphNode(CUgraphNode *phGraphNode, CUgraph hGraph, const CUgraphNode *dependencies, size_t numDependencies, CUgraph childGraph);
14668
14669/**
14670 * \brief Gets a handle to the embedded graph of a child graph node
14671 *
14672 * Gets a handle to the embedded graph in a child graph node. This call
14673 * does not clone the graph. Changes to the graph will be reflected in
14674 * the node, and the node retains ownership of the graph.
14675 *
14676 * Allocation and free nodes cannot be added to the returned graph.
14677 * Attempting to do so will return an error.
14678 *
14679 * \param hNode - Node to get the embedded graph for
14680 * \param phGraph - Location to store a handle to the graph
14681 *
14682 * \return
14683 * ::CUDA_SUCCESS,
14684 * ::CUDA_ERROR_DEINITIALIZED,
14685 * ::CUDA_ERROR_NOT_INITIALIZED,
14686 * ::CUDA_ERROR_INVALID_VALUE,
14687 * \note_graph_thread_safety
14688 * \notefnerr
14689 *
14690 * \sa
14691 * ::cuGraphAddChildGraphNode,
14692 * ::cuGraphNodeFindInClone
14693 */
14694CUresult CUDAAPI cuGraphChildGraphNodeGetGraph(CUgraphNode hNode, CUgraph *phGraph);
14695
14696/**
14697 * \brief Creates an empty node and adds it to a graph
14698 *
14699 * Creates a new node which performs no operation, and adds it to \p hGraph with
14700 * \p numDependencies dependencies specified via \p dependencies.
14701 * It is possible for \p numDependencies to be 0, in which case the node will be placed
14702 * at the root of the graph. \p dependencies may not have any duplicate entries.
14703 * A handle to the new node will be returned in \p phGraphNode.
14704 *
14705 * An empty node performs no operation during execution, but can be used for
14706 * transitive ordering. For example, a phased execution graph with 2 groups of n
14707 * nodes with a barrier between them can be represented using an empty node and
14708 * 2*n dependency edges, rather than no empty node and n^2 dependency edges.
14709 *
14710 * \param phGraphNode - Returns newly created node
14711 * \param hGraph - Graph to which to add the node
14712 * \param dependencies - Dependencies of the node
14713 * \param numDependencies - Number of dependencies
14714 *
14715 * \return
14716 * ::CUDA_SUCCESS,
14717 * ::CUDA_ERROR_DEINITIALIZED,
14718 * ::CUDA_ERROR_NOT_INITIALIZED,
14719 * ::CUDA_ERROR_INVALID_VALUE,
14720 * \note_graph_thread_safety
14721 * \notefnerr
14722 *
14723 * \sa
14724 * ::cuGraphCreate,
14725 * ::cuGraphDestroyNode,
14726 * ::cuGraphAddChildGraphNode,
14727 * ::cuGraphAddKernelNode,
14728 * ::cuGraphAddHostNode,
14729 * ::cuGraphAddMemcpyNode,
14730 * ::cuGraphAddMemsetNode
14731 */
14732CUresult CUDAAPI cuGraphAddEmptyNode(CUgraphNode *phGraphNode, CUgraph hGraph, const CUgraphNode *dependencies, size_t numDependencies);
14733
14734/**
14735 * \brief Creates an event record node and adds it to a graph
14736 *
14737 * Creates a new event record node and adds it to \p hGraph with \p numDependencies
14738 * dependencies specified via \p dependencies and event specified in \p event.
14739 * It is possible for \p numDependencies to be 0, in which case the node will be placed
14740 * at the root of the graph. \p dependencies may not have any duplicate entries.
14741 * A handle to the new node will be returned in \p phGraphNode.
14742 *
14743 * Each launch of the graph will record \p event to capture execution of the
14744 * node's dependencies.
14745 *
14746 * \param phGraphNode - Returns newly created node
14747 * \param hGraph - Graph to which to add the node
14748 * \param dependencies - Dependencies of the node
14749 * \param numDependencies - Number of dependencies
14750 * \param event - Event for the node
14751 *
14752 * \return
14753 * ::CUDA_SUCCESS,
14754 * ::CUDA_ERROR_DEINITIALIZED,
14755 * ::CUDA_ERROR_NOT_INITIALIZED,
14756 * ::CUDA_ERROR_NOT_SUPPORTED,
14757 * ::CUDA_ERROR_INVALID_VALUE
14758 * \note_graph_thread_safety
14759 * \notefnerr
14760 *
14761 * \sa
14762 * ::cuGraphAddEventWaitNode,
14763 * ::cuEventRecordWithFlags,
14764 * ::cuStreamWaitEvent,
14765 * ::cuGraphCreate,
14766 * ::cuGraphDestroyNode,
14767 * ::cuGraphAddChildGraphNode,
14768 * ::cuGraphAddEmptyNode,
14769 * ::cuGraphAddKernelNode,
14770 * ::cuGraphAddMemcpyNode,
14771 * ::cuGraphAddMemsetNode,
14772 */
14773CUresult CUDAAPI cuGraphAddEventRecordNode(CUgraphNode *phGraphNode, CUgraph hGraph, const CUgraphNode *dependencies, size_t numDependencies, CUevent event);
14774
14775/**
14776 * \brief Returns the event associated with an event record node
14777 *
14778 * Returns the event of event record node \p hNode in \p event_out.
14779 *
14780 * \param hNode - Node to get the event for
14781 * \param event_out - Pointer to return the event
14782 *
14783 * \return
14784 * ::CUDA_SUCCESS,
14785 * ::CUDA_ERROR_DEINITIALIZED,
14786 * ::CUDA_ERROR_NOT_INITIALIZED,
14787 * ::CUDA_ERROR_INVALID_VALUE
14788 * \note_graph_thread_safety
14789 * \notefnerr
14790 *
14791 * \sa
14792 * ::cuGraphAddEventRecordNode,
14793 * ::cuGraphEventRecordNodeSetEvent,
14794 * ::cuGraphEventWaitNodeGetEvent,
14795 * ::cuEventRecordWithFlags,
14796 * ::cuStreamWaitEvent
14797 */
14798CUresult CUDAAPI cuGraphEventRecordNodeGetEvent(CUgraphNode hNode, CUevent *event_out);
14799
14800/**
14801 * \brief Sets an event record node's event
14802 *
14803 * Sets the event of event record node \p hNode to \p event.
14804 *
14805 * \param hNode - Node to set the event for
14806 * \param event - Event to use
14807 *
14808 * \return
14809 * ::CUDA_SUCCESS,
14810 * ::CUDA_ERROR_INVALID_VALUE,
14811 * ::CUDA_ERROR_INVALID_HANDLE,
14812 * ::CUDA_ERROR_OUT_OF_MEMORY
14813 * \note_graph_thread_safety
14814 * \notefnerr
14815 *
14816 * \sa
14817 * ::cuGraphAddEventRecordNode,
14818 * ::cuGraphEventRecordNodeGetEvent,
14819 * ::cuGraphEventWaitNodeSetEvent,
14820 * ::cuEventRecordWithFlags,
14821 * ::cuStreamWaitEvent
14822 */
14823CUresult CUDAAPI cuGraphEventRecordNodeSetEvent(CUgraphNode hNode, CUevent event);
14824
14825/**
14826 * \brief Creates an event wait node and adds it to a graph
14827 *
14828 * Creates a new event wait node and adds it to \p hGraph with \p numDependencies
14829 * dependencies specified via \p dependencies and event specified in \p event.
14830 * It is possible for \p numDependencies to be 0, in which case the node will be placed
14831 * at the root of the graph. \p dependencies may not have any duplicate entries.
14832 * A handle to the new node will be returned in \p phGraphNode.
14833 *
14834 * The graph node will wait for all work captured in \p event. See ::cuEventRecord()
14835 * for details on what is captured by an event. \p event may be from a different context
14836 * or device than the launch stream.
14837 *
14838 * \param phGraphNode - Returns newly created node
14839 * \param hGraph - Graph to which to add the node
14840 * \param dependencies - Dependencies of the node
14841 * \param numDependencies - Number of dependencies
14842 * \param event - Event for the node
14843 *
14844 * \return
14845 * ::CUDA_SUCCESS,
14846 * ::CUDA_ERROR_DEINITIALIZED,
14847 * ::CUDA_ERROR_NOT_INITIALIZED,
14848 * ::CUDA_ERROR_NOT_SUPPORTED,
14849 * ::CUDA_ERROR_INVALID_VALUE
14850 * \note_graph_thread_safety
14851 * \notefnerr
14852 *
14853 * \sa
14854 * ::cuGraphAddEventRecordNode,
14855 * ::cuEventRecordWithFlags,
14856 * ::cuStreamWaitEvent,
14857 * ::cuGraphCreate,
14858 * ::cuGraphDestroyNode,
14859 * ::cuGraphAddChildGraphNode,
14860 * ::cuGraphAddEmptyNode,
14861 * ::cuGraphAddKernelNode,
14862 * ::cuGraphAddMemcpyNode,
14863 * ::cuGraphAddMemsetNode,
14864 */
14865CUresult CUDAAPI cuGraphAddEventWaitNode(CUgraphNode *phGraphNode, CUgraph hGraph, const CUgraphNode *dependencies, size_t numDependencies, CUevent event);
14866
14867/**
14868 * \brief Returns the event associated with an event wait node
14869 *
14870 * Returns the event of event wait node \p hNode in \p event_out.
14871 *
14872 * \param hNode - Node to get the event for
14873 * \param event_out - Pointer to return the event
14874 *
14875 * \return
14876 * ::CUDA_SUCCESS,
14877 * ::CUDA_ERROR_DEINITIALIZED,
14878 * ::CUDA_ERROR_NOT_INITIALIZED,
14879 * ::CUDA_ERROR_INVALID_VALUE
14880 * \note_graph_thread_safety
14881 * \notefnerr
14882 *
14883 * \sa
14884 * ::cuGraphAddEventWaitNode,
14885 * ::cuGraphEventWaitNodeSetEvent,
14886 * ::cuGraphEventRecordNodeGetEvent,
14887 * ::cuEventRecordWithFlags,
14888 * ::cuStreamWaitEvent
14889 */
14890CUresult CUDAAPI cuGraphEventWaitNodeGetEvent(CUgraphNode hNode, CUevent *event_out);
14891
14892/**
14893 * \brief Sets an event wait node's event
14894 *
14895 * Sets the event of event wait node \p hNode to \p event.
14896 *
14897 * \param hNode - Node to set the event for
14898 * \param event - Event to use
14899 *
14900 * \return
14901 * ::CUDA_SUCCESS,
14902 * ::CUDA_ERROR_INVALID_VALUE,
14903 * ::CUDA_ERROR_INVALID_HANDLE,
14904 * ::CUDA_ERROR_OUT_OF_MEMORY
14905 * \note_graph_thread_safety
14906 * \notefnerr
14907 *
14908 * \sa
14909 * ::cuGraphAddEventWaitNode,
14910 * ::cuGraphEventWaitNodeGetEvent,
14911 * ::cuGraphEventRecordNodeSetEvent,
14912 * ::cuEventRecordWithFlags,
14913 * ::cuStreamWaitEvent
14914 */
14915CUresult CUDAAPI cuGraphEventWaitNodeSetEvent(CUgraphNode hNode, CUevent event);
14916
14917/**
14918 * \brief Creates an external semaphore signal node and adds it to a graph
14919 *
14920 * Creates a new external semaphore signal node and adds it to \p hGraph with \p
14921 * numDependencies dependencies specified via \p dependencies and arguments specified
14922 * in \p nodeParams. It is possible for \p numDependencies to be 0, in which case the
14923 * node will be placed at the root of the graph. \p dependencies may not have any
14924 * duplicate entries. A handle to the new node will be returned in \p phGraphNode.
14925 *
14926 * Performs a signal operation on a set of externally allocated semaphore objects
14927 * when the node is launched. The operation(s) will occur after all of the node's
14928 * dependencies have completed.
14929 *
14930 * \param phGraphNode - Returns newly created node
14931 * \param hGraph - Graph to which to add the node
14932 * \param dependencies - Dependencies of the node
14933 * \param numDependencies - Number of dependencies
14934 * \param nodeParams - Parameters for the node
14935 *
14936 * \return
14937 * ::CUDA_SUCCESS,
14938 * ::CUDA_ERROR_DEINITIALIZED,
14939 * ::CUDA_ERROR_NOT_INITIALIZED,
14940 * ::CUDA_ERROR_NOT_SUPPORTED,
14941 * ::CUDA_ERROR_INVALID_VALUE
14942 * \note_graph_thread_safety
14943 * \notefnerr
14944 *
14945 * \sa
14946 * ::cuGraphExternalSemaphoresSignalNodeGetParams,
14947 * ::cuGraphExternalSemaphoresSignalNodeSetParams,
14948 * ::cuGraphExecExternalSemaphoresSignalNodeSetParams,
14949 * ::cuGraphAddExternalSemaphoresWaitNode,
14950 * ::cuImportExternalSemaphore,
14951 * ::cuSignalExternalSemaphoresAsync,
14952 * ::cuWaitExternalSemaphoresAsync,
14953 * ::cuGraphCreate,
14954 * ::cuGraphDestroyNode,
14955 * ::cuGraphAddEventRecordNode,
14956 * ::cuGraphAddEventWaitNode,
14957 * ::cuGraphAddChildGraphNode,
14958 * ::cuGraphAddEmptyNode,
14959 * ::cuGraphAddKernelNode,
14960 * ::cuGraphAddMemcpyNode,
14961 * ::cuGraphAddMemsetNode,
14962 */
14963CUresult CUDAAPI cuGraphAddExternalSemaphoresSignalNode(CUgraphNode *phGraphNode, CUgraph hGraph, const CUgraphNode *dependencies, size_t numDependencies, const CUDA_EXT_SEM_SIGNAL_NODE_PARAMS *nodeParams);
14964
14965/**
14966 * \brief Returns an external semaphore signal node's parameters
14967 *
14968 * Returns the parameters of an external semaphore signal node \p hNode in \p params_out.
14969 * The \p extSemArray and \p paramsArray returned in \p params_out,
14970 * are owned by the node. This memory remains valid until the node is destroyed or its
14971 * parameters are modified, and should not be modified
14972 * directly. Use ::cuGraphExternalSemaphoresSignalNodeSetParams to update the
14973 * parameters of this node.
14974 *
14975 * \param hNode - Node to get the parameters for
14976 * \param params_out - Pointer to return the parameters
14977 *
14978 * \return
14979 * ::CUDA_SUCCESS,
14980 * ::CUDA_ERROR_DEINITIALIZED,
14981 * ::CUDA_ERROR_NOT_INITIALIZED,
14982 * ::CUDA_ERROR_INVALID_VALUE
14983 * \note_graph_thread_safety
14984 * \notefnerr
14985 *
14986 * \sa
14987 * ::cuLaunchKernel,
14988 * ::cuGraphAddExternalSemaphoresSignalNode,
14989 * ::cuGraphExternalSemaphoresSignalNodeSetParams,
14990 * ::cuGraphAddExternalSemaphoresWaitNode,
14991 * ::cuSignalExternalSemaphoresAsync,
14992 * ::cuWaitExternalSemaphoresAsync
14993 */
14994CUresult CUDAAPI cuGraphExternalSemaphoresSignalNodeGetParams(CUgraphNode hNode, CUDA_EXT_SEM_SIGNAL_NODE_PARAMS *params_out);
14995
14996/**
14997 * \brief Sets an external semaphore signal node's parameters
14998 *
14999 * Sets the parameters of an external semaphore signal node \p hNode to \p nodeParams.
15000 *
15001 * \param hNode - Node to set the parameters for
15002 * \param nodeParams - Parameters to copy
15003 *
15004 * \return
15005 * ::CUDA_SUCCESS,
15006 * ::CUDA_ERROR_INVALID_VALUE,
15007 * ::CUDA_ERROR_INVALID_HANDLE,
15008 * ::CUDA_ERROR_OUT_OF_MEMORY
15009 * \note_graph_thread_safety
15010 * \notefnerr
15011 *
15012 * \sa
15013 * ::cuGraphAddExternalSemaphoresSignalNode,
15014 * ::cuGraphExternalSemaphoresSignalNodeSetParams,
15015 * ::cuGraphAddExternalSemaphoresWaitNode,
15016 * ::cuSignalExternalSemaphoresAsync,
15017 * ::cuWaitExternalSemaphoresAsync
15018 */
15019CUresult CUDAAPI cuGraphExternalSemaphoresSignalNodeSetParams(CUgraphNode hNode, const CUDA_EXT_SEM_SIGNAL_NODE_PARAMS *nodeParams);
15020
15021/**
15022 * \brief Creates an external semaphore wait node and adds it to a graph
15023 *
15024 * Creates a new external semaphore wait node and adds it to \p hGraph with \p numDependencies
15025 * dependencies specified via \p dependencies and arguments specified in \p nodeParams.
15026 * It is possible for \p numDependencies to be 0, in which case the node will be placed
15027 * at the root of the graph. \p dependencies may not have any duplicate entries. A handle
15028 * to the new node will be returned in \p phGraphNode.
15029 *
15030 * Performs a wait operation on a set of externally allocated semaphore objects
15031 * when the node is launched. The node's dependencies will not be launched until
15032 * the wait operation has completed.
15033 *
15034 * \param phGraphNode - Returns newly created node
15035 * \param hGraph - Graph to which to add the node
15036 * \param dependencies - Dependencies of the node
15037 * \param numDependencies - Number of dependencies
15038 * \param nodeParams - Parameters for the node
15039 *
15040 * \return
15041 * ::CUDA_SUCCESS,
15042 * ::CUDA_ERROR_DEINITIALIZED,
15043 * ::CUDA_ERROR_NOT_INITIALIZED,
15044 * ::CUDA_ERROR_NOT_SUPPORTED,
15045 * ::CUDA_ERROR_INVALID_VALUE
15046 * \note_graph_thread_safety
15047 * \notefnerr
15048 *
15049 * \sa
15050 * ::cuGraphExternalSemaphoresWaitNodeGetParams,
15051 * ::cuGraphExternalSemaphoresWaitNodeSetParams,
15052 * ::cuGraphExecExternalSemaphoresWaitNodeSetParams,
15053 * ::cuGraphAddExternalSemaphoresSignalNode,
15054 * ::cuImportExternalSemaphore,
15055 * ::cuSignalExternalSemaphoresAsync,
15056 * ::cuWaitExternalSemaphoresAsync,
15057 * ::cuGraphCreate,
15058 * ::cuGraphDestroyNode,
15059 * ::cuGraphAddEventRecordNode,
15060 * ::cuGraphAddEventWaitNode,
15061 * ::cuGraphAddChildGraphNode,
15062 * ::cuGraphAddEmptyNode,
15063 * ::cuGraphAddKernelNode,
15064 * ::cuGraphAddMemcpyNode,
15065 * ::cuGraphAddMemsetNode,
15066 */
15067CUresult CUDAAPI cuGraphAddExternalSemaphoresWaitNode(CUgraphNode *phGraphNode, CUgraph hGraph, const CUgraphNode *dependencies, size_t numDependencies, const CUDA_EXT_SEM_WAIT_NODE_PARAMS *nodeParams);
15068
15069/**
15070 * \brief Returns an external semaphore wait node's parameters
15071 *
15072 * Returns the parameters of an external semaphore wait node \p hNode in \p params_out.
15073 * The \p extSemArray and \p paramsArray returned in \p params_out,
15074 * are owned by the node. This memory remains valid until the node is destroyed or its
15075 * parameters are modified, and should not be modified
15076 * directly. Use ::cuGraphExternalSemaphoresSignalNodeSetParams to update the
15077 * parameters of this node.
15078 *
15079 * \param hNode - Node to get the parameters for
15080 * \param params_out - Pointer to return the parameters
15081 *
15082 * \return
15083 * ::CUDA_SUCCESS,
15084 * ::CUDA_ERROR_DEINITIALIZED,
15085 * ::CUDA_ERROR_NOT_INITIALIZED,
15086 * ::CUDA_ERROR_INVALID_VALUE
15087 * \note_graph_thread_safety
15088 * \notefnerr
15089 *
15090 * \sa
15091 * ::cuLaunchKernel,
15092 * ::cuGraphAddExternalSemaphoresWaitNode,
15093 * ::cuGraphExternalSemaphoresWaitNodeSetParams,
15094 * ::cuGraphAddExternalSemaphoresWaitNode,
15095 * ::cuSignalExternalSemaphoresAsync,
15096 * ::cuWaitExternalSemaphoresAsync
15097 */
15098CUresult CUDAAPI cuGraphExternalSemaphoresWaitNodeGetParams(CUgraphNode hNode, CUDA_EXT_SEM_WAIT_NODE_PARAMS *params_out);
15099
15100/**
15101 * \brief Sets an external semaphore wait node's parameters
15102 *
15103 * Sets the parameters of an external semaphore wait node \p hNode to \p nodeParams.
15104 *
15105 * \param hNode - Node to set the parameters for
15106 * \param nodeParams - Parameters to copy
15107 *
15108 * \return
15109 * ::CUDA_SUCCESS,
15110 * ::CUDA_ERROR_INVALID_VALUE,
15111 * ::CUDA_ERROR_INVALID_HANDLE,
15112 * ::CUDA_ERROR_OUT_OF_MEMORY
15113 * \note_graph_thread_safety
15114 * \notefnerr
15115 *
15116 * \sa
15117 * ::cuGraphAddExternalSemaphoresWaitNode,
15118 * ::cuGraphExternalSemaphoresWaitNodeSetParams,
15119 * ::cuGraphAddExternalSemaphoresWaitNode,
15120 * ::cuSignalExternalSemaphoresAsync,
15121 * ::cuWaitExternalSemaphoresAsync
15122 */
15123CUresult CUDAAPI cuGraphExternalSemaphoresWaitNodeSetParams(CUgraphNode hNode, const CUDA_EXT_SEM_WAIT_NODE_PARAMS *nodeParams);
15124
15125/**
15126 * \brief Creates an allocation node and adds it to a graph
15127 *
15128 * Creates a new allocation node and adds it to \p hGraph with \p numDependencies
15129 * dependencies specified via \p dependencies and arguments specified in \p nodeParams.
15130 * It is possible for \p numDependencies to be 0, in which case the node will be placed
15131 * at the root of the graph. \p dependencies may not have any duplicate entries. A handle
15132 * to the new node will be returned in \p phGraphNode.
15133 *
15134 * \param phGraphNode - Returns newly created node
15135 * \param hGraph - Graph to which to add the node
15136 * \param dependencies - Dependencies of the node
15137 * \param numDependencies - Number of dependencies
15138 * \param nodeParams - Parameters for the node
15139 *
15140 * When ::cuGraphAddMemAllocNode creates an allocation node, it returns the address of the allocation in
15141 * \param nodeParams.dptr. The allocation's address remains fixed across instantiations and launches.
15142 *
15143 * If the allocation is freed in the same graph, by creating a free node using ::cuGraphAddMemFreeNode,
15144 * the allocation can be accessed by nodes ordered after the allocation node but before the free node.
15145 * These allocations cannot be freed outside the owning graph, and they can only be freed once in the
15146 * owning graph.
15147 *
15148 * If the allocation is not freed in the same graph, then it can be accessed not only by nodes in the
15149 * graph which are ordered after the allocation node, but also by stream operations ordered after the
15150 * graph's execution but before the allocation is freed.
15151 *
15152 * Allocations which are not freed in the same graph can be freed by:
15153 * - passing the allocation to ::cuMemFreeAsync or ::cuMemFree;
15154 * - launching a graph with a free node for that allocation; or
15155 * - specifying ::CUDA_GRAPH_INSTANTIATE_FLAG_AUTO_FREE_ON_LAUNCH during instantiation, which makes
15156 * each launch behave as though it called ::cuMemFreeAsync for every unfreed allocation.
15157 *
15158 * It is not possible to free an allocation in both the owning graph and another graph. If the allocation
15159 * is freed in the same graph, a free node cannot be added to another graph. If the allocation is freed
15160 * in another graph, a free node can no longer be added to the owning graph.
15161 *
15162 * The following restrictions apply to graphs which contain allocation and/or memory free nodes:
15163 * - Nodes and edges of the graph cannot be deleted.
15164 * - The graph cannot be used in a child node.
15165 * - Only one instantiation of the graph may exist at any point in time.
15166 * - The graph cannot be cloned.
15167 *
15168 * \return
15169 * ::CUDA_SUCCESS,
15170 * ::CUDA_ERROR_DEINITIALIZED,
15171 * ::CUDA_ERROR_NOT_INITIALIZED,
15172 * ::CUDA_ERROR_NOT_SUPPORTED,
15173 * ::CUDA_ERROR_INVALID_VALUE
15174 * \note_graph_thread_safety
15175 * \notefnerr
15176 *
15177 * \sa
15178 * ::cuGraphAddMemFreeNode,
15179 * ::cuGraphMemAllocNodeGetParams,
15180 * ::cuDeviceGraphMemTrim,
15181 * ::cuDeviceGetGraphMemAttribute,
15182 * ::cuDeviceSetGraphMemAttribute,
15183 * ::cuMemAllocAsync,
15184 * ::cuMemFreeAsync,
15185 * ::cuGraphCreate,
15186 * ::cuGraphDestroyNode,
15187 * ::cuGraphAddChildGraphNode,
15188 * ::cuGraphAddEmptyNode,
15189 * ::cuGraphAddEventRecordNode,
15190 * ::cuGraphAddEventWaitNode,
15191 * ::cuGraphAddExternalSemaphoresSignalNode,
15192 * ::cuGraphAddExternalSemaphoresWaitNode,
15193 * ::cuGraphAddKernelNode,
15194 * ::cuGraphAddMemcpyNode,
15195 * ::cuGraphAddMemsetNode
15196 */
15197CUresult CUDAAPI cuGraphAddMemAllocNode(CUgraphNode *phGraphNode, CUgraph hGraph, const CUgraphNode *dependencies, size_t numDependencies, CUDA_MEM_ALLOC_NODE_PARAMS *nodeParams);
15198
15199/**
15200 * \brief Returns a memory alloc node's parameters
15201 *
15202 * Returns the parameters of a memory alloc node \p hNode in \p params_out.
15203 * The \p poolProps and \p accessDescs returned in \p params_out, are owned by the
15204 * node. This memory remains valid until the node is destroyed. The returned
15205 * parameters must not be modified.
15206 *
15207 * \param hNode - Node to get the parameters for
15208 * \param params_out - Pointer to return the parameters
15209 *
15210 * \return
15211 * ::CUDA_SUCCESS,
15212 * ::CUDA_ERROR_DEINITIALIZED,
15213 * ::CUDA_ERROR_NOT_INITIALIZED,
15214 * ::CUDA_ERROR_INVALID_VALUE
15215 * \note_graph_thread_safety
15216 * \notefnerr
15217 *
15218 * \sa
15219 * ::cuGraphAddMemAllocNode,
15220 * ::cuGraphMemFreeNodeGetParams
15221 */
15222CUresult CUDAAPI cuGraphMemAllocNodeGetParams(CUgraphNode hNode, CUDA_MEM_ALLOC_NODE_PARAMS *params_out);
15223
15224/**
15225 * \brief Creates a memory free node and adds it to a graph
15226 *
15227 * Creates a new memory free node and adds it to \p hGraph with \p numDependencies
15228 * dependencies specified via \p dependencies and arguments specified in \p nodeParams.
15229 * It is possible for \p numDependencies to be 0, in which case the node will be placed
15230 * at the root of the graph. \p dependencies may not have any duplicate entries. A handle
15231 * to the new node will be returned in \p phGraphNode.
15232 *
15233 * \param phGraphNode - Returns newly created node
15234 * \param hGraph - Graph to which to add the node
15235 * \param dependencies - Dependencies of the node
15236 * \param numDependencies - Number of dependencies
15237 * \param dptr - Address of memory to free
15238 *
15239 * ::cuGraphAddMemFreeNode will return ::CUDA_ERROR_INVALID_VALUE if the user attempts to free:
15240 * - an allocation twice in the same graph.
15241 * - an address that was not returned by an allocation node.
15242 * - an invalid address.
15243 *
15244 * The following restrictions apply to graphs which contain allocation and/or memory free nodes:
15245 * - Nodes and edges of the graph cannot be deleted.
15246 * - The graph cannot be used in a child node.
15247 * - Only one instantiation of the graph may exist at any point in time.
15248 * - The graph cannot be cloned.
15249 *
15250 * \return
15251 * ::CUDA_SUCCESS,
15252 * ::CUDA_ERROR_DEINITIALIZED,
15253 * ::CUDA_ERROR_NOT_INITIALIZED,
15254 * ::CUDA_ERROR_NOT_SUPPORTED,
15255 * ::CUDA_ERROR_INVALID_VALUE
15256 * \note_graph_thread_safety
15257 * \notefnerr
15258 *
15259 * \sa
15260 * ::cuGraphAddMemAllocNode,
15261 * ::cuGraphMemFreeNodeGetParams,
15262 * ::cuDeviceGraphMemTrim,
15263 * ::cuDeviceGetGraphMemAttribute,
15264 * ::cuDeviceSetGraphMemAttribute,
15265 * ::cuMemAllocAsync,
15266 * ::cuMemFreeAsync,
15267 * ::cuGraphCreate,
15268 * ::cuGraphDestroyNode,
15269 * ::cuGraphAddChildGraphNode,
15270 * ::cuGraphAddEmptyNode,
15271 * ::cuGraphAddEventRecordNode,
15272 * ::cuGraphAddEventWaitNode,
15273 * ::cuGraphAddExternalSemaphoresSignalNode,
15274 * ::cuGraphAddExternalSemaphoresWaitNode,
15275 * ::cuGraphAddKernelNode,
15276 * ::cuGraphAddMemcpyNode,
15277 * ::cuGraphAddMemsetNode
15278 */
15279CUresult CUDAAPI cuGraphAddMemFreeNode(CUgraphNode *phGraphNode, CUgraph hGraph, const CUgraphNode *dependencies, size_t numDependencies, CUdeviceptr dptr);
15280
15281/**
15282 * \brief Returns a memory free node's parameters
15283 *
15284 * Returns the address of a memory free node \p hNode in \p dptr_out.
15285 *
15286 * \param hNode - Node to get the parameters for
15287 * \param dptr_out - Pointer to return the device address
15288 *
15289 * \return
15290 * ::CUDA_SUCCESS,
15291 * ::CUDA_ERROR_DEINITIALIZED,
15292 * ::CUDA_ERROR_NOT_INITIALIZED,
15293 * ::CUDA_ERROR_INVALID_VALUE
15294 * \note_graph_thread_safety
15295 * \notefnerr
15296 *
15297 * \sa
15298 * ::cuGraphAddMemFreeNode,
15299 * ::cuGraphMemAllocNodeGetParams
15300 */
15301CUresult CUDAAPI cuGraphMemFreeNodeGetParams(CUgraphNode hNode, CUdeviceptr *dptr_out);
15302
15303/**
15304 * \brief Free unused memory that was cached on the specified device for use with graphs back to the OS.
15305 *
15306 * Blocks which are not in use by a graph that is either currently executing or scheduled to execute are
15307 * freed back to the operating system.
15308 *
15309 * \param device - The device for which cached memory should be freed.
15310 *
15311 * \return
15312 * ::CUDA_SUCCESS,
15313 * ::CUDA_ERROR_INVALID_DEVICE
15314 *
15315 * \sa
15316 * ::cuGraphAddMemAllocNode,
15317 * ::cuGraphAddMemFreeNode
15318 */
15319CUresult CUDAAPI cuDeviceGraphMemTrim(CUdevice device);
15320
15321/**
15322 * \brief Query asynchronous allocation attributes related to graphs
15323 *
15324 * Valid attributes are:
15325 *
15326 * - ::CU_GRAPH_MEM_ATTR_USED_MEM_CURRENT: Amount of memory, in bytes, currently associated with graphs
15327 * - ::CU_GRAPH_MEM_ATTR_USED_MEM_HIGH: High watermark of memory, in bytes, associated with graphs since the
15328 * last time it was reset. High watermark can only be reset to zero.
15329 * - ::CU_GRAPH_MEM_ATTR_RESERVED_MEM_CURRENT: Amount of memory, in bytes, currently allocated for use by
15330 * the CUDA graphs asynchronous allocator.
15331 * - ::CU_GRAPH_MEM_ATTR_RESERVED_MEM_HIGH: High watermark of memory, in bytes, currently allocated for use by
15332 * the CUDA graphs asynchronous allocator.
15333 *
15334 * \param device - Specifies the scope of the query
15335 * \param attr - attribute to get
15336 * \param value - retrieved value
15337 *
15338 * \return
15339 * ::CUDA_SUCCESS,
15340 * ::CUDA_ERROR_INVALID_DEVICE
15341 *
15342 * \sa
15343 * ::cuGraphAddMemAllocNode,
15344 * ::cuGraphAddMemFreeNode
15345 */
15346CUresult CUDAAPI cuDeviceGetGraphMemAttribute(CUdevice device, CUgraphMem_attribute attr, void* value);
15347
15348/**
15349 * \brief Set asynchronous allocation attributes related to graphs
15350 *
15351 * Valid attributes are:
15352 *
15353 * - ::CU_GRAPH_MEM_ATTR_USED_MEM_HIGH: High watermark of memory, in bytes, associated with graphs since the
15354 * last time it was reset. High watermark can only be reset to zero.
15355 * - ::CU_GRAPH_MEM_ATTR_RESERVED_MEM_HIGH: High watermark of memory, in bytes, currently allocated for use by
15356 * the CUDA graphs asynchronous allocator.
15357 *
15358 * \param device - Specifies the scope of the query
15359 * \param attr - attribute to get
15360 * \param value - pointer to value to set
15361 *
15362 * \return
15363 * ::CUDA_SUCCESS,
15364 * ::CUDA_ERROR_INVALID_DEVICE
15365 *
15366 * \sa
15367 * ::cuGraphAddMemAllocNode,
15368 * ::cuGraphAddMemFreeNode
15369 */
15370CUresult CUDAAPI cuDeviceSetGraphMemAttribute(CUdevice device, CUgraphMem_attribute attr, void* value);
15371
15372/**
15373 * \brief Clones a graph
15374 *
15375 * This function creates a copy of \p originalGraph and returns it in \p phGraphClone.
15376 * All parameters are copied into the cloned graph. The original graph may be modified
15377 * after this call without affecting the clone.
15378 *
15379 * Child graph nodes in the original graph are recursively copied into the clone.
15380 *
15381 * \param phGraphClone - Returns newly created cloned graph
15382 * \param originalGraph - Graph to clone
15383 *
15384 * \return
15385 * ::CUDA_SUCCESS,
15386 * ::CUDA_ERROR_INVALID_VALUE,
15387 * ::CUDA_ERROR_OUT_OF_MEMORY
15388 * \note_graph_thread_safety
15389 * \notefnerr
15390 *
15391 * \sa
15392 * ::cuGraphCreate,
15393 * ::cuGraphNodeFindInClone
15394 */
15395CUresult CUDAAPI cuGraphClone(CUgraph *phGraphClone, CUgraph originalGraph);
15396
15397/**
15398 * \brief Finds a cloned version of a node
15399 *
15400 * This function returns the node in \p hClonedGraph corresponding to \p hOriginalNode
15401 * in the original graph.
15402 *
15403 * \p hClonedGraph must have been cloned from \p hOriginalGraph via ::cuGraphClone.
15404 * \p hOriginalNode must have been in \p hOriginalGraph at the time of the call to
15405 * ::cuGraphClone, and the corresponding cloned node in \p hClonedGraph must not have
15406 * been removed. The cloned node is then returned via \p phClonedNode.
15407 *
15408 * \param phNode - Returns handle to the cloned node
15409 * \param hOriginalNode - Handle to the original node
15410 * \param hClonedGraph - Cloned graph to query
15411 *
15412 * \return
15413 * ::CUDA_SUCCESS,
15414 * ::CUDA_ERROR_INVALID_VALUE,
15415 * \note_graph_thread_safety
15416 * \notefnerr
15417 *
15418 * \sa
15419 * ::cuGraphClone
15420 */
15421CUresult CUDAAPI cuGraphNodeFindInClone(CUgraphNode *phNode, CUgraphNode hOriginalNode, CUgraph hClonedGraph);
15422
15423/**
15424 * \brief Returns a node's type
15425 *
15426 * Returns the node type of \p hNode in \p type.
15427 *
15428 * \param hNode - Node to query
15429 * \param type - Pointer to return the node type
15430 *
15431 * \return
15432 * ::CUDA_SUCCESS,
15433 * ::CUDA_ERROR_DEINITIALIZED,
15434 * ::CUDA_ERROR_NOT_INITIALIZED,
15435 * ::CUDA_ERROR_INVALID_VALUE
15436 * \note_graph_thread_safety
15437 * \notefnerr
15438 *
15439 * \sa
15440 * ::cuGraphGetNodes,
15441 * ::cuGraphGetRootNodes,
15442 * ::cuGraphChildGraphNodeGetGraph,
15443 * ::cuGraphKernelNodeGetParams,
15444 * ::cuGraphKernelNodeSetParams,
15445 * ::cuGraphHostNodeGetParams,
15446 * ::cuGraphHostNodeSetParams,
15447 * ::cuGraphMemcpyNodeGetParams,
15448 * ::cuGraphMemcpyNodeSetParams,
15449 * ::cuGraphMemsetNodeGetParams,
15450 * ::cuGraphMemsetNodeSetParams
15451 */
15452CUresult CUDAAPI cuGraphNodeGetType(CUgraphNode hNode, CUgraphNodeType *type);
15453
15454/**
15455 * \brief Returns a graph's nodes
15456 *
15457 * Returns a list of \p hGraph's nodes. \p nodes may be NULL, in which case this
15458 * function will return the number of nodes in \p numNodes. Otherwise,
15459 * \p numNodes entries will be filled in. If \p numNodes is higher than the actual
15460 * number of nodes, the remaining entries in \p nodes will be set to NULL, and the
15461 * number of nodes actually obtained will be returned in \p numNodes.
15462 *
15463 * \param hGraph - Graph to query
15464 * \param nodes - Pointer to return the nodes
15465 * \param numNodes - See description
15466 *
15467 * \return
15468 * ::CUDA_SUCCESS,
15469 * ::CUDA_ERROR_DEINITIALIZED,
15470 * ::CUDA_ERROR_NOT_INITIALIZED,
15471 * ::CUDA_ERROR_INVALID_VALUE
15472 * \note_graph_thread_safety
15473 * \notefnerr
15474 *
15475 * \sa
15476 * ::cuGraphCreate,
15477 * ::cuGraphGetRootNodes,
15478 * ::cuGraphGetEdges,
15479 * ::cuGraphNodeGetType,
15480 * ::cuGraphNodeGetDependencies,
15481 * ::cuGraphNodeGetDependentNodes
15482 */
15483CUresult CUDAAPI cuGraphGetNodes(CUgraph hGraph, CUgraphNode *nodes, size_t *numNodes);
15484
15485/**
15486 * \brief Returns a graph's root nodes
15487 *
15488 * Returns a list of \p hGraph's root nodes. \p rootNodes may be NULL, in which case this
15489 * function will return the number of root nodes in \p numRootNodes. Otherwise,
15490 * \p numRootNodes entries will be filled in. If \p numRootNodes is higher than the actual
15491 * number of root nodes, the remaining entries in \p rootNodes will be set to NULL, and the
15492 * number of nodes actually obtained will be returned in \p numRootNodes.
15493 *
15494 * \param hGraph - Graph to query
15495 * \param rootNodes - Pointer to return the root nodes
15496 * \param numRootNodes - See description
15497 *
15498 * \return
15499 * ::CUDA_SUCCESS,
15500 * ::CUDA_ERROR_DEINITIALIZED,
15501 * ::CUDA_ERROR_NOT_INITIALIZED,
15502 * ::CUDA_ERROR_INVALID_VALUE
15503 * \note_graph_thread_safety
15504 * \notefnerr
15505 *
15506 * \sa
15507 * ::cuGraphCreate,
15508 * ::cuGraphGetNodes,
15509 * ::cuGraphGetEdges,
15510 * ::cuGraphNodeGetType,
15511 * ::cuGraphNodeGetDependencies,
15512 * ::cuGraphNodeGetDependentNodes
15513 */
15514CUresult CUDAAPI cuGraphGetRootNodes(CUgraph hGraph, CUgraphNode *rootNodes, size_t *numRootNodes);
15515
15516/**
15517 * \brief Returns a graph's dependency edges
15518 *
15519 * Returns a list of \p hGraph's dependency edges. Edges are returned via corresponding
15520 * indices in \p from and \p to; that is, the node in \p to[i] has a dependency on the
15521 * node in \p from[i]. \p from and \p to may both be NULL, in which
15522 * case this function only returns the number of edges in \p numEdges. Otherwise,
15523 * \p numEdges entries will be filled in. If \p numEdges is higher than the actual
15524 * number of edges, the remaining entries in \p from and \p to will be set to NULL, and
15525 * the number of edges actually returned will be written to \p numEdges.
15526 *
15527 * \param hGraph - Graph to get the edges from
15528 * \param from - Location to return edge endpoints
15529 * \param to - Location to return edge endpoints
15530 * \param numEdges - See description
15531 *
15532 * \return
15533 * ::CUDA_SUCCESS,
15534 * ::CUDA_ERROR_DEINITIALIZED,
15535 * ::CUDA_ERROR_NOT_INITIALIZED,
15536 * ::CUDA_ERROR_INVALID_VALUE
15537 * \note_graph_thread_safety
15538 * \notefnerr
15539 *
15540 * \sa
15541 * ::cuGraphGetNodes,
15542 * ::cuGraphGetRootNodes,
15543 * ::cuGraphAddDependencies,
15544 * ::cuGraphRemoveDependencies,
15545 * ::cuGraphNodeGetDependencies,
15546 * ::cuGraphNodeGetDependentNodes
15547 */
15548CUresult CUDAAPI cuGraphGetEdges(CUgraph hGraph, CUgraphNode *from, CUgraphNode *to, size_t *numEdges);
15549
15550/**
15551 * \brief Returns a node's dependencies
15552 *
15553 * Returns a list of \p node's dependencies. \p dependencies may be NULL, in which case this
15554 * function will return the number of dependencies in \p numDependencies. Otherwise,
15555 * \p numDependencies entries will be filled in. If \p numDependencies is higher than the actual
15556 * number of dependencies, the remaining entries in \p dependencies will be set to NULL, and the
15557 * number of nodes actually obtained will be returned in \p numDependencies.
15558 *
15559 * \param hNode - Node to query
15560 * \param dependencies - Pointer to return the dependencies
15561 * \param numDependencies - See description
15562 *
15563 * \return
15564 * ::CUDA_SUCCESS,
15565 * ::CUDA_ERROR_DEINITIALIZED,
15566 * ::CUDA_ERROR_NOT_INITIALIZED,
15567 * ::CUDA_ERROR_INVALID_VALUE
15568 * \note_graph_thread_safety
15569 * \notefnerr
15570 *
15571 * \sa
15572 * ::cuGraphNodeGetDependentNodes,
15573 * ::cuGraphGetNodes,
15574 * ::cuGraphGetRootNodes,
15575 * ::cuGraphGetEdges,
15576 * ::cuGraphAddDependencies,
15577 * ::cuGraphRemoveDependencies
15578 */
15579CUresult CUDAAPI cuGraphNodeGetDependencies(CUgraphNode hNode, CUgraphNode *dependencies, size_t *numDependencies);
15580
15581/**
15582 * \brief Returns a node's dependent nodes
15583 *
15584 * Returns a list of \p node's dependent nodes. \p dependentNodes may be NULL, in which
15585 * case this function will return the number of dependent nodes in \p numDependentNodes.
15586 * Otherwise, \p numDependentNodes entries will be filled in. If \p numDependentNodes is
15587 * higher than the actual number of dependent nodes, the remaining entries in
15588 * \p dependentNodes will be set to NULL, and the number of nodes actually obtained will
15589 * be returned in \p numDependentNodes.
15590 *
15591 * \param hNode - Node to query
15592 * \param dependentNodes - Pointer to return the dependent nodes
15593 * \param numDependentNodes - See description
15594 *
15595 * \return
15596 * ::CUDA_SUCCESS,
15597 * ::CUDA_ERROR_DEINITIALIZED,
15598 * ::CUDA_ERROR_NOT_INITIALIZED,
15599 * ::CUDA_ERROR_INVALID_VALUE
15600 * \note_graph_thread_safety
15601 * \notefnerr
15602 *
15603 * \sa
15604 * ::cuGraphNodeGetDependencies,
15605 * ::cuGraphGetNodes,
15606 * ::cuGraphGetRootNodes,
15607 * ::cuGraphGetEdges,
15608 * ::cuGraphAddDependencies,
15609 * ::cuGraphRemoveDependencies
15610 */
15611CUresult CUDAAPI cuGraphNodeGetDependentNodes(CUgraphNode hNode, CUgraphNode *dependentNodes, size_t *numDependentNodes);
15612
15613/**
15614 * \brief Adds dependency edges to a graph
15615 *
15616 * The number of dependencies to be added is defined by \p numDependencies
15617 * Elements in \p from and \p to at corresponding indices define a dependency.
15618 * Each node in \p from and \p to must belong to \p hGraph.
15619 *
15620 * If \p numDependencies is 0, elements in \p from and \p to will be ignored.
15621 * Specifying an existing dependency will return an error.
15622 *
15623 * \param hGraph - Graph to which dependencies are added
15624 * \param from - Array of nodes that provide the dependencies
15625 * \param to - Array of dependent nodes
15626 * \param numDependencies - Number of dependencies to be added
15627 *
15628 * \return
15629 * ::CUDA_SUCCESS,
15630 * ::CUDA_ERROR_INVALID_VALUE
15631 * \note_graph_thread_safety
15632 * \notefnerr
15633 *
15634 * \sa
15635 * ::cuGraphRemoveDependencies,
15636 * ::cuGraphGetEdges,
15637 * ::cuGraphNodeGetDependencies,
15638 * ::cuGraphNodeGetDependentNodes
15639 */
15640CUresult CUDAAPI cuGraphAddDependencies(CUgraph hGraph, const CUgraphNode *from, const CUgraphNode *to, size_t numDependencies);
15641
15642/**
15643 * \brief Removes dependency edges from a graph
15644 *
15645 * The number of \p dependencies to be removed is defined by \p numDependencies.
15646 * Elements in \p from and \p to at corresponding indices define a dependency.
15647 * Each node in \p from and \p to must belong to \p hGraph.
15648 *
15649 * If \p numDependencies is 0, elements in \p from and \p to will be ignored.
15650 * Specifying a non-existing dependency will return an error.
15651 *
15652 * Dependencies cannot be removed from graphs which contain allocation or free nodes.
15653 * Any attempt to do so will return an error.
15654 *
15655 * \param hGraph - Graph from which to remove dependencies
15656 * \param from - Array of nodes that provide the dependencies
15657 * \param to - Array of dependent nodes
15658 * \param numDependencies - Number of dependencies to be removed
15659 *
15660 * \return
15661 * ::CUDA_SUCCESS,
15662 * ::CUDA_ERROR_INVALID_VALUE
15663 * \note_graph_thread_safety
15664 * \notefnerr
15665 *
15666 * \sa
15667 * ::cuGraphAddDependencies,
15668 * ::cuGraphGetEdges,
15669 * ::cuGraphNodeGetDependencies,
15670 * ::cuGraphNodeGetDependentNodes
15671 */
15672CUresult CUDAAPI cuGraphRemoveDependencies(CUgraph hGraph, const CUgraphNode *from, const CUgraphNode *to, size_t numDependencies);
15673
15674/**
15675 * \brief Remove a node from the graph
15676 *
15677 * Removes \p hNode from its graph. This operation also severs any dependencies of other nodes
15678 * on \p hNode and vice versa.
15679 *
15680 * Nodes which belong to a graph which contains allocation or free nodes cannot be destroyed.
15681 * Any attempt to do so will return an error.
15682 *
15683 * \param hNode - Node to remove
15684 *
15685 * \return
15686 * ::CUDA_SUCCESS,
15687 * ::CUDA_ERROR_INVALID_VALUE
15688 * \note_graph_thread_safety
15689 * \notefnerr
15690 *
15691 * \sa
15692 * ::cuGraphAddChildGraphNode,
15693 * ::cuGraphAddEmptyNode,
15694 * ::cuGraphAddKernelNode,
15695 * ::cuGraphAddHostNode,
15696 * ::cuGraphAddMemcpyNode,
15697 * ::cuGraphAddMemsetNode
15698 */
15699CUresult CUDAAPI cuGraphDestroyNode(CUgraphNode hNode);
15700
15701/**
15702 * \brief Creates an executable graph from a graph
15703 *
15704 * Instantiates \p hGraph as an executable graph. The graph is validated for any
15705 * structural constraints or intra-node constraints which were not previously
15706 * validated. If instantiation is successful, a handle to the instantiated graph
15707 * is returned in \p phGraphExec.
15708 *
15709 * If there are any errors, diagnostic information may be returned in \p errorNode and
15710 * \p logBuffer. This is the primary way to inspect instantiation errors. The output
15711 * will be null terminated unless the diagnostics overflow
15712 * the buffer. In this case, they will be truncated, and the last byte can be
15713 * inspected to determine if truncation occurred.
15714 *
15715 * \param phGraphExec - Returns instantiated graph
15716 * \param hGraph - Graph to instantiate
15717 * \param phErrorNode - In case of an instantiation error, this may be modified to
15718 * indicate a node contributing to the error
15719 * \param logBuffer - A character buffer to store diagnostic messages
15720 * \param bufferSize - Size of the log buffer in bytes
15721 *
15722 * \return
15723 * ::CUDA_SUCCESS,
15724 * ::CUDA_ERROR_DEINITIALIZED,
15725 * ::CUDA_ERROR_NOT_INITIALIZED,
15726 * ::CUDA_ERROR_INVALID_VALUE
15727 * \note_graph_thread_safety
15728 * \notefnerr
15729 *
15730 * \sa
15731 * ::cuGraphInstantiateWithFlags,
15732 * ::cuGraphCreate,
15733 * ::cuGraphUpload,
15734 * ::cuGraphLaunch,
15735 * ::cuGraphExecDestroy
15736 */
15737CUresult CUDAAPI cuGraphInstantiate(CUgraphExec *phGraphExec, CUgraph hGraph, CUgraphNode *phErrorNode, char *logBuffer, size_t bufferSize);
15738
15739/**
15740 * \brief Creates an executable graph from a graph
15741 *
15742 * Instantiates \p hGraph as an executable graph. The graph is validated for any
15743 * structural constraints or intra-node constraints which were not previously
15744 * validated. If instantiation is successful, a handle to the instantiated graph
15745 * is returned in \p phGraphExec.
15746 *
15747 * The \p flags parameter controls the behavior of instantiation and subsequent
15748 * graph launches. Valid flags are:
15749 *
15750 * - ::CUDA_GRAPH_INSTANTIATE_FLAG_AUTO_FREE_ON_LAUNCH, which configures a
15751 * graph containing memory allocation nodes to automatically free any
15752 * unfreed memory allocations before the graph is relaunched.
15753 *
15754 * If \p hGraph contains any allocation or free nodes, there can be at most one
15755 * executable graph in existence for that graph at a time.
15756 *
15757 * An attempt to instantiate a second executable graph before destroying the first
15758 * with ::cuGraphExecDestroy will result in an error.
15759 *
15760 * \param phGraphExec - Returns instantiated graph
15761 * \param hGraph - Graph to instantiate
15762 * \param flags - Flags to control instantiation. See ::CUgraphInstantiate_flags.
15763 *
15764 * \return
15765 * ::CUDA_SUCCESS,
15766 * ::CUDA_ERROR_DEINITIALIZED,
15767 * ::CUDA_ERROR_NOT_INITIALIZED,
15768 * ::CUDA_ERROR_INVALID_VALUE
15769 * \note_graph_thread_safety
15770 * \notefnerr
15771 *
15772 * \sa
15773 * ::cuGraphInstantiate,
15774 * ::cuGraphCreate,
15775 * ::cuGraphUpload,
15776 * ::cuGraphLaunch,
15777 * ::cuGraphExecDestroy
15778 */
15779CUresult CUDAAPI cuGraphInstantiateWithFlags(CUgraphExec *phGraphExec, CUgraph hGraph, unsigned long long flags);
15780
15781/**
15782 * \brief Sets the parameters for a kernel node in the given graphExec
15783 *
15784 * Sets the parameters of a kernel node in an executable graph \p hGraphExec.
15785 * The node is identified by the corresponding node \p hNode in the
15786 * non-executable graph, from which the executable graph was instantiated.
15787 *
15788 * \p hNode must not have been removed from the original graph. The \p func field
15789 * of \p nodeParams cannot be modified and must match the original value.
15790 * All other values can be modified.
15791 *
15792 * The modifications only affect future launches of \p hGraphExec. Already
15793 * enqueued or running launches of \p hGraphExec are not affected by this call.
15794 * \p hNode is also not modified by this call.
15795 *
15796 * \param hGraphExec - The executable graph in which to set the specified node
15797 * \param hNode - kernel node from the graph from which graphExec was instantiated
15798 * \param nodeParams - Updated Parameters to set
15799 *
15800 * \return
15801 * ::CUDA_SUCCESS,
15802 * ::CUDA_ERROR_INVALID_VALUE,
15803 * \note_graph_thread_safety
15804 * \notefnerr
15805 *
15806 * \sa
15807 * ::cuGraphAddKernelNode,
15808 * ::cuGraphKernelNodeSetParams,
15809 * ::cuGraphExecMemcpyNodeSetParams,
15810 * ::cuGraphExecMemsetNodeSetParams,
15811 * ::cuGraphExecHostNodeSetParams,
15812 * ::cuGraphExecChildGraphNodeSetParams,
15813 * ::cuGraphExecEventRecordNodeSetEvent,
15814 * ::cuGraphExecEventWaitNodeSetEvent,
15815 * ::cuGraphExecExternalSemaphoresSignalNodeSetParams,
15816 * ::cuGraphExecExternalSemaphoresWaitNodeSetParams,
15817 * ::cuGraphExecUpdate,
15818 * ::cuGraphInstantiate
15819 */
15820CUresult CUDAAPI cuGraphExecKernelNodeSetParams(CUgraphExec hGraphExec, CUgraphNode hNode, const CUDA_KERNEL_NODE_PARAMS *nodeParams);
15821
15822/**
15823 * \brief Sets the parameters for a memcpy node in the given graphExec.
15824 *
15825 * Updates the work represented by \p hNode in \p hGraphExec as though \p hNode had
15826 * contained \p copyParams at instantiation. hNode must remain in the graph which was
15827 * used to instantiate \p hGraphExec. Changed edges to and from hNode are ignored.
15828 *
15829 * The source and destination memory in \p copyParams must be allocated from the same
15830 * contexts as the original source and destination memory. Both the instantiation-time
15831 * memory operands and the memory operands in \p copyParams must be 1-dimensional.
15832 * Zero-length operations are not supported.
15833 *
15834 * The modifications only affect future launches of \p hGraphExec. Already enqueued
15835 * or running launches of \p hGraphExec are not affected by this call. hNode is also
15836 * not modified by this call.
15837 *
15838 * Returns CUDA_ERROR_INVALID_VALUE if the memory operands' mappings changed or
15839 * either the original or new memory operands are multidimensional.
15840 *
15841 * \param hGraphExec - The executable graph in which to set the specified node
15842 * \param hNode - Memcpy node from the graph which was used to instantiate graphExec
15843 * \param copyParams - The updated parameters to set
15844 * \param ctx - Context on which to run the node
15845 *
15846 * \return
15847 * ::CUDA_SUCCESS,
15848 * ::CUDA_ERROR_INVALID_VALUE,
15849 * \note_graph_thread_safety
15850 * \notefnerr
15851 *
15852 * \sa
15853 * ::cuGraphAddMemcpyNode,
15854 * ::cuGraphMemcpyNodeSetParams,
15855 * ::cuGraphExecKernelNodeSetParams,
15856 * ::cuGraphExecMemsetNodeSetParams,
15857 * ::cuGraphExecHostNodeSetParams,
15858 * ::cuGraphExecChildGraphNodeSetParams,
15859 * ::cuGraphExecEventRecordNodeSetEvent,
15860 * ::cuGraphExecEventWaitNodeSetEvent,
15861 * ::cuGraphExecExternalSemaphoresSignalNodeSetParams,
15862 * ::cuGraphExecExternalSemaphoresWaitNodeSetParams,
15863 * ::cuGraphExecUpdate,
15864 * ::cuGraphInstantiate
15865 */
15866CUresult CUDAAPI cuGraphExecMemcpyNodeSetParams(CUgraphExec hGraphExec, CUgraphNode hNode, const CUDA_MEMCPY3D *copyParams, CUcontext ctx);
15867
15868/**
15869 * \brief Sets the parameters for a memset node in the given graphExec.
15870 *
15871 * Updates the work represented by \p hNode in \p hGraphExec as though \p hNode had
15872 * contained \p memsetParams at instantiation. hNode must remain in the graph which was
15873 * used to instantiate \p hGraphExec. Changed edges to and from hNode are ignored.
15874 *
15875 * The destination memory in \p memsetParams must be allocated from the same
15876 * contexts as the original destination memory. Both the instantiation-time
15877 * memory operand and the memory operand in \p memsetParams must be 1-dimensional.
15878 * Zero-length operations are not supported.
15879 *
15880 * The modifications only affect future launches of \p hGraphExec. Already enqueued
15881 * or running launches of \p hGraphExec are not affected by this call. hNode is also
15882 * not modified by this call.
15883 *
15884 * Returns CUDA_ERROR_INVALID_VALUE if the memory operand's mappings changed or
15885 * either the original or new memory operand are multidimensional.
15886 *
15887 * \param hGraphExec - The executable graph in which to set the specified node
15888 * \param hNode - Memset node from the graph which was used to instantiate graphExec
15889 * \param memsetParams - The updated parameters to set
15890 * \param ctx - Context on which to run the node
15891 *
15892 * \return
15893 * ::CUDA_SUCCESS,
15894 * ::CUDA_ERROR_INVALID_VALUE,
15895 * \note_graph_thread_safety
15896 * \notefnerr
15897 *
15898 * \sa
15899 * ::cuGraphAddMemsetNode,
15900 * ::cuGraphMemsetNodeSetParams,
15901 * ::cuGraphExecKernelNodeSetParams,
15902 * ::cuGraphExecMemcpyNodeSetParams,
15903 * ::cuGraphExecHostNodeSetParams,
15904 * ::cuGraphExecChildGraphNodeSetParams,
15905 * ::cuGraphExecEventRecordNodeSetEvent,
15906 * ::cuGraphExecEventWaitNodeSetEvent,
15907 * ::cuGraphExecExternalSemaphoresSignalNodeSetParams,
15908 * ::cuGraphExecExternalSemaphoresWaitNodeSetParams,
15909 * ::cuGraphExecUpdate,
15910 * ::cuGraphInstantiate
15911 */
15912CUresult CUDAAPI cuGraphExecMemsetNodeSetParams(CUgraphExec hGraphExec, CUgraphNode hNode, const CUDA_MEMSET_NODE_PARAMS *memsetParams, CUcontext ctx);
15913
15914/**
15915 * \brief Sets the parameters for a host node in the given graphExec.
15916 *
15917 * Updates the work represented by \p hNode in \p hGraphExec as though \p hNode had
15918 * contained \p nodeParams at instantiation. hNode must remain in the graph which was
15919 * used to instantiate \p hGraphExec. Changed edges to and from hNode are ignored.
15920 *
15921 * The modifications only affect future launches of \p hGraphExec. Already enqueued
15922 * or running launches of \p hGraphExec are not affected by this call. hNode is also
15923 * not modified by this call.
15924 *
15925 * \param hGraphExec - The executable graph in which to set the specified node
15926 * \param hNode - Host node from the graph which was used to instantiate graphExec
15927 * \param nodeParams - The updated parameters to set
15928 *
15929 * \return
15930 * ::CUDA_SUCCESS,
15931 * ::CUDA_ERROR_INVALID_VALUE,
15932 * \note_graph_thread_safety
15933 * \notefnerr
15934 *
15935 * \sa
15936 * ::cuGraphAddHostNode,
15937 * ::cuGraphHostNodeSetParams,
15938 * ::cuGraphExecKernelNodeSetParams,
15939 * ::cuGraphExecMemcpyNodeSetParams,
15940 * ::cuGraphExecMemsetNodeSetParams,
15941 * ::cuGraphExecChildGraphNodeSetParams,
15942 * ::cuGraphExecEventRecordNodeSetEvent,
15943 * ::cuGraphExecEventWaitNodeSetEvent,
15944 * ::cuGraphExecExternalSemaphoresSignalNodeSetParams,
15945 * ::cuGraphExecExternalSemaphoresWaitNodeSetParams,
15946 * ::cuGraphExecUpdate,
15947 * ::cuGraphInstantiate
15948 */
15949CUresult CUDAAPI cuGraphExecHostNodeSetParams(CUgraphExec hGraphExec, CUgraphNode hNode, const CUDA_HOST_NODE_PARAMS *nodeParams);
15950
15951/**
15952 * \brief Updates node parameters in the child graph node in the given graphExec.
15953 *
15954 * Updates the work represented by \p hNode in \p hGraphExec as though the nodes contained
15955 * in \p hNode's graph had the parameters contained in \p childGraph's nodes at instantiation.
15956 * \p hNode must remain in the graph which was used to instantiate \p hGraphExec.
15957 * Changed edges to and from \p hNode are ignored.
15958 *
15959 * The modifications only affect future launches of \p hGraphExec. Already enqueued
15960 * or running launches of \p hGraphExec are not affected by this call. \p hNode is also
15961 * not modified by this call.
15962 *
15963 * The topology of \p childGraph, as well as the node insertion order, must match that
15964 * of the graph contained in \p hNode. See ::cuGraphExecUpdate() for a list of restrictions
15965 * on what can be updated in an instantiated graph. The update is recursive, so child graph
15966 * nodes contained within the top level child graph will also be updated.
15967 *
15968 * \param hGraphExec - The executable graph in which to set the specified node
15969 * \param hNode - Host node from the graph which was used to instantiate graphExec
15970 * \param childGraph - The graph supplying the updated parameters
15971 *
15972 * \return
15973 * ::CUDA_SUCCESS,
15974 * ::CUDA_ERROR_INVALID_VALUE,
15975 * \note_graph_thread_safety
15976 * \notefnerr
15977 *
15978 * \sa
15979 * ::cuGraphAddChildGraphNode,
15980 * ::cuGraphChildGraphNodeGetGraph,
15981 * ::cuGraphExecKernelNodeSetParams,
15982 * ::cuGraphExecMemcpyNodeSetParams,
15983 * ::cuGraphExecMemsetNodeSetParams,
15984 * ::cuGraphExecHostNodeSetParams,
15985 * ::cuGraphExecEventRecordNodeSetEvent,
15986 * ::cuGraphExecEventWaitNodeSetEvent,
15987 * ::cuGraphExecExternalSemaphoresSignalNodeSetParams,
15988 * ::cuGraphExecExternalSemaphoresWaitNodeSetParams,
15989 * ::cuGraphExecUpdate,
15990 * ::cuGraphInstantiate
15991 */
15992CUresult CUDAAPI cuGraphExecChildGraphNodeSetParams(CUgraphExec hGraphExec, CUgraphNode hNode, CUgraph childGraph);
15993
15994/**
15995 * \brief Sets the event for an event record node in the given graphExec
15996 *
15997 * Sets the event of an event record node in an executable graph \p hGraphExec.
15998 * The node is identified by the corresponding node \p hNode in the
15999 * non-executable graph, from which the executable graph was instantiated.
16000 *
16001 * The modifications only affect future launches of \p hGraphExec. Already
16002 * enqueued or running launches of \p hGraphExec are not affected by this call.
16003 * \p hNode is also not modified by this call.
16004 *
16005 * \param hGraphExec - The executable graph in which to set the specified node
16006 * \param hNode - event record node from the graph from which graphExec was instantiated
16007 * \param event - Updated event to use
16008 *
16009 * \return
16010 * ::CUDA_SUCCESS,
16011 * ::CUDA_ERROR_INVALID_VALUE,
16012 * \note_graph_thread_safety
16013 * \notefnerr
16014 *
16015 * \sa
16016 * ::cuGraphAddEventRecordNode,
16017 * ::cuGraphEventRecordNodeGetEvent,
16018 * ::cuGraphEventWaitNodeSetEvent,
16019 * ::cuEventRecordWithFlags,
16020 * ::cuStreamWaitEvent,
16021 * ::cuGraphExecKernelNodeSetParams,
16022 * ::cuGraphExecMemcpyNodeSetParams,
16023 * ::cuGraphExecMemsetNodeSetParams,
16024 * ::cuGraphExecHostNodeSetParams,
16025 * ::cuGraphExecChildGraphNodeSetParams,
16026 * ::cuGraphExecEventWaitNodeSetEvent,
16027 * ::cuGraphExecExternalSemaphoresSignalNodeSetParams,
16028 * ::cuGraphExecExternalSemaphoresWaitNodeSetParams,
16029 * ::cuGraphExecUpdate,
16030 * ::cuGraphInstantiate
16031 */
16032CUresult CUDAAPI cuGraphExecEventRecordNodeSetEvent(CUgraphExec hGraphExec, CUgraphNode hNode, CUevent event);
16033
16034/**
16035 * \brief Sets the event for an event wait node in the given graphExec
16036 *
16037 * Sets the event of an event wait node in an executable graph \p hGraphExec.
16038 * The node is identified by the corresponding node \p hNode in the
16039 * non-executable graph, from which the executable graph was instantiated.
16040 *
16041 * The modifications only affect future launches of \p hGraphExec. Already
16042 * enqueued or running launches of \p hGraphExec are not affected by this call.
16043 * \p hNode is also not modified by this call.
16044 *
16045 * \param hGraphExec - The executable graph in which to set the specified node
16046 * \param hNode - event wait node from the graph from which graphExec was instantiated
16047 * \param event - Updated event to use
16048 *
16049 * \return
16050 * ::CUDA_SUCCESS,
16051 * ::CUDA_ERROR_INVALID_VALUE,
16052 * \note_graph_thread_safety
16053 * \notefnerr
16054 *
16055 * \sa
16056 * ::cuGraphAddEventWaitNode,
16057 * ::cuGraphEventWaitNodeGetEvent,
16058 * ::cuGraphEventRecordNodeSetEvent,
16059 * ::cuEventRecordWithFlags,
16060 * ::cuStreamWaitEvent,
16061 * ::cuGraphExecKernelNodeSetParams,
16062 * ::cuGraphExecMemcpyNodeSetParams,
16063 * ::cuGraphExecMemsetNodeSetParams,
16064 * ::cuGraphExecHostNodeSetParams,
16065 * ::cuGraphExecChildGraphNodeSetParams,
16066 * ::cuGraphExecEventRecordNodeSetEvent,
16067 * ::cuGraphExecExternalSemaphoresSignalNodeSetParams,
16068 * ::cuGraphExecExternalSemaphoresWaitNodeSetParams,
16069 * ::cuGraphExecUpdate,
16070 * ::cuGraphInstantiate
16071 */
16072CUresult CUDAAPI cuGraphExecEventWaitNodeSetEvent(CUgraphExec hGraphExec, CUgraphNode hNode, CUevent event);
16073
16074/**
16075 * \brief Sets the parameters for an external semaphore signal node in the given graphExec
16076 *
16077 * Sets the parameters of an external semaphore signal node in an executable graph \p hGraphExec.
16078 * The node is identified by the corresponding node \p hNode in the
16079 * non-executable graph, from which the executable graph was instantiated.
16080 *
16081 * \p hNode must not have been removed from the original graph.
16082 *
16083 * The modifications only affect future launches of \p hGraphExec. Already
16084 * enqueued or running launches of \p hGraphExec are not affected by this call.
16085 * \p hNode is also not modified by this call.
16086 *
16087 * Changing \p nodeParams->numExtSems is not supported.
16088 *
16089 * \param hGraphExec - The executable graph in which to set the specified node
16090 * \param hNode - semaphore signal node from the graph from which graphExec was instantiated
16091 * \param nodeParams - Updated Parameters to set
16092 *
16093 * \return
16094 * ::CUDA_SUCCESS,
16095 * ::CUDA_ERROR_INVALID_VALUE,
16096 * \note_graph_thread_safety
16097 * \notefnerr
16098 *
16099 * \sa
16100 * ::cuGraphAddExternalSemaphoresSignalNode,
16101 * ::cuImportExternalSemaphore,
16102 * ::cuSignalExternalSemaphoresAsync,
16103 * ::cuWaitExternalSemaphoresAsync,
16104 * ::cuGraphExecKernelNodeSetParams,
16105 * ::cuGraphExecMemcpyNodeSetParams,
16106 * ::cuGraphExecMemsetNodeSetParams,
16107 * ::cuGraphExecHostNodeSetParams,
16108 * ::cuGraphExecChildGraphNodeSetParams,
16109 * ::cuGraphExecEventRecordNodeSetEvent,
16110 * ::cuGraphExecEventWaitNodeSetEvent,
16111 * ::cuGraphExecExternalSemaphoresWaitNodeSetParams,
16112 * ::cuGraphExecUpdate,
16113 * ::cuGraphInstantiate
16114 */
16115CUresult CUDAAPI cuGraphExecExternalSemaphoresSignalNodeSetParams(CUgraphExec hGraphExec, CUgraphNode hNode, const CUDA_EXT_SEM_SIGNAL_NODE_PARAMS *nodeParams);
16116
16117/**
16118 * \brief Sets the parameters for an external semaphore wait node in the given graphExec
16119 *
16120 * Sets the parameters of an external semaphore wait node in an executable graph \p hGraphExec.
16121 * The node is identified by the corresponding node \p hNode in the
16122 * non-executable graph, from which the executable graph was instantiated.
16123 *
16124 * \p hNode must not have been removed from the original graph.
16125 *
16126 * The modifications only affect future launches of \p hGraphExec. Already
16127 * enqueued or running launches of \p hGraphExec are not affected by this call.
16128 * \p hNode is also not modified by this call.
16129 *
16130 * Changing \p nodeParams->numExtSems is not supported.
16131 *
16132 * \param hGraphExec - The executable graph in which to set the specified node
16133 * \param hNode - semaphore wait node from the graph from which graphExec was instantiated
16134 * \param nodeParams - Updated Parameters to set
16135 *
16136 * \return
16137 * ::CUDA_SUCCESS,
16138 * ::CUDA_ERROR_INVALID_VALUE,
16139 * \note_graph_thread_safety
16140 * \notefnerr
16141 *
16142 * \sa
16143 * ::cuGraphAddExternalSemaphoresWaitNode,
16144 * ::cuImportExternalSemaphore,
16145 * ::cuSignalExternalSemaphoresAsync,
16146 * ::cuWaitExternalSemaphoresAsync,
16147 * ::cuGraphExecKernelNodeSetParams,
16148 * ::cuGraphExecMemcpyNodeSetParams,
16149 * ::cuGraphExecMemsetNodeSetParams,
16150 * ::cuGraphExecHostNodeSetParams,
16151 * ::cuGraphExecChildGraphNodeSetParams,
16152 * ::cuGraphExecEventRecordNodeSetEvent,
16153 * ::cuGraphExecEventWaitNodeSetEvent,
16154 * ::cuGraphExecExternalSemaphoresSignalNodeSetParams,
16155 * ::cuGraphExecUpdate,
16156 * ::cuGraphInstantiate
16157 */
16158CUresult CUDAAPI cuGraphExecExternalSemaphoresWaitNodeSetParams(CUgraphExec hGraphExec, CUgraphNode hNode, const CUDA_EXT_SEM_WAIT_NODE_PARAMS *nodeParams);
16159
16160/**
16161 * \brief Uploads an executable graph in a stream
16162 *
16163 * Uploads \p hGraphExec to the device in \p hStream without executing it. Uploads of
16164 * the same \p hGraphExec will be serialized. Each upload is ordered behind both any
16165 * previous work in \p hStream and any previous launches of \p hGraphExec.
16166 * Uses memory cached by \p stream to back the allocations owned by \p hGraphExec.
16167 *
16168 * \param hGraphExec - Executable graph to upload
16169 * \param hStream - Stream in which to upload the graph
16170 *
16171 * \return
16172 * ::CUDA_SUCCESS,
16173 * ::CUDA_ERROR_DEINITIALIZED,
16174 * ::CUDA_ERROR_NOT_INITIALIZED,
16175 * ::CUDA_ERROR_INVALID_VALUE
16176 * \note_graph_thread_safety
16177 * \notefnerr
16178 *
16179 * \sa
16180 * ::cuGraphInstantiate,
16181 * ::cuGraphLaunch,
16182 * ::cuGraphExecDestroy
16183 */
16184CUresult CUDAAPI cuGraphUpload(CUgraphExec hGraphExec, CUstream hStream);
16185
16186/**
16187 * \brief Launches an executable graph in a stream
16188 *
16189 * Executes \p hGraphExec in \p hStream. Only one instance of \p hGraphExec may be executing
16190 * at a time. Each launch is ordered behind both any previous work in \p hStream
16191 * and any previous launches of \p hGraphExec. To execute a graph concurrently, it must be
16192 * instantiated multiple times into multiple executable graphs.
16193 *
16194 * If any allocations created by \p hGraphExec remain unfreed (from a previous launch) and
16195 * \p hGraphExec was not instantiated with ::CUDA_GRAPH_INSTANTIATE_FLAG_AUTO_FREE_ON_LAUNCH,
16196 * the launch will fail with ::CUDA_ERROR_INVALID_VALUE.
16197 *
16198 * \param hGraphExec - Executable graph to launch
16199 * \param hStream - Stream in which to launch the graph
16200 *
16201 * \return
16202 * ::CUDA_SUCCESS,
16203 * ::CUDA_ERROR_DEINITIALIZED,
16204 * ::CUDA_ERROR_NOT_INITIALIZED,
16205 * ::CUDA_ERROR_INVALID_VALUE
16206 * \note_graph_thread_safety
16207 * \notefnerr
16208 *
16209 * \sa
16210 * ::cuGraphInstantiate,
16211 * ::cuGraphUpload,
16212 * ::cuGraphExecDestroy
16213 */
16214CUresult CUDAAPI cuGraphLaunch(CUgraphExec hGraphExec, CUstream hStream);
16215
16216/**
16217 * \brief Destroys an executable graph
16218 *
16219 * Destroys the executable graph specified by \p hGraphExec, as well
16220 * as all of its executable nodes. If the executable graph is
16221 * in-flight, it will not be terminated, but rather freed
16222 * asynchronously on completion.
16223 *
16224 * \param hGraphExec - Executable graph to destroy
16225 *
16226 * \return
16227 * ::CUDA_SUCCESS,
16228 * ::CUDA_ERROR_DEINITIALIZED,
16229 * ::CUDA_ERROR_NOT_INITIALIZED,
16230 * ::CUDA_ERROR_INVALID_VALUE
16231 * \note_graph_thread_safety
16232 * \notefnerr
16233 *
16234 * \sa
16235 * ::cuGraphInstantiate,
16236 * ::cuGraphUpload,
16237 * ::cuGraphLaunch
16238 */
16239CUresult CUDAAPI cuGraphExecDestroy(CUgraphExec hGraphExec);
16240
16241/**
16242 * \brief Destroys a graph
16243 *
16244 * Destroys the graph specified by \p hGraph, as well as all of its nodes.
16245 *
16246 * \param hGraph - Graph to destroy
16247 *
16248 * \return
16249 * ::CUDA_SUCCESS,
16250 * ::CUDA_ERROR_DEINITIALIZED,
16251 * ::CUDA_ERROR_NOT_INITIALIZED,
16252 * ::CUDA_ERROR_INVALID_VALUE
16253 * \note_graph_thread_safety
16254 * \notefnerr
16255 *
16256 * \sa
16257 * ::cuGraphCreate
16258 */
16259CUresult CUDAAPI cuGraphDestroy(CUgraph hGraph);
16260
16261/**
16262 * \brief Check whether an executable graph can be updated with a graph and perform the update if possible
16263 *
16264 * Updates the node parameters in the instantiated graph specified by \p hGraphExec with the
16265 * node parameters in a topologically identical graph specified by \p hGraph.
16266 *
16267 * Limitations:
16268 *
16269 * - Kernel nodes:
16270 * - The owning context of the function cannot change.
16271 * - A node whose function originally did not use CUDA dynamic parallelism cannot be updated
16272 * to a function which uses CDP
16273 * - Memset and memcpy nodes:
16274 * - The CUDA device(s) to which the operand(s) was allocated/mapped cannot change.
16275 * - The source/destination memory must be allocated from the same contexts as the original
16276 * source/destination memory.
16277 * - Only 1D memsets can be changed.
16278 * - Additional memcpy node restrictions:
16279 * - Changing either the source or destination memory type(i.e. CU_MEMORYTYPE_DEVICE,
16280 * CU_MEMORYTYPE_ARRAY, etc.) is not supported.
16281 * - External semaphore wait nodes and record nodes:
16282 * - Changing the number of semaphores is not supported.
16283 *
16284 * Note: The API may add further restrictions in future releases. The return code should always be checked.
16285 *
16286 * cuGraphExecUpdate sets \p updateResult_out to CU_GRAPH_EXEC_UPDATE_ERROR_TOPOLOGY_CHANGED under
16287 * the following conditions:
16288 *
16289 * - The count of nodes directly in \p hGraphExec and \p hGraph differ, in which case \p hErrorNode_out
16290 * is NULL.
16291 * - A node is deleted in \p hGraph but not not its pair from \p hGraphExec, in which case \p hErrorNode_out
16292 * is NULL.
16293 * - A node is deleted in \p hGraphExec but not its pair from \p hGraph, in which case \p hErrorNode_out is
16294 * the pairless node from \p hGraph.
16295 * - The dependent nodes of a pair differ, in which case \p hErrorNode_out is the node from \p hGraph.
16296 *
16297 * cuGraphExecUpdate sets \p updateResult_out to:
16298 * - CU_GRAPH_EXEC_UPDATE_ERROR if passed an invalid value.
16299 * - CU_GRAPH_EXEC_UPDATE_ERROR_TOPOLOGY_CHANGED if the graph topology changed
16300 * - CU_GRAPH_EXEC_UPDATE_ERROR_NODE_TYPE_CHANGED if the type of a node changed, in which case
16301 * \p hErrorNode_out is set to the node from \p hGraph.
16302 * - CU_GRAPH_EXEC_UPDATE_ERROR_UNSUPPORTED_FUNCTION_CHANGE if the function changed in an unsupported
16303 * way(see note above), in which case \p hErrorNode_out is set to the node from \p hGraph
16304 * - CU_GRAPH_EXEC_UPDATE_ERROR_PARAMETERS_CHANGED if any parameters to a node changed in a way
16305 * that is not supported, in which case \p hErrorNode_out is set to the node from \p hGraph.
16306 * - CU_GRAPH_EXEC_UPDATE_ERROR_NOT_SUPPORTED if something about a node is unsupported, like
16307 * the node's type or configuration, in which case \p hErrorNode_out is set to the node from \p hGraph
16308 *
16309 * If \p updateResult_out isn't set in one of the situations described above, the update check passes
16310 * and cuGraphExecUpdate updates \p hGraphExec to match the contents of \p hGraph. If an error happens
16311 * during the update, \p updateResult_out will be set to CU_GRAPH_EXEC_UPDATE_ERROR; otherwise,
16312 * \p updateResult_out is set to CU_GRAPH_EXEC_UPDATE_SUCCESS.
16313 *
16314 * cuGraphExecUpdate returns CUDA_SUCCESS when the updated was performed successfully. It returns
16315 * CUDA_ERROR_GRAPH_EXEC_UPDATE_FAILURE if the graph update was not performed because it included
16316 * changes which violated constraints specific to instantiated graph update.
16317 *
16318 * \param hGraphExec The instantiated graph to be updated
16319 * \param hGraph The graph containing the updated parameters
16320 * \param hErrorNode_out The node which caused the permissibility check to forbid the update, if any
16321 * \param updateResult_out Whether the graph update was permitted. If was forbidden, the reason why
16322 *
16323 * \return
16324 * ::CUDA_SUCCESS,
16325 * ::CUDA_ERROR_GRAPH_EXEC_UPDATE_FAILURE,
16326 * \note_graph_thread_safety
16327 * \notefnerr
16328 *
16329 * \sa
16330 * ::cuGraphInstantiate,
16331 */
16332CUresult CUDAAPI cuGraphExecUpdate(CUgraphExec hGraphExec, CUgraph hGraph, CUgraphNode *hErrorNode_out, CUgraphExecUpdateResult *updateResult_out);
16333
16334/**
16335 * \brief Copies attributes from source node to destination node.
16336 *
16337 * Copies attributes from source node \p src to destination node \p dst.
16338 * Both node must have the same context.
16339 *
16340 * \param[out] dst Destination node
16341 * \param[in] src Source node
16342 * For list of attributes see ::CUkernelNodeAttrID
16343 *
16344 * \return
16345 * ::CUDA_SUCCESS,
16346 * ::CUDA_ERROR_INVALID_VALUE
16347 * \notefnerr
16348 *
16349 * \sa
16350 * ::CUaccessPolicyWindow
16351 */
16352CUresult CUDAAPI cuGraphKernelNodeCopyAttributes(CUgraphNode dst, CUgraphNode src);
16353
16354/**
16355 * \brief Queries node attribute.
16356 *
16357 * Queries attribute \p attr from node \p hNode and stores it in corresponding
16358 * member of \p value_out.
16359 *
16360 * \param[in] hNode
16361 * \param[in] attr
16362 * \param[out] value_out
16363 *
16364 * \return
16365 * ::CUDA_SUCCESS,
16366 * ::CUDA_ERROR_INVALID_VALUE,
16367 * ::CUDA_ERROR_INVALID_HANDLE
16368 * \notefnerr
16369 *
16370 * \sa
16371 * ::CUaccessPolicyWindow
16372 */
16373CUresult CUDAAPI cuGraphKernelNodeGetAttribute(CUgraphNode hNode, CUkernelNodeAttrID attr,
16374 CUkernelNodeAttrValue *value_out);
16375
16376/**
16377 * \brief Sets node attribute.
16378 *
16379 * Sets attribute \p attr on node \p hNode from corresponding attribute of
16380 * \p value.
16381 *
16382 * \param[out] hNode
16383 * \param[in] attr
16384 * \param[out] value
16385 *
16386 * \return
16387 * ::CUDA_SUCCESS,
16388 * ::CUDA_ERROR_INVALID_VALUE,
16389 * ::CUDA_ERROR_INVALID_HANDLE
16390 * \notefnerr
16391 *
16392 * \sa
16393 * ::CUaccessPolicyWindow
16394 */
16395CUresult CUDAAPI cuGraphKernelNodeSetAttribute(CUgraphNode hNode, CUkernelNodeAttrID attr,
16396 const CUkernelNodeAttrValue *value);
16397
16398/**
16399 * \brief Write a DOT file describing graph structure
16400 *
16401 * Using the provided \p hGraph, write to \p path a DOT formatted description of the graph.
16402 * By default this includes the graph topology, node types, node id, kernel names and memcpy direction.
16403 * \p flags can be specified to write more detailed information about each node type such as
16404 * parameter values, kernel attributes, node and function handles.
16405 *
16406 * \param hGraph - The graph to create a DOT file from
16407 * \param path - The path to write the DOT file to
16408 * \param flags - Flags from CUgraphDebugDot_flags for specifying which additional node information to write
16409 *
16410 * \return
16411 * ::CUDA_SUCCESS,
16412 * ::CUDA_ERROR_INVALID_VALUE,
16413 * ::CUDA_ERROR_OPERATING_SYSTEM
16414 */
16415CUresult CUDAAPI cuGraphDebugDotPrint(CUgraph hGraph, const char *path, unsigned int flags);
16416
16417/**
16418 * \brief Create a user object
16419 *
16420 * Create a user object with the specified destructor callback and initial reference count. The
16421 * initial references are owned by the caller.
16422 *
16423 * Destructor callbacks cannot make CUDA API calls and should avoid blocking behavior, as they
16424 * are executed by a shared internal thread. Another thread may be signaled to perform such
16425 * actions, if it does not block forward progress of tasks scheduled through CUDA.
16426 *
16427 * See CUDA User Objects in the CUDA C++ Programming Guide for more information on user objects.
16428 *
16429 * \param object_out - Location to return the user object handle
16430 * \param ptr - The pointer to pass to the destroy function
16431 * \param destroy - Callback to free the user object when it is no longer in use
16432 * \param initialRefcount - The initial refcount to create the object with, typically 1. The
16433 * initial references are owned by the calling thread.
16434 * \param flags - Currently it is required to pass ::CU_USER_OBJECT_NO_DESTRUCTOR_SYNC,
16435 * which is the only defined flag. This indicates that the destroy
16436 * callback cannot be waited on by any CUDA API. Users requiring
16437 * synchronization of the callback should signal its completion
16438 * manually.
16439 *
16440 * \return
16441 * ::CUDA_SUCCESS,
16442 * ::CUDA_ERROR_INVALID_VALUE
16443 *
16444 * \sa
16445 * ::cuUserObjectRetain,
16446 * ::cuUserObjectRelease,
16447 * ::cuGraphRetainUserObject,
16448 * ::cuGraphReleaseUserObject,
16449 * ::cuGraphCreate
16450 */
16451CUresult CUDAAPI cuUserObjectCreate(CUuserObject *object_out, void *ptr, CUhostFn destroy,
16452 unsigned int initialRefcount, unsigned int flags);
16453
16454/**
16455 * \brief Retain a reference to a user object
16456 *
16457 * Retains new references to a user object. The new references are owned by the caller.
16458 *
16459 * See CUDA User Objects in the CUDA C++ Programming Guide for more information on user objects.
16460 *
16461 * \param object - The object to retain
16462 * \param count - The number of references to retain, typically 1. Must be nonzero
16463 * and not larger than INT_MAX.
16464 *
16465 * \return
16466 * ::CUDA_SUCCESS,
16467 * ::CUDA_ERROR_INVALID_VALUE
16468 *
16469 * \sa
16470 * ::cuUserObjectCreate,
16471 * ::cuUserObjectRelease,
16472 * ::cuGraphRetainUserObject,
16473 * ::cuGraphReleaseUserObject,
16474 * ::cuGraphCreate
16475 */
16476CUresult CUDAAPI cuUserObjectRetain(CUuserObject object, unsigned int count);
16477
16478/**
16479 * \brief Release a reference to a user object
16480 *
16481 * Releases user object references owned by the caller. The object's destructor is invoked if
16482 * the reference count reaches zero.
16483 *
16484 * It is undefined behavior to release references not owned by the caller, or to use a user
16485 * object handle after all references are released.
16486 *
16487 * See CUDA User Objects in the CUDA C++ Programming Guide for more information on user objects.
16488 *
16489 * \param object - The object to release
16490 * \param count - The number of references to release, typically 1. Must be nonzero
16491 * and not larger than INT_MAX.
16492 *
16493 * \return
16494 * ::CUDA_SUCCESS,
16495 * ::CUDA_ERROR_INVALID_VALUE
16496 *
16497 * \sa
16498 * ::cuUserObjectCreate,
16499 * ::cuUserObjectRetain,
16500 * ::cuGraphRetainUserObject,
16501 * ::cuGraphReleaseUserObject,
16502 * ::cuGraphCreate
16503 */
16504CUresult CUDAAPI cuUserObjectRelease(CUuserObject object, unsigned int count);
16505
16506/**
16507 * \brief Retain a reference to a user object from a graph
16508 *
16509 * Creates or moves user object references that will be owned by a CUDA graph.
16510 *
16511 * See CUDA User Objects in the CUDA C++ Programming Guide for more information on user objects.
16512 *
16513 * \param graph - The graph to associate the reference with
16514 * \param object - The user object to retain a reference for
16515 * \param count - The number of references to add to the graph, typically 1. Must be
16516 * nonzero and not larger than INT_MAX.
16517 * \param flags - The optional flag ::CU_GRAPH_USER_OBJECT_MOVE transfers references
16518 * from the calling thread, rather than create new references. Pass 0
16519 * to create new references.
16520 *
16521 * \return
16522 * ::CUDA_SUCCESS,
16523 * ::CUDA_ERROR_INVALID_VALUE
16524 *
16525 * \sa
16526 * ::cuUserObjectCreate,
16527 * ::cuUserObjectRetain,
16528 * ::cuUserObjectRelease,
16529 * ::cuGraphReleaseUserObject,
16530 * ::cuGraphCreate
16531 */
16532CUresult CUDAAPI cuGraphRetainUserObject(CUgraph graph, CUuserObject object, unsigned int count, unsigned int flags);
16533
16534/**
16535 * \brief Release a user object reference from a graph
16536 *
16537 * Releases user object references owned by a graph.
16538 *
16539 * See CUDA User Objects in the CUDA C++ Programming Guide for more information on user objects.
16540 *
16541 * \param graph - The graph that will release the reference
16542 * \param object - The user object to release a reference for
16543 * \param count - The number of references to release, typically 1. Must be nonzero
16544 * and not larger than INT_MAX.
16545 *
16546 * \return
16547 * ::CUDA_SUCCESS,
16548 * ::CUDA_ERROR_INVALID_VALUE
16549 *
16550 * \sa
16551 * ::cuUserObjectCreate,
16552 * ::cuUserObjectRetain,
16553 * ::cuUserObjectRelease,
16554 * ::cuGraphRetainUserObject,
16555 * ::cuGraphCreate
16556 */
16557CUresult CUDAAPI cuGraphReleaseUserObject(CUgraph graph, CUuserObject object, unsigned int count);
16558
16559/** @} */ /* END CUDA_GRAPH */
16560
16561/**
16562 * \defgroup CUDA_OCCUPANCY Occupancy
16563 *
16564 * ___MANBRIEF___ occupancy calculation functions of the low-level CUDA driver
16565 * API (___CURRENT_FILE___) ___ENDMANBRIEF___
16566 *
16567 * This section describes the occupancy calculation functions of the low-level CUDA
16568 * driver application programming interface.
16569 *
16570 * @{
16571 */
16572
16573/**
16574 * \brief Returns occupancy of a function
16575 *
16576 * Returns in \p *numBlocks the number of the maximum active blocks per
16577 * streaming multiprocessor.
16578 *
16579 * \param numBlocks - Returned occupancy
16580 * \param func - Kernel for which occupancy is calculated
16581 * \param blockSize - Block size the kernel is intended to be launched with
16582 * \param dynamicSMemSize - Per-block dynamic shared memory usage intended, in bytes
16583 *
16584 * \return
16585 * ::CUDA_SUCCESS,
16586 * ::CUDA_ERROR_DEINITIALIZED,
16587 * ::CUDA_ERROR_NOT_INITIALIZED,
16588 * ::CUDA_ERROR_INVALID_CONTEXT,
16589 * ::CUDA_ERROR_INVALID_VALUE,
16590 * ::CUDA_ERROR_UNKNOWN
16591 * \notefnerr
16592 *
16593 * \sa
16594 * ::cudaOccupancyMaxActiveBlocksPerMultiprocessor
16595 */
16596CUresult CUDAAPI cuOccupancyMaxActiveBlocksPerMultiprocessor(int *numBlocks, CUfunction func, int blockSize, size_t dynamicSMemSize);
16597
16598/**
16599 * \brief Returns occupancy of a function
16600 *
16601 * Returns in \p *numBlocks the number of the maximum active blocks per
16602 * streaming multiprocessor.
16603 *
16604 * The \p Flags parameter controls how special cases are handled. The
16605 * valid flags are:
16606 *
16607 * - ::CU_OCCUPANCY_DEFAULT, which maintains the default behavior as
16608 * ::cuOccupancyMaxActiveBlocksPerMultiprocessor;
16609 *
16610 * - ::CU_OCCUPANCY_DISABLE_CACHING_OVERRIDE, which suppresses the
16611 * default behavior on platform where global caching affects
16612 * occupancy. On such platforms, if caching is enabled, but
16613 * per-block SM resource usage would result in zero occupancy, the
16614 * occupancy calculator will calculate the occupancy as if caching
16615 * is disabled. Setting ::CU_OCCUPANCY_DISABLE_CACHING_OVERRIDE makes
16616 * the occupancy calculator to return 0 in such cases. More information
16617 * can be found about this feature in the "Unified L1/Texture Cache"
16618 * section of the Maxwell tuning guide.
16619 *
16620 * \param numBlocks - Returned occupancy
16621 * \param func - Kernel for which occupancy is calculated
16622 * \param blockSize - Block size the kernel is intended to be launched with
16623 * \param dynamicSMemSize - Per-block dynamic shared memory usage intended, in bytes
16624 * \param flags - Requested behavior for the occupancy calculator
16625 *
16626 * \return
16627 * ::CUDA_SUCCESS,
16628 * ::CUDA_ERROR_DEINITIALIZED,
16629 * ::CUDA_ERROR_NOT_INITIALIZED,
16630 * ::CUDA_ERROR_INVALID_CONTEXT,
16631 * ::CUDA_ERROR_INVALID_VALUE,
16632 * ::CUDA_ERROR_UNKNOWN
16633 * \notefnerr
16634 *
16635 * \sa
16636 * ::cudaOccupancyMaxActiveBlocksPerMultiprocessorWithFlags
16637 */
16638CUresult CUDAAPI cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags(int *numBlocks, CUfunction func, int blockSize, size_t dynamicSMemSize, unsigned int flags);
16639
16640/**
16641 * \brief Suggest a launch configuration with reasonable occupancy
16642 *
16643 * Returns in \p *blockSize a reasonable block size that can achieve
16644 * the maximum occupancy (or, the maximum number of active warps with
16645 * the fewest blocks per multiprocessor), and in \p *minGridSize the
16646 * minimum grid size to achieve the maximum occupancy.
16647 *
16648 * If \p blockSizeLimit is 0, the configurator will use the maximum
16649 * block size permitted by the device / function instead.
16650 *
16651 * If per-block dynamic shared memory allocation is not needed, the
16652 * user should leave both \p blockSizeToDynamicSMemSize and \p
16653 * dynamicSMemSize as 0.
16654 *
16655 * If per-block dynamic shared memory allocation is needed, then if
16656 * the dynamic shared memory size is constant regardless of block
16657 * size, the size should be passed through \p dynamicSMemSize, and \p
16658 * blockSizeToDynamicSMemSize should be NULL.
16659 *
16660 * Otherwise, if the per-block dynamic shared memory size varies with
16661 * different block sizes, the user needs to provide a unary function
16662 * through \p blockSizeToDynamicSMemSize that computes the dynamic
16663 * shared memory needed by \p func for any given block size. \p
16664 * dynamicSMemSize is ignored. An example signature is:
16665 *
16666 * \code
16667 * // Take block size, returns dynamic shared memory needed
16668 * size_t blockToSmem(int blockSize);
16669 * \endcode
16670 *
16671 * \param minGridSize - Returned minimum grid size needed to achieve the maximum occupancy
16672 * \param blockSize - Returned maximum block size that can achieve the maximum occupancy
16673 * \param func - Kernel for which launch configuration is calculated
16674 * \param blockSizeToDynamicSMemSize - A function that calculates how much per-block dynamic shared memory \p func uses based on the block size
16675 * \param dynamicSMemSize - Dynamic shared memory usage intended, in bytes
16676 * \param blockSizeLimit - The maximum block size \p func is designed to handle
16677 *
16678 * \return
16679 * ::CUDA_SUCCESS,
16680 * ::CUDA_ERROR_DEINITIALIZED,
16681 * ::CUDA_ERROR_NOT_INITIALIZED,
16682 * ::CUDA_ERROR_INVALID_CONTEXT,
16683 * ::CUDA_ERROR_INVALID_VALUE,
16684 * ::CUDA_ERROR_UNKNOWN
16685 * \notefnerr
16686 *
16687 * \sa
16688 * ::cudaOccupancyMaxPotentialBlockSize
16689 */
16690CUresult CUDAAPI cuOccupancyMaxPotentialBlockSize(int *minGridSize, int *blockSize, CUfunction func, CUoccupancyB2DSize blockSizeToDynamicSMemSize, size_t dynamicSMemSize, int blockSizeLimit);
16691
16692/**
16693 * \brief Suggest a launch configuration with reasonable occupancy
16694 *
16695 * An extended version of ::cuOccupancyMaxPotentialBlockSize. In
16696 * addition to arguments passed to ::cuOccupancyMaxPotentialBlockSize,
16697 * ::cuOccupancyMaxPotentialBlockSizeWithFlags also takes a \p Flags
16698 * parameter.
16699 *
16700 * The \p Flags parameter controls how special cases are handled. The
16701 * valid flags are:
16702 *
16703 * - ::CU_OCCUPANCY_DEFAULT, which maintains the default behavior as
16704 * ::cuOccupancyMaxPotentialBlockSize;
16705 *
16706 * - ::CU_OCCUPANCY_DISABLE_CACHING_OVERRIDE, which suppresses the
16707 * default behavior on platform where global caching affects
16708 * occupancy. On such platforms, the launch configurations that
16709 * produces maximal occupancy might not support global
16710 * caching. Setting ::CU_OCCUPANCY_DISABLE_CACHING_OVERRIDE
16711 * guarantees that the the produced launch configuration is global
16712 * caching compatible at a potential cost of occupancy. More information
16713 * can be found about this feature in the "Unified L1/Texture Cache"
16714 * section of the Maxwell tuning guide.
16715 *
16716 * \param minGridSize - Returned minimum grid size needed to achieve the maximum occupancy
16717 * \param blockSize - Returned maximum block size that can achieve the maximum occupancy
16718 * \param func - Kernel for which launch configuration is calculated
16719 * \param blockSizeToDynamicSMemSize - A function that calculates how much per-block dynamic shared memory \p func uses based on the block size
16720 * \param dynamicSMemSize - Dynamic shared memory usage intended, in bytes
16721 * \param blockSizeLimit - The maximum block size \p func is designed to handle
16722 * \param flags - Options
16723 *
16724 * \return
16725 * ::CUDA_SUCCESS,
16726 * ::CUDA_ERROR_DEINITIALIZED,
16727 * ::CUDA_ERROR_NOT_INITIALIZED,
16728 * ::CUDA_ERROR_INVALID_CONTEXT,
16729 * ::CUDA_ERROR_INVALID_VALUE,
16730 * ::CUDA_ERROR_UNKNOWN
16731 * \notefnerr
16732 *
16733 * \sa
16734 * ::cudaOccupancyMaxPotentialBlockSizeWithFlags
16735 */
16736CUresult CUDAAPI cuOccupancyMaxPotentialBlockSizeWithFlags(int *minGridSize, int *blockSize, CUfunction func, CUoccupancyB2DSize blockSizeToDynamicSMemSize, size_t dynamicSMemSize, int blockSizeLimit, unsigned int flags);
16737
16738/**
16739 * \brief Returns dynamic shared memory available per block when launching \p numBlocks blocks on SM
16740 *
16741 * Returns in \p *dynamicSmemSize the maximum size of dynamic shared memory to allow \p numBlocks blocks per SM.
16742 *
16743 * \param dynamicSmemSize - Returned maximum dynamic shared memory
16744 * \param func - Kernel function for which occupancy is calculated
16745 * \param numBlocks - Number of blocks to fit on SM
16746 * \param blockSize - Size of the blocks
16747 *
16748 * \return
16749 * ::CUDA_SUCCESS,
16750 * ::CUDA_ERROR_DEINITIALIZED,
16751 * ::CUDA_ERROR_NOT_INITIALIZED,
16752 * ::CUDA_ERROR_INVALID_CONTEXT,
16753 * ::CUDA_ERROR_INVALID_VALUE,
16754 * ::CUDA_ERROR_UNKNOWN
16755 * \notefnerr
16756 *
16757 * \sa
16758 */
16759CUresult CUDAAPI cuOccupancyAvailableDynamicSMemPerBlock(size_t *dynamicSmemSize, CUfunction func, int numBlocks, int blockSize);
16760
16761/** @} */ /* END CUDA_OCCUPANCY */
16762
16763/**
16764 * \defgroup CUDA_TEXREF_DEPRECATED Texture Reference Management [DEPRECATED]
16765 *
16766 * ___MANBRIEF___ deprecated texture reference management functions of the
16767 * low-level CUDA driver API (___CURRENT_FILE___) ___ENDMANBRIEF___
16768 *
16769 * This section describes the deprecated texture reference management
16770 * functions of the low-level CUDA driver application programming interface.
16771 *
16772 * @{
16773 */
16774
16775/**
16776 * \brief Binds an array as a texture reference
16777 *
16778 * \deprecated
16779 *
16780 * Binds the CUDA array \p hArray to the texture reference \p hTexRef. Any
16781 * previous address or CUDA array state associated with the texture reference
16782 * is superseded by this function. \p Flags must be set to
16783 * ::CU_TRSA_OVERRIDE_FORMAT. Any CUDA array previously bound to \p hTexRef is
16784 * unbound.
16785 *
16786 * \param hTexRef - Texture reference to bind
16787 * \param hArray - Array to bind
16788 * \param Flags - Options (must be ::CU_TRSA_OVERRIDE_FORMAT)
16789 *
16790 * \return
16791 * ::CUDA_SUCCESS,
16792 * ::CUDA_ERROR_DEINITIALIZED,
16793 * ::CUDA_ERROR_NOT_INITIALIZED,
16794 * ::CUDA_ERROR_INVALID_CONTEXT,
16795 * ::CUDA_ERROR_INVALID_VALUE
16796 *
16797 * \sa ::cuTexRefSetAddress,
16798 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode,
16799 * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags, ::cuTexRefSetFormat,
16800 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
16801 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat,
16802 * ::cudaBindTextureToArray
16803 */
16804__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefSetArray(CUtexref hTexRef, CUarray hArray, unsigned int Flags);
16805
16806/**
16807 * \brief Binds a mipmapped array to a texture reference
16808 *
16809 * \deprecated
16810 *
16811 * Binds the CUDA mipmapped array \p hMipmappedArray to the texture reference \p hTexRef.
16812 * Any previous address or CUDA array state associated with the texture reference
16813 * is superseded by this function. \p Flags must be set to ::CU_TRSA_OVERRIDE_FORMAT.
16814 * Any CUDA array previously bound to \p hTexRef is unbound.
16815 *
16816 * \param hTexRef - Texture reference to bind
16817 * \param hMipmappedArray - Mipmapped array to bind
16818 * \param Flags - Options (must be ::CU_TRSA_OVERRIDE_FORMAT)
16819 *
16820 * \return
16821 * ::CUDA_SUCCESS,
16822 * ::CUDA_ERROR_DEINITIALIZED,
16823 * ::CUDA_ERROR_NOT_INITIALIZED,
16824 * ::CUDA_ERROR_INVALID_CONTEXT,
16825 * ::CUDA_ERROR_INVALID_VALUE
16826 *
16827 * \sa ::cuTexRefSetAddress,
16828 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode,
16829 * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags, ::cuTexRefSetFormat,
16830 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
16831 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat,
16832 * ::cudaBindTextureToMipmappedArray
16833 */
16834__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefSetMipmappedArray(CUtexref hTexRef, CUmipmappedArray hMipmappedArray, unsigned int Flags);
16835
16836/**
16837 * \brief Binds an address as a texture reference
16838 *
16839 * \deprecated
16840 *
16841 * Binds a linear address range to the texture reference \p hTexRef. Any
16842 * previous address or CUDA array state associated with the texture reference
16843 * is superseded by this function. Any memory previously bound to \p hTexRef
16844 * is unbound.
16845 *
16846 * Since the hardware enforces an alignment requirement on texture base
16847 * addresses, ::cuTexRefSetAddress() passes back a byte offset in
16848 * \p *ByteOffset that must be applied to texture fetches in order to read from
16849 * the desired memory. This offset must be divided by the texel size and
16850 * passed to kernels that read from the texture so they can be applied to the
16851 * ::tex1Dfetch() function.
16852 *
16853 * If the device memory pointer was returned from ::cuMemAlloc(), the offset
16854 * is guaranteed to be 0 and NULL may be passed as the \p ByteOffset parameter.
16855 *
16856 * The total number of elements (or texels) in the linear address range
16857 * cannot exceed ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LINEAR_WIDTH.
16858 * The number of elements is computed as (\p bytes / bytesPerElement),
16859 * where bytesPerElement is determined from the data format and number of
16860 * components set using ::cuTexRefSetFormat().
16861 *
16862 * \param ByteOffset - Returned byte offset
16863 * \param hTexRef - Texture reference to bind
16864 * \param dptr - Device pointer to bind
16865 * \param bytes - Size of memory to bind in bytes
16866 *
16867 * \return
16868 * ::CUDA_SUCCESS,
16869 * ::CUDA_ERROR_DEINITIALIZED,
16870 * ::CUDA_ERROR_NOT_INITIALIZED,
16871 * ::CUDA_ERROR_INVALID_CONTEXT,
16872 * ::CUDA_ERROR_INVALID_VALUE
16873 *
16874 * \sa ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
16875 * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags, ::cuTexRefSetFormat,
16876 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
16877 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat,
16878 * ::cudaBindTexture
16879 */
16880__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefSetAddress(size_t *ByteOffset, CUtexref hTexRef, CUdeviceptr dptr, size_t bytes);
16881
16882/**
16883 * \brief Binds an address as a 2D texture reference
16884 *
16885 * \deprecated
16886 *
16887 * Binds a linear address range to the texture reference \p hTexRef. Any
16888 * previous address or CUDA array state associated with the texture reference
16889 * is superseded by this function. Any memory previously bound to \p hTexRef
16890 * is unbound.
16891 *
16892 * Using a ::tex2D() function inside a kernel requires a call to either
16893 * ::cuTexRefSetArray() to bind the corresponding texture reference to an
16894 * array, or ::cuTexRefSetAddress2D() to bind the texture reference to linear
16895 * memory.
16896 *
16897 * Function calls to ::cuTexRefSetFormat() cannot follow calls to
16898 * ::cuTexRefSetAddress2D() for the same texture reference.
16899 *
16900 * It is required that \p dptr be aligned to the appropriate hardware-specific
16901 * texture alignment. You can query this value using the device attribute
16902 * ::CU_DEVICE_ATTRIBUTE_TEXTURE_ALIGNMENT. If an unaligned \p dptr is
16903 * supplied, ::CUDA_ERROR_INVALID_VALUE is returned.
16904 *
16905 * \p Pitch has to be aligned to the hardware-specific texture pitch alignment.
16906 * This value can be queried using the device attribute
16907 * ::CU_DEVICE_ATTRIBUTE_TEXTURE_PITCH_ALIGNMENT. If an unaligned \p Pitch is
16908 * supplied, ::CUDA_ERROR_INVALID_VALUE is returned.
16909 *
16910 * Width and Height, which are specified in elements (or texels), cannot exceed
16911 * ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_WIDTH and
16912 * ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_HEIGHT respectively.
16913 * \p Pitch, which is specified in bytes, cannot exceed
16914 * ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_PITCH.
16915 *
16916 * \param hTexRef - Texture reference to bind
16917 * \param desc - Descriptor of CUDA array
16918 * \param dptr - Device pointer to bind
16919 * \param Pitch - Line pitch in bytes
16920 *
16921 * \return
16922 * ::CUDA_SUCCESS,
16923 * ::CUDA_ERROR_DEINITIALIZED,
16924 * ::CUDA_ERROR_NOT_INITIALIZED,
16925 * ::CUDA_ERROR_INVALID_CONTEXT,
16926 * ::CUDA_ERROR_INVALID_VALUE
16927 *
16928 * \sa ::cuTexRefSetAddress,
16929 * ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
16930 * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags, ::cuTexRefSetFormat,
16931 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
16932 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat,
16933 * ::cudaBindTexture2D
16934 */
16935__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefSetAddress2D(CUtexref hTexRef, const CUDA_ARRAY_DESCRIPTOR *desc, CUdeviceptr dptr, size_t Pitch);
16936
16937/**
16938 * \brief Sets the format for a texture reference
16939 *
16940 * \deprecated
16941 *
16942 * Specifies the format of the data to be read by the texture reference
16943 * \p hTexRef. \p fmt and \p NumPackedComponents are exactly analogous to the
16944 * ::Format and ::NumChannels members of the ::CUDA_ARRAY_DESCRIPTOR structure:
16945 * They specify the format of each component and the number of components per
16946 * array element.
16947 *
16948 * \param hTexRef - Texture reference
16949 * \param fmt - Format to set
16950 * \param NumPackedComponents - Number of components per array element
16951 *
16952 * \return
16953 * ::CUDA_SUCCESS,
16954 * ::CUDA_ERROR_DEINITIALIZED,
16955 * ::CUDA_ERROR_NOT_INITIALIZED,
16956 * ::CUDA_ERROR_INVALID_CONTEXT,
16957 * ::CUDA_ERROR_INVALID_VALUE
16958 *
16959 * \sa ::cuTexRefSetAddress,
16960 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
16961 * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags,
16962 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
16963 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat,
16964 * ::cudaCreateChannelDesc,
16965 * ::cudaBindTexture,
16966 * ::cudaBindTexture2D,
16967 * ::cudaBindTextureToArray,
16968 * ::cudaBindTextureToMipmappedArray
16969 */
16970__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefSetFormat(CUtexref hTexRef, CUarray_format fmt, int NumPackedComponents);
16971
16972/**
16973 * \brief Sets the addressing mode for a texture reference
16974 *
16975 * \deprecated
16976 *
16977 * Specifies the addressing mode \p am for the given dimension \p dim of the
16978 * texture reference \p hTexRef. If \p dim is zero, the addressing mode is
16979 * applied to the first parameter of the functions used to fetch from the
16980 * texture; if \p dim is 1, the second, and so on. ::CUaddress_mode is defined
16981 * as:
16982 * \code
16983 typedef enum CUaddress_mode_enum {
16984 CU_TR_ADDRESS_MODE_WRAP = 0,
16985 CU_TR_ADDRESS_MODE_CLAMP = 1,
16986 CU_TR_ADDRESS_MODE_MIRROR = 2,
16987 CU_TR_ADDRESS_MODE_BORDER = 3
16988 } CUaddress_mode;
16989 * \endcode
16990 *
16991 * Note that this call has no effect if \p hTexRef is bound to linear memory.
16992 * Also, if the flag, ::CU_TRSF_NORMALIZED_COORDINATES, is not set, the only
16993 * supported address mode is ::CU_TR_ADDRESS_MODE_CLAMP.
16994 *
16995 * \param hTexRef - Texture reference
16996 * \param dim - Dimension
16997 * \param am - Addressing mode to set
16998 *
16999 * \return
17000 * ::CUDA_SUCCESS,
17001 * ::CUDA_ERROR_DEINITIALIZED,
17002 * ::CUDA_ERROR_NOT_INITIALIZED,
17003 * ::CUDA_ERROR_INVALID_CONTEXT,
17004 * ::CUDA_ERROR_INVALID_VALUE
17005 *
17006 * \sa ::cuTexRefSetAddress,
17007 * ::cuTexRefSetAddress2D, ::cuTexRefSetArray,
17008 * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags, ::cuTexRefSetFormat,
17009 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
17010 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat,
17011 * ::cudaBindTexture,
17012 * ::cudaBindTexture2D,
17013 * ::cudaBindTextureToArray,
17014 * ::cudaBindTextureToMipmappedArray
17015 */
17016__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefSetAddressMode(CUtexref hTexRef, int dim, CUaddress_mode am);
17017
17018/**
17019 * \brief Sets the filtering mode for a texture reference
17020 *
17021 * \deprecated
17022 *
17023 * Specifies the filtering mode \p fm to be used when reading memory through
17024 * the texture reference \p hTexRef. ::CUfilter_mode_enum is defined as:
17025 *
17026 * \code
17027 typedef enum CUfilter_mode_enum {
17028 CU_TR_FILTER_MODE_POINT = 0,
17029 CU_TR_FILTER_MODE_LINEAR = 1
17030 } CUfilter_mode;
17031 * \endcode
17032 *
17033 * Note that this call has no effect if \p hTexRef is bound to linear memory.
17034 *
17035 * \param hTexRef - Texture reference
17036 * \param fm - Filtering mode to set
17037 *
17038 * \return
17039 * ::CUDA_SUCCESS,
17040 * ::CUDA_ERROR_DEINITIALIZED,
17041 * ::CUDA_ERROR_NOT_INITIALIZED,
17042 * ::CUDA_ERROR_INVALID_CONTEXT,
17043 * ::CUDA_ERROR_INVALID_VALUE
17044 *
17045 * \sa ::cuTexRefSetAddress,
17046 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
17047 * ::cuTexRefSetFlags, ::cuTexRefSetFormat,
17048 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
17049 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat,
17050 * ::cudaBindTextureToArray
17051 */
17052__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefSetFilterMode(CUtexref hTexRef, CUfilter_mode fm);
17053
17054/**
17055 * \brief Sets the mipmap filtering mode for a texture reference
17056 *
17057 * \deprecated
17058 *
17059 * Specifies the mipmap filtering mode \p fm to be used when reading memory through
17060 * the texture reference \p hTexRef. ::CUfilter_mode_enum is defined as:
17061 *
17062 * \code
17063 typedef enum CUfilter_mode_enum {
17064 CU_TR_FILTER_MODE_POINT = 0,
17065 CU_TR_FILTER_MODE_LINEAR = 1
17066 } CUfilter_mode;
17067 * \endcode
17068 *
17069 * Note that this call has no effect if \p hTexRef is not bound to a mipmapped array.
17070 *
17071 * \param hTexRef - Texture reference
17072 * \param fm - Filtering mode to set
17073 *
17074 * \return
17075 * ::CUDA_SUCCESS,
17076 * ::CUDA_ERROR_DEINITIALIZED,
17077 * ::CUDA_ERROR_NOT_INITIALIZED,
17078 * ::CUDA_ERROR_INVALID_CONTEXT,
17079 * ::CUDA_ERROR_INVALID_VALUE
17080 *
17081 * \sa ::cuTexRefSetAddress,
17082 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
17083 * ::cuTexRefSetFlags, ::cuTexRefSetFormat,
17084 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
17085 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat,
17086 * ::cudaBindTextureToMipmappedArray
17087 */
17088__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefSetMipmapFilterMode(CUtexref hTexRef, CUfilter_mode fm);
17089
17090/**
17091 * \brief Sets the mipmap level bias for a texture reference
17092 *
17093 * \deprecated
17094 *
17095 * Specifies the mipmap level bias \p bias to be added to the specified mipmap level when
17096 * reading memory through the texture reference \p hTexRef.
17097 *
17098 * Note that this call has no effect if \p hTexRef is not bound to a mipmapped array.
17099 *
17100 * \param hTexRef - Texture reference
17101 * \param bias - Mipmap level bias
17102 *
17103 * \return
17104 * ::CUDA_SUCCESS,
17105 * ::CUDA_ERROR_DEINITIALIZED,
17106 * ::CUDA_ERROR_NOT_INITIALIZED,
17107 * ::CUDA_ERROR_INVALID_CONTEXT,
17108 * ::CUDA_ERROR_INVALID_VALUE
17109 *
17110 * \sa ::cuTexRefSetAddress,
17111 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
17112 * ::cuTexRefSetFlags, ::cuTexRefSetFormat,
17113 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
17114 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat,
17115 * ::cudaBindTextureToMipmappedArray
17116 */
17117__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefSetMipmapLevelBias(CUtexref hTexRef, float bias);
17118
17119/**
17120 * \brief Sets the mipmap min/max mipmap level clamps for a texture reference
17121 *
17122 * \deprecated
17123 *
17124 * Specifies the min/max mipmap level clamps, \p minMipmapLevelClamp and \p maxMipmapLevelClamp
17125 * respectively, to be used when reading memory through the texture reference
17126 * \p hTexRef.
17127 *
17128 * Note that this call has no effect if \p hTexRef is not bound to a mipmapped array.
17129 *
17130 * \param hTexRef - Texture reference
17131 * \param minMipmapLevelClamp - Mipmap min level clamp
17132 * \param maxMipmapLevelClamp - Mipmap max level clamp
17133 *
17134 * \return
17135 * ::CUDA_SUCCESS,
17136 * ::CUDA_ERROR_DEINITIALIZED,
17137 * ::CUDA_ERROR_NOT_INITIALIZED,
17138 * ::CUDA_ERROR_INVALID_CONTEXT,
17139 * ::CUDA_ERROR_INVALID_VALUE
17140 *
17141 * \sa ::cuTexRefSetAddress,
17142 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
17143 * ::cuTexRefSetFlags, ::cuTexRefSetFormat,
17144 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
17145 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat,
17146 * ::cudaBindTextureToMipmappedArray
17147 */
17148__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefSetMipmapLevelClamp(CUtexref hTexRef, float minMipmapLevelClamp, float maxMipmapLevelClamp);
17149
17150/**
17151 * \brief Sets the maximum anisotropy for a texture reference
17152 *
17153 * \deprecated
17154 *
17155 * Specifies the maximum anisotropy \p maxAniso to be used when reading memory through
17156 * the texture reference \p hTexRef.
17157 *
17158 * Note that this call has no effect if \p hTexRef is bound to linear memory.
17159 *
17160 * \param hTexRef - Texture reference
17161 * \param maxAniso - Maximum anisotropy
17162 *
17163 * \return
17164 * ::CUDA_SUCCESS,
17165 * ::CUDA_ERROR_DEINITIALIZED,
17166 * ::CUDA_ERROR_NOT_INITIALIZED,
17167 * ::CUDA_ERROR_INVALID_CONTEXT,
17168 * ::CUDA_ERROR_INVALID_VALUE
17169 *
17170 * \sa ::cuTexRefSetAddress,
17171 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
17172 * ::cuTexRefSetFlags, ::cuTexRefSetFormat,
17173 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
17174 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat,
17175 * ::cudaBindTextureToArray,
17176 * ::cudaBindTextureToMipmappedArray
17177 */
17178__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefSetMaxAnisotropy(CUtexref hTexRef, unsigned int maxAniso);
17179
17180/**
17181 * \brief Sets the border color for a texture reference
17182 *
17183 * \deprecated
17184 *
17185 * Specifies the value of the RGBA color via the \p pBorderColor to the texture reference
17186 * \p hTexRef. The color value supports only float type and holds color components in
17187 * the following sequence:
17188 * pBorderColor[0] holds 'R' component
17189 * pBorderColor[1] holds 'G' component
17190 * pBorderColor[2] holds 'B' component
17191 * pBorderColor[3] holds 'A' component
17192 *
17193 * Note that the color values can be set only when the Address mode is set to
17194 * CU_TR_ADDRESS_MODE_BORDER using ::cuTexRefSetAddressMode.
17195 * Applications using integer border color values have to "reinterpret_cast" their values to float.
17196 *
17197 * \param hTexRef - Texture reference
17198 * \param pBorderColor - RGBA color
17199 *
17200 * \return
17201 * ::CUDA_SUCCESS,
17202 * ::CUDA_ERROR_DEINITIALIZED,
17203 * ::CUDA_ERROR_NOT_INITIALIZED,
17204 * ::CUDA_ERROR_INVALID_CONTEXT,
17205 * ::CUDA_ERROR_INVALID_VALUE
17206 *
17207 * \sa ::cuTexRefSetAddressMode,
17208 * ::cuTexRefGetAddressMode, ::cuTexRefGetBorderColor,
17209 * ::cudaBindTexture,
17210 * ::cudaBindTexture2D,
17211 * ::cudaBindTextureToArray,
17212 * ::cudaBindTextureToMipmappedArray
17213 */
17214__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefSetBorderColor(CUtexref hTexRef, float *pBorderColor);
17215
17216/**
17217 * \brief Sets the flags for a texture reference
17218 *
17219 * \deprecated
17220 *
17221 * Specifies optional flags via \p Flags to specify the behavior of data
17222 * returned through the texture reference \p hTexRef. The valid flags are:
17223 *
17224 * - ::CU_TRSF_READ_AS_INTEGER, which suppresses the default behavior of
17225 * having the texture promote integer data to floating point data in the
17226 * range [0, 1]. Note that texture with 32-bit integer format
17227 * would not be promoted, regardless of whether or not this
17228 * flag is specified;
17229 * - ::CU_TRSF_NORMALIZED_COORDINATES, which suppresses the
17230 * default behavior of having the texture coordinates range
17231 * from [0, Dim) where Dim is the width or height of the CUDA
17232 * array. Instead, the texture coordinates [0, 1.0) reference
17233 * the entire breadth of the array dimension;
17234 * - ::CU_TRSF_DISABLE_TRILINEAR_OPTIMIZATION, which disables any trilinear
17235 * filtering optimizations. Trilinear optimizations improve texture filtering
17236 * performance by allowing bilinear filtering on textures in scenarios where
17237 * it can closely approximate the expected results.
17238 *
17239 * \param hTexRef - Texture reference
17240 * \param Flags - Optional flags to set
17241 *
17242 * \return
17243 * ::CUDA_SUCCESS,
17244 * ::CUDA_ERROR_DEINITIALIZED,
17245 * ::CUDA_ERROR_NOT_INITIALIZED,
17246 * ::CUDA_ERROR_INVALID_CONTEXT,
17247 * ::CUDA_ERROR_INVALID_VALUE
17248 *
17249 * \sa ::cuTexRefSetAddress,
17250 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
17251 * ::cuTexRefSetFilterMode, ::cuTexRefSetFormat,
17252 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
17253 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat,
17254 * ::cudaBindTexture,
17255 * ::cudaBindTexture2D,
17256 * ::cudaBindTextureToArray,
17257 * ::cudaBindTextureToMipmappedArray
17258 */
17259__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefSetFlags(CUtexref hTexRef, unsigned int Flags);
17260
17261/**
17262 * \brief Gets the address associated with a texture reference
17263 *
17264 * \deprecated
17265 *
17266 * Returns in \p *pdptr the base address bound to the texture reference
17267 * \p hTexRef, or returns ::CUDA_ERROR_INVALID_VALUE if the texture reference
17268 * is not bound to any device memory range.
17269 *
17270 * \param pdptr - Returned device address
17271 * \param hTexRef - Texture reference
17272 *
17273 * \return
17274 * ::CUDA_SUCCESS,
17275 * ::CUDA_ERROR_DEINITIALIZED,
17276 * ::CUDA_ERROR_NOT_INITIALIZED,
17277 * ::CUDA_ERROR_INVALID_CONTEXT,
17278 * ::CUDA_ERROR_INVALID_VALUE
17279 *
17280 * \sa ::cuTexRefSetAddress,
17281 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
17282 * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags, ::cuTexRefSetFormat,
17283 * ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
17284 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat
17285 */
17286__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefGetAddress(CUdeviceptr *pdptr, CUtexref hTexRef);
17287
17288/**
17289 * \brief Gets the array bound to a texture reference
17290 *
17291 * \deprecated
17292 *
17293 * Returns in \p *phArray the CUDA array bound to the texture reference
17294 * \p hTexRef, or returns ::CUDA_ERROR_INVALID_VALUE if the texture reference
17295 * is not bound to any CUDA array.
17296 *
17297 * \param phArray - Returned array
17298 * \param hTexRef - Texture reference
17299 *
17300 * \return
17301 * ::CUDA_SUCCESS,
17302 * ::CUDA_ERROR_DEINITIALIZED,
17303 * ::CUDA_ERROR_NOT_INITIALIZED,
17304 * ::CUDA_ERROR_INVALID_CONTEXT,
17305 * ::CUDA_ERROR_INVALID_VALUE
17306 *
17307 * \sa ::cuTexRefSetAddress,
17308 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
17309 * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags, ::cuTexRefSetFormat,
17310 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode,
17311 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat
17312 */
17313__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefGetArray(CUarray *phArray, CUtexref hTexRef);
17314
17315/**
17316 * \brief Gets the mipmapped array bound to a texture reference
17317 *
17318 * \deprecated
17319 *
17320 * Returns in \p *phMipmappedArray the CUDA mipmapped array bound to the texture
17321 * reference \p hTexRef, or returns ::CUDA_ERROR_INVALID_VALUE if the texture reference
17322 * is not bound to any CUDA mipmapped array.
17323 *
17324 * \param phMipmappedArray - Returned mipmapped array
17325 * \param hTexRef - Texture reference
17326 *
17327 * \return
17328 * ::CUDA_SUCCESS,
17329 * ::CUDA_ERROR_DEINITIALIZED,
17330 * ::CUDA_ERROR_NOT_INITIALIZED,
17331 * ::CUDA_ERROR_INVALID_CONTEXT,
17332 * ::CUDA_ERROR_INVALID_VALUE
17333 *
17334 * \sa ::cuTexRefSetAddress,
17335 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
17336 * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags, ::cuTexRefSetFormat,
17337 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode,
17338 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat
17339 */
17340__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefGetMipmappedArray(CUmipmappedArray *phMipmappedArray, CUtexref hTexRef);
17341
17342/**
17343 * \brief Gets the addressing mode used by a texture reference
17344 *
17345 * \deprecated
17346 *
17347 * Returns in \p *pam the addressing mode corresponding to the
17348 * dimension \p dim of the texture reference \p hTexRef. Currently, the only
17349 * valid value for \p dim are 0 and 1.
17350 *
17351 * \param pam - Returned addressing mode
17352 * \param hTexRef - Texture reference
17353 * \param dim - Dimension
17354 *
17355 * \return
17356 * ::CUDA_SUCCESS,
17357 * ::CUDA_ERROR_DEINITIALIZED,
17358 * ::CUDA_ERROR_NOT_INITIALIZED,
17359 * ::CUDA_ERROR_INVALID_CONTEXT,
17360 * ::CUDA_ERROR_INVALID_VALUE
17361 *
17362 * \sa ::cuTexRefSetAddress,
17363 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
17364 * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags, ::cuTexRefSetFormat,
17365 * ::cuTexRefGetAddress, ::cuTexRefGetArray,
17366 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat
17367 */
17368__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefGetAddressMode(CUaddress_mode *pam, CUtexref hTexRef, int dim);
17369
17370/**
17371 * \brief Gets the filter-mode used by a texture reference
17372 *
17373 * \deprecated
17374 *
17375 * Returns in \p *pfm the filtering mode of the texture reference
17376 * \p hTexRef.
17377 *
17378 * \param pfm - Returned filtering mode
17379 * \param hTexRef - Texture reference
17380 *
17381 * \return
17382 * ::CUDA_SUCCESS,
17383 * ::CUDA_ERROR_DEINITIALIZED,
17384 * ::CUDA_ERROR_NOT_INITIALIZED,
17385 * ::CUDA_ERROR_INVALID_CONTEXT,
17386 * ::CUDA_ERROR_INVALID_VALUE
17387 *
17388 * \sa ::cuTexRefSetAddress,
17389 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
17390 * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags, ::cuTexRefSetFormat,
17391 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
17392 * ::cuTexRefGetFlags, ::cuTexRefGetFormat
17393 */
17394__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefGetFilterMode(CUfilter_mode *pfm, CUtexref hTexRef);
17395
17396/**
17397 * \brief Gets the format used by a texture reference
17398 *
17399 * \deprecated
17400 *
17401 * Returns in \p *pFormat and \p *pNumChannels the format and number
17402 * of components of the CUDA array bound to the texture reference \p hTexRef.
17403 * If \p pFormat or \p pNumChannels is NULL, it will be ignored.
17404 *
17405 * \param pFormat - Returned format
17406 * \param pNumChannels - Returned number of components
17407 * \param hTexRef - Texture reference
17408 *
17409 * \return
17410 * ::CUDA_SUCCESS,
17411 * ::CUDA_ERROR_DEINITIALIZED,
17412 * ::CUDA_ERROR_NOT_INITIALIZED,
17413 * ::CUDA_ERROR_INVALID_CONTEXT,
17414 * ::CUDA_ERROR_INVALID_VALUE
17415 *
17416 * \sa ::cuTexRefSetAddress,
17417 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
17418 * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags, ::cuTexRefSetFormat,
17419 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
17420 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags
17421 */
17422__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefGetFormat(CUarray_format *pFormat, int *pNumChannels, CUtexref hTexRef);
17423
17424/**
17425 * \brief Gets the mipmap filtering mode for a texture reference
17426 *
17427 * \deprecated
17428 *
17429 * Returns the mipmap filtering mode in \p pfm that's used when reading memory through
17430 * the texture reference \p hTexRef.
17431 *
17432 * \param pfm - Returned mipmap filtering mode
17433 * \param hTexRef - Texture reference
17434 *
17435 * \return
17436 * ::CUDA_SUCCESS,
17437 * ::CUDA_ERROR_DEINITIALIZED,
17438 * ::CUDA_ERROR_NOT_INITIALIZED,
17439 * ::CUDA_ERROR_INVALID_CONTEXT,
17440 * ::CUDA_ERROR_INVALID_VALUE
17441 *
17442 * \sa ::cuTexRefSetAddress,
17443 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
17444 * ::cuTexRefSetFlags, ::cuTexRefSetFormat,
17445 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
17446 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat
17447 */
17448__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefGetMipmapFilterMode(CUfilter_mode *pfm, CUtexref hTexRef);
17449
17450/**
17451 * \brief Gets the mipmap level bias for a texture reference
17452 *
17453 * \deprecated
17454 *
17455 * Returns the mipmap level bias in \p pBias that's added to the specified mipmap
17456 * level when reading memory through the texture reference \p hTexRef.
17457 *
17458 * \param pbias - Returned mipmap level bias
17459 * \param hTexRef - Texture reference
17460 *
17461 * \return
17462 * ::CUDA_SUCCESS,
17463 * ::CUDA_ERROR_DEINITIALIZED,
17464 * ::CUDA_ERROR_NOT_INITIALIZED,
17465 * ::CUDA_ERROR_INVALID_CONTEXT,
17466 * ::CUDA_ERROR_INVALID_VALUE
17467 *
17468 * \sa ::cuTexRefSetAddress,
17469 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
17470 * ::cuTexRefSetFlags, ::cuTexRefSetFormat,
17471 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
17472 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat
17473 */
17474__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefGetMipmapLevelBias(float *pbias, CUtexref hTexRef);
17475
17476/**
17477 * \brief Gets the min/max mipmap level clamps for a texture reference
17478 *
17479 * \deprecated
17480 *
17481 * Returns the min/max mipmap level clamps in \p pminMipmapLevelClamp and \p pmaxMipmapLevelClamp
17482 * that's used when reading memory through the texture reference \p hTexRef.
17483 *
17484 * \param pminMipmapLevelClamp - Returned mipmap min level clamp
17485 * \param pmaxMipmapLevelClamp - Returned mipmap max level clamp
17486 * \param hTexRef - Texture reference
17487 *
17488 * \return
17489 * ::CUDA_SUCCESS,
17490 * ::CUDA_ERROR_DEINITIALIZED,
17491 * ::CUDA_ERROR_NOT_INITIALIZED,
17492 * ::CUDA_ERROR_INVALID_CONTEXT,
17493 * ::CUDA_ERROR_INVALID_VALUE
17494 *
17495 * \sa ::cuTexRefSetAddress,
17496 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
17497 * ::cuTexRefSetFlags, ::cuTexRefSetFormat,
17498 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
17499 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat
17500 */
17501__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefGetMipmapLevelClamp(float *pminMipmapLevelClamp, float *pmaxMipmapLevelClamp, CUtexref hTexRef);
17502
17503/**
17504 * \brief Gets the maximum anisotropy for a texture reference
17505 *
17506 * \deprecated
17507 *
17508 * Returns the maximum anisotropy in \p pmaxAniso that's used when reading memory through
17509 * the texture reference \p hTexRef.
17510 *
17511 * \param pmaxAniso - Returned maximum anisotropy
17512 * \param hTexRef - Texture reference
17513 *
17514 * \return
17515 * ::CUDA_SUCCESS,
17516 * ::CUDA_ERROR_DEINITIALIZED,
17517 * ::CUDA_ERROR_NOT_INITIALIZED,
17518 * ::CUDA_ERROR_INVALID_CONTEXT,
17519 * ::CUDA_ERROR_INVALID_VALUE
17520 *
17521 * \sa ::cuTexRefSetAddress,
17522 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
17523 * ::cuTexRefSetFlags, ::cuTexRefSetFormat,
17524 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
17525 * ::cuTexRefGetFilterMode, ::cuTexRefGetFlags, ::cuTexRefGetFormat
17526 */
17527__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefGetMaxAnisotropy(int *pmaxAniso, CUtexref hTexRef);
17528
17529/**
17530 * \brief Gets the border color used by a texture reference
17531 *
17532 * \deprecated
17533 *
17534 * Returns in \p pBorderColor, values of the RGBA color used by
17535 * the texture reference \p hTexRef.
17536 * The color value is of type float and holds color components in
17537 * the following sequence:
17538 * pBorderColor[0] holds 'R' component
17539 * pBorderColor[1] holds 'G' component
17540 * pBorderColor[2] holds 'B' component
17541 * pBorderColor[3] holds 'A' component
17542 *
17543 * \param hTexRef - Texture reference
17544 * \param pBorderColor - Returned Type and Value of RGBA color
17545 *
17546 * \return
17547 * ::CUDA_SUCCESS,
17548 * ::CUDA_ERROR_DEINITIALIZED,
17549 * ::CUDA_ERROR_NOT_INITIALIZED,
17550 * ::CUDA_ERROR_INVALID_CONTEXT,
17551 * ::CUDA_ERROR_INVALID_VALUE
17552 *
17553 * \sa ::cuTexRefSetAddressMode,
17554 * ::cuTexRefSetAddressMode, ::cuTexRefSetBorderColor
17555 */
17556__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefGetBorderColor(float *pBorderColor, CUtexref hTexRef);
17557
17558/**
17559 * \brief Gets the flags used by a texture reference
17560 *
17561 * \deprecated
17562 *
17563 * Returns in \p *pFlags the flags of the texture reference \p hTexRef.
17564 *
17565 * \param pFlags - Returned flags
17566 * \param hTexRef - Texture reference
17567 *
17568 * \return
17569 * ::CUDA_SUCCESS,
17570 * ::CUDA_ERROR_DEINITIALIZED,
17571 * ::CUDA_ERROR_NOT_INITIALIZED,
17572 * ::CUDA_ERROR_INVALID_CONTEXT,
17573 * ::CUDA_ERROR_INVALID_VALUE
17574 *
17575 * \sa ::cuTexRefSetAddress,
17576 * ::cuTexRefSetAddress2D, ::cuTexRefSetAddressMode, ::cuTexRefSetArray,
17577 * ::cuTexRefSetFilterMode, ::cuTexRefSetFlags, ::cuTexRefSetFormat,
17578 * ::cuTexRefGetAddress, ::cuTexRefGetAddressMode, ::cuTexRefGetArray,
17579 * ::cuTexRefGetFilterMode, ::cuTexRefGetFormat
17580 */
17581__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefGetFlags(unsigned int *pFlags, CUtexref hTexRef);
17582
17583/**
17584 * \brief Creates a texture reference
17585 *
17586 * \deprecated
17587 *
17588 * Creates a texture reference and returns its handle in \p *pTexRef. Once
17589 * created, the application must call ::cuTexRefSetArray() or
17590 * ::cuTexRefSetAddress() to associate the reference with allocated memory.
17591 * Other texture reference functions are used to specify the format and
17592 * interpretation (addressing, filtering, etc.) to be used when the memory is
17593 * read through this texture reference.
17594 *
17595 * \param pTexRef - Returned texture reference
17596 *
17597 * \return
17598 * ::CUDA_SUCCESS,
17599 * ::CUDA_ERROR_DEINITIALIZED,
17600 * ::CUDA_ERROR_NOT_INITIALIZED,
17601 * ::CUDA_ERROR_INVALID_CONTEXT,
17602 * ::CUDA_ERROR_INVALID_VALUE
17603 *
17604 * \sa ::cuTexRefDestroy
17605 */
17606__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefCreate(CUtexref *pTexRef);
17607
17608/**
17609 * \brief Destroys a texture reference
17610 *
17611 * \deprecated
17612 *
17613 * Destroys the texture reference specified by \p hTexRef.
17614 *
17615 * \param hTexRef - Texture reference to destroy
17616 *
17617 * \return
17618 * ::CUDA_SUCCESS,
17619 * ::CUDA_ERROR_DEINITIALIZED,
17620 * ::CUDA_ERROR_NOT_INITIALIZED,
17621 * ::CUDA_ERROR_INVALID_CONTEXT,
17622 * ::CUDA_ERROR_INVALID_VALUE
17623 *
17624 * \sa ::cuTexRefCreate
17625 */
17626__CUDA_DEPRECATED CUresult CUDAAPI cuTexRefDestroy(CUtexref hTexRef);
17627
17628/** @} */ /* END CUDA_TEXREF_DEPRECATED */
17629
17630
17631/**
17632 * \defgroup CUDA_SURFREF_DEPRECATED Surface Reference Management [DEPRECATED]
17633 *
17634 * ___MANBRIEF___ surface reference management functions of the low-level CUDA
17635 * driver API (___CURRENT_FILE___) ___ENDMANBRIEF___
17636 *
17637 * This section describes the surface reference management functions of the
17638 * low-level CUDA driver application programming interface.
17639 *
17640 * @{
17641 */
17642
17643/**
17644 * \brief Sets the CUDA array for a surface reference.
17645 *
17646 * \deprecated
17647 *
17648 * Sets the CUDA array \p hArray to be read and written by the surface reference
17649 * \p hSurfRef. Any previous CUDA array state associated with the surface
17650 * reference is superseded by this function. \p Flags must be set to 0.
17651 * The ::CUDA_ARRAY3D_SURFACE_LDST flag must have been set for the CUDA array.
17652 * Any CUDA array previously bound to \p hSurfRef is unbound.
17653
17654 * \param hSurfRef - Surface reference handle
17655 * \param hArray - CUDA array handle
17656 * \param Flags - set to 0
17657 *
17658 * \return
17659 * ::CUDA_SUCCESS,
17660 * ::CUDA_ERROR_DEINITIALIZED,
17661 * ::CUDA_ERROR_NOT_INITIALIZED,
17662 * ::CUDA_ERROR_INVALID_CONTEXT,
17663 * ::CUDA_ERROR_INVALID_VALUE
17664 *
17665 * \sa
17666 * ::cuModuleGetSurfRef,
17667 * ::cuSurfRefGetArray,
17668 * ::cudaBindSurfaceToArray
17669 */
17670__CUDA_DEPRECATED CUresult CUDAAPI cuSurfRefSetArray(CUsurfref hSurfRef, CUarray hArray, unsigned int Flags);
17671
17672/**
17673 * \brief Passes back the CUDA array bound to a surface reference.
17674 *
17675 * \deprecated
17676 *
17677 * Returns in \p *phArray the CUDA array bound to the surface reference
17678 * \p hSurfRef, or returns ::CUDA_ERROR_INVALID_VALUE if the surface reference
17679 * is not bound to any CUDA array.
17680
17681 * \param phArray - Surface reference handle
17682 * \param hSurfRef - Surface reference handle
17683 *
17684 * \return
17685 * ::CUDA_SUCCESS,
17686 * ::CUDA_ERROR_DEINITIALIZED,
17687 * ::CUDA_ERROR_NOT_INITIALIZED,
17688 * ::CUDA_ERROR_INVALID_CONTEXT,
17689 * ::CUDA_ERROR_INVALID_VALUE
17690 *
17691 * \sa ::cuModuleGetSurfRef, ::cuSurfRefSetArray
17692 */
17693__CUDA_DEPRECATED CUresult CUDAAPI cuSurfRefGetArray(CUarray *phArray, CUsurfref hSurfRef);
17694
17695/** @} */ /* END CUDA_SURFREF_DEPRECATED */
17696
17697/**
17698 * \defgroup CUDA_TEXOBJECT Texture Object Management
17699 *
17700 * ___MANBRIEF___ texture object management functions of the low-level CUDA
17701 * driver API (___CURRENT_FILE___) ___ENDMANBRIEF___
17702 *
17703 * This section describes the texture object management functions of the
17704 * low-level CUDA driver application programming interface. The texture
17705 * object API is only supported on devices of compute capability 3.0 or higher.
17706 *
17707 * @{
17708 */
17709
17710/**
17711 * \brief Creates a texture object
17712 *
17713 * Creates a texture object and returns it in \p pTexObject. \p pResDesc describes
17714 * the data to texture from. \p pTexDesc describes how the data should be sampled.
17715 * \p pResViewDesc is an optional argument that specifies an alternate format for
17716 * the data described by \p pResDesc, and also describes the subresource region
17717 * to restrict access to when texturing. \p pResViewDesc can only be specified if
17718 * the type of resource is a CUDA array or a CUDA mipmapped array.
17719 *
17720 * Texture objects are only supported on devices of compute capability 3.0 or higher.
17721 * Additionally, a texture object is an opaque value, and, as such, should only be
17722 * accessed through CUDA API calls.
17723 *
17724 * The ::CUDA_RESOURCE_DESC structure is defined as:
17725 * \code
17726 typedef struct CUDA_RESOURCE_DESC_st
17727 {
17728 CUresourcetype resType;
17729
17730 union {
17731 struct {
17732 CUarray hArray;
17733 } array;
17734 struct {
17735 CUmipmappedArray hMipmappedArray;
17736 } mipmap;
17737 struct {
17738 CUdeviceptr devPtr;
17739 CUarray_format format;
17740 unsigned int numChannels;
17741 size_t sizeInBytes;
17742 } linear;
17743 struct {
17744 CUdeviceptr devPtr;
17745 CUarray_format format;
17746 unsigned int numChannels;
17747 size_t width;
17748 size_t height;
17749 size_t pitchInBytes;
17750 } pitch2D;
17751 } res;
17752
17753 unsigned int flags;
17754 } CUDA_RESOURCE_DESC;
17755
17756 * \endcode
17757 * where:
17758 * - ::CUDA_RESOURCE_DESC::resType specifies the type of resource to texture from.
17759 * CUresourceType is defined as:
17760 * \code
17761 typedef enum CUresourcetype_enum {
17762 CU_RESOURCE_TYPE_ARRAY = 0x00,
17763 CU_RESOURCE_TYPE_MIPMAPPED_ARRAY = 0x01,
17764 CU_RESOURCE_TYPE_LINEAR = 0x02,
17765 CU_RESOURCE_TYPE_PITCH2D = 0x03
17766 } CUresourcetype;
17767 * \endcode
17768 *
17769 * \par
17770 * If ::CUDA_RESOURCE_DESC::resType is set to ::CU_RESOURCE_TYPE_ARRAY, ::CUDA_RESOURCE_DESC::res::array::hArray
17771 * must be set to a valid CUDA array handle.
17772 *
17773 * \par
17774 * If ::CUDA_RESOURCE_DESC::resType is set to ::CU_RESOURCE_TYPE_MIPMAPPED_ARRAY, ::CUDA_RESOURCE_DESC::res::mipmap::hMipmappedArray
17775 * must be set to a valid CUDA mipmapped array handle.
17776 *
17777 * \par
17778 * If ::CUDA_RESOURCE_DESC::resType is set to ::CU_RESOURCE_TYPE_LINEAR, ::CUDA_RESOURCE_DESC::res::linear::devPtr
17779 * must be set to a valid device pointer, that is aligned to ::CU_DEVICE_ATTRIBUTE_TEXTURE_ALIGNMENT.
17780 * ::CUDA_RESOURCE_DESC::res::linear::format and ::CUDA_RESOURCE_DESC::res::linear::numChannels
17781 * describe the format of each component and the number of components per array element. ::CUDA_RESOURCE_DESC::res::linear::sizeInBytes
17782 * specifies the size of the array in bytes. The total number of elements in the linear address range cannot exceed
17783 * ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_LINEAR_WIDTH. The number of elements is computed as (sizeInBytes / (sizeof(format) * numChannels)).
17784 *
17785 * \par
17786 * If ::CUDA_RESOURCE_DESC::resType is set to ::CU_RESOURCE_TYPE_PITCH2D, ::CUDA_RESOURCE_DESC::res::pitch2D::devPtr
17787 * must be set to a valid device pointer, that is aligned to ::CU_DEVICE_ATTRIBUTE_TEXTURE_ALIGNMENT.
17788 * ::CUDA_RESOURCE_DESC::res::pitch2D::format and ::CUDA_RESOURCE_DESC::res::pitch2D::numChannels
17789 * describe the format of each component and the number of components per array element. ::CUDA_RESOURCE_DESC::res::pitch2D::width
17790 * and ::CUDA_RESOURCE_DESC::res::pitch2D::height specify the width and height of the array in elements, and cannot exceed
17791 * ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_WIDTH and ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_HEIGHT respectively.
17792 * ::CUDA_RESOURCE_DESC::res::pitch2D::pitchInBytes specifies the pitch between two rows in bytes and has to be aligned to
17793 * ::CU_DEVICE_ATTRIBUTE_TEXTURE_PITCH_ALIGNMENT. Pitch cannot exceed ::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE2D_LINEAR_PITCH.
17794 *
17795 * - ::flags must be set to zero.
17796 *
17797 *
17798 * The ::CUDA_TEXTURE_DESC struct is defined as
17799 * \code
17800 typedef struct CUDA_TEXTURE_DESC_st {
17801 CUaddress_mode addressMode[3];
17802 CUfilter_mode filterMode;
17803 unsigned int flags;
17804 unsigned int maxAnisotropy;
17805 CUfilter_mode mipmapFilterMode;
17806 float mipmapLevelBias;
17807 float minMipmapLevelClamp;
17808 float maxMipmapLevelClamp;
17809 } CUDA_TEXTURE_DESC;
17810 * \endcode
17811 * where
17812 * - ::CUDA_TEXTURE_DESC::addressMode specifies the addressing mode for each dimension of the texture data. ::CUaddress_mode is defined as:
17813 * \code
17814 typedef enum CUaddress_mode_enum {
17815 CU_TR_ADDRESS_MODE_WRAP = 0,
17816 CU_TR_ADDRESS_MODE_CLAMP = 1,
17817 CU_TR_ADDRESS_MODE_MIRROR = 2,
17818 CU_TR_ADDRESS_MODE_BORDER = 3
17819 } CUaddress_mode;
17820 * \endcode
17821 * This is ignored if ::CUDA_RESOURCE_DESC::resType is ::CU_RESOURCE_TYPE_LINEAR. Also, if the flag, ::CU_TRSF_NORMALIZED_COORDINATES
17822 * is not set, the only supported address mode is ::CU_TR_ADDRESS_MODE_CLAMP.
17823 *
17824 * - ::CUDA_TEXTURE_DESC::filterMode specifies the filtering mode to be used when fetching from the texture. CUfilter_mode is defined as:
17825 * \code
17826 typedef enum CUfilter_mode_enum {
17827 CU_TR_FILTER_MODE_POINT = 0,
17828 CU_TR_FILTER_MODE_LINEAR = 1
17829 } CUfilter_mode;
17830 * \endcode
17831 * This is ignored if ::CUDA_RESOURCE_DESC::resType is ::CU_RESOURCE_TYPE_LINEAR.
17832 *
17833 * - ::CUDA_TEXTURE_DESC::flags can be any combination of the following:
17834 * - ::CU_TRSF_READ_AS_INTEGER, which suppresses the default behavior of
17835 * having the texture promote integer data to floating point data in the
17836 * range [0, 1]. Note that texture with 32-bit integer format would not be
17837 * promoted, regardless of whether or not this flag is specified.
17838 * - ::CU_TRSF_NORMALIZED_COORDINATES, which suppresses the default behavior
17839 * of having the texture coordinates range from [0, Dim) where Dim is the
17840 * width or height of the CUDA array. Instead, the texture coordinates
17841 * [0, 1.0) reference the entire breadth of the array dimension; Note that
17842 * for CUDA mipmapped arrays, this flag has to be set.
17843 * - ::CU_TRSF_DISABLE_TRILINEAR_OPTIMIZATION, which disables any trilinear
17844 * filtering optimizations. Trilinear optimizations improve texture filtering
17845 * performance by allowing bilinear filtering on textures in scenarios where
17846 * it can closely approximate the expected results.
17847 *
17848 * - ::CUDA_TEXTURE_DESC::maxAnisotropy specifies the maximum anisotropy ratio to be used when doing anisotropic filtering. This value will be
17849 * clamped to the range [1,16].
17850 *
17851 * - ::CUDA_TEXTURE_DESC::mipmapFilterMode specifies the filter mode when the calculated mipmap level lies between two defined mipmap levels.
17852 *
17853 * - ::CUDA_TEXTURE_DESC::mipmapLevelBias specifies the offset to be applied to the calculated mipmap level.
17854 *
17855 * - ::CUDA_TEXTURE_DESC::minMipmapLevelClamp specifies the lower end of the mipmap level range to clamp access to.
17856 *
17857 * - ::CUDA_TEXTURE_DESC::maxMipmapLevelClamp specifies the upper end of the mipmap level range to clamp access to.
17858 *
17859 *
17860 * The ::CUDA_RESOURCE_VIEW_DESC struct is defined as
17861 * \code
17862 typedef struct CUDA_RESOURCE_VIEW_DESC_st
17863 {
17864 CUresourceViewFormat format;
17865 size_t width;
17866 size_t height;
17867 size_t depth;
17868 unsigned int firstMipmapLevel;
17869 unsigned int lastMipmapLevel;
17870 unsigned int firstLayer;
17871 unsigned int lastLayer;
17872 } CUDA_RESOURCE_VIEW_DESC;
17873 * \endcode
17874 * where:
17875 * - ::CUDA_RESOURCE_VIEW_DESC::format specifies how the data contained in the CUDA array or CUDA mipmapped array should
17876 * be interpreted. Note that this can incur a change in size of the texture data. If the resource view format is a block
17877 * compressed format, then the underlying CUDA array or CUDA mipmapped array has to have a base of format ::CU_AD_FORMAT_UNSIGNED_INT32.
17878 * with 2 or 4 channels, depending on the block compressed format. For ex., BC1 and BC4 require the underlying CUDA array to have
17879 * a format of ::CU_AD_FORMAT_UNSIGNED_INT32 with 2 channels. The other BC formats require the underlying resource to have the same base
17880 * format but with 4 channels.
17881 *
17882 * - ::CUDA_RESOURCE_VIEW_DESC::width specifies the new width of the texture data. If the resource view format is a block
17883 * compressed format, this value has to be 4 times the original width of the resource. For non block compressed formats,
17884 * this value has to be equal to that of the original resource.
17885 *
17886 * - ::CUDA_RESOURCE_VIEW_DESC::height specifies the new height of the texture data. If the resource view format is a block
17887 * compressed format, this value has to be 4 times the original height of the resource. For non block compressed formats,
17888 * this value has to be equal to that of the original resource.
17889 *
17890 * - ::CUDA_RESOURCE_VIEW_DESC::depth specifies the new depth of the texture data. This value has to be equal to that of the
17891 * original resource.
17892 *
17893 * - ::CUDA_RESOURCE_VIEW_DESC::firstMipmapLevel specifies the most detailed mipmap level. This will be the new mipmap level zero.
17894 * For non-mipmapped resources, this value has to be zero.::CUDA_TEXTURE_DESC::minMipmapLevelClamp and ::CUDA_TEXTURE_DESC::maxMipmapLevelClamp
17895 * will be relative to this value. For ex., if the firstMipmapLevel is set to 2, and a minMipmapLevelClamp of 1.2 is specified,
17896 * then the actual minimum mipmap level clamp will be 3.2.
17897 *
17898 * - ::CUDA_RESOURCE_VIEW_DESC::lastMipmapLevel specifies the least detailed mipmap level. For non-mipmapped resources, this value
17899 * has to be zero.
17900 *
17901 * - ::CUDA_RESOURCE_VIEW_DESC::firstLayer specifies the first layer index for layered textures. This will be the new layer zero.
17902 * For non-layered resources, this value has to be zero.
17903 *
17904 * - ::CUDA_RESOURCE_VIEW_DESC::lastLayer specifies the last layer index for layered textures. For non-layered resources,
17905 * this value has to be zero.
17906 *
17907 *
17908 * \param pTexObject - Texture object to create
17909 * \param pResDesc - Resource descriptor
17910 * \param pTexDesc - Texture descriptor
17911 * \param pResViewDesc - Resource view descriptor
17912 *
17913 * \return
17914 * ::CUDA_SUCCESS,
17915 * ::CUDA_ERROR_DEINITIALIZED,
17916 * ::CUDA_ERROR_NOT_INITIALIZED,
17917 * ::CUDA_ERROR_INVALID_CONTEXT,
17918 * ::CUDA_ERROR_INVALID_VALUE
17919 *
17920 * \sa
17921 * ::cuTexObjectDestroy,
17922 * ::cudaCreateTextureObject
17923 */
17924CUresult CUDAAPI cuTexObjectCreate(CUtexObject *pTexObject, const CUDA_RESOURCE_DESC *pResDesc, const CUDA_TEXTURE_DESC *pTexDesc, const CUDA_RESOURCE_VIEW_DESC *pResViewDesc);
17925
17926/**
17927 * \brief Destroys a texture object
17928 *
17929 * Destroys the texture object specified by \p texObject.
17930 *
17931 * \param texObject - Texture object to destroy
17932 *
17933 * \return
17934 * ::CUDA_SUCCESS,
17935 * ::CUDA_ERROR_DEINITIALIZED,
17936 * ::CUDA_ERROR_NOT_INITIALIZED,
17937 * ::CUDA_ERROR_INVALID_CONTEXT,
17938 * ::CUDA_ERROR_INVALID_VALUE
17939 *
17940 * \sa
17941 * ::cuTexObjectCreate,
17942 * ::cudaDestroyTextureObject
17943 */
17944CUresult CUDAAPI cuTexObjectDestroy(CUtexObject texObject);
17945
17946/**
17947 * \brief Returns a texture object's resource descriptor
17948 *
17949 * Returns the resource descriptor for the texture object specified by \p texObject.
17950 *
17951 * \param pResDesc - Resource descriptor
17952 * \param texObject - Texture object
17953 *
17954 * \return
17955 * ::CUDA_SUCCESS,
17956 * ::CUDA_ERROR_DEINITIALIZED,
17957 * ::CUDA_ERROR_NOT_INITIALIZED,
17958 * ::CUDA_ERROR_INVALID_CONTEXT,
17959 * ::CUDA_ERROR_INVALID_VALUE
17960 *
17961 * \sa
17962 * ::cuTexObjectCreate,
17963 * ::cudaGetTextureObjectResourceDesc,
17964 */
17965CUresult CUDAAPI cuTexObjectGetResourceDesc(CUDA_RESOURCE_DESC *pResDesc, CUtexObject texObject);
17966
17967/**
17968 * \brief Returns a texture object's texture descriptor
17969 *
17970 * Returns the texture descriptor for the texture object specified by \p texObject.
17971 *
17972 * \param pTexDesc - Texture descriptor
17973 * \param texObject - Texture object
17974 *
17975 * \return
17976 * ::CUDA_SUCCESS,
17977 * ::CUDA_ERROR_DEINITIALIZED,
17978 * ::CUDA_ERROR_NOT_INITIALIZED,
17979 * ::CUDA_ERROR_INVALID_CONTEXT,
17980 * ::CUDA_ERROR_INVALID_VALUE
17981 *
17982 * \sa
17983 * ::cuTexObjectCreate,
17984 * ::cudaGetTextureObjectTextureDesc
17985 */
17986CUresult CUDAAPI cuTexObjectGetTextureDesc(CUDA_TEXTURE_DESC *pTexDesc, CUtexObject texObject);
17987
17988/**
17989 * \brief Returns a texture object's resource view descriptor
17990 *
17991 * Returns the resource view descriptor for the texture object specified by \p texObject.
17992 * If no resource view was set for \p texObject, the ::CUDA_ERROR_INVALID_VALUE is returned.
17993 *
17994 * \param pResViewDesc - Resource view descriptor
17995 * \param texObject - Texture object
17996 *
17997 * \return
17998 * ::CUDA_SUCCESS,
17999 * ::CUDA_ERROR_DEINITIALIZED,
18000 * ::CUDA_ERROR_NOT_INITIALIZED,
18001 * ::CUDA_ERROR_INVALID_CONTEXT,
18002 * ::CUDA_ERROR_INVALID_VALUE
18003 *
18004 * \sa
18005 * ::cuTexObjectCreate,
18006 * ::cudaGetTextureObjectResourceViewDesc
18007 */
18008CUresult CUDAAPI cuTexObjectGetResourceViewDesc(CUDA_RESOURCE_VIEW_DESC *pResViewDesc, CUtexObject texObject);
18009
18010/** @} */ /* END CUDA_TEXOBJECT */
18011
18012/**
18013 * \defgroup CUDA_SURFOBJECT Surface Object Management
18014 *
18015 * ___MANBRIEF___ surface object management functions of the low-level CUDA
18016 * driver API (___CURRENT_FILE___) ___ENDMANBRIEF___
18017 *
18018 * This section describes the surface object management functions of the
18019 * low-level CUDA driver application programming interface. The surface
18020 * object API is only supported on devices of compute capability 3.0 or higher.
18021 *
18022 * @{
18023 */
18024
18025/**
18026 * \brief Creates a surface object
18027 *
18028 * Creates a surface object and returns it in \p pSurfObject. \p pResDesc describes
18029 * the data to perform surface load/stores on. ::CUDA_RESOURCE_DESC::resType must be
18030 * ::CU_RESOURCE_TYPE_ARRAY and ::CUDA_RESOURCE_DESC::res::array::hArray
18031 * must be set to a valid CUDA array handle. ::CUDA_RESOURCE_DESC::flags must be set to zero.
18032 *
18033 * Surface objects are only supported on devices of compute capability 3.0 or higher.
18034 * Additionally, a surface object is an opaque value, and, as such, should only be
18035 * accessed through CUDA API calls.
18036 *
18037 * \param pSurfObject - Surface object to create
18038 * \param pResDesc - Resource descriptor
18039 *
18040 * \return
18041 * ::CUDA_SUCCESS,
18042 * ::CUDA_ERROR_DEINITIALIZED,
18043 * ::CUDA_ERROR_NOT_INITIALIZED,
18044 * ::CUDA_ERROR_INVALID_CONTEXT,
18045 * ::CUDA_ERROR_INVALID_VALUE
18046 *
18047 * \sa
18048 * ::cuSurfObjectDestroy,
18049 * ::cudaCreateSurfaceObject
18050 */
18051CUresult CUDAAPI cuSurfObjectCreate(CUsurfObject *pSurfObject, const CUDA_RESOURCE_DESC *pResDesc);
18052
18053/**
18054 * \brief Destroys a surface object
18055 *
18056 * Destroys the surface object specified by \p surfObject.
18057 *
18058 * \param surfObject - Surface object to destroy
18059 *
18060 * \return
18061 * ::CUDA_SUCCESS,
18062 * ::CUDA_ERROR_DEINITIALIZED,
18063 * ::CUDA_ERROR_NOT_INITIALIZED,
18064 * ::CUDA_ERROR_INVALID_CONTEXT,
18065 * ::CUDA_ERROR_INVALID_VALUE
18066 *
18067 * \sa
18068 * ::cuSurfObjectCreate,
18069 * ::cudaDestroySurfaceObject
18070 */
18071CUresult CUDAAPI cuSurfObjectDestroy(CUsurfObject surfObject);
18072
18073/**
18074 * \brief Returns a surface object's resource descriptor
18075 *
18076 * Returns the resource descriptor for the surface object specified by \p surfObject.
18077 *
18078 * \param pResDesc - Resource descriptor
18079 * \param surfObject - Surface object
18080 *
18081 * \return
18082 * ::CUDA_SUCCESS,
18083 * ::CUDA_ERROR_DEINITIALIZED,
18084 * ::CUDA_ERROR_NOT_INITIALIZED,
18085 * ::CUDA_ERROR_INVALID_CONTEXT,
18086 * ::CUDA_ERROR_INVALID_VALUE
18087 *
18088 * \sa
18089 * ::cuSurfObjectCreate,
18090 * ::cudaGetSurfaceObjectResourceDesc
18091 */
18092CUresult CUDAAPI cuSurfObjectGetResourceDesc(CUDA_RESOURCE_DESC *pResDesc, CUsurfObject surfObject);
18093
18094/** @} */ /* END CUDA_SURFOBJECT */
18095
18096/**
18097 * \defgroup CUDA_PEER_ACCESS Peer Context Memory Access
18098 *
18099 * ___MANBRIEF___ direct peer context memory access functions of the low-level
18100 * CUDA driver API (___CURRENT_FILE___) ___ENDMANBRIEF___
18101 *
18102 * This section describes the direct peer context memory access functions
18103 * of the low-level CUDA driver application programming interface.
18104 *
18105 * @{
18106 */
18107
18108/**
18109 * \brief Queries if a device may directly access a peer device's memory.
18110 *
18111 * Returns in \p *canAccessPeer a value of 1 if contexts on \p dev are capable of
18112 * directly accessing memory from contexts on \p peerDev and 0 otherwise.
18113 * If direct access of \p peerDev from \p dev is possible, then access may be
18114 * enabled on two specific contexts by calling ::cuCtxEnablePeerAccess().
18115 *
18116 * \param canAccessPeer - Returned access capability
18117 * \param dev - Device from which allocations on \p peerDev are to
18118 * be directly accessed.
18119 * \param peerDev - Device on which the allocations to be directly accessed
18120 * by \p dev reside.
18121 *
18122 * \return
18123 * ::CUDA_SUCCESS,
18124 * ::CUDA_ERROR_DEINITIALIZED,
18125 * ::CUDA_ERROR_NOT_INITIALIZED,
18126 * ::CUDA_ERROR_INVALID_DEVICE
18127 * \notefnerr
18128 *
18129 * \sa
18130 * ::cuCtxEnablePeerAccess,
18131 * ::cuCtxDisablePeerAccess,
18132 * ::cudaDeviceCanAccessPeer
18133 */
18134CUresult CUDAAPI cuDeviceCanAccessPeer(int *canAccessPeer, CUdevice dev, CUdevice peerDev);
18135
18136/**
18137 * \brief Enables direct access to memory allocations in a peer context.
18138 *
18139 * If both the current context and \p peerContext are on devices which support unified
18140 * addressing (as may be queried using ::CU_DEVICE_ATTRIBUTE_UNIFIED_ADDRESSING) and same
18141 * major compute capability, then on success all allocations from \p peerContext will
18142 * immediately be accessible by the current context. See \ref CUDA_UNIFIED for additional
18143 * details.
18144 *
18145 * Note that access granted by this call is unidirectional and that in order to access
18146 * memory from the current context in \p peerContext, a separate symmetric call
18147 * to ::cuCtxEnablePeerAccess() is required.
18148 *
18149 * Note that there are both device-wide and system-wide limitations per system
18150 * configuration, as noted in the CUDA Programming Guide under the section
18151 * "Peer-to-Peer Memory Access".
18152 *
18153 * Returns ::CUDA_ERROR_PEER_ACCESS_UNSUPPORTED if ::cuDeviceCanAccessPeer() indicates
18154 * that the ::CUdevice of the current context cannot directly access memory
18155 * from the ::CUdevice of \p peerContext.
18156 *
18157 * Returns ::CUDA_ERROR_PEER_ACCESS_ALREADY_ENABLED if direct access of
18158 * \p peerContext from the current context has already been enabled.
18159 *
18160 * Returns ::CUDA_ERROR_TOO_MANY_PEERS if direct peer access is not possible
18161 * because hardware resources required for peer access have been exhausted.
18162 *
18163 * Returns ::CUDA_ERROR_INVALID_CONTEXT if there is no current context, \p peerContext
18164 * is not a valid context, or if the current context is \p peerContext.
18165 *
18166 * Returns ::CUDA_ERROR_INVALID_VALUE if \p Flags is not 0.
18167 *
18168 * \param peerContext - Peer context to enable direct access to from the current context
18169 * \param Flags - Reserved for future use and must be set to 0
18170 *
18171 * \return
18172 * ::CUDA_SUCCESS,
18173 * ::CUDA_ERROR_DEINITIALIZED,
18174 * ::CUDA_ERROR_NOT_INITIALIZED,
18175 * ::CUDA_ERROR_PEER_ACCESS_ALREADY_ENABLED,
18176 * ::CUDA_ERROR_TOO_MANY_PEERS,
18177 * ::CUDA_ERROR_INVALID_CONTEXT,
18178 * ::CUDA_ERROR_PEER_ACCESS_UNSUPPORTED,
18179 * ::CUDA_ERROR_INVALID_VALUE
18180 * \notefnerr
18181 *
18182 * \sa
18183 * ::cuDeviceCanAccessPeer,
18184 * ::cuCtxDisablePeerAccess,
18185 * ::cudaDeviceEnablePeerAccess
18186 */
18187CUresult CUDAAPI cuCtxEnablePeerAccess(CUcontext peerContext, unsigned int Flags);
18188
18189/**
18190 * \brief Disables direct access to memory allocations in a peer context and
18191 * unregisters any registered allocations.
18192 *
18193 Returns ::CUDA_ERROR_PEER_ACCESS_NOT_ENABLED if direct peer access has
18194 * not yet been enabled from \p peerContext to the current context.
18195 *
18196 * Returns ::CUDA_ERROR_INVALID_CONTEXT if there is no current context, or if
18197 * \p peerContext is not a valid context.
18198 *
18199 * \param peerContext - Peer context to disable direct access to
18200 *
18201 * \return
18202 * ::CUDA_SUCCESS,
18203 * ::CUDA_ERROR_DEINITIALIZED,
18204 * ::CUDA_ERROR_NOT_INITIALIZED,
18205 * ::CUDA_ERROR_PEER_ACCESS_NOT_ENABLED,
18206 * ::CUDA_ERROR_INVALID_CONTEXT,
18207 * \notefnerr
18208 *
18209 * \sa
18210 * ::cuDeviceCanAccessPeer,
18211 * ::cuCtxEnablePeerAccess,
18212 * ::cudaDeviceDisablePeerAccess
18213 */
18214CUresult CUDAAPI cuCtxDisablePeerAccess(CUcontext peerContext);
18215
18216/**
18217 * \brief Queries attributes of the link between two devices.
18218 *
18219 * Returns in \p *value the value of the requested attribute \p attrib of the
18220 * link between \p srcDevice and \p dstDevice. The supported attributes are:
18221 * - ::CU_DEVICE_P2P_ATTRIBUTE_PERFORMANCE_RANK: A relative value indicating the
18222 * performance of the link between two devices.
18223 * - ::CU_DEVICE_P2P_ATTRIBUTE_ACCESS_SUPPORTED P2P: 1 if P2P Access is enable.
18224 * - ::CU_DEVICE_P2P_ATTRIBUTE_NATIVE_ATOMIC_SUPPORTED: 1 if Atomic operations over
18225 * the link are supported.
18226 * - ::CU_DEVICE_P2P_ATTRIBUTE_CUDA_ARRAY_ACCESS_SUPPORTED: 1 if cudaArray can
18227 * be accessed over the link.
18228 *
18229 * Returns ::CUDA_ERROR_INVALID_DEVICE if \p srcDevice or \p dstDevice are not valid
18230 * or if they represent the same device.
18231 *
18232 * Returns ::CUDA_ERROR_INVALID_VALUE if \p attrib is not valid or if \p value is
18233 * a null pointer.
18234 *
18235 * \param value - Returned value of the requested attribute
18236 * \param attrib - The requested attribute of the link between \p srcDevice and \p dstDevice.
18237 * \param srcDevice - The source device of the target link.
18238 * \param dstDevice - The destination device of the target link.
18239 *
18240 * \return
18241 * ::CUDA_SUCCESS,
18242 * ::CUDA_ERROR_DEINITIALIZED,
18243 * ::CUDA_ERROR_NOT_INITIALIZED,
18244 * ::CUDA_ERROR_INVALID_DEVICE,
18245 * ::CUDA_ERROR_INVALID_VALUE
18246 * \notefnerr
18247 *
18248 * \sa
18249 * ::cuCtxEnablePeerAccess,
18250 * ::cuCtxDisablePeerAccess,
18251 * ::cuDeviceCanAccessPeer,
18252 * ::cudaDeviceGetP2PAttribute
18253 */
18254CUresult CUDAAPI cuDeviceGetP2PAttribute(int* value, CUdevice_P2PAttribute attrib, CUdevice srcDevice, CUdevice dstDevice);
18255
18256/** @} */ /* END CUDA_PEER_ACCESS */
18257
18258/**
18259 * \defgroup CUDA_GRAPHICS Graphics Interoperability
18260 *
18261 * ___MANBRIEF___ graphics interoperability functions of the low-level CUDA
18262 * driver API (___CURRENT_FILE___) ___ENDMANBRIEF___
18263 *
18264 * This section describes the graphics interoperability functions of the
18265 * low-level CUDA driver application programming interface.
18266 *
18267 * @{
18268 */
18269
18270/**
18271 * \brief Unregisters a graphics resource for access by CUDA
18272 *
18273 * Unregisters the graphics resource \p resource so it is not accessible by
18274 * CUDA unless registered again.
18275 *
18276 * If \p resource is invalid then ::CUDA_ERROR_INVALID_HANDLE is
18277 * returned.
18278 *
18279 * \param resource - Resource to unregister
18280 *
18281 * \return
18282 * ::CUDA_SUCCESS,
18283 * ::CUDA_ERROR_DEINITIALIZED,
18284 * ::CUDA_ERROR_NOT_INITIALIZED,
18285 * ::CUDA_ERROR_INVALID_CONTEXT,
18286 * ::CUDA_ERROR_INVALID_HANDLE,
18287 * ::CUDA_ERROR_UNKNOWN
18288 * \notefnerr
18289 *
18290 * \sa
18291 * ::cuGraphicsD3D9RegisterResource,
18292 * ::cuGraphicsD3D10RegisterResource,
18293 * ::cuGraphicsD3D11RegisterResource,
18294 * ::cuGraphicsGLRegisterBuffer,
18295 * ::cuGraphicsGLRegisterImage,
18296 * ::cudaGraphicsUnregisterResource
18297 */
18298CUresult CUDAAPI cuGraphicsUnregisterResource(CUgraphicsResource resource);
18299
18300/**
18301 * \brief Get an array through which to access a subresource of a mapped graphics resource.
18302 *
18303 * Returns in \p *pArray an array through which the subresource of the mapped
18304 * graphics resource \p resource which corresponds to array index \p arrayIndex
18305 * and mipmap level \p mipLevel may be accessed. The value set in \p *pArray may
18306 * change every time that \p resource is mapped.
18307 *
18308 * If \p resource is not a texture then it cannot be accessed via an array and
18309 * ::CUDA_ERROR_NOT_MAPPED_AS_ARRAY is returned.
18310 * If \p arrayIndex is not a valid array index for \p resource then
18311 * ::CUDA_ERROR_INVALID_VALUE is returned.
18312 * If \p mipLevel is not a valid mipmap level for \p resource then
18313 * ::CUDA_ERROR_INVALID_VALUE is returned.
18314 * If \p resource is not mapped then ::CUDA_ERROR_NOT_MAPPED is returned.
18315 *
18316 * \param pArray - Returned array through which a subresource of \p resource may be accessed
18317 * \param resource - Mapped resource to access
18318 * \param arrayIndex - Array index for array textures or cubemap face
18319 * index as defined by ::CUarray_cubemap_face for
18320 * cubemap textures for the subresource to access
18321 * \param mipLevel - Mipmap level for the subresource to access
18322 *
18323 * \return
18324 * ::CUDA_SUCCESS,
18325 * ::CUDA_ERROR_DEINITIALIZED,
18326 * ::CUDA_ERROR_NOT_INITIALIZED,
18327 * ::CUDA_ERROR_INVALID_CONTEXT,
18328 * ::CUDA_ERROR_INVALID_VALUE,
18329 * ::CUDA_ERROR_INVALID_HANDLE,
18330 * ::CUDA_ERROR_NOT_MAPPED,
18331 * ::CUDA_ERROR_NOT_MAPPED_AS_ARRAY
18332 * \notefnerr
18333 *
18334 * \sa
18335 * ::cuGraphicsResourceGetMappedPointer,
18336 * ::cudaGraphicsSubResourceGetMappedArray
18337 */
18338CUresult CUDAAPI cuGraphicsSubResourceGetMappedArray(CUarray *pArray, CUgraphicsResource resource, unsigned int arrayIndex, unsigned int mipLevel);
18339
18340/**
18341 * \brief Get a mipmapped array through which to access a mapped graphics resource.
18342 *
18343 * Returns in \p *pMipmappedArray a mipmapped array through which the mapped graphics
18344 * resource \p resource. The value set in \p *pMipmappedArray may change every time
18345 * that \p resource is mapped.
18346 *
18347 * If \p resource is not a texture then it cannot be accessed via a mipmapped array and
18348 * ::CUDA_ERROR_NOT_MAPPED_AS_ARRAY is returned.
18349 * If \p resource is not mapped then ::CUDA_ERROR_NOT_MAPPED is returned.
18350 *
18351 * \param pMipmappedArray - Returned mipmapped array through which \p resource may be accessed
18352 * \param resource - Mapped resource to access
18353 *
18354 * \return
18355 * ::CUDA_SUCCESS,
18356 * ::CUDA_ERROR_DEINITIALIZED,
18357 * ::CUDA_ERROR_NOT_INITIALIZED,
18358 * ::CUDA_ERROR_INVALID_CONTEXT,
18359 * ::CUDA_ERROR_INVALID_VALUE,
18360 * ::CUDA_ERROR_INVALID_HANDLE,
18361 * ::CUDA_ERROR_NOT_MAPPED,
18362 * ::CUDA_ERROR_NOT_MAPPED_AS_ARRAY
18363 * \notefnerr
18364 *
18365 * \sa
18366 * ::cuGraphicsResourceGetMappedPointer,
18367 * ::cudaGraphicsResourceGetMappedMipmappedArray
18368 */
18369CUresult CUDAAPI cuGraphicsResourceGetMappedMipmappedArray(CUmipmappedArray *pMipmappedArray, CUgraphicsResource resource);
18370
18371/**
18372 * \brief Get a device pointer through which to access a mapped graphics resource.
18373 *
18374 * Returns in \p *pDevPtr a pointer through which the mapped graphics resource
18375 * \p resource may be accessed.
18376 * Returns in \p pSize the size of the memory in bytes which may be accessed from that pointer.
18377 * The value set in \p pPointer may change every time that \p resource is mapped.
18378 *
18379 * If \p resource is not a buffer then it cannot be accessed via a pointer and
18380 * ::CUDA_ERROR_NOT_MAPPED_AS_POINTER is returned.
18381 * If \p resource is not mapped then ::CUDA_ERROR_NOT_MAPPED is returned.
18382 * *
18383 * \param pDevPtr - Returned pointer through which \p resource may be accessed
18384 * \param pSize - Returned size of the buffer accessible starting at \p *pPointer
18385 * \param resource - Mapped resource to access
18386 *
18387 * \return
18388 * ::CUDA_SUCCESS,
18389 * ::CUDA_ERROR_DEINITIALIZED,
18390 * ::CUDA_ERROR_NOT_INITIALIZED,
18391 * ::CUDA_ERROR_INVALID_CONTEXT,
18392 * ::CUDA_ERROR_INVALID_VALUE,
18393 * ::CUDA_ERROR_INVALID_HANDLE,
18394 * ::CUDA_ERROR_NOT_MAPPED,
18395 * ::CUDA_ERROR_NOT_MAPPED_AS_POINTER
18396 * \notefnerr
18397 *
18398 * \sa
18399 * ::cuGraphicsMapResources,
18400 * ::cuGraphicsSubResourceGetMappedArray,
18401 * ::cudaGraphicsResourceGetMappedPointer
18402 */
18403CUresult CUDAAPI cuGraphicsResourceGetMappedPointer(CUdeviceptr *pDevPtr, size_t *pSize, CUgraphicsResource resource);
18404
18405/**
18406 * \brief Set usage flags for mapping a graphics resource
18407 *
18408 * Set \p flags for mapping the graphics resource \p resource.
18409 *
18410 * Changes to \p flags will take effect the next time \p resource is mapped.
18411 * The \p flags argument may be any of the following:
18412
18413 * - ::CU_GRAPHICS_MAP_RESOURCE_FLAGS_NONE: Specifies no hints about how this
18414 * resource will be used. It is therefore assumed that this resource will be
18415 * read from and written to by CUDA kernels. This is the default value.
18416 * - ::CU_GRAPHICS_MAP_RESOURCE_FLAGS_READONLY: Specifies that CUDA kernels which
18417 * access this resource will not write to this resource.
18418 * - ::CU_GRAPHICS_MAP_RESOURCE_FLAGS_WRITEDISCARD: Specifies that CUDA kernels
18419 * which access this resource will not read from this resource and will
18420 * write over the entire contents of the resource, so none of the data
18421 * previously stored in the resource will be preserved.
18422 *
18423 * If \p resource is presently mapped for access by CUDA then
18424 * ::CUDA_ERROR_ALREADY_MAPPED is returned.
18425 * If \p flags is not one of the above values then ::CUDA_ERROR_INVALID_VALUE is returned.
18426 *
18427 * \param resource - Registered resource to set flags for
18428 * \param flags - Parameters for resource mapping
18429 *
18430 * \return
18431 * ::CUDA_SUCCESS,
18432 * ::CUDA_ERROR_DEINITIALIZED,
18433 * ::CUDA_ERROR_NOT_INITIALIZED,
18434 * ::CUDA_ERROR_INVALID_CONTEXT,
18435 * ::CUDA_ERROR_INVALID_VALUE,
18436 * ::CUDA_ERROR_INVALID_HANDLE,
18437 * ::CUDA_ERROR_ALREADY_MAPPED
18438 * \notefnerr
18439 *
18440 * \sa
18441 * ::cuGraphicsMapResources,
18442 * ::cudaGraphicsResourceSetMapFlags
18443 */
18444CUresult CUDAAPI cuGraphicsResourceSetMapFlags(CUgraphicsResource resource, unsigned int flags);
18445
18446/**
18447 * \brief Map graphics resources for access by CUDA
18448 *
18449 * Maps the \p count graphics resources in \p resources for access by CUDA.
18450 *
18451 * The resources in \p resources may be accessed by CUDA until they
18452 * are unmapped. The graphics API from which \p resources were registered
18453 * should not access any resources while they are mapped by CUDA. If an
18454 * application does so, the results are undefined.
18455 *
18456 * This function provides the synchronization guarantee that any graphics calls
18457 * issued before ::cuGraphicsMapResources() will complete before any subsequent CUDA
18458 * work issued in \p stream begins.
18459 *
18460 * If \p resources includes any duplicate entries then ::CUDA_ERROR_INVALID_HANDLE is returned.
18461 * If any of \p resources are presently mapped for access by CUDA then ::CUDA_ERROR_ALREADY_MAPPED is returned.
18462 *
18463 * \param count - Number of resources to map
18464 * \param resources - Resources to map for CUDA usage
18465 * \param hStream - Stream with which to synchronize
18466 *
18467 * \return
18468 * ::CUDA_SUCCESS,
18469 * ::CUDA_ERROR_DEINITIALIZED,
18470 * ::CUDA_ERROR_NOT_INITIALIZED,
18471 * ::CUDA_ERROR_INVALID_CONTEXT,
18472 * ::CUDA_ERROR_INVALID_HANDLE,
18473 * ::CUDA_ERROR_ALREADY_MAPPED,
18474 * ::CUDA_ERROR_UNKNOWN
18475 * \note_null_stream
18476 * \notefnerr
18477 *
18478 * \sa
18479 * ::cuGraphicsResourceGetMappedPointer,
18480 * ::cuGraphicsSubResourceGetMappedArray,
18481 * ::cuGraphicsUnmapResources,
18482 * ::cudaGraphicsMapResources
18483 */
18484CUresult CUDAAPI cuGraphicsMapResources(unsigned int count, CUgraphicsResource *resources, CUstream hStream);
18485
18486/**
18487 * \brief Unmap graphics resources.
18488 *
18489 * Unmaps the \p count graphics resources in \p resources.
18490 *
18491 * Once unmapped, the resources in \p resources may not be accessed by CUDA
18492 * until they are mapped again.
18493 *
18494 * This function provides the synchronization guarantee that any CUDA work issued
18495 * in \p stream before ::cuGraphicsUnmapResources() will complete before any
18496 * subsequently issued graphics work begins.
18497 *
18498 *
18499 * If \p resources includes any duplicate entries then ::CUDA_ERROR_INVALID_HANDLE is returned.
18500 * If any of \p resources are not presently mapped for access by CUDA then ::CUDA_ERROR_NOT_MAPPED is returned.
18501 *
18502 * \param count - Number of resources to unmap
18503 * \param resources - Resources to unmap
18504 * \param hStream - Stream with which to synchronize
18505 *
18506 * \return
18507 * ::CUDA_SUCCESS,
18508 * ::CUDA_ERROR_DEINITIALIZED,
18509 * ::CUDA_ERROR_NOT_INITIALIZED,
18510 * ::CUDA_ERROR_INVALID_CONTEXT,
18511 * ::CUDA_ERROR_INVALID_HANDLE,
18512 * ::CUDA_ERROR_NOT_MAPPED,
18513 * ::CUDA_ERROR_UNKNOWN
18514 * \note_null_stream
18515 * \notefnerr
18516 *
18517 * \sa
18518 * ::cuGraphicsMapResources,
18519 * ::cudaGraphicsUnmapResources
18520 */
18521CUresult CUDAAPI cuGraphicsUnmapResources(unsigned int count, CUgraphicsResource *resources, CUstream hStream);
18522
18523/** @} */ /* END CUDA_GRAPHICS */
18524
18525/**
18526 * \defgroup CUDA_DRIVER_ENTRY_POINT Driver Entry Point Access
18527 *
18528 * ___MANBRIEF___ driver entry point access functions of the low-level CUDA driver API
18529 * (___CURRENT_FILE___) ___ENDMANBRIEF___
18530 *
18531 * This section describes the driver entry point access functions of the low-level CUDA
18532 * driver application programming interface.
18533 *
18534 * @{
18535 */
18536
18537/**
18538 * \brief Returns the requested driver API function pointer
18539 *
18540 * Returns in \p **pfn the address of the CUDA driver function for the requested
18541 * CUDA version and flags.
18542 *
18543 * The CUDA version is specified as (1000 * major + 10 * minor), so CUDA 11.2
18544 * should be specified as 11020. For a requested driver symbol, if the specified
18545 * CUDA version is greater than or equal to the CUDA version in which the driver symbol
18546 * was introduced, this API will return the function pointer to the corresponding
18547 * versioned function.
18548 *
18549 * The pointer returned by the API should be cast to a function pointer matching the
18550 * requested driver function's definition in the API header file. The function pointer
18551 * typedef can be picked up from the corresponding typedefs header file. For example,
18552 * cudaTypedefs.h consists of function pointer typedefs for driver APIs defined in cuda.h.
18553 *
18554 * The API will return ::CUDA_ERROR_NOT_FOUND if the requested driver function is not
18555 * supported on the platform, no ABI compatible driver function exists for the specified
18556 * \p cudaVersion or if the driver symbol is invalid.
18557 *
18558 * The requested flags can be:
18559 * - ::CU_GET_PROC_ADDRESS_DEFAULT: This is the default mode. This is equivalent to
18560 * ::CU_GET_PROC_ADDRESS_PER_THREAD_DEFAULT_STREAM if the code is compiled with
18561 * --default-stream per-thread compilation flag or the macro CUDA_API_PER_THREAD_DEFAULT_STREAM
18562 * is defined; ::CU_GET_PROC_ADDRESS_LEGACY_STREAM otherwise.
18563 * - ::CU_GET_PROC_ADDRESS_LEGACY_STREAM: This will enable the search for all driver symbols
18564 * that match the requested driver symbol name except the corresponding per-thread versions.
18565 * - ::CU_GET_PROC_ADDRESS_PER_THREAD_DEFAULT_STREAM: This will enable the search for all
18566 * driver symbols that match the requested driver symbol name including the per-thread
18567 * versions. If a per-thread version is not found, the API will return the legacy version
18568 * of the driver function.
18569 *
18570 * \param symbol - The base name of the driver API function to look for. As an example,
18571 * for the driver API ::cuMemAlloc_v2, \p symbol would be cuMemAlloc and
18572 * \p cudaVersion would be the ABI compatible CUDA version for the _v2 variant.
18573 * \param pfn - Location to return the function pointer to the requested driver function
18574 * \param cudaVersion - The CUDA version to look for the requested driver symbol
18575 * \param flags - Flags to specify search options.
18576 *
18577 * \return
18578 * ::CUDA_SUCCESS,
18579 * ::CUDA_ERROR_INVALID_VALUE,
18580 * ::CUDA_ERROR_NOT_SUPPORTED,
18581 * ::CUDA_ERROR_NOT_FOUND
18582 * \note_version_mixing
18583 *
18584 * \sa
18585 * ::cudaGetDriverEntryPoint
18586 */
18587CUresult CUDAAPI cuGetProcAddress(const char *symbol, void **pfn, int cudaVersion, cuuint64_t flags);
18588
18589/** @} */ /* END CUDA_DRIVER_ENTRY_POINT */
18590
18591CUresult CUDAAPI cuGetExportTable(const void **ppExportTable, const CUuuid *pExportTableId);
18592
18593/**
18594 * CUDA API versioning support
18595 */
18596#if defined(__CUDA_API_VERSION_INTERNAL)
18597 #undef cuMemHostRegister
18598 #undef cuGraphicsResourceSetMapFlags
18599 #undef cuLinkCreate
18600 #undef cuLinkAddData
18601 #undef cuLinkAddFile
18602 #undef cuDeviceTotalMem
18603 #undef cuCtxCreate
18604 #undef cuModuleGetGlobal
18605 #undef cuMemGetInfo
18606 #undef cuMemAlloc
18607 #undef cuMemAllocPitch
18608 #undef cuMemFree
18609 #undef cuMemGetAddressRange
18610 #undef cuMemAllocHost
18611 #undef cuMemHostGetDevicePointer
18612 #undef cuMemcpyHtoD
18613 #undef cuMemcpyDtoH
18614 #undef cuMemcpyDtoD
18615 #undef cuMemcpyDtoA
18616 #undef cuMemcpyAtoD
18617 #undef cuMemcpyHtoA
18618 #undef cuMemcpyAtoH
18619 #undef cuMemcpyAtoA
18620 #undef cuMemcpyHtoAAsync
18621 #undef cuMemcpyAtoHAsync
18622 #undef cuMemcpy2D
18623 #undef cuMemcpy2DUnaligned
18624 #undef cuMemcpy3D
18625 #undef cuMemcpyHtoDAsync
18626 #undef cuMemcpyDtoHAsync
18627 #undef cuMemcpyDtoDAsync
18628 #undef cuMemcpy2DAsync
18629 #undef cuMemcpy3DAsync
18630 #undef cuMemsetD8
18631 #undef cuMemsetD16
18632 #undef cuMemsetD32
18633 #undef cuMemsetD2D8
18634 #undef cuMemsetD2D16
18635 #undef cuMemsetD2D32
18636 #undef cuArrayCreate
18637 #undef cuArrayGetDescriptor
18638 #undef cuArray3DCreate
18639 #undef cuArray3DGetDescriptor
18640 #undef cuTexRefSetAddress
18641 #undef cuTexRefSetAddress2D
18642 #undef cuTexRefGetAddress
18643 #undef cuGraphicsResourceGetMappedPointer
18644 #undef cuCtxDestroy
18645 #undef cuCtxPopCurrent
18646 #undef cuCtxPushCurrent
18647 #undef cuStreamDestroy
18648 #undef cuEventDestroy
18649 #undef cuMemcpy
18650 #undef cuMemcpyAsync
18651 #undef cuMemcpyPeer
18652 #undef cuMemcpyPeerAsync
18653 #undef cuMemcpy3DPeer
18654 #undef cuMemcpy3DPeerAsync
18655 #undef cuMemsetD8Async
18656 #undef cuMemsetD16Async
18657 #undef cuMemsetD32Async
18658 #undef cuMemsetD2D8Async
18659 #undef cuMemsetD2D16Async
18660 #undef cuMemsetD2D32Async
18661 #undef cuStreamGetPriority
18662 #undef cuStreamGetFlags
18663 #undef cuStreamGetCtx
18664 #undef cuStreamWaitEvent
18665 #undef cuStreamAddCallback
18666 #undef cuStreamAttachMemAsync
18667 #undef cuStreamQuery
18668 #undef cuStreamSynchronize
18669 #undef cuEventRecord
18670 #undef cuEventRecordWithFlags
18671 #undef cuLaunchKernel
18672 #undef cuLaunchHostFunc
18673 #undef cuGraphicsMapResources
18674 #undef cuGraphicsUnmapResources
18675 #undef cuStreamWriteValue32
18676 #undef cuStreamWaitValue32
18677 #undef cuStreamWriteValue64
18678 #undef cuStreamWaitValue64
18679 #undef cuStreamBatchMemOp
18680 #undef cuMemPrefetchAsync
18681 #undef cuLaunchCooperativeKernel
18682 #undef cuSignalExternalSemaphoresAsync
18683 #undef cuWaitExternalSemaphoresAsync
18684 #undef cuStreamBeginCapture
18685 #undef cuStreamEndCapture
18686 #undef cuStreamIsCapturing
18687 #undef cuStreamGetCaptureInfo
18688 #undef cuStreamGetCaptureInfo_v2
18689 #undef cuGraphUpload
18690 #undef cuGraphLaunch
18691 #undef cuDevicePrimaryCtxRelease
18692 #undef cuDevicePrimaryCtxReset
18693 #undef cuDevicePrimaryCtxSetFlags
18694 #undef cuIpcOpenMemHandle
18695 #undef cuStreamCopyAttributes
18696 #undef cuStreamSetAttribute
18697 #undef cuStreamGetAttribute
18698 #undef cuGraphInstantiate
18699 #undef cuMemMapArrayAsync
18700 #undef cuMemFreeAsync
18701 #undef cuMemAllocAsync
18702 #undef cuMemAllocFromPoolAsync
18703 #undef cuStreamUpdateCaptureDependencies
18704
18705 CUresult CUDAAPI cuMemHostRegister(void *p, size_t bytesize, unsigned int Flags);
18706 CUresult CUDAAPI cuGraphicsResourceSetMapFlags(CUgraphicsResource resource, unsigned int flags);
18707 CUresult CUDAAPI cuLinkCreate(unsigned int numOptions, CUjit_option *options, void **optionValues, CUlinkState *stateOut);
18708 CUresult CUDAAPI cuLinkAddData(CUlinkState state, CUjitInputType type, void *data, size_t size, const char *name,
18709 unsigned int numOptions, CUjit_option *options, void **optionValues);
18710 CUresult CUDAAPI cuLinkAddFile(CUlinkState state, CUjitInputType type, const char *path,
18711 unsigned int numOptions, CUjit_option *options, void **optionValues);
18712 CUresult CUDAAPI cuTexRefSetAddress2D_v2(CUtexref hTexRef, const CUDA_ARRAY_DESCRIPTOR *desc, CUdeviceptr dptr, size_t Pitch);
18713
18714 typedef unsigned int CUdeviceptr_v1;
18715
18716 typedef struct CUDA_MEMCPY2D_v1_st
18717 {
18718 unsigned int srcXInBytes; /**< Source X in bytes */
18719 unsigned int srcY; /**< Source Y */
18720 CUmemorytype srcMemoryType; /**< Source memory type (host, device, array) */
18721 const void *srcHost; /**< Source host pointer */
18722 CUdeviceptr_v1 srcDevice; /**< Source device pointer */
18723 CUarray srcArray; /**< Source array reference */
18724 unsigned int srcPitch; /**< Source pitch (ignored when src is array) */
18725
18726 unsigned int dstXInBytes; /**< Destination X in bytes */
18727 unsigned int dstY; /**< Destination Y */
18728 CUmemorytype dstMemoryType; /**< Destination memory type (host, device, array) */
18729 void *dstHost; /**< Destination host pointer */
18730 CUdeviceptr_v1 dstDevice; /**< Destination device pointer */
18731 CUarray dstArray; /**< Destination array reference */
18732 unsigned int dstPitch; /**< Destination pitch (ignored when dst is array) */
18733
18734 unsigned int WidthInBytes; /**< Width of 2D memory copy in bytes */
18735 unsigned int Height; /**< Height of 2D memory copy */
18736 } CUDA_MEMCPY2D_v1;
18737
18738 typedef struct CUDA_MEMCPY3D_v1_st
18739 {
18740 unsigned int srcXInBytes; /**< Source X in bytes */
18741 unsigned int srcY; /**< Source Y */
18742 unsigned int srcZ; /**< Source Z */
18743 unsigned int srcLOD; /**< Source LOD */
18744 CUmemorytype srcMemoryType; /**< Source memory type (host, device, array) */
18745 const void *srcHost; /**< Source host pointer */
18746 CUdeviceptr_v1 srcDevice; /**< Source device pointer */
18747 CUarray srcArray; /**< Source array reference */
18748 void *reserved0; /**< Must be NULL */
18749 unsigned int srcPitch; /**< Source pitch (ignored when src is array) */
18750 unsigned int srcHeight; /**< Source height (ignored when src is array; may be 0 if Depth==1) */
18751
18752 unsigned int dstXInBytes; /**< Destination X in bytes */
18753 unsigned int dstY; /**< Destination Y */
18754 unsigned int dstZ; /**< Destination Z */
18755 unsigned int dstLOD; /**< Destination LOD */
18756 CUmemorytype dstMemoryType; /**< Destination memory type (host, device, array) */
18757 void *dstHost; /**< Destination host pointer */
18758 CUdeviceptr_v1 dstDevice; /**< Destination device pointer */
18759 CUarray dstArray; /**< Destination array reference */
18760 void *reserved1; /**< Must be NULL */
18761 unsigned int dstPitch; /**< Destination pitch (ignored when dst is array) */
18762 unsigned int dstHeight; /**< Destination height (ignored when dst is array; may be 0 if Depth==1) */
18763
18764 unsigned int WidthInBytes; /**< Width of 3D memory copy in bytes */
18765 unsigned int Height; /**< Height of 3D memory copy */
18766 unsigned int Depth; /**< Depth of 3D memory copy */
18767 } CUDA_MEMCPY3D_v1;
18768
18769 typedef struct CUDA_ARRAY_DESCRIPTOR_v1_st
18770 {
18771 unsigned int Width; /**< Width of array */
18772 unsigned int Height; /**< Height of array */
18773
18774 CUarray_format Format; /**< Array format */
18775 unsigned int NumChannels; /**< Channels per array element */
18776 } CUDA_ARRAY_DESCRIPTOR_v1;
18777
18778 typedef struct CUDA_ARRAY3D_DESCRIPTOR_v1_st
18779 {
18780 unsigned int Width; /**< Width of 3D array */
18781 unsigned int Height; /**< Height of 3D array */
18782 unsigned int Depth; /**< Depth of 3D array */
18783
18784 CUarray_format Format; /**< Array format */
18785 unsigned int NumChannels; /**< Channels per array element */
18786 unsigned int Flags; /**< Flags */
18787 } CUDA_ARRAY3D_DESCRIPTOR_v1;
18788
18789 CUresult CUDAAPI cuDeviceTotalMem(unsigned int *bytes, CUdevice dev);
18790 CUresult CUDAAPI cuCtxCreate(CUcontext *pctx, unsigned int flags, CUdevice dev);
18791 CUresult CUDAAPI cuModuleGetGlobal(CUdeviceptr_v1 *dptr, unsigned int *bytes, CUmodule hmod, const char *name);
18792 CUresult CUDAAPI cuMemGetInfo(unsigned int *free, unsigned int *total);
18793 CUresult CUDAAPI cuMemAlloc(CUdeviceptr_v1 *dptr, unsigned int bytesize);
18794 CUresult CUDAAPI cuMemAllocPitch(CUdeviceptr_v1 *dptr, unsigned int *pPitch, unsigned int WidthInBytes, unsigned int Height, unsigned int ElementSizeBytes);
18795 CUresult CUDAAPI cuMemFree(CUdeviceptr_v1 dptr);
18796 CUresult CUDAAPI cuMemGetAddressRange(CUdeviceptr_v1 *pbase, unsigned int *psize, CUdeviceptr_v1 dptr);
18797 CUresult CUDAAPI cuMemAllocHost(void **pp, unsigned int bytesize);
18798 CUresult CUDAAPI cuMemHostGetDevicePointer(CUdeviceptr_v1 *pdptr, void *p, unsigned int Flags);
18799 CUresult CUDAAPI cuMemcpyHtoD(CUdeviceptr_v1 dstDevice, const void *srcHost, unsigned int ByteCount);
18800 CUresult CUDAAPI cuMemcpyDtoH(void *dstHost, CUdeviceptr_v1 srcDevice, unsigned int ByteCount);
18801 CUresult CUDAAPI cuMemcpyDtoD(CUdeviceptr_v1 dstDevice, CUdeviceptr_v1 srcDevice, unsigned int ByteCount);
18802 CUresult CUDAAPI cuMemcpyDtoA(CUarray dstArray, unsigned int dstOffset, CUdeviceptr_v1 srcDevice, unsigned int ByteCount);
18803 CUresult CUDAAPI cuMemcpyAtoD(CUdeviceptr_v1 dstDevice, CUarray srcArray, unsigned int srcOffset, unsigned int ByteCount);
18804 CUresult CUDAAPI cuMemcpyHtoA(CUarray dstArray, unsigned int dstOffset, const void *srcHost, unsigned int ByteCount);
18805 CUresult CUDAAPI cuMemcpyAtoH(void *dstHost, CUarray srcArray, unsigned int srcOffset, unsigned int ByteCount);
18806 CUresult CUDAAPI cuMemcpyAtoA(CUarray dstArray, unsigned int dstOffset, CUarray srcArray, unsigned int srcOffset, unsigned int ByteCount);
18807 CUresult CUDAAPI cuMemcpyHtoAAsync(CUarray dstArray, unsigned int dstOffset, const void *srcHost, unsigned int ByteCount, CUstream hStream);
18808 CUresult CUDAAPI cuMemcpyAtoHAsync(void *dstHost, CUarray srcArray, unsigned int srcOffset, unsigned int ByteCount, CUstream hStream);
18809 CUresult CUDAAPI cuMemcpy2D(const CUDA_MEMCPY2D_v1 *pCopy);
18810 CUresult CUDAAPI cuMemcpy2DUnaligned(const CUDA_MEMCPY2D_v1 *pCopy);
18811 CUresult CUDAAPI cuMemcpy3D(const CUDA_MEMCPY3D_v1 *pCopy);
18812 CUresult CUDAAPI cuMemcpyHtoDAsync(CUdeviceptr_v1 dstDevice, const void *srcHost, unsigned int ByteCount, CUstream hStream);
18813 CUresult CUDAAPI cuMemcpyDtoHAsync(void *dstHost, CUdeviceptr_v1 srcDevice, unsigned int ByteCount, CUstream hStream);
18814 CUresult CUDAAPI cuMemcpyDtoDAsync(CUdeviceptr_v1 dstDevice, CUdeviceptr_v1 srcDevice, unsigned int ByteCount, CUstream hStream);
18815 CUresult CUDAAPI cuMemcpy2DAsync(const CUDA_MEMCPY2D_v1 *pCopy, CUstream hStream);
18816 CUresult CUDAAPI cuMemcpy3DAsync(const CUDA_MEMCPY3D_v1 *pCopy, CUstream hStream);
18817 CUresult CUDAAPI cuMemsetD8(CUdeviceptr_v1 dstDevice, unsigned char uc, unsigned int N);
18818 CUresult CUDAAPI cuMemsetD16(CUdeviceptr_v1 dstDevice, unsigned short us, unsigned int N);
18819 CUresult CUDAAPI cuMemsetD32(CUdeviceptr_v1 dstDevice, unsigned int ui, unsigned int N);
18820 CUresult CUDAAPI cuMemsetD2D8(CUdeviceptr_v1 dstDevice, unsigned int dstPitch, unsigned char uc, unsigned int Width, unsigned int Height);
18821 CUresult CUDAAPI cuMemsetD2D16(CUdeviceptr_v1 dstDevice, unsigned int dstPitch, unsigned short us, unsigned int Width, unsigned int Height);
18822 CUresult CUDAAPI cuMemsetD2D32(CUdeviceptr_v1 dstDevice, unsigned int dstPitch, unsigned int ui, unsigned int Width, unsigned int Height);
18823 CUresult CUDAAPI cuArrayCreate(CUarray *pHandle, const CUDA_ARRAY_DESCRIPTOR_v1 *pAllocateArray);
18824 CUresult CUDAAPI cuArrayGetDescriptor(CUDA_ARRAY_DESCRIPTOR_v1 *pArrayDescriptor, CUarray hArray);
18825 CUresult CUDAAPI cuArray3DCreate(CUarray *pHandle, const CUDA_ARRAY3D_DESCRIPTOR_v1 *pAllocateArray);
18826 CUresult CUDAAPI cuArray3DGetDescriptor(CUDA_ARRAY3D_DESCRIPTOR_v1 *pArrayDescriptor, CUarray hArray);
18827 CUresult CUDAAPI cuTexRefSetAddress(unsigned int *ByteOffset, CUtexref hTexRef, CUdeviceptr_v1 dptr, unsigned int bytes);
18828 CUresult CUDAAPI cuTexRefSetAddress2D(CUtexref hTexRef, const CUDA_ARRAY_DESCRIPTOR_v1 *desc, CUdeviceptr_v1 dptr, unsigned int Pitch);
18829 CUresult CUDAAPI cuTexRefGetAddress(CUdeviceptr_v1 *pdptr, CUtexref hTexRef);
18830 CUresult CUDAAPI cuGraphicsResourceGetMappedPointer(CUdeviceptr_v1 *pDevPtr, unsigned int *pSize, CUgraphicsResource resource);
18831
18832 CUresult CUDAAPI cuCtxDestroy(CUcontext ctx);
18833 CUresult CUDAAPI cuCtxPopCurrent(CUcontext *pctx);
18834 CUresult CUDAAPI cuCtxPushCurrent(CUcontext ctx);
18835 CUresult CUDAAPI cuStreamDestroy(CUstream hStream);
18836 CUresult CUDAAPI cuEventDestroy(CUevent hEvent);
18837 CUresult CUDAAPI cuDevicePrimaryCtxRelease(CUdevice dev);
18838 CUresult CUDAAPI cuDevicePrimaryCtxReset(CUdevice dev);
18839 CUresult CUDAAPI cuDevicePrimaryCtxSetFlags(CUdevice dev, unsigned int flags);
18840
18841 CUresult CUDAAPI cuMemcpyHtoD_v2(CUdeviceptr dstDevice, const void *srcHost, size_t ByteCount);
18842 CUresult CUDAAPI cuMemcpyDtoH_v2(void *dstHost, CUdeviceptr srcDevice, size_t ByteCount);
18843 CUresult CUDAAPI cuMemcpyDtoD_v2(CUdeviceptr dstDevice, CUdeviceptr srcDevice, size_t ByteCount);
18844 CUresult CUDAAPI cuMemcpyDtoA_v2(CUarray dstArray, size_t dstOffset, CUdeviceptr srcDevice, size_t ByteCount);
18845 CUresult CUDAAPI cuMemcpyAtoD_v2(CUdeviceptr dstDevice, CUarray srcArray, size_t srcOffset, size_t ByteCount);
18846 CUresult CUDAAPI cuMemcpyHtoA_v2(CUarray dstArray, size_t dstOffset, const void *srcHost, size_t ByteCount);
18847 CUresult CUDAAPI cuMemcpyAtoH_v2(void *dstHost, CUarray srcArray, size_t srcOffset, size_t ByteCount);
18848 CUresult CUDAAPI cuMemcpyAtoA_v2(CUarray dstArray, size_t dstOffset, CUarray srcArray, size_t srcOffset, size_t ByteCount);
18849 CUresult CUDAAPI cuMemcpyHtoAAsync_v2(CUarray dstArray, size_t dstOffset, const void *srcHost, size_t ByteCount, CUstream hStream);
18850 CUresult CUDAAPI cuMemcpyAtoHAsync_v2(void *dstHost, CUarray srcArray, size_t srcOffset, size_t ByteCount, CUstream hStream);
18851 CUresult CUDAAPI cuMemcpy2D_v2(const CUDA_MEMCPY2D *pCopy);
18852 CUresult CUDAAPI cuMemcpy2DUnaligned_v2(const CUDA_MEMCPY2D *pCopy);
18853 CUresult CUDAAPI cuMemcpy3D_v2(const CUDA_MEMCPY3D *pCopy);
18854 CUresult CUDAAPI cuMemcpyHtoDAsync_v2(CUdeviceptr dstDevice, const void *srcHost, size_t ByteCount, CUstream hStream);
18855 CUresult CUDAAPI cuMemcpyDtoHAsync_v2(void *dstHost, CUdeviceptr srcDevice, size_t ByteCount, CUstream hStream);
18856 CUresult CUDAAPI cuMemcpyDtoDAsync_v2(CUdeviceptr dstDevice, CUdeviceptr srcDevice, size_t ByteCount, CUstream hStream);
18857 CUresult CUDAAPI cuMemcpy2DAsync_v2(const CUDA_MEMCPY2D *pCopy, CUstream hStream);
18858 CUresult CUDAAPI cuMemcpy3DAsync_v2(const CUDA_MEMCPY3D *pCopy, CUstream hStream);
18859 CUresult CUDAAPI cuMemsetD8_v2(CUdeviceptr dstDevice, unsigned char uc, size_t N);
18860 CUresult CUDAAPI cuMemsetD16_v2(CUdeviceptr dstDevice, unsigned short us, size_t N);
18861 CUresult CUDAAPI cuMemsetD32_v2(CUdeviceptr dstDevice, unsigned int ui, size_t N);
18862 CUresult CUDAAPI cuMemsetD2D8_v2(CUdeviceptr dstDevice, size_t dstPitch, unsigned char uc, size_t Width, size_t Height);
18863 CUresult CUDAAPI cuMemsetD2D16_v2(CUdeviceptr dstDevice, size_t dstPitch, unsigned short us, size_t Width, size_t Height);
18864 CUresult CUDAAPI cuMemsetD2D32_v2(CUdeviceptr dstDevice, size_t dstPitch, unsigned int ui, size_t Width, size_t Height);
18865 CUresult CUDAAPI cuMemcpy(CUdeviceptr dst, CUdeviceptr src, size_t ByteCount);
18866 CUresult CUDAAPI cuMemcpyAsync(CUdeviceptr dst, CUdeviceptr src, size_t ByteCount, CUstream hStream);
18867 CUresult CUDAAPI cuMemcpyPeer(CUdeviceptr dstDevice, CUcontext dstContext, CUdeviceptr srcDevice, CUcontext srcContext, size_t ByteCount);
18868 CUresult CUDAAPI cuMemcpyPeerAsync(CUdeviceptr dstDevice, CUcontext dstContext, CUdeviceptr srcDevice, CUcontext srcContext, size_t ByteCount, CUstream hStream);
18869 CUresult CUDAAPI cuMemcpy3DPeer(const CUDA_MEMCPY3D_PEER *pCopy);
18870 CUresult CUDAAPI cuMemcpy3DPeerAsync(const CUDA_MEMCPY3D_PEER *pCopy, CUstream hStream);
18871
18872 CUresult CUDAAPI cuMemsetD8Async(CUdeviceptr dstDevice, unsigned char uc, size_t N, CUstream hStream);
18873 CUresult CUDAAPI cuMemsetD16Async(CUdeviceptr dstDevice, unsigned short us, size_t N, CUstream hStream);
18874 CUresult CUDAAPI cuMemsetD32Async(CUdeviceptr dstDevice, unsigned int ui, size_t N, CUstream hStream);
18875 CUresult CUDAAPI cuMemsetD2D8Async(CUdeviceptr dstDevice, size_t dstPitch, unsigned char uc, size_t Width, size_t Height, CUstream hStream);
18876 CUresult CUDAAPI cuMemsetD2D16Async(CUdeviceptr dstDevice, size_t dstPitch, unsigned short us, size_t Width, size_t Height, CUstream hStream);
18877 CUresult CUDAAPI cuMemsetD2D32Async(CUdeviceptr dstDevice, size_t dstPitch, unsigned int ui, size_t Width, size_t Height, CUstream hStream);
18878
18879 CUresult CUDAAPI cuStreamGetPriority(CUstream hStream, int *priority);
18880 CUresult CUDAAPI cuStreamGetFlags(CUstream hStream, unsigned int *flags);
18881 CUresult CUDAAPI cuStreamGetCtx(CUstream hStream, CUcontext *pctx);
18882 CUresult CUDAAPI cuStreamWaitEvent(CUstream hStream, CUevent hEvent, unsigned int Flags);
18883 CUresult CUDAAPI cuStreamAddCallback(CUstream hStream, CUstreamCallback callback, void *userData, unsigned int flags);
18884 CUresult CUDAAPI cuStreamAttachMemAsync(CUstream hStream, CUdeviceptr dptr, size_t length, unsigned int flags);
18885 CUresult CUDAAPI cuStreamQuery(CUstream hStream);
18886 CUresult CUDAAPI cuStreamSynchronize(CUstream hStream);
18887 CUresult CUDAAPI cuEventRecord(CUevent hEvent, CUstream hStream);
18888 CUresult CUDAAPI cuEventRecordWithFlags(CUevent hEvent, CUstream hStream, unsigned int flags);
18889 CUresult CUDAAPI cuLaunchKernel(CUfunction f, unsigned int gridDimX, unsigned int gridDimY, unsigned int gridDimZ, unsigned int blockDimX, unsigned int blockDimY, unsigned int blockDimZ, unsigned int sharedMemBytes, CUstream hStream, void **kernelParams, void **extra);
18890 CUresult CUDAAPI cuLaunchHostFunc(CUstream hStream, CUhostFn fn, void *userData);
18891 CUresult CUDAAPI cuGraphicsMapResources(unsigned int count, CUgraphicsResource *resources, CUstream hStream);
18892 CUresult CUDAAPI cuGraphicsUnmapResources(unsigned int count, CUgraphicsResource *resources, CUstream hStream);
18893 CUresult CUDAAPI cuStreamWriteValue32(CUstream stream, CUdeviceptr addr, cuuint32_t value, unsigned int flags);
18894 CUresult CUDAAPI cuStreamWaitValue32(CUstream stream, CUdeviceptr addr, cuuint32_t value, unsigned int flags);
18895 CUresult CUDAAPI cuStreamWriteValue64(CUstream stream, CUdeviceptr addr, cuuint64_t value, unsigned int flags);
18896 CUresult CUDAAPI cuStreamWaitValue64(CUstream stream, CUdeviceptr addr, cuuint64_t value, unsigned int flags);
18897 CUresult CUDAAPI cuStreamBatchMemOp(CUstream stream, unsigned int count, CUstreamBatchMemOpParams *paramArray, unsigned int flags);
18898 CUresult CUDAAPI cuMemPrefetchAsync(CUdeviceptr devPtr, size_t count, CUdevice dstDevice, CUstream hStream);
18899 CUresult CUDAAPI cuLaunchCooperativeKernel(CUfunction f, unsigned int gridDimX, unsigned int gridDimY, unsigned int gridDimZ, unsigned int blockDimX, unsigned int blockDimY, unsigned int blockDimZ, unsigned int sharedMemBytes, CUstream hStream, void **kernelParams);
18900 CUresult CUDAAPI cuSignalExternalSemaphoresAsync(const CUexternalSemaphore *extSemArray, const CUDA_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS *paramsArray, unsigned int numExtSems, CUstream stream);
18901 CUresult CUDAAPI cuWaitExternalSemaphoresAsync(const CUexternalSemaphore *extSemArray, const CUDA_EXTERNAL_SEMAPHORE_WAIT_PARAMS *paramsArray, unsigned int numExtSems, CUstream stream);
18902 CUresult CUDAAPI cuStreamBeginCapture(CUstream hStream);
18903 CUresult CUDAAPI cuStreamBeginCapture_ptsz(CUstream hStream);
18904 CUresult CUDAAPI cuStreamBeginCapture_v2(CUstream hStream, CUstreamCaptureMode mode);
18905 CUresult CUDAAPI cuStreamEndCapture(CUstream hStream, CUgraph *phGraph);
18906 CUresult CUDAAPI cuStreamIsCapturing(CUstream hStream, CUstreamCaptureStatus *captureStatus);
18907 CUresult CUDAAPI cuStreamGetCaptureInfo(CUstream hStream, CUstreamCaptureStatus *captureStatus_out, cuuint64_t *id_out);
18908 CUresult CUDAAPI cuStreamGetCaptureInfo_v2(CUstream hStream, CUstreamCaptureStatus *captureStatus_out, cuuint64_t *id_out, CUgraph *graph_out, const CUgraphNode **dependencies_out, size_t *numDependencies_out);
18909 CUresult CUDAAPI cuGraphUpload(CUgraphExec hGraph, CUstream hStream);
18910 CUresult CUDAAPI cuGraphLaunch(CUgraphExec hGraph, CUstream hStream);
18911 CUresult CUDAAPI cuStreamCopyAttributes(CUstream dstStream, CUstream srcStream);
18912 CUresult CUDAAPI cuStreamGetAttribute(CUstream hStream, CUstreamAttrID attr, CUstreamAttrValue *value);
18913 CUresult CUDAAPI cuStreamSetAttribute(CUstream hStream, CUstreamAttrID attr, const CUstreamAttrValue *param);
18914
18915 CUresult CUDAAPI cuIpcOpenMemHandle(CUdeviceptr *pdptr, CUipcMemHandle handle, unsigned int Flags);
18916 CUresult CUDAAPI cuGraphInstantiate(CUgraphExec *phGraphExec, CUgraph hGraph, CUgraphNode *phErrorNode, char *logBuffer, size_t bufferSize);
18917 CUresult CUDAAPI cuMemMapArrayAsync(CUarrayMapInfo *mapInfoList, unsigned int count, CUstream hStream);
18918
18919 CUresult CUDAAPI cuMemFreeAsync(CUdeviceptr dptr, CUstream hStream);
18920 CUresult CUDAAPI cuMemAllocAsync(CUdeviceptr *dptr, size_t bytesize, CUstream hStream);
18921 CUresult CUDAAPI cuMemAllocFromPoolAsync(CUdeviceptr *dptr, size_t bytesize, CUmemoryPool pool, CUstream hStream);
18922
18923 CUresult CUDAAPI cuStreamUpdateCaptureDependencies(CUstream hStream, CUgraphNode *dependencies, size_t numDependencies, unsigned int flags);
18924#elif defined(__CUDA_API_PER_THREAD_DEFAULT_STREAM)
18925static inline CUresult cuGetProcAddress_ptsz(const char *symbol, void **funcPtr, int driverVersion, cuuint64_t flags) {
18926 const int procAddressMask = (CU_GET_PROC_ADDRESS_LEGACY_STREAM|
18927 CU_GET_PROC_ADDRESS_PER_THREAD_DEFAULT_STREAM);
18928 if ((flags & procAddressMask) == 0) {
18929 flags |= CU_GET_PROC_ADDRESS_PER_THREAD_DEFAULT_STREAM;
18930 }
18931 return cuGetProcAddress(symbol, funcPtr, driverVersion, flags);
18932}
18933#define cuGetProcAddress cuGetProcAddress_ptsz
18934#endif
18935
18936#ifdef __cplusplus
18937}
18938#endif
18939
18940#if defined(__GNUC__)
18941 #if defined(__CUDA_API_PUSH_VISIBILITY_DEFAULT)
18942 #pragma GCC visibility pop
18943 #endif
18944#endif
18945
18946#undef __CUDA_DEPRECATED
18947
18948#endif /* __cuda_cuda_h__ */
18949