1 | /******************************************************************************* |
2 | * Copyright 2019-2021 Intel Corporation |
3 | * |
4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
5 | * you may not use this file except in compliance with the License. |
6 | * You may obtain a copy of the License at |
7 | * |
8 | * http://www.apache.org/licenses/LICENSE-2.0 |
9 | * |
10 | * Unless required by applicable law or agreed to in writing, software |
11 | * distributed under the License is distributed on an "AS IS" BASIS, |
12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
13 | * See the License for the specific language governing permissions and |
14 | * limitations under the License. |
15 | *******************************************************************************/ |
16 | |
17 | #include <CL/cl.h> |
18 | |
19 | #include "gpu/ocl/ocl_buffer_memory_storage.hpp" |
20 | #include "gpu/ocl/ocl_engine.hpp" |
21 | #include "gpu/ocl/ocl_stream.hpp" |
22 | #include "gpu/ocl/ocl_usm_utils.hpp" |
23 | |
24 | namespace dnnl { |
25 | namespace impl { |
26 | namespace gpu { |
27 | namespace ocl { |
28 | |
29 | status_t ocl_buffer_memory_storage_t::init_allocate(size_t size) { |
30 | auto *ocl_engine = utils::downcast<ocl_gpu_engine_t *>(engine()); |
31 | cl_int err; |
32 | mem_object_ = clCreateBuffer_wrapper( |
33 | ocl_engine->context(), CL_MEM_READ_WRITE, size, nullptr, &err); |
34 | |
35 | OCL_CHECK(err); |
36 | return status::success; |
37 | } |
38 | |
39 | namespace { |
40 | status_t get_map_queue( |
41 | cl_command_queue &queue, engine_t *engine, stream_t *stream) { |
42 | ocl_stream_t *ocl_stream; |
43 | if (stream == nullptr) { |
44 | auto *ocl_engine = utils::downcast<ocl_gpu_engine_t *>(engine); |
45 | status_t status = ocl_engine->get_service_stream(stream); |
46 | if (status != status::success) { return status::runtime_error; } |
47 | } |
48 | ocl_stream = utils::downcast<ocl_stream_t *>(stream); |
49 | queue = ocl_stream->queue(); |
50 | return status::success; |
51 | } |
52 | } // namespace |
53 | |
54 | status_t ocl_buffer_memory_storage_t::map_data( |
55 | void **mapped_ptr, stream_t *stream, size_t) const { |
56 | if (!mem_object()) { |
57 | *mapped_ptr = nullptr; |
58 | return status::success; |
59 | } |
60 | |
61 | cl_mem_flags mem_flags; |
62 | OCL_CHECK(clGetMemObjectInfo(mem_object(), CL_MEM_FLAGS, sizeof(mem_flags), |
63 | &mem_flags, nullptr)); |
64 | |
65 | size_t mem_bytes; |
66 | OCL_CHECK(clGetMemObjectInfo( |
67 | mem_object(), CL_MEM_SIZE, sizeof(mem_bytes), &mem_bytes, nullptr)); |
68 | |
69 | cl_map_flags map_flags = 0; |
70 | if (mem_flags & CL_MEM_READ_WRITE) { |
71 | map_flags |= CL_MAP_READ; |
72 | map_flags |= CL_MAP_WRITE; |
73 | } else if (mem_flags & CL_MEM_READ_ONLY) { |
74 | map_flags |= CL_MAP_READ; |
75 | } else if (mem_flags & CL_MEM_WRITE_ONLY) { |
76 | map_flags |= CL_MAP_WRITE; |
77 | } |
78 | |
79 | cl_command_queue queue; |
80 | CHECK(get_map_queue(queue, engine(), stream)); |
81 | |
82 | // Use blocking operation to simplify the implementation and API |
83 | cl_int err; |
84 | *mapped_ptr = clEnqueueMapBuffer(queue, mem_object(), CL_TRUE, map_flags, 0, |
85 | mem_bytes, 0, nullptr, nullptr, &err); |
86 | return convert_to_dnnl(err); |
87 | } |
88 | |
89 | status_t ocl_buffer_memory_storage_t::unmap_data( |
90 | void *mapped_ptr, stream_t *stream) const { |
91 | if (!mapped_ptr) return status::success; |
92 | cl_command_queue queue; |
93 | CHECK(get_map_queue(queue, engine(), stream)); |
94 | OCL_CHECK(clEnqueueUnmapMemObject(queue, mem_object_, |
95 | const_cast<void *>(mapped_ptr), 0, nullptr, nullptr)); |
96 | OCL_CHECK(clFinish(queue)); |
97 | return status::success; |
98 | } |
99 | |
100 | std::unique_ptr<memory_storage_t> ocl_buffer_memory_storage_t::get_sub_storage( |
101 | size_t offset, size_t size) const { |
102 | // Fast return on size = 0. |
103 | // It also seems clCreateSubBuffer() does not work properly for such case. |
104 | // Assumption: returned sub-storage won't be used for extracting cl_mem. |
105 | if (size == 0) return nullptr; |
106 | |
107 | cl_mem_flags mem_flags; |
108 | cl_int err; |
109 | err = clGetMemObjectInfo( |
110 | mem_object(), CL_MEM_FLAGS, sizeof(mem_flags), &mem_flags, nullptr); |
111 | assert(err == CL_SUCCESS); |
112 | if (err != CL_SUCCESS) return nullptr; |
113 | |
114 | assert(size != 0); |
115 | cl_buffer_region buffer_region = {base_offset_ + offset, size}; |
116 | ocl_wrapper_t<cl_mem> sub_buffer = clCreateSubBuffer(parent_mem_object(), |
117 | mem_flags, CL_BUFFER_CREATE_TYPE_REGION, &buffer_region, &err); |
118 | assert(err == CL_SUCCESS); |
119 | if (err != CL_SUCCESS) return nullptr; |
120 | |
121 | auto sub_storage |
122 | = new ocl_buffer_memory_storage_t(this->engine(), parent_storage()); |
123 | if (sub_storage) { |
124 | sub_storage->init(memory_flags_t::use_runtime_ptr, size, sub_buffer); |
125 | sub_storage->base_offset_ = base_offset_ + offset; |
126 | } |
127 | return std::unique_ptr<memory_storage_t>(sub_storage); |
128 | } |
129 | |
130 | std::unique_ptr<memory_storage_t> ocl_buffer_memory_storage_t::clone() const { |
131 | auto storage = new ocl_buffer_memory_storage_t(engine()); |
132 | if (storage) storage->init(memory_flags_t::use_runtime_ptr, 0, mem_object_); |
133 | return std::unique_ptr<memory_storage_t>(storage); |
134 | } |
135 | |
136 | cl_mem ocl_buffer_memory_storage_t::parent_mem_object() const { |
137 | return utils::downcast<const ocl_buffer_memory_storage_t *>( |
138 | parent_storage()) |
139 | ->mem_object(); |
140 | } |
141 | |
142 | } // namespace ocl |
143 | } // namespace gpu |
144 | } // namespace impl |
145 | } // namespace dnnl |
146 | |