1 | // Generated from "/code/pytorch/third_party/nvfuser/runtime/helpers.cu" |
2 | // 2023-02-12 08:01:26 |
3 | |
4 | namespace nvfuser_resources { |
5 | |
6 | constexpr const char* helpers_cu = R"( |
7 | #define NVFUSER_DEFINE_MAGIC_ZERO \ |
8 | __shared__ int nvfuser_zero_s; \ |
9 | if (threadIdx.x == 0) \ |
10 | nvfuser_zero_s = 0; \ |
11 | __syncthreads(); \ |
12 | atomicMin(&nvfuser_zero_s, threadIdx.x); \ |
13 | int nvfuser_zero = nvfuser_zero_s; |
14 | |
15 | #define NVFUSER_UPDATE_MAGIC_ZERO \ |
16 | do { \ |
17 | nvfuser_zero <<= 1; \ |
18 | } while (0); |
19 | |
20 | __device__ constexpr int ceilDiv(int a, int b) { |
21 | return (a + b - 1) / b; |
22 | } |
23 | |
24 | __device__ constexpr int64_t ceilDiv(int64_t a, int64_t b) { |
25 | return (a + b - 1) / b; |
26 | } |
27 | |
28 | __device__ constexpr int64_t ceilDiv(int64_t a, int b) { |
29 | return ceilDiv(a, (int64_t)b); |
30 | } |
31 | |
32 | __device__ constexpr int64_t ceilDiv(int a, int64_t b) { |
33 | return ceilDiv((int64_t)a, b); |
34 | } |
35 | |
36 | __device__ constexpr double ceilDiv(double a, double b) { |
37 | return std::ceil(a / b); |
38 | } |
39 | |
40 | __device__ constexpr double ceilDiv(double a, int64_t b) { |
41 | return std::ceil(a / b); |
42 | } |
43 | |
44 | __device__ constexpr double ceilDiv(int64_t a, double b) { |
45 | return std::ceil(a / b); |
46 | } |
47 | |
48 | // Monotonic and precise lerp is described here: |
49 | // https://math.stackexchange.com/a/1798323 |
50 | __device__ double lerp(double start, double end, double weight) { |
51 | if (weight < 0.5) { |
52 | return start + weight * (end - start); |
53 | } else { |
54 | return end - (end - start) * (1.0 - weight); |
55 | } |
56 | } |
57 | |
58 | __device__ float lerp(float start, float end, float weight) { |
59 | if (weight < 0.5f) { |
60 | return start + weight * (end - start); |
61 | } else { |
62 | return end - (end - start) * (1.0f - weight); |
63 | } |
64 | } |
65 | |
66 | __device__ std::complex<double> lerp( |
67 | std::complex<double> start, |
68 | std::complex<double> end, |
69 | std::complex<double> weight) { |
70 | if (abs(weight) < 0.5) { |
71 | return start + weight * (end - start); |
72 | } else { |
73 | return end - (end - start) * (1.0 - weight); |
74 | } |
75 | } |
76 | |
77 | __device__ std::complex<float> lerp( |
78 | std::complex<float> start, |
79 | std::complex<float> end, |
80 | std::complex<float> weight) { |
81 | if (abs(weight) < 0.5f) { |
82 | return start + weight * (end - start); |
83 | } else { |
84 | return end - (end - start) * (1.0f - weight); |
85 | } |
86 | } |
87 | |
88 | __device__ float lerp(float start, float end, double weight) { |
89 | return lerp(start, end, static_cast<float>(weight)); |
90 | } |
91 | |
92 | __device__ constexpr int max(int a, int b) { |
93 | return a > b ? a : b; |
94 | } |
95 | |
96 | __device__ constexpr int64_t max(int64_t a, int b) { |
97 | return a > (int64_t)b ? a : (int64_t)b; |
98 | } |
99 | |
100 | __device__ constexpr int64_t max(int a, int64_t b) { |
101 | return (int64_t)a > b ? (int64_t)a : b; |
102 | } |
103 | |
104 | __device__ constexpr int64_t max(int64_t a, int64_t b) { |
105 | return a > b ? a : b; |
106 | } |
107 | |
108 | __device__ double fmax(double a, double b) { |
109 | // check and propagate NaN |
110 | if (a != a) { |
111 | return a; |
112 | } else if (b != b) { |
113 | return b; |
114 | } else { |
115 | return a > b ? a : b; |
116 | } |
117 | } |
118 | |
119 | __device__ float fmax(float a, float b) { |
120 | // check and propagate NaN |
121 | if (a != a) { |
122 | return a; |
123 | } else if (b != b) { |
124 | return b; |
125 | } else { |
126 | return a > b ? a : b; |
127 | } |
128 | } |
129 | |
130 | __device__ constexpr int min(int a, int b) { |
131 | return a > b ? b : a; |
132 | } |
133 | |
134 | __device__ constexpr int64_t min(int64_t a, int b) { |
135 | return (int64_t)a > b ? b : (int64_t)a; |
136 | } |
137 | |
138 | __device__ constexpr int64_t min(int a, int64_t b) { |
139 | return a > (int64_t)b ? (int64_t)b : a; |
140 | } |
141 | |
142 | __device__ constexpr int64_t min(int64_t a, int64_t b) { |
143 | return a > b ? b : a; |
144 | } |
145 | |
146 | __device__ double fmin(double a, double b) { |
147 | // check and propagate NaN |
148 | if (a != a) { |
149 | return a; |
150 | } else if (b != b) { |
151 | return b; |
152 | } else { |
153 | return a > b ? b : a; |
154 | } |
155 | } |
156 | |
157 | __device__ float fmin(float a, float b) { |
158 | // check and propagate NaN |
159 | if (a != a) { |
160 | return a; |
161 | } else if (b != b) { |
162 | return b; |
163 | } else { |
164 | return a > b ? b : a; |
165 | } |
166 | } |
167 | |
168 | __device__ constexpr int alignBufferSize(int buffer, int size) { |
169 | return (buffer + (size - 1)) & ~(size - 1); |
170 | } |
171 | |
172 | __device__ double clamp(double x, double minv, double maxv) { |
173 | return fmin(fmax(x, minv), maxv); |
174 | } |
175 | |
176 | __device__ float clamp(float x, double minv, double maxv) { |
177 | return fmin(fmax((double)x, minv), maxv); |
178 | } |
179 | |
180 | __device__ int clamp(int x, int64_t minv, int64_t maxv) { |
181 | return min(max((int64_t)x, minv), maxv); |
182 | } |
183 | |
184 | __device__ int64_t clamp(int64_t x, int64_t minv, int64_t maxv) { |
185 | return min(max(x, minv), maxv); |
186 | } |
187 | |
188 | __device__ double frac(double x) { |
189 | return x - trunc(x); |
190 | } |
191 | |
192 | __device__ float frac(float x) { |
193 | return x - trunc(x); |
194 | } |
195 | |
196 | __device__ double reciprocal(double x) { |
197 | return 1 / x; |
198 | } |
199 | |
200 | __device__ float reciprocal(float x) { |
201 | return 1 / x; |
202 | } |
203 | |
204 | __device__ std::complex<double> reciprocal(std::complex<double> x) { |
205 | return 1.0 / x; |
206 | } |
207 | |
208 | __device__ std::complex<float> reciprocal(std::complex<float> x) { |
209 | return 1.0f / x; |
210 | } |
211 | |
212 | __device__ double relu(double x) { |
213 | return x <= 0 ? 0 : x; |
214 | } |
215 | |
216 | __device__ float relu(float x) { |
217 | return x <= 0 ? 0 : x; |
218 | } |
219 | |
220 | __device__ float relu(int64_t x) { |
221 | return x <= 0 ? 0 : x; |
222 | } |
223 | |
224 | __device__ float relu(int x) { |
225 | return x <= 0 ? 0 : x; |
226 | } |
227 | |
228 | __device__ double remainder(double a, double b) { |
229 | auto mod = ::fmod(a, b); |
230 | if ((mod != 0) && ((b < 0) != (mod < 0))) |
231 | mod += b; |
232 | return mod; |
233 | } |
234 | |
235 | __device__ float remainder(float a, float b) { |
236 | auto mod = ::fmod(a, b); |
237 | if ((mod != 0) && ((b < 0) != (mod < 0))) |
238 | mod += b; |
239 | return mod; |
240 | } |
241 | |
242 | __device__ double sigmoid(double x) { |
243 | return 1.0 / (1.0 + exp(-x)); |
244 | } |
245 | |
246 | __device__ float sigmoid(float x) { |
247 | return 1.0f / (1.0f + exp(-x)); |
248 | } |
249 | |
250 | __device__ std::complex<double> sigmoid(std::complex<double> x) { |
251 | return 1.0 / (1.0 + exp(-x)); |
252 | } |
253 | |
254 | __device__ std::complex<float> sigmoid(std::complex<float> x) { |
255 | return 1.0f / (1.0f + exp(-x)); |
256 | } |
257 | |
258 | __device__ double silu(double x) { |
259 | return x * sigmoid(x); |
260 | } |
261 | |
262 | __device__ float silu(float x) { |
263 | return x * sigmoid(x); |
264 | } |
265 | |
266 | __device__ double threshold(double x, double t, double v) { |
267 | return x <= t ? v : x; |
268 | } |
269 | |
270 | __device__ float threshold(float x, double t, double v) { |
271 | return x <= t ? v : x; |
272 | } |
273 | |
274 | __device__ std::complex<double> where( |
275 | bool c, |
276 | std::complex<double> a, |
277 | std::complex<double> b) { |
278 | return c ? a : b; |
279 | } |
280 | |
281 | __device__ std::complex<float> where( |
282 | bool c, |
283 | std::complex<float> a, |
284 | std::complex<float> b) { |
285 | return c ? a : b; |
286 | } |
287 | |
288 | __device__ int threshold(int x, int64_t t, int64_t v) { |
289 | return x <= t ? v : x; |
290 | } |
291 | |
292 | __device__ int64_t threshold(int64_t x, int64_t t, int64_t v) { |
293 | return x <= t ? v : x; |
294 | } |
295 | |
296 | __device__ double where(bool c, double a, double b) { |
297 | return c ? a : b; |
298 | } |
299 | |
300 | __device__ float where(bool c, float a, float b) { |
301 | return c ? a : b; |
302 | } |
303 | |
304 | __device__ int64_t where(bool c, int64_t a, int64_t b) { |
305 | return c ? a : b; |
306 | } |
307 | |
308 | __device__ int where(bool c, int a, int b) { |
309 | return c ? a : b; |
310 | } |
311 | |
312 | __device__ int64_t where(bool c, int64_t a, int b) { |
313 | return c ? a : b; |
314 | } |
315 | |
316 | __device__ int64_t where(bool c, int a, int64_t b) { |
317 | return c ? a : b; |
318 | } |
319 | |
320 | __device__ constexpr int64_t remainder(int64_t a, int64_t b) { |
321 | auto mod = a % b; |
322 | if ((mod != 0) && ((b < 0) != (mod < 0))) |
323 | mod += b; |
324 | return mod; |
325 | } |
326 | |
327 | __device__ constexpr int remainder(int a, int b) { |
328 | auto mod = a % b; |
329 | if ((mod != 0) && ((b < 0) != (mod < 0))) |
330 | mod += b; |
331 | return mod; |
332 | } |
333 | |
334 | __device__ constexpr int64_t fmod(int64_t a, int64_t b) { |
335 | return a % b; |
336 | } |
337 | |
338 | __device__ constexpr int fmod(int a, int b) { |
339 | return a % b; |
340 | } |
341 | |
342 | __device__ constexpr double fmod(double a, double b) { |
343 | return ::fmod(a, b); |
344 | } |
345 | |
346 | __device__ constexpr float fmod(float a, float b) { |
347 | return ::fmod(a, b); |
348 | } |
349 | |
350 | template <typename T> |
351 | __device__ T pow(T a, T b) { |
352 | if (b < 0) { |
353 | if (a == 1) { |
354 | return 1; |
355 | } else if (a == -1) { |
356 | auto negative = (-b) % static_cast<T>(2); |
357 | return negative ? -1 : 1; |
358 | } else { |
359 | return 0; |
360 | } |
361 | } else { |
362 | T result = 1; |
363 | while (b) { |
364 | if (b & 1) { |
365 | result *= a; |
366 | } |
367 | b /= 2; |
368 | a *= a; |
369 | } |
370 | return result; |
371 | } |
372 | } |
373 | |
374 | template __device__ int pow<int>(int a, int b); |
375 | template __device__ int64_t pow<int64_t>(int64_t a, int64_t b); |
376 | |
377 | template <> |
378 | __device__ float pow<float>(float a, float b) { |
379 | return ::pow(a, b); |
380 | } |
381 | |
382 | template <> |
383 | __device__ double pow<double>(double a, double b) { |
384 | return ::pow(a, b); |
385 | } |
386 | |
387 | __device__ float pow(float a, int b) { |
388 | return pow(a, (float)b); |
389 | } |
390 | |
391 | __device__ double pow(double a, int b) { |
392 | return pow(a, (double)b); |
393 | } |
394 | |
395 | __device__ float pow(float a, int64_t b) { |
396 | return pow(a, (float)b); |
397 | } |
398 | |
399 | __device__ double pow(double a, int64_t b) { |
400 | return pow(a, (double)b); |
401 | } |
402 | |
403 | int64_t pow(int64_t a, int b) { |
404 | return pow(a, (int64_t)b); |
405 | } |
406 | |
407 | int64_t pow(int a, int64_t b) { |
408 | return pow((int64_t)a, b); |
409 | } |
410 | |
411 | template <int size, int align = size> |
412 | struct alignas(align) TypelessData { |
413 | int8_t data[size]; |
414 | |
415 | template <typename T, std::enable_if_t<sizeof(T) == size, int> _ = 0> |
416 | TypelessData(T x) { |
417 | *reinterpret_cast<T*>(data) = x; |
418 | } |
419 | |
420 | template <typename T, std::enable_if_t<sizeof(T) == size, int> _ = 0> |
421 | operator T() { |
422 | return *reinterpret_cast<T*>(data); |
423 | } |
424 | }; |
425 | |
426 | template <typename T> |
427 | TypelessData<sizeof(T), alignof(T)> erase_type(T x) { |
428 | return x; |
429 | } |
430 | |
431 | template <typename T> |
432 | bool isfinite(T x) { |
433 | return ::isfinite(x); |
434 | } |
435 | |
436 | template <typename T> |
437 | bool isfinite(std::complex<T> x) { |
438 | return ::isfinite(std::real(x)) && ::isfinite(std::imag(x)); |
439 | } |
440 | |
441 | template <typename T> |
442 | bool isinf(T x) { |
443 | return ::isinf(x); |
444 | } |
445 | |
446 | template <typename T> |
447 | bool isinf(std::complex<T> x) { |
448 | return ::isinf(std::real(x)) || ::isinf(std::imag(x)); |
449 | } |
450 | |
451 | //////////////////////////////////////////////////////////// |
452 | // TODO: the following overloads are only needed for CUDA // |
453 | // 10.2 Please remove when CUDA 10.2 support is dropped // |
454 | //////////////////////////////////////////////////////////// |
455 | |
456 | bool isinf(int64_t x) { |
457 | return false; |
458 | } |
459 | |
460 | bool isinf(int x) { |
461 | return false; |
462 | } |
463 | |
464 | bool isinf(short x) { |
465 | return false; |
466 | } |
467 | |
468 | bool isinf(char x) { |
469 | return false; |
470 | } |
471 | |
472 | bool isinf(unsigned char x) { |
473 | return false; |
474 | } |
475 | |
476 | bool isinf(bool x) { |
477 | return false; |
478 | } |
479 | |
480 | bool isfinite(int64_t x) { |
481 | return true; |
482 | } |
483 | |
484 | bool isfinite(int x) { |
485 | return true; |
486 | } |
487 | |
488 | bool isfinite(short x) { |
489 | return true; |
490 | } |
491 | |
492 | bool isfinite(char x) { |
493 | return true; |
494 | } |
495 | |
496 | bool isfinite(unsigned char x) { |
497 | return true; |
498 | } |
499 | |
500 | bool isfinite(bool x) { |
501 | return true; |
502 | } |
503 | |
504 | //////////////////////////////////////////////////////////// |
505 | // End TODO // |
506 | //////////////////////////////////////////////////////////// |
507 | |
508 | template <typename T> |
509 | bool isnan(T x) { |
510 | return x != x; |
511 | } |
512 | |
513 | template <typename T> |
514 | bool isneginf(T x) { |
515 | return x < 0 && isinf(x); |
516 | } |
517 | |
518 | template <typename T> |
519 | bool isposinf(T x) { |
520 | return x > 0 && isinf(x); |
521 | } |
522 | |
523 | template <typename T> |
524 | bool isreal(T x) { |
525 | return true; |
526 | } |
527 | |
528 | template <typename T> |
529 | bool isreal(std::complex<T> x) { |
530 | return std::imag(x) == 0; |
531 | } |
532 | |
533 | // Return the current value of the cycle counter |
534 | __device__ inline int64_t readCycleCounter() { |
535 | // Ensures preceding memory operations are completed. Doing this |
536 | // would make sense for measuring elapsed times enclosed with this |
537 | // function. |
538 | __threadfence(); |
539 | return clock64(); |
540 | } |
541 | |
542 | __device__ float print_impl(const char* name, float value) { |
543 | printf( |
544 | "%s = %f @ threadIdx=(%d,%d,%d), blockIdx=(%d,%d,%d)\n", |
545 | name, |
546 | value, |
547 | (int)threadIdx.x, |
548 | (int)threadIdx.y, |
549 | (int)threadIdx.z, |
550 | (int)blockIdx.x, |
551 | (int)blockIdx.y, |
552 | (int)blockIdx.z); |
553 | return value; |
554 | } |
555 | |
556 | __device__ double print_impl(const char* name, double value) { |
557 | printf( |
558 | "%s = %lf @ threadIdx=(%d,%d,%d), blockIdx=(%d,%d,%d)\n", |
559 | name, |
560 | value, |
561 | (int)threadIdx.x, |
562 | (int)threadIdx.y, |
563 | (int)threadIdx.z, |
564 | (int)blockIdx.x, |
565 | (int)blockIdx.y, |
566 | (int)blockIdx.z); |
567 | return value; |
568 | } |
569 | |
570 | __device__ int print_impl(const char* name, int value) { |
571 | printf( |
572 | "%s = %d @ threadIdx=(%d,%d,%d), blockIdx=(%d,%d,%d)\n", |
573 | name, |
574 | value, |
575 | (int)threadIdx.x, |
576 | (int)threadIdx.y, |
577 | (int)threadIdx.z, |
578 | (int)blockIdx.x, |
579 | (int)blockIdx.y, |
580 | (int)blockIdx.z); |
581 | return value; |
582 | } |
583 | |
584 | __device__ int64_t print_impl(const char* name, int64_t value) { |
585 | printf( |
586 | "%s = %ld @ threadIdx=(%d,%d,%d), blockIdx=(%d,%d,%d)\n", |
587 | name, |
588 | value, |
589 | (int)threadIdx.x, |
590 | (int)threadIdx.y, |
591 | (int)threadIdx.z, |
592 | (int)blockIdx.x, |
593 | (int)blockIdx.y, |
594 | (int)blockIdx.z); |
595 | return value; |
596 | } |
597 | |
598 | __device__ bool print_impl(const char* name, bool value) { |
599 | printf( |
600 | "%s = %s @ threadIdx=(%d,%d,%d), blockIdx=(%d,%d,%d)\n", |
601 | name, |
602 | value ? "true" : "false", |
603 | (int)threadIdx.x, |
604 | (int)threadIdx.y, |
605 | (int)threadIdx.z, |
606 | (int)blockIdx.x, |
607 | (int)blockIdx.y, |
608 | (int)blockIdx.z); |
609 | return value; |
610 | } |
611 | |
612 | __device__ __half print_impl(const char* name, __half value) { |
613 | printf( |
614 | "%s = %f @ threadIdx=(%d,%d,%d), blockIdx=(%d,%d,%d)\n", |
615 | name, |
616 | __half2float(value), |
617 | (int)threadIdx.x, |
618 | (int)threadIdx.y, |
619 | (int)threadIdx.z, |
620 | (int)blockIdx.x, |
621 | (int)blockIdx.y, |
622 | (int)blockIdx.z); |
623 | return value; |
624 | } |
625 | |
626 | __device__ __bfloat print_impl(const char* name, __bfloat value) { |
627 | printf( |
628 | "%s = %f @ threadIdx=(%d,%d,%d), blockIdx=(%d,%d,%d)\n", |
629 | name, |
630 | __bfloat2float(value), |
631 | (int)threadIdx.x, |
632 | (int)threadIdx.y, |
633 | (int)threadIdx.z, |
634 | (int)blockIdx.x, |
635 | (int)blockIdx.y, |
636 | (int)blockIdx.z); |
637 | return value; |
638 | } |
639 | |
640 | #define print(...) print_impl(#__VA_ARGS__, (__VA_ARGS__)) |
641 | |
642 | template <typename OutT, typename IndexT, typename InputT> |
643 | __device__ OutT arange(IndexT index, InputT start, InputT step) { |
644 | return start + step * index; |
645 | } |
646 | )" ; |
647 | |
648 | } // namespace nvfuser_resources |
649 | |