cpu_vec.h 14.1 KB
Newer Older
T
tensor-tang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

#pragma once
T
tensor-tang 已提交
16
#include <cmath>
T
tensor-tang 已提交
17
#include <functional>
18
#include <string>
T
tensor-tang 已提交
19
#include "paddle/fluid/platform/cpu_info.h"
T
tensor-tang 已提交
20
#include "paddle/fluid/platform/enforce.h"
T
tensor-tang 已提交
21 22 23 24 25 26 27
#ifdef __AVX__
#include <immintrin.h>
#endif

#ifdef PADDLE_WITH_MKLML
#include "paddle/fluid/platform/dynload/mklml.h"
#endif
T
tensor-tang 已提交
28 29 30 31 32 33 34 35

namespace paddle {
namespace operators {
namespace math {

#define SIGMOID_THRESHOLD_MIN -40.0
#define SIGMOID_THRESHOLD_MAX 13.0

T
tensor-tang 已提交
36 37 38 39 40 41 42
#define AVX_FLOAT_BLOCK 8
#define AVX_DOUBLE_BLOCK 4
#define AVX2_FLOAT_BLOCK 8
#define AVX2_DOUBLE_BLOCK 4
#define AVX512_FLOAT_BLOCK 16
#define AVX512_DOUBLE_BLOCK 8

T
tensor-tang 已提交
43
template <typename T>
T
tensor-tang 已提交
44 45 46 47
inline void vec_exp(const int n, const T* x, T* y) {
  for (int i = 0; i < n; ++i) {
    y[i] = std::exp(x[i]);
  }
T
tensor-tang 已提交
48 49
}

50 51 52 53 54 55 56
template <typename T>
inline void vec_scal(const int n, const T a, T* x) {
  for (int i = 0; i < n; ++i) {
    x[i] = a * x[i];
  }
}

T
tensor-tang 已提交
57 58 59 60
#ifdef PADDLE_WITH_MKLML
template <>
inline void vec_exp<float>(const int n, const float* x, float* y) {
  platform::dynload::vsExp(n, x, y);
T
tensor-tang 已提交
61 62
}

T
tensor-tang 已提交
63 64 65 66
template <>
inline void vec_exp<double>(const int n, const double* x, double* y) {
  platform::dynload::vdExp(n, x, y);
}
67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91

template <>
inline void vec_scal<float>(const int n, const float a, float* x) {
  platform::dynload::cblas_sscal(n, a, x, 1);
}

template <>
inline void vec_scal<double>(const int n, const double a, double* x) {
  platform::dynload::cblas_dscal(n, a, x, 1);
}
#endif

// MKL scal only support inplace, choose this if src and dst are not equal
template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
inline void vec_scal(const int n, const T a, const T* x, T* y) {
  for (int i = 0; i < n; ++i) {
    y[i] = a * x[i];
  }
}

template <>
inline void vec_scal<float, platform::jit::avx>(const int n, const float a,
                                                const float* x, float* y) {
#ifdef __AVX__
  constexpr int block = AVX_FLOAT_BLOCK;
T
tensor-tang 已提交
92
  if (n < block) {
93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117
    vec_scal<float, platform::jit::isa_any>(n, a, x, y);
    return;
  }
  const int rest = n % block;
  const int end = n - rest;
  int i = 0;
  __m256 scalar = _mm256_set1_ps(a);
  __m256 tmp;
#define MOVE_ONE_STEP               \
  tmp = _mm256_loadu_ps(x + i);     \
  tmp = _mm256_mul_ps(tmp, scalar); \
  _mm256_storeu_ps(y + i, tmp)
  for (i = 0; i < end; i += block) {
    MOVE_ONE_STEP;
  }
#undef MOVE_ONE_STEP
  if (rest == 0) {
    return;
  }
  // can not continue move step if src and dst are inplace
  for (i = n - rest; i < n; ++i) {
    y[i] = a * x[i];
  }
#else
  vec_scal<float, platform::jit::isa_any>(n, a, x, y);
T
tensor-tang 已提交
118
#endif
119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134
}

template <>
inline void vec_scal<float, platform::jit::avx2>(const int n, const float a,
                                                 const float* x, float* y) {
  vec_scal<float, platform::jit::avx>(n, a, x, y);
}

template <>
inline void vec_scal<float, platform::jit::avx512_common>(const int n,
                                                          const float a,
                                                          const float* x,
                                                          float* y) {
  // TODO(TJ): enable me
  vec_scal<float, platform::jit::avx2>(n, a, x, y);
}
T
tensor-tang 已提交
135

T
tensor-tang 已提交
136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191
template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
inline void vec_bias_sub(const int n, const T a, const T* x, T* y) {
  for (int i = 0; i < n; ++i) {
    y[i] = a - x[i];
  }
}

template <>
inline void vec_bias_sub<float, platform::jit::avx>(const int n, const float a,
                                                    const float* x, float* y) {
#ifdef __AVX__
  constexpr int block = AVX_FLOAT_BLOCK;
  if (n < block) {
    vec_bias_sub<float, platform::jit::isa_any>(n, a, x, y);
    return;
  }
  const int rest = n % block;
  const int end = n - rest;
  int i = 0;
  __m256 bias = _mm256_set1_ps(a);
  __m256 tmp;
#define MOVE_ONE_STEP             \
  tmp = _mm256_loadu_ps(x + i);   \
  tmp = _mm256_sub_ps(bias, tmp); \
  _mm256_storeu_ps(y + i, tmp)
  for (i = 0; i < end; i += block) {
    MOVE_ONE_STEP;
  }
#undef MOVE_ONE_STEP
  if (rest == 0) {
    return;
  }
  // can not continue move step if src and dst are inplace
  for (i = n - rest; i < n; ++i) {
    y[i] = a - x[i];
  }
#else
  vec_bias_sub<float, platform::jit::isa_any>(n, a, x, y);
#endif
}

template <>
inline void vec_bias_sub<float, platform::jit::avx2>(const int n, const float a,
                                                     const float* x, float* y) {
  vec_bias_sub<float, platform::jit::avx>(n, a, x, y);
}

template <>
inline void vec_bias_sub<float, platform::jit::avx512_common>(const int n,
                                                              const float a,
                                                              const float* x,
                                                              float* y) {
  // TODO(TJ): enable me
  vec_bias_sub<float, platform::jit::avx2>(n, a, x, y);
}

T
tensor-tang 已提交
192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250
// out = x*y + (1-x)*z
template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
inline void vec_cross(const int n, const T* x, const T* y, const T* z, T* out) {
  for (int i = 0; i < n; ++i) {
    out[i] = x[i] * y[i] + (static_cast<T>(1) - x[i]) * z[i];
  }
}

template <>
inline void vec_cross<float, platform::jit::avx>(const int n, const float* x,
                                                 const float* y, const float* z,
                                                 float* out) {
#ifdef __AVX__
  constexpr int block = AVX_FLOAT_BLOCK;
  if (n < block) {
    vec_cross<float, platform::jit::isa_any>(n, x, y, z, out);
    return;
  }
  const int rest = n % block;
  const int end = n - rest;
  int i = 0;
  __m256 bias = _mm256_set1_ps(1.f);
  __m256 tmpx, tmpy, tmpz;
  for (i = 0; i < end; i += block) {
    tmpx = _mm256_loadu_ps(x + i);
    tmpy = _mm256_loadu_ps(y + i);
    tmpz = _mm256_loadu_ps(z + i);
    tmpy = _mm256_mul_ps(tmpx, tmpy);
    tmpx = _mm256_sub_ps(bias, tmpx);
    tmpz = _mm256_mul_ps(tmpx, tmpz);
    tmpz = _mm256_add_ps(tmpy, tmpz);
    _mm256_storeu_ps(out + i, tmpz);
  }
  if (rest == 0) {
    return;
  }
  // can not continue move step if src and dst are inplace
  for (i = n - rest; i < n; ++i) {
    out[i] = x[i] * y[i] + (1.f - x[i]) * z[i];
  }
#else
  vec_cross<float, platform::jit::isa_any>(n, x, y, z, out);
#endif
}

template <>
inline void vec_cross<float, platform::jit::avx2>(const int n, const float* x,
                                                  const float* y,
                                                  const float* z, float* out) {
  vec_cross<float, platform::jit::avx>(n, x, y, z, out);
}

template <>
inline void vec_cross<float, platform::jit::avx512_common>(
    const int n, const float* x, const float* y, const float* z, float* out) {
  // TODO(TJ): enable me
  vec_cross<float, platform::jit::avx>(n, x, y, z, out);
}

T
tensor-tang 已提交
251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306
template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
inline void vec_add_bias(const int n, const T a, const T* x, T* y) {
  for (int i = 0; i < n; ++i) {
    y[i] = x[i] + a;
  }
}

template <>
inline void vec_add_bias<float, platform::jit::avx>(const int n, const float a,
                                                    const float* x, float* y) {
#ifdef __AVX__
  constexpr int block = AVX_FLOAT_BLOCK;
  if (n < block) {
    vec_add_bias<float, platform::jit::isa_any>(n, a, x, y);
    return;
  }
  const int rest = n % block;
  const int end = n - rest;
  int i = 0;
  __m256 bias = _mm256_set1_ps(a);
  __m256 tmp;
#define MOVE_ONE_STEP             \
  tmp = _mm256_loadu_ps(x + i);   \
  tmp = _mm256_add_ps(tmp, bias); \
  _mm256_storeu_ps(y + i, tmp)
  for (i = 0; i < end; i += block) {
    MOVE_ONE_STEP;
  }
#undef MOVE_ONE_STEP
  if (rest == 0) {
    return;
  }
  // can not continue move step if src and dst are inplace
  for (i = n - rest; i < n; ++i) {
    y[i] = x[i] + a;
  }
#else
  vec_add_bias<float, platform::jit::isa_any>(n, a, x, y);
#endif
}

template <>
inline void vec_add_bias<float, platform::jit::avx2>(const int n, const float a,
                                                     const float* x, float* y) {
  vec_add_bias<float, platform::jit::avx>(n, a, x, y);
}

template <>
inline void vec_add_bias<float, platform::jit::avx512_common>(const int n,
                                                              const float a,
                                                              const float* x,
                                                              float* y) {
  // TODO(TJ): enable me
  vec_add_bias<float, platform::jit::avx2>(n, a, x, y);
}

307 308 309 310 311 312
template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
inline void vec_identity(const int n, const T* x, T* y) {
  // do nothing
  return;
}

T
tensor-tang 已提交
313 314 315 316 317
template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
inline void vec_sigmoid(const int n, const T* x, T* y) {
  const T min = SIGMOID_THRESHOLD_MIN;
  const T max = SIGMOID_THRESHOLD_MAX;
  for (int i = 0; i < n; ++i) {
T
tensor-tang 已提交
318 319 320 321 322 323
    y[i] = (x[i] < min) ? min : ((x[i] > max) ? max : x[i]);
    y[i] = static_cast<T>(0) - y[i];
  }
  vec_exp<T>(n, y, y);
  for (int i = 0; i < n; ++i) {
    y[i] = static_cast<T>(1) / (static_cast<T>(1) + y[i]);
T
tensor-tang 已提交
324 325 326
  }
}

327 328 329 330 331
template <>
inline void vec_sigmoid<float, platform::jit::avx>(const int n, const float* x,
                                                   float* y) {
#ifdef __AVX__
  constexpr int block = AVX_FLOAT_BLOCK;
332
  if (n < block) {
333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351
    vec_sigmoid<float, platform::jit::isa_any>(n, x, y);
    return;
  }
  const int rest = n % block;
  const int end = n - rest;
  int i = 0;
  __m256 max = _mm256_set1_ps(SIGMOID_THRESHOLD_MAX);
  __m256 min = _mm256_set1_ps(SIGMOID_THRESHOLD_MIN);
  __m256 zeros = _mm256_setzero_ps();
  __m256 tmp;
#define MOVE_ONE_STEP              \
  tmp = _mm256_loadu_ps(x + i);    \
  tmp = _mm256_max_ps(tmp, min);   \
  tmp = _mm256_min_ps(tmp, max);   \
  tmp = _mm256_sub_ps(zeros, tmp); \
  _mm256_storeu_ps(y + i, tmp)
  for (i = 0; i < end; i += block) {
    MOVE_ONE_STEP;
  }
352
#undef MOVE_ONE_STEP
353
  if (rest != 0) {
354 355 356 357 358 359
    // can not continue move step since the src and dst address could be equal
    const float xmin = SIGMOID_THRESHOLD_MIN;
    const float xmax = SIGMOID_THRESHOLD_MAX;
    for (i = n - rest; i < n; ++i) {
      y[i] = 0.f - ((x[i] < xmin) ? xmin : ((x[i] > xmax) ? xmax : x[i]));
    }
360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395
  }

  vec_exp<float>(n, y, y);

  __m256 ones = _mm256_set1_ps(1.0f);
#define MOVE_ONE_STEP             \
  tmp = _mm256_loadu_ps(y + i);   \
  tmp = _mm256_add_ps(ones, tmp); \
  tmp = _mm256_div_ps(ones, tmp); \
  _mm256_storeu_ps(y + i, tmp)
  for (i = 0; i < end; i += block) {
    MOVE_ONE_STEP;
  }
#undef MOVE_ONE_STEP
  if (rest == 0) {
    return;
  }
  // can not continue move step
  for (i = n - rest; i < n; ++i) {
    y[i] = 1.f / (1.f + y[i]);
  }
#else
  vec_sigmoid<float, platform::jit::isa_any>(n, x, y);
#endif
}

template <>
inline void vec_sigmoid<float, platform::jit::avx2>(const int n, const float* x,
                                                    float* y) {
  vec_sigmoid<float, platform::jit::avx>(n, x, y);
}

template <>
inline void vec_sigmoid<float, platform::jit::avx512_common>(const int n,
                                                             const float* x,
                                                             float* y) {
396 397
  // TODO(TJ): enable me
  vec_sigmoid<float, platform::jit::avx2>(n, x, y);
398 399
}

T
tensor-tang 已提交
400 401
template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
inline void vec_tanh(const int n, const T* x, T* y) {
402 403 404
  vec_scal<T, isa>(n, static_cast<T>(2), x, y);
  vec_sigmoid<T, isa>(n, y, y);
  vec_scal<T>(n, static_cast<T>(2), y);
T
tensor-tang 已提交
405
  vec_add_bias<T, isa>(n, static_cast<T>(-1), y, y);
T
tensor-tang 已提交
406 407
}

T
tensor-tang 已提交
408
// TODO(TJ): make relu clip
T
tensor-tang 已提交
409 410 411 412 413 414 415
template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
inline void vec_relu(const int n, const T* x, T* y) {
  for (int i = 0; i < n; ++i) {
    y[i] = x[i] > 0 ? x[i] : 0;
  }
}

T
tensor-tang 已提交
416 417 418 419 420
template <>
inline void vec_relu<float, platform::jit::avx>(const int n, const float* x,
                                                float* y) {
#ifdef __AVX__
  constexpr int block = AVX_FLOAT_BLOCK;
T
tensor-tang 已提交
421
  if (n < block * 4) {
T
tensor-tang 已提交
422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449
    vec_relu<float, platform::jit::isa_any>(n, x, y);
    return;
  }

  const int rest = n % block;
  const int end = n - rest;
  int i = 0;
  __m256 zeros = _mm256_setzero_ps();
  __m256 tmp;
#define MOVE_ONE_STEP              \
  tmp = _mm256_loadu_ps(x + i);    \
  tmp = _mm256_max_ps(tmp, zeros); \
  _mm256_storeu_ps(y + i, tmp)
  for (i = 0; i < end; i += block) {
    MOVE_ONE_STEP;
  }
  if (rest == 0) {
    return;
  }
  i = n - block;
  MOVE_ONE_STEP;
#undef MOVE_ONE_STEP

#else
  vec_relu<float, platform::jit::isa_any>(n, x, y);
#endif
}

T
tensor-tang 已提交
450 451 452
template <>
inline void vec_relu<float, platform::jit::avx2>(const int n, const float* x,
                                                 float* y) {
T
tensor-tang 已提交
453
  vec_relu<float, platform::jit::avx>(n, x, y);
T
tensor-tang 已提交
454 455 456
}

template <>
T
tensor-tang 已提交
457 458 459
inline void vec_relu<float, platform::jit::avx512_common>(const int n,
                                                          const float* x,
                                                          float* y) {
460
  // TODO(TJ): enable me
T
tensor-tang 已提交
461
  vec_relu<float, platform::jit::avx2>(n, x, y);
T
tensor-tang 已提交
462 463
}

T
tensor-tang 已提交
464 465
// TODO(TJ): optimize double of sigmoid, tanh and relu if necessary

466 467 468 469 470 471 472 473 474 475 476 477 478 479
template <typename T, platform::jit::cpu_isa_t isa = platform::jit::isa_any>
class VecActivations {
 public:
  std::function<void(const int, const T*, T*)> operator()(
      const std::string& type) {
    if (type == "sigmoid") {
      return vec_sigmoid<T, isa>;
    } else if (type == "relu") {
      return vec_relu<T, isa>;
    } else if (type == "tanh") {
      return vec_tanh<T, isa>;
    } else if (type == "identity" || type == "") {
      return vec_identity<T, isa>;
    }
T
tensor-tang 已提交
480
    PADDLE_THROW("Not support type: %s", type);
481 482 483
  }
};

T
tensor-tang 已提交
484 485 486
}  // namespace math
}  // namespace operators
}  // namespace paddle