jit_kernel_test.cc 17.5 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) 2018 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. */

#include "paddle/fluid/operators/math/jit_kernel.h"
T
tensor-tang 已提交
16
#include <sys/time.h>
T
tensor-tang 已提交
17
#include <cstring>  // for memcpy
T
tensor-tang 已提交
18 19 20 21 22 23
#include <string>
#include <vector>
#include "gflags/gflags.h"
#include "glog/logging.h"
#include "gtest/gtest.h"

T
tensor-tang 已提交
24 25 26 27 28 29 30 31
#ifdef PADDLE_WITH_MKLML
#include "paddle/fluid/platform/dynload/mklml.h"
#endif

#ifdef __AVX__
#include <immintrin.h>
#endif

T
tensor-tang 已提交
32 33
constexpr int repeat = 20000;

T
tensor-tang 已提交
34 35 36 37 38 39 40
inline double GetCurrentUS() {
  struct timeval time;
  gettimeofday(&time, NULL);
  return 1e+6 * time.tv_sec + time.tv_usec;
}

template <typename T>
T
tensor-tang 已提交
41 42
void RandomVec(const int n, T* a, const T lower = static_cast<T>(-20.f),
               const T upper = static_cast<T>(20.f)) {
T
tensor-tang 已提交
43 44 45 46 47 48 49 50
  static unsigned int seed = 100;
  std::mt19937 rng(seed++);
  std::uniform_real_distribution<double> uniform_dist(0, 1);
  for (int i = 0; i < n; ++i) {
    a[i] = static_cast<T>(uniform_dist(rng) * (upper - lower) + lower);
  }
}

T
tensor-tang 已提交
51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87
void vaddbias_ref(const int n, const float a, const float* x, float* y) {
  for (int i = 0; i < n; ++i) {
    y[i] = x[i] + a;
  }
}

TEST(JitKernel, vaddbias) {
  namespace jit = paddle::operators::math::jitkernel;
  for (int d : {7, 8, 15, 16, 30, 64, 100, 128, 256}) {
    std::vector<float> x(d);
    std::vector<float> zref(d), ztgt(d);
    RandomVec<float>(d, x.data(), -2.f, 2.f);
    const auto& ker =
        jit::KernelPool::Instance().template Get<jit::VAddBiasKernel<float>>(d);
    const float a = 2.f;
    const float* x_data = x.data();
    float* ztgt_data = ztgt.data();
    float* zref_data = zref.data();
    auto trefs = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      vaddbias_ref(d, a, x_data, zref_data);
    }
    auto trefe = GetCurrentUS();
    auto ttgts = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      ker->Compute(d, a, x_data, ztgt_data);
    }
    auto ttgte = GetCurrentUS();

    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
            << " us, tgt takes: " << (ttgte - ttgts) / repeat;
    for (int i = 0; i < d; ++i) {
      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
    }
  }
}

T
tensor-tang 已提交
88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139
void vexp_ref(const int n, const float* x, float* y) {
  for (int i = 0; i < n; ++i) {
    y[i] = std::exp(x[i]);
  }
}

#ifdef PADDLE_WITH_MKLML
void vexp_mkl(const int n, const float* x, float* y) {
  paddle::platform::dynload::vsExp(n, x, y);
}
#endif

TEST(JitKernel, vexp) {
  namespace jit = paddle::operators::math::jitkernel;
  for (int d : {7, 8, 15, 16, 30, 128}) {
    std::vector<float> x(d);
    std::vector<float> zref(d), ztgt(d);
    RandomVec<float>(d, x.data(), -2.f, 2.f);
    const auto& ker =
        jit::KernelPool::Instance().template Get<jit::VExpKernel<float>>(d);
    const float* x_data = x.data();
    float* ztgt_data = ztgt.data();
    float* zref_data = zref.data();
    auto trefs = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      vexp_ref(d, x_data, zref_data);
    }
    auto trefe = GetCurrentUS();

#ifdef PADDLE_WITH_MKLML
    auto tmkls = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      vexp_mkl(d, x_data, zref_data);
    }
    auto tmkle = GetCurrentUS();
#endif

    auto ttgts = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      ker->Compute(d, x_data, ztgt_data);
    }
    auto ttgte = GetCurrentUS();

    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
#ifdef PADDLE_WITH_MKLML
            << " us, mkl takes: " << (tmkle - tmkls) / repeat << " us, "
#else
            << " us, "
#endif
            << "tgt takes: " << (ttgte - ttgts) / repeat;
    for (int i = 0; i < d; ++i) {
      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
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
    }
  }
}

inline float _sigmoid(float x) {
  const float min = SIGMOID_THRESHOLD_MIN;
  const float max = SIGMOID_THRESHOLD_MAX;
  float tmp = (x < min) ? min : ((x > max) ? max : x);
  return 1.f / (1.f + std::exp(-tmp));
}

void vsigmoid_ref(const int n, const float* x, float* y) {
  for (int i = 0; i < n; ++i) {
    y[i] = _sigmoid(x[i]);
  }
}

void vsigmoid_better(
    const std::shared_ptr<
        const paddle::operators::math::jitkernel::VExpKernel<float>>& vexp,
    const int n, const float* x, float* y) {
  const float min = SIGMOID_THRESHOLD_MIN;
  const float max = SIGMOID_THRESHOLD_MAX;
  for (int i = 0; i < n; ++i) {
    y[i] = (x[i] < min) ? min : ((x[i] > max) ? max : x[i]);
    y[i] = 0.f - y[i];
  }
  vexp->Compute(n, y, y);
  for (int i = 0; i < n; ++i) {
    y[i] = 1.f / (1.f + y[i]);
  }
}

TEST(JitKernel, vsigmoid) {
  namespace jit = paddle::operators::math::jitkernel;
T
tensor-tang 已提交
175
  for (int d : {7, 8, 15, 16, 30, 32, 64, 100, 128, 256}) {
176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201
    std::vector<float> x(d);
    std::vector<float> zref(d), ztgt(d);
    RandomVec<float>(d, x.data(), -2.f, 2.f);
    const auto& ker =
        jit::KernelPool::Instance().template Get<jit::VSigmoidKernel<float>>(d);
    const auto& vexp =
        jit::KernelPool::Instance().template Get<jit::VExpKernel<float>>(d);
    const float* x_data = x.data();
    float* ztgt_data = ztgt.data();
    float* zref_data = zref.data();
    auto tmkls = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      vsigmoid_better(vexp, d, x_data, zref_data);
    }
    auto tmkle = GetCurrentUS();
    auto trefs = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      vsigmoid_ref(d, x_data, zref_data);
    }
    auto trefe = GetCurrentUS();
    auto ttgts = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      ker->Compute(d, x_data, ztgt_data);
    }
    auto ttgte = GetCurrentUS();

T
tensor-tang 已提交
202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267
    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
            << " us, better(jit exp) takes: " << (tmkle - tmkls) / repeat
            << " us, tgt takes: " << (ttgte - ttgts) / repeat;
    for (int i = 0; i < d; ++i) {
      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
    }
  }
}

inline float _tanh(float x) { return 2.f * _sigmoid(2.f * x) - 1.f; }

void vtanh_ref(const int n, const float* x, float* y) {
  for (int i = 0; i < n; ++i) {
    y[i] = _tanh(x[i]);
  }
}

void vtanh_better(
    const std::shared_ptr<
        const paddle::operators::math::jitkernel::VScalKernel<float>>& vscal,
    const std::shared_ptr<
        const paddle::operators::math::jitkernel::VSigmoidKernel<float>>&
        vsigmoid,
    const std::shared_ptr<
        const paddle::operators::math::jitkernel::VAddBiasKernel<float>>&
        vaddbias,
    const int n, const float* x, float* y) {
  vscal->Compute(n, 2.f, x, y);
  vsigmoid->Compute(n, y, y);
  vscal->Compute(n, 2.f, y);
  vaddbias->Compute(n, -1.f, y, y);
}

TEST(JitKernel, vtanh) {
  namespace jit = paddle::operators::math::jitkernel;
  for (int d : {7, 8, 15, 16, 30, 32, 64, 100, 128, 256}) {
    std::vector<float> x(d);
    std::vector<float> zref(d), ztgt(d);
    RandomVec<float>(d, x.data(), -2.f, 2.f);
    const auto& ker =
        jit::KernelPool::Instance().template Get<jit::VTanhKernel<float>>(d);
    const auto& vscal =
        jit::KernelPool::Instance().template Get<jit::VScalKernel<float>>(d);
    const auto& vsigmoid =
        jit::KernelPool::Instance().template Get<jit::VSigmoidKernel<float>>(d);
    const auto& vaddbias =
        jit::KernelPool::Instance().template Get<jit::VAddBiasKernel<float>>(d);
    const float* x_data = x.data();
    float* ztgt_data = ztgt.data();
    float* zref_data = zref.data();
    auto tmkls = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      vtanh_better(vscal, vsigmoid, vaddbias, d, x_data, zref_data);
    }
    auto tmkle = GetCurrentUS();
    auto trefs = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      vtanh_ref(d, x_data, zref_data);
    }
    auto trefe = GetCurrentUS();
    auto ttgts = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      ker->Compute(d, x_data, ztgt_data);
    }
    auto ttgte = GetCurrentUS();

268 269 270 271 272
    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
            << " us, better(jit exp) takes: " << (tmkle - tmkls) / repeat
            << " us, tgt takes: " << (ttgte - ttgts) / repeat;
    for (int i = 0; i < d; ++i) {
      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
T
tensor-tang 已提交
273 274 275 276
    }
  }
}

T
tensor-tang 已提交
277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383
void vscal_ref(const int n, const float a, const float* x, float* y) {
  for (int i = 0; i < n; ++i) {
    y[i] = a * x[i];
  }
}
void vscal_inp_ref(const int n, const float a, float* x) {
  for (int i = 0; i < n; ++i) {
    x[i] = a * x[i];
  }
}
#if defined __AVX__ || defined __AVX2__
void vscal_intri8(const int n, const float a, const float* x, float* y) {
  __m256 tmp;
  __m256 scalar = _mm256_set1_ps(a);
  tmp = _mm256_loadu_ps(x);
  tmp = _mm256_mul_ps(tmp, scalar);
  _mm256_storeu_ps(y, tmp);
}
void vscal_inp_intri8(const int n, const float a, float* x) {
  __m256 tmp;
  __m256 scalar = _mm256_set1_ps(a);
  tmp = _mm256_loadu_ps(x);
  tmp = _mm256_mul_ps(tmp, scalar);
  _mm256_storeu_ps(x, tmp);
}
#endif

#ifdef PADDLE_WITH_MKLML
void vscal_inp_mkl(const int n, const float a, float* x) {
  paddle::platform::dynload::cblas_sscal(n, a, x, 1);
}
#endif

TEST(JitKernel, vscal) {
  namespace jit = paddle::operators::math::jitkernel;
  for (int d : {7, 8, 15, 16, 30, 256, 512}) {
    std::vector<float> x(d), y(d);
    std::vector<float> zref(d), ztgt(d);
    RandomVec<float>(d, x.data());
    std::memcpy(y.data(), x.data(), sizeof(float) * d);
    float a = 2.f;
    const auto& ker =
        jit::KernelPool::Instance().template Get<jit::VScalKernel<float>>(d);
    const float* x_data = x.data();
    float* y_data = y.data();
    float* ztgt_data = ztgt.data();
    float* zref_data = zref.data();
    auto trefs = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      vscal_ref(d, a, x_data, zref_data);
    }
    auto trefe = GetCurrentUS();
    auto trefs1 = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      vscal_inp_ref(d, a, y_data);
    }
    auto trefe1 = GetCurrentUS();

#ifdef PADDLE_WITH_MKLML
    auto tmkls = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      vscal_inp_mkl(d, a, y_data);
    }
    auto tmkle = GetCurrentUS();
#endif

#if defined __AVX__ || defined __AVX2__
    if (d == 8) {
      auto si0 = GetCurrentUS();
      for (int i = 0; i < repeat; ++i) {
        vscal_intri8(d, a, x_data, zref_data);
      }
      auto si1 = GetCurrentUS();
      auto si2 = GetCurrentUS();
      for (int i = 0; i < repeat; ++i) {
        vscal_inp_intri8(d, a, y_data);
      }
      auto si3 = GetCurrentUS();
      VLOG(3) << "Vec size 8 intr takes: " << (si1 - si0) / repeat
              << " us, inplace: " << (si3 - si2) / repeat;
    }
#endif

    auto ttgts = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      ker->Compute(d, a, x_data, ztgt_data);
    }
    auto ttgte = GetCurrentUS();
    auto ttgts1 = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      ker->Compute(d, a, y_data);
    }
    auto ttgte1 = GetCurrentUS();
    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
            << " us, inplace takes: " << (trefe1 - trefs1) / repeat
#ifdef PADDLE_WITH_MKLML
            << " us, mkl inplace takes: " << (tmkle - tmkls) / repeat << " us, "
#else
            << " us, "
#endif
            << "tgt takes: " << (ttgte - ttgts) / repeat
            << "us, tgt inplace takes: " << (ttgte1 - ttgts1) / repeat;
    for (int i = 0; i < d; ++i) {
      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
    }
  }
}
T
tensor-tang 已提交
384

T
tensor-tang 已提交
385 386 387 388 389 390
void vmul_ref(const int n, const float* x, const float* y, float* z) {
  for (int i = 0; i < n; ++i) {
    z[i] = x[i] * y[i];
  }
}

T
tensor-tang 已提交
391
#if defined __AVX__ || defined __AVX2__
T
tensor-tang 已提交
392
void vmul_intri8(const int n, const float* x, const float* y, float* z) {
T
tensor-tang 已提交
393 394 395 396 397 398 399
  __m256 tmpx, tmpy;
  tmpx = _mm256_loadu_ps(x);
  tmpy = _mm256_loadu_ps(y);
  tmpx = _mm256_mul_ps(tmpx, tmpy);
  _mm256_storeu_ps(z, tmpx);
}
#endif
T
tensor-tang 已提交
400

T
tensor-tang 已提交
401 402 403
#ifdef PADDLE_WITH_MKLML
void vmul_mkl(const int n, const float* x, const float* y, float* z) {
  paddle::platform::dynload::vsMul(n, x, y, z);
T
tensor-tang 已提交
404
}
T
tensor-tang 已提交
405
#endif
T
tensor-tang 已提交
406

T
tensor-tang 已提交
407 408
TEST(JitKernel, vmul) {
  namespace jit = paddle::operators::math::jitkernel;
T
tensor-tang 已提交
409
  for (int d : {7, 8, 15, 16, 30, 256, 512}) {
T
tensor-tang 已提交
410 411 412 413 414 415 416 417 418 419
    std::vector<float> x(d), y(d);
    std::vector<float> zref(d), ztgt(d);
    RandomVec<float>(d, x.data());
    RandomVec<float>(d, y.data());
    const auto& ker =
        jit::KernelPool::Instance().template Get<jit::VMulKernel<float>>(d);
    const float* x_data = x.data();
    const float* y_data = y.data();
    float* ztgt_data = ztgt.data();
    float* zref_data = zref.data();
T
tensor-tang 已提交
420
    auto trefs = GetCurrentUS();
T
tensor-tang 已提交
421
    for (int i = 0; i < repeat; ++i) {
T
tensor-tang 已提交
422
      vmul_ref(d, x_data, y_data, zref_data);
T
tensor-tang 已提交
423
    }
T
tensor-tang 已提交
424
    auto trefe = GetCurrentUS();
T
tensor-tang 已提交
425

T
tensor-tang 已提交
426 427
#ifdef PADDLE_WITH_MKLML
    auto tmkls = GetCurrentUS();
T
tensor-tang 已提交
428
    for (int i = 0; i < repeat; ++i) {
T
tensor-tang 已提交
429
      vmul_mkl(d, x_data, y_data, zref_data);
T
tensor-tang 已提交
430
    }
T
tensor-tang 已提交
431 432
    auto tmkle = GetCurrentUS();
#endif
T
tensor-tang 已提交
433

T
tensor-tang 已提交
434 435 436 437
#if defined __AVX__ || defined __AVX2__
    if (d == 8) {
      auto si0 = GetCurrentUS();
      for (int i = 0; i < repeat; ++i) {
T
tensor-tang 已提交
438
        vmul_intri8(d, x_data, y_data, zref_data);
T
tensor-tang 已提交
439 440
      }
      auto si1 = GetCurrentUS();
T
tensor-tang 已提交
441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519
      VLOG(3) << "Vec size 8 intr takes: " << (si1 - si0) / repeat;
    }
#endif

    auto ttgts = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      ker->Compute(d, x_data, y_data, ztgt_data);
    }
    auto ttgte = GetCurrentUS();

    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
#ifdef PADDLE_WITH_MKLML
            << " us, mkl takes: " << (tmkle - tmkls) / repeat << " us, "
#else
            << " us, "
#endif
            << "tgt takes: " << (ttgte - ttgts) / repeat;
    for (int i = 0; i < d; ++i) {
      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
    }
  }
}

void vadd_ref(const int n, const float* x, const float* y, float* z) {
  for (int i = 0; i < n; ++i) {
    z[i] = x[i] + y[i];
  }
}

#if defined __AVX__ || defined __AVX2__
void vadd_intri8(const int n, const float* x, const float* y, float* z) {
  __m256 tmpx, tmpy;
  tmpx = _mm256_loadu_ps(x);
  tmpy = _mm256_loadu_ps(y);
  tmpx = _mm256_add_ps(tmpx, tmpy);
  _mm256_storeu_ps(z, tmpx);
}
#endif

#ifdef PADDLE_WITH_MKLML
void vadd_mkl(const int n, const float* x, const float* y, float* z) {
  paddle::platform::dynload::vsAdd(n, x, y, z);
}
#endif

TEST(JitKernel, vadd) {
  namespace jit = paddle::operators::math::jitkernel;
  for (int d : {7, 8, 15, 16, 30, 256, 512}) {
    std::vector<float> x(d), y(d);
    std::vector<float> zref(d), ztgt(d);
    RandomVec<float>(d, x.data());
    RandomVec<float>(d, y.data());
    const auto& ker =
        jit::KernelPool::Instance().template Get<jit::VAddKernel<float>>(d);
    const float* x_data = x.data();
    const float* y_data = y.data();
    float* ztgt_data = ztgt.data();
    float* zref_data = zref.data();
    auto trefs = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      vadd_ref(d, x_data, y_data, zref_data);
    }
    auto trefe = GetCurrentUS();

#ifdef PADDLE_WITH_MKLML
    auto tmkls = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      vadd_mkl(d, x_data, y_data, zref_data);
    }
    auto tmkle = GetCurrentUS();
#endif

#if defined __AVX__ || defined __AVX2__
    if (d == 8) {
      auto si0 = GetCurrentUS();
      for (int i = 0; i < repeat; ++i) {
        vadd_intri8(d, x_data, y_data, zref_data);
      }
      auto si1 = GetCurrentUS();
T
tensor-tang 已提交
520 521 522 523
      VLOG(3) << "Vec size 8 intr takes: " << (si1 - si0) / repeat;
    }
#endif

T
tensor-tang 已提交
524 525 526 527 528 529 530
    auto ttgts = GetCurrentUS();
    for (int i = 0; i < repeat; ++i) {
      ker->Compute(d, x_data, y_data, ztgt_data);
    }
    auto ttgte = GetCurrentUS();

    VLOG(3) << "Vec size " << d << ": refer takes: " << (trefe - trefs) / repeat
T
tensor-tang 已提交
531
#ifdef PADDLE_WITH_MKLML
T
tensor-tang 已提交
532
            << " us, mkl takes: " << (tmkle - tmkls) / repeat << " us, "
T
tensor-tang 已提交
533
#else
T
tensor-tang 已提交
534
            << " us, "
T
tensor-tang 已提交
535
#endif
T
tensor-tang 已提交
536
            << "tgt takes: " << (ttgte - ttgts) / repeat;
T
tensor-tang 已提交
537 538 539 540 541 542
    for (int i = 0; i < d; ++i) {
      EXPECT_NEAR(ztgt_data[i], zref_data[i], 1e-3);
    }
  }
}

T
tensor-tang 已提交
543 544 545 546
TEST(JitKernel, pool) {
  namespace jit = paddle::operators::math::jitkernel;
  const int frame_size = 4;
  std::string act_gate = "sigmoid", act_cand = "tanh", act_cell = "tanh";
T
tensor-tang 已提交
547
  const auto& plstm1 =
T
tensor-tang 已提交
548 549 550 551
      jit::KernelPool::Instance()
          .template Get<jit::LSTMKernel<float>, int, const std::string&,
                        const std::string&, const std::string&>(
              frame_size, act_gate, act_cand, act_cell);
T
tensor-tang 已提交
552
  const auto& plstm2 =
T
tensor-tang 已提交
553 554 555 556
      jit::KernelPool::Instance()
          .template Get<jit::LSTMKernel<float>, int, const std::string&,
                        const std::string&, const std::string&>(
              frame_size, act_gate, act_cand, act_cell);
T
tensor-tang 已提交
557
  EXPECT_EQ(plstm1, plstm2);
T
tensor-tang 已提交
558

T
tensor-tang 已提交
559
  const auto& pvmul_f =
T
tensor-tang 已提交
560
      jit::KernelPool::Instance().template Get<jit::VMulKernel<float>>(4);
T
tensor-tang 已提交
561 562
  EXPECT_TRUE(std::dynamic_pointer_cast<const jit::Kernel>(plstm2) !=
              std::dynamic_pointer_cast<const jit::Kernel>(pvmul_f));
T
tensor-tang 已提交
563

T
tensor-tang 已提交
564
  const auto& pvmul_d =
T
tensor-tang 已提交
565
      jit::KernelPool::Instance().template Get<jit::VMulKernel<double>>(4);
T
tensor-tang 已提交
566 567
  EXPECT_TRUE(std::dynamic_pointer_cast<const jit::Kernel>(pvmul_f) !=
              std::dynamic_pointer_cast<const jit::Kernel>(pvmul_d));
T
tensor-tang 已提交
568 569

  const auto& pvmul_from_key = jit::KernelPool::Instance().Get("vmulf4");
T
tensor-tang 已提交
570
  EXPECT_EQ(pvmul_f, pvmul_from_key);
T
tensor-tang 已提交
571 572
  const auto& pvmul_from_key2 = jit::KernelPool::Instance().Get("vmulf5");
  EXPECT_TRUE(pvmul_from_key2 == nullptr);
T
tensor-tang 已提交
573
}