test.cc 23.6 KB
Newer Older
T
tensor-tang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
/* 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 <random>
#include <string>
#include <vector>
#include "gflags/gflags.h"
#include "glog/logging.h"
#include "gtest/gtest.h"
T
tensor-tang 已提交
21
#include "paddle/fluid/operators/jit/kernels.h"
T
tensor-tang 已提交
22
#include "paddle/fluid/platform/cpu_info.h"
T
tensor-tang 已提交
23
#include "paddle/fluid/platform/place.h"
T
tensor-tang 已提交
24

T
tensor-tang 已提交
25 26 27 28 29 30 31 32 33 34 35 36 37 38 39
template <typename T>
void RandomVec(const int n, T* a, const T lower = static_cast<T>(-20.f),
               const T upper = static_cast<T>(20.f)) {
  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);
  }
}

template <typename T>
void ExpectEQ(const T* target, const T* refer, int n) {
  if (std::is_floating_point<T>::value) {
    for (int i = 0; i < n; ++i) {
40
      EXPECT_NEAR(target[i], refer[i], 1e-5);
T
tensor-tang 已提交
41 42 43 44 45 46 47 48
    }
  } else {
    for (int i = 0; i < n; ++i) {
      EXPECT_EQ(target[i], refer[i]);
    }
  }
}

T
tensor-tang 已提交
49 50
std::vector<int> TestSizes() {
  std::vector<int> s;
T
tensor-tang 已提交
51
  for (int i = 1; i < 32; ++i) {
T
tensor-tang 已提交
52 53
    s.push_back(i);
  }
T
tensor-tang 已提交
54 55 56 57
  // test some large size
  s.push_back(100);
  s.push_back(1000);
  s.push_back(2000);
T
tensor-tang 已提交
58 59 60
  return s;
}

T
tensor-tang 已提交
61 62 63 64 65 66 67 68 69 70 71
namespace jit = paddle::operators::jit;

template <typename KernelTuples, typename... Args>
struct TestFuncWithRefer {
  void operator()(const typename KernelTuples::func_type tgt, Args... args) {}
};

template <typename T>
struct TestFuncWithRefer<jit::XYZNTuples<T>, std::vector<T>, std::vector<T>,
                         std::vector<T>> {
  void operator()(const typename jit::XYZNTuples<T>::func_type tgt,
72 73
                  const std::vector<T>& x, const std::vector<T>& y,
                  const std::vector<T>& zref) {
T
tensor-tang 已提交
74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213
    EXPECT_TRUE(tgt != nullptr);
    EXPECT_EQ(zref.size(), x.size());
    EXPECT_EQ(zref.size(), y.size());
    const T* x_data = x.data();
    const T* y_data = y.data();
    const T* zref_data = zref.data();
    const int d = zref.size();

    std::vector<T> ztgt(d);
    T* ztgt_data = ztgt.data();
    // test normal
    tgt(x_data, y_data, ztgt_data, d);
    ExpectEQ<T>(ztgt_data, zref_data, d);
    // test inplace x
    std::copy(x.begin(), x.end(), ztgt.begin());
    tgt(ztgt_data, y_data, ztgt_data, d);
    ExpectEQ<T>(ztgt_data, zref_data, d);
    // test inplace y
    std::copy(y.begin(), y.end(), ztgt.begin());
    tgt(x_data, ztgt_data, ztgt_data, d);
    ExpectEQ<T>(ztgt_data, zref_data, d);
  }
};

template <typename T>
struct TestFuncWithRefer<jit::AXYNTuples<T>, T, std::vector<T>,
                         std::vector<T>> {
  void operator()(const typename jit::AXYNTuples<T>::func_type tgt, const T a,
                  const std::vector<T>& x, const std::vector<T>& yref) {
    EXPECT_TRUE(tgt != nullptr);
    EXPECT_EQ(yref.size(), x.size());
    const T* x_data = x.data();
    const T* yref_data = yref.data();
    const int d = yref.size();
    std::vector<T> ytgt(d);
    T* ytgt_data = ytgt.data();
    // test normal
    tgt(&a, x_data, ytgt_data, d);
    ExpectEQ<T>(ytgt_data, yref_data, d);
    // test inplace x
    std::copy(x.begin(), x.end(), ytgt.begin());
    tgt(&a, ytgt_data, ytgt_data, d);
    ExpectEQ<T>(ytgt_data, yref_data, d);
  }
};

template <typename T>
struct TestFuncWithRefer<jit::XYNTuples<T>, std::vector<T>, std::vector<T>> {
  void operator()(const typename jit::XYNTuples<T>::func_type tgt,
                  const std::vector<T>& x, const std::vector<T>& yref) {
    EXPECT_TRUE(tgt != nullptr);
    EXPECT_EQ(yref.size(), x.size());
    const T* x_data = x.data();
    const T* yref_data = yref.data();
    const int d = yref.size();
    std::vector<T> ytgt(d);
    T* ytgt_data = ytgt.data();
    // test normal
    tgt(x_data, ytgt_data, d);
    ExpectEQ<T>(ytgt_data, yref_data, d);
    // test inplace x
    std::copy(x.begin(), x.end(), ytgt.begin());
    tgt(ytgt_data, ytgt_data, d);
    ExpectEQ<T>(ytgt_data, yref_data, d);
  }
};

template <typename T>
struct TestFuncWithRefer<jit::LSTMTuples<T>, std::vector<T>, std::vector<T>,
                         std::vector<T>, std::vector<T>, std::vector<T>> {
  void operator()(const typename jit::LSTMTuples<T>::func_type tgt,
                  const std::vector<T>& xsrc, const std::vector<T>& wp,
                  const std::vector<T>& ct_1, const std::vector<T>& ct_ref,
                  const std::vector<T>& ht_ref,
                  const typename jit::LSTMTuples<T>::attr_type& attr) {
    EXPECT_TRUE(tgt != nullptr);
    EXPECT_EQ(ct_ref.size(), ht_ref.size());
    EXPECT_EQ(ct_1.size(), ht_ref.size());
    EXPECT_EQ(xsrc.size(), 4 * ht_ref.size());
    EXPECT_EQ(wp.size(), 3 * ht_ref.size());

    // x could be changed after compute, so copy to save src
    int d = ht_ref.size();
    std::vector<T> x(xsrc.size()), ct(ct_ref.size()), ht(ht_ref.size());
    std::vector<T> checked(2 * d);
    std::copy(xsrc.begin(), xsrc.end(), x.begin());

    const T* ct_1_data = ct_1.data();
    const T* wp_data = wp.data();
    const T* ct_ref_data = ct_ref.data();
    const T* ht_ref_data = ht_ref.data();
    T* x_data = x.data();
    T* ct_data = ct.data();
    T* ht_data = ht.data();
    T* checked_data = checked.data();

    paddle::operators::jit::lstm_t step;
    step.gates = x_data;
    step.ct_1 = ct_1_data;
    step.ct = ct_data;
    step.ht = ht_data;
    if (attr.use_peephole) {
      step.wp = wp_data;
      step.checked = checked_data;
    }

    tgt(&step, &attr);
    ExpectEQ<T>(ct_data, ct_ref_data, d);
    ExpectEQ<T>(ht_data, ht_ref_data, d);
  }
};

template <typename T>
struct TestFuncWithRefer<jit::GRUTuples<T>, std::vector<T>, std::vector<T>,
                         std::vector<T>> {
  void operator()(const typename jit::GRUTuples<T>::func_type tgt,
                  const std::vector<T>& xsrc, const std::vector<T>& ht_1,
                  const std::vector<T>& ht_ref,
                  const typename jit::GRUTuples<T>::attr_type& attr) {
    EXPECT_TRUE(tgt != nullptr);
    EXPECT_EQ(ht_1.size(), ht_ref.size());
    EXPECT_EQ(xsrc.size(), 3 * ht_ref.size());

    // x could be changed after compute, so copy to save src
    int d = ht_ref.size();
    std::vector<T> x(xsrc.size()), ht(ht_ref.size());
    std::copy(xsrc.begin(), xsrc.end(), x.begin());
    const T* ht_1_data = ht_1.data();
    const T* ht_ref_data = ht_ref.data();
    T* x_data = x.data();
    T* ht_data = ht.data();
    paddle::operators::jit::gru_t step;
    step.gates = x_data;
    step.ht_1 = ht_1_data;
    step.ht = ht_data;
    tgt(&step, &attr);
    ExpectEQ<T>(ht_data, ht_ref_data, d);
  }
};

214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231
template <typename T>
struct TestFuncWithRefer<jit::SeqPoolTuples<T>, std::vector<T>,
                         std::vector<T>> {
  void operator()(const typename jit::SeqPoolTuples<T>::func_type tgt,
                  const std::vector<T>& x, const std::vector<T>& yref,
                  const typename jit::SeqPoolTuples<T>::attr_type& attr) {
    EXPECT_TRUE(tgt != nullptr);
    EXPECT_EQ(x.size() % yref.size(), 0);
    int w = yref.size();
    std::vector<T> y(w);
    const T* x_data = x.data();
    const T* yref_data = yref.data();
    T* y_data = y.data();
    tgt(x_data, y_data, &attr);
    ExpectEQ<T>(y_data, yref_data, w);
  }
};

T
tensor-tang 已提交
232
template <typename T>
233 234
struct TestFuncWithRefer<jit::MatMulTuples<T>, std::vector<T>, std::vector<T>,
                         std::vector<T>, int, int, int> {
T
tensor-tang 已提交
235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251
  void operator()(const typename jit::MatMulTuples<T>::func_type tgt,
                  const std::vector<T>& a, const std::vector<T>& b,
                  const std::vector<T>& cref, int m, int n, int k) {
    EXPECT_TRUE(tgt != nullptr);
    EXPECT_EQ(a.size(), static_cast<size_t>(m * k));
    EXPECT_EQ(b.size(), static_cast<size_t>(k * n));
    EXPECT_EQ(cref.size(), static_cast<size_t>(m * n));
    std::vector<T> c(cref.size());
    const T* a_data = a.data();
    const T* b_data = b.data();
    const T* cref_data = cref.data();
    T* c_data = c.data();
    tgt(a_data, b_data, c_data, m, n, k);
    ExpectEQ<T>(c_data, cref_data, m * n);
  }
};

T
tensor-tang 已提交
252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268
template <paddle::operators::jit::KernelType KT, typename KernelTuples,
          typename PlaceType, typename... Args>
void TestAllImpls(const typename KernelTuples::attr_type& attr, Args... args) {
  TestFuncWithRefer<KernelTuples, Args...> test;
  // test jitcode
  auto jitcode = jit::GetJitCode<KT, KernelTuples, PlaceType>(attr);
  if (jitcode) {
    VLOG(10) << "Test Jitcode Kernel ";
    test(jitcode, args...);
  }
  // test all impls in more
  jit::KernelKey kkey(KT, PlaceType());
  auto& pool = jit::KernelPool().Instance().AllKernels();
  auto iter = pool.find(kkey);
  if (iter != pool.end()) {
    auto& impls = iter->second;
    for (auto& impl : impls) {
T
tensor-tang 已提交
269
      auto i = dynamic_cast<const jit::KernelMore<KernelTuples>*>(impl.get());
T
tensor-tang 已提交
270 271
      if (i && i->UseMe(attr)) {
        auto more = i->GetFunc();
T
tensor-tang 已提交
272
        VLOG(10) << "Test More Kernel : " << i->ImplType();
T
tensor-tang 已提交
273 274 275 276 277
        test(more, args...);
      }
    }
  }
  // test result from Get function
T
tensor-tang 已提交
278
  // VLOG(10) << "Test Get function ";
T
tensor-tang 已提交
279 280
  auto tgt = jit::Get<KT, KernelTuples, PlaceType>(attr);
  test(tgt, args...);
T
tensor-tang 已提交
281 282
}

283 284
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
void TestXYZNKernel() {
T
tensor-tang 已提交
285
  namespace jit = paddle::operators::jit;
286
  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
T
tensor-tang 已提交
287
  for (int d : TestSizes()) {
288
    auto ref = jit::GetRefer<KT, jit::XYZNTuples<T>>();
T
tensor-tang 已提交
289 290
    EXPECT_TRUE(ref != nullptr);

T
tensor-tang 已提交
291
    std::vector<T> x(d), y(d), zref(d);
T
tensor-tang 已提交
292 293 294
    RandomVec<T>(d, x.data());
    RandomVec<T>(d, y.data());

T
tensor-tang 已提交
295 296 297 298 299 300 301 302 303 304 305
    std::vector<T> xinp(d), yinp(d);  // inplace test
    std::copy(x.begin(), x.end(), xinp.begin());
    std::copy(y.begin(), y.end(), yinp.begin());

    const T* x_data = x.data();
    const T* y_data = y.data();
    T* zref_data = zref.data();
    T* xinp_data = xinp.data();
    T* yinp_data = yinp.data();

    // test refer code inplace
T
tensor-tang 已提交
306
    ref(x_data, y_data, zref_data, d);
T
tensor-tang 已提交
307 308 309 310 311
    ref(x_data, yinp_data, yinp_data, d);
    ref(xinp_data, y_data, xinp_data, d);
    ExpectEQ<T>(xinp_data, zref_data, d);
    ExpectEQ<T>(yinp_data, zref_data, d);

T
tensor-tang 已提交
312 313
    TestAllImpls<KT, jit::XYZNTuples<T>, PlaceType, std::vector<T>,
                 std::vector<T>, std::vector<T>>(d, x, y, zref);
T
tensor-tang 已提交
314 315
  }
}
T
tensor-tang 已提交
316

317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
void TestAXYNKernel() {
  namespace jit = paddle::operators::jit;
  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
  for (int d : TestSizes()) {
    auto ref = jit::GetRefer<KT, jit::AXYNTuples<T>>();
    EXPECT_TRUE(ref != nullptr);

    const T a = static_cast<T>(3);
    std::vector<T> x(d), yref(d);
    std::vector<T> xinp(d);  // inplace test
    RandomVec<T>(d, x.data());
    std::copy(x.begin(), x.end(), xinp.begin());

    const T* x_data = x.data();
    T* yref_data = yref.data();
    T* xinp_data = xinp.data();
    // test refer code inplace
    ref(&a, x_data, yref_data, d);
    ref(&a, xinp_data, xinp_data, d);
    ExpectEQ<T>(xinp_data, yref_data, d);

T
tensor-tang 已提交
339 340
    TestAllImpls<KT, jit::AXYNTuples<T>, PlaceType, T, std::vector<T>,
                 std::vector<T>>(d, a, x, yref);
341 342 343
  }
}

344 345 346 347 348 349 350 351 352 353
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
void TestXYNKernel() {
  namespace jit = paddle::operators::jit;
  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
  for (int d : TestSizes()) {
    auto ref = jit::GetRefer<KT, jit::XYNTuples<T>>();
    EXPECT_TRUE(ref != nullptr);

    std::vector<T> x(d), yref(d);
    std::vector<T> xinp(d);  // inplace test
354
    RandomVec<T>(d, x.data(), -2.f, 2.f);
355 356 357 358 359 360 361 362 363 364
    std::copy(x.begin(), x.end(), xinp.begin());

    const T* x_data = x.data();
    T* yref_data = yref.data();
    T* xinp_data = xinp.data();
    // test refer code inplace
    ref(x_data, yref_data, d);
    ref(xinp_data, xinp_data, d);
    ExpectEQ<T>(xinp_data, yref_data, d);

T
tensor-tang 已提交
365 366
    TestAllImpls<KT, jit::XYNTuples<T>, PlaceType, std::vector<T>,
                 std::vector<T>>(d, x, yref);
367 368 369
  }
}

T
tensor-tang 已提交
370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
void TestLSTMKernel() {
  namespace jit = paddle::operators::jit;
  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
  std::vector<std::string> all_acts = {"sigmoid", "tanh", "relu", "identity"};
  for (int d : TestSizes()) {
    for (bool use_peephole : {true, false}) {
      for (auto& act_gate : all_acts) {
        for (auto& act_cand : all_acts) {
          for (auto& act_cell : all_acts) {
            const jit::lstm_attr_t attr(
                d, jit::to_kerneltype(act_gate), jit::to_kerneltype(act_cand),
                jit::to_kerneltype(act_cell), use_peephole);
            auto ref = jit::GetRefer<KT, jit::LSTMTuples<T>>();
            EXPECT_TRUE(ref != nullptr);
            std::vector<T> xsrc(4 * d), wp(3 * d), ct_1(d);
            std::vector<T> ct_ref(d), ht_ref(d), checked(2 * d);
            RandomVec<T>(4 * d, xsrc.data(), -2.f, 2.f);
            RandomVec<T>(3 * d, wp.data(), -2.f, 2.f);
            RandomVec<T>(d, ct_1.data(), -2.f, 2.f);
            // x could be changed after compute, so copy to save src
            std::vector<T> x(xsrc.size());
            std::copy(xsrc.begin(), xsrc.end(), x.begin());
            const T* ct_1_data = ct_1.data();
            const T* wp_data = wp.data();
            T* x_data = x.data();
            T* checked_data = checked.data();
            T* ct_ref_data = ct_ref.data();
            T* ht_ref_data = ht_ref.data();
            jit::lstm_t step;
            step.gates = x_data;
            step.ct_1 = ct_1_data;
            step.ct = ct_ref_data;
            step.ht = ht_ref_data;
            if (use_peephole) {
              step.wp = wp_data;
              step.checked = checked_data;
            }
            ref(&step, &attr);
T
tensor-tang 已提交
409
            VLOG(10) << attr;
T
tensor-tang 已提交
410 411 412 413
            TestAllImpls<KT, jit::LSTMTuples<T>, PlaceType, std::vector<T>,
                         std::vector<T>, std::vector<T>, std::vector<T>,
                         std::vector<T>>(attr, xsrc, wp, ct_1, ct_ref, ht_ref,
                                         attr);
T
tensor-tang 已提交
414 415 416 417 418 419 420
          }
        }
      }
    }
  }
}

421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
void TestGRUKernel() {
  namespace jit = paddle::operators::jit;
  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
  std::vector<std::string> all_acts = {"sigmoid", "tanh", "relu", "identity"};
  for (int d : TestSizes()) {
    for (auto& act_gate : all_acts) {
      for (auto& act_cand : all_acts) {
        const jit::gru_attr_t attr(d, jit::to_kerneltype(act_gate),
                                   jit::to_kerneltype(act_cand));
        auto ref = jit::GetRefer<KT, jit::GRUTuples<T>>();
        EXPECT_TRUE(ref != nullptr);
        std::vector<T> xsrc(3 * d), ht_1(d), ht_ref(d);
        RandomVec<T>(3 * d, xsrc.data(), -2.f, 2.f);
        RandomVec<T>(d, ht_1.data(), -2.f, 2.f);
        // x could be changed after compute, so copy to save src
        std::vector<T> x(xsrc.size());
        std::copy(xsrc.begin(), xsrc.end(), x.begin());
        const T* ht_1_data = ht_1.data();
        T* x_data = x.data();
        T* ht_ref_data = ht_ref.data();
        jit::gru_t step;
        step.gates = x_data;
        step.ht_1 = ht_1_data;
        step.ht = ht_ref_data;
        ref(&step, &attr);
T
tensor-tang 已提交
447
        VLOG(10) << attr;
T
tensor-tang 已提交
448 449 450
        TestAllImpls<KT, jit::GRUTuples<T>, PlaceType, std::vector<T>,
                     std::vector<T>, std::vector<T>>(attr, xsrc, ht_1, ht_ref,
                                                     attr);
451 452 453 454 455
      }
    }
  }
}

456 457 458
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
void TestSeqPoolKernel() {
  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
459 460
  std::vector<jit::SeqPoolType> pool_types = {
      jit::SeqPoolType::kSum, jit::SeqPoolType::kAvg, jit::SeqPoolType::kSqrt};
461
  for (auto type : pool_types) {
T
tensor-tang 已提交
462 463 464 465
    for (int w : TestSizes()) {
      jit::seq_pool_attr_t attr(w, type);
      for (int h : TestSizes()) {
        attr.h = h;
466 467 468 469 470 471 472 473 474 475 476 477 478 479 480
        auto ref = jit::GetRefer<KT, jit::SeqPoolTuples<T>>();
        EXPECT_TRUE(ref != nullptr);
        std::vector<T> x(h * w), yref(w);
        RandomVec<T>(h * w, x.data(), -2.f, 2.f);
        const T* x_data = x.data();
        T* yref_data = yref.data();
        ref(x_data, yref_data, &attr);
        VLOG(10) << attr;
        TestAllImpls<KT, jit::SeqPoolTuples<T>, PlaceType, std::vector<T>,
                     std::vector<T>>(attr, x, yref, attr);
      }
    }
  }
}

T
tensor-tang 已提交
481 482 483 484 485 486 487 488 489
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
void TestMatMulKernel() {
  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
  for (int m : {1, 2, 3, 4}) {
    for (int n : {1, 2, 3, 4}) {
      for (int k : TestSizes()) {
        auto ref = jit::GetRefer<KT, jit::MatMulTuples<T>>();
        EXPECT_TRUE(ref != nullptr);
        std::vector<T> a(m * k), b(k * n), c(m * n);
490 491
        RandomVec<T>(m * k, a.data(), -0.2f, 0.2f);
        RandomVec<T>(k * n, b.data(), -0.2f, 0.2f);
T
tensor-tang 已提交
492 493 494 495 496 497 498 499 500 501 502
        const T* a_data = a.data();
        const T* b_data = b.data();
        T* c_data = c.data();
        ref(a_data, b_data, c_data, m, n, k);
        TestAllImpls<KT, jit::MatMulTuples<T>, PlaceType, std::vector<T>,
                     std::vector<T>, std::vector<T>>(k, a, b, c, m, n, k);
      }
    }
  }
}

T
tensor-tang 已提交
503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555
template <paddle::operators::jit::KernelType KT, typename T, typename PlaceType>
void TestNCHW16CMulNCKernel() {
  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
  const int n = 3, c = 16 * 4, h = 10, w = 10;
  auto ref = jit::GetRefer<KT, jit::NCHW16CMulNCTuples<T>>();
  EXPECT_TRUE(ref != nullptr);
  int sz = n * c * h * w;
  std::vector<T> x(sz), y(n * c), zref(sz);
  std::vector<T> ztgt(sz), zjit(sz);
  RandomVec<T>(sz, x.data(), -2.f, 2.f);
  RandomVec<T>(n * c, y.data(), -2.f, 2.f);

  const T* x_data = x.data();
  const T* y_data = y.data();
  T* zref_data = zref.data();
  T* ztgt_data = ztgt.data();
  T* zjit_data = zjit.data();
  constexpr int simd_width = ZMM_FLOAT_BLOCK;
  int C = c / simd_width;
  auto tgt = jit::Get<KT, jit::NCHW16CMulNCTuples<T>, PlaceType>(0);
  auto jitcode = jit::GetJitCode<KT, jit::NCHW16CMulNCTuples<T>, PlaceType>(0);
  EXPECT_TRUE(tgt != nullptr);

  if (std::is_same<T, float>::value &&
      paddle::platform::MayIUse(paddle::platform::avx512f)) {
    EXPECT_TRUE(jitcode != nullptr);
  }
  for (int ni = 0; ni < n; ni++) {
    for (int ci = 0; ci < C; ci++) {
      auto ptr_x =
          x_data + ni * C * h * w * simd_width + ci * h * w * simd_width;
      auto ptr_y = y_data + ni * C * simd_width + ci * simd_width;
      auto ptr_zref =
          zref_data + ni * C * h * w * simd_width + ci * h * w * simd_width;
      auto ptr_ztgt =
          ztgt_data + ni * C * h * w * simd_width + ci * h * w * simd_width;

      ref(ptr_x, ptr_y, ptr_zref, h, w);
      tgt(ptr_x, ptr_y, ptr_ztgt, h, w);

      if (jitcode) {
        auto ptr_zjit =
            zjit_data + ni * C * h * w * simd_width + ci * h * w * simd_width;
        jitcode(ptr_x, ptr_y, ptr_zjit, h, w);
      }
    }
  }
  ExpectEQ<T>(ztgt_data, zref_data, sz);
  if (jitcode) {
    ExpectEQ<T>(zjit_data, zref_data, sz);
  }
}

T
tensor-tang 已提交
556
// XYZNTuple
T
tensor-tang 已提交
557
TEST(JITKernel, kVMul) {
T
tensor-tang 已提交
558
  namespace jit = paddle::operators::jit;
T
tensor-tang 已提交
559 560
  TestXYZNKernel<jit::kVMul, float, paddle::platform::CPUPlace>();
  TestXYZNKernel<jit::kVMul, double, paddle::platform::CPUPlace>();
T
tensor-tang 已提交
561 562
}

T
tensor-tang 已提交
563
TEST(JITKernel, kVAdd) {
T
tensor-tang 已提交
564
  namespace jit = paddle::operators::jit;
T
tensor-tang 已提交
565 566
  TestXYZNKernel<jit::kVAdd, float, paddle::platform::CPUPlace>();
  TestXYZNKernel<jit::kVAdd, double, paddle::platform::CPUPlace>();
T
tensor-tang 已提交
567 568
}

T
tensor-tang 已提交
569
TEST(JITKernel, kVAddRelu) {
T
tensor-tang 已提交
570
  namespace jit = paddle::operators::jit;
T
tensor-tang 已提交
571 572
  TestXYZNKernel<jit::kVAddRelu, float, paddle::platform::CPUPlace>();
  TestXYZNKernel<jit::kVAddRelu, double, paddle::platform::CPUPlace>();
T
tensor-tang 已提交
573 574
}

T
tensor-tang 已提交
575
TEST(JITKernel, kVSub) {
T
tensor-tang 已提交
576
  namespace jit = paddle::operators::jit;
T
tensor-tang 已提交
577 578
  TestXYZNKernel<jit::kVSub, float, paddle::platform::CPUPlace>();
  TestXYZNKernel<jit::kVSub, double, paddle::platform::CPUPlace>();
T
tensor-tang 已提交
579 580 581
}

// AXYNTuples
T
tensor-tang 已提交
582
TEST(JITKernel, kVScal) {
T
tensor-tang 已提交
583
  namespace jit = paddle::operators::jit;
T
tensor-tang 已提交
584 585
  TestAXYNKernel<jit::kVScal, float, paddle::platform::CPUPlace>();
  TestAXYNKernel<jit::kVScal, double, paddle::platform::CPUPlace>();
T
tensor-tang 已提交
586 587
}

T
tensor-tang 已提交
588
TEST(JITKernel, kVAddBias) {
T
tensor-tang 已提交
589
  namespace jit = paddle::operators::jit;
T
tensor-tang 已提交
590 591
  TestAXYNKernel<jit::kVAddBias, float, paddle::platform::CPUPlace>();
  TestAXYNKernel<jit::kVAddBias, double, paddle::platform::CPUPlace>();
T
tensor-tang 已提交
592 593 594
}

// XYNTuples
T
tensor-tang 已提交
595
TEST(JITKernel, kVRelu) {
T
tensor-tang 已提交
596
  namespace jit = paddle::operators::jit;
T
tensor-tang 已提交
597 598
  TestXYNKernel<jit::kVRelu, float, paddle::platform::CPUPlace>();
  TestXYNKernel<jit::kVRelu, double, paddle::platform::CPUPlace>();
T
tensor-tang 已提交
599 600
}

T
tensor-tang 已提交
601
TEST(JITKernel, kVIdentity) {
T
tensor-tang 已提交
602
  namespace jit = paddle::operators::jit;
T
tensor-tang 已提交
603 604
  TestXYNKernel<jit::kVIdentity, float, paddle::platform::CPUPlace>();
  TestXYNKernel<jit::kVIdentity, double, paddle::platform::CPUPlace>();
T
tensor-tang 已提交
605 606
}

T
tensor-tang 已提交
607
TEST(JITKernel, kVExp) {
T
tensor-tang 已提交
608
  namespace jit = paddle::operators::jit;
T
tensor-tang 已提交
609 610
  TestXYNKernel<jit::kVExp, float, paddle::platform::CPUPlace>();
  TestXYNKernel<jit::kVExp, double, paddle::platform::CPUPlace>();
T
tensor-tang 已提交
611 612
}

T
tensor-tang 已提交
613
TEST(JITKernel, kVSigmoid) {
T
tensor-tang 已提交
614
  namespace jit = paddle::operators::jit;
T
tensor-tang 已提交
615 616
  TestXYNKernel<jit::kVSigmoid, float, paddle::platform::CPUPlace>();
  TestXYNKernel<jit::kVSigmoid, double, paddle::platform::CPUPlace>();
T
tensor-tang 已提交
617 618
}

T
tensor-tang 已提交
619
TEST(JITKernel, kVTanh) {
T
tensor-tang 已提交
620
  namespace jit = paddle::operators::jit;
T
tensor-tang 已提交
621 622
  TestXYNKernel<jit::kVTanh, float, paddle::platform::CPUPlace>();
  TestXYNKernel<jit::kVTanh, double, paddle::platform::CPUPlace>();
T
tensor-tang 已提交
623 624 625
}

// LSTM
T
tensor-tang 已提交
626
TEST(JITKernel, kLSTMCtHt) {
T
tensor-tang 已提交
627
  namespace jit = paddle::operators::jit;
T
tensor-tang 已提交
628 629
  TestLSTMKernel<jit::kLSTMCtHt, float, paddle::platform::CPUPlace>();
  TestLSTMKernel<jit::kLSTMCtHt, double, paddle::platform::CPUPlace>();
T
tensor-tang 已提交
630 631
}

T
tensor-tang 已提交
632
TEST(JITKernel, kLSTMC1H1) {
T
tensor-tang 已提交
633
  namespace jit = paddle::operators::jit;
T
tensor-tang 已提交
634 635
  TestLSTMKernel<jit::kLSTMC1H1, float, paddle::platform::CPUPlace>();
  TestLSTMKernel<jit::kLSTMC1H1, double, paddle::platform::CPUPlace>();
T
tensor-tang 已提交
636 637 638
}

// GRU
T
tensor-tang 已提交
639
TEST(JITKernel, kGRUH1) {
640
  namespace jit = paddle::operators::jit;
T
tensor-tang 已提交
641 642
  TestGRUKernel<jit::kGRUH1, float, paddle::platform::CPUPlace>();
  TestGRUKernel<jit::kGRUH1, double, paddle::platform::CPUPlace>();
643 644
}

T
tensor-tang 已提交
645
TEST(JITKernel, kGRUHtPart1) {
646
  namespace jit = paddle::operators::jit;
T
tensor-tang 已提交
647 648
  TestGRUKernel<jit::kGRUHtPart1, float, paddle::platform::CPUPlace>();
  TestGRUKernel<jit::kGRUHtPart1, double, paddle::platform::CPUPlace>();
649 650
}

T
tensor-tang 已提交
651
TEST(JITKernel, kGRUHtPart2) {
652
  namespace jit = paddle::operators::jit;
T
tensor-tang 已提交
653 654
  TestGRUKernel<jit::kGRUHtPart2, float, paddle::platform::CPUPlace>();
  TestGRUKernel<jit::kGRUHtPart2, double, paddle::platform::CPUPlace>();
655 656
}

657 658 659 660 661 662
TEST(JITKernel, kSeqPool) {
  namespace jit = paddle::operators::jit;
  TestSeqPoolKernel<jit::kSeqPool, float, paddle::platform::CPUPlace>();
  TestSeqPoolKernel<jit::kSeqPool, double, paddle::platform::CPUPlace>();
}

T
tensor-tang 已提交
663 664 665 666 667 668
TEST(JITKernel, kMatMul) {
  namespace jit = paddle::operators::jit;
  TestMatMulKernel<jit::kMatMul, float, paddle::platform::CPUPlace>();
  TestMatMulKernel<jit::kMatMul, double, paddle::platform::CPUPlace>();
}

T
tensor-tang 已提交
669
TEST(JITKernel, kNCHW16CMulNC) {
T
tensor-tang 已提交
670
  namespace jit = paddle::operators::jit;
T
tensor-tang 已提交
671
  TestNCHW16CMulNCKernel<jit::kNCHW16CMulNC, float,
T
tensor-tang 已提交
672
                         paddle::platform::CPUPlace>();
T
tensor-tang 已提交
673
  TestNCHW16CMulNCKernel<jit::kNCHW16CMulNC, double,
T
tensor-tang 已提交
674 675 676
                         paddle::platform::CPUPlace>();
}

677
// TODO(yihua/TJ): add crf decoding and layer norm unit tests
T
tensor-tang 已提交
678

679 680 681
TEST(JITKernel, pool) {
  // TODO(TJ): add some test
}