test.cc 25.1 KB
Newer Older
T
tensor-tang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 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

25 26
static double acc = 1e-5;

T
tensor-tang 已提交
27 28 29 30 31 32 33 34 35 36 37 38 39 40 41
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) {
42
      EXPECT_NEAR(target[i], refer[i], acc);
T
tensor-tang 已提交
43 44 45 46 47 48 49 50
    }
  } else {
    for (int i = 0; i < n; ++i) {
      EXPECT_EQ(target[i], refer[i]);
    }
  }
}

T
tensor-tang 已提交
51 52
std::vector<int> TestSizes() {
  std::vector<int> s;
T
tensor-tang 已提交
53
  for (int i = 1; i < 32; ++i) {
T
tensor-tang 已提交
54 55
    s.push_back(i);
  }
T
tensor-tang 已提交
56 57 58 59
  // test some large size
  s.push_back(100);
  s.push_back(1000);
  s.push_back(2000);
T
tensor-tang 已提交
60 61 62
  return s;
}

T
tensor-tang 已提交
63
namespace jit = paddle::operators::jit;
64
using CPUPlace = paddle::platform::CPUPlace;
T
tensor-tang 已提交
65 66 67

template <typename KernelTuples, typename... Args>
struct TestFuncWithRefer {
68 69 70
  void operator()(const typename KernelTuples::func_type tgt, Args... args) {
    LOG(FATAL) << "Should specify this function.";
  }
T
tensor-tang 已提交
71 72 73 74 75 76
};

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,
77 78
                  const std::vector<T>& x, const std::vector<T>& y,
                  const std::vector<T>& zref) {
T
tensor-tang 已提交
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
    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);
  }
};

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

template <typename T>
struct TestFuncWithRefer<jit::XRNTuples<T>, std::vector<T>, T> {
  void operator()(const typename jit::XRNTuples<T>::func_type tgt,
                  const std::vector<T>& x, const T ref_res) {
    EXPECT_TRUE(tgt != nullptr);
    T tgt_res;
    tgt(x.data(), &tgt_res, x.size());
    ExpectEQ<T>(&tgt_res, &ref_res, 1);
  }
};

T
tensor-tang 已提交
159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181
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>,
182 183
                         std::vector<T>, std::vector<T>, std::vector<T>,
                         typename jit::LSTMTuples<T>::attr_type> {
T
tensor-tang 已提交
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
  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();

210
    jit::lstm_t step;
T
tensor-tang 已提交
211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227
    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>,
228 229
                         std::vector<T>,
                         typename jit::GRUTuples<T>::attr_type> {
T
tensor-tang 已提交
230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245
  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();
246
    jit::gru_t step;
T
tensor-tang 已提交
247 248 249 250 251 252 253 254
    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);
  }
};

255
template <typename T>
256 257
struct TestFuncWithRefer<jit::SeqPoolTuples<T>, std::vector<T>, std::vector<T>,
                         typename jit::SeqPoolTuples<T>::attr_type> {
258 259 260 261 262 263 264 265 266 267 268 269 270 271 272
  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
273
template <typename T>
274 275
struct TestFuncWithRefer<jit::MatMulTuples<T>, std::vector<T>, std::vector<T>,
                         std::vector<T>, int, int, int> {
T
tensor-tang 已提交
276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292
  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);
  }
};

293 294
template <jit::KernelType KT, typename KernelTuples, typename PlaceType,
          typename... Args>
T
tensor-tang 已提交
295 296 297 298 299 300 301 302 303 304 305 306 307 308 309
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
310
      auto i = dynamic_cast<const jit::KernelMore<KernelTuples>*>(impl.get());
T
tensor-tang 已提交
311 312
      if (i && i->UseMe(attr)) {
        auto more = i->GetFunc();
T
tensor-tang 已提交
313
        VLOG(10) << "Test More Kernel : " << i->ImplType();
T
tensor-tang 已提交
314 315 316 317 318
        test(more, args...);
      }
    }
  }
  // test result from Get function
T
tensor-tang 已提交
319
  // VLOG(10) << "Test Get function ";
T
tensor-tang 已提交
320 321
  auto tgt = jit::Get<KT, KernelTuples, PlaceType>(attr);
  test(tgt, args...);
T
tensor-tang 已提交
322 323
}

324
template <jit::KernelType KT, typename T, typename PlaceType>
325
void TestXYZNKernel() {
326
  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
T
tensor-tang 已提交
327
  for (int d : TestSizes()) {
328
    auto ref = jit::GetRefer<KT, jit::XYZNTuples<T>>();
T
tensor-tang 已提交
329 330
    EXPECT_TRUE(ref != nullptr);

T
tensor-tang 已提交
331
    std::vector<T> x(d), y(d), zref(d);
T
tensor-tang 已提交
332 333 334
    RandomVec<T>(d, x.data());
    RandomVec<T>(d, y.data());

T
tensor-tang 已提交
335 336 337 338 339 340 341 342 343 344 345
    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
346
    ref(x_data, y_data, zref_data, d);
T
tensor-tang 已提交
347 348 349 350 351
    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
352 353
    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
354 355
  }
}
T
tensor-tang 已提交
356

357
template <jit::KernelType KT, typename T, typename PlaceType>
358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377
void TestAXYNKernel() {
  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
378 379
    TestAllImpls<KT, jit::AXYNTuples<T>, PlaceType, T, std::vector<T>,
                 std::vector<T>>(d, a, x, yref);
380 381 382
  }
}

383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398
template <jit::KernelType KT, typename T, typename PlaceType>
void TestXRNKernel() {
  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
  for (int d : TestSizes()) {
    auto ref = jit::GetRefer<KT, jit::XRNTuples<T>>();
    EXPECT_TRUE(ref != nullptr);
    std::vector<T> x(d);
    RandomVec<T>(d, x.data());
    T ref_res;
    ref(x.data(), &ref_res, d);
    TestAllImpls<KT, jit::XRNTuples<T>, PlaceType, std::vector<T>, T>(d, x,
                                                                      ref_res);
  }
}

template <jit::KernelType KT, typename T, typename PlaceType>
399 400 401 402 403 404 405 406
void TestXYNKernel() {
  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
407
    RandomVec<T>(d, x.data(), -2.f, 2.f);
408 409 410 411 412 413 414 415 416 417
    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
418 419
    TestAllImpls<KT, jit::XYNTuples<T>, PlaceType, std::vector<T>,
                 std::vector<T>>(d, x, yref);
420 421 422
  }
}

423
template <jit::KernelType KT, typename T, typename PlaceType>
T
tensor-tang 已提交
424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439
void TestLSTMKernel() {
  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);
440 441
            RandomVec<T>(3 * d, wp.data(), -1.f, 1.f);
            RandomVec<T>(d, ct_1.data(), -1.f, 1.f);
T
tensor-tang 已提交
442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460
            // 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
461
            VLOG(10) << attr;
T
tensor-tang 已提交
462 463 464 465
            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
466 467 468 469 470 471 472
          }
        }
      }
    }
  }
}

473
template <jit::KernelType KT, typename T, typename PlaceType>
474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497
void TestGRUKernel() {
  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
498
        VLOG(10) << attr;
T
tensor-tang 已提交
499 500 501
        TestAllImpls<KT, jit::GRUTuples<T>, PlaceType, std::vector<T>,
                     std::vector<T>, std::vector<T>>(attr, xsrc, ht_1, ht_ref,
                                                     attr);
502 503 504 505 506
      }
    }
  }
}

507
template <jit::KernelType KT, typename T, typename PlaceType>
508 509
void TestSeqPoolKernel() {
  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
510 511
  std::vector<jit::SeqPoolType> pool_types = {
      jit::SeqPoolType::kSum, jit::SeqPoolType::kAvg, jit::SeqPoolType::kSqrt};
512
  for (auto type : pool_types) {
T
tensor-tang 已提交
513 514 515 516
    for (int w : TestSizes()) {
      jit::seq_pool_attr_t attr(w, type);
      for (int h : TestSizes()) {
        attr.h = h;
517 518 519 520 521 522 523 524 525 526 527 528 529 530 531
        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);
      }
    }
  }
}

532
template <jit::KernelType KT, typename T, typename PlaceType>
T
tensor-tang 已提交
533 534
void TestMatMulKernel() {
  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
535 536 537
  auto last_acc = acc;
  // TODO(intel): this should be acc issue of MKL
  acc = 1e-3;
T
tensor-tang 已提交
538 539 540 541 542 543
  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);
544 545
        RandomVec<T>(m * k, a.data(), -2.f, 2.f);
        RandomVec<T>(k * n, b.data(), -2.f, 2.f);
T
tensor-tang 已提交
546 547 548 549 550 551 552 553 554
        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);
      }
    }
  }
555
  acc = last_acc;
T
tensor-tang 已提交
556 557
}

558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583
template <jit::KernelType KT, typename T, typename PlaceType>
void TestSoftmaxKernel() {
  VLOG(10) << "===== Test JITKernel " << jit::to_string(KT);
  for (int bs : {1, 2, 10}) {
    for (int n : TestSizes()) {
      auto ref = jit::GetRefer<KT, jit::SoftmaxTuples<T>>();
      EXPECT_TRUE(ref != nullptr);
      std::vector<T> x(bs * n), y(bs * n);
      RandomVec<T>(bs * n, x.data(), -2.f, 2.f);
      const T* x_data = x.data();
      T* y_data = y.data();

      std::vector<T> xinp(x.size());  // inplace test
      std::copy(x.begin(), x.end(), xinp.begin());
      ref(x_data, y_data, n, bs);
      T* xinp_data = xinp.data();
      ref(xinp_data, xinp_data, n, bs);
      ExpectEQ<T>(xinp_data, y_data, n * bs);

      TestAllImpls<KT, jit::SoftmaxTuples<T>, PlaceType, std::vector<T>,
                   std::vector<T>>(n, x, y, n, bs);
    }
  }
}

template <jit::KernelType KT, typename T, typename PlaceType>
T
tensor-tang 已提交
584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635
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
636
// XYZNTuple
T
tensor-tang 已提交
637
TEST(JITKernel, kVMul) {
638 639
  TestXYZNKernel<jit::kVMul, float, CPUPlace>();
  TestXYZNKernel<jit::kVMul, double, CPUPlace>();
T
tensor-tang 已提交
640 641
}

T
tensor-tang 已提交
642
TEST(JITKernel, kVAdd) {
643 644
  TestXYZNKernel<jit::kVAdd, float, CPUPlace>();
  TestXYZNKernel<jit::kVAdd, double, CPUPlace>();
T
tensor-tang 已提交
645 646
}

T
tensor-tang 已提交
647
TEST(JITKernel, kVAddRelu) {
648 649
  TestXYZNKernel<jit::kVAddRelu, float, CPUPlace>();
  TestXYZNKernel<jit::kVAddRelu, double, CPUPlace>();
T
tensor-tang 已提交
650 651
}

T
tensor-tang 已提交
652
TEST(JITKernel, kVSub) {
653 654
  TestXYZNKernel<jit::kVSub, float, CPUPlace>();
  TestXYZNKernel<jit::kVSub, double, CPUPlace>();
T
tensor-tang 已提交
655 656 657
}

// AXYNTuples
T
tensor-tang 已提交
658
TEST(JITKernel, kVScal) {
659 660
  TestAXYNKernel<jit::kVScal, float, CPUPlace>();
  TestAXYNKernel<jit::kVScal, double, CPUPlace>();
T
tensor-tang 已提交
661 662
}

T
tensor-tang 已提交
663
TEST(JITKernel, kVAddBias) {
664 665 666 667 668 669 670 671 672 673 674 675 676
  TestAXYNKernel<jit::kVAddBias, float, CPUPlace>();
  TestAXYNKernel<jit::kVAddBias, double, CPUPlace>();
}

// XRNTuples
TEST(JITKernel, kHMax) {
  TestXRNKernel<jit::kHMax, float, CPUPlace>();
  TestXRNKernel<jit::kHMax, double, CPUPlace>();
}

TEST(JITKernel, kHSum) {
  TestXRNKernel<jit::kHSum, float, CPUPlace>();
  TestXRNKernel<jit::kHSum, double, CPUPlace>();
T
tensor-tang 已提交
677 678 679
}

// XYNTuples
T
tensor-tang 已提交
680
TEST(JITKernel, kVRelu) {
681 682
  TestXYNKernel<jit::kVRelu, float, CPUPlace>();
  TestXYNKernel<jit::kVRelu, double, CPUPlace>();
T
tensor-tang 已提交
683 684
}

T
tensor-tang 已提交
685
TEST(JITKernel, kVIdentity) {
686 687
  TestXYNKernel<jit::kVIdentity, float, CPUPlace>();
  TestXYNKernel<jit::kVIdentity, double, CPUPlace>();
T
tensor-tang 已提交
688 689
}

T
tensor-tang 已提交
690
TEST(JITKernel, kVSquare) {
691 692
  TestXYNKernel<jit::kVSquare, float, CPUPlace>();
  TestXYNKernel<jit::kVSquare, double, CPUPlace>();
T
tensor-tang 已提交
693 694
}

T
tensor-tang 已提交
695
TEST(JITKernel, kVExp) {
696 697
  TestXYNKernel<jit::kVExp, float, CPUPlace>();
  TestXYNKernel<jit::kVExp, double, CPUPlace>();
T
tensor-tang 已提交
698 699
}

T
tensor-tang 已提交
700
TEST(JITKernel, kVSigmoid) {
701 702
  TestXYNKernel<jit::kVSigmoid, float, CPUPlace>();
  TestXYNKernel<jit::kVSigmoid, double, CPUPlace>();
T
tensor-tang 已提交
703 704
}

T
tensor-tang 已提交
705
TEST(JITKernel, kVTanh) {
706 707
  TestXYNKernel<jit::kVTanh, float, CPUPlace>();
  TestXYNKernel<jit::kVTanh, double, CPUPlace>();
T
tensor-tang 已提交
708 709 710
}

// LSTM
T
tensor-tang 已提交
711
TEST(JITKernel, kLSTMCtHt) {
712 713
  TestLSTMKernel<jit::kLSTMCtHt, float, CPUPlace>();
  TestLSTMKernel<jit::kLSTMCtHt, double, CPUPlace>();
T
tensor-tang 已提交
714 715
}

T
tensor-tang 已提交
716
TEST(JITKernel, kLSTMC1H1) {
717 718
  TestLSTMKernel<jit::kLSTMC1H1, float, CPUPlace>();
  TestLSTMKernel<jit::kLSTMC1H1, double, CPUPlace>();
T
tensor-tang 已提交
719 720 721
}

// GRU
T
tensor-tang 已提交
722
TEST(JITKernel, kGRUH1) {
723 724
  TestGRUKernel<jit::kGRUH1, float, CPUPlace>();
  TestGRUKernel<jit::kGRUH1, double, CPUPlace>();
725 726
}

T
tensor-tang 已提交
727
TEST(JITKernel, kGRUHtPart1) {
728 729
  TestGRUKernel<jit::kGRUHtPart1, float, CPUPlace>();
  TestGRUKernel<jit::kGRUHtPart1, double, CPUPlace>();
730 731
}

T
tensor-tang 已提交
732
TEST(JITKernel, kGRUHtPart2) {
733 734
  TestGRUKernel<jit::kGRUHtPart2, float, CPUPlace>();
  TestGRUKernel<jit::kGRUHtPart2, double, CPUPlace>();
735 736
}

737
TEST(JITKernel, kSeqPool) {
738 739
  TestSeqPoolKernel<jit::kSeqPool, float, CPUPlace>();
  TestSeqPoolKernel<jit::kSeqPool, double, CPUPlace>();
740 741
}

T
tensor-tang 已提交
742
TEST(JITKernel, kMatMul) {
743 744 745 746 747 748 749
  TestMatMulKernel<jit::kMatMul, float, CPUPlace>();
  TestMatMulKernel<jit::kMatMul, double, CPUPlace>();
}

TEST(JITKernel, kSoftmax) {
  TestSoftmaxKernel<jit::kSoftmax, float, CPUPlace>();
  TestSoftmaxKernel<jit::kSoftmax, double, CPUPlace>();
T
tensor-tang 已提交
750 751
}

T
tensor-tang 已提交
752
TEST(JITKernel, kNCHW16CMulNC) {
753 754
  TestNCHW16CMulNCKernel<jit::kNCHW16CMulNC, float, CPUPlace>();
  TestNCHW16CMulNCKernel<jit::kNCHW16CMulNC, double, CPUPlace>();
T
tensor-tang 已提交
755 756
}

757
// TODO(yihua/TJ): add crf decoding and layer norm unit tests
T
tensor-tang 已提交
758

759 760 761
TEST(JITKernel, pool) {
  // TODO(TJ): add some test
}