eager_utils_test.cc 10.5 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
// Copyright (c) 2021 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.

15
#include <sstream>
16 17 18

#include "gtest/gtest.h"

19
#include "paddle/fluid/eager/accumulation/accumulation_node.h"
20
#include "paddle/fluid/eager/eager_tensor.h"
21 22 23 24 25
#include "paddle/fluid/eager/grad_node_info.h"
#include "paddle/fluid/eager/tests/data_structure_tests/grad_node_test.h"
#include "paddle/fluid/eager/tests/test_utils.h"
#include "paddle/fluid/eager/utils.h"

26
#include "paddle/phi/api/lib/utils/allocator.h"
27 28 29
#include "paddle/phi/core/kernel_registry.h"

PD_DECLARE_KERNEL(full, CPU, ALL_LAYOUT);
30

31
namespace egr {
32 33 34

TEST(EagerUtils, AutoGradMeta) {
  // Construct Eager Tensor
35 36 37
  phi::DenseTensorMeta meta =
      phi::DenseTensorMeta(phi::DataType::FLOAT32, phi::make_ddim({1, 1}));
  std::shared_ptr<phi::DenseTensor> dt0 = std::make_shared<phi::DenseTensor>(
38 39 40
      std::make_unique<paddle::experimental::DefaultAllocator>(
          paddle::platform::CPUPlace())
          .get(),
41
      meta);
42
  dt0->mutable_data<float>(paddle::platform::CPUPlace())[0] = 10.0;
43
  paddle::experimental::Tensor et0 = paddle::experimental::Tensor(dt0);
44

45
  std::shared_ptr<phi::DenseTensor> dt1 = std::make_shared<phi::DenseTensor>(
46 47 48
      std::make_unique<paddle::experimental::DefaultAllocator>(
          paddle::platform::CPUPlace())
          .get(),
49
      meta);
50
  dt1->mutable_data<float>(paddle::platform::CPUPlace())[0] = 20.0;
51
  paddle::experimental::Tensor et1 = paddle::experimental::Tensor(dt1);
52 53 54 55 56 57 58 59 60 61

  // unsafe_autograd_meta()
  // autograd_meta()
  AutogradMeta* autograd_meta0 = EagerUtils::autograd_meta(&et0);
  AutogradMeta* autograd_meta1 = EagerUtils::autograd_meta(&et1);

  AutogradMeta* unsafe_autograd_meta_after =
      EagerUtils::unsafe_autograd_meta(et0);
  CHECK_NOTNULL(unsafe_autograd_meta_after);

62 63 64 65
  // NOTE: Since autograd_meta will be copied make sure it's not null
  std::vector<paddle::experimental::Tensor> ets = {et0, et1};
  auto test_node = std::make_shared<eager_test::GradTestNode>();

66
  std::vector<AutogradMeta*> autograd_metas = EagerUtils::autograd_meta(&ets);
67
  std::vector<AutogradMeta*> unsafe_autograd_metas =
68
      EagerUtils::unsafe_autograd_meta(ets);
69 70 71 72 73 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
  CHECK_NOTNULL(unsafe_autograd_metas[0]);
  CHECK_NOTNULL(unsafe_autograd_metas[1]);

  // Set Autograd Meta
  autograd_meta0->SetSingleOutRankWithSlot(0, 1);

  autograd_meta0->SetGradNode(test_node);

  // OutRankInfo()
  std::pair<size_t, size_t> out_rank_info0 = EagerUtils::OutRankInfo(et0);
  CHECK_EQ(static_cast<int>(out_rank_info0.first), 0);
  CHECK_EQ(static_cast<int>(out_rank_info0.second), 1);

  // grad_node()
  std::shared_ptr<GradNodeBase> grad_node0 = EagerUtils::grad_node(et0);
  CHECK_NOTNULL(grad_node0.get());

  EagerUtils::SetHistory(autograd_meta1, test_node);
  EagerUtils::SetHistory({autograd_meta1}, test_node);
  std::shared_ptr<GradNodeBase> grad_node1 = EagerUtils::grad_node(et1);
  CHECK_NOTNULL(grad_node1.get());

  // SetOutRankWithSlot()
  EagerUtils::SetOutRankWithSlot(autograd_meta1, 0);
  std::pair<size_t, size_t> out_rank_info1 = EagerUtils::OutRankInfo(et1);
  CHECK_EQ(static_cast<int>(out_rank_info1.first), 0);
  CHECK_EQ(static_cast<int>(out_rank_info1.second), 0);

  EagerUtils::SetOutRankWithSlot(&autograd_metas, 0);
  std::pair<size_t, size_t> out_rank_info2 = EagerUtils::OutRankInfo(et0);
  CHECK_EQ(static_cast<int>(out_rank_info2.first), 0);
  CHECK_EQ(static_cast<int>(out_rank_info2.second), 0);

  std::pair<size_t, size_t> out_rank_info3 = EagerUtils::OutRankInfo(et1);
  CHECK_EQ(static_cast<int>(out_rank_info3.first), 0);
  CHECK_EQ(static_cast<int>(out_rank_info3.second), 1);
}

107
template <typename T>
108 109
paddle::experimental::Tensor CreateTestCPUTensor(
    T val, const paddle::framework::DDim& ddim) {
110 111
  phi::DenseTensorMeta meta =
      phi::DenseTensorMeta(phi::DataType::FLOAT32, ddim);
112
  paddle::experimental::Tensor tensor;
113
  std::shared_ptr<phi::DenseTensor> dt = std::make_shared<phi::DenseTensor>(
114 115 116
      std::make_unique<paddle::experimental::DefaultAllocator>(
          paddle::platform::CPUPlace())
          .get(),
117
      meta);
118
  auto* dt_ptr = dt->mutable_data<T>(paddle::platform::CPUPlace());
119 120 121 122 123 124
  for (int64_t i = 0; i < dt->numel(); i++) {
    dt_ptr[i] = val;
  }
  tensor.set_impl(dt);
  return tensor;
}
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
TEST(EagerUtils, ComputeRequireGrad) {
  auto auto_grad0 = std::make_shared<egr::AutogradMeta>();
  auto auto_grad1 = std::make_shared<egr::AutogradMeta>();
  auto auto_grad2 = std::make_shared<egr::AutogradMeta>();
  auto auto_grad3 = std::make_shared<egr::AutogradMeta>();
  CHECK_EQ(auto_grad0->NumericStopGradient(), -1);
  VLOG(6) << "Single Test ComputeRequireGrad";
  auto_grad0->SetStopGradient(true);
  CHECK(egr::EagerUtils::ComputeRequireGrad(true, auto_grad0.get()) == false);
  CHECK(egr::EagerUtils::ComputeRequireGrad(false, auto_grad0.get()) == false);
  auto_grad0->SetStopGradient(false);
  CHECK(egr::EagerUtils::ComputeRequireGrad(false, auto_grad0.get()) == false);
  CHECK(egr::EagerUtils::ComputeRequireGrad(true, auto_grad0.get()) == true);

  VLOG(6) << "Multi Test ComputeRequireGrad";
  auto_grad0->SetStopGradient(false);
  auto_grad1->SetStopGradient(true);
  CHECK(egr::EagerUtils::ComputeRequireGrad(true, auto_grad0.get(),
                                            auto_grad1.get()) == true);
  CHECK(egr::EagerUtils::ComputeRequireGrad(false, auto_grad0.get(),
                                            auto_grad1.get()) == false);
  auto_grad0->SetStopGradient(true);
  CHECK(egr::EagerUtils::ComputeRequireGrad(true, auto_grad0.get(),
                                            auto_grad1.get()) == false);
  CHECK(egr::EagerUtils::ComputeRequireGrad(false, auto_grad0.get(),
                                            auto_grad1.get()) == false);
}

TEST(EagerUtils, PassStopGradient) {
  auto auto_grad0 = std::make_shared<egr::AutogradMeta>();
  auto auto_grad1 = std::make_shared<egr::AutogradMeta>();
  auto auto_grad2 = std::make_shared<egr::AutogradMeta>();
  auto auto_grad3 = std::make_shared<egr::AutogradMeta>();
  CHECK_EQ(auto_grad0->NumericStopGradient(), -1);
  VLOG(6) << "Test PassStopGradient";
  egr::EagerUtils::PassStopGradient(false, auto_grad0.get());
  CHECK(auto_grad0->StopGradient() == false);
  egr::EagerUtils::PassStopGradient(true, auto_grad0.get(), auto_grad1.get(),
                                    auto_grad2.get(), auto_grad3.get());
165
  CHECK(auto_grad0->StopGradient() == false);
166 167 168 169 170
  CHECK(auto_grad1->StopGradient() == true);
  CHECK(auto_grad2->StopGradient() == true);
  CHECK(auto_grad3->StopGradient() == true);
}

171
TEST(EagerUtils, TrySyncToVar) {
172
  paddle::framework::DDim ddim = phi::make_ddim({2, 4, 4, 4});
173
  auto tensor = CreateTestCPUTensor(5.0f, ddim);
174
  std::vector<std::shared_ptr<egr::EagerVariable>> var_bases = {
175
      egr::EagerUtils::TrySyncToVar(tensor)};
176 177 178 179 180 181 182 183 184 185 186 187 188

  paddle::framework::Variable* var = var_bases[0]->MutableVar();
  const auto& framework_tensor = var->Get<paddle::framework::LoDTensor>();

  const float* ptr = framework_tensor.data<float>();
  VLOG(6) << "Check Value for SyncToVarsSingle";
  CHECK_EQ(framework_tensor.numel(), tensor.numel());

  for (int i = 0; i < framework_tensor.numel(); i++) {
    CHECK_EQ(ptr[i], 5.0f);
  }
}

189
TEST(EagerUtils, TrySyncToVars) {
190
  paddle::framework::DDim ddim = phi::make_ddim({2, 4, 4, 4});
191 192
  std::vector<paddle::experimental::Tensor> tensors = {
      CreateTestCPUTensor(1.0f, ddim), CreateTestCPUTensor(2.0f, ddim)};
193

194
  std::vector<std::shared_ptr<egr::EagerVariable>> var_bases =
195
      egr::EagerUtils::TrySyncToVars(tensors);
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

  {
    paddle::framework::Variable* var = var_bases[0]->MutableVar();
    const auto& framework_tensor = var->Get<paddle::framework::LoDTensor>();

    const float* ptr = framework_tensor.data<float>();
    CHECK_EQ(framework_tensor.numel(), tensors[0].numel());

    for (int i = 0; i < framework_tensor.numel(); i++) {
      CHECK_EQ(ptr[i], 1.0);
    }
  }

  {
    paddle::framework::Variable* var = var_bases[1]->MutableVar();
    const auto& framework_tensor = var->Get<paddle::framework::LoDTensor>();

    const float* ptr = framework_tensor.data<float>();
    VLOG(6) << "Check Value for SyncToVarsMultiple";
    CHECK_EQ(framework_tensor.numel(), tensors[0].numel());

    for (int i = 0; i < framework_tensor.numel(); i++) {
      CHECK_EQ(ptr[i], 2.0);
    }
  }
}

223 224
TEST(EagerUtils, CreateVars) {
  VLOG(6) << "Check CreateVars";
225
  std::vector<std::shared_ptr<egr::EagerVariable>> outs =
226
      egr::EagerUtils::CreateVars(2);
227
  CHECK_EQ(outs.size(), size_t(2));
228
  CHECK(outs[0]->Var().IsInitialized() == false);
229
}
230

231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252
TEST(EagerUtils, GetGradAccumulationNode) {
  VLOG(6) << "Check GetGradAccumulationNode";
  paddle::experimental::Tensor t0("test_tensor");
  ASSERT_EQ(egr::EagerUtils::GetGradAccumulationNode(t0), nullptr);
  auto autograd_ptr0 = egr::EagerUtils::autograd_meta(&t0);
  autograd_ptr0->SetStopGradient(true);
  ASSERT_EQ(egr::EagerUtils::GetGradAccumulationNode(t0), nullptr);
  autograd_ptr0->SetStopGradient(false);
  auto res = std::dynamic_pointer_cast<egr::GradNodeAccumulation>(
      egr::EagerUtils::GetGradAccumulationNode(t0));
  ASSERT_TRUE(res != nullptr);
  auto res2 = egr::EagerUtils::GetGradAccumulationNode(t0);
  ASSERT_EQ(res2.get(), res.get());
  autograd_ptr0->SetStopGradient(true);
  auto res3 = egr::EagerUtils::GetGradAccumulationNode(t0);
  ASSERT_EQ(res3, nullptr);
  autograd_ptr0->SetStopGradient(false);
  autograd_ptr0->SetGradNode(
      std::make_shared<eager_test::GradTestNode>(1, 2.0, 3));
  ASSERT_ANY_THROW(egr::EagerUtils::GetGradAccumulationNode(t0));
}

253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268
TEST(EagerUtils, FillZeroForEmptyGradInputs) {
  std::vector<std::vector<paddle::experimental::Tensor>> grads = {
      std::vector<paddle::experimental::Tensor>(1)};
  std::vector<std::vector<GradSlotMeta>> slot_metas = {
      std::vector<GradSlotMeta>(1)};

  phi::DenseTensorMeta tensor_meta;
  tensor_meta.dtype = paddle::experimental::DataType::FLOAT32;
  tensor_meta.dims = {2, 4};
  slot_metas[0][0].SetTensorMeta(tensor_meta);
  slot_metas[0][0].SetPlace(phi::CPUPlace());

  EagerUtils::FillZeroForEmptyGradInputs(&grads, slot_metas);
  eager_test::CompareTensorWithValue<float>(grads[0][0], 0.0);
}

269
}  // namespace egr