nccl_op_test.cu 8.9 KB
Newer Older
D
Dong Zhihong 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

   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. */

D
Dong Zhihong 已提交
15 16
#include <glog/logging.h>
#include <gtest/gtest.h>
D
Dong Zhihong 已提交
17
#include <algorithm>
D
Dong Zhihong 已提交
18
#include <memory>
D
Dong Zhihong 已提交
19 20 21
#include <mutex>
#include <thread>
#include <utility>
D
Dong Zhihong 已提交
22
#include <vector>
D
Dong Zhihong 已提交
23

D
Dong Zhihong 已提交
24 25
#include "paddle/framework/block_desc.h"
#include "paddle/framework/op_desc.h"
D
Dong Zhihong 已提交
26
#include "paddle/framework/op_registry.h"
D
Dong Zhihong 已提交
27 28
#include "paddle/framework/program_desc.h"
#include "paddle/framework/var_desc.h"
D
Dong Zhihong 已提交
29
#include "paddle/operators/nccl/nccl_gpu_common.h"
D
Dong Zhihong 已提交
30 31 32
#include "paddle/platform/device_context.h"
#include "paddle/platform/enforce.h"
#include "paddle/platform/gpu_info.h"
D
Dong Zhihong 已提交
33
#include "paddle/platform/place.h"
D
Dong Zhihong 已提交
34

D
Dong Zhihong 已提交
35
USE_NO_KERNEL_OP(ncclInit);
D
Dong Zhihong 已提交
36 37
USE_GPU_ONLY_OP(ncclAllReduce);
USE_GPU_ONLY_OP(ncclReduce);
D
Dong Zhihong 已提交
38
USE_GPU_ONLY_OP(ncclBcast);
D
Dong Zhihong 已提交
39

D
Dong Zhihong 已提交
40 41 42
namespace f = paddle::framework;
namespace p = paddle::platform;

D
Dong Zhihong 已提交
43
static std::vector<int> gpu_list;
D
Dong Zhihong 已提交
44 45 46

// test data amount
const f::DDim kDims = {100, 100};
D
Dong Zhihong 已提交
47

D
Dong Zhihong 已提交
48 49 50 51 52 53 54 55 56 57 58 59
// nccl op common tester, init communicator.
class NCCLTester : public ::testing::Test {
 public:
  virtual void SetUp() override {
    cpu_ctx = new p::CPUDeviceContext(p::CPUPlace());
    for (size_t i = 0; i < gpu_list.size(); ++i) {
      p::GPUPlace place(i);
      dev_ctxs.emplace_back(new p::CUDADeviceContext(place));
    }

    NCCLInitOp();
  }
D
Dong Zhihong 已提交
60

D
Dong Zhihong 已提交
61 62 63 64 65
  virtual void TearDown() override {
    for (auto &device_context : dev_ctxs) {
      delete device_context;
    }
  }
D
Dong Zhihong 已提交
66

D
Dong Zhihong 已提交
67 68
  void NCCLInitOp() {
    std::unique_ptr<f::OpDescBind> op1(new f::OpDescBind);
D
Dong Zhihong 已提交
69

D
Dong Zhihong 已提交
70 71 72
    op1->SetType("ncclInit");
    op1->SetOutput("Communicator", {"comm"});
    op1->SetAttr("gpus", {gpu_list});
D
Dong Zhihong 已提交
73

D
Dong Zhihong 已提交
74 75
    auto *var = g_scope.Var("comm");
    var->GetMutable<p::Communicator>();
76

D
Dong Zhihong 已提交
77 78 79 80 81 82 83 84 85 86
    auto op = f::OpRegistry::CreateOp(*op1);
    VLOG(1) << "invoke NCCLInitOp.";
    op->Run(g_scope, *cpu_ctx);
    VLOG(1) << "NCCLInitOp finished.";
  }

  template <class T>
  void PerThreadProgram(int gpu_id, const f::OpDescBind &op_desc,
                        f::Scope *scope) {
    std::unique_lock<std::mutex> lk(mu);
D
Dong Zhihong 已提交
87
    const f::OpDescBind *op1 = &op_desc;
D
Dong Zhihong 已提交
88 89 90 91 92 93 94

    p::GPUPlace place(gpu_id);
    auto &ctx = dev_ctxs.at(gpu_id);

    auto *send_tensor = scope->Var("st")->GetMutable<f::LoDTensor>();
    auto *recv_tensor = scope->Var("rt")->GetMutable<f::LoDTensor>();

D
Dong Zhihong 已提交
95 96 97 98 99 100 101 102 103 104
    if (!send_tensor->numel()) {
      send_tensor->Resize(kDims);
      send_tensor->mutable_data<T>(kDims, place);

      std::vector<T> send_vector(f::product(kDims), gpu_id);
      send_tensor->CopyFromVector<T>(send_vector, *ctx);
      ctx->Wait();
      VLOG(1) << "Send Tensor filled with elements " << send_tensor->numel();
    }

D
Dong Zhihong 已提交
105
    lk.unlock();
D
Dong Zhihong 已提交
106

D
Dong Zhihong 已提交
107 108 109 110
    PADDLE_ENFORCE(send_tensor->numel() == f::product(kDims),
                   "Tensor numel not match!");

    auto op = f::OpRegistry::CreateOp(*op1);
D
Dong Zhihong 已提交
111

D
Dong Zhihong 已提交
112
    VLOG(1) << "Device : " << gpu_id << " invoke " << op_desc.Type();
D
Dong Zhihong 已提交
113 114
    VLOG(1) << " send_tensor : " << send_tensor->numel()
            << " recv_tensor : " << recv_tensor->numel();
D
Dong Zhihong 已提交
115 116 117
    op->Run(*scope, *ctx);
    VLOG(1) << "Device : " << gpu_id << " finished " << op_desc.Type();
  }
118

D
Dong Zhihong 已提交
119 120 121 122 123 124 125
 public:
  std::vector<p::DeviceContext *> dev_ctxs;
  p::DeviceContext *cpu_ctx;
  f::Scope g_scope;
  std::mutex mu;
};

126 127 128 129 130 131 132 133 134 135 136 137 138
// ncclInitOp with desc
TEST(NCCL, ncclInitOp) {
  std::unique_ptr<f::OpDescBind> op_desc(new f::OpDescBind);

  op_desc->SetType("ncclInit");
  op_desc->SetOutput("Communicator", {"x1"});
  op_desc->SetAttr("gpus", {gpu_list});

  f::Scope g_scope;
  std::unique_ptr<p::DeviceContext> ctx(new p::CPUDeviceContext(p::CPUPlace()));

  auto *var = g_scope.Var("x1");
  var->GetMutable<p::Communicator>();
D
Dong Zhihong 已提交
139

140 141 142 143 144 145 146 147
  auto op = f::OpRegistry::CreateOp(*op_desc);
  VLOG(1) << "invoke NCCLInitOp.";
  op->Run(g_scope, *ctx.get());
  VLOG(1) << "NCCLInitOp finished.";
}

// ncclAllReduceOp with desc
TEST_F(NCCLTester, ncclAllReduceOp) {
D
Dong Zhihong 已提交
148
  std::unique_ptr<f::OpDescBind> op2(new f::OpDescBind);
149 150
  op2->SetType("ncclAllReduce");
  op2->SetInput("X", {"st"});
D
Dong Zhihong 已提交
151
  op2->SetInput("Communicator", {"comm"});
152
  op2->SetOutput("Out", {"rt"});
D
Dong Zhihong 已提交
153 154

  std::vector<f::Scope *> dev_scopes;
D
Dong Zhihong 已提交
155 156

  std::vector<std::thread> ths;
D
Dong Zhihong 已提交
157 158 159 160 161 162 163

  for (size_t i = 0; i < gpu_list.size(); ++i) {
    dev_scopes.emplace_back(&g_scope.NewScope());
    std::thread th(&NCCLTester::PerThreadProgram<float>, this, gpu_list[i],
                   *op2.get(), dev_scopes[i]);
    ths.emplace_back(std::move(th));
  }
D
Dong Zhihong 已提交
164

D
Dong Zhihong 已提交
165 166 167
  for (size_t i = 0; i < gpu_list.size(); ++i) {
    ths[i].join();
  }
D
Dong Zhihong 已提交
168

D
Dong Zhihong 已提交
169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186
  // check results
  float result = std::accumulate(gpu_list.begin(), gpu_list.end(), 0);

  for (size_t i = 0; i < dev_scopes.size(); ++i) {
    p::CPUPlace cpu_place;
    p::GPUPlace gpu_place(gpu_list[i]);

    auto &recv_tensor = dev_scopes[i]->FindVar("rt")->Get<f::LoDTensor>();
    auto *rt = recv_tensor.data<float>();
    auto *result_tensor = dev_scopes[i]->Var("ct")->GetMutable<f::LoDTensor>();
    result_tensor->Resize(kDims);
    auto *ct = result_tensor->mutable_data<float>(cpu_place);

    paddle::memory::Copy(
        cpu_place, ct, p::GPUPlace(gpu_list[i]), rt,
        recv_tensor.numel() * sizeof(float),
        static_cast<p::CUDADeviceContext *>(dev_ctxs[i])->stream());

D
dangqingqing 已提交
187
    for (int64_t j = 0; j < f::product(kDims); ++j) {
D
Dong Zhihong 已提交
188 189 190 191
      ASSERT_NEAR(ct[j], result, 1e-5);
    }
  }
}
D
Dong Zhihong 已提交
192

D
dzhwinter 已提交
193
// ncclReduceOp with desc
194 195 196 197 198 199 200
TEST_F(NCCLTester, ncclReduceOp) {
  std::unique_ptr<f::OpDescBind> op2(new f::OpDescBind);
  const int kRoot = 0;
  op2->SetType("ncclReduce");
  op2->SetInput("X", {"st"});
  op2->SetInput("Communicator", {"comm"});
  op2->SetOutput("Out", {"rt"});
D
dzhwinter 已提交
201
  op2->SetAttr("root", kRoot);
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

  std::vector<f::Scope *> dev_scopes;

  std::vector<std::thread> ths;

  for (size_t i = 0; i < gpu_list.size(); ++i) {
    dev_scopes.emplace_back(&g_scope.NewScope());
    std::thread th(&NCCLTester::PerThreadProgram<float>, this, gpu_list[i],
                   *op2.get(), dev_scopes[i]);
    ths.emplace_back(std::move(th));
  }

  for (size_t i = 0; i < gpu_list.size(); ++i) {
    ths[i].join();
  }

  // check results on
  float result = std::accumulate(gpu_list.begin(), gpu_list.end(), 0);

  p::CPUPlace cpu_place;
  p::GPUPlace gpu_place(gpu_list[kRoot]);

  auto &recv_tensor = dev_scopes[kRoot]->FindVar("rt")->Get<f::LoDTensor>();
  auto *rt = recv_tensor.data<float>();
  auto *result_tensor =
      dev_scopes[kRoot]->Var("ct")->GetMutable<f::LoDTensor>();
  result_tensor->Resize(kDims);
  auto *ct = result_tensor->mutable_data<float>(cpu_place);

  paddle::memory::Copy(
      cpu_place, ct, p::GPUPlace(gpu_list[kRoot]), rt,
      recv_tensor.numel() * sizeof(float),
      static_cast<p::CUDADeviceContext *>(dev_ctxs[kRoot])->stream());

D
dangqingqing 已提交
236
  for (int64_t j = 0; j < f::product(kDims); ++j) {
237 238 239 240
    ASSERT_NEAR(ct[j], result, 1e-5);
  }
}

D
dzhwinter 已提交
241
// ncclBcastOp with desc
242 243 244 245 246 247 248
TEST_F(NCCLTester, ncclBcastOp) {
  std::unique_ptr<f::OpDescBind> op2(new f::OpDescBind);
  const int kRoot = 5;
  op2->SetType("ncclBcast");
  op2->SetInput("X", {"st"});
  op2->SetInput("Communicator", {"comm"});
  op2->SetOutput("Out", {"rt"});
D
dzhwinter 已提交
249
  op2->SetAttr("root", kRoot);
250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283

  std::vector<f::Scope *> dev_scopes;

  std::vector<std::thread> ths;

  for (size_t i = 0; i < gpu_list.size(); ++i) {
    dev_scopes.emplace_back(&g_scope.NewScope());
    std::thread th(&NCCLTester::PerThreadProgram<float>, this, gpu_list[i],
                   *op2.get(), dev_scopes[i]);
    ths.emplace_back(std::move(th));
  }

  for (size_t i = 0; i < gpu_list.size(); ++i) {
    ths[i].join();
  }

  const int idx = 1;
  // check results on
  float result = kRoot;

  p::CPUPlace cpu_place;
  p::GPUPlace gpu_place(gpu_list[idx]);

  auto &recv_tensor = dev_scopes[idx]->FindVar("rt")->Get<f::LoDTensor>();
  auto *rt = recv_tensor.data<float>();
  auto *result_tensor = dev_scopes[idx]->Var("ct")->GetMutable<f::LoDTensor>();
  result_tensor->Resize(kDims);
  auto *ct = result_tensor->mutable_data<float>(cpu_place);

  paddle::memory::Copy(
      cpu_place, ct, p::GPUPlace(gpu_list[idx]), rt,
      recv_tensor.numel() * sizeof(float),
      static_cast<p::CUDADeviceContext *>(dev_ctxs[idx])->stream());

D
dangqingqing 已提交
284
  for (int64_t j = 0; j < f::product(kDims); ++j) {
285 286 287 288
    ASSERT_NEAR(ct[j], result, 1e-5);
  }
}

D
Dong Zhihong 已提交
289
int main(int argc, char **argv) {
D
Dong Zhihong 已提交
290
  const int dev_count = p::GetCUDADeviceCount();
D
Dong Zhihong 已提交
291 292 293 294 295 296
  if (dev_count <= 1) {
    LOG(WARNING)
        << "Cannot test multi-gpu nccl, because the CUDA device count is "
        << dev_count;
    return 0;
  }
D
Dong Zhihong 已提交
297 298 299 300

  for (int i = 0; i < dev_count; ++i) {
    gpu_list.emplace_back(i);
  }
D
Dong Zhihong 已提交
301
  testing::InitGoogleTest(&argc, argv);
D
Dong Zhihong 已提交
302 303 304

  // device context should be release before scope.
  // otherwise driver will down.
D
Dong Zhihong 已提交
305 306
  return RUN_ALL_TESTS();
}