/* Copyright (c) 2016 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 #include #include #include "glog/logging.h" #include "gtest/gtest.h" #include "paddle/fluid/framework/init.h" #include "paddle/fluid/platform/device_context.h" #include "paddle/fluid/platform/dynload/nccl.h" #include "paddle/fluid/platform/enforce.h" #include "paddle/fluid/platform/gpu_info.h" static int dev_count = 0; namespace paddle { namespace platform { TEST(NCCL, init) { std::vector comms; comms.resize(dev_count); PADDLE_ENFORCE(dynload::ncclCommInitAll(comms.data(), dev_count, nullptr)); for (int i = 0; i < dev_count; ++i) { dynload::ncclCommDestroy(comms[i]); } } template struct PerThreadData { thrust::device_vector send_buff; thrust::device_vector recv_buff; CUDADeviceContext dev_ctx; T* SendBuff() { return thrust::raw_pointer_cast(send_buff.data()); } T* RecvBuff() { return thrust::raw_pointer_cast(recv_buff.data()); } PerThreadData(int gpu_id, size_t size) : dev_ctx(CUDAPlace(gpu_id)) { send_buff.resize(size); for (size_t i = 0; i < size; ++i) { send_buff[i] = static_cast(i); } recv_buff.resize(size); } }; static constexpr int ELEM_COUNT = 10000; TEST(NCCL, all_reduce) { std::vector comms; comms.resize(dev_count); VLOG(1) << "Initializing ncclComm"; dynload::ncclCommInitAll(comms.data(), dev_count, nullptr); VLOG(1) << "ncclComm initialized"; VLOG(1) << "Creating thread data"; std::vector>> data; data.reserve(dev_count); for (int i = 0; i < dev_count; ++i) { VLOG(1) << "Creating thread data for device " << i; SetDeviceId(i); data.emplace_back(new PerThreadData(i, ELEM_COUNT)); } VLOG(1) << "Thread data created"; VLOG(1) << "Check send_buf data"; for (int i = 0; i < dev_count; ++i) { VLOG(1) << "Check on device " << i; SetDeviceId(i); thrust::host_vector tmp = data[i]->send_buff; for (size_t j = 0; j < tmp.size(); ++j) { ASSERT_NEAR(static_cast(j), tmp[j], 1e-5); } } VLOG(1) << "Invoking ncclAllReduce"; dynload::ncclGroupStart(); for (int i = 0; i < dev_count; ++i) { VLOG(1) << "Invoking ncclAllReduce with device " << i; SetDeviceId(i); PADDLE_ENFORCE(dynload::ncclAllReduce( data[i]->SendBuff(), data[i]->RecvBuff(), ELEM_COUNT, ncclDouble, ncclSum, comms[i], data[i]->dev_ctx.stream())); VLOG(1) << "Invoked ncclAllReduce for device " << i; } dynload::ncclGroupEnd(); VLOG(1) << "Invoked ncclAllReduce"; VLOG(1) << "Sync devices"; for (int i = 0; i < dev_count; ++i) { VLOG(1) << "Sync device " << i; SetDeviceId(i); data[i]->dev_ctx.Wait(); } VLOG(1) << "device synced"; for (int i = 0; i < dev_count; ++i) { SetDeviceId(i); VLOG(1) << "Checking vector on device " << i; thrust::host_vector tmp = data[i]->recv_buff; for (size_t j = 0; j < tmp.size(); ++j) { auto elem = static_cast(j); elem *= dev_count; ASSERT_NEAR(tmp[j], elem, 1e-4); } } for (int i = 0; i < dev_count; ++i) { dynload::ncclCommDestroy(comms[i]); } } } // namespace platform } // namespace paddle int main(int argc, char** argv) { // FIXME(tonyyang-svail): // Due to the driver issue on our CI, disable for now return 0; dev_count = paddle::platform::GetCUDADeviceCount(); if (dev_count <= 1) { LOG(WARNING) << "Cannot test multi-gpu nccl, because the CUDA device count is " << dev_count; return 0; } std::vector places; places.emplace_back(paddle::platform::CPUPlace()); int count = paddle::platform::GetCUDADeviceCount(); for (int i = 0; i < count; ++i) { places.emplace_back(paddle::platform::CUDAPlace(i)); } VLOG(0) << " DeviceCount " << count; paddle::platform::DeviceContextPool::Init(places); testing::InitGoogleTest(&argc, argv); return RUN_ALL_TESTS(); }