merge conflict, test=develop

revert-16734-refine/test_imperative_transformer
lujun 6 years ago
commit 92c8ac8a74

@ -75,7 +75,6 @@ option(WITH_INFERENCE_API_TEST "Test fluid inference high-level api interface"
option(WITH_SYSTEM_BLAS "Use system blas library" OFF)
option(PY_VERSION "Compile PaddlePaddle with python3 support" ${PY_VERSION})
option(WITH_FAST_MATH "Make use of fast math library, might affect the precision to some extent" ON)
option(WITH_WBAES "Compile PaddlePaddle with WBAES support" ON)
# PY_VERSION
if(NOT PY_VERSION)
@ -149,7 +148,6 @@ include(external/dlpack)
include(external/snappy) # download snappy
include(external/snappystream) # download snappystream
include(external/warpctc) # download, build, install warpctc
include(external/wbaes) # download wbaes
if (NOT WIN32)
# there is no official support of nccl, cupti in windows

@ -157,7 +157,3 @@ endif(WITH_BRPC_RDMA)
if(ON_INFER)
add_definitions(-DPADDLE_ON_INFERENCE)
endif(ON_INFER)
if(WITH_WBAES)
add_definitions(-DPADDLE_WITH_WBAES)
endif(WITH_WBAES)

@ -1,71 +0,0 @@
# Copyright (c) 2019 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.
IF(NOT ${WITH_WBAES})
return()
ENDIF(NOT ${WITH_WBAES})
INCLUDE(ExternalProject)
SET(WBAES_DST_DIR "wbaes")
SET(WBAES_INSTALL_ROOT "${THIRD_PARTY_PATH}/install")
SET(WBAES_INSTALL_DIR ${WBAES_INSTALL_ROOT}/${WBAES_DST_DIR})
SET(WBAES_ROOT ${WBAES_INSTALL_DIR})
SET(WBAES_INC_DIR ${WBAES_ROOT}/include)
SET(WBAES_LIB_DIR ${WBAES_ROOT}/lib)
SET(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_RPATH}" "${WBAES_ROOT}/lib")
SET(CMAKE_INSTALL_RPATH_USE_LINK_PATH TRUE)
IF(APPLE)
SET(WBAES_TAG "v1.0.0" CACHE STRING "" FORCE)
SET(WBAES_URL "http://paddlepaddledeps.bj.bcebos.com/wbaes-sdk.mac.${WBAES_TAG}.tgz" CACHE STRING "" FORCE)
SET(WBAES_LIB ${WBAES_LIB_DIR}/libwbaes.dylib)
SET(WBAES_SHARED_LIB ${WBAES_LIB_DIR}/libwbaes.dylib)
ELSEIF(WIN32)
SET(WBAES_TAG "v1.0.0" CACHE STRING "" FORCE)
SET(WBAES_URL "http://paddlepaddledeps.bj.bcebos.com/wbaes-sdk.windows-x64.${WBAES_TAG}.tgz" CACHE STRING "" FORCE)
SET(WBAES_LIB ${WBAES_LIB_DIR}/libwbaes.lib)
SET(WBAES_SHARED_LIB ${WBAES_LIB_DIR}/libwbaes.dll)
ELSE()
SET(WBAES_TAG "v1.0.2" CACHE STRING "" FORCE)
SET(WBAES_URL "http://paddlepaddledeps.bj.bcebos.com/wbaes-sdk.linux-x86_64.${WBAES_TAG}.tgz" CACHE STRING "" FORCE)
SET(WBAES_LIB ${WBAES_LIB_DIR}/libwbaes.so)
SET(WBAES_SHARED_LIB ${WBAES_LIB_DIR}/libwbaes.so)
ENDIF()
SET(WBAES_PROJECT "extern_wbaes")
MESSAGE(STATUS "WBAES_URL: ${WBAES_URL}, WBAES_LIB: ${WBAES_LIB}")
SET(WBAES_SOURCE_DIR "${THIRD_PARTY_PATH}/wbaes")
SET(WBAES_DOWNLOAD_DIR "${WBAES_SOURCE_DIR}/src/${WBAES_PROJECT}")
ExternalProject_Add(
${WBAES_PROJECT}
${EXTERNAL_PROJECT_LOG_ARGS}
PREFIX ${WBAES_SOURCE_DIR}
URL ${WBAES_URL}
DOWNLOAD_DIR ${WBAES_DOWNLOAD_DIR}
DOWNLOAD_NO_PROGRESS 1
CONFIGURE_COMMAND ""
BUILD_COMMAND ""
INSTALL_COMMAND ""
${CMAKE_COMMAND} -E copy_directory ${WBAES_DOWNLOAD_DIR}/include ${WBAES_INC_DIR} &&
${CMAKE_COMMAND} -E copy_directory ${WBAES_DOWNLOAD_DIR}/lib ${WBAES_LIB_DIR}
)
INCLUDE_DIRECTORIES(${WBAES_INC_DIR})
ADD_LIBRARY(wbaes SHARED IMPORTED GLOBAL)
SET_PROPERTY(TARGET wbaes PROPERTY IMPORTED_LOCATION ${WBAES_LIB})
SET_PROPERTY(TARGET wbaes PROPERTY IMPORTED_NO_SONAME 1)
ADD_DEPENDENCIES(wbaes ${WBAES_PROJECT})

@ -264,14 +264,6 @@ function(cc_library TARGET_NAME)
list(REMOVE_ITEM cc_library_DEPS warpctc)
add_dependencies(${TARGET_NAME} warpctc)
endif()
# Only deps libwbaes.so, not link
if("${cc_library_DEPS};" MATCHES "wbaes;")
list(REMOVE_ITEM cc_library_DEPS wbaes)
if(NOT "${TARGET_NAME}" MATCHES "dynload_wbaes")
list(APPEND cc_library_DEPS dynload_wbaes)
endif()
add_dependencies(${TARGET_NAME} wbaes)
endif()
# Only deps libmklml.so, not link
if("${cc_library_DEPS};" MATCHES "mklml;")
list(REMOVE_ITEM cc_library_DEPS mklml)

@ -170,14 +170,6 @@ copy(snappystream_lib
DSTS ${dst_dir} ${dst_dir}/lib
DEPS snappystream)
if (WITH_WBAES)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/wbaes")
copy(wbaes_lib
SRCS ${WBAES_INC_DIR} ${WBAES_LIB}
DSTS ${dst_dir} ${dst_dir}/lib
DEPS wbaes)
endif ()
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/zlib")
copy(zlib_lib
SRCS ${ZLIB_INCLUDE_DIR} ${ZLIB_LIBRARIES}

@ -236,6 +236,7 @@ paddle.fluid.layers.huber_loss (ArgSpec(args=['input', 'label', 'delta'], vararg
paddle.fluid.layers.kldiv_loss (ArgSpec(args=['x', 'target', 'reduction', 'name'], varargs=None, keywords=None, defaults=('mean', None)), ('document', '776d536cac47c89073abc7ee524d5aec'))
paddle.fluid.layers.tree_conv (ArgSpec(args=['nodes_vector', 'edge_set', 'output_size', 'num_filters', 'max_depth', 'act', 'param_attr', 'bias_attr', 'name'], varargs=None, keywords=None, defaults=(1, 2, 'tanh', None, None, None)), ('document', '34ea12ac9f10a65dccbc50100d12e607'))
paddle.fluid.layers.npair_loss (ArgSpec(args=['anchor', 'positive', 'labels', 'l2_reg'], varargs=None, keywords=None, defaults=(0.002,)), ('document', '46994d10276dd4cb803b4062b5d14329'))
paddle.fluid.layers.pixel_shuffle (ArgSpec(args=['x', 'upscale_factor'], varargs=None, keywords=None, defaults=None), ('document', 'ad669cdf83e72a69ebc5ed79e36486de'))
paddle.fluid.layers.fsp_matrix (ArgSpec(args=['x', 'y'], varargs=None, keywords=None, defaults=None), ('document', 'b76ccca3735bea4a58a0dbf0d77c5393'))
paddle.fluid.layers.data (ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True)), ('document', '33bbd42027d872b3818b3d64ec52e139'))
paddle.fluid.layers.open_files (ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None)), ('document', 'b1ae2e1cc0750e58726374061ea90ecc'))

@ -242,6 +242,11 @@ void InMemoryDataFeed<T>::SetTrainerNum(int trainer_num) {
trainer_num_ = trainer_num;
}
template <typename T>
void InMemoryDataFeed<T>::SetFleetSendBatchSize(int64_t size) {
fleet_send_batch_size_ = size;
}
template <typename T>
void InMemoryDataFeed<T>::PutInsToChannel(const std::string& ins_str) {
#ifdef _LINUX
@ -361,8 +366,13 @@ void InMemoryDataFeed<T>::GlobalShuffle() {
VLOG(3) << "GlobalShuffle() begin, thread_id=" << thread_id_;
auto fleet_ptr = FleetWrapper::GetInstance();
std::vector<std::vector<T*>> send_vec(trainer_num_);
std::vector<int> send_index(trainer_num_);
uint64_t reserve_len = fleet_send_batch_size_ / trainer_num_;
for (auto& vec : send_vec) {
vec.reserve(fleet_send_batch_size_);
vec.reserve(reserve_len);
}
for (int i = 0; i < trainer_num_; ++i) {
send_index[i] = i;
}
std::vector<std::future<int32_t>> total_status;
auto interval = GetMemoryDataInterval();
@ -375,7 +385,10 @@ void InMemoryDataFeed<T>::GlobalShuffle() {
int64_t node_id = random_num % trainer_num_;
send_vec[node_id].push_back(&((*memory_data_)[i]));
if (i % fleet_send_batch_size_ == 0 && i != 0) {
for (int j = 0; j < send_vec.size(); ++j) {
// shuffle the sequence of sending to avoid network timeout error
std::random_shuffle(send_index.begin(), send_index.end());
for (int index = 0; index < send_index.size(); ++index) {
int j = send_index[index];
std::string send_str;
SerializeIns(send_vec[j], &send_str);
VLOG(3) << "send str_length=" << send_str.length()
@ -388,7 +401,10 @@ void InMemoryDataFeed<T>::GlobalShuffle() {
}
}
}
for (int j = 0; j < send_vec.size(); ++j) {
// shuffle the sequence of sending to avoid network timeout error
std::random_shuffle(send_index.begin(), send_index.end());
for (int index = 0; index < send_index.size(); ++index) {
int j = send_index[index];
if (send_vec[j].size() != 0) {
std::string send_str;
SerializeIns(send_vec[j], &send_str);

@ -94,6 +94,8 @@ class DataFeed {
virtual void SetThreadNum(int thread_num) {}
// This function will do nothing at default
virtual void SetTrainerNum(int trainer_num) {}
// This function will do nothing at default
virtual void SetFleetSendBatchSize(int64_t size) {}
virtual void SetFileListMutex(std::mutex* mutex) {
mutex_for_pick_file_ = mutex;
}
@ -212,6 +214,7 @@ class InMemoryDataFeed : public PrivateQueueDataFeed<T> {
virtual void SetThreadId(int thread_id);
virtual void SetThreadNum(int thread_num);
virtual void SetTrainerNum(int trainer_num);
virtual void SetFleetSendBatchSize(int64_t size);
virtual void PutInsToChannel(const std::string& ins_str);
virtual void FillMemoryDataToChannel();
virtual void FillChannelToMemoryData();

@ -64,6 +64,17 @@ void DatasetImpl<T>::SetTrainerNum(int trainer_num) {
}
}
// if you run distributed, and want to do global shuffle,
// set this before global shuffle.
// be sure you call CreateReaders before SetFleetSendBatchSize
template <typename T>
void DatasetImpl<T>::SetFleetSendBatchSize(int64_t size) {
fleet_send_batch_size_ = size;
for (auto reader : readers_) {
reader->SetFleetSendBatchSize(size);
}
}
template <typename T>
void DatasetImpl<T>::SetHdfsConfig(const std::string& fs_name,
const std::string& fs_ugi) {

@ -47,6 +47,8 @@ class Dataset {
virtual void SetThreadNum(int thread_num) = 0;
// set workers' num
virtual void SetTrainerNum(int trainer_num) = 0;
// set fleet send batch size
virtual void SetFleetSendBatchSize(int64_t size) = 0;
// set fs name and ugi
virtual void SetHdfsConfig(const std::string& fs_name,
const std::string& fs_ugi) = 0;
@ -59,6 +61,8 @@ class Dataset {
virtual int GetThreadNum() = 0;
// get worker num
virtual int GetTrainerNum() = 0;
// get fleet send batch size
virtual int64_t GetFleetSendBatchSize() = 0;
// get hdfs config
virtual std::pair<std::string, std::string> GetHdfsConfig() = 0;
// get data fedd desc
@ -98,6 +102,7 @@ class DatasetImpl : public Dataset {
virtual void SetFileList(const std::vector<std::string>& filelist);
virtual void SetThreadNum(int thread_num);
virtual void SetTrainerNum(int trainer_num);
virtual void SetFleetSendBatchSize(int64_t size);
virtual void SetHdfsConfig(const std::string& fs_name,
const std::string& fs_ugi);
virtual void SetDataFeedDesc(const std::string& data_feed_desc_str);
@ -105,6 +110,7 @@ class DatasetImpl : public Dataset {
virtual const std::vector<std::string>& GetFileList() { return filelist_; }
virtual int GetThreadNum() { return thread_num_; }
virtual int GetTrainerNum() { return trainer_num_; }
virtual int64_t GetFleetSendBatchSize() { return fleet_send_batch_size_; }
virtual std::pair<std::string, std::string> GetHdfsConfig() {
return std::make_pair(fs_name_, fs_ugi_);
}
@ -137,6 +143,7 @@ class DatasetImpl : public Dataset {
std::string fs_name_;
std::string fs_ugi_;
unsigned int rand_seed;
int64_t fleet_send_batch_size_;
};
// use std::vector<MultiSlotType> as data type

@ -53,6 +53,10 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
this->SetDeviceContext(p, nccl_ctxs_->DevCtx(p));
}
}
// TODO(gongwb) :polish them!
if (is_encoded) {
VLOG(1) << "Use dgc allreduce mode";
}
}
#else
AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
@ -86,7 +90,7 @@ void AllReduceOpHandle::RunImplEncoded() {
paddle::framework::GradOriginalVarName(in_var_handles[i]->name());
auto encode_var_name = original_name + g_dgc_encoded;
auto *in_var = local_scope->FindVar(encode_var_name);
PADDLE_ENFORCE_NOT_NULL(in_var);
PADDLE_ENFORCE_NOT_NULL(in_var, "%s should not be null", encode_var_name);
auto &in = in_var->Get<LoDTensor>();
ins.emplace_back(&in);

@ -0,0 +1,79 @@
// Copyright (c) 2019 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.
#pragma once
#include <algorithm>
#include <string>
#include <unordered_map>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/details/build_strategy.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
namespace details {
void SetFuseParameterGroupsSize(int group_size);
int GetFuseParameterGroupsSize();
void SetFuseParameterMemorySize(uint64_t memory_size);
uint64_t GetFuseParameterMemorySize();
class AllocContinuousSpaceForGradPass : public ir::Pass {
protected:
void ApplyImpl(ir::Graph *graph) const override;
template <typename AttrType>
void ResetAttribute(const std::string &attr_name, ir::Graph *graph) const;
void SetGroupGradsAndParams(
const std::unordered_map<std::string, ir::Node *> &var_nodes,
const ParamsAndGrads &params_grads,
GroupGradsAndParams *group_grads_params) const;
void SetGroupAccordingToLayers(
const std::unordered_map<std::string, ir::Node *> &var_nodes,
const ParamsAndGrads &params_grads,
GroupGradsAndParams *group_grads_params) const;
void SetGroupAccordingToMemorySize(
const std::unordered_map<std::string, ir::Node *> &var_nodes,
GroupGradsAndParams *group_grads_params) const;
void SetGroupAccordingToGroupSize(
const std::unordered_map<std::string, ir::Node *> &var_nodes,
GroupGradsAndParams *group_grads_params) const;
private:
bool IsSupportedVarType(const proto::VarType::Type &type) const;
void RecordParamsAndGrads(ir::Node *node, ParamsAndGrads *params_grads) const;
void InitFusedVarsAndAllocSpaceForVars(
const std::vector<platform::Place> &places,
const std::vector<Scope *> &local_scopes,
const std::unordered_map<std::string, ir::Node *> &vars,
const std::string &fused_var_name,
const ParamsAndGrads &params_grads) const;
void AppendAllocSpaceForVarsOp(const std::vector<std::string> &params_name,
const std::vector<std::string> &grads_name,
const std::string &fused_var_name,
BlockDesc *global_block) const;
};
} // namespace details
} // namespace framework
} // namespace paddle

@ -142,6 +142,14 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
AppendPass("memory_optimize_pass");
}
// runtime_context_cache pass should be the last pass to enable the attr of
// all original and fused operators. But no operators can be enabled this
// attr if putting it after MultiDevPass.
if (strategy_.cache_runtime_context_) {
VLOG(10) << "Add runtime_context_cache_pass";
AppendPass("runtime_context_cache_pass");
}
AppendMultiDevPass(strategy_);
if (strategy_.fuse_all_reduce_ops_) {
@ -243,7 +251,7 @@ ir::Graph *BuildStrategy::Apply(ir::Graph *graph,
CreatePassesFromStrategy(false);
for (std::shared_ptr<ir::Pass> &pass : pass_builder_->AllPasses()) {
VLOG(3) << "apply " << pass->Type();
VLOG(3) << "BuildStrategy::Apply pass:" << pass->Type();
if (IsMultiDevPass(pass->Type())) {
pass->Erase(kPlaces);
pass->SetNotOwned<const std::vector<platform::Place>>(kPlaces, &places);
@ -328,3 +336,4 @@ USE_PASS(graph_to_program_pass);
USE_PASS(fuse_adam_op_pass);
USE_PASS(fuse_sgd_op_pass);
USE_PASS(fuse_all_reduce_op_pass);
USE_PASS(runtime_context_cache_pass);

@ -107,6 +107,8 @@ struct BuildStrategy {
std::vector<std::string> trainers_endpoints_;
bool remove_unnecessary_lock_{true};
bool cache_runtime_context_{false};
// NOTE:
// Before you add new options, think if it's a general strategy that works
// with other strategy. If not, the strategy should be created through

@ -237,6 +237,7 @@ void FleetWrapper::PushDenseParamSync(
std::vector<paddle::ps::Region> regions;
for (auto& t : var_names) {
Variable* var = scope.FindVar(t);
CHECK(var != nullptr) << "var[" << t << "] not found";
LoDTensor* tensor = var->GetMutable<LoDTensor>();
float* g = tensor->mutable_data<float>(place);
paddle::ps::Region reg(g, tensor->numel());

@ -126,7 +126,7 @@ static int shell_popen_fork_internal(const char* real_cmd, bool do_read,
}
close_open_fds_internal();
if (execl("/bin/sh", "sh", "-c", real_cmd, NULL) < 0) {
if (execl("/bin/bash", "bash", "-c", real_cmd, NULL) < 0) {
return -1;
}
exit(127);

@ -68,6 +68,7 @@ pass_library(transpose_flatten_concat_fuse_pass inference)
pass_library(identity_scale_op_clean_pass base)
pass_library(sync_batch_norm_pass base)
pass_library(runtime_context_cache_pass base)
pass_library(expected_kernel_cache_pass base)
pass_library(quant_conv2d_dequant_fuse_pass inference)
pass_library(fillconstant_elementwisemul_fuse inference)

@ -0,0 +1,37 @@
/* Copyright (c) 2019 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 "paddle/fluid/framework/ir/expected_kernel_cache_pass.h"
#include <memory>
#include "paddle/fluid/framework/operator.h"
namespace paddle {
namespace framework {
namespace ir {
void ExpectedKernelCachePass::ApplyImpl(ir::Graph* graph) const {
VLOG(3) << "Applies Expected Kernel Cache strategy.";
for (const Node* n : graph->Nodes()) {
if (n->IsOp()) {
n->Op()->SetAttr(kEnableCacheExpectedKernel, true);
}
}
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(expected_kernel_cache_pass,
paddle::framework::ir::ExpectedKernelCachePass);

@ -12,23 +12,20 @@ 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. */
#ifdef PADDLE_WITH_WBAES
#pragma once
#include "paddle/fluid/platform/dynload/wbaes.h"
#include <memory>
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace platform {
namespace dynload {
namespace framework {
namespace ir {
std::once_flag wbaes_dso_flag;
void *wbaes_dso_handle = nullptr;
class ExpectedKernelCachePass : public Pass {
protected:
void ApplyImpl(ir::Graph* graph) const override;
};
#define DEFINE_WRAP(__name) DynLoad__##__name __name
WBAES_ROUTINE_EACH(DEFINE_WRAP);
} // namespace dynload
} // namespace platform
} // namespace ir
} // namespace framework
} // namespace paddle
#endif

@ -23,7 +23,7 @@ namespace ir {
void RuntimeContextCachePass::ApplyImpl(ir::Graph* graph) const {
VLOG(3) << "Applies Runtime Context Cache strategy.";
for (const Node* n : graph->Nodes()) {
if (n->IsOp()) {
if (n->IsOp() && n->Op()) {
n->Op()->SetAttr(kEnableCacheRuntimeContext, true);
}
}

@ -899,50 +899,23 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
auto* dev_ctx = pool.Get(place);
// check if op[type] has kernel registered.
auto& all_op_kernels = AllOpKernels();
auto kernels_iter = all_op_kernels.find(type_);
if (kernels_iter == all_op_kernels.end()) {
PADDLE_THROW(
"There are no kernels which are registered in the %s operator.", type_);
if (!HasAttr(kEnableCacheExpectedKernel) || !kernel_type_) {
ChooseKernel(*runtime_ctx, scope, place);
}
OpKernelMap& kernels = kernels_iter->second;
auto expected_kernel_key = this->GetExpectedKernelType(
ExecutionContext(*this, scope, *dev_ctx, *runtime_ctx, nullptr));
VLOG(3) << "expected_kernel_key:" << expected_kernel_key;
auto kernel_iter = kernels.find(expected_kernel_key);
#ifdef PADDLE_WITH_MKLDNN
// workaround for missing MKLDNN kernel when FLAGS_use_mkldnn env var is set
if (kernel_iter == kernels.end() &&
expected_kernel_key.library_type_ == LibraryType::kMKLDNN) {
VLOG(3) << "missing MKLDNN kernel: fallbacking to PLAIN one";
expected_kernel_key.library_type_ = LibraryType::kPlain;
expected_kernel_key.data_layout_ = DataLayout::kAnyLayout;
kernel_iter = kernels.find(expected_kernel_key);
}
#endif
if (kernel_iter == kernels.end()) {
PADDLE_THROW("op %s does not have kernel for %s", type_,
KernelTypeToString(expected_kernel_key));
}
std::vector<KernelConfig>* kernel_configs =
GetKernelConfig(expected_kernel_key);
std::vector<KernelConfig>* kernel_configs = GetKernelConfig(*kernel_type_);
// do data transformScope &transfer_scope;
std::vector<std::string> transfered_inplace_vars;
auto* transfer_scope = PrepareData(scope, expected_kernel_key,
&transfered_inplace_vars, runtime_ctx);
auto* transfer_scope =
PrepareData(scope, *kernel_type_, &transfered_inplace_vars, runtime_ctx);
// exec scope is the scope that kernel actually executed on.
const Scope& exec_scope =
(transfer_scope == nullptr ? scope : *transfer_scope);
if (!(expected_kernel_key.place_ == dev_ctx->GetPlace())) {
dev_ctx = pool.Get(expected_kernel_key.place_);
if (!(kernel_type_->place_ == dev_ctx->GetPlace())) {
dev_ctx = pool.Get(kernel_type_->place_);
}
if (!HasAttr(kAllKernelsMustComputeRuntimeShape)) {
@ -951,8 +924,8 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
}
// TODO(panyx0718): ExecutionContext should only depend on RuntimeContext
// not Scope. Imperative mode only pass inputs and get outputs.
kernel_iter->second(ExecutionContext(*this, exec_scope, *dev_ctx,
*runtime_ctx, kernel_configs));
(*kernel_func_)(ExecutionContext(*this, exec_scope, *dev_ctx, *runtime_ctx,
kernel_configs));
if (!transfered_inplace_vars.empty()) {
// there is inplace variable has been transfered.
@ -978,6 +951,46 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
}
}
void OperatorWithKernel::ChooseKernel(const RuntimeContext& ctx,
const Scope& scope,
const platform::Place& place) const {
platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
auto* dev_ctx = pool.Get(place);
// check if op[type] has kernel registered.
auto& all_op_kernels = AllOpKernels();
auto kernels_iter = all_op_kernels.find(type_);
if (kernels_iter == all_op_kernels.end()) {
PADDLE_THROW(
"There are no kernels which are registered in the %s operator.", type_);
}
OpKernelMap& kernels = kernels_iter->second;
auto expected_kernel_key = this->GetExpectedKernelType(
ExecutionContext(*this, scope, *dev_ctx, ctx, nullptr));
VLOG(3) << "expected_kernel_key:" << expected_kernel_key;
auto kernel_iter = kernels.find(expected_kernel_key);
#ifdef PADDLE_WITH_MKLDNN
// workaround for missing MKLDNN kernel when FLAGS_use_mkldnn env var is set
if (kernel_iter == kernels.end() &&
expected_kernel_key.library_type_ == LibraryType::kMKLDNN) {
VLOG(3) << "missing MKLDNN kernel: fallbacking to PLAIN one";
expected_kernel_key.library_type_ = LibraryType::kPlain;
expected_kernel_key.data_layout_ = DataLayout::kAnyLayout;
kernel_iter = kernels.find(expected_kernel_key);
}
#endif
if (kernel_iter == kernels.end()) {
PADDLE_THROW("op %s does not have kernel for %s", type_,
KernelTypeToString(expected_kernel_key));
}
kernel_type_.reset(new OpKernelType(expected_kernel_key));
kernel_func_.reset(new OpKernelFunc(kernel_iter->second));
}
void OperatorWithKernel::TransferInplaceVarsBack(
const Scope& scope, const std::vector<std::string>& inplace_vars,
const Scope& transfer_scope) const {

@ -70,6 +70,12 @@ constexpr char kNewGradSuffix[] = "@NEWGRAD@";
/// this Op's execution to save the elapsed time.
constexpr char kEnableCacheRuntimeContext[] = "@ENABLE_CACHE_RUNTIME_CONTEXT@";
/// If an Op has attribtue kEnableCacheExpectedKernel, it means that in a same
/// name scope and same place, since the expected kerenl of this Op does not
/// change in the execution, it could be recorded only at the first iteration of
/// this Op's execution to save the elapsed time.
constexpr char kEnableCacheExpectedKernel[] = "@ENABLE_CACHE_EXPECTED_KERNEL@";
/// If an Op has this attribute, all its kernels should calculate output
/// variable's shape in the corresponding Compute() function. And
/// OperatorWithKernel::RunImpl() would skip call this Op's InferShape()
@ -491,8 +497,13 @@ class OperatorWithKernel : public OperatorBase {
const std::vector<std::string>& inplace_vars,
const Scope& exec_scope) const;
void ChooseKernel(const RuntimeContext& ctx, const Scope& scope,
const platform::Place& place) const;
protected:
mutable OpKernelConfigsMap kernel_configs_map_;
mutable std::unique_ptr<OpKernelType> kernel_type_;
mutable std::unique_ptr<OpKernelFunc> kernel_func_;
mutable std::unique_ptr<RuntimeContext> runtime_ctx_;
mutable const Scope* pre_scope_ = nullptr;
};

@ -3,4 +3,7 @@ cc_library(layer SRCS layer.cc DEPS proto_desc operator device_context blas pybi
cc_library(tracer SRCS tracer.cc DEPS proto_desc device_context pybind)
cc_library(engine SRCS engine.cc)
cc_library(imperative_profiler SRCS profiler.cc)
cc_library(nccl_context SRCS nccl_context.cc DEPS device_context)
cc_test(nccl_context_test SRCS nccl_context_test.cc DEPS nccl_context)
endif()

@ -0,0 +1,133 @@
// Copyright (c) 2019 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 "paddle/fluid/imperative/nccl_context.h"
namespace paddle {
namespace imperative {
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
void NCCLParallelContext::RecvNCCLID(const std::string &ep,
ncclUniqueId *nccl_id) {
auto addr = paddle::string::Split(ep, ':');
PADDLE_ENFORCE_EQ(addr.size(), 2UL,
"The endpoint should contain host and port: %s", ep);
std::string host = addr[0];
int port = std::stoi(addr[1]);
int server_fd, new_socket;
struct sockaddr_in address;
int addrlen = sizeof(address);
char buffer[1024] = {0};
int opt = 0;
// creating socket fd
if ((server_fd = socket(AF_INET, SOCK_STREAM, 0)) == 0)
PADDLE_THROW("create server fd failed");
if (setsockopt(server_fd, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof(opt)))
PADDLE_THROW("set socket opt failed");
address.sin_family = AF_INET;
address.sin_addr.s_addr = INADDR_ANY;
address.sin_port = htons(port);
if (bind(server_fd, (struct sockaddr *)&address, sizeof(address)) < 0)
PADDLE_THROW("binding failed on ep: %s", ep);
VLOG(3) << "listening on: " << ep;
if (listen(server_fd, 3) < 0) PADDLE_THROW("listen on server fd failed");
if ((new_socket =
accept(server_fd, reinterpret_cast<struct sockaddr *>(&address),
reinterpret_cast<socklen_t *>(&addrlen))) < 0)
PADDLE_THROW("accept the new socket fd failed");
if (read(new_socket, buffer, 1024) < 0)
PADDLE_THROW("reading the ncclUniqueId from socket failed");
VLOG(3) << "recevived the ncclUniqueId";
memcpy(nccl_id, buffer, NCCL_UNIQUE_ID_BYTES);
VLOG(3) << "closing the socket server: " << ep;
close(server_fd);
}
void NCCLParallelContext::SendNCCLID(const std::string &ep,
ncclUniqueId *nccl_id) {
auto addr = paddle::string::Split(ep, ':');
PADDLE_ENFORCE_EQ(addr.size(), 2UL,
"The endpoint should contain host and port: %s", ep);
std::string host = addr[0];
int port = std::stoi(addr[1]);
// struct sockaddr_in address;
int sock = 0;
struct sockaddr_in serv_addr;
char buffer[1024] = {0};
memcpy(buffer, nccl_id, NCCL_UNIQUE_ID_BYTES);
if ((sock = socket(AF_INET, SOCK_STREAM, 0)) < 0)
PADDLE_THROW("create socket failed");
memset(&serv_addr, '0', sizeof(serv_addr));
serv_addr.sin_family = AF_INET;
serv_addr.sin_port = htons(port);
if (inet_pton(AF_INET, host.c_str(), &serv_addr.sin_addr) <= 0)
PADDLE_THROW("invalied address: %s", ep);
while (true) {
if (connect(sock, (struct sockaddr *)&serv_addr, sizeof(serv_addr)) < 0) {
VLOG(0) << "worker: " << ep
<< " is not ready, will retry after 3 seconds...";
std::this_thread::sleep_for(std::chrono::seconds(3));
continue;
}
VLOG(3) << "sending the ncclUniqueId to " << ep;
send(sock, buffer, NCCL_UNIQUE_ID_BYTES, 0);
break;
}
}
void NCCLParallelContext::BcastNCCLId(ncclUniqueId *nccl_id, int root) {
if (strategy_.local_rank_ == root) {
for (auto ep : strategy_.trainer_endpoints_) {
if (ep != strategy_.current_endpoint_) SendNCCLID(ep, nccl_id);
}
} else {
RecvNCCLID(strategy_.current_endpoint_, nccl_id);
}
}
void NCCLParallelContext::Init() {
ncclUniqueId nccl_id;
ncclComm_t comm;
if (strategy_.local_rank_ == 0) {
// generate the unique ncclid on the root worker
platform::dynload::ncclGetUniqueId(&nccl_id);
BcastNCCLId(&nccl_id, 0);
} else {
BcastNCCLId(&nccl_id, 0);
}
int gpu_id = boost::get<platform::CUDAPlace>(place_).device;
VLOG(0) << "init nccl context nranks: " << strategy_.nranks_
<< " local rank: " << strategy_.local_rank_ << " gpu id: " << gpu_id;
PADDLE_ENFORCE(cudaSetDevice(gpu_id));
PADDLE_ENFORCE(platform::dynload::ncclCommInitRank(
&comm, strategy_.nranks_, nccl_id, strategy_.local_rank_));
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
auto *dev_ctx = static_cast<platform::CUDADeviceContext *>(pool.Get(place_));
dev_ctx->set_nccl_comm(comm);
}
#endif
} // namespace imperative
} // namespace paddle

Some files were not shown because too many files have changed in this diff Show More

Loading…
Cancel
Save