m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into revert_vlog

Browse Source 
test=develop
local_add_cudnn_lstm
minqiyang 6 years ago
parent 53433d7f2e 05b7ee7eeb
commit be04d99fe4
56 changed files with 157 additions and 139 deletions
Show Stats Download Patch File Download Diff File

5
cmake/cuda.cmake
Unescape View File

@ -199,8 +199,11 @@ elseif(CMAKE_BUILD_TYPE STREQUAL "MinSizeRel")
list(APPEND CUDA_NVCC_FLAGS ${CMAKE_CXX_FLAGS_RELEASE})
endif()
else(NOT WIN32)
list(APPEND CUDA_NVCC_FLAGS "--compiler-options;/bigobj")
if(CMAKE_BUILD_TYPE STREQUAL "Debug")
list(APPEND CUDA_NVCC_FLAGS "-g -G")
list(APPEND CUDA_NVCC_FLAGS "-g -G")
# match the cl's _ITERATOR_DEBUG_LEVEL
list(APPEND CUDA_NVCC_FLAGS "-D_DEBUG")
elseif(CMAKE_BUILD_TYPE STREQUAL "Release")
list(APPEND CUDA_NVCC_FLAGS "-O3 -DNDEBUG")
else()

2
cmake/external/pybind11.cmake vendored
Unescape View File

@ -26,7 +26,7 @@ ExternalProject_Add(
extern_pybind
${EXTERNAL_PROJECT_LOG_ARGS}
GIT_REPOSITORY "https://github.com/pybind/pybind11.git"
GIT_TAG "v2.1.1"
GIT_TAG "v2.2.4"
PREFIX ${PYBIND_SOURCE_DIR}
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""

11
cmake/generic.cmake
Unescape View File

@ -349,10 +349,17 @@ function(cc_test TARGET_NAME)
set(oneValueArgs "")
set(multiValueArgs SRCS DEPS ARGS)
cmake_parse_arguments(cc_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
if(WIN32)
list(APPEND win32_deps shlwapi)
if("${cc_test_DEPS};" MATCHES "python;")
list(REMOVE_ITEM cc_test_DEPS python)
list(APPEND win32_deps ${PYTHON_LIBRARIES})
endif()
endif(WIN32)
add_executable(${TARGET_NAME} ${cc_test_SRCS})
target_link_libraries(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
if(WIN32)
target_link_libraries(${TARGET_NAME} shlwapi)
target_link_libraries(${TARGET_NAME} ${win32_deps})
endif(WIN32)
add_dependencies(${TARGET_NAME} ${cc_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
add_test(NAME ${TARGET_NAME}
@ -683,7 +690,7 @@ function(py_test TARGET_NAME)
set(multiValueArgs SRCS DEPS ARGS ENVS)
cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
add_test(NAME ${TARGET_NAME}
COMMAND env FLAGS_init_allocated_mem=true FLAGS_cudnn_deterministic=true
COMMAND ${CMAKE_COMMAND} -E env FLAGS_init_allocated_mem=true FLAGS_cudnn_deterministic=true
FLAGS_cpu_deterministic=true
PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}

8
paddle/fluid/API.spec
Unescape View File

@ -26,10 +26,10 @@ paddle.fluid.release_memory ArgSpec(args=['input_program', 'skip_opt_set'], vara
paddle.fluid.DistributeTranspilerConfig.__init__
paddle.fluid.ParallelExecutor.__init__ ArgSpec(args=['self', 'use_cuda', 'loss_name', 'main_program', 'share_vars_from', 'exec_strategy', 'build_strategy', 'num_trainers', 'trainer_id', 'scope'], varargs=None, keywords=None, defaults=(None, None, None, None, None, 1, 0, None))
paddle.fluid.ParallelExecutor.run ArgSpec(args=['self', 'fetch_list', 'feed', 'feed_dict', 'return_numpy'], varargs=None, keywords=None, defaults=(None, None, True))
paddle.fluid.ExecutionStrategy.__init__ __init__(self: paddle.fluid.core.ExecutionStrategy) -> None
paddle.fluid.BuildStrategy.GradientScaleStrategy.__init__ __init__(self: paddle.fluid.core.GradientScaleStrategy, arg0: int) -> None
paddle.fluid.BuildStrategy.ReduceStrategy.__init__ __init__(self: paddle.fluid.core.ReduceStrategy, arg0: int) -> None
paddle.fluid.BuildStrategy.__init__ __init__(self: paddle.fluid.core.BuildStrategy) -> None
paddle.fluid.ExecutionStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.ExecutionStrategy) -> None
paddle.fluid.BuildStrategy.GradientScaleStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.BuildStrategy.GradientScaleStrategy, arg0: int) -> None
paddle.fluid.BuildStrategy.ReduceStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.BuildStrategy.ReduceStrategy, arg0: int) -> None
paddle.fluid.BuildStrategy.__init__ __init__(self: paddle.fluid.core.ParallelExecutor.BuildStrategy) -> None
paddle.fluid.create_lod_tensor ArgSpec(args=['data', 'recursive_seq_lens', 'place'], varargs=None, keywords=None, defaults=None)
paddle.fluid.create_random_int_lodtensor ArgSpec(args=['recursive_seq_lens', 'base_shape', 'place', 'low', 'high'], varargs=None, keywords=None, defaults=None)
paddle.fluid.io.save_vars ArgSpec(args=['executor', 'dirname', 'main_program', 'vars', 'predicate', 'filename'], varargs=None, keywords=None, defaults=(None, None, None, None))

5
paddle/fluid/framework/CMakeLists.txt
Unescape View File

@ -116,14 +116,9 @@ cc_test(op_proto_maker_test SRCS op_proto_maker_test.cc DEPS op_proto_maker)
cc_library(op_info SRCS op_info.cc DEPS attribute framework_proto)
cc_library(shape_inference SRCS shape_inference.cc DEPS ddim attribute device_context)
if (NOT WIN32)
cc_library(transfer_scope_cache SRCS transfer_scope_cache.cc DEPS scope framework_proto device_context)
cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
shape_inference data_transform lod_tensor profiler transfer_scope_cache)
else()
cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope glog
shape_inference data_transform lod_tensor)
endif(NOT WIN32)
cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry device_context)

4
paddle/fluid/framework/details/all_reduce_op_handle.cc
Unescape View File

@ -23,7 +23,7 @@ namespace paddle {
namespace framework {
namespace details {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
@ -74,7 +74,7 @@ void AllReduceOpHandle::RunImpl() {
}
if (platform::is_gpu_place(lod_tensors[0]->place())) {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
PADDLE_ENFORCE(nccl_ctxs_, "nccl_ctxs should not be nullptr.");
int dtype = -1;
size_t numel = 0;

6
paddle/fluid/framework/details/all_reduce_op_handle.h
Unescape View File

@ -20,7 +20,7 @@
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/scope.h"
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
#include "paddle/fluid/platform/nccl_helper.h"
#endif
@ -29,7 +29,7 @@ namespace framework {
namespace details {
struct AllReduceOpHandle : public OpHandleBase {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
AllReduceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
const platform::NCCLContextMap *ctxs);
@ -49,7 +49,7 @@ struct AllReduceOpHandle : public OpHandleBase {
private:
std::vector<Scope *> local_scopes_;
std::vector<platform::Place> places_;
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
const platform::NCCLContextMap *nccl_ctxs_;
#endif
};

2
paddle/fluid/framework/details/broadcast_op_handle.cc
Unescape View File

@ -82,7 +82,7 @@ void BroadcastOpHandle::BroadcastOneVar(
});
}
} else {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
VarHandle *out_handle = nullptr;
int root_id = boost::get<platform::CUDAPlace>(in_tensor.place()).device;
std::vector<std::function<void()>> broadcast_calls;

6
paddle/fluid/framework/details/broadcast_op_handle.h
Unescape View File

@ -24,7 +24,7 @@
#include "paddle/fluid/framework/selected_rows.h"
#include "paddle/fluid/platform/device_context.h"
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
#include "paddle/fluid/platform/nccl_helper.h"
#endif
@ -34,7 +34,7 @@ namespace details {
struct BroadcastOpHandle : public OpHandleBase {
public:
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
BroadcastOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
const platform::NCCLContextMap *nccl_ctxs)
@ -68,7 +68,7 @@ struct BroadcastOpHandle : public OpHandleBase {
std::vector<Scope *> local_scopes_;
std::vector<platform::Place> places_;
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
const platform::NCCLContextMap *nccl_ctxs_;
#endif

12
paddle/fluid/framework/details/broadcast_op_handle_test.h
Unescape View File

@ -42,7 +42,7 @@ struct TestBroadcastOpHandle {
std::vector<std::unique_ptr<ir::Node>> nodes_;
std::vector<p::Place> place_list_;
bool use_gpu_;
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
#endif
@ -50,7 +50,7 @@ struct TestBroadcastOpHandle {
for (size_t j = 0; j < ctxs_.size(); ++j) {
ctxs_[j]->Wait();
}
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
if (nccl_ctxs_) {
nccl_ctxs_->WaitAll();
}
@ -60,7 +60,7 @@ struct TestBroadcastOpHandle {
void InitCtxOnGpu(bool use_gpu) {
use_gpu_ = use_gpu;
if (use_gpu_) {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
int count = p::GetCUDADeviceCount();
if (count <= 1) {
LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA "
@ -84,7 +84,7 @@ struct TestBroadcastOpHandle {
place_list_.push_back(p);
ctxs_.emplace_back(new p::CPUDeviceContext(p));
}
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
nccl_ctxs_.reset(nullptr);
#endif
}
@ -106,14 +106,14 @@ struct TestBroadcastOpHandle {
nodes_.emplace_back(
ir::CreateNodeForTest("node0", ir::Node::Type::kOperation));
if (use_gpu_) {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
op_handle_ = new BroadcastOpHandle(nodes_.back().get(), local_scopes_,
place_list_, nccl_ctxs_.get());
#else
PADDLE_THROW("CUDA is not support.");
#endif
} else {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
op_handle_ = new BroadcastOpHandle(nodes_.back().get(), local_scopes_,
place_list_, nccl_ctxs_.get());
#else

4
paddle/fluid/framework/details/build_strategy.cc
Unescape View File

@ -96,7 +96,7 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
const std::string &loss_var_name,
const std::unordered_set<std::string> &param_names,
const std::vector<Scope *> &local_scopes,
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
const bool use_cuda, platform::NCCLContextMap *nccl_ctxs) const {
#else
const bool use_cuda) const {
@ -118,7 +118,7 @@ std::unique_ptr<ir::Graph> BuildStrategy::Apply(
pass->Erase("local_scopes");
pass->SetNotOwned<const std::vector<Scope *>>("local_scopes",
&local_scopes);
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
platform::NCCLContextMap *nctx = use_cuda ? nccl_ctxs : nullptr;
pass->Erase("nccl_ctxs");
pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx);

4
paddle/fluid/framework/details/build_strategy.h
Unescape View File

@ -23,7 +23,7 @@
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/enforce.h"
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
#include "paddle/fluid/platform/nccl_helper.h"
#endif
@ -98,7 +98,7 @@ struct BuildStrategy {
const std::string &loss_var_name,
const std::unordered_set<std::string> &param_names,
const std::vector<Scope *> &local_scopes,
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
const bool use_cuda, platform::NCCLContextMap *nccl_ctxs) const;
#else
const bool use_cuda) const;

2
paddle/fluid/framework/details/data_balance_op_handle.cc
Unescape View File

@ -20,7 +20,7 @@ namespace paddle {
namespace framework {
namespace details {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
DataBalanceOpHandle::DataBalanceOpHandle(
ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,

4
paddle/fluid/framework/details/data_balance_op_handle.h
Unescape View File

@ -19,7 +19,7 @@
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/scope.h"
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
#include "paddle/fluid/platform/nccl_helper.h"
#endif
@ -29,7 +29,7 @@ namespace details {
struct DataBalanceOpHandle : public OpHandleBase {
public:
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
DataBalanceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
const platform::NCCLContextMap *ctxs);

4
paddle/fluid/framework/details/fused_broadcast_op_handle.h
Unescape View File

@ -25,7 +25,7 @@
#include "paddle/fluid/framework/selected_rows.h"
#include "paddle/fluid/platform/device_context.h"
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
#include "paddle/fluid/platform/nccl_helper.h"
#endif
@ -35,7 +35,7 @@ namespace details {
struct FusedBroadcastOpHandle : public BroadcastOpHandle {
public:
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
FusedBroadcastOpHandle(ir::Node *node,
const std::vector<Scope *> local_scopes,
const std::vector<platform::Place> &places,

4
paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc
Unescape View File

@ -44,14 +44,14 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
nodes_.emplace_back(
ir::CreateNodeForTest("fused_broadcast", ir::Node::Type::kOperation));
if (use_gpu_) {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
op_handle_ = new FusedBroadcastOpHandle(
nodes_.back().get(), local_scopes_, place_list_, nccl_ctxs_.get());
#else
PADDLE_THROW("CUDA is not supported.");
#endif
} else {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
op_handle_ = new FusedBroadcastOpHandle(
nodes_.back().get(), local_scopes_, place_list_, nccl_ctxs_.get());
#else

16
paddle/fluid/framework/details/multi_devices_graph_pass.cc
Unescape View File

@ -142,7 +142,7 @@ void MultiDevSSAGraphBuilder::Init() const {
places_ = Get<const std::vector<platform::Place>>(kPlaces);
local_scopes_ = Get<const std::vector<Scope *>>(kLocalScopes);
strategy_ = Get<const BuildStrategy>(kStrategy);
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
nccl_ctxs_ = &Get<platform::NCCLContextMap>("nccl_ctxs");
#endif
@ -431,7 +431,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
}
}
bool use_gpu = false;
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
use_gpu = nccl_ctxs_ != nullptr;
#endif
@ -478,7 +478,7 @@ bool MultiDevSSAGraphBuilder::IsSparseGradient(const std::string &og) const {
void MultiDevSSAGraphBuilder::SetCommunicationContext(
OpHandleBase *op_handle, const platform::Place &p) const {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
if (nccl_ctxs_ == nullptr) {
op_handle->SetDeviceContext(p,
platform::DeviceContextPool::Instance().Get(p));
@ -492,7 +492,7 @@ void MultiDevSSAGraphBuilder::SetCommunicationContext(
void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result,
const std::string &p_name,
size_t src_dev_id) const {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
auto *op_handle = new BroadcastOpHandle(
result->CreateEmptyNode("broadcast", ir::Node::Type::kOperation),
local_scopes_, places_, nccl_ctxs_);
@ -522,7 +522,7 @@ void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result,
void MultiDevSSAGraphBuilder::CreateFusedBroadcastOp(
ir::Graph *result,
const std::vector<std::unordered_set<std::string>> &bcast_varnames) const {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
auto *op_handle = new FusedBroadcastOpHandle(
result->CreateEmptyNode("fused_broadcast", ir::Node::Type::kOperation),
local_scopes_, places_, nccl_ctxs_);
@ -568,7 +568,7 @@ void MultiDevSSAGraphBuilder::CreateComputationalOp(ir::Graph *result,
void MultiDevSSAGraphBuilder::InsertAllReduceOp(ir::Graph *result,
const std::string &og) const {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
local_scopes_, places_, nccl_ctxs_));
@ -597,7 +597,7 @@ void MultiDevSSAGraphBuilder::InsertAllReduceOp(ir::Graph *result,
void MultiDevSSAGraphBuilder::InsertDataBalanceOp(
ir::Graph *result, const std::vector<std::string> &datas) const {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
result->Get<GraphOps>(kGraphOps).emplace_back(new DataBalanceOpHandle(
result->CreateEmptyNode("data_balance", ir::Node::Type::kOperation),
local_scopes_, places_, nccl_ctxs_));
@ -694,7 +694,7 @@ void MultiDevSSAGraphBuilder::CreateComputationalOps(ir::Graph *result,
VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(ir::Graph *result,
const std::string &og,
int dst_dev_id) const {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
result->Get<GraphOps>(kGraphOps).emplace_back(new ReduceOpHandle(
result->CreateEmptyNode("reduce", ir::Node::Type::kOperation),
local_scopes_, places_, nccl_ctxs_));

2
paddle/fluid/framework/details/multi_devices_graph_pass.h
Unescape View File

@ -40,7 +40,7 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
size_t device_id) const;
void Init() const;
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
mutable platform::NCCLContextMap *nccl_ctxs_;
#endif

2
paddle/fluid/framework/details/reduce_op_handle.cc
Unescape View File

@ -125,7 +125,7 @@ void ReduceOpHandle::RunImpl() {
}
});
} else if (paddle::platform::is_gpu_place(lod_tensors[0]->place())) {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
auto pre_in = pre_in_var->Get<framework::LoDTensor>();
VariableVisitor::ShareDimsAndLoD(*pre_in_var, out_var);
VariableVisitor::GetMutableTensor(out_var).mutable_data(

4
paddle/fluid/framework/details/reduce_op_handle.h
Unescape View File

@ -23,7 +23,7 @@
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/selected_rows.h"
#include "paddle/fluid/platform/device_context.h"
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
#include "paddle/fluid/platform/nccl_helper.h"
#endif
@ -35,7 +35,7 @@ struct ReduceOpHandle : public OpHandleBase {
std::vector<Scope *> local_scopes_;
std::vector<platform::Place> places_;
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
const platform::NCCLContextMap *nccl_ctxs_;
ReduceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,

12
paddle/fluid/framework/details/reduce_op_handle_test.cc
Unescape View File

@ -35,7 +35,7 @@ struct TestReduceOpHandle {
std::vector<p::Place> gpu_list_;
std::vector<std::unique_ptr<p::DeviceContext>> ctxs_;
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
std::unique_ptr<platform::NCCLContextMap> nccl_ctxs_;
#endif
@ -43,7 +43,7 @@ struct TestReduceOpHandle {
for (size_t j = 0; j < ctxs_.size(); ++j) {
ctxs_[j]->Wait();
}
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
if (nccl_ctxs_) {
nccl_ctxs_->WaitAll();
}
@ -53,7 +53,7 @@ struct TestReduceOpHandle {
void InitCtxOnGpu(bool use_gpu) {
use_gpu_ = use_gpu;
if (use_gpu) {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
int count = p::GetCUDADeviceCount();
if (count <= 1) {
LOG(WARNING) << "Cannot test multi-gpu Broadcast, because the CUDA "
@ -77,7 +77,7 @@ struct TestReduceOpHandle {
gpu_list_.push_back(p);
ctxs_.emplace_back(new p::CPUDeviceContext(p));
}
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
nccl_ctxs_.reset(nullptr);
#endif
}
@ -99,14 +99,14 @@ struct TestReduceOpHandle {
nodes.emplace_back(new ir::Node("node"));
if (use_gpu_) {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
op_handle_.reset(new ReduceOpHandle(nodes.back().get(), local_scopes_,
gpu_list_, nccl_ctxs_.get()));
#else
PADDLE_THROW("CUDA is not support.");
#endif
} else {
#ifdef PADDLE_WITH_CUDA
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
op_handle_.reset(new ReduceOpHandle(nodes.back().get(), local_scopes_,
gpu_list_, nccl_ctxs_.get()));
#else

5
paddle/fluid/framework/ir/is_test_pass_tester.cc
Unescape View File

@ -15,7 +15,10 @@
#include "paddle/fluid/framework/ir/is_test_pass.h"
#include <gtest/gtest.h>
#ifdef _WIN32
#undef FALSE
#undef TRUE
#endif
namespace paddle {
namespace framework {
namespace ir {

17
paddle/fluid/framework/lod_tensor.cc
Unescape View File

@ -26,10 +26,8 @@ limitations under the License. */
#include "paddle/fluid/memory/memcpy.h"
#include "paddle/fluid/memory/memory.h"
#if !defined(_WIN32)
#include "paddle/fluid/recordio/scanner.h"
#include "paddle/fluid/recordio/writer.h"
#endif // _WIN32
namespace paddle {
namespace framework {
@ -305,7 +303,6 @@ void DeserializeFromStream(std::istream &is, LoDTensor *tensor,
TensorFromStream(is, static_cast<Tensor *>(tensor), dev_ctx);
}
#if !defined(_WIN32)
void WriteToRecordIO(recordio::Writer *writer,
const std::vector<LoDTensor> &tensor,
const platform::DeviceContext &dev_ctx) {
@ -335,19 +332,7 @@ bool ReadFromRecordIO(recordio::Scanner *scanner,
return true;
}
#else
class Writer {};
class Scanner {};
void WriteToRecordIO(recordio::Writer *writer,
const std::vector<LoDTensor> &tensor,
const platform::DeviceContext &dev_ctx) {}
bool ReadFromRecordIO(recordio::Scanner *scanner,
const platform::DeviceContext &dev_ctx,
std::vector<LoDTensor> *result_ptr) {
PADDLE_ENFORCE("windows didn't supported recordio!.");
return true;
}
#endif // _WIN32
std::vector<LoDTensor> LoDTensor::SplitLoDTensor(
const std::vector<platform::Place> places) const {
check_memory_size();

2
paddle/fluid/framework/lod_tensor_test.cc
Unescape View File

@ -274,7 +274,6 @@ TEST(LoD, ConvertToOffsetBasedLoD) {
EXPECT_EQ(offset_lod, expected);
}
#if !defined(_WIN32)
template <typename T>
static void TestRecordIO() {
LoDTensor tensor;
@ -321,7 +320,6 @@ TEST(LoDTensor, RecordIO) {
TestRecordIO<float>();
TestRecordIO<double>();
}
#endif // !defined(_WIN32)
} // namespace framework
} // namespace paddle

11
paddle/fluid/framework/operator.cc
Unescape View File

@ -149,17 +149,14 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
#endif
}
// The profile has a process-wide mutex, results in serious performance issue
// in concurrency scenerio. Here use an `if` to fix this issue.
// Please not remove the `if`, ask @Superjomn if there are any concern.
#ifndef _WIN32
// The profile has a process-wide mutex, results in serious performance issue
// in concurrency scenerio. Here use an `if` to fix this issue.
// Please not remove the `if`, ask @Superjomn if there are any concern.
if (platform::IsProfileEnabled()) {
platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
platform::RecordEvent record_event(Type(), pool.Get(place));
RunImpl(scope, place);
} else // NOLINT
#endif
{
} else {
RunImpl(scope, place);
}
VLOG(3) << place << " " << DebugStringEx(&scope);

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

Write Preview
Loading…
Cancel
Save