Skip to content

P
PaddleDetection

s920243400 / PaddleDetection
与 Fork 源项目一致

Fork自 PaddlePaddle / PaddleDetection

通知 2
Star 0
Fork 0
P
PaddleDetection
未验证 提交 cbdb8a17 编写于 作者: G gongweibao 提交者： GitHub
浏览文件
操作

Polish DGC code (#16818)

上级 dbf66dd0
隐藏空白更改
内联 并排
Showing with 369 addition and 253 deletion
CMakeLists.txt
浏览文件 @ cbdb8a17
......@@ -77,6 +77,7 @@ option(WITH_INFERENCE_API_TEST "Test fluid inference C++ high-level api interf
option(WITH_HIGH_LEVEL_API_TEST "Test fluid python high-level api interface" OFF)
option(PY_VERSION "Compile PaddlePaddle with python3 support" ${PY_VERSION})
option(WITH_FAST_MATH "Make use of fast math library, might affect the precision to some extent" ON)
option(WITH_DGC "Use DGC(Deep Gradient Compression) or not" ON)
# PY_VERSION
if(NOT PY_VERSION)
......@@ -196,9 +197,14 @@ if(WITH_GPU)
include(anakin_subgraph)
endif()
if(WITH_GPU AND NOT WIN32)
if(WIN32 OR APPLE OR NOT WITH_GPU OR ON_INFER)
set(WITH_DGC OFF)
endif()
if(WITH_DGC)
message(STATUS "add dgc lib.")
include(external/dgc)
add_definitions(-DPADDLE_WITH_DGC)
endif()
if(WITH_MKL OR WITH_MKLML)
......
cmake/inference_lib.cmake
浏览文件 @ cbdb8a17
......@@ -131,15 +131,6 @@ elseif (NOT CBLAS_FOUND OR WIN32)
)
endif ()
if (WITH_GPU AND NOT WIN32)
set(dgc_dir "${FLUID_INSTALL_DIR}/third_party/install/dgc")
copy(dgc_lib
SRCS ${DGC_INSTALL_DIR}/lib ${DGC_INSTALL_DIR}/include
DSTS ${dgc_dir} ${dgc_dir}
DEPS dgc)
endif()
if (WITH_MKLDNN)
set(dst_dir "${FLUID_INSTALL_DIR}/third_party/install/mkldnn")
copy(mkldnn_lib
......
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @ cbdb8a17
......@@ -24,15 +24,19 @@ if(WITH_DISTRIBUTE)
endif()
endif()
set(all_reduce_deps all_reduce_op_handle)
if(WITH_GPU)
set(dgc_deps "")
if(NOT WIN32)
set(dgc_deps dgc)
endif()
nv_library(all_reduce_op_handle SRCS all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
dynload_cuda variable_visitor ${dgc_deps})
dynload_cuda variable_visitor)
nv_library(fused_all_reduce_op_handle SRCS fused_all_reduce_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory
dynload_cuda variable_visitor)
if(WITH_DGC)
nv_library(sparse_all_reduce_op_handle SRCS sparse_all_reduce_op_handle.cc DEPS op_handle_base scope
lod_tensor ddim memory dynload_cuda variable_visitor dgc all_reduce_op_handle)
set(all_reduce_deps sparse_all_reduce_op_handle)
endif()
if(WITH_DISTRIBUTE)
nv_library(reduce_op_handle SRCS reduce_op_handle.cc DEPS op_handle_base variable_visitor scope
ddim dynload_cuda selected_rows_functor sendrecvop_rpc)
......@@ -80,7 +84,7 @@ cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS grap
cc_library(all_reduce_deps_pass SRCS all_reduce_deps_pass.cc DEPS graph graph_helper pass)
cc_library(multi_devices_graph_pass SRCS multi_devices_graph_pass.cc DEPS multi_devices_helper computation_op_handle
scale_loss_grad_op_handle rpc_op_handle fetch_barrier_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle fused_broadcast_op_handle)
scale_loss_grad_op_handle rpc_op_handle fetch_barrier_op_handle ${all_reduce_deps} reduce_op_handle broadcast_op_handle fused_broadcast_op_handle)
cc_library(fuse_all_reduce_op_pass SRCS fuse_all_reduce_op_pass.cc DEPS graph graph_helper fused_all_reduce_op_handle)
......
paddle/fluid/framework/details/all_reduce_op_handle.cc
浏览文件 @ cbdb8a17
......@@ -17,11 +17,6 @@
#include "paddle/fluid/framework/details/reduce_and_gather.h"
#include "paddle/fluid/framework/details/variable_visitor.h"
#include "paddle/fluid/framework/operator.h"
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
#include "dgc/dgc.h"
#endif
#include "paddle/fluid/platform/gpu_info.h"
#include "paddle/fluid/platform/profiler.h"
......@@ -40,23 +35,16 @@ namespace details {
AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
const platform::NCCLContextMap *ctxs,
bool is_encoded, int nranks)
const platform::NCCLContextMap *ctxs)
: OpHandleBase(node),
local_scopes_(local_scopes),
places_(places),
nccl_ctxs_(ctxs),
is_encoded_(is_encoded),
nranks_(nranks) {
nccl_ctxs_(ctxs) {
if (nccl_ctxs_) {
for (auto &p : places_) {
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,
......@@ -66,92 +54,8 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
#endif
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
void AllReduceOpHandle::RunImplEncoded() {
platform::RecordEvent record_event(Name());
WaitInputVarGenerated();
auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
PADDLE_ENFORCE_EQ(
in_var_handles.size(), places_.size(),
"The NoDummyInputSize should be equal to the number of places.");
PADDLE_ENFORCE_EQ(
in_var_handles.size(), out_var_handles.size(),
"The NoDummyInputSize and NoDummyOutputSize should be equal.");
std::vector<const LoDTensor *> ins;
std::vector<LoDTensor *> outs;
int k = -1;
for (size_t i = 0; i < local_scopes_.size(); ++i) {
auto &local_scope =
local_scopes_[i]->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto original_name =
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, "%s should not be null", encode_var_name);
auto &in = in_var->Get<LoDTensor>();
ins.emplace_back(&in);
auto *out = local_scope->FindVar(out_var_handles[i]->name())
->GetMutable<LoDTensor>();
outs.emplace_back(out);
if (k < 0) {
k = GetKValue(in_var_handles[i]->name());
}
}
PADDLE_ENFORCE(platform::is_gpu_place(ins[0]->place()));
PADDLE_ENFORCE(platform::is_gpu_place(outs[0]->place()));
PADDLE_ENFORCE(nccl_ctxs_, "nccl_ctxs should not be nullptr.");
int dtype = -1;
size_t in_numel = 0;
size_t out_numel = 0;
PADDLE_ENFORCE(nranks_ > 1);
std::vector<std::function<void()>> all_reduce_calls;
for (size_t i = 0; i < local_scopes_.size(); ++i) {
auto &place = places_[i];
auto &in = *ins[i];
void *in_tensor_buf = const_cast<void *>(in.data<void>());
auto &out = *outs[i];
float *out_tensor_buf = out.data<float>();
dtype = (dtype == -1) ? platform::ToNCCLDataType(in.type()) : dtype;
in_numel = (in_numel == 0) ? static_cast<size_t>(in.numel()) : in_numel;
PADDLE_ENFORCE(in_numel % 2 == 0);
PADDLE_ENFORCE(in_numel / 2 == static_cast<size_t>(k));
out_numel = (out_numel == 0) ? static_cast<size_t>(out.numel()) : out_numel;
int dev_id = boost::get<platform::CUDAPlace>(place).device;
auto &nccl_ctx = nccl_ctxs_->at(dev_id);
auto stream = nccl_ctx.stream();
auto comm = nccl_ctx.comm_;
auto &allocator =
platform::DeviceTemporaryAllocator::Instance().Get(place, stream);
int encode_size = 2 * k * sizeof(int);
// dgc use ncclAllGather to get all the encoded data
// so the buffer need nranks.
int buf_size = nranks_ * encode_size;
auto tmp_ious_data = allocator.Allocate(buf_size);
void *gather_buff = reinterpret_cast<void *>(tmp_ious_data->ptr());
VLOG(10) << "in_numel:" << in_numel << ", out_numel:" << out_numel
<< ", nranks:" << nranks_ << ", gather_buf size:" << buf_size
<< ", k:" << k << ", place:" << place << ", dtype:" << dtype;
all_reduce_calls.emplace_back([=] {
PADDLE_ENFORCE(paddle::communication::dgc::sparseAllGReduce(
in_tensor_buf, gather_buff, k, out_tensor_buf, out_numel, comm,
stream));
});
}
void AllReduceOpHandle::RunAllReduceFuncs(
const std::vector<std::function<void()>> &all_reduce_calls) {
this->RunAndRecordEvent([&] {
if (all_reduce_calls.size() == 1UL) {
// Do not use NCCLGroup when manage NCCL by per thread per device
......@@ -182,68 +86,9 @@ void AllReduceOpHandle::RunImplEncoded() {
}
}
}
int AllReduceOpHandle::GetKValue(const std::string &grad_name) {
auto original_name = paddle::framework::GradOriginalVarName(grad_name);
auto var_name = original_name + g_dgc_k;
PADDLE_ENFORCE(local_scopes_.size() > 0);
auto *scope = local_scopes_[0];
auto &local_scope = scope->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto var = local_scope->FindVar(var_name);
PADDLE_ENFORCE_NOT_NULL(var);
auto tensor = var->Get<LoDTensor>().data<float>();
return *tensor;
}
#endif
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
bool AllReduceOpHandle::IsEncoded() {
if (!is_encoded_) {
return false;
}
auto counter_name = g_dgc_counter_name;
auto step_name = g_dgc_rampup_begin_step;
PADDLE_ENFORCE(local_scopes_.size() > 0);
auto *scope = local_scopes_[0];
auto &local_scope = scope->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto count_var = local_scope->FindVar(counter_name);
auto step_var = local_scope->FindVar(step_name);
if (count_var == nullptr || step_var == nullptr) {
PADDLE_THROW("not find count_var:%s or step_var:%s", counter_name,
step_var);
}
float count = *count_var->Get<LoDTensor>().data<float>();
float step = *step_var->Get<LoDTensor>().data<float>();
if (static_cast<int>(count) < static_cast<int>(step)) {
VLOG(10) << "in all_reduce currentstep:" << count
<< " < rampup_begin_step:" << step
<< " so not use sparse all reduce";
return false;
}
return true;
}
#else
bool AllReduceOpHandle::IsEncoded() { return false; }
#endif
void AllReduceOpHandle::RunImpl() {
if (!IsEncoded()) {
RunImplNormal();
return;
}
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
RunImplEncoded();
#else
PADDLE_THROW("Not compiled with CUDA");
#endif
}
void AllReduceOpHandle::RunImplNormal() {
platform::RecordEvent record_event(Name());
WaitInputVarGenerated();
......@@ -304,27 +149,7 @@ void AllReduceOpHandle::RunImplNormal() {
comm, stream));
});
}
this->RunAndRecordEvent([&] {
if (all_reduce_calls.size() == 1UL) {
// Do not use NCCLGroup when manage NCCL by per thread per device
all_reduce_calls[0]();
} else {
platform::NCCLGroupGuard guard;
for (auto &call : all_reduce_calls) {
call();
}
}
});
if (FLAGS_sync_nccl_allreduce) {
for (auto &p : places_) {
int dev_id = boost::get<platform::CUDAPlace>(p).device;
auto &nccl_ctx = nccl_ctxs_->at(dev_id);
auto stream = nccl_ctx.stream();
cudaStreamSynchronize(stream);
}
}
RunAllReduceFuncs(all_reduce_calls);
#else
PADDLE_THROW("Not compiled with CUDA");
#endif
......
paddle/fluid/framework/details/all_reduce_op_handle.h
浏览文件 @ cbdb8a17
......@@ -28,19 +28,12 @@ namespace paddle {
namespace framework {
namespace details {
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
constexpr char g_dgc_counter_name[] = "__g_dgc_counter__";
constexpr char g_dgc_rampup_begin_step[] = "__g_rampup_begin_step__";
constexpr char g_dgc_encoded[] = "__dgc_encoded__";
constexpr char g_dgc_k[] = "__dgc_k__";
#endif
struct AllReduceOpHandle : public OpHandleBase {
class AllReduceOpHandle : public OpHandleBase {
public:
#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,
bool is_encoded = false, int nranks = -1);
const platform::NCCLContextMap *ctxs);
#else
AllReduceOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places);
......@@ -54,18 +47,13 @@ struct AllReduceOpHandle : public OpHandleBase {
protected:
void RunImpl() override;
private:
std::vector<Scope *> local_scopes_;
std::vector<platform::Place> places_;
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
void RunImplEncoded();
void RunAllReduceFuncs(
const std::vector<std::function<void()>> &all_reduce_calls);
const platform::NCCLContextMap *nccl_ctxs_;
bool is_encoded_{false};
int nranks_{-1};
int GetKValue(const std::string &grad_name);
#endif
void RunImplNormal();
bool IsEncoded();
};
} // namespace details
......
paddle/fluid/framework/details/dgc_const_values.h 0 → 100644
浏览文件 @ cbdb8a17
// Copyright (c) 2018 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 <string>
namespace paddle {
namespace framework {
namespace details {
constexpr char g_dgc_counter_name[] = "__g_dgc_counter__";
constexpr char g_dgc_rampup_begin_step[] = "__g_rampup_begin_step__";
constexpr char g_dgc_u[] = "__dgc_u__";
constexpr char g_dgc_v[] = "__dgc_v__";
constexpr char g_dgc_k[] = "__dgc_k__";
constexpr char g_dgc_encoded[] = "__dgc_encoded__";
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/multi_devices_graph_pass.cc
浏览文件 @ cbdb8a17
......@@ -34,6 +34,10 @@
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/operators/math/math_function.h"
#if defined(PADDLE_WITH_DGC)
#include "paddle/fluid/framework/details/sparse_all_reduce_op_handle.h"
#endif
namespace paddle {
namespace framework {
namespace details {
......@@ -438,12 +442,22 @@ void MultiDevSSAGraphBuilderBase::CreateAllReduceOp(ir::Graph *result,
auto append_allreduce_op = [&](
const std::vector<Scope *> &scopes,
const std::vector<platform::Place> &places) -> OpHandleBase * {
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
#if defined(PADDLE_WITH_DGC)
if (is_encoded) {
result->Get<GraphOps>(kGraphOps).emplace_back(new SparseAllReduceOpHandle(
result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
scopes, places, nccl_ctxs_, is_encoded,
static_cast<int>(strategy_.trainers_endpoints_.size()) *
places_.size()));
} else {
result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
scopes, places, nccl_ctxs_));
}
#elif defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
scopes, places, nccl_ctxs_, is_encoded,
static_cast<int>(strategy_.trainers_endpoints_.size()) *
places_.size()));
scopes, places, nccl_ctxs_));
#else
result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
......@@ -561,7 +575,11 @@ void AllReduceSSAGraphBuilder::InsertCollectiveOp(
CreateReduceOp(result, g_name, 0);
CreateBroadcastOp(result, g_name, 0);
} else {
#if defined(PADDLE_WITH_DGC)
CreateAllReduceOp(result, g_name, IsEncoded(p_name));
#else
CreateAllReduceOp(result, g_name);
#endif
}
}
......@@ -965,8 +983,9 @@ int DistSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
return op_dev_id;
}
bool DistSSAGraphBuilder::IsEncoded(const std::string &p_name) const {
auto u_name = p_name + "__dgc_u__";
#if defined(PADDLE_WITH_DGC)
bool AllReduceSSAGraphBuilder::IsEncoded(const std::string &p_name) const {
auto u_name = p_name + g_dgc_u;
auto it = all_vars_.find(u_name);
if (it == all_vars_.end()) {
VLOG(10) << "can't find u_name, so it's not encoded:" << u_name;
......@@ -975,6 +994,11 @@ bool DistSSAGraphBuilder::IsEncoded(const std::string &p_name) const {
return true;
}
#else
bool AllReduceSSAGraphBuilder::IsEncoded(const std::string &p_name) const {
return false;
}
#endif
void DistSSAGraphBuilder::InsertCollectiveOp(ir::Graph *result,
const std::string &p_name,
......@@ -992,11 +1016,7 @@ void DistSSAGraphBuilder::InsertCollectiveOp(ir::Graph *result,
CreateReduceOp(result, g_name, 0);
CreateBroadcastOp(result, g_name, 0);
} else {
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
CreateAllReduceOp(result, g_name, IsEncoded(p_name));
#else
PADDLE_ENFORCE(false, "Compiled withoud cuda!");
#endif
CreateAllReduceOp(result, g_name);
}
break;
default:
......
paddle/fluid/framework/details/multi_devices_graph_pass.h
浏览文件 @ cbdb8a17
......@@ -113,6 +113,8 @@ class AllReduceSSAGraphBuilder : public MultiDevSSAGraphBuilderBase {
const std::string &g_name) const;
virtual void InsertPostprocessOps(ir::Graph *result) const {}
bool IsEncoded(const std::string &p_name) const;
};
class AsyncSSAGraphBuilder : public MultiDevSSAGraphBuilderBase {
......@@ -203,8 +205,6 @@ class DistSSAGraphBuilder : public BalanceVarSSAGraphBuilder {
mutable std::vector<std::unordered_set<std::string>> bcast_var_name_set_;
mutable bool need_broadcast_var_{false};
bool IsEncoded(const std::string &p_name) const;
};
std::unordered_set<std::string> &MultiDevSSAGraphBuilder();
......
paddle/fluid/framework/details/sparse_all_reduce_op_handle.cc 0 → 100644
浏览文件 @ cbdb8a17
// Copyright (c) 2018 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/details/sparse_all_reduce_op_handle.h"
#include <algorithm>
#include "dgc/dgc.h"
#include "paddle/fluid/framework/details/container_cast.h"
#include "paddle/fluid/framework/details/reduce_and_gather.h"
#include "paddle/fluid/framework/details/variable_visitor.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/platform/gpu_info.h"
#include "paddle/fluid/platform/profiler.h"
DECLARE_bool(sync_nccl_allreduce);
namespace paddle {
namespace framework {
namespace details {
SparseAllReduceOpHandle::SparseAllReduceOpHandle(
ir::Node *node, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
const platform::NCCLContextMap *ctxs, bool is_encoded, int nranks)
: AllReduceOpHandle(node, local_scopes, places, ctxs),
is_encoded_(is_encoded),
nranks_(nranks) {
// TODO(gongwb) :polish them!
if (is_encoded) {
VLOG(1) << "Use dgc allreduce mode";
}
}
void SparseAllReduceOpHandle::RunImplEncoded() {
platform::RecordEvent record_event(Name());
WaitInputVarGenerated();
auto in_var_handles = DynamicCast<VarHandle>(this->Inputs());
auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
PADDLE_ENFORCE_EQ(
in_var_handles.size(), places_.size(),
"The NoDummyInputSize should be equal to the number of places.");
PADDLE_ENFORCE_EQ(
in_var_handles.size(), out_var_handles.size(),
"The NoDummyInputSize and NoDummyOutputSize should be equal.");
std::vector<const LoDTensor *> ins;
std::vector<LoDTensor *> outs;
int k = -1;
for (size_t i = 0; i < local_scopes_.size(); ++i) {
auto &local_scope =
local_scopes_[i]->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto original_name =
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, "%s should not be null", encode_var_name);
auto &in = in_var->Get<LoDTensor>();
ins.emplace_back(&in);
auto *out = local_scope->FindVar(out_var_handles[i]->name())
->GetMutable<LoDTensor>();
outs.emplace_back(out);
if (k < 0) {
k = GetKValue(in_var_handles[i]->name());
}
}
PADDLE_ENFORCE(platform::is_gpu_place(ins[0]->place()));
PADDLE_ENFORCE(platform::is_gpu_place(outs[0]->place()));
PADDLE_ENFORCE(nccl_ctxs_, "nccl_ctxs should not be nullptr.");
int dtype = -1;
size_t in_numel = 0;
size_t out_numel = 0;
PADDLE_ENFORCE(nranks_ > 1);
std::vector<std::function<void()>> all_reduce_calls;
for (size_t i = 0; i < local_scopes_.size(); ++i) {
auto &place = places_[i];
auto &in = *ins[i];
void *in_tensor_buf = const_cast<void *>(in.data<void>());
auto &out = *outs[i];
float *out_tensor_buf = out.data<float>();
dtype = (dtype == -1) ? platform::ToNCCLDataType(in.type()) : dtype;
in_numel = (in_numel == 0) ? static_cast<size_t>(in.numel()) : in_numel;
PADDLE_ENFORCE(in_numel % 2 == 0);
PADDLE_ENFORCE(in_numel / 2 == static_cast<size_t>(k));
out_numel = (out_numel == 0) ? static_cast<size_t>(out.numel()) : out_numel;
int dev_id = boost::get<platform::CUDAPlace>(place).device;
auto &nccl_ctx = nccl_ctxs_->at(dev_id);
auto stream = nccl_ctx.stream();
auto comm = nccl_ctx.comm_;
auto &allocator =
platform::DeviceTemporaryAllocator::Instance().Get(place, stream);
int encode_size = 2 * k * sizeof(int);
// dgc use ncclAllGather to get all the encoded data
// so the buffer need nranks.
int buf_size = nranks_ * encode_size;
auto tmp_ious_data = allocator.Allocate(buf_size);
void *gather_buff = reinterpret_cast<void *>(tmp_ious_data->ptr());
VLOG(10) << "in_numel:" << in_numel << ", out_numel:" << out_numel
<< ", nranks:" << nranks_ << ", gather_buf size:" << buf_size
<< ", k:" << k << ", place:" << place << ", dtype:" << dtype;
all_reduce_calls.emplace_back([=] {
PADDLE_ENFORCE(paddle::communication::dgc::sparseAllGReduce(
in_tensor_buf, gather_buff, k, out_tensor_buf, out_numel, comm,
stream));
});
}
RunAllReduceFuncs(all_reduce_calls);
}
int SparseAllReduceOpHandle::GetKValue(const std::string &grad_name) {
auto original_name = paddle::framework::GradOriginalVarName(grad_name);
auto var_name = original_name + g_dgc_k;
PADDLE_ENFORCE(local_scopes_.size() > 0);
auto *scope = local_scopes_[0];
auto &local_scope = scope->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto var = local_scope->FindVar(var_name);
PADDLE_ENFORCE_NOT_NULL(var);
auto tensor = var->Get<LoDTensor>().data<float>();
return *tensor;
}
bool SparseAllReduceOpHandle::IsEncoded() {
if (!is_encoded_) {
return false;
}
auto counter_name = g_dgc_counter_name;
auto step_name = g_dgc_rampup_begin_step;
PADDLE_ENFORCE(local_scopes_.size() > 0);
auto *scope = local_scopes_[0];
auto &local_scope = scope->FindVar(kLocalExecScopeName)->Get<Scope *>();
auto count_var = local_scope->FindVar(counter_name);
auto step_var = local_scope->FindVar(step_name);
if (count_var == nullptr || step_var == nullptr) {
PADDLE_THROW("not find count_var:%s or step_var:%s", counter_name,
step_var);
}
float count = *count_var->Get<LoDTensor>().data<float>();
float step = *step_var->Get<LoDTensor>().data<float>();
if (static_cast<int>(count) < static_cast<int>(step)) {
VLOG(10) << "in all_reduce currentstep:" << count
<< " < rampup_begin_step:" << step
<< " so not use sparse all reduce";
return false;
}
return true;
}
void SparseAllReduceOpHandle::RunImpl() {
if (!IsEncoded()) {
AllReduceOpHandle::RunImpl();
return;
}
RunImplEncoded();
}
std::string SparseAllReduceOpHandle::Name() const {
return "sparse_all_reduce";
}
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/framework/details/sparse_all_reduce_op_handle.h 0 → 100644
浏览文件 @ cbdb8a17
// Copyright (c) 2018 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 <string>
#include <vector>
#include "paddle/fluid/framework/details/all_reduce_op_handle.h"
#include "paddle/fluid/framework/details/dgc_const_values.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/platform/nccl_helper.h"
namespace paddle {
namespace framework {
namespace details {
class SparseAllReduceOpHandle : public AllReduceOpHandle {
public:
SparseAllReduceOpHandle(ir::Node *node,
const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
const platform::NCCLContextMap *ctxs,
bool is_encoded = false, int nranks = -1);
std::string Name() const override;
protected:
void RunImpl() override;
int GetKValue(const std::string &grad_name);
bool IsEncoded();
void RunImplEncoded();
private:
bool is_encoded_{false};
int nranks_{-1};
};
} // namespace details
} // namespace framework
} // namespace paddle
paddle/fluid/inference/CMakeLists.txt
浏览文件 @ cbdb8a17
......@@ -49,11 +49,6 @@ set(SHARED_INFERENCE_SRCS
${mkldnn_quantizer_src}
${CMAKE_CURRENT_SOURCE_DIR}/api/details/zero_copy_tensor.cc)
# FIXME(gongwb): hidden libdgc.a
if(WITH_GPU AND NOT WIN32)
set(fluid_modules ${fluid_modules} dgc)
endif()
if(WIN32)
sep_library(paddle_fluid DEPS ${fluid_modules} ${STATIC_INFERENCE_APIS} zero_copy_tensor reset_tensor_array
analysis_config ${mkldnn_quantizer_cfg} paddle_pass_builder)
......
paddle/fluid/operators/CMakeLists.txt
浏览文件 @ cbdb8a17
......@@ -72,7 +72,7 @@ endif()
set(COMMON_OP_DEPS ${OP_HEADER_DEPS})
if (WITH_GPU AND NOT WIN32)
if (WITH_DGC)
op_library(dgc_op DEPS dgc)
file(APPEND ${pybind_file} "USE_CUDA_ONLY_OP(dgc);\n")
set(COMMON_OP_DEPS ${COMMON_OP_DEPS} dgc)
......
paddle/fluid/platform/CMakeLists.txt
浏览文件 @ cbdb8a17
......@@ -45,13 +45,12 @@ cc_library(cpu_helper SRCS cpu_helper.cc DEPS cblas enforce)
cc_test(cpu_helper_test SRCS cpu_helper_test.cc DEPS cpu_helper)
set(dgc_deps "")
IF(WITH_DGC)
set(dgc_deps dgc)
ENDIF()
IF(WITH_GPU)
set(GPU_CTX_DEPS dynload_cuda dynamic_loader)
if(NOT WIN32)
set(dgc_deps dgc)
endif()
ELSE()
set(dgc_deps)
ENDIF()
IF(WITH_MKLDNN)
......
paddle/fluid/platform/init.cc
浏览文件 @ cbdb8a17
......@@ -31,7 +31,7 @@ limitations under the License. */
#include "paddle/fluid/platform/place.h"
#include "paddle/fluid/string/piece.h"
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
#if defined(PADDLE_WITH_DGC)
#include "dgc/dgc.h"
#endif
......@@ -211,7 +211,7 @@ void InitGLOG(const std::string &prog_name) {
#endif
}
#if defined(PADDLE_WITH_CUDA) && !defined(_WIN32)
#if defined(PADDLE_WITH_DGC)
void InitDGC() {
std::call_once(dgc_init_flag, []() {
PADDLE_ENFORCE(paddle::communication::dgc::dynloadNcclLib());
......
paddle/fluid/pybind/const_value.cc
浏览文件 @ cbdb8a17
......@@ -17,6 +17,11 @@ limitations under the License. */
#include "paddle/fluid/framework/op_proto_maker.h"
#include "paddle/fluid/framework/operator.h"
#if defined(PADDLE_WITH_DGC)
#include "paddle/fluid/framework/details/dgc_const_values.h"
#include "paddle/fluid/framework/details/sparse_all_reduce_op_handle.h"
#endif
namespace paddle {
namespace pybind {
......@@ -52,6 +57,17 @@ void BindConstValue(pybind11::module* m) {
op_proto_and_checker_maker.def(
"kOpCreationCallstackAttrName",
framework::OpProtoAndCheckerMaker::OpCreationCallstackAttrName);
#if defined(PADDLE_WITH_DGC)
auto dgc = m->def_submodule("dgc");
dgc.def("kDGCUName", [] { return framework::details::g_dgc_u; });
dgc.def("kDGCVName", [] { return framework::details::g_dgc_v; });
dgc.def("kDGCKName", [] { return framework::details::g_dgc_k; });
dgc.def("kDGCEncodedName", [] { return framework::details::g_dgc_encoded; });
dgc.def("kDGCCounterName",
[] { return framework::details::g_dgc_counter_name; });
dgc.def("kDGCRampUpBeginStepName",
[] { return framework::details::g_dgc_rampup_begin_step; });
#endif
}
} // namespace pybind
......
python/paddle/fluid/optimizer.py
浏览文件 @ cbdb8a17
......@@ -751,14 +751,14 @@ class DGCMomentumOptimizer(MomentumOptimizer):
# step counter
self._global_step_var = self._add_auto_increment_var(
counter_name='__g_dgc_counter__', begin=0)
counter_name=core.dgc.kDGCCounterName(), begin=0)
# rampup begin step var for all_reduce_op_handle
self._rampup_begin_step_var = tensor.create_global_var(
shape=[1],
dtype=core.VarDesc.VarType.FP32,
persistable=True,
name='__g_rampup_begin_step__',
name=core.dgc.kDGCRampUpBeginStepName(),
value=self._rampup_begin_step * 1.0,
force_cpu=True)
......@@ -774,20 +774,20 @@ class DGCMomentumOptimizer(MomentumOptimizer):
shape=param_var.shape,
dtype=param_var.dtype,
persistable=True,
name=param_var.name + "__dgc_u__",
name=param_var.name + core.dgc.kDGCUName(),
value=0.0)
v_var = tensor.create_global_var(
shape=param_var.shape,
dtype=param_var.dtype,
persistable=True,
name=param_var.name + "__dgc_v__",
name=param_var.name + core.dgc.kDGCVName(),
value=0.0)
k_var = tensor.create_global_var(
shape=[1],
dtype=param_var.dtype,
persistable=True,
name=param_var.name + "__dgc_k__",
name=param_var.name + core.dgc.kDGCKName(),
value=0.0,
force_cpu=True)
......@@ -795,7 +795,7 @@ class DGCMomentumOptimizer(MomentumOptimizer):
shape=[1],
dtype=param_var.dtype,
persistable=True,
name=param_var.name + "__dgc_encoded__",
name=param_var.name + core.dgc.kDGCEncodedName(),
value=0.0,
force_cpu=False)
......
python/paddle/fluid/parallel_executor.py
浏览文件 @ cbdb8a17
......@@ -104,11 +104,13 @@ class ParallelExecutor(object):
self._scope = scope if scope is not None else executor.global_scope()
if main_program is not None and main_program._enable_dgc:
assert num_trainers > 1
assert build_strategy.reduce_strategy == BuildStrategy.ReduceStrategy.AllReduce
assert num_trainers > 1, "dgc is not useful when num_trainers <= 1"
assert build_strategy.reduce_strategy == BuildStrategy.ReduceStrategy.AllReduce, "dgc \
only used for allreduce"
assert num_trainers * len(
self._places) > 1, "dgc is not useful for single card training"
assert use_cuda
assert use_cuda, "dgc only used under cuda"
main_program = main_program if main_program is not None \
else framework.default_main_program()
......@@ -125,10 +127,6 @@ class ParallelExecutor(object):
share_vars_from=share_vars_from._compiled_program
if share_vars_from else None)
# FIXME(gongwb): I will move dgc from dist mode to allreduce mode in next pr.
if main_program._enable_dgc:
self._compiled_program._build_strategy.is_distribution = True
self._place = core.CUDAPlace(0) if use_cuda else core.CPUPlace()
self._exe = executor.Executor(self._place)
self._compiled_program._compile(place=self._place, scope=self._scope)
......
python/paddle/fluid/tests/unittests/CMakeLists.txt
浏览文件 @ cbdb8a17
......@@ -97,11 +97,13 @@ py_test_modules(test_imperative_se_resnext MODULES test_imperative_se_resnext EN
if(WITH_DISTRIBUTE)
py_test_modules(test_dist_train MODULES test_dist_train SERIAL)
set_tests_properties(test_listen_and_serv_op PROPERTIES TIMEOUT 20)
if(WITH_DGC)
py_test_modules(test_dgc_op MODULES test_dgc_op)
endif()
if(NOT APPLE)
set_tests_properties(test_dist_mnist PROPERTIES TIMEOUT 200)
set_tests_properties(test_dist_word2vec PROPERTIES TIMEOUT 200)
py_test_modules(test_dist_se_resnext MODULES test_dist_se_resnext)
py_test_modules(test_dgc_op MODULES test_dgc_op)
set_tests_properties(test_dist_se_resnext PROPERTIES TIMEOUT 1000)
py_test_modules(test_dist_se_resnext_nccl MODULES test_dist_se_resnext_nccl SERIAL)
set_tests_properties(test_dist_se_resnext_nccl PROPERTIES TIMEOUT 1000)
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或