Skip to content
体验新版
项目
组织
正在加载...
登录
切换导航
打开侧边栏
PaddlePaddle
Paddle
提交
02aaecca
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2298
Star
20931
Fork
5422
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1423
列表
看板
标记
里程碑
合并请求
543
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
Paddle
项目概览
项目概览
详情
发布
仓库
仓库
文件
提交
分支
标签
贡献者
分支图
比较
Issue
1,423
Issue
1,423
列表
看板
标记
里程碑
合并请求
543
合并请求
543
Pages
分析
分析
仓库分析
DevOps
Wiki
0
Wiki
成员
成员
收起侧边栏
关闭侧边栏
动态
分支图
创建新Issue
提交
Issue看板
体验新版 GitCode,发现更多精彩内容 >>
提交
02aaecca
编写于
3月 26, 2018
作者:
Y
Yu Yang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Fix CPU compile
上级
54bd17fe
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
62 addition
and
13 deletion
+62
-13
paddle/fluid/framework/details/CMakeLists.txt
paddle/fluid/framework/details/CMakeLists.txt
+7
-1
paddle/fluid/framework/details/multi_devices_graph_builder.cc
...le/fluid/framework/details/multi_devices_graph_builder.cc
+31
-6
paddle/fluid/framework/details/multi_devices_graph_builder.h
paddle/fluid/framework/details/multi_devices_graph_builder.h
+11
-1
paddle/fluid/framework/parallel_executor.cc
paddle/fluid/framework/parallel_executor.cc
+11
-3
paddle/fluid/framework/parallel_executor.h
paddle/fluid/framework/parallel_executor.h
+0
-2
paddle/fluid/operators/reader/create_recordio_file_reader_op.cc
.../fluid/operators/reader/create_recordio_file_reader_op.cc
+2
-0
未找到文件。
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @
02aaecca
...
...
@@ -8,8 +8,14 @@ cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_pr
cc_library
(
ssa_graph SRCS ssa_graph.cc DEPS var_handle op_handle_base
)
cc_library
(
ssa_graph_builder SRCS ssa_graph_builder.cc DEPS ssa_graph
)
if
(
WITH_GPU
)
set
(
multi_devices_graph_builder_deps nccl_all_reduce_op_handle
)
else
()
set
(
multi_devices_graph_builder_deps
)
endif
()
cc_library
(
multi_devices_graph_builder SRCS multi_devices_graph_builder.cc DEPS ssa_graph_builder computation_op_handle
nccl_all_reduce_op_handle scale_loss_grad_op_handle
)
scale_loss_grad_op_handle
${
multi_devices_graph_builder_deps
}
)
cc_library
(
ssa_graph_executor SRCS ssa_graph_executor.cc DEPS ssa_graph
)
cc_library
(
threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope
simple_threadpool device_context
)
paddle/fluid/framework/details/multi_devices_graph_builder.cc
浏览文件 @
02aaecca
...
...
@@ -14,14 +14,18 @@
#include "paddle/fluid/framework/details/multi_devices_graph_builder.h"
#include "paddle/fluid/framework/details/computation_op_handle.h"
#include "paddle/fluid/framework/details/nccl_all_reduce_op_handle.h"
#include "paddle/fluid/framework/details/scale_loss_grad_op_handle.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/platform/nccl_helper.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/framework/details/nccl_all_reduce_op_handle.h"
#endif
namespace
paddle
{
namespace
framework
{
namespace
details
{
#ifdef PADDLE_WITH_CUDA
MultiDevSSAGraphBuilder
::
MultiDevSSAGraphBuilder
(
const
std
::
vector
<
platform
::
Place
>
&
places
,
const
std
::
string
&
loss_var_name
,
...
...
@@ -32,6 +36,16 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
places_
(
places
),
local_scopes_
(
local_scopes
),
nccl_ctxs_
(
nccl_ctxs
)
{
#else
MultiDevSSAGraphBuilder
::
MultiDevSSAGraphBuilder
(
const
std
::
vector
<
platform
::
Place
>
&
places
,
const
std
::
string
&
loss_var_name
,
const
std
::
unordered_set
<
std
::
string
>
&
params
,
const
std
::
vector
<
Scope
*>
&
local_scopes
)
:
loss_var_name_
(
loss_var_name
),
places_
(
places
),
local_scopes_
(
local_scopes
)
{
#endif
for
(
auto
&
p
:
params
)
{
grad_names_
.
insert
(
GradVarName
(
p
));
}
...
...
@@ -78,9 +92,16 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
if
(
is_forwarding
)
{
if
(
var_names
.
size
()
==
1
&&
var_names
[
0
]
==
loss_var_name_
)
{
// Insert ScaleCost OpHandle
// Insert ScaleCost OpHandle
#ifdef PADDLE_WITH_CUDA
auto
*
communication_dev_ctx
=
nccl_ctxs_
->
DevCtx
(
p
);
#else
auto
*
communication_dev_ctx
=
platform
::
DeviceContextPool
::
Instance
().
Get
(
platform
::
CPUPlace
());
#endif
op_handle
=
new
ScaleLossGradOpHandle
(
local_scopes_
.
size
(),
s
,
p
,
nccl_ctxs_
->
DevCtx
(
p
)
);
communication_dev_ctx
);
result
.
ops_
.
emplace_back
(
op_handle
);
// FIXME: Currently ScaleLossGradOp only use device_count as scale
...
...
@@ -103,7 +124,8 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
auto
var_names
=
op
->
OutputArgumentNames
();
for
(
auto
&
og
:
var_names
)
{
if
(
grad_names_
.
count
(
og
)
!=
0
)
{
// is param grad
// Insert NCCL AllReduce Op
// Insert NCCL AllReduce Op
#ifdef PADDLE_WITH_CUDA
result
.
ops_
.
emplace_back
(
new
NCCLAllReduceOpHandle
(
local_scopes_
,
places_
,
*
nccl_ctxs_
));
auto
*
op_handle
=
result
.
ops_
.
back
().
get
();
...
...
@@ -125,6 +147,9 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
op_handle
->
AddOutput
(
&
var
);
}
#else
PADDLE_ENFORCE
(
"Not implemented"
);
#endif
}
}
}
...
...
@@ -143,7 +168,7 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
}
return
std
::
unique_ptr
<
SSAGraph
>
(
graph
);
}
}
// namespace details
}
// namespace details
}
// namespace framework
}
// namespace paddle
paddle/fluid/framework/details/multi_devices_graph_builder.h
浏览文件 @
02aaecca
...
...
@@ -26,11 +26,18 @@ class Scope;
namespace
details
{
class
MultiDevSSAGraphBuilder
:
public
SSAGraphBuilder
{
public:
#ifdef PADDLE_WITH_CUDA
MultiDevSSAGraphBuilder
(
const
std
::
vector
<
platform
::
Place
>
&
places
,
const
std
::
string
&
loss_var_name
,
const
std
::
unordered_set
<
std
::
string
>
&
params
,
const
std
::
vector
<
Scope
*>
&
local_scopes
,
platform
::
NCCLContextMap
*
nccl_ctxs
);
#else
MultiDevSSAGraphBuilder
(
const
std
::
vector
<
platform
::
Place
>
&
places
,
const
std
::
string
&
loss_var_name
,
const
std
::
unordered_set
<
std
::
string
>
&
params
,
const
std
::
vector
<
Scope
*>
&
local_scopes
);
#endif
std
::
unique_ptr
<
SSAGraph
>
Build
(
const
ProgramDesc
&
program
)
const
override
;
...
...
@@ -38,8 +45,11 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
std
::
string
loss_var_name_
;
const
std
::
vector
<
platform
::
Place
>
&
places_
;
const
std
::
vector
<
Scope
*>
&
local_scopes_
;
platform
::
NCCLContextMap
*
nccl_ctxs_
;
std
::
unordered_set
<
std
::
string
>
grad_names_
;
#ifdef PADDLE_WITH_CUDA
platform
::
NCCLContextMap
*
nccl_ctxs_
;
#endif
};
}
// namespace details
}
// namespace framework
...
...
paddle/fluid/framework/parallel_executor.cc
浏览文件 @
02aaecca
...
...
@@ -16,7 +16,9 @@ limitations under the License. */
#include "ThreadPool.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/nccl_helper.h"
#endif
#include "paddle/fluid/framework/details/multi_devices_graph_builder.h"
#include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
...
...
@@ -64,13 +66,18 @@ ParallelExecutor::ParallelExecutor(
member_
->
local_scopes_
.
size
()
!=
1
)
{
// Is CUDA
BCastParamsToGPUs
(
startup_program
);
}
// Startup Program has been run. All local scopes has correct parameters.
// Startup Program has been run. All local scopes has correct parameters.
// Step 2. Convert main_program to SSA form and dependency graph. Also, insert
// ncclOp
// Step 2. Convert main_program to SSA form and dependency graph. Also, insert
// ncclOp
#ifdef PADDLE_WITH_CUDA
details
::
MultiDevSSAGraphBuilder
builder
(
member_
->
places_
,
loss_var_name
,
params
,
member_
->
local_scopes_
,
member_
->
nccl_ctxs_
.
get
());
#else
details
::
MultiDevSSAGraphBuilder
builder
(
member_
->
places_
,
loss_var_name
,
params
,
member_
->
local_scopes_
);
#endif
auto
graph
=
builder
.
Build
(
main_program
);
member_
->
executor_
.
reset
(
new
details
::
ThreadedSSAGraphExecutor
(
...
...
@@ -137,3 +144,4 @@ void ParallelExecutor::Run(const std::vector<std::string> &fetch_tensors,
}
// namespace framework
}
// namespace paddle
A
\ No newline at end of file
paddle/fluid/framework/parallel_executor.h
浏览文件 @
02aaecca
...
...
@@ -21,8 +21,6 @@ limitations under the License. */
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/operators/nccl/nccl_gpu_common.h"
#include "paddle/fluid/platform/device_context.h"
namespace
paddle
{
...
...
paddle/fluid/operators/reader/create_recordio_file_reader_op.cc
浏览文件 @
02aaecca
...
...
@@ -12,6 +12,8 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include <mutex>
#include <thread>
#include "paddle/fluid/operators/reader/reader_op_registry.h"
#include "paddle/fluid/recordio/scanner.h"
...
...
编辑
预览
Markdown
is supported
0%
请重试
或
添加新附件
.
添加附件
取消
You are about to add
0
people
to the discussion. Proceed with caution.
先完成此消息的编辑!
取消
想要评论请
注册
或
登录