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
d215293c
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 >>
未验证
提交
d215293c
编写于
11月 29, 2018
作者:
Z
Zhaolong Xing
提交者:
GitHub
11月 29, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #14649 from NHZlX/add_params_sync_pass
Add params sync pass
上级
055da6e0
49c28b8c
变更
9
隐藏空白更改
内联
并排
Showing
9 changed file
with
128 addition
and
16 deletion
+128
-16
paddle/fluid/inference/analysis/passes/CMakeLists.txt
paddle/fluid/inference/analysis/passes/CMakeLists.txt
+2
-1
paddle/fluid/inference/analysis/passes/ir_analysis_compose_pass.cc
...uid/inference/analysis/passes/ir_analysis_compose_pass.cc
+1
-0
paddle/fluid/inference/analysis/passes/ir_graph_build_pass.cc
...le/fluid/inference/analysis/passes/ir_graph_build_pass.cc
+1
-6
paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.cc
...ence/analysis/passes/ir_params_sync_among_devices_pass.cc
+74
-0
paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.h
...rence/analysis/passes/ir_params_sync_among_devices_pass.h
+39
-0
paddle/fluid/inference/analysis/passes/passes.cc
paddle/fluid/inference/analysis/passes/passes.cc
+4
-0
paddle/fluid/inference/api/paddle_pass_builder.h
paddle/fluid/inference/api/paddle_pass_builder.h
+1
-5
paddle/fluid/inference/utils/benchmark.cc
paddle/fluid/inference/utils/benchmark.cc
+1
-1
paddle/fluid/inference/utils/benchmark.h
paddle/fluid/inference/utils/benchmark.h
+5
-3
未找到文件。
paddle/fluid/inference/analysis/passes/CMakeLists.txt
浏览文件 @
d215293c
cc_library
(
ir_graph_build_pass SRCS ir_graph_build_pass.cc DEPS analysis_pass argument ir_pass_manager
)
cc_library
(
ir_analysis_pass SRCS ir_analysis_pass.cc DEPS analysis_pass argument ir_pass_manager
)
cc_library
(
analysis_passes SRCS passes.cc DEPS ir_graph_build_pass ir_analysis_pass
)
cc_library
(
ir_params_sync_among_devices_pass SRCS ir_params_sync_among_devices_pass.cc DEPS analysis_pass argument ir_pass_manager
)
cc_library
(
analysis_passes SRCS passes.cc DEPS ir_graph_build_pass ir_analysis_pass ir_params_sync_among_devices_pass
)
set
(
analysis_deps
${
analysis_deps
}
ir_graph_build_pass
...
...
paddle/fluid/inference/analysis/passes/ir_analysis_compose_pass.cc
浏览文件 @
d215293c
...
...
@@ -61,6 +61,7 @@ void IrAnalysisComposePass::InitTensorRTAttrs(Argument *argument) {
void
IrAnalysisComposePass
::
ApplyIrPasses
(
Argument
*
argument
)
{
std
::
vector
<
std
::
string
>
passes
({
"ir_graph_build_pass"
,
"ir_analysis_pass"
,
"ir_params_sync_among_devices_pass"
,
});
for
(
const
auto
&
pass
:
passes
)
{
VLOG
(
2
)
<<
"Run pass "
<<
pass
;
...
...
paddle/fluid/inference/analysis/passes/ir_graph_build_pass.cc
浏览文件 @
d215293c
...
...
@@ -36,12 +36,7 @@ void IrGraphBuildPass::RunImpl(Argument *argument) {
// so that the parameters will on the same device, or they will keep copying
// between difference devices.
platform
::
Place
place
;
if
(
argument
->
use_gpu
())
{
PADDLE_ENFORCE
(
argument
->
gpu_device_id_valid
());
place
=
platform
::
CUDAPlace
(
argument
->
gpu_device_id
());
}
else
{
place
=
platform
::
CPUPlace
();
}
place
=
platform
::
CPUPlace
();
if
(
argument
->
model_dir_valid
())
{
auto
program
=
...
...
paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.cc
0 → 100644
浏览文件 @
d215293c
// 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/inference/analysis/passes/ir_params_sync_among_devices_pass.h"
#include "paddle/fluid/framework/data_layout.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/tensor_util.h"
#include "paddle/fluid/platform/enforce.h"
namespace
paddle
{
namespace
inference
{
namespace
analysis
{
void
IrParamsSyncAmongDevicesPass
::
RunImpl
(
Argument
*
argument
)
{
PADDLE_ENFORCE
(
argument
->
scope_valid
());
PADDLE_ENFORCE
(
argument
->
use_gpu_valid
());
platform
::
Place
place
;
// The parameters are on the cpu, therefore, synchronization is not necessary.
if
(
!
argument
->
use_gpu
())
return
;
LOG
(
INFO
)
<<
"Sync params from CPU to GPU"
;
PADDLE_ENFORCE
(
argument
->
gpu_device_id_valid
());
place
=
platform
::
CUDAPlace
(
argument
->
gpu_device_id
());
auto
*
scope
=
argument
->
scope_ptr
();
std
::
vector
<
std
::
string
>
all_vars
=
scope
->
LocalVarNames
();
// We get all the vars from local_scope instead of the ProgramDesc.
// Because there exists the case that new parameter variables are not added to
// the program in the analysis pass.
for
(
auto
&
var_name
:
all_vars
)
{
auto
*
var
=
scope
->
FindLocalVar
(
var_name
);
PADDLE_ENFORCE
(
var
!=
nullptr
);
if
(
var
->
IsType
<
framework
::
LoDTensor
>
()
||
var
->
IsType
<
framework
::
Tensor
>
())
{
auto
*
t
=
var
->
GetMutable
<
framework
::
LoDTensor
>
();
platform
::
CPUPlace
cpu_place
;
framework
::
LoDTensor
temp_tensor
;
temp_tensor
.
Resize
(
t
->
dims
());
temp_tensor
.
mutable_data
<
float
>
(
cpu_place
);
// Copy the parameter data to a tmp tensor.
TensorCopySync
(
*
t
,
cpu_place
,
&
temp_tensor
);
// Reallocation the space on GPU
t
->
mutable_data
<
float
>
(
place
);
// Copy parameter data to newly allocated GPU space.
TensorCopySync
(
temp_tensor
,
place
,
t
);
}
}
}
std
::
string
IrParamsSyncAmongDevicesPass
::
repr
()
const
{
return
"ir-params-sync-among-devices-pass"
;
}
}
// namespace analysis
}
// namespace inference
}
// namespace paddle
paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.h
0 → 100644
浏览文件 @
d215293c
// 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/scope.h"
#include "paddle/fluid/inference/analysis/analysis_pass.h"
#include "paddle/fluid/platform/place.h"
namespace
paddle
{
namespace
inference
{
namespace
analysis
{
/*
* Sync parameter from CPU to GPU.
*/
class
IrParamsSyncAmongDevicesPass
:
public
AnalysisPass
{
public:
void
RunImpl
(
Argument
*
argument
)
override
;
std
::
string
repr
()
const
override
;
};
}
// namespace analysis
}
// namespace inference
}
// namespace paddle
paddle/fluid/inference/analysis/passes/passes.cc
浏览文件 @
d215293c
...
...
@@ -16,6 +16,7 @@
#include "paddle/fluid/inference/analysis/passes/ir_analysis_compose_pass.cc"
#include "paddle/fluid/inference/analysis/passes/ir_analysis_pass.h"
#include "paddle/fluid/inference/analysis/passes/ir_graph_build_pass.h"
#include "paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.h"
namespace
paddle
{
namespace
inference
{
...
...
@@ -27,6 +28,9 @@ PassRegistry::PassRegistry() {
std
::
unique_ptr
<
AnalysisPass
>
(
new
IrGraphBuildPass
));
passes_
.
emplace
(
"ir_analysis_compose_pass"
,
std
::
unique_ptr
<
AnalysisPass
>
(
new
IrAnalysisComposePass
));
passes_
.
emplace
(
"ir_params_sync_among_devices_pass"
,
std
::
unique_ptr
<
AnalysisPass
>
(
new
IrParamsSyncAmongDevicesPass
));
}
}
// namespace analysis
...
...
paddle/fluid/inference/api/paddle_pass_builder.h
浏览文件 @
d215293c
...
...
@@ -116,12 +116,8 @@ class CpuPassStrategy : public PassStrategy {
class
GpuPassStrategy
:
public
PassStrategy
{
public:
GpuPassStrategy
()
:
PassStrategy
({})
{
// TODO(NHZlX) Problem with Data synchronization between GPU and CPU
// When running in GPU mode, the parameters are all on GPU. But the
// opearations of "conv_bn_fuse_pass" are on CPU.
passes_
.
assign
({
"infer_clean_graph_pass"
,
// "infer_clean_graph_pass", "conv_bn_fuse_pass",
"infer_clean_graph_pass"
,
"conv_bn_fuse_pass"
,
});
}
...
...
paddle/fluid/inference/utils/benchmark.cc
浏览文件 @
d215293c
...
...
@@ -33,7 +33,7 @@ std::string Benchmark::SerializeToString() const {
ss
<<
batch_size_
<<
"
\t
"
;
ss
<<
num_threads_
<<
"
\t
"
;
ss
<<
latency_
<<
"
\t
"
;
ss
<<
1000
/
latency_
;
ss
<<
1000
.0
/
latency_
;
ss
<<
'\n'
;
return
ss
.
str
();
}
...
...
paddle/fluid/inference/utils/benchmark.h
浏览文件 @
d215293c
...
...
@@ -11,9 +11,11 @@
// 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 <fstream>
#include <iostream>
#include <string>
namespace
paddle
{
namespace
inference
{
...
...
@@ -31,8 +33,8 @@ struct Benchmark {
bool
use_gpu
()
const
{
return
use_gpu_
;
}
void
SetUseGpu
()
{
use_gpu_
=
true
;
}
in
t
latency
()
const
{
return
latency_
;
}
void
SetLatency
(
in
t
x
)
{
latency_
=
x
;
}
floa
t
latency
()
const
{
return
latency_
;
}
void
SetLatency
(
floa
t
x
)
{
latency_
=
x
;
}
const
std
::
string
&
name
()
const
{
return
name_
;
}
void
SetName
(
const
std
::
string
&
name
)
{
name_
=
name
;
}
...
...
@@ -43,7 +45,7 @@ struct Benchmark {
private:
bool
use_gpu_
{
false
};
int
batch_size_
{
0
};
in
t
latency_
;
floa
t
latency_
;
int
num_threads_
{
1
};
std
::
string
name_
;
};
...
...
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