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
8e4ad008
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 个月 前同步成功
通知
2292
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
8e4ad008
编写于
3月 18, 2019
作者:
Q
Qiyang Min
提交者:
GitHub
3月 18, 2019
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #16198 from velconia/imperative_train_speed
Improve imperative mode training speed
上级
2579ade4
36225373
变更
16
隐藏空白更改
内联
并排
Showing
16 changed file
with
170 addition
and
57 deletion
+170
-57
CMakeLists.txt
CMakeLists.txt
+2
-0
paddle/fluid/framework/grad_op_desc_maker.h
paddle/fluid/framework/grad_op_desc_maker.h
+5
-3
paddle/fluid/imperative/CMakeLists.txt
paddle/fluid/imperative/CMakeLists.txt
+1
-0
paddle/fluid/imperative/layer.cc
paddle/fluid/imperative/layer.cc
+3
-3
paddle/fluid/imperative/layer.h
paddle/fluid/imperative/layer.h
+17
-11
paddle/fluid/imperative/profiler.cc
paddle/fluid/imperative/profiler.cc
+62
-0
paddle/fluid/imperative/profiler.h
paddle/fluid/imperative/profiler.h
+25
-0
paddle/fluid/imperative/tracer.cc
paddle/fluid/imperative/tracer.cc
+3
-37
paddle/fluid/inference/CMakeLists.txt
paddle/fluid/inference/CMakeLists.txt
+1
-1
paddle/fluid/pybind/CMakeLists.txt
paddle/fluid/pybind/CMakeLists.txt
+1
-1
paddle/fluid/pybind/imperative.cc
paddle/fluid/pybind/imperative.cc
+2
-0
paddle/fluid/pybind/pybind.cc
paddle/fluid/pybind/pybind.cc
+6
-0
python/paddle/fluid/__init__.py
python/paddle/fluid/__init__.py
+2
-1
python/paddle/fluid/imperative/__init__.py
python/paddle/fluid/imperative/__init__.py
+4
-0
python/paddle/fluid/imperative/profiler.py
python/paddle/fluid/imperative/profiler.py
+30
-0
tools/manylinux1/build_scripts/build.sh
tools/manylinux1/build_scripts/build.sh
+6
-0
未找到文件。
CMakeLists.txt
浏览文件 @
8e4ad008
...
...
@@ -24,6 +24,8 @@ message(STATUS "CXX compiler: ${CMAKE_CXX_COMPILER}, version: "
"
${
CMAKE_CXX_COMPILER_ID
}
${
CMAKE_CXX_COMPILER_VERSION
}
"
)
message
(
STATUS
"C compiler:
${
CMAKE_C_COMPILER
}
, version: "
"
${
CMAKE_C_COMPILER_ID
}
${
CMAKE_C_COMPILER_VERSION
}
"
)
message
(
STATUS
"AR tools:
${
CMAKE_AR
}
"
)
if
(
WIN32
)
set
(
CMAKE_SUPPRESS_REGENERATION ON
)
set
(
CMAKE_STATIC_LIBRARY_PREFIX lib
)
...
...
paddle/fluid/framework/grad_op_desc_maker.h
浏览文件 @
8e4ad008
...
...
@@ -14,7 +14,9 @@ limitations under the License. */
#pragma once
#include <algorithm>
#include <memory>
#include <string>
#include <unordered_map>
#include <unordered_set>
#include <vector>
#include "paddle/fluid/framework/op_desc.h"
...
...
@@ -55,11 +57,11 @@ class GradOpDescMakerBase {
std
::
back_inserter
(
ret_val
),
[
this
](
const
std
::
string
&
fwd_var_name
)
->
std
::
string
{
auto
g_name
=
GradVarName
(
fwd_var_name
);
if
(
no_grad_set_
.
count
(
g_name
))
{
return
kEmptyVarName
;
}
else
{
if
(
no_grad_set_
.
empty
()
||
!
no_grad_set_
.
count
(
g_name
))
{
(
*
this
->
grad_to_var_
)[
g_name
]
=
fwd_var_name
;
return
g_name
;
}
else
{
return
kEmptyVarName
;
}
});
if
(
!
drop_empty_grad
)
{
...
...
paddle/fluid/imperative/CMakeLists.txt
浏览文件 @
8e4ad008
...
...
@@ -2,4 +2,5 @@ if(WITH_PYTHON)
cc_library
(
layer SRCS layer.cc DEPS proto_desc operator device_context blas pybind
)
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
)
endif
()
paddle/fluid/imperative/layer.cc
浏览文件 @
8e4ad008
...
...
@@ -239,7 +239,7 @@ std::map<std::string, std::vector<VarBase*>> OpBase::ApplyGrad() {
VLOG
(
3
)
<<
"apply grad op "
<<
grad_op_desc
->
Type
();
// Allocate tmp grad output variable
for
(
auto
it
:
grad_output_variable_map
)
{
for
(
const
auto
&
it
:
grad_output_variable_map
)
{
auto
&
outputs
=
tmp_grad_outputs
[
k
][
it
.
first
];
outputs
.
reserve
(
it
.
second
.
size
());
for
(
size_t
i
=
0
;
i
<
it
.
second
.
size
();
++
i
)
{
...
...
@@ -273,9 +273,9 @@ std::map<std::string, std::vector<VarBase*>> OpBase::ApplyGrad() {
// Add tmp grad outputs to original grad vars
for
(
size_t
k
=
0
;
k
<
grad_output_vars_
.
size
();
++
k
)
{
for
(
auto
it
:
grad_output_vars_
[
k
])
{
for
(
const
auto
&
it
:
grad_output_vars_
[
k
])
{
auto
&
outputs
=
tmp_grad_outputs
[
k
][
it
.
first
];
auto
&
origin_outputs
=
it
.
second
;
const
auto
&
origin_outputs
=
it
.
second
;
PADDLE_ENFORCE_EQ
(
outputs
.
size
(),
origin_outputs
.
size
());
for
(
size_t
i
=
0
;
i
<
outputs
.
size
();
++
i
)
{
...
...
paddle/fluid/imperative/layer.h
浏览文件 @
8e4ad008
...
...
@@ -294,17 +294,23 @@ class PYBIND11_HIDDEN OpBase {
void
InvokeBackwardHooks
();
void
TrackPreOp
(
const
VarBase
*
inp_var
,
const
std
::
string
&
inp_name
)
{
if
(
inp_var
->
PreOp
()
&&
!
inp_var
->
IsStopGradient
())
{
VLOG
(
3
)
<<
"add pre op "
<<
inp_var
->
PreOp
()
->
Type
()
<<
" in slot "
<<
inp_name
;
pre_ops_
[
inp_name
].
push_back
(
inp_var
->
PreOp
());
pre_ops_out_idx_
[
inp_name
].
push_back
(
inp_var
->
PreOpOutIdx
());
}
else
{
VLOG
(
3
)
<<
"no pre op in slot "
<<
inp_name
<<
" input var stop_gradient: "
<<
inp_var
->
IsStopGradient
();
pre_ops_
[
inp_name
].
push_back
(
nullptr
);
// pre_ops_out_idx_[inp_name].push_back(-1);
void
TrackPreOp
(
const
std
::
string
&
inp_name
,
const
std
::
vector
<
VarBase
*>&
inputs
)
{
auto
&
pre_ops_list
=
pre_ops_
[
inp_name
];
pre_ops_list
.
reserve
(
inputs
.
size
());
auto
&
pre_ops_out_idx_list
=
pre_ops_out_idx_
[
inp_name
];
for
(
VarBase
*
inp_var
:
inputs
)
{
if
(
inp_var
->
PreOp
()
&&
!
inp_var
->
IsStopGradient
())
{
VLOG
(
3
)
<<
"add pre op "
<<
inp_var
->
PreOp
()
->
Type
()
<<
" in slot "
<<
inp_name
;
pre_ops_list
.
emplace_back
(
inp_var
->
PreOp
());
pre_ops_out_idx_list
.
push_back
(
inp_var
->
PreOpOutIdx
());
}
else
{
VLOG
(
3
)
<<
"no pre op in slot "
<<
inp_name
<<
" input var stop_gradient: "
<<
inp_var
->
IsStopGradient
();
pre_ops_list
.
emplace_back
(
nullptr
);
// pre_ops_out_idx_list.push_back(-1);
}
}
}
...
...
paddle/fluid/imperative/profiler.cc
0 → 100644
浏览文件 @
8e4ad008
// 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/imperative/profiler.h"
#ifdef WITH_GPERFTOOLS
#include "gperftools/profiler.h"
#endif
#include <gflags/gflags.h>
#include <glog/logging.h>
#include <mutex> // NOLINT
#include <thread> // NOLINT
DEFINE_string
(
tracer_profile_fname
,
"xxgperf"
,
"Profiler filename for imperative tracer, which generated by gperftools."
"Only valid when compiled `WITH_PROFILER=ON`. Empty if disable."
);
namespace
paddle
{
namespace
imperative
{
static
std
::
once_flag
gTracerProfileOnce
;
#ifdef WITH_GPERFTOOLS
static
bool
gTracerProfilerStarted
=
false
;
#endif
void
StartProfile
()
{
if
(
!
FLAGS_tracer_profile_fname
.
empty
())
{
std
::
call_once
(
gTracerProfileOnce
,
[]
{
#ifdef WITH_GPERFTOOLS
ProfilerStart
(
FLAGS_tracer_profile_fname
.
c_str
());
gTracerProfilerStarted
=
true
;
#else
LOG
(
WARNING
)
<<
"Paddle is not compiled with gperftools. "
"FLAGS_tracer_profile_fname will be ignored"
;
#endif
});
}
}
void
StopProfile
()
{
#ifdef WITH_GPERFTOOLS
ProfilerFlush
();
#else
LOG
(
WARNING
)
<<
"Paddle is not compiled with gperftools. "
"FLAGS_tracer_profile_fname will be ignored"
;
#endif
}
}
// namespace imperative
}
// namespace paddle
paddle/fluid/imperative/profiler.h
0 → 100644
浏览文件 @
8e4ad008
// 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
namespace
paddle
{
namespace
imperative
{
extern
void
StartProfile
();
extern
void
StopProfile
();
}
// namespace imperative
}
// namespace paddle
paddle/fluid/imperative/tracer.cc
浏览文件 @
8e4ad008
...
...
@@ -23,23 +23,9 @@
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/enforce.h"
#ifdef WITH_GPERFTOOLS
#include "gperftools/profiler.h"
#endif
DEFINE_string
(
tracer_profile_fname
,
""
,
"Profiler filename for imperative tracer, which generated by gperftools."
"Only valid when compiled `WITH_PROFILER=ON`. Empty if disable."
);
namespace
paddle
{
namespace
imperative
{
static
std
::
once_flag
gTracerProfileOnce
;
#ifdef WITH_GPERFTOOLS
static
bool
gTracerProfilerStarted
=
false
;
#endif
void
CreateGradOp
(
const
framework
::
OpDesc
&
op_desc
,
const
std
::
unordered_set
<
std
::
string
>&
no_grad_set
,
const
std
::
vector
<
framework
::
BlockDesc
*>&
grad_sub_block
,
...
...
@@ -145,31 +131,13 @@ framework::VariableNameMap CreateOutputVarNameMap(
return
result
;
}
Tracer
::
Tracer
(
framework
::
BlockDesc
*
root_block
)
:
root_block_
(
root_block
)
{
if
(
!
FLAGS_tracer_profile_fname
.
empty
())
{
std
::
call_once
(
gTracerProfileOnce
,
[]
{
#ifdef WITH_GPERFTOOLS
ProfilerStart
(
FLAGS_tracer_profile_fname
.
c_str
());
gTracerProfilerStarted
=
true
;
#else
LOG
(
WARNING
)
<<
"Paddle is not compiled with gperftools. "
"FLAGS_tracer_profile_fname will be ignored"
;
#endif
});
}
}
Tracer
::
Tracer
(
framework
::
BlockDesc
*
root_block
)
:
root_block_
(
root_block
)
{}
std
::
set
<
std
::
string
>
Tracer
::
Trace
(
OpBase
*
op
,
const
VarBasePtrMap
&
inputs
,
const
VarBasePtrMap
&
outputs
,
framework
::
AttributeMap
attrs_map
,
const
platform
::
Place
expected_place
,
const
bool
stop_gradient
)
{
#ifdef WITH_GPERFTOOLS
if
(
gTracerProfilerStarted
)
{
ProfilerFlush
();
}
#endif
framework
::
VariableValueMap
invars_map
;
framework
::
VariableValueMap
outvars_map
;
...
...
@@ -184,7 +152,6 @@ std::set<std::string> Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
inp
->
Name
());
invars
.
emplace_back
(
inp
->
var_
);
op
->
TrackPreOp
(
inp
,
it
.
first
);
if
(
!
stop_gradient
)
{
current_vars_map
[
inp
->
Name
()]
=
inp
;
}
...
...
@@ -192,6 +159,7 @@ std::set<std::string> Tracer::Trace(OpBase* op, const VarBasePtrMap& inputs,
<<
" inited: "
<<
inp
->
var_
->
IsInitialized
()
<<
" stop_grad: "
<<
inp
->
IsStopGradient
();
}
op
->
TrackPreOp
(
it
.
first
,
it
.
second
);
}
op
->
output_vars_
=
outputs
;
...
...
@@ -319,9 +287,7 @@ std::vector<VarBase*> Tracer::PyTrace(OpBase* op,
std
::
vector
<
framework
::
Variable
*>
ret_vars
=
PyLayer
::
Apply
(
op
->
forward_id_
,
inputs
);
for
(
VarBase
*
inp
:
inputs
)
{
op
->
TrackPreOp
(
inp
,
PyLayer
::
kFwdInp
);
}
op
->
TrackPreOp
(
PyLayer
::
kFwdInp
,
inputs
);
std
::
vector
<
VarBase
*>&
outputs
=
op
->
output_vars_
[
PyLayer
::
kFwdOut
];
outputs
.
reserve
(
ret_vars
.
size
());
...
...
paddle/fluid/inference/CMakeLists.txt
浏览文件 @
8e4ad008
...
...
@@ -91,5 +91,5 @@ if(WITH_TESTING)
add_subdirectory
(
tests/book
)
if
(
WITH_INFERENCE_API_TEST
)
add_subdirectory
(
tests/api
)
endif
()
endif
()
endif
()
paddle/fluid/pybind/CMakeLists.txt
浏览文件 @
8e4ad008
set
(
PYBIND_DEPS pybind python proto_desc memory executor async_executor prune
feed_fetch_method pass_builder parallel_executor profiler layer scope_pool
tracer analysis_predictor
)
tracer analysis_predictor
imperative_profiler
)
if
(
WITH_PYTHON
)
list
(
APPEND PYBIND_DEPS py_func_op
)
...
...
paddle/fluid/pybind/imperative.cc
浏览文件 @
8e4ad008
...
...
@@ -42,6 +42,7 @@ void BindTracer(pybind11::module* m) {
framework
::
AttributeMap
attrs_map
,
const
platform
::
CPUPlace
expected_place
,
const
bool
stop_gradient
=
false
)
{
pybind11
::
gil_scoped_release
release
;
return
self
.
Trace
(
op
,
inputs
,
outputs
,
attrs_map
,
expected_place
,
stop_gradient
);
})
...
...
@@ -52,6 +53,7 @@ void BindTracer(pybind11::module* m) {
framework
::
AttributeMap
attrs_map
,
const
platform
::
CUDAPlace
expected_place
,
const
bool
stop_gradient
=
false
)
{
pybind11
::
gil_scoped_release
release
;
return
self
.
Trace
(
op
,
inputs
,
outputs
,
attrs_map
,
expected_place
,
stop_gradient
);
})
...
...
paddle/fluid/pybind/pybind.cc
浏览文件 @
8e4ad008
...
...
@@ -36,6 +36,7 @@ limitations under the License. */
#include "paddle/fluid/framework/selected_rows.h"
#include "paddle/fluid/framework/version.h"
#include "paddle/fluid/imperative/layer.h"
#include "paddle/fluid/imperative/profiler.h"
#include "paddle/fluid/memory/allocation/allocator_strategy.h"
#include "paddle/fluid/memory/allocation/legacy_allocator.h"
#include "paddle/fluid/operators/activation_op.h"
...
...
@@ -156,6 +157,11 @@ PYBIND11_MODULE(core, m) {
m
.
def
(
"print_mem_usage"
,
[]()
{
return
memory
::
allocation
::
GPUMemMonitor
.
PrintMemUsage
();
});
m
.
def
(
"start_imperative_gperf_profiler"
,
[]()
{
imperative
::
StartProfile
();
});
m
.
def
(
"stop_imperative_gperf_profiler"
,
[]()
{
imperative
::
StopProfile
();
});
py
::
class_
<
imperative
::
VarBase
>
(
m
,
"VarBase"
,
R"DOC()DOC"
)
.
def
(
py
::
init
<
const
std
::
string
&
,
paddle
::
framework
::
proto
::
VarType
::
Type
,
...
...
python/paddle/fluid/__init__.py
浏览文件 @
8e4ad008
...
...
@@ -132,7 +132,8 @@ def __bootstrap__():
'allocator_strategy'
,
'reader_queue_speed_test_mode'
,
'print_sub_graph_dir'
,
'pe_profile_fname'
,
'warpctc_dir'
,
'inner_op_parallelism'
,
'enable_parallel_graph'
,
'multiple_of_cupti_buffer_size'
,
'enable_subgraph_optimize'
'multiple_of_cupti_buffer_size'
,
'enable_subgraph_optimize'
,
'tracer_profile_fname'
]
if
'Darwin'
not
in
sysstr
:
read_env_flags
.
append
(
'use_pinned_memory'
)
...
...
python/paddle/fluid/imperative/__init__.py
浏览文件 @
8e4ad008
...
...
@@ -26,8 +26,12 @@ from .nn import *
from
.
import
tracer
from
.tracer
import
*
from
.
import
profiler
from
.profiler
import
*
__all__
=
[]
__all__
+=
layers
.
__all__
__all__
+=
base
.
__all__
__all__
+=
nn
.
__all__
__all__
+=
tracer
.
__all__
__all__
+=
profiler
.
__all__
python/paddle/fluid/imperative/profiler.py
0 → 100644
浏览文件 @
8e4ad008
# 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.
from
__future__
import
print_function
from
..
import
core
__all__
=
[
'start_gperf_profiler'
,
'stop_gperf_profiler'
,
]
def
start_gperf_profiler
():
core
.
start_imperative_gperf_profiler
()
def
stop_gperf_profiler
():
core
.
stop_imperative_gperf_profiler
()
tools/manylinux1/build_scripts/build.sh
浏览文件 @
8e4ad008
...
...
@@ -153,3 +153,9 @@ done
# Restore LD_LIBRARY_PATH
LD_LIBRARY_PATH
=
"
${
ORIGINAL_LD_LIBRARY_PATH
}
"
# According to ar issues: https://lists.gnu.org/archive/html/bug-binutils/2016-05/msg00211.html
# we should install new version ar with 64-bit supported here
wget https://ftp.gnu.org/gnu/binutils/binutils-2.27.tar.gz
tar
xzf binutils-2.27.tar.gz
&&
cd
binutils-2.27
./configure
--prefix
=
/opt/rh/devtoolset-2/root/usr/
--enable-64-bit-archive
&&
make
-j
`
nproc
`
&&
make
install
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