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
88a607c3
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
未验证
提交
88a607c3
编写于
8月 06, 2018
作者:
Q
Qiao Longfei
提交者:
GitHub
8月 06, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #12541 from jacquesqiao/optimize-profiler
optimize profiler
上级
0fd2f713
954d680b
变更
9
显示空白变更内容
内联
并排
Showing
9 changed file
with
8 addition
and
26 deletion
+8
-26
paddle/fluid/framework/operator.cc
paddle/fluid/framework/operator.cc
+2
-3
paddle/fluid/operators/feed_op.cc
paddle/fluid/operators/feed_op.cc
+0
-1
paddle/fluid/operators/fetch_barrier_op.cc
paddle/fluid/operators/fetch_barrier_op.cc
+0
-6
paddle/fluid/operators/fetch_op.cc
paddle/fluid/operators/fetch_op.cc
+0
-3
paddle/fluid/operators/load_op.cc
paddle/fluid/operators/load_op.cc
+0
-3
paddle/fluid/operators/recv_op.cc
paddle/fluid/operators/recv_op.cc
+0
-2
paddle/fluid/operators/send_barrier_op.cc
paddle/fluid/operators/send_barrier_op.cc
+0
-5
paddle/fluid/operators/send_op.cc
paddle/fluid/operators/send_op.cc
+0
-3
paddle/fluid/platform/profiler.cc
paddle/fluid/platform/profiler.cc
+6
-0
未找到文件。
paddle/fluid/framework/operator.cc
浏览文件 @
88a607c3
...
...
@@ -136,6 +136,8 @@ void OperatorBase::Run(const Scope& scope, const platform::Place& place) {
platform
::
SetDeviceId
(
dev_id
);
#endif
}
platform
::
DeviceContextPool
&
pool
=
platform
::
DeviceContextPool
::
Instance
();
platform
::
RecordEvent
record_event
(
Type
(),
pool
.
Get
(
place
));
RunImpl
(
scope
,
place
);
VLOG
(
10
)
<<
"+ "
<<
DebugStringEx
(
&
scope
);
}
...
...
@@ -639,9 +641,6 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
platform
::
DeviceContextPool
&
pool
=
platform
::
DeviceContextPool
::
Instance
();
auto
*
dev_ctx
=
pool
.
Get
(
place
);
// For profiling, don't move out of this function because that will result
// in the failure of multi-GPU profiling.
platform
::
RecordEvent
record_event
(
Type
(),
dev_ctx
);
// check if op[type] has kernel registered.
auto
&
all_op_kernels
=
AllOpKernels
();
auto
kernels_iter
=
all_op_kernels
.
find
(
type_
);
...
...
paddle/fluid/operators/feed_op.cc
浏览文件 @
88a607c3
...
...
@@ -31,7 +31,6 @@ class FeedOp : public framework::OperatorBase {
const
platform
::
Place
&
place
)
const
override
{
// get device context from pool
auto
*
dev_ctx
=
platform
::
DeviceContextPool
::
Instance
().
Get
(
place
);
platform
::
RecordEvent
record_event
(
Type
(),
dev_ctx
);
auto
feed_var_name
=
Input
(
"X"
);
auto
*
feed_var
=
scope
.
FindVar
(
feed_var_name
);
...
...
paddle/fluid/operators/fetch_barrier_op.cc
浏览文件 @
88a607c3
...
...
@@ -36,12 +36,6 @@ class FetchBarrierOp : public framework::OperatorBase {
void
RunImpl
(
const
framework
::
Scope
&
scope
,
const
platform
::
Place
&
place
)
const
override
{
std
::
vector
<
std
::
string
>
eps
=
Attr
<
std
::
vector
<
std
::
string
>>
(
"endpoints"
);
platform
::
DeviceContextPool
&
pool
=
platform
::
DeviceContextPool
::
Instance
();
auto
&
ctx
=
*
pool
.
Get
(
place
);
// For profiling
platform
::
RecordEvent
record_event
(
Type
(),
&
ctx
);
distributed
::
RPCClient
*
rpc_client
=
distributed
::
RPCClient
::
GetInstance
<
RPCCLIENT_T
>
();
...
...
paddle/fluid/operators/fetch_op.cc
浏览文件 @
88a607c3
...
...
@@ -30,9 +30,6 @@ class FetchOp : public framework::OperatorBase {
private:
void
RunImpl
(
const
framework
::
Scope
&
scope
,
const
platform
::
Place
&
place
)
const
override
{
platform
::
DeviceContextPool
&
pool
=
platform
::
DeviceContextPool
::
Instance
();
platform
::
RecordEvent
record_event
(
Type
(),
pool
.
Get
(
place
));
auto
fetch_var_name
=
Input
(
"X"
);
auto
*
fetch_var
=
scope
.
FindVar
(
fetch_var_name
);
PADDLE_ENFORCE
(
fetch_var
!=
nullptr
,
...
...
paddle/fluid/operators/load_op.cc
浏览文件 @
88a607c3
...
...
@@ -31,9 +31,6 @@ class LoadOp : public framework::OperatorBase {
private:
void
RunImpl
(
const
framework
::
Scope
&
scope
,
const
platform
::
Place
&
place
)
const
override
{
auto
*
dev_ctx
=
platform
::
DeviceContextPool
::
Instance
().
Get
(
place
);
platform
::
RecordEvent
record_event
(
Type
(),
dev_ctx
);
// FIXME(yuyang18): We save variable to local file now, but we should change
// it to save an output stream.
auto
filename
=
Attr
<
std
::
string
>
(
"file_path"
);
...
...
paddle/fluid/operators/recv_op.cc
浏览文件 @
88a607c3
...
...
@@ -40,8 +40,6 @@ class RecvOp : public framework::OperatorBase {
platform
::
DeviceContextPool
&
pool
=
platform
::
DeviceContextPool
::
Instance
();
auto
&
ctx
=
*
pool
.
Get
(
place
);
// For profiling
platform
::
RecordEvent
record_event
(
Type
(),
&
ctx
);
distributed
::
RPCClient
*
rpc_client
=
distributed
::
RPCClient
::
GetInstance
<
RPCCLIENT_T
>
();
...
...
paddle/fluid/operators/send_barrier_op.cc
浏览文件 @
88a607c3
...
...
@@ -39,11 +39,6 @@ class SendBarrierOp : public framework::OperatorBase {
std
::
vector
<
std
::
string
>
eps
=
Attr
<
std
::
vector
<
std
::
string
>>
(
"endpoints"
);
bool
sync_mode
=
Attr
<
bool
>
(
"sync_mode"
);
platform
::
DeviceContextPool
&
pool
=
platform
::
DeviceContextPool
::
Instance
();
auto
&
ctx
=
*
pool
.
Get
(
place
);
// For profiling
platform
::
RecordEvent
record_event
(
Type
(),
&
ctx
);
distributed
::
RPCClient
*
rpc_client
=
distributed
::
RPCClient
::
GetInstance
<
RPCCLIENT_T
>
();
...
...
paddle/fluid/operators/send_op.cc
浏览文件 @
88a607c3
...
...
@@ -42,9 +42,6 @@ class SendOp : public framework::OperatorBase {
platform
::
DeviceContextPool
&
pool
=
platform
::
DeviceContextPool
::
Instance
();
auto
&
ctx
=
*
pool
.
Get
(
place
);
// For profiling
platform
::
RecordEvent
record_event
(
Type
(),
&
ctx
);
distributed
::
RPCClient
*
rpc_client
=
distributed
::
RPCClient
::
GetInstance
<
RPCCLIENT_T
>
();
...
...
paddle/fluid/platform/profiler.cc
浏览文件 @
88a607c3
...
...
@@ -110,6 +110,8 @@ Event::Event(EventType type, std::string name, uint32_t thread_id,
has_cuda_
=
dev_ctx
?
platform
::
is_gpu_place
(
dev_ctx
->
GetPlace
())
:
false
;
if
(
has_cuda_
)
{
auto
*
cuda_dev_ctx
=
static_cast
<
const
CUDADeviceContext
*>
(
dev_ctx
);
PADDLE_ENFORCE
(
cudaSetDevice
(
boost
::
get
<
platform
::
CUDAPlace
>
(
cuda_dev_ctx
->
GetPlace
()).
device
));
PADDLE_ENFORCE
(
cudaGetDevice
(
&
device_
));
PADDLE_ENFORCE
(
cudaEventCreate
(
&
event_
));
auto
stream
=
cuda_dev_ctx
->
stream
();
...
...
@@ -176,6 +178,7 @@ void PopEvent(const std::string& name, const DeviceContext* dev_ctx) {
RecordEvent
::
RecordEvent
(
const
std
::
string
&
name
,
const
DeviceContext
*
dev_ctx
)
:
is_enabled_
(
false
),
start_ns_
(
PosixInNsec
())
{
std
::
lock_guard
<
std
::
mutex
>
l
(
profiler_mu
);
if
(
g_state
==
ProfilerState
::
kDisabled
)
return
;
is_enabled_
=
true
;
dev_ctx_
=
dev_ctx
;
...
...
@@ -186,6 +189,7 @@ RecordEvent::RecordEvent(const std::string& name, const DeviceContext* dev_ctx)
}
RecordEvent
::~
RecordEvent
()
{
std
::
lock_guard
<
std
::
mutex
>
l
(
profiler_mu
);
if
(
g_state
==
ProfilerState
::
kDisabled
||
!
is_enabled_
)
return
;
DeviceTracer
*
tracer
=
GetDeviceTracer
();
if
(
tracer
)
{
...
...
@@ -198,6 +202,7 @@ RecordEvent::~RecordEvent() {
RecordBlock
::
RecordBlock
(
int
block_id
)
:
is_enabled_
(
false
),
start_ns_
(
PosixInNsec
())
{
std
::
lock_guard
<
std
::
mutex
>
l
(
profiler_mu
);
if
(
g_state
==
ProfilerState
::
kDisabled
)
return
;
is_enabled_
=
true
;
SetCurBlock
(
block_id
);
...
...
@@ -205,6 +210,7 @@ RecordBlock::RecordBlock(int block_id)
}
RecordBlock
::~
RecordBlock
()
{
std
::
lock_guard
<
std
::
mutex
>
l
(
profiler_mu
);
if
(
g_state
==
ProfilerState
::
kDisabled
||
!
is_enabled_
)
return
;
DeviceTracer
*
tracer
=
GetDeviceTracer
();
if
(
tracer
)
{
...
...
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