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
1dad8cea
P
Paddle
项目概览
PaddlePaddle
/
Paddle
接近 2 年 前同步成功
通知
2323
Star
20933
Fork
5424
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
1dad8cea
编写于
11月 20, 2020
作者:
G
gongweibao
提交者:
GitHub
11月 20, 2020
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Fix gpu memory allocation bug. (#28703)
上级
91bab752
变更
20
隐藏空白更改
内联
并排
Showing
20 changed file
with
80 addition
and
61 deletion
+80
-61
paddle/fluid/framework/data_device_transform_test.cu
paddle/fluid/framework/data_device_transform_test.cu
+1
-1
paddle/fluid/framework/lod_tensor_test.cu
paddle/fluid/framework/lod_tensor_test.cu
+2
-2
paddle/fluid/framework/operator_test.cc
paddle/fluid/framework/operator_test.cc
+10
-10
paddle/fluid/framework/parallel_executor.cc
paddle/fluid/framework/parallel_executor.cc
+44
-0
paddle/fluid/inference/api/analysis_predictor.cc
paddle/fluid/inference/api/analysis_predictor.cc
+1
-1
paddle/fluid/inference/api/api_impl.cc
paddle/fluid/inference/api/api_impl.cc
+1
-1
paddle/fluid/inference/io.cc
paddle/fluid/inference/io.cc
+1
-2
paddle/fluid/memory/allocation/allocator_facade.cc
paddle/fluid/memory/allocation/allocator_facade.cc
+1
-0
paddle/fluid/operators/benchmark/op_tester.cc
paddle/fluid/operators/benchmark/op_tester.cc
+1
-1
paddle/fluid/operators/fused/fusion_group_op_test.cc
paddle/fluid/operators/fused/fusion_group_op_test.cc
+1
-1
paddle/fluid/platform/device_code_test.cc
paddle/fluid/platform/device_code_test.cc
+2
-2
paddle/fluid/platform/init.cc
paddle/fluid/platform/init.cc
+4
-29
paddle/fluid/platform/init.h
paddle/fluid/platform/init.h
+2
-2
paddle/fluid/platform/init_test.cc
paddle/fluid/platform/init_test.cc
+3
-3
paddle/fluid/pybind/pybind.cc
paddle/fluid/pybind/pybind.cc
+1
-1
paddle/fluid/train/demo/demo_trainer.cc
paddle/fluid/train/demo/demo_trainer.cc
+1
-1
paddle/fluid/train/imdb_demo/demo_trainer.cc
paddle/fluid/train/imdb_demo/demo_trainer.cc
+1
-1
paddle/fluid/train/test_train_recognize_digits.cc
paddle/fluid/train/test_train_recognize_digits.cc
+1
-1
paddle/testing/paddle_gtest_main.cc
paddle/testing/paddle_gtest_main.cc
+1
-1
python/paddle/fluid/__init__.py
python/paddle/fluid/__init__.py
+1
-1
未找到文件。
paddle/fluid/framework/data_device_transform_test.cu
浏览文件 @
1dad8cea
...
@@ -103,7 +103,7 @@ static void BuildVar(const std::string& param_name,
...
@@ -103,7 +103,7 @@ static void BuildVar(const std::string& param_name,
}
}
TEST
(
Operator
,
CPUtoGPU
)
{
TEST
(
Operator
,
CPUtoGPU
)
{
paddle
::
framework
::
InitDevices
(
true
);
paddle
::
framework
::
InitDevices
();
paddle
::
framework
::
Scope
scope
;
paddle
::
framework
::
Scope
scope
;
paddle
::
platform
::
CPUPlace
cpu_place
;
paddle
::
platform
::
CPUPlace
cpu_place
;
...
...
paddle/fluid/framework/lod_tensor_test.cu
浏览文件 @
1dad8cea
...
@@ -26,7 +26,7 @@ __global__ void test(size_t* a, int size) {
...
@@ -26,7 +26,7 @@ __global__ void test(size_t* a, int size) {
}
}
TEST
(
LoD
,
data
)
{
TEST
(
LoD
,
data
)
{
paddle
::
framework
::
InitDevices
(
true
);
paddle
::
framework
::
InitDevices
();
paddle
::
framework
::
LoD
lod
{{
0
,
1
,
2
}};
paddle
::
framework
::
LoD
lod
{{
0
,
1
,
2
}};
lod
.
push_back
({
0
,
2
,
4
,
5
});
lod
.
push_back
({
0
,
2
,
4
,
5
});
...
@@ -42,7 +42,7 @@ TEST(LoD, data) {
...
@@ -42,7 +42,7 @@ TEST(LoD, data) {
}
}
TEST
(
LoDTensor
,
LoDInGPU
)
{
TEST
(
LoDTensor
,
LoDInGPU
)
{
paddle
::
framework
::
InitDevices
(
true
);
paddle
::
framework
::
InitDevices
();
paddle
::
framework
::
LoDTensor
lod_tensor
;
paddle
::
framework
::
LoDTensor
lod_tensor
;
paddle
::
platform
::
CUDAPlace
place
(
0
);
paddle
::
platform
::
CUDAPlace
place
(
0
);
...
...
paddle/fluid/framework/operator_test.cc
浏览文件 @
1dad8cea
...
@@ -76,7 +76,7 @@ REGISTER_OP_WITHOUT_GRADIENT(test_operator,
...
@@ -76,7 +76,7 @@ REGISTER_OP_WITHOUT_GRADIENT(test_operator,
paddle
::
framework
::
OpWithoutKernelCheckerMaker
);
paddle
::
framework
::
OpWithoutKernelCheckerMaker
);
TEST
(
OperatorBase
,
all
)
{
TEST
(
OperatorBase
,
all
)
{
paddle
::
framework
::
InitDevices
(
true
);
paddle
::
framework
::
InitDevices
();
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
op_desc
.
set_type
(
"test_operator"
);
op_desc
.
set_type
(
"test_operator"
);
BuildVar
(
"input"
,
{
"IN1"
},
op_desc
.
add_inputs
());
BuildVar
(
"input"
,
{
"IN1"
},
op_desc
.
add_inputs
());
...
@@ -228,7 +228,7 @@ REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(
...
@@ -228,7 +228,7 @@ REGISTER_OP_KERNEL_WITH_CUSTOM_TYPE(
// test with single input
// test with single input
TEST
(
OpKernel
,
all
)
{
TEST
(
OpKernel
,
all
)
{
paddle
::
framework
::
InitDevices
(
true
);
paddle
::
framework
::
InitDevices
();
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
op_desc
.
set_type
(
"op_with_kernel"
);
op_desc
.
set_type
(
"op_with_kernel"
);
BuildVar
(
"x"
,
{
"IN1"
},
op_desc
.
add_inputs
());
BuildVar
(
"x"
,
{
"IN1"
},
op_desc
.
add_inputs
());
...
@@ -268,7 +268,7 @@ REGISTER_OP_CPU_KERNEL(op_multi_inputs_with_kernel,
...
@@ -268,7 +268,7 @@ REGISTER_OP_CPU_KERNEL(op_multi_inputs_with_kernel,
// test with multi inputs
// test with multi inputs
TEST
(
OpKernel
,
multi_inputs
)
{
TEST
(
OpKernel
,
multi_inputs
)
{
paddle
::
framework
::
InitDevices
(
true
);
paddle
::
framework
::
InitDevices
();
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
op_desc
.
set_type
(
"op_multi_inputs_with_kernel"
);
op_desc
.
set_type
(
"op_multi_inputs_with_kernel"
);
...
@@ -419,7 +419,7 @@ REGISTER_OP_CPU_KERNEL(indicate_other_data_type_test,
...
@@ -419,7 +419,7 @@ REGISTER_OP_CPU_KERNEL(indicate_other_data_type_test,
paddle
::
platform
::
CPUDeviceContext
,
int
>
);
paddle
::
platform
::
CPUDeviceContext
,
int
>
);
TEST
(
IndicateVarDataTypeTest
,
lodtensor
)
{
TEST
(
IndicateVarDataTypeTest
,
lodtensor
)
{
paddle
::
framework
::
InitDevices
(
true
);
paddle
::
framework
::
InitDevices
();
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
op_desc
.
set_type
(
"indicate_lod_tensor_data_type_test"
);
op_desc
.
set_type
(
"indicate_lod_tensor_data_type_test"
);
BuildVar
(
"LoDTensor"
,
{
"lodtensor_1"
},
op_desc
.
add_inputs
());
BuildVar
(
"LoDTensor"
,
{
"lodtensor_1"
},
op_desc
.
add_inputs
());
...
@@ -447,7 +447,7 @@ TEST(IndicateVarDataTypeTest, lodtensor) {
...
@@ -447,7 +447,7 @@ TEST(IndicateVarDataTypeTest, lodtensor) {
}
}
TEST
(
IndicateVarDataTypeTest
,
selectedrows
)
{
TEST
(
IndicateVarDataTypeTest
,
selectedrows
)
{
paddle
::
framework
::
InitDevices
(
true
);
paddle
::
framework
::
InitDevices
();
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
op_desc
.
set_type
(
"indicate_selected_rows_data_type_test"
);
op_desc
.
set_type
(
"indicate_selected_rows_data_type_test"
);
BuildVar
(
"SelectedRows"
,
{
"selected_rows_1"
},
op_desc
.
add_inputs
());
BuildVar
(
"SelectedRows"
,
{
"selected_rows_1"
},
op_desc
.
add_inputs
());
...
@@ -474,7 +474,7 @@ TEST(IndicateVarDataTypeTest, selectedrows) {
...
@@ -474,7 +474,7 @@ TEST(IndicateVarDataTypeTest, selectedrows) {
}
}
TEST
(
IndicateVarDataTypeTest
,
other
)
{
TEST
(
IndicateVarDataTypeTest
,
other
)
{
paddle
::
framework
::
InitDevices
(
true
);
paddle
::
framework
::
InitDevices
();
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
op_desc
.
set_type
(
"indicate_other_data_type_test"
);
op_desc
.
set_type
(
"indicate_other_data_type_test"
);
BuildVar
(
"Other"
,
{
"lod_rank_table_1"
},
op_desc
.
add_inputs
());
BuildVar
(
"Other"
,
{
"lod_rank_table_1"
},
op_desc
.
add_inputs
());
...
@@ -504,7 +504,7 @@ TEST(IndicateVarDataTypeTest, other) {
...
@@ -504,7 +504,7 @@ TEST(IndicateVarDataTypeTest, other) {
}
}
TEST
(
ExecutionContextAttrAndInOut
,
new_api
)
{
TEST
(
ExecutionContextAttrAndInOut
,
new_api
)
{
paddle
::
framework
::
InitDevices
(
true
);
paddle
::
framework
::
InitDevices
();
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
op_desc
.
set_type
(
"test_operator"
);
op_desc
.
set_type
(
"test_operator"
);
BuildVar
(
"input"
,
{
"IN1"
},
op_desc
.
add_inputs
());
BuildVar
(
"input"
,
{
"IN1"
},
op_desc
.
add_inputs
());
...
@@ -596,7 +596,7 @@ REGISTER_OP_CPU_KERNEL(set_lod_level_test,
...
@@ -596,7 +596,7 @@ REGISTER_OP_CPU_KERNEL(set_lod_level_test,
paddle
::
platform
::
CPUDeviceContext
,
float
>
);
paddle
::
platform
::
CPUDeviceContext
,
float
>
);
void
SetGetLoDLevelTestMain
(
std
::
string
op_type
)
{
void
SetGetLoDLevelTestMain
(
std
::
string
op_type
)
{
paddle
::
framework
::
InitDevices
(
false
,
{});
paddle
::
framework
::
InitDevices
({});
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
op_desc
.
set_type
(
op_type
);
op_desc
.
set_type
(
op_type
);
BuildVar
(
"X"
,
{
"x.0"
},
op_desc
.
add_inputs
());
BuildVar
(
"X"
,
{
"x.0"
},
op_desc
.
add_inputs
());
...
@@ -701,7 +701,7 @@ REGISTER_OP_CPU_KERNEL(op_without_unused_var,
...
@@ -701,7 +701,7 @@ REGISTER_OP_CPU_KERNEL(op_without_unused_var,
TEST
(
OpWithUnusedVar
,
all
)
{
TEST
(
OpWithUnusedVar
,
all
)
{
// enable the unused_var_check
// enable the unused_var_check
FLAGS_enable_unused_var_check
=
true
;
FLAGS_enable_unused_var_check
=
true
;
paddle
::
framework
::
InitDevices
(
true
);
paddle
::
framework
::
InitDevices
();
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
op_desc
.
set_type
(
"op_with_unused_var"
);
op_desc
.
set_type
(
"op_with_unused_var"
);
BuildVar
(
"X"
,
{
"X"
},
op_desc
.
add_inputs
());
BuildVar
(
"X"
,
{
"X"
},
op_desc
.
add_inputs
());
...
@@ -726,7 +726,7 @@ TEST(OpWithoutUnusedVar, all) {
...
@@ -726,7 +726,7 @@ TEST(OpWithoutUnusedVar, all) {
// enable the unused_var_check
// enable the unused_var_check
FLAGS_enable_unused_var_check
=
true
;
FLAGS_enable_unused_var_check
=
true
;
paddle
::
framework
::
InitDevices
(
true
);
paddle
::
framework
::
InitDevices
();
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
paddle
::
framework
::
proto
::
OpDesc
op_desc
;
op_desc
.
set_type
(
"op_without_unused_var"
);
op_desc
.
set_type
(
"op_without_unused_var"
);
BuildVar
(
"X"
,
{
"X"
},
op_desc
.
add_inputs
());
BuildVar
(
"X"
,
{
"X"
},
op_desc
.
add_inputs
());
...
...
paddle/fluid/framework/parallel_executor.cc
浏览文件 @
1dad8cea
...
@@ -36,6 +36,10 @@ limitations under the License. */
...
@@ -36,6 +36,10 @@ limitations under the License. */
#include "paddle/fluid/platform/event.h"
#include "paddle/fluid/platform/event.h"
#include "paddle/fluid/platform/profiler.h"
#include "paddle/fluid/platform/profiler.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/cuda_device_guard.h"
#endif
DECLARE_double
(
eager_delete_tensor_gb
);
DECLARE_double
(
eager_delete_tensor_gb
);
#ifdef WITH_GPERFTOOLS
#ifdef WITH_GPERFTOOLS
...
@@ -55,6 +59,10 @@ static std::once_flag gProfileOnce;
...
@@ -55,6 +59,10 @@ static std::once_flag gProfileOnce;
static
bool
gProfileStarted
=
false
;
static
bool
gProfileStarted
=
false
;
#endif
#endif
#ifdef PADDLE_WITH_CUDA
std
::
once_flag
p2p_init_flag
;
#endif
class
ParallelExecutorPrivate
{
class
ParallelExecutorPrivate
{
public:
public:
ParallelExecutorPrivate
(
const
std
::
vector
<
platform
::
Place
>
&
places
,
ParallelExecutorPrivate
(
const
std
::
vector
<
platform
::
Place
>
&
places
,
...
@@ -458,6 +466,41 @@ bool ParallelExecutor::NeedCreateLocalExeScope() {
...
@@ -458,6 +466,41 @@ bool ParallelExecutor::NeedCreateLocalExeScope() {
return
executor
&&
executor
->
NeedCreateLocalExeScope
();
return
executor
&&
executor
->
NeedCreateLocalExeScope
();
}
}
void
InitP2P
(
const
std
::
vector
<
platform
::
Place
>
&
places
)
{
#ifdef PADDLE_WITH_CUDA
std
::
call_once
(
p2p_init_flag
,
[
&
]()
{
int
count
=
places
.
size
();
if
(
count
<=
1
)
return
;
std
::
vector
<
int
>
devices
;
for
(
int
i
=
0
;
i
<
count
;
i
++
)
{
if
(
!
is_gpu_place
(
places
[
i
]))
return
;
platform
::
CUDAPlace
device
=
BOOST_GET_CONST
(
platform
::
CUDAPlace
,
places
[
i
]);
devices
.
push_back
(
device
.
GetDeviceId
());
}
for
(
int
i
=
0
;
i
<
count
;
++
i
)
{
for
(
int
j
=
0
;
j
<
count
;
++
j
)
{
if
(
devices
[
i
]
==
devices
[
j
])
continue
;
int
can_acess
=
-
1
;
cudaError_t
ret
=
cudaDeviceCanAccessPeer
(
&
can_acess
,
devices
[
i
],
devices
[
j
]);
if
(
ret
!=
cudaSuccess
||
can_acess
!=
1
)
{
LOG
(
WARNING
)
<<
"Cannot enable P2P access from "
<<
devices
[
i
]
<<
" to "
<<
devices
[
j
];
}
else
{
platform
::
CUDADeviceGuard
guard
(
devices
[
i
]);
cudaDeviceEnablePeerAccess
(
devices
[
j
],
0
);
}
}
}
VLOG
(
1
)
<<
"init p2p"
;
});
#endif
}
ParallelExecutor
::
ParallelExecutor
(
const
std
::
vector
<
platform
::
Place
>
&
places
,
ParallelExecutor
::
ParallelExecutor
(
const
std
::
vector
<
platform
::
Place
>
&
places
,
const
std
::
vector
<
std
::
string
>
&
bcast_vars
,
const
std
::
vector
<
std
::
string
>
&
bcast_vars
,
const
std
::
string
&
loss_var_name
,
const
std
::
string
&
loss_var_name
,
...
@@ -470,6 +513,7 @@ ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
...
@@ -470,6 +513,7 @@ ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
PADDLE_ENFORCE
(
places
.
size
()
>
0
&&
!
is_xpu_place
(
places
[
0
]),
PADDLE_ENFORCE
(
places
.
size
()
>
0
&&
!
is_xpu_place
(
places
[
0
]),
platform
::
errors
::
Unavailable
(
platform
::
errors
::
Unavailable
(
"XPU is not supported in ParallelExecutor"
));
"XPU is not supported in ParallelExecutor"
));
InitP2P
(
places
);
ir
::
InitReaderQueueDeviceCount
(
graph
,
*
(
member_
->
global_scope_
),
ir
::
InitReaderQueueDeviceCount
(
graph
,
*
(
member_
->
global_scope_
),
member_
->
places_
.
size
());
member_
->
places_
.
size
());
member_
->
use_cuda_
=
exec_strategy
.
use_cuda_
;
member_
->
use_cuda_
=
exec_strategy
.
use_cuda_
;
...
...
paddle/fluid/inference/api/analysis_predictor.cc
浏览文件 @
1dad8cea
...
@@ -174,7 +174,7 @@ bool AnalysisPredictor::PrepareScope(
...
@@ -174,7 +174,7 @@ bool AnalysisPredictor::PrepareScope(
scope_
=
parent_scope
;
scope_
=
parent_scope
;
status_is_cloned_
=
true
;
status_is_cloned_
=
true
;
}
else
{
}
else
{
paddle
::
framework
::
InitDevices
(
false
);
paddle
::
framework
::
InitDevices
();
scope_
.
reset
(
new
paddle
::
framework
::
Scope
(),
[](
framework
::
Scope
*
scope
)
{
scope_
.
reset
(
new
paddle
::
framework
::
Scope
(),
[](
framework
::
Scope
*
scope
)
{
delete
scope
;
delete
scope
;
#ifdef PADDLE_WITH_CUDA
#ifdef PADDLE_WITH_CUDA
...
...
paddle/fluid/inference/api/api_impl.cc
浏览文件 @
1dad8cea
...
@@ -91,7 +91,7 @@ bool NativePaddlePredictor::Init(
...
@@ -91,7 +91,7 @@ bool NativePaddlePredictor::Init(
platform
::
errors
::
PreconditionNotMet
(
platform
::
errors
::
PreconditionNotMet
(
"The sub_scope should not be nullptr."
));
"The sub_scope should not be nullptr."
));
}
else
{
}
else
{
paddle
::
framework
::
InitDevices
(
false
);
paddle
::
framework
::
InitDevices
();
scope_
.
reset
(
new
paddle
::
framework
::
Scope
());
scope_
.
reset
(
new
paddle
::
framework
::
Scope
());
}
}
...
...
paddle/fluid/inference/io.cc
浏览文件 @
1dad8cea
...
@@ -25,7 +25,6 @@ limitations under the License. */
...
@@ -25,7 +25,6 @@ limitations under the License. */
#include "paddle/fluid/pybind/pybind.h"
#include "paddle/fluid/pybind/pybind.h"
DEFINE_string
(
devices
,
""
,
"The devices to be used which is joined by comma."
);
DEFINE_string
(
devices
,
""
,
"The devices to be used which is joined by comma."
);
DEFINE_bool
(
init_p2p
,
false
,
"Whether to init p2p."
);
DEFINE_int32
(
math_num_threads
,
1
,
DEFINE_int32
(
math_num_threads
,
1
,
"Number of threads used to run math functions."
);
"Number of threads used to run math functions."
);
...
@@ -42,7 +41,7 @@ void Init(const std::vector<std::string> argv) {
...
@@ -42,7 +41,7 @@ void Init(const std::vector<std::string> argv) {
while
(
std
::
getline
(
tokenStream
,
token
,
','
))
{
while
(
std
::
getline
(
tokenStream
,
token
,
','
))
{
devices
.
push_back
(
std
::
stoi
(
token
));
devices
.
push_back
(
std
::
stoi
(
token
));
}
}
framework
::
InitDevices
(
FLAGS_init_p2p
,
devices
);
framework
::
InitDevices
(
devices
);
}
}
void
ReadBinaryFile
(
const
std
::
string
&
filename
,
std
::
string
*
contents
)
{
void
ReadBinaryFile
(
const
std
::
string
&
filename
,
std
::
string
*
contents
)
{
...
...
paddle/fluid/memory/allocation/allocator_facade.cc
浏览文件 @
1dad8cea
...
@@ -37,6 +37,7 @@
...
@@ -37,6 +37,7 @@
#include "paddle/fluid/memory/allocation/pinned_allocator.h"
#include "paddle/fluid/memory/allocation/pinned_allocator.h"
#include "paddle/fluid/memory/allocation/thread_local_allocator.h"
#include "paddle/fluid/memory/allocation/thread_local_allocator.h"
#include "paddle/fluid/platform/cuda_device_guard.h"
#include "paddle/fluid/platform/cuda_device_guard.h"
#include "paddle/fluid/platform/dynload/cupti.h"
#include "paddle/fluid/platform/gpu_info.h"
#include "paddle/fluid/platform/gpu_info.h"
#endif
#endif
#ifdef PADDLE_WITH_XPU
#ifdef PADDLE_WITH_XPU
...
...
paddle/fluid/operators/benchmark/op_tester.cc
浏览文件 @
1dad8cea
...
@@ -57,7 +57,7 @@ void OpTester::Init(const OpTesterConfig &config) {
...
@@ -57,7 +57,7 @@ void OpTester::Init(const OpTesterConfig &config) {
place_
=
paddle
::
platform
::
CPUPlace
();
place_
=
paddle
::
platform
::
CPUPlace
();
}
}
framework
::
InitDevices
(
false
);
framework
::
InitDevices
();
scope_
.
reset
(
new
paddle
::
framework
::
Scope
());
scope_
.
reset
(
new
paddle
::
framework
::
Scope
());
op_
=
framework
::
OpRegistry
::
CreateOp
(
op_desc_
);
op_
=
framework
::
OpRegistry
::
CreateOp
(
op_desc_
);
...
...
paddle/fluid/operators/fused/fusion_group_op_test.cc
浏览文件 @
1dad8cea
...
@@ -140,7 +140,7 @@ void TestMain(const std::vector<std::string>& input_names,
...
@@ -140,7 +140,7 @@ void TestMain(const std::vector<std::string>& input_names,
std
::
string
func_name
,
std
::
string
cuda_kernel_str
,
std
::
string
func_name
,
std
::
string
cuda_kernel_str
,
CPUKernelFunc
cpu_kernel_func
)
{
CPUKernelFunc
cpu_kernel_func
)
{
// Compile the device code
// Compile the device code
paddle
::
framework
::
InitDevices
(
false
,
{
0
});
paddle
::
framework
::
InitDevices
({
0
});
platform
::
CUDAPlace
place
=
platform
::
CUDAPlace
(
0
);
platform
::
CUDAPlace
place
=
platform
::
CUDAPlace
(
0
);
PrepareDeviceCode
(
place
,
func_name
,
cuda_kernel_str
);
PrepareDeviceCode
(
place
,
func_name
,
cuda_kernel_str
);
...
...
paddle/fluid/platform/device_code_test.cc
浏览文件 @
1dad8cea
...
@@ -35,7 +35,7 @@ TEST(DeviceCode, cuda) {
...
@@ -35,7 +35,7 @@ TEST(DeviceCode, cuda) {
return
;
return
;
}
}
paddle
::
framework
::
InitDevices
(
false
,
{
0
});
paddle
::
framework
::
InitDevices
({
0
});
paddle
::
platform
::
CUDAPlace
place
=
paddle
::
platform
::
CUDAPlace
(
0
);
paddle
::
platform
::
CUDAPlace
place
=
paddle
::
platform
::
CUDAPlace
(
0
);
paddle
::
platform
::
CUDADeviceCode
code
(
place
,
"saxpy_kernel"
,
saxpy_code
);
paddle
::
platform
::
CUDADeviceCode
code
(
place
,
"saxpy_kernel"
,
saxpy_code
);
...
@@ -90,7 +90,7 @@ TEST(DeviceCodePool, cuda) {
...
@@ -90,7 +90,7 @@ TEST(DeviceCodePool, cuda) {
return
;
return
;
}
}
paddle
::
framework
::
InitDevices
(
false
,
{
0
});
paddle
::
framework
::
InitDevices
({
0
});
paddle
::
platform
::
CUDAPlace
place
=
paddle
::
platform
::
CUDAPlace
(
0
);
paddle
::
platform
::
CUDAPlace
place
=
paddle
::
platform
::
CUDAPlace
(
0
);
paddle
::
platform
::
DeviceCodePool
&
pool
=
paddle
::
platform
::
DeviceCodePool
&
pool
=
paddle
::
platform
::
DeviceCodePool
::
Init
({
place
});
paddle
::
platform
::
DeviceCodePool
::
Init
({
place
});
...
...
paddle/fluid/platform/init.cc
浏览文件 @
1dad8cea
...
@@ -63,7 +63,6 @@ namespace framework {
...
@@ -63,7 +63,6 @@ namespace framework {
std
::
once_flag
gflags_init_flag
;
std
::
once_flag
gflags_init_flag
;
std
::
once_flag
glog_init_flag
;
std
::
once_flag
glog_init_flag
;
std
::
once_flag
p2p_init_flag
;
bool
InitGflags
(
std
::
vector
<
std
::
string
>
args
)
{
bool
InitGflags
(
std
::
vector
<
std
::
string
>
args
)
{
bool
successed
=
false
;
bool
successed
=
false
;
...
@@ -95,28 +94,7 @@ bool InitGflags(std::vector<std::string> args) {
...
@@ -95,28 +94,7 @@ bool InitGflags(std::vector<std::string> args) {
return
successed
;
return
successed
;
}
}
void
InitP2P
(
std
::
vector
<
int
>
devices
)
{
#ifdef PADDLE_WITH_CUDA
std
::
call_once
(
p2p_init_flag
,
[
&
]()
{
int
count
=
devices
.
size
();
for
(
int
i
=
0
;
i
<
count
;
++
i
)
{
for
(
int
j
=
0
;
j
<
count
;
++
j
)
{
if
(
devices
[
i
]
==
devices
[
j
])
continue
;
int
can_acess
=
-
1
;
PADDLE_ENFORCE_CUDA_SUCCESS
(
cudaDeviceCanAccessPeer
(
&
can_acess
,
devices
[
i
],
devices
[
j
]));
if
(
can_acess
!=
1
)
{
VLOG
(
2
)
<<
"Cannot enable P2P access from "
<<
devices
[
i
]
<<
" to "
<<
devices
[
j
];
}
else
{
platform
::
CUDADeviceGuard
guard
(
devices
[
i
]);
cudaDeviceEnablePeerAccess
(
devices
[
j
],
0
);
}
}
}
});
#endif
}
void
InitCupti
()
{
void
InitCupti
()
{
#ifdef PADDLE_WITH_CUPTI
#ifdef PADDLE_WITH_CUPTI
...
@@ -144,7 +122,7 @@ void InitCupti() {
...
@@ -144,7 +122,7 @@ void InitCupti() {
#endif
#endif
}
}
void
InitDevices
(
bool
init_p2p
)
{
void
InitDevices
()
{
// CUPTI attribute should be set before any CUDA context is created (see CUPTI
// CUPTI attribute should be set before any CUDA context is created (see CUPTI
// documentation about CUpti_ActivityAttribute).
// documentation about CUpti_ActivityAttribute).
InitCupti
();
InitCupti
();
...
@@ -166,10 +144,10 @@ void InitDevices(bool init_p2p) {
...
@@ -166,10 +144,10 @@ void InitDevices(bool init_p2p) {
LOG
(
WARNING
)
<<
"Compiled with WITH_XPU, but no XPU found in runtime."
;
LOG
(
WARNING
)
<<
"Compiled with WITH_XPU, but no XPU found in runtime."
;
}
}
#endif
#endif
InitDevices
(
init_p2p
,
devices
);
InitDevices
(
devices
);
}
}
void
InitDevices
(
bool
init_p2p
,
const
std
::
vector
<
int
>
devices
)
{
void
InitDevices
(
const
std
::
vector
<
int
>
devices
)
{
std
::
vector
<
platform
::
Place
>
places
;
std
::
vector
<
platform
::
Place
>
places
;
for
(
size_t
i
=
0
;
i
<
devices
.
size
();
++
i
)
{
for
(
size_t
i
=
0
;
i
<
devices
.
size
();
++
i
)
{
...
@@ -187,9 +165,6 @@ void InitDevices(bool init_p2p, const std::vector<int> devices) {
...
@@ -187,9 +165,6 @@ void InitDevices(bool init_p2p, const std::vector<int> devices) {
places
.
emplace_back
(
platform
::
XPUPlace
(
devices
[
i
]));
places
.
emplace_back
(
platform
::
XPUPlace
(
devices
[
i
]));
#endif
#endif
}
}
if
(
init_p2p
)
{
InitP2P
(
devices
);
}
places
.
emplace_back
(
platform
::
CPUPlace
());
places
.
emplace_back
(
platform
::
CPUPlace
());
#ifdef PADDLE_WITH_CUDA
#ifdef PADDLE_WITH_CUDA
places
.
emplace_back
(
platform
::
CUDAPinnedPlace
());
places
.
emplace_back
(
platform
::
CUDAPinnedPlace
());
...
...
paddle/fluid/platform/init.h
浏览文件 @
1dad8cea
...
@@ -35,9 +35,9 @@ bool InitGflags(std::vector<std::string> argv);
...
@@ -35,9 +35,9 @@ bool InitGflags(std::vector<std::string> argv);
void
InitGLOG
(
const
std
::
string
&
prog_name
);
void
InitGLOG
(
const
std
::
string
&
prog_name
);
void
InitDevices
(
bool
init_p2p
);
void
InitDevices
();
void
InitDevices
(
bool
init_p2p
,
const
std
::
vector
<
int
>
devices
);
void
InitDevices
(
const
std
::
vector
<
int
>
devices
);
#ifndef _WIN32
#ifndef _WIN32
class
SignalMessageDumper
{
class
SignalMessageDumper
{
...
...
paddle/fluid/platform/init_test.cc
浏览文件 @
1dad8cea
...
@@ -22,7 +22,7 @@ TEST(InitDevices, CPU) {
...
@@ -22,7 +22,7 @@ TEST(InitDevices, CPU) {
using
paddle
::
platform
::
DeviceContextPool
;
using
paddle
::
platform
::
DeviceContextPool
;
#if !defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_XPU)
#if !defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_XPU)
InitDevices
(
true
);
InitDevices
();
DeviceContextPool
&
pool
=
DeviceContextPool
::
Instance
();
DeviceContextPool
&
pool
=
DeviceContextPool
::
Instance
();
ASSERT_EQ
(
pool
.
size
(),
1U
);
ASSERT_EQ
(
pool
.
size
(),
1U
);
#endif
#endif
...
@@ -34,7 +34,7 @@ TEST(InitDevices, CUDA) {
...
@@ -34,7 +34,7 @@ TEST(InitDevices, CUDA) {
#ifdef PADDLE_WITH_CUDA
#ifdef PADDLE_WITH_CUDA
int
count
=
paddle
::
platform
::
GetCUDADeviceCount
();
int
count
=
paddle
::
platform
::
GetCUDADeviceCount
();
InitDevices
(
true
);
InitDevices
();
DeviceContextPool
&
pool
=
DeviceContextPool
::
Instance
();
DeviceContextPool
&
pool
=
DeviceContextPool
::
Instance
();
ASSERT_EQ
(
pool
.
size
(),
2U
+
static_cast
<
unsigned
>
(
count
));
ASSERT_EQ
(
pool
.
size
(),
2U
+
static_cast
<
unsigned
>
(
count
));
#endif
#endif
...
@@ -46,7 +46,7 @@ TEST(InitDevices, XPU) {
...
@@ -46,7 +46,7 @@ TEST(InitDevices, XPU) {
#ifdef PADDLE_WITH_XPU
#ifdef PADDLE_WITH_XPU
int
count
=
paddle
::
platform
::
GetXPUDeviceCount
();
int
count
=
paddle
::
platform
::
GetXPUDeviceCount
();
InitDevices
(
true
);
InitDevices
();
DeviceContextPool
&
pool
=
DeviceContextPool
::
Instance
();
DeviceContextPool
&
pool
=
DeviceContextPool
::
Instance
();
ASSERT_EQ
(
pool
.
size
(),
1U
+
static_cast
<
unsigned
>
(
count
));
ASSERT_EQ
(
pool
.
size
(),
1U
+
static_cast
<
unsigned
>
(
count
));
#endif
#endif
...
...
paddle/fluid/pybind/pybind.cc
浏览文件 @
1dad8cea
...
@@ -1715,7 +1715,7 @@ All parameter, weight, gradient are variables in Paddle.
...
@@ -1715,7 +1715,7 @@ All parameter, weight, gradient are variables in Paddle.
m
.
def
(
"init_glog"
,
framework
::
InitGLOG
);
m
.
def
(
"init_glog"
,
framework
::
InitGLOG
);
m
.
def
(
"load_op_library"
,
framework
::
LoadOpLib
);
m
.
def
(
"load_op_library"
,
framework
::
LoadOpLib
);
m
.
def
(
"init_devices"
,
m
.
def
(
"init_devices"
,
[](
bool
init_p2p
)
{
framework
::
InitDevices
(
init_p2p
);
});
[](
)
{
framework
::
InitDevices
(
);
});
m
.
def
(
"is_compiled_with_cuda"
,
IsCompiledWithCUDA
);
m
.
def
(
"is_compiled_with_cuda"
,
IsCompiledWithCUDA
);
m
.
def
(
"is_compiled_with_xpu"
,
IsCompiledWithXPU
);
m
.
def
(
"is_compiled_with_xpu"
,
IsCompiledWithXPU
);
...
...
paddle/fluid/train/demo/demo_trainer.cc
浏览文件 @
1dad8cea
...
@@ -55,7 +55,7 @@ std::unique_ptr<paddle::framework::ProgramDesc> Load(
...
@@ -55,7 +55,7 @@ std::unique_ptr<paddle::framework::ProgramDesc> Load(
}
// namespace paddle
}
// namespace paddle
int
main
()
{
int
main
()
{
paddle
::
framework
::
InitDevices
(
false
);
paddle
::
framework
::
InitDevices
();
const
auto
cpu_place
=
paddle
::
platform
::
CPUPlace
();
const
auto
cpu_place
=
paddle
::
platform
::
CPUPlace
();
...
...
paddle/fluid/train/imdb_demo/demo_trainer.cc
浏览文件 @
1dad8cea
...
@@ -105,7 +105,7 @@ int main(int argc, char* argv[]) {
...
@@ -105,7 +105,7 @@ int main(int argc, char* argv[]) {
platform
::
errors
::
InvalidArgument
(
platform
::
errors
::
InvalidArgument
(
"At least one file to train, but received number of file is %d."
,
"At least one file to train, but received number of file is %d."
,
file_vec
.
size
()));
file_vec
.
size
()));
paddle
::
framework
::
InitDevices
(
false
);
paddle
::
framework
::
InitDevices
();
const
auto
cpu_place
=
paddle
::
platform
::
CPUPlace
();
const
auto
cpu_place
=
paddle
::
platform
::
CPUPlace
();
paddle
::
framework
::
Executor
executor
(
cpu_place
);
paddle
::
framework
::
Executor
executor
(
cpu_place
);
paddle
::
framework
::
Scope
scope
;
paddle
::
framework
::
Scope
scope
;
...
...
paddle/fluid/train/test_train_recognize_digits.cc
浏览文件 @
1dad8cea
...
@@ -33,7 +33,7 @@ DEFINE_string(dirname, "", "Directory of the train model.");
...
@@ -33,7 +33,7 @@ DEFINE_string(dirname, "", "Directory of the train model.");
namespace
paddle
{
namespace
paddle
{
void
Train
(
std
::
string
model_dir
)
{
void
Train
(
std
::
string
model_dir
)
{
framework
::
InitDevices
(
false
);
framework
::
InitDevices
();
const
auto
cpu_place
=
platform
::
CPUPlace
();
const
auto
cpu_place
=
platform
::
CPUPlace
();
framework
::
Executor
executor
(
cpu_place
);
framework
::
Executor
executor
(
cpu_place
);
framework
::
Scope
scope
;
framework
::
Scope
scope
;
...
...
paddle/testing/paddle_gtest_main.cc
浏览文件 @
1dad8cea
...
@@ -121,7 +121,7 @@ int main(int argc, char** argv) {
...
@@ -121,7 +121,7 @@ int main(int argc, char** argv) {
int
internal_argc
=
internal_argv
.
size
();
int
internal_argc
=
internal_argv
.
size
();
char
**
arr
=
internal_argv
.
data
();
char
**
arr
=
internal_argv
.
data
();
paddle
::
platform
::
ParseCommandLineFlags
(
internal_argc
,
arr
,
true
);
paddle
::
platform
::
ParseCommandLineFlags
(
internal_argc
,
arr
,
true
);
paddle
::
framework
::
InitDevices
(
true
);
paddle
::
framework
::
InitDevices
();
int
ret
=
RUN_ALL_TESTS
();
int
ret
=
RUN_ALL_TESTS
();
...
...
python/paddle/fluid/__init__.py
浏览文件 @
1dad8cea
...
@@ -254,7 +254,7 @@ def __bootstrap__():
...
@@ -254,7 +254,7 @@ def __bootstrap__():
core
.
init_gflags
([
"--tryfromenv="
+
","
.
join
(
read_env_flags
)])
core
.
init_gflags
([
"--tryfromenv="
+
","
.
join
(
read_env_flags
)])
core
.
init_glog
(
sys
.
argv
[
0
])
core
.
init_glog
(
sys
.
argv
[
0
])
# don't init_p2p when in unittest to save time.
# don't init_p2p when in unittest to save time.
core
.
init_devices
(
not
in_test
)
core
.
init_devices
()
# TODO(panyx0718): Avoid doing complex initialization logic in __init__.py.
# TODO(panyx0718): Avoid doing complex initialization logic in __init__.py.
...
...
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