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
兽拳
Paddle
提交
9141bee1
P
Paddle
项目概览
兽拳
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
1
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1
列表
看板
标记
里程碑
合并请求
0
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
Issue
1
列表
看板
标记
里程碑
合并请求
0
合并请求
0
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
9141bee1
编写于
6月 07, 2018
作者:
C
cuichaowen
提交者:
Yan Chunwei
6月 07, 2018
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add Anakin api for paddle (#11228)
上级
d48172f2
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
123 addition
and
38 deletion
+123
-38
paddle/contrib/inference/CMakeLists.txt
paddle/contrib/inference/CMakeLists.txt
+23
-17
paddle/contrib/inference/paddle_inference_api.h
paddle/contrib/inference/paddle_inference_api.h
+0
-1
paddle/contrib/inference/paddle_inference_api_anakin_engine.cc
...e/contrib/inference/paddle_inference_api_anakin_engine.cc
+43
-11
paddle/contrib/inference/paddle_inference_api_anakin_engine.h
...le/contrib/inference/paddle_inference_api_anakin_engine.h
+16
-6
paddle/contrib/inference/paddle_inference_api_anakin_engine_tester.cc
...ib/inference/paddle_inference_api_anakin_engine_tester.cc
+41
-3
未找到文件。
paddle/contrib/inference/CMakeLists.txt
浏览文件 @
9141bee1
...
@@ -24,31 +24,37 @@ set(ANAKIN_LIBRARY "" CACHE STRING "path of Anakin library")
...
@@ -24,31 +24,37 @@ set(ANAKIN_LIBRARY "" CACHE STRING "path of Anakin library")
set
(
inference_deps paddle_inference_api paddle_fluid_api
)
set
(
inference_deps paddle_inference_api paddle_fluid_api
)
# if anakin is set enable anakin api implementation
# if anakin is set enable anakin api implementation
if
(
ANAKIN_INCLUDE
_DIR
AND ANAKIN_LIBRARY
)
if
(
ANAKIN_INCLUDE AND ANAKIN_LIBRARY
)
set
(
ANAKIN_FOUND ON
)
set
(
ANAKIN_FOUND ON
)
else
()
else
()
set
(
ANAKIN_FOUND OFF
)
set
(
ANAKIN_FOUND OFF
)
endif
()
endif
()
function
(
fetch_include_recursively root_dir
)
if
(
IS_DIRECTORY
${
root_dir
}
)
include_directories
(
${
root_dir
}
)
endif
()
file
(
GLOB ALL_SUB RELATIVE
${
root_dir
}
${
root_dir
}
/*
)
foreach
(
sub
${
ALL_SUB
}
)
if
(
IS_DIRECTORY
${
root_dir
}
/
${
sub
}
)
fetch_include_recursively
(
${
root_dir
}
/
${
sub
}
)
endif
()
endforeach
()
endfunction
()
if
(
ANAKIN_FOUND
)
if
(
ANAKIN_FOUND
)
# Anakin's code style doesn't follow google c style.
# Anakin's code style doesn't follow google c style.
set
(
CMAKE_CXX_FLAGS
"
${
CMAKE_CXX_FLAGS
}
-Wno-error=comment
set
(
CMAKE_CXX_FLAGS
"
${
CMAKE_CXX_FLAGS
}
-Wno-error=unused-variable -Wno-error=format-extra-args -Wno-error=comment -Wno-error=format -Wno-error=switch -Wno-error=return-type -Wno-error=non-virtual-dtor -Wno-reorder -Wno-error=cpp"
)
-Wno-error=reorder
-Wno-error=format
-Wno-error=switch
-Wno-error=return-type
-Wno-error=non-virtual-dtor
-Wno-error=cpp"
)
message
(
STATUS
"Anakin for inference is enabled"
)
message
(
STATUS
"Anakin for inference is enabled"
)
message
(
STATUS
"Anakin is set INCLUDE:
${
ANAKIN_INCLUDE
}
LIBRARY:
${
ANAKIN_LIBRARY
}
"
)
message
(
STATUS
"Anakin is set INCLUDE:
${
ANAKIN_INCLUDE
}
LIBRARY:
${
ANAKIN_LIBRARY
}
"
)
include_directories
(
"
${
ANAKIN_INCLUDE
}
"
)
fetch_include_recursively
(
${
ANAKIN_INCLUDE
}
)
# Anakin's source path is a mass, need to set sub-directories trivially.
include_directories
(
"
${
ANAKIN_INCLUDE
}
/saber"
)
link_directories
(
${
ANAKIN_LIBRARY
}
)
link_directories
(
"
${
ANAKIN_LIBRARY
}
"
)
nv_library
(
inference_anakin_api S
RCS
paddle_inference_api_anakin_engine.cc
)
nv_library
(
inference_anakin_api S
HARED SRCS paddle_inference_api.cc
paddle_inference_api_anakin_engine.cc
)
target_link_libraries
(
inference_anakin_api anakin
)
target_link_libraries
(
inference_anakin_api anakin
anakin_saber_common
)
list
(
APPEND inference_deps inference_anakin_api
)
list
(
APPEND inference_deps inference_anakin_api
)
endif
()
endif
()
...
@@ -73,7 +79,7 @@ function(inference_api_test TARGET_NAME)
...
@@ -73,7 +79,7 @@ function(inference_api_test TARGET_NAME)
endfunction
(
inference_api_test
)
endfunction
(
inference_api_test
)
cc_library
(
paddle_inference_api
cc_library
(
paddle_inference_api
SRCS paddle_inference_api.cc paddle_inference_api_impl.cc
SRCS paddle_inference_api.cc paddle_inference_api_impl.cc
DEPS
${
FLUID_CORE_MODULES
}
${
GLOB_OP_LIB
}
)
DEPS
${
FLUID_CORE_MODULES
}
${
GLOB_OP_LIB
}
)
cc_test
(
test_paddle_inference_api
cc_test
(
test_paddle_inference_api
...
@@ -84,8 +90,8 @@ inference_api_test(test_paddle_inference_api_impl
...
@@ -84,8 +90,8 @@ inference_api_test(test_paddle_inference_api_impl
ARGS test_word2vec test_image_classification
)
ARGS test_word2vec test_image_classification
)
if
(
ANAKIN_FOUND
)
if
(
ANAKIN_FOUND
)
nv
_test
(
inference_anakin_test SRCS paddle_inference_api_anakin_engine_tester.cc
cc
_test
(
inference_anakin_test SRCS paddle_inference_api_anakin_engine_tester.cc
DEPS
${
inference_deps
}
protobuf
)
DEPS
${
inference_deps
}
)
endif
()
endif
()
if
(
WITH_TESTING
)
if
(
WITH_TESTING
)
...
...
paddle/contrib/inference/paddle_inference_api.h
浏览文件 @
9141bee1
...
@@ -113,5 +113,4 @@ struct AnakinConfig : public PaddlePredictor::Config {
...
@@ -113,5 +113,4 @@ struct AnakinConfig : public PaddlePredictor::Config {
// Similarly, each engine kind should map to a unique predictor implementation.
// Similarly, each engine kind should map to a unique predictor implementation.
template
<
typename
ConfigT
,
PaddleEngineKind
engine
=
PaddleEngineKind
::
kNative
>
template
<
typename
ConfigT
,
PaddleEngineKind
engine
=
PaddleEngineKind
::
kNative
>
std
::
unique_ptr
<
PaddlePredictor
>
CreatePaddlePredictor
(
const
ConfigT
&
config
);
std
::
unique_ptr
<
PaddlePredictor
>
CreatePaddlePredictor
(
const
ConfigT
&
config
);
}
// namespace paddle
}
// namespace paddle
paddle/contrib/inference/paddle_inference_api_anakin_engine.cc
浏览文件 @
9141bee1
...
@@ -24,8 +24,16 @@ PaddleInferenceAnakinPredictor::PaddleInferenceAnakinPredictor(
...
@@ -24,8 +24,16 @@ PaddleInferenceAnakinPredictor::PaddleInferenceAnakinPredictor(
}
}
bool
PaddleInferenceAnakinPredictor
::
Init
(
const
AnakinConfig
&
config
)
{
bool
PaddleInferenceAnakinPredictor
::
Init
(
const
AnakinConfig
&
config
)
{
// TODO(Superjomn) Tell anakin to support return code.
if
(
!
(
graph_
.
load
(
config
.
model_file
)))
{
engine_
.
Build
(
config
.
model_file
,
config
.
max_batch_size
);
return
false
;
}
graph_
.
ResetBatchSize
(
"input_0"
,
config
.
max_batch_size
);
// optimization for graph
if
(
!
(
graph_
.
Optimize
()))
{
return
false
;
}
// construct executer
executor_
.
init
(
graph_
);
return
true
;
return
true
;
}
}
...
@@ -38,24 +46,30 @@ bool PaddleInferenceAnakinPredictor::Run(
...
@@ -38,24 +46,30 @@ bool PaddleInferenceAnakinPredictor::Run(
<<
"'s type is not float"
;
<<
"'s type is not float"
;
return
false
;
return
false
;
}
}
engine_
.
SetInputFromCPU
(
auto
d_tensor_in_p
=
executor_
.
get_in
(
input
.
name
);
input
.
name
,
static_cast
<
float
*>
(
input
.
data
.
data
),
input
.
data
.
length
);
float
*
d_data_p
=
d_tensor_in_p
->
mutable_data
();
if
(
cudaMemcpy
(
d_data_p
,
static_cast
<
float
*>
(
input
.
data
.
data
),
d_tensor_in_p
->
valid_size
()
*
sizeof
(
float
),
cudaMemcpyHostToDevice
)
!=
0
)
{
LOG
(
ERROR
)
<<
"copy data from CPU to GPU error"
;
return
false
;
}
}
}
// TODO(Superjomn) Tell anakin to support return code.
executor_
.
prediction
();
engine_
.
Execute
();
if
(
output_data
->
empty
())
{
if
(
output_data
->
empty
())
{
LOG
(
ERROR
)
<<
"At least one output should be set with tensors' names."
;
LOG
(
ERROR
)
<<
"At least one output should be set with tensors' names."
;
return
false
;
return
false
;
}
}
for
(
auto
&
output
:
*
output_data
)
{
for
(
auto
&
output
:
*
output_data
)
{
auto
*
tensor
=
e
ngine_
.
GetOutputInGPU
(
output
.
name
);
auto
*
tensor
=
e
xecutor_
.
get_out
(
output
.
name
);
output
.
shape
=
tensor
->
shape
();
output
.
shape
=
tensor
->
shape
();
// Copy data from GPU -> CPU
// Copy data from GPU -> CPU
if
(
cudaMemcpy
(
output
.
data
.
data
,
if
(
cudaMemcpy
(
output
.
data
.
data
,
tensor
->
data
(),
tensor
->
mutable_
data
(),
tensor
->
size
(
),
tensor
->
valid_size
()
*
sizeof
(
float
),
cudaMemcpyDeviceToHost
)
!=
0
)
{
cudaMemcpyDeviceToHost
)
!=
0
)
{
LOG
(
ERROR
)
<<
"copy data from GPU to CPU error"
;
LOG
(
ERROR
)
<<
"copy data from GPU to CPU error"
;
return
false
;
return
false
;
...
@@ -64,9 +78,26 @@ bool PaddleInferenceAnakinPredictor::Run(
...
@@ -64,9 +78,26 @@ bool PaddleInferenceAnakinPredictor::Run(
return
true
;
return
true
;
}
}
// TODO(Superjomn) To implement latter.
anakin
::
Net
<
anakin
::
NV
,
anakin
::
saber
::
AK_FLOAT
,
anakin
::
Precision
::
FP32
>
&
PaddleInferenceAnakinPredictor
::
get_executer
()
{
return
executor_
;
}
// the cloned new Predictor of anakin share the same net weights from original
// Predictor
std
::
unique_ptr
<
PaddlePredictor
>
PaddleInferenceAnakinPredictor
::
Clone
()
{
std
::
unique_ptr
<
PaddlePredictor
>
PaddleInferenceAnakinPredictor
::
Clone
()
{
return
nullptr
;
VLOG
(
3
)
<<
"Anakin Predictor::clone"
;
std
::
unique_ptr
<
PaddlePredictor
>
cls
(
new
PaddleInferenceAnakinPredictor
());
// construct executer from other graph
auto
anakin_predictor_p
=
dynamic_cast
<
PaddleInferenceAnakinPredictor
*>
(
cls
.
get
());
if
(
!
anakin_predictor_p
)
{
LOG
(
ERROR
)
<<
"fail to call Init"
;
return
nullptr
;
}
anakin_predictor_p
->
get_executer
().
init
(
graph_
);
return
std
::
move
(
cls
);
}
}
// A factory to help create difference predictor.
// A factory to help create difference predictor.
...
@@ -74,6 +105,7 @@ template <>
...
@@ -74,6 +105,7 @@ template <>
std
::
unique_ptr
<
PaddlePredictor
>
std
::
unique_ptr
<
PaddlePredictor
>
CreatePaddlePredictor
<
AnakinConfig
,
PaddleEngineKind
::
kAnakin
>
(
CreatePaddlePredictor
<
AnakinConfig
,
PaddleEngineKind
::
kAnakin
>
(
const
AnakinConfig
&
config
)
{
const
AnakinConfig
&
config
)
{
VLOG
(
3
)
<<
"Anakin Predictor create."
;
std
::
unique_ptr
<
PaddlePredictor
>
x
(
std
::
unique_ptr
<
PaddlePredictor
>
x
(
new
PaddleInferenceAnakinPredictor
(
config
));
new
PaddleInferenceAnakinPredictor
(
config
));
return
x
;
return
x
;
...
...
paddle/contrib/inference/paddle_inference_api_anakin_engine.h
浏览文件 @
9141bee1
...
@@ -20,32 +20,42 @@ limitations under the License. */
...
@@ -20,32 +20,42 @@ limitations under the License. */
#pragma once
#pragma once
// NOTE This header file do not have namespace.
// NOTE This header file do not have namespace.
// TODO(Superjomn) Tell Anakin to provide better APIs.
//#include <test/framework/net/paddle_api.h>
#include <test/framework/net/paddle_api.h>
#include "paddle/contrib/inference/paddle_inference_api.h"
#include "paddle/contrib/inference/paddle_inference_api.h"
#include "framework/core/net/net.h"
#include "saber/saber_types.h"
namespace
paddle
{
namespace
paddle
{
class
PaddleInferenceAnakinPredictor
:
public
PaddlePredictor
{
class
PaddleInferenceAnakinPredictor
:
public
PaddlePredictor
{
public:
public:
PaddleInferenceAnakinPredictor
()
{}
PaddleInferenceAnakinPredictor
(
const
AnakinConfig
&
config
);
PaddleInferenceAnakinPredictor
(
const
AnakinConfig
&
config
);
// NOTE Unlike the native engine, the buffers of anakin engine's output_data
// NOTE Unlike the native engine, the buffers of anakin engine's output_data
// should be allocated first.
// should be allocated first.
// TODO(Superjomn) should unify all the behaviors of output_data accross all
// the engines.
bool
Run
(
const
std
::
vector
<
PaddleTensor
>&
inputs
,
bool
Run
(
const
std
::
vector
<
PaddleTensor
>&
inputs
,
std
::
vector
<
PaddleTensor
>*
output_data
)
override
;
std
::
vector
<
PaddleTensor
>*
output_data
)
override
;
std
::
unique_ptr
<
PaddlePredictor
>
Clone
()
override
;
std
::
unique_ptr
<
PaddlePredictor
>
Clone
()
override
;
anakin
::
Net
<
anakin
::
NV
,
anakin
::
saber
::
AK_FLOAT
,
anakin
::
Precision
::
FP32
>&
get_executer
();
~
PaddleInferenceAnakinPredictor
()
override
{};
private:
private:
bool
Init
(
const
AnakinConfig
&
config
);
bool
Init
(
const
AnakinConfig
&
config
);
anakin
::
AnakinEngine
<
anakin
::
NV
,
anakin
::
graph
::
Graph
<
anakin
::
NV
,
anakin
::
saber
::
AK_FLOAT
,
anakin
::
saber
::
AK_FLOAT
,
anakin
::
Precision
::
FP32
>
anakin
::
Precision
::
FP32
>
engine_
;
graph_
;
anakin
::
Net
<
anakin
::
NV
,
anakin
::
saber
::
AK_FLOAT
,
anakin
::
Precision
::
FP32
>
executor_
;
AnakinConfig
config_
;
};
};
}
// namespace paddle
}
// namespace paddle
paddle/contrib/inference/paddle_inference_api_anakin_engine_tester.cc
浏览文件 @
9141bee1
...
@@ -12,16 +12,54 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
...
@@ -12,16 +12,54 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
See the License for the specific language governing permissions and
limitations under the License. */
limitations under the License. */
#include
"paddle/contrib/inference/paddle_inference_api.h"
#include
<glog/logging.h>
#include <gtest/gtest.h>
#include <gtest/gtest.h>
#include "gflags/gflags.h"
#include "paddle/contrib/inference/paddle_inference_api.h"
namespace
paddle
{
namespace
paddle
{
TEST
(
inference
,
anakin
)
{
AnakinConfig
GetConfig
(
)
{
AnakinConfig
config
;
AnakinConfig
config
;
config
.
model_file
=
"./mobilenet_v2.anakin.bin"
;
config
.
device
=
0
;
config
.
max_batch_size
=
1
;
return
config
;
}
auto
engine
=
TEST
(
inference
,
anakin
)
{
AnakinConfig
config
=
GetConfig
();
auto
predictor
=
CreatePaddlePredictor
<
AnakinConfig
,
PaddleEngineKind
::
kAnakin
>
(
config
);
CreatePaddlePredictor
<
AnakinConfig
,
PaddleEngineKind
::
kAnakin
>
(
config
);
float
data
[
1
*
3
*
224
*
224
]
=
{
1.0
f
};
PaddleBuf
buf
{.
data
=
data
,
.
length
=
sizeof
(
data
)};
PaddleTensor
tensor
{.
name
=
"input_0"
,
.
shape
=
std
::
vector
<
int
>
({
1
,
3
,
224
,
224
}),
.
data
=
buf
,
.
dtype
=
PaddleDType
::
FLOAT32
};
// For simplicity, we set all the slots with the same data.
std
::
vector
<
PaddleTensor
>
paddle_tensor_feeds
(
1
,
tensor
);
float
data_out
[
1000
];
PaddleBuf
buf_out
{.
data
=
data_out
,
.
length
=
sizeof
(
data
)};
PaddleTensor
tensor_out
{.
name
=
"prob_out"
,
.
shape
=
std
::
vector
<
int
>
({
1000
,
1
}),
.
data
=
buf_out
,
.
dtype
=
PaddleDType
::
FLOAT32
};
std
::
vector
<
PaddleTensor
>
outputs
(
1
,
tensor_out
);
ASSERT_TRUE
(
predictor
->
Run
(
paddle_tensor_feeds
,
&
outputs
));
float
*
data_o
=
static_cast
<
float
*>
(
outputs
[
0
].
data
.
data
);
for
(
size_t
j
=
0
;
j
<
1000
;
++
j
)
{
LOG
(
INFO
)
<<
"output["
<<
j
<<
"]: "
<<
data_o
[
j
];
}
}
}
}
// namespace paddle
}
// namespace paddle
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