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
wk1234123
Paddle
提交
211e1315
P
Paddle
项目概览
wk1234123
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
4
Star
0
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
0
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
0
Issue
0
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 >>
未验证
提交
211e1315
编写于
5月 30, 2018
作者:
Y
Yan Chunwei
提交者:
GitHub
5月 30, 2018
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
feature/tensorrt engine op (#11001)
上级
49449205
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
213 addition
and
4 deletion
+213
-4
paddle/fluid/inference/tensorrt/engine.cc
paddle/fluid/inference/tensorrt/engine.cc
+25
-1
paddle/fluid/inference/tensorrt/engine.h
paddle/fluid/inference/tensorrt/engine.h
+5
-3
paddle/fluid/operators/CMakeLists.txt
paddle/fluid/operators/CMakeLists.txt
+3
-0
paddle/fluid/operators/tensorrt_engine_op.cc
paddle/fluid/operators/tensorrt_engine_op.cc
+70
-0
paddle/fluid/operators/tensorrt_engine_op.h
paddle/fluid/operators/tensorrt_engine_op.h
+110
-0
未找到文件。
paddle/fluid/inference/tensorrt/engine.cc
浏览文件 @
211e1315
...
...
@@ -131,6 +131,20 @@ void* TensorRTEngine::GetOutputInGPU(const std::string& name) {
return
buffer
(
name
).
buffer
;
}
void
TensorRTEngine
::
GetOutputInGPU
(
const
std
::
string
&
name
,
void
*
dst
,
size_t
max_size
)
{
// determine data size
auto
it
=
buffer_sizes_
.
find
(
name
);
PADDLE_ENFORCE
(
it
!=
buffer_sizes_
.
end
());
PADDLE_ENFORCE_GT
(
it
->
second
,
0
);
PADDLE_ENFORCE_GE
(
max_size
,
it
->
second
);
auto
&
buf
=
buffer
(
name
);
PADDLE_ENFORCE_NOT_NULL
(
buf
.
buffer
,
"buffer should be allocated before"
);
PADDLE_ENFORCE_EQ
(
cudaMemcpyAsync
(
dst
,
buf
.
buffer
,
it
->
second
,
cudaMemcpyDeviceToDevice
,
*
stream_
),
0
);
}
void
TensorRTEngine
::
GetOutputInCPU
(
const
std
::
string
&
name
,
void
*
dst
,
size_t
max_size
)
{
// determine data size
...
...
@@ -152,7 +166,7 @@ Buffer& TensorRTEngine::buffer(const std::string& name) {
return
buffers_
[
slot_offset
];
}
void
TensorRTEngine
::
SetInputFromCPU
(
const
std
::
string
&
name
,
void
*
data
,
void
TensorRTEngine
::
SetInputFromCPU
(
const
std
::
string
&
name
,
const
void
*
data
,
size_t
size
)
{
auto
&
buf
=
buffer
(
name
);
PADDLE_ENFORCE_NOT_NULL
(
buf
.
buffer
);
...
...
@@ -162,6 +176,16 @@ void TensorRTEngine::SetInputFromCPU(const std::string& name, void* data,
cudaMemcpyHostToDevice
,
*
stream_
));
}
void
TensorRTEngine
::
SetInputFromGPU
(
const
std
::
string
&
name
,
const
void
*
data
,
size_t
size
)
{
auto
&
buf
=
buffer
(
name
);
PADDLE_ENFORCE_NOT_NULL
(
buf
.
buffer
);
PADDLE_ENFORCE_LE
(
size
,
buf
.
max_size
,
"buffer is too small"
);
PADDLE_ENFORCE
(
buf
.
device
==
DeviceType
::
GPU
);
PADDLE_ENFORCE_EQ
(
0
,
cudaMemcpyAsync
(
buf
.
buffer
,
data
,
size
,
cudaMemcpyDeviceToDevice
,
*
stream_
));
}
void
TensorRTEngine
::
SetITensor
(
const
std
::
string
&
name
,
nvinfer1
::
ITensor
*
tensor
)
{
PADDLE_ENFORCE
(
tensor
!=
nullptr
);
...
...
paddle/fluid/inference/tensorrt/engine.h
浏览文件 @
211e1315
...
...
@@ -92,13 +92,15 @@ class TensorRTEngine : public EngineBase {
cudaStream_t
*
stream
()
{
return
stream_
;
}
// Fill an input from CPU memory with name and size.
void
SetInputFromCPU
(
const
std
::
string
&
name
,
void
*
data
,
size_t
size
);
void
SetInputFromCPU
(
const
std
::
string
&
name
,
const
void
*
data
,
size_t
size
);
// TODO(Superjomn) is this method necessary given that buffer(xxx) can be
// accessed directly. Fill an input from GPU memory with name and size.
void
SetInputFromGPU
(
const
std
::
string
&
name
,
void
*
data
,
size_t
size
);
void
SetInputFromGPU
(
const
std
::
string
&
name
,
const
void
*
data
,
size_t
size
);
// Get an output called name, the output of tensorrt is in GPU, so this method
//
will just return the output's GPU memory address
.
//
Return the output's GPU memory address without copy
.
void
*
GetOutputInGPU
(
const
std
::
string
&
name
);
// Copy data into dst inside the GPU device.
void
GetOutputInGPU
(
const
std
::
string
&
name
,
void
*
dst
,
size_t
max_size
);
// LOW EFFICENCY! Get output to CPU, this will trigger a memory copy from GPU
// to CPU.
void
GetOutputInCPU
(
const
std
::
string
&
name
,
void
*
dst
,
size_t
max_size
);
...
...
paddle/fluid/operators/CMakeLists.txt
浏览文件 @
211e1315
...
...
@@ -225,6 +225,9 @@ op_library(cross_entropy_op DEPS cross_entropy)
op_library
(
softmax_with_cross_entropy_op DEPS cross_entropy softmax
)
op_library
(
softmax_op DEPS softmax
)
op_library
(
sequence_softmax_op DEPS softmax
)
if
(
WITH_GPU AND TENSORRT_FOUND
)
op_library
(
tensorrt_engine_op DEPS tensorrt_engine
)
endif
()
op_library
(
sum_op DEPS selected_rows_functor
)
op_library
(
sgd_op DEPS selected_rows_functor
)
op_library
(
print_op DEPS lod_tensor
)
...
...
paddle/fluid/operators/tensorrt_engine_op.cc
0 → 100644
浏览文件 @
211e1315
/* 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. */
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/operators/tensorrt_engine_op.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
#include "paddle/fluid/inference/utils/singleton.h"
namespace
paddle
{
namespace
operators
{
template
<
typename
DeviceContext
,
typename
T
>
void
paddle
::
operators
::
TensorRTEngineKernel
<
DeviceContext
,
T
>::
Prepare
(
const
framework
::
ExecutionContext
&
context
)
const
{
// Get the ProgramDesc and pass to convert.
const
auto
&
block
=
context
.
Attr
<
framework
::
proto
::
BlockDesc
>
(
"subgraph"
);
max_batch_
=
context
.
Attr
<
int
>
(
"max_batch"
);
auto
max_workspace
=
context
.
Attr
<
int
>
(
"max_workspace"
);
engine_
.
reset
(
new
inference
::
tensorrt
::
TensorRTEngine
(
max_batch_
,
max_workspace
,
nullptr
));
inference
::
Singleton
<
inference
::
tensorrt
::
OpConverter
>::
Global
().
ConvertBlock
(
block
,
engine_
.
get
());
engine_
->
FreezeNetwork
();
}
class
TensorRTEngineOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
void
Make
()
override
{
AddInput
(
"Xs"
,
"A list of inputs."
).
AsDuplicable
();
AddOutput
(
"Ys"
,
"A list of outputs"
).
AsDuplicable
();
AddAttr
<
std
::
string
>
(
"subgraph"
,
"the subgraph"
);
AddComment
(
"TensorRT engine operator."
);
}
};
class
TensorRTEngineInferVarType
:
public
framework
::
VarTypeInference
{
public:
void
operator
()(
const
framework
::
OpDesc
&
op_desc
,
framework
::
BlockDesc
*
block
)
const
override
{}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OPERATOR
(
tensorrt_engine
,
ops
::
TensorRTEngineOp
,
ops
::
TensorRTEngineOpMaker
,
ops
::
TensorRTEngineOpMaker
);
REGISTER_OP_CPU_KERNEL
(
tensorrt_engine
,
ops
::
TensorRTEngineKernel
<
paddle
::
platform
::
CPUDeviceContext
,
float
>
,
ops
::
TensorRTEngineKernel
<
paddle
::
platform
::
CPUDeviceContext
,
double
>
,
ops
::
TensorRTEngineKernel
<
paddle
::
platform
::
CPUDeviceContext
,
int
>
,
ops
::
TensorRTEngineKernel
<
paddle
::
platform
::
CPUDeviceContext
,
int64_t
>
);
#endif // PADDLE_WITH_CUDA
paddle/fluid/operators/tensorrt_engine_op.h
0 → 100644
浏览文件 @
211e1315
/* 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
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/inference/analysis/helper.h"
#include "paddle/fluid/inference/tensorrt/engine.h"
namespace
paddle
{
namespace
operators
{
class
TensorRTEngineOp
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
protected:
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{}
framework
::
OpKernelType
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
framework
::
OpKernelType
kt
=
framework
::
OpKernelType
(
framework
::
ToDataType
(
ctx
.
Input
<
framework
::
LoDTensor
>
(
"pre_ids"
)
->
type
()),
platform
::
CPUPlace
());
return
kt
;
}
};
template
<
typename
DeviceContext
,
typename
T
>
class
TensorRTEngineKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
if
(
!
engine_
)
{
Prepare
(
context
);
}
auto
input_names
=
context
.
op
().
Inputs
(
"Xs"
);
PADDLE_ENFORCE
(
!
input_names
.
empty
(),
"should pass more than one inputs"
);
// Try to determine a batch_size
auto
*
tensor0
=
context
.
Input
<
framework
::
LoDTensor
>
(
input_names
.
front
());
PADDLE_ENFORCE_NOT_NULL
(
tensor0
);
int
batch_size
=
tensor0
->
dims
()[
0
];
PADDLE_ENFORCE_LE
(
batch_size
,
max_batch_
);
// Convert input tensor from fluid to engine.
for
(
const
auto
&
x
:
context
.
Inputs
(
"Xs"
))
{
// convert input and copy to TRT engine's buffer
auto
*
v
=
context
.
scope
().
FindVar
(
x
);
PADDLE_ENFORCE_NOT_NULL
(
v
,
"no variable called %s"
,
x
);
auto
&
t
=
v
->
Get
<
framework
::
LoDTensor
>
();
if
(
platform
::
is_cpu_place
(
t
.
place
()))
{
engine_
->
SetInputFromCPU
(
x
,
static_cast
<
const
void
*>
(
t
.
data
<
void
>
()),
t
.
memory_size
());
}
else
{
engine_
->
SetInputFromGPU
(
x
,
static_cast
<
const
void
*>
(
t
.
data
<
void
>
()),
t
.
memory_size
());
}
}
// Execute the engine.
PADDLE_ENFORCE_GT
(
batch_size
,
0
);
engine_
->
Execute
(
batch_size
);
// Convert output tensor from engine to fluid
for
(
const
auto
&
y
:
context
.
Outputs
(
"Ys"
))
{
// convert output and copy to fluid.
nvinfer1
::
ITensor
*
trt_t
=
engine_
->
GetITensor
(
y
);
auto
dims
=
trt_t
->
getDimensions
();
// Use the output ITensor's dims to reshape the Fluid Tensor.
std
::
vector
<
int
>
ddim
(
dims
.
d
,
dims
.
d
+
dims
.
nbDims
);
auto
*
fluid_v
=
context
.
scope
().
FindVar
(
y
);
PADDLE_ENFORCE_NOT_NULL
(
fluid_v
,
"no output variable called %s"
,
y
);
auto
*
fluid_t
=
fluid_v
->
GetMutable
<
framework
::
LoDTensor
>
();
fluid_t
->
Resize
(
framework
::
make_ddim
(
ddim
));
auto
size
=
inference
::
analysis
::
AccuDims
(
dims
.
d
,
dims
.
nbDims
);
if
(
platform
::
is_cpu_place
(
fluid_t
->
place
()))
{
engine_
->
GetOutputInCPU
(
y
,
fluid_t
->
mutable_data
<
float
>
(
platform
::
CPUPlace
()),
size
);
}
else
{
engine_
->
GetOutputInGPU
(
y
,
fluid_t
->
mutable_data
<
float
>
(
platform
::
CUDAPlace
()),
size
);
}
}
}
protected:
// Build the engine.
void
Prepare
(
const
framework
::
ExecutionContext
&
context
)
const
;
private:
mutable
std
::
unique_ptr
<
inference
::
tensorrt
::
TensorRTEngine
>
engine_
;
mutable
int
max_batch_
{
0
};
};
}
// namespace operators
}
// namespace paddle
#endif // PADDLE_WITH_CUDA
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