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
urhero
Paddle
提交
96216052
P
Paddle
项目概览
urhero
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
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 >>
提交
96216052
编写于
12月 14, 2018
作者:
N
nhzlx
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
1. fix trt multi thread bug
上级
30aad884
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
81 addition
and
125 deletion
+81
-125
paddle/fluid/inference/tensorrt/convert/op_converter.h
paddle/fluid/inference/tensorrt/convert/op_converter.h
+2
-0
paddle/fluid/operators/tensorrt/CMakeLists.txt
paddle/fluid/operators/tensorrt/CMakeLists.txt
+1
-1
paddle/fluid/operators/tensorrt/tensorrt_engine_op.cc
paddle/fluid/operators/tensorrt/tensorrt_engine_op.cc
+1
-3
paddle/fluid/operators/tensorrt/tensorrt_engine_op.cu.cc
paddle/fluid/operators/tensorrt/tensorrt_engine_op.cu.cc
+0
-24
paddle/fluid/operators/tensorrt/tensorrt_engine_op.h
paddle/fluid/operators/tensorrt/tensorrt_engine_op.h
+77
-95
paddle/fluid/operators/tensorrt/tensorrt_engine_op_test.cc
paddle/fluid/operators/tensorrt/tensorrt_engine_op_test.cc
+0
-2
未找到文件。
paddle/fluid/inference/tensorrt/convert/op_converter.h
浏览文件 @
96216052
...
...
@@ -103,6 +103,7 @@ class OpConverter {
void
ConvertBlock
(
const
framework
::
proto
::
BlockDesc
&
block
,
const
std
::
unordered_set
<
std
::
string
>&
parameters
,
const
framework
::
Scope
&
scope
,
TensorRTEngine
*
engine
)
{
std
::
unique_lock
<
std
::
mutex
>
lk
(
mut_
);
for
(
int
i
=
0
;
i
<
block
.
ops_size
();
i
++
)
{
const
auto
&
op
=
block
.
ops
(
i
);
ConvertOp
(
op
,
parameters
,
scope
,
engine
);
...
...
@@ -125,6 +126,7 @@ class OpConverter {
std
::
unordered_map
<
std
::
string
,
OpConverter
*>
converters_
;
// fluid inference scope
framework
::
Scope
*
scope_
{
nullptr
};
std
::
mutex
mut_
;
};
}
// namespace tensorrt
...
...
paddle/fluid/operators/tensorrt/CMakeLists.txt
浏览文件 @
96216052
op_library
(
tensorrt_engine_op DEPS tensorrt_engine tensorrt_converter
)
file
(
APPEND
${
pybind_file
}
"USE_
CUDA_ONLY
_OP(tensorrt_engine);
\n
"
)
file
(
APPEND
${
pybind_file
}
"USE_
NO_KERNEL
_OP(tensorrt_engine);
\n
"
)
nv_test
(
test_tensorrt_engine_op SRCS tensorrt_engine_op_test.cc
DEPS tensorrt_engine_op
analysis
)
paddle/fluid/operators/tensorrt/tensorrt_engine_op.cc
浏览文件 @
96216052
...
...
@@ -21,8 +21,6 @@
namespace
paddle
{
DEFINE_int32
(
tensorrt_engine_batch_size
,
1
,
"the batch_size of TensorRT"
);
namespace
operators
{
class
TensorRTEngineOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
...
...
@@ -50,6 +48,6 @@ class TensorRTEngineInferVarType : public framework::VarTypeInference {
namespace
ops
=
paddle
::
operators
;
REGISTER_OPERATOR
(
tensorrt_engine
,
ops
::
TensorRTEngineOp
,
ops
::
TensorRTEngineOpMaker
,
ops
::
TensorRTEngineOpMaker
);
ops
::
TensorRTEngineOpMaker
);
#endif // PADDLE_WITH_CUDA
paddle/fluid/operators/tensorrt/tensorrt_engine_op.cu.cc
已删除
100644 → 0
浏览文件 @
30aad884
/* Copyright (c) 2016 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/operators/tensorrt/tensorrt_engine_op.h"
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_CUDA_KERNEL
(
tensorrt_engine
,
ops
::
TensorRTEngineKernel
<
paddle
::
platform
::
CUDADeviceContext
,
float
>
,
ops
::
TensorRTEngineKernel
<
paddle
::
platform
::
CUDADeviceContext
,
double
>
,
ops
::
TensorRTEngineKernel
<
paddle
::
platform
::
CUDADeviceContext
,
int
>
,
ops
::
TensorRTEngineKernel
<
paddle
::
platform
::
CUDADeviceContext
,
int64_t
>
);
paddle/fluid/operators/tensorrt/tensorrt_engine_op.h
浏览文件 @
96216052
...
...
@@ -27,8 +27,6 @@
namespace
paddle
{
DECLARE_int32
(
tensorrt_engine_batch_size
);
namespace
operators
{
using
FluidDT
=
framework
::
proto
::
VarType_Type
;
...
...
@@ -49,7 +47,7 @@ TRT_DT FluidDataType2TRT(FluidDT type) {
return
TRT_DT
::
kINT32
;
}
nvinfer1
::
Dims
Vec2TRT_Dims
(
const
std
::
vector
<
int64_t
>
&
shape
)
{
nvinfer1
::
Dims
Vec2TRT_Dims
(
const
std
::
vector
<
int64_t
>
&
shape
)
{
PADDLE_ENFORCE_GT
(
shape
.
size
(),
1UL
,
"TensorRT' tensor input requires at least 2 dimensions"
);
PADDLE_ENFORCE_LE
(
shape
.
size
(),
4UL
,
...
...
@@ -63,131 +61,121 @@ nvinfer1::Dims Vec2TRT_Dims(const std::vector<int64_t>& shape) {
}
// namespace // NOLINT
using
inference
::
Singleton
;
using
inference
::
tensorrt
::
TRT_EngineManager
;
using
inference
::
tensorrt
::
TensorRTEngine
;
class
TensorRTEngineOp
:
public
framework
::
OperatorBase
{
private:
std
::
string
engine_name_
;
std
::
vector
<
std
::
string
>
input_names_
;
std
::
unordered_set
<
std
::
string
>
param_names_
;
mutable
std
::
unique_ptr
<
TensorRTEngine
>
trt_engine_
;
int
max_batch_size_
;
int
workspace_size_
;
class
TensorRTEngineOp
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
TensorRTEngineOp
(
const
std
::
string
&
type
,
const
framework
::
VariableNameMap
&
inputs
,
const
framework
::
VariableNameMap
&
outputs
,
const
framework
::
AttributeMap
&
attrs
)
:
framework
::
OperatorBase
(
type
,
inputs
,
outputs
,
attrs
)
{
engine_name_
=
Attr
<
std
::
string
>
(
"engine_uniq_key"
);
input_names_
=
Inputs
(
"Xs"
);
max_batch_size_
=
Attr
<
int
>
(
"max_batch_size"
);
workspace_size_
=
Attr
<
int
>
(
"workspace_size"
);
auto
params
=
Attr
<
std
::
vector
<
std
::
string
>>
(
"parameters"
);
for
(
const
auto
&
param
:
params
)
{
param_names_
.
insert
(
param
);
}
}
protected:
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{}
framework
::
OpKernelType
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
input0
=
ctx
.
Inputs
(
"Xs"
).
front
();
framework
::
OpKernelType
kt
=
framework
::
OpKernelType
(
framework
::
ToDataType
(
ctx
.
scope
()
.
FindVar
(
input0
)
->
GetMutable
<
framework
::
LoDTensor
>
()
->
type
()),
ctx
.
GetPlace
());
return
kt
;
void
RunImpl
(
const
framework
::
Scope
&
scope
,
const
platform
::
Place
&
dev_place
)
const
override
{
RunTrt
(
scope
,
dev_place
);
}
};
template
<
typename
DeviceContext
,
typename
T
>
class
TensorRTEngineKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
auto
engine_name
=
context
.
Attr
<
std
::
string
>
(
"engine_uniq_key"
);
int
max_batch_size
=
context
.
Attr
<
int
>
(
"max_batch_size"
);
if
(
!
Singleton
<
TRT_EngineManager
>::
Global
().
HasEngine
(
engine_name
))
{
Prepare
(
context
);
void
RunTrt
(
const
framework
::
Scope
&
scope
,
const
platform
::
Place
&
dev_place
)
const
{
int
runtime_batch
=
1
;
if
(
trt_engine_
.
get
()
==
nullptr
)
{
trt_engine_
.
reset
(
new
TensorRTEngine
(
max_batch_size_
,
workspace_size_
,
nullptr
,
boost
::
get
<
platform
::
CUDAPlace
>
(
dev_place
).
device
));
Prepare
(
scope
,
dev_place
,
trt_engine_
.
get
()
);
}
auto
*
engine
=
Singleton
<
TRT_EngineManager
>::
Global
().
Get
(
engine_name
);
auto
input_names
=
context
.
op
().
Inputs
(
"Xs"
);
PADDLE_ENFORCE
(
!
input_names
.
empty
(),
"should pass more than one inputs"
);
PADDLE_ENFORCE_LE
(
FLAGS_tensorrt_engine_batch_size
,
max_batch_size
);
auto
*
engine
=
trt_engine_
.
get
();
PADDLE_ENFORCE
(
!
input_names_
.
empty
(),
"should pass more than one inputs"
);
std
::
vector
<
std
::
string
>
output_maps
=
context
.
Attr
<
std
::
vector
<
std
::
string
>>
(
"output_name_mapping"
);
Attr
<
std
::
vector
<
std
::
string
>>
(
"output_name_mapping"
);
auto
params
=
context
.
Attr
<
std
::
vector
<
std
::
string
>>
(
"parameters"
);
std
::
unordered_set
<
std
::
string
>
parameters
;
for
(
const
auto
&
param
:
params
)
{
parameters
.
insert
(
param
);
}
// Convert input tensor from fluid to engine.
for
(
const
auto
&
x
:
context
.
Inputs
(
"Xs"
))
{
if
(
param
eters
.
count
(
x
))
continue
;
for
(
const
auto
&
x
:
Inputs
(
"Xs"
))
{
if
(
param
_names_
.
count
(
x
))
continue
;
// convert input and copy to TRT engine's buffer
auto
&
t
=
inference
::
analysis
::
GetFromScope
<
framework
::
LoDTensor
>
(
context
.
scope
(),
x
);
auto
&
t
=
inference
::
analysis
::
GetFromScope
<
framework
::
LoDTensor
>
(
scope
,
x
);
auto
t_shape
=
framework
::
vectorize
(
t
.
dims
());
runtime_batch
=
t_shape
[
0
];
if
(
platform
::
is_cpu_place
(
t
.
place
()))
{
engine
->
SetInputFromCPU
(
x
,
static_cast
<
const
void
*>
(
t
.
data
<
void
>
()),
engine
->
SetInputFromCPU
(
x
,
static_cast
<
const
void
*>
(
t
.
data
<
void
>
()),
t
.
memory_size
());
}
else
{
engine
->
SetInputFromGPU
(
x
,
static_cast
<
const
void
*>
(
t
.
data
<
void
>
()),
engine
->
SetInputFromGPU
(
x
,
static_cast
<
const
void
*>
(
t
.
data
<
void
>
()),
t
.
memory_size
());
}
}
PADDLE_ENFORCE_LE
(
runtime_batch
,
max_batch_size_
);
// Execute the engine.
PADDLE_ENFORCE_GT
(
FLAGS_tensorrt_engine_batch_size
,
0
);
engine
->
Execute
(
FLAGS_tensorrt_engine_batch_size
);
engine
->
Execute
(
runtime_batch
);
// Convert output tensor from engine to fluid
int
output_index
=
0
;
VLOG
(
4
)
<<
"TensorRT Engine Op Outputs:"
;
for
(
const
auto
&
y
:
context
.
Outputs
(
"Ys"
))
{
for
(
const
auto
&
y
:
Outputs
(
"Ys"
))
{
VLOG
(
4
)
<<
y
;
// convert output and copy to fluid.
nvinfer1
::
ITensor
*
trt_t
=
engine
->
GetITensor
(
output_maps
[
output_index
]);
nvinfer1
::
ITensor
*
trt_t
=
engine
->
GetITensor
(
output_maps
[
output_index
]);
auto
dims
=
trt_t
->
getDimensions
();
// Use the output ITensor's dims to reshape the Fluid Tensor.
// The ITensor doesn't contain the batch size dim.
std
::
vector
<
int
>
ddim
;
ddim
.
push_back
(
FLAGS_tensorrt_engine_batch_size
);
ddim
.
push_back
(
runtime_batch
);
for
(
int
i
=
0
;
i
<
dims
.
nbDims
;
i
++
)
{
ddim
.
push_back
(
dims
.
d
[
i
]);
}
auto
*
fluid_v
=
context
.
scope
()
.
FindVar
(
y
);
auto
*
fluid_v
=
scope
.
FindVar
(
y
);
PADDLE_ENFORCE_NOT_NULL
(
fluid_v
,
"no output variable called %s"
,
y
);
auto
*
fluid_t
=
fluid_v
->
GetMutable
<
framework
::
LoDTensor
>
();
auto
*
fluid_t
=
fluid_v
->
GetMutable
<
framework
::
LoDTensor
>
();
fluid_t
->
Resize
(
framework
::
make_ddim
(
ddim
));
// TODO(Superjomn) find some way to determine which device to output the
// tensor.
// if (platform::is_cpu_place(fluid_t->place())) {
// TODO(Superjomn) change this float to dtype size.
auto
size
=
inference
::
analysis
::
AccuDims
(
dims
.
d
,
dims
.
nbDims
)
*
FLAGS_tensorrt_engine_batch_size
;
auto
size
=
inference
::
analysis
::
AccuDims
(
dims
.
d
,
dims
.
nbDims
)
*
runtime_batch
;
engine
->
GetOutputInGPU
(
output_maps
[
output_index
],
fluid_t
->
mutable_data
<
float
>
(
platform
::
CUDAPlace
(
boost
::
get
<
platform
::
CUDAPlace
>
(
context
.
GetPlace
()
).
device
)),
boost
::
get
<
platform
::
CUDAPlace
>
(
dev_place
).
device
)),
size
*
sizeof
(
float
));
output_index
+=
1
;
}
cudaStreamSynchronize
(
*
engine
->
stream
());
}
protected:
void
Prepare
(
const
framework
::
ExecutionContext
&
context
)
const
{
void
Prepare
(
const
framework
::
Scope
&
scope
,
const
platform
::
Place
&
dev_place
,
TensorRTEngine
*
engine
)
const
{
VLOG
(
4
)
<<
"Prepare engine"
;
// Get the ProgramDesc and pass to convert.
framework
::
proto
::
BlockDesc
block_desc
;
block_desc
.
ParseFromString
(
context
.
Attr
<
std
::
string
>
(
"subgraph"
));
int
max_batch_size
=
context
.
Attr
<
int
>
(
"max_batch_size"
);
int
workspace_size
=
context
.
Attr
<
int
>
(
"workspace_size"
);
auto
params
=
context
.
Attr
<
std
::
vector
<
std
::
string
>>
(
"parameters"
);
std
::
unordered_set
<
std
::
string
>
parameters
;
for
(
const
auto
&
param
:
params
)
{
parameters
.
insert
(
param
);
}
block_desc
.
ParseFromString
(
Attr
<
std
::
string
>
(
"subgraph"
));
std
::
vector
<
std
::
string
>
output_maps
=
context
.
Attr
<
std
::
vector
<
std
::
string
>>
(
"output_name_mapping"
);
// TODO(Superjomn) replace this with a different stream
auto
*
engine
=
Singleton
<
TRT_EngineManager
>::
Global
().
Create
(
max_batch_size
,
workspace_size
,
nullptr
/*engine hold its own stream*/
,
context
.
Attr
<
std
::
string
>
(
"engine_uniq_key"
),
boost
::
get
<
platform
::
CUDAPlace
>
(
context
.
GetPlace
()).
device
);
Attr
<
std
::
vector
<
std
::
string
>>
(
"output_name_mapping"
);
engine
->
InitNetwork
();
...
...
@@ -195,39 +183,33 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
VLOG
(
4
)
<<
"parsed var size "
<<
block
.
AllVars
().
size
();
// Add inputs
VLOG
(
4
)
<<
"declare inputs"
;
for
(
auto
&
input
:
context
.
Inputs
(
"Xs"
))
{
if
(
param
eters
.
count
(
input
))
continue
;
for
(
auto
&
input
:
Inputs
(
"Xs"
))
{
if
(
param
_names_
.
count
(
input
))
continue
;
VLOG
(
4
)
<<
"declare input "
<<
input
;
auto
*
var
=
block
.
FindVar
(
input
);
auto
&
t
=
inference
::
analysis
::
GetFromScope
<
framework
::
LoDTensor
>
(
scope
,
input
);
auto
t_shape
=
framework
::
vectorize
(
t
.
dims
());
auto
*
var
=
block
.
FindVar
(
input
);
// TensorRT engine need to create parameters. The parameter's description
// should be set in
PADDLE_ENFORCE
(
var
,
"no variable called %s"
,
input
);
PADDLE_ENFORCE_EQ
(
var
->
GetType
(),
FluidDT
::
VarType_Type_LOD_TENSOR
,
"TensorRT engine only takes LoDTensor as input"
);
auto
shape
=
var
->
GetShape
();
// For the special batch_size placeholder -1, drop it and pass the real
// shape of data.
// TODO(Superjomn) fix this with batch broadcast, or it can't handle
// variational batch size.
if
(
shape
[
0
]
==
-
1
)
{
shape
[
0
]
=
FLAGS_tensorrt_engine_batch_size
;
}
engine
->
DeclareInput
(
input
,
FluidDataType2TRT
(
var
->
Proto
()
->
type
().
lod_tensor
().
tensor
().
data_type
()),
Vec2TRT_Dims
(
shape
));
Vec2TRT_Dims
(
t_
shape
));
}
inference
::
Singleton
<
inference
::
tensorrt
::
OpConverter
>::
Global
()
.
ConvertBlock
(
block_desc
,
param
eters
,
context
.
scope
()
,
engine
);
.
ConvertBlock
(
block_desc
,
param
_names_
,
scope
,
engine
);
// Add outputs
for
(
auto
&
output
:
output_maps
)
{
if
(
!
engine
->
HasDeclared
(
output
))
{
engine
->
DeclareOutput
(
output
);
}
for
(
auto
&
output
:
output_maps
)
{
engine
->
DeclareOutput
(
output
);
}
engine
->
FreezeNetwork
();
}
};
...
...
paddle/fluid/operators/tensorrt/tensorrt_engine_op_test.cc
浏览文件 @
96216052
...
...
@@ -24,8 +24,6 @@ limitations under the License. */
#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h"
USE_CUDA_ONLY_OP
(
tensorrt_engine
);
namespace
paddle
{
namespace
operators
{
...
...
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