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
PaddleDetection
提交
d7dd0868
P
PaddleDetection
项目概览
PaddlePaddle
/
PaddleDetection
大约 1 年 前同步成功
通知
694
Star
11112
Fork
2696
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
184
列表
看板
标记
里程碑
合并请求
40
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
PaddleDetection
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
184
Issue
184
列表
看板
标记
里程碑
合并请求
40
合并请求
40
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
d7dd0868
编写于
8月 01, 2018
作者:
Z
Zhaolong Xing
提交者:
GitHub
8月 01, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #12449 from NHZlX/add_tensorrt_elementwise_add
Add tensorrt elementwise add
上级
900d61dd
64a08f84
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
313 addition
and
2 deletion
+313
-2
paddle/fluid/inference/tensorrt/convert/CMakeLists.txt
paddle/fluid/inference/tensorrt/convert/CMakeLists.txt
+4
-1
paddle/fluid/inference/tensorrt/convert/elementwise_op.cc
paddle/fluid/inference/tensorrt/convert/elementwise_op.cc
+210
-0
paddle/fluid/inference/tensorrt/convert/op_converter.h
paddle/fluid/inference/tensorrt/convert/op_converter.h
+25
-0
paddle/fluid/inference/tensorrt/convert/test_elementwise_op.cc
...e/fluid/inference/tensorrt/convert/test_elementwise_op.cc
+73
-0
paddle/fluid/inference/tensorrt/convert/ut_helper.h
paddle/fluid/inference/tensorrt/convert/ut_helper.h
+1
-1
未找到文件。
paddle/fluid/inference/tensorrt/convert/CMakeLists.txt
浏览文件 @
d7dd0868
# Add TRT tests
# Add TRT tests
nv_library
(
tensorrt_converter
nv_library
(
tensorrt_converter
SRCS mul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc
SRCS mul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc
elementwise_op.cc
DEPS tensorrt_engine operator scope framework_proto op_registry
)
DEPS tensorrt_engine operator scope framework_proto op_registry
)
nv_test
(
test_op_converter SRCS test_op_converter.cc DEPS
nv_test
(
test_op_converter SRCS test_op_converter.cc DEPS
...
@@ -17,3 +17,6 @@ nv_test(test_trt_conv_op SRCS test_conv2d_op.cc conv2d_op.cc
...
@@ -17,3 +17,6 @@ nv_test(test_trt_conv_op SRCS test_conv2d_op.cc conv2d_op.cc
DEPS
${
FLUID_CORE_MODULES
}
tensorrt_engine conv_op SERIAL
)
DEPS
${
FLUID_CORE_MODULES
}
tensorrt_engine conv_op SERIAL
)
nv_test
(
test_trt_pool2d_op SRCS test_pool2d_op.cc pool2d_op.cc
nv_test
(
test_trt_pool2d_op SRCS test_pool2d_op.cc pool2d_op.cc
DEPS
${
FLUID_CORE_MODULES
}
tensorrt_engine pool_op SERIAL
)
DEPS
${
FLUID_CORE_MODULES
}
tensorrt_engine pool_op SERIAL
)
nv_test
(
test_trt_elementwise_op SRCS test_elementwise_op.cc elementwise_op.cc
DEPS
${
FLUID_CORE_MODULES
}
tensorrt_engine elementwise_add_op SERIAL
)
paddle/fluid/inference/tensorrt/convert/elementwise_op.cc
0 → 100644
浏览文件 @
d7dd0868
/* 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. */
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
namespace
paddle
{
namespace
inference
{
namespace
tensorrt
{
class
ElementwiseWeightOpConverter
:
public
OpConverter
{
public:
ElementwiseWeightOpConverter
()
{}
void
operator
()(
const
framework
::
proto
::
OpDesc
&
op
,
const
framework
::
Scope
&
scope
,
bool
test_mode
)
override
{
// Here the two nullptr looks strange, that's because the
// framework::OpDesc's constructor is strange.
framework
::
OpDesc
op_desc
(
op
,
nullptr
);
LOG
(
INFO
)
<<
"convert a fluid elementwise op to tensorrt IScaleLayer"
;
PADDLE_ENFORCE_EQ
(
op_desc
.
Input
(
"X"
).
size
(),
1
);
PADDLE_ENFORCE_EQ
(
op_desc
.
Input
(
"Y"
).
size
(),
1
);
// Y is a weight
PADDLE_ENFORCE_EQ
(
op_desc
.
Output
(
"Out"
).
size
(),
1
);
auto
*
X
=
engine_
->
GetITensor
(
op_desc
.
Input
(
"X"
).
front
());
nvinfer1
::
Dims
dims_x
=
X
->
getDimensions
();
PADDLE_ENFORCE
(
dims_x
.
nbDims
>=
3
);
auto
*
Y_v
=
scope
.
FindVar
(
op_desc
.
Input
(
"Y"
).
front
());
PADDLE_ENFORCE_NOT_NULL
(
Y_v
);
auto
*
Y_t
=
Y_v
->
GetMutable
<
framework
::
LoDTensor
>
();
auto
*
weight_data
=
Y_t
->
mutable_data
<
float
>
(
platform
::
CPUPlace
());
auto
scale_mode
=
nvinfer1
::
ScaleMode
::
kELEMENTWISE
;
std
::
vector
<
int
>
dims_y
=
framework
::
vectorize2int
(
Y_t
->
dims
());
if
(
static_cast
<
int
>
(
dims_y
.
size
())
==
dims_x
.
nbDims
+
1
)
{
if
(
dims_y
[
0
]
==
1
)
dims_y
.
erase
(
dims_y
.
begin
());
}
if
(
static_cast
<
int
>
(
dims_y
.
size
())
==
1
&&
dims_y
[
0
]
==
dims_x
.
d
[
0
])
{
scale_mode
=
nvinfer1
::
ScaleMode
::
kCHANNEL
;
}
else
if
(
static_cast
<
int
>
(
dims_y
.
size
())
==
dims_x
.
nbDims
&&
dims_y
[
0
]
==
dims_x
.
d
[
0
])
{
scale_mode
=
nvinfer1
::
ScaleMode
::
kELEMENTWISE
;
for
(
int
i
=
1
;
i
<
dims_x
.
nbDims
;
i
++
)
{
if
(
dims_y
[
i
]
!=
dims_x
.
d
[
i
])
{
scale_mode
=
nvinfer1
::
ScaleMode
::
kCHANNEL
;
break
;
}
}
if
(
scale_mode
==
nvinfer1
::
ScaleMode
::
kCHANNEL
)
{
for
(
int
i
=
1
;
i
<
dims_x
.
nbDims
;
i
++
)
{
if
(
dims_y
[
i
]
!=
1
)
PADDLE_THROW
(
"TensorRT unsupported weight shape for Elementwise op!"
);
}
}
}
else
{
PADDLE_THROW
(
"TensorRT unsupported weight Shape for Elementwise op!"
);
}
TensorRTEngine
::
Weight
shift_weights
{
nvinfer1
::
DataType
::
kFLOAT
,
static_cast
<
void
*>
(
weight_data
),
Y_t
->
memory_size
()
/
sizeof
(
float
)};
TensorRTEngine
::
Weight
scale_weights
{
nvinfer1
::
DataType
::
kFLOAT
,
nullptr
,
0
};
TensorRTEngine
::
Weight
power_weights
{
nvinfer1
::
DataType
::
kFLOAT
,
nullptr
,
0
};
nvinfer1
::
IScaleLayer
*
layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Scale
,
*
const_cast
<
nvinfer1
::
ITensor
*>
(
X
),
scale_mode
,
shift_weights
.
get
(),
scale_weights
.
get
(),
power_weights
.
get
());
auto
output_name
=
op_desc
.
Output
(
"Out"
)[
0
];
engine_
->
SetITensor
(
output_name
,
layer
->
getOutput
(
0
));
if
(
test_mode
)
{
// the test framework can not determine which is the
// output, so place the declaration inside.
engine_
->
DeclareOutput
(
output_name
);
}
}
};
class
ElementwiseTensorOpConverter
:
public
OpConverter
{
public:
ElementwiseTensorOpConverter
()
{}
void
operator
()(
const
framework
::
proto
::
OpDesc
&
op
,
const
framework
::
Scope
&
scope
,
bool
test_mode
)
override
{
// Here the two nullptr looks strange, that's because the
// framework::OpDesc's constructor is strange.
framework
::
OpDesc
op_desc
(
op
,
nullptr
);
LOG
(
INFO
)
<<
"convert a fluid elementwise op to tensorrt IScaleLayer"
;
PADDLE_ENFORCE_EQ
(
op_desc
.
Input
(
"X"
).
size
(),
1
);
PADDLE_ENFORCE_EQ
(
op_desc
.
Input
(
"Y"
).
size
(),
1
);
// Y is a weight
PADDLE_ENFORCE_EQ
(
op_desc
.
Output
(
"Out"
).
size
(),
1
);
auto
*
X
=
engine_
->
GetITensor
(
op_desc
.
Input
(
"X"
).
front
());
auto
*
Y
=
engine_
->
GetITensor
(
op_desc
.
Input
(
"Y"
).
front
());
nvinfer1
::
Dims
dims_x
=
X
->
getDimensions
();
nvinfer1
::
Dims
dims_y
=
Y
->
getDimensions
();
// The two input tensor should have the same dims
PADDLE_ENFORCE
(
dims_x
.
nbDims
>=
3
);
if
(
dims_x
.
nbDims
==
dims_y
.
nbDims
)
{
for
(
int
i
=
0
;
i
<
dims_x
.
nbDims
;
i
++
)
{
if
(
dims_x
.
d
[
i
]
!=
dims_y
.
d
[
i
])
PADDLE_THROW
(
"TensorRT unsupported tensor shape for Elementwise op!"
);
}
}
else
{
PADDLE_THROW
(
"TensorRT unsupported tensor shape for Elementwise op!"
);
}
auto
op_pair
=
ops
.
find
(
op_type_
);
if
(
op_pair
==
ops
.
end
())
{
PADDLE_THROW
(
"Wrong elementwise op type!"
);
}
nvinfer1
::
IElementWiseLayer
*
layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
ElementWise
,
*
const_cast
<
nvinfer1
::
ITensor
*>
(
X
),
*
const_cast
<
nvinfer1
::
ITensor
*>
(
Y
),
op_pair
->
second
);
auto
output_name
=
op_desc
.
Output
(
"Out"
)[
0
];
engine_
->
SetITensor
(
output_name
,
layer
->
getOutput
(
0
));
if
(
test_mode
)
{
// the test framework can not determine which is the
// output, so place the declaration inside.
engine_
->
DeclareOutput
(
output_name
);
}
}
protected:
static
const
std
::
unordered_map
<
std
::
string
,
nvinfer1
::
ElementWiseOperation
>
ops
;
std
::
string
op_type_
;
};
const
std
::
unordered_map
<
std
::
string
,
nvinfer1
::
ElementWiseOperation
>
ElementwiseTensorOpConverter
::
ops
=
{
{
"add"
,
nvinfer1
::
ElementWiseOperation
::
kSUM
},
{
"mul"
,
nvinfer1
::
ElementWiseOperation
::
kPROD
},
{
"sub"
,
nvinfer1
::
ElementWiseOperation
::
kSUB
},
{
"div"
,
nvinfer1
::
ElementWiseOperation
::
kDIV
},
{
"min"
,
nvinfer1
::
ElementWiseOperation
::
kMIN
},
{
"pow"
,
nvinfer1
::
ElementWiseOperation
::
kPOW
},
{
"max"
,
nvinfer1
::
ElementWiseOperation
::
kMAX
},
};
class
ElementwiseTensorAddOpConverter
:
public
ElementwiseTensorOpConverter
{
public:
ElementwiseTensorAddOpConverter
()
{
op_type_
=
"add"
;
}
};
class
ElementwiseTensorMulOpConverter
:
public
ElementwiseTensorOpConverter
{
public:
ElementwiseTensorMulOpConverter
()
{
op_type_
=
"mul"
;
}
};
class
ElementwiseTensorSubOpConverter
:
public
ElementwiseTensorOpConverter
{
public:
ElementwiseTensorSubOpConverter
()
{
op_type_
=
"sub"
;
}
};
class
ElementwiseTensorDivOpConverter
:
public
ElementwiseTensorOpConverter
{
public:
ElementwiseTensorDivOpConverter
()
{
op_type_
=
"div"
;
}
};
class
ElementwiseTensorMinOpConverter
:
public
ElementwiseTensorOpConverter
{
public:
ElementwiseTensorMinOpConverter
()
{
op_type_
=
"min"
;
}
};
class
ElementwiseTensorMaxOpConverter
:
public
ElementwiseTensorOpConverter
{
public:
ElementwiseTensorMaxOpConverter
()
{
op_type_
=
"max"
;
}
};
class
ElementwiseTensorPowOpConverter
:
public
ElementwiseTensorOpConverter
{
public:
ElementwiseTensorPowOpConverter
()
{
op_type_
=
"pow"
;
}
};
}
// namespace tensorrt
}
// namespace inference
}
// namespace paddle
REGISTER_TRT_OP_CONVERTER
(
elementwise_add_weight
,
ElementwiseWeightOpConverter
);
REGISTER_TRT_OP_CONVERTER
(
elementwise_add_tensor
,
ElementwiseTensorAddOpConverter
);
REGISTER_TRT_OP_CONVERTER
(
elementwise_sub_tensor
,
ElementwiseTensorSubOpConverter
);
REGISTER_TRT_OP_CONVERTER
(
elementwise_div_tensor
,
ElementwiseTensorDivOpConverter
);
REGISTER_TRT_OP_CONVERTER
(
elementwise_mul_tensor
,
ElementwiseTensorMulOpConverter
);
REGISTER_TRT_OP_CONVERTER
(
elementwise_max_tensor
,
ElementwiseTensorMaxOpConverter
);
REGISTER_TRT_OP_CONVERTER
(
elementwise_min_tensor
,
ElementwiseTensorMinOpConverter
);
REGISTER_TRT_OP_CONVERTER
(
elementwise_pow_tensor
,
ElementwiseTensorPowOpConverter
);
paddle/fluid/inference/tensorrt/convert/op_converter.h
浏览文件 @
d7dd0868
...
@@ -55,6 +55,31 @@ class OpConverter {
...
@@ -55,6 +55,31 @@ class OpConverter {
it
=
Registry
<
OpConverter
>::
Lookup
(
"fc"
);
it
=
Registry
<
OpConverter
>::
Lookup
(
"fc"
);
}
}
}
}
if
(
op_desc
.
Type
().
find
(
"elementwise"
)
!=
std
::
string
::
npos
)
{
static
std
::
unordered_set
<
std
::
string
>
add_tensor_op_set
{
"add"
,
"mul"
,
"sub"
,
"div"
,
"max"
,
"min"
,
"pow"
};
// TODO(xingzhaolong): all mul, sub, div
// static std::unordered_set<std::string> add_weight_op_set {"add", "mul",
// "sub", "div"};
static
std
::
unordered_set
<
std
::
string
>
add_weight_op_set
{
"add"
};
PADDLE_ENFORCE_EQ
(
op_desc
.
Input
(
"Y"
).
size
(),
1UL
);
int
op_type_len
=
op_desc
.
Type
().
size
();
std
::
string
op_type
=
op_desc
.
Type
().
substr
(
op_type_len
-
3
,
op_type_len
);
std
::
string
Y
=
op_desc
.
Input
(
"Y"
)[
0
];
if
(
parameters
.
count
(
Y
))
{
PADDLE_ENFORCE
(
add_weight_op_set
.
count
(
op_type
)
>
0
,
"Unsupported elementwise type"
+
op_type
);
it
=
Registry
<
OpConverter
>::
Lookup
(
"elementwise_"
+
op_type
+
"_weight"
);
}
else
{
PADDLE_ENFORCE
(
add_tensor_op_set
.
count
(
op_type
)
>
0
,
"Unsupported elementwise type"
+
op_type
);
it
=
Registry
<
OpConverter
>::
Lookup
(
"elementwise_"
+
op_type
+
"_tensor"
);
}
}
if
(
!
it
)
{
if
(
!
it
)
{
it
=
Registry
<
OpConverter
>::
Lookup
(
op_desc
.
Type
());
it
=
Registry
<
OpConverter
>::
Lookup
(
op_desc
.
Type
());
}
}
...
...
paddle/fluid/inference/tensorrt/convert/test_elementwise_op.cc
0 → 100644
浏览文件 @
d7dd0868
/* 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. */
#include <gtest/gtest.h>
#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
#include "paddle/fluid/inference/tensorrt/convert/ut_helper.h"
namespace
paddle
{
namespace
inference
{
namespace
tensorrt
{
TEST
(
elementwise_op
,
add_weight_test
)
{
std
::
unordered_set
<
std
::
string
>
parameters
({
"elementwise_add-Y"
});
framework
::
Scope
scope
;
TRTConvertValidation
validator
(
10
,
parameters
,
scope
,
1
<<
15
);
validator
.
DeclInputVar
(
"elementwise_add-X"
,
nvinfer1
::
DimsCHW
(
10
,
3
,
3
));
validator
.
DeclParamVar
(
"elementwise_add-Y"
,
nvinfer1
::
Dims3
(
10
,
1
,
1
));
// validator.DeclParamVar("mul-Y", nvinfer1::Dims2(8, 2));
validator
.
DeclOutputVar
(
"elementwise_add-Out"
,
nvinfer1
::
DimsCHW
(
10
,
3
,
3
));
// Prepare Op description
framework
::
OpDesc
desc
;
desc
.
SetType
(
"elementwise_add"
);
desc
.
SetInput
(
"X"
,
{
"elementwise_add-X"
});
desc
.
SetInput
(
"Y"
,
{
"elementwise_add-Y"
});
desc
.
SetOutput
(
"Out"
,
{
"elementwise_add-Out"
});
int
axis
=
1
;
desc
.
SetAttr
(
"axis"
,
axis
);
validator
.
SetOp
(
*
desc
.
Proto
());
validator
.
Execute
(
8
);
}
TEST
(
elementwise_op
,
add_tensor_test
)
{
std
::
unordered_set
<
std
::
string
>
parameters
;
framework
::
Scope
scope
;
TRTConvertValidation
validator
(
8
,
parameters
,
scope
,
1
<<
15
);
validator
.
DeclInputVar
(
"elementwise_add-X"
,
nvinfer1
::
DimsCHW
(
10
,
3
,
3
));
validator
.
DeclInputVar
(
"elementwise_add-Y"
,
nvinfer1
::
Dims3
(
10
,
3
,
3
));
// validator.DeclParamVar("mul-Y", nvinfer1::Dims2(8, 2));
validator
.
DeclOutputVar
(
"elementwise_add-Out"
,
nvinfer1
::
DimsCHW
(
10
,
3
,
3
));
// Prepare Op description
framework
::
OpDesc
desc
;
desc
.
SetType
(
"elementwise_add"
);
desc
.
SetInput
(
"X"
,
{
"elementwise_add-X"
});
desc
.
SetInput
(
"Y"
,
{
"elementwise_add-Y"
});
desc
.
SetOutput
(
"Out"
,
{
"elementwise_add-Out"
});
// the defalut axis of elementwise op is -1
validator
.
SetOp
(
*
desc
.
Proto
());
validator
.
Execute
(
8
);
}
}
// namespace tensorrt
}
// namespace inference
}
// namespace paddle
USE_OP
(
elementwise_add
);
paddle/fluid/inference/tensorrt/convert/ut_helper.h
浏览文件 @
d7dd0868
...
@@ -149,7 +149,7 @@ class TRTConvertValidation {
...
@@ -149,7 +149,7 @@ class TRTConvertValidation {
cudaStreamSynchronize
(
*
engine_
->
stream
());
cudaStreamSynchronize
(
*
engine_
->
stream
());
ASSERT_FALSE
(
op_desc_
->
OutputArgumentNames
().
empty
());
ASSERT_FALSE
(
op_desc_
->
OutputArgumentNames
().
empty
());
const
size_t
output_space_size
=
2
000
;
const
size_t
output_space_size
=
3
000
;
for
(
const
auto
&
output
:
op_desc_
->
OutputArgumentNames
())
{
for
(
const
auto
&
output
:
op_desc_
->
OutputArgumentNames
())
{
std
::
vector
<
float
>
fluid_out
;
std
::
vector
<
float
>
fluid_out
;
std
::
vector
<
float
>
trt_out
(
output_space_size
);
std
::
vector
<
float
>
trt_out
(
output_space_size
);
...
...
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