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
d77c4955
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2297
Star
20931
Fork
5422
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
体验新版 GitCode,发现更多精彩内容 >>
未验证
提交
d77c4955
编写于
6月 24, 2022
作者:
Z
zhoutianzi666
提交者:
GitHub
6月 24, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[Inference] rewrite elementwise trt layer (#43615)
* rewrite elementwise
上级
2bef8a48
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
116 addition
and
232 deletion
+116
-232
paddle/fluid/inference/tensorrt/convert/elementwise_op.cc
paddle/fluid/inference/tensorrt/convert/elementwise_op.cc
+94
-232
paddle/fluid/inference/tensorrt/op_teller.cc
paddle/fluid/inference/tensorrt/op_teller.cc
+22
-0
未找到文件。
paddle/fluid/inference/tensorrt/convert/elementwise_op.cc
浏览文件 @
d77c4955
...
...
@@ -19,228 +19,115 @@ namespace paddle {
namespace
inference
{
namespace
tensorrt
{
static
bool
CheckDims
(
const
nvinfer1
::
Dims
&
dims_x
,
const
nvinfer1
::
Dims
&
dims_y
)
{
if
(
dims_x
.
nbDims
!=
dims_y
.
nbDims
)
{
return
false
;
}
for
(
int
i
=
0
;
i
<
dims_x
.
nbDims
;
i
++
)
{
if
(
dims_x
.
d
[
i
]
!=
dims_y
.
d
[
i
])
{
return
false
;
}
}
return
true
;
}
class
ElementwiseWeightOpConverter
:
public
OpConverter
{
class
ElementwiseTensorOpConverter
:
public
OpConverter
{
public:
Elementwise
Weight
OpConverter
()
{}
Elementwise
Tensor
OpConverter
()
{}
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.
nvinfer1
::
ILayer
*
layer
=
nullptr
;
const
framework
::
Scope
&
scope
,
bool
test_mode
)
override
{
VLOG
(
3
)
<<
"Convert a fluid elementwise op to TensorRT IElementWiseLayer"
;
framework
::
OpDesc
op_desc
(
op
,
nullptr
);
VLOG
(
3
)
<<
"Convert a fluid elementwise op to TensorRT IScaleLayer"
;
auto
*
X
=
engine_
->
GetITensor
(
op_desc
.
Input
(
"X"
).
front
());
nvinfer1
::
ITensor
*
Y
=
nullptr
;
auto
*
Y_v
=
scope
.
FindVar
(
op_desc
.
Input
(
"Y"
).
front
());
PADDLE_ENFORCE_NOT_NULL
(
Y_v
,
platform
::
errors
::
NotFound
(
"Variable %s not found in scope."
,
op_desc
.
Input
(
"Y"
).
front
().
c_str
()));
auto
*
Y_t
=
Y_v
->
GetMutable
<
framework
::
LoDTensor
>
();
float
*
weight_data
=
nullptr
;
auto
output_name
=
op_desc
.
Output
(
"Out"
)[
0
];
weight_data
=
engine_
->
GetWeightCPUData
(
op_desc
.
Input
(
"Y"
).
front
(),
Y_t
);
if
(
Y_v
)
{
// Y is weight
auto
*
Y_t
=
Y_v
->
GetMutable
<
framework
::
LoDTensor
>
();
float
*
weight_data
=
engine_
->
GetWeightCPUData
(
op_desc
.
Input
(
"Y"
).
front
(),
Y_t
);
std
::
vector
<
int
>
dims_y
=
phi
::
vectorize
<
int
>
(
Y_t
->
dims
());
TensorRTEngine
::
Weight
y_weight
{
nvinfer1
::
DataType
::
kFLOAT
,
static_cast
<
void
*>
(
weight_data
),
static_cast
<
size_t
>
(
Y_t
->
numel
())};
nvinfer1
::
Dims
trt_dims_y
;
trt_dims_y
.
nbDims
=
dims_y
.
size
();
for
(
int
i
=
0
;
i
<
trt_dims_y
.
nbDims
;
i
++
)
{
trt_dims_y
.
d
[
i
]
=
dims_y
[
i
];
}
Y
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Constant
,
trt_dims_y
,
y_weight
.
get
())
->
getOutput
(
0
);
}
else
{
Y
=
engine_
->
GetITensor
(
op_desc
.
Input
(
"Y"
).
front
());
}
if
(
X
->
getDimensions
().
nbDims
<
Y
->
getDimensions
().
nbDims
)
{
auto
*
tmp
=
X
;
X
=
Y
;
Y
=
tmp
;
}
nvinfer1
::
Dims
dims_x
=
X
->
getDimensions
();
std
::
vector
<
int
>
dims_y
=
phi
::
vectorize
<
int
>
(
Y_t
->
dims
());
nvinfer1
::
Dims
dims_y
=
Y
->
getDimensions
();
auto
output_name
=
op_desc
.
Output
(
"Out"
)[
0
];
auto
regist_eltwise_weight
=
[
&
](
nvinfer1
::
ScaleMode
scale_mode
)
{
nvinfer1
::
IShuffleLayer
*
expand_layer
=
nullptr
;
nvinfer1
::
IShuffleLayer
*
squeeze_layer
=
nullptr
;
int
dynamic_shape_offset
=
engine_
->
with_dynamic_shape
()
?
1
:
0
;
auto
input_dim
=
X
->
getDimensions
();
// reshape
if
(
input_dim
.
nbDims
<
3
+
dynamic_shape_offset
)
{
nvinfer1
::
Dims
expand_shape
;
expand_shape
.
nbDims
=
3
+
dynamic_shape_offset
;
for
(
int
i
=
0
;
i
<
expand_shape
.
nbDims
;
i
++
)
{
if
(
i
<
input_dim
.
nbDims
)
{
expand_shape
.
d
[
i
]
=
input_dim
.
d
[
i
]
<
0
?
0
:
input_dim
.
d
[
i
];
}
else
{
expand_shape
.
d
[
i
]
=
1
;
}
}
expand_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
X
);
expand_layer
->
setReshapeDimensions
(
expand_shape
);
X
=
expand_layer
->
getOutput
(
0
);
expand_layer
->
getOutput
(
0
)
->
setName
(
(
"elementwise_reshape_out: "
+
output_name
).
c_str
());
expand_layer
->
setName
(
(
"Elewise: Shuffle: (Output: "
+
output_name
+
")"
).
c_str
());
}
// eltwise_ops
TensorRTEngine
::
Weight
shift_weights
{
nvinfer1
::
DataType
::
kFLOAT
,
nullptr
,
0
};
TensorRTEngine
::
Weight
scale_weights
{
nvinfer1
::
DataType
::
kFLOAT
,
nullptr
,
0
};
TensorRTEngine
::
Weight
power_weights
{
nvinfer1
::
DataType
::
kFLOAT
,
nullptr
,
0
};
if
(
op_type_
==
"add"
)
{
shift_weights
=
TensorRTEngine
::
Weight
(
nvinfer1
::
DataType
::
kFLOAT
,
static_cast
<
void
*>
(
weight_data
),
static_cast
<
size_t
>
(
Y_t
->
numel
()));
}
else
if
(
op_type_
==
"sub"
)
{
for
(
int
i
=
0
;
i
<
Y_t
->
numel
();
i
++
)
{
weight_data
[
i
]
=
-
weight_data
[
i
];
}
shift_weights
=
TensorRTEngine
::
Weight
(
nvinfer1
::
DataType
::
kFLOAT
,
static_cast
<
void
*>
(
weight_data
),
static_cast
<
size_t
>
(
Y_t
->
numel
()));
}
else
if
(
op_type_
==
"mul"
)
{
scale_weights
=
TensorRTEngine
::
Weight
(
nvinfer1
::
DataType
::
kFLOAT
,
static_cast
<
void
*>
(
weight_data
),
static_cast
<
size_t
>
(
Y_t
->
numel
()));
}
else
if
(
op_type_
==
"div"
)
{
for
(
int
i
=
0
;
i
<
Y_t
->
numel
();
i
++
)
{
weight_data
[
i
]
=
1.
f
/
weight_data
[
i
];
// axis here is relative to explicit batch
int
axis
=
BOOST_GET_CONST
(
int
,
op_desc
.
GetAttr
(
"axis"
));
int
real_x_rank
=
dims_x
.
nbDims
;
int
real_y_rank
=
dims_y
.
nbDims
;
if
(
!
engine_
->
with_dynamic_shape
())
{
real_x_rank
++
;
real_y_rank
++
;
if
(
Y_v
)
real_y_rank
--
;
}
if
(
axis
==
-
1
)
{
axis
=
real_x_rank
-
real_y_rank
;
}
if
(
!
engine_
->
with_dynamic_shape
()
&&
axis
>
0
)
{
axis
--
;
}
// X: - - - - - - -
// axis
// Y: - - -
// we need expand Y's rank = X's rank
int
left_one_num
=
axis
;
int
right_one_num
=
dims_x
.
nbDims
-
axis
-
dims_y
.
nbDims
;
nvinfer1
::
IShuffleLayer
*
reshape_layer
;
nvinfer1
::
ITensor
*
reshape_y_tensor
;
if
(
left_one_num
>
0
||
right_one_num
>
0
)
{
if
(
engine_
->
with_dynamic_shape
())
{
auto
*
y_shape_tensor
=
Shape
(
Y
);
auto
*
new_y_shape_tensor
=
y_shape_tensor
;
if
(
axis
>
0
)
{
std
::
vector
<
int32_t
>
left_one
(
left_one_num
,
1
);
auto
*
left_one_tensor
=
Add1DConstantLayer
(
left_one
);
new_y_shape_tensor
=
Concat
(
std
::
vector
<
nvinfer1
::
ITensor
*>
{
left_one_tensor
,
new_y_shape_tensor
});
}
scale_weights
=
TensorRTEngine
::
Weight
(
nvinfer1
::
DataType
::
kFLOAT
,
static_cast
<
void
*>
(
weight_data
),
static_cast
<
size_t
>
(
Y_t
->
numel
()));
}
else
if
(
op_type_
==
"pow"
)
{
power_weights
=
TensorRTEngine
::
Weight
(
nvinfer1
::
DataType
::
kFLOAT
,
static_cast
<
void
*>
(
weight_data
),
static_cast
<
size_t
>
(
Y_t
->
numel
()));
}
nvinfer1
::
IScaleLayer
*
scale_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
ScaleNd
,
*
X
,
scale_mode
,
shift_weights
.
get
(),
scale_weights
.
get
(),
power_weights
.
get
(),
dynamic_shape_offset
);
layer
=
scale_layer
;
// reshape
if
(
input_dim
.
nbDims
<
3
+
dynamic_shape_offset
)
{
nvinfer1
::
Dims
squeeze_shape
;
squeeze_shape
.
nbDims
=
input_dim
.
nbDims
;
for
(
int
i
=
0
;
i
<
squeeze_shape
.
nbDims
;
i
++
)
{
squeeze_shape
.
d
[
i
]
=
input_dim
.
d
[
i
]
<
0
?
0
:
input_dim
.
d
[
i
];
if
(
right_one_num
>
0
)
{
std
::
vector
<
int32_t
>
right_one
(
right_one_num
,
1
);
auto
*
right_one_tensor
=
Add1DConstantLayer
(
right_one
);
new_y_shape_tensor
=
Concat
(
std
::
vector
<
nvinfer1
::
ITensor
*>
{
new_y_shape_tensor
,
right_one_tensor
});
}
squeeze_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
(
layer
->
getOutput
(
0
)));
squeeze_layer
->
setReshapeDimensions
(
squeeze_shape
);
RreplenishLayerAndOutput
(
squeeze_layer
,
"elementwise_"
+
op_type_
,
{
output_name
},
test_mode
);
reshape_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
Y
);
reshape_layer
->
setInput
(
1
,
*
new_y_shape_tensor
);
}
else
{
RreplenishLayerAndOutput
(
layer
,
"elementwise_"
+
op_type_
,
{
output_name
},
test_mode
);
nvinfer1
::
Dims
new_y_dims
;
new_y_dims
.
nbDims
=
left_one_num
+
dims_y
.
nbDims
+
right_one_num
;
for
(
int
i
=
0
;
i
<
new_y_dims
.
nbDims
;
i
++
)
new_y_dims
.
d
[
i
]
=
1
;
for
(
int
i
=
0
;
i
<
dims_y
.
nbDims
;
i
++
)
new_y_dims
.
d
[
left_one_num
+
i
]
=
dims_y
.
d
[
i
];
reshape_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
Y
);
reshape_layer
->
setReshapeDimensions
(
new_y_dims
);
}
};
// dynamic shape
if
(
engine_
->
with_dynamic_shape
())
{
if
(
dims_y
.
size
()
==
1
&&
dims_y
[
0
]
==
dims_x
.
d
[
1
])
{
regist_eltwise_weight
(
nvinfer1
::
ScaleMode
::
kCHANNEL
);
}
else
if
(
dims_y
.
size
()
==
1
&&
dims_y
[
0
]
==
1
)
{
regist_eltwise_weight
(
nvinfer1
::
ScaleMode
::
kUNIFORM
);
}
else
if
(
dims_y
.
size
()
==
static_cast
<
size_t
>
(
dims_x
.
nbDims
))
{
regist_eltwise_weight
(
nvinfer1
::
ScaleMode
::
kELEMENTWISE
);
}
else
{
PADDLE_THROW
(
platform
::
errors
::
InvalidArgument
(
"The size of input_y's dims is %d, but TensorRT dynamic shape "
"only support size = 1 or size = input_x.size() for Elementwise "
"op!"
,
dims_y
.
size
()));
}
return
;
}
// static shape with dynamic batch
std
::
vector
<
int
>
no_batch_dims
;
int
start_index
=
0
;
for
(;
start_index
<
dims_x
.
nbDims
;
start_index
++
)
{
no_batch_dims
.
push_back
(
dims_x
.
d
[
start_index
]);
}
if
(
dims_y
.
size
()
==
1
&&
dims_y
[
0
]
==
no_batch_dims
[
0
])
{
regist_eltwise_weight
(
nvinfer1
::
ScaleMode
::
kCHANNEL
);
}
else
if
(
dims_y
.
size
()
==
1
&&
dims_y
[
0
]
==
1
)
{
regist_eltwise_weight
(
nvinfer1
::
ScaleMode
::
kUNIFORM
);
}
else
if
(
dims_y
.
size
()
==
no_batch_dims
.
size
()
+
1
)
{
regist_eltwise_weight
(
nvinfer1
::
ScaleMode
::
kELEMENTWISE
);
reshape_y_tensor
=
reshape_layer
->
getOutput
(
0
);
}
else
{
PADDLE_THROW
(
platform
::
errors
::
InvalidArgument
(
"The size of input_y's dims is %d, but TensorRT dynamic shape "
"only support size = 1 or size = input_x.size() for Elementwise "
"op!"
,
dims_y
.
size
()));
// In fact , we can remove this `else`, but -> rt_resnet50_test CI in trt
// 6015 faling, how ridiculous!
reshape_y_tensor
=
Y
;
}
}
protected:
std
::
string
op_type_
;
};
class
ElementwiseTensorOpConverter
:
public
OpConverter
{
public:
ElementwiseTensorOpConverter
()
{}
void
operator
()(
const
framework
::
proto
::
OpDesc
&
op
,
const
framework
::
Scope
&
scope
,
bool
test_mode
)
override
{
auto
op_pair
=
ops
.
find
(
op_type_
);
PADDLE_ENFORCE_NE
(
op_pair
,
ops
.
end
(),
PADDLE_ENFORCE_NE
(
op_pair
,
ops
.
end
(),
platform
::
errors
::
InvalidArgument
(
"Elementwise op's type(%s) is not supported. Please "
"check if the op_type is correct."
,
op_type_
));
// Here the two nullptr looks strange, that's because the
// framework::OpDesc's constructor is strange.
framework
::
OpDesc
op_desc
(
op
,
nullptr
);
nvinfer1
::
ILayer
*
layer
=
nullptr
;
auto
*
X
=
engine_
->
GetITensor
(
op_desc
.
Input
(
"X"
).
front
());
auto
*
Y
=
engine_
->
GetITensor
(
op_desc
.
Input
(
"Y"
).
front
());
std
::
vector
<
nvinfer1
::
ITensor
*>
itensors
;
itensors
.
push_back
(
X
);
itensors
.
push_back
(
Y
);
nvinfer1
::
Dims
dims_x
=
X
->
getDimensions
();
nvinfer1
::
Dims
dims_y
=
Y
->
getDimensions
();
int
axis
=
BOOST_GET_CONST
(
int
,
op_desc
.
GetAttr
(
"axis"
));
auto
output_name
=
op_desc
.
Output
(
"Out"
)[
0
];
auto
common_func
=
[
&
](
nvinfer1
::
ILayer
*
layer
)
{
RreplenishLayerAndOutput
(
layer
,
"elementwise"
,
{
output_name
},
test_mode
);
};
if
(
dims_x
.
nbDims
==
dims_y
.
nbDims
)
{
// The two input tensor should have the same dims
VLOG
(
3
)
<<
"Convert a fluid elementwise op to TensorRT IElementWiseLayer"
;
nvinfer1
::
IElementWiseLayer
*
elet_layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
ElementWise
,
*
X
,
*
Y
,
op_pair
->
second
);
layer
=
elet_layer
;
}
else
{
VLOG
(
3
)
<<
"Convert a fluid elementwise op to TensorRT "
"ElementWisePluginLayer"
;
if
(
engine_
->
with_dynamic_shape
())
{
#if IS_TRT_VERSION_GE(6000)
plugin
::
ElementwisePluginDynamic
*
plugin
=
new
plugin
::
ElementwisePluginDynamic
(
op_type_
,
axis
);
layer
=
engine_
->
AddDynamicPlugin
(
itensors
.
data
(),
2
,
plugin
);
#else
PADDLE_THROW
(
platform
::
errors
::
Fatal
(
"You are running the TRT Dynamic Shape mode, need to confirm that "
"your TRT version is no less than 6.0"
));
#endif
}
else
{
plugin
::
ElementWisePlugin
*
plugin
=
new
plugin
::
ElementWisePlugin
(
op_type_
,
dims_x
,
dims_y
,
axis
);
std
::
vector
<
nvinfer1
::
ITensor
*>
inputs
{
X
,
Y
};
auto
*
plugin_layer
=
engine_
->
AddPlugin
(
inputs
.
data
(),
inputs
.
size
(),
reinterpret_cast
<
plugin
::
PluginTensorRT
*>
(
plugin
));
layer
=
plugin_layer
;
}
}
common_func
(
layer
);
auto
*
layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
ElementWise
,
*
X
,
*
reshape_y_tensor
,
op_pair
->
second
);
RreplenishLayerAndOutput
(
layer
,
"elementwise"
,
{
output_name
},
test_mode
);
}
protected:
...
...
@@ -260,31 +147,6 @@ const std::unordered_map<std::string, nvinfer1::ElementWiseOperation>
{
"max"
,
nvinfer1
::
ElementWiseOperation
::
kMAX
},
};
class
ElementwiseWeightAddOpConverter
:
public
ElementwiseWeightOpConverter
{
public:
ElementwiseWeightAddOpConverter
()
{
op_type_
=
"add"
;
}
};
class
ElementwiseWeightMulOpConverter
:
public
ElementwiseWeightOpConverter
{
public:
ElementwiseWeightMulOpConverter
()
{
op_type_
=
"mul"
;
}
};
class
ElementwiseWeightSubOpConverter
:
public
ElementwiseWeightOpConverter
{
public:
ElementwiseWeightSubOpConverter
()
{
op_type_
=
"sub"
;
}
};
class
ElementwiseWeightDivOpConverter
:
public
ElementwiseWeightOpConverter
{
public:
ElementwiseWeightDivOpConverter
()
{
op_type_
=
"div"
;
}
};
class
ElementwiseWeightPowOpConverter
:
public
ElementwiseWeightOpConverter
{
public:
ElementwiseWeightPowOpConverter
()
{
op_type_
=
"pow"
;
}
};
class
ElementwiseTensorAddOpConverter
:
public
ElementwiseTensorOpConverter
{
public:
ElementwiseTensorAddOpConverter
()
{
op_type_
=
"add"
;
}
...
...
@@ -325,15 +187,15 @@ class ElementwiseTensorPowOpConverter : public ElementwiseTensorOpConverter {
}
// namespace paddle
REGISTER_TRT_OP_CONVERTER
(
elementwise_add_weight
,
Elementwise
Weight
AddOpConverter
);
Elementwise
Tensor
AddOpConverter
);
REGISTER_TRT_OP_CONVERTER
(
elementwise_mul_weight
,
Elementwise
Weight
MulOpConverter
);
Elementwise
Tensor
MulOpConverter
);
REGISTER_TRT_OP_CONVERTER
(
elementwise_sub_weight
,
Elementwise
Weight
SubOpConverter
);
Elementwise
Tensor
SubOpConverter
);
REGISTER_TRT_OP_CONVERTER
(
elementwise_div_weight
,
Elementwise
Weight
DivOpConverter
);
Elementwise
Tensor
DivOpConverter
);
REGISTER_TRT_OP_CONVERTER
(
elementwise_pow_weight
,
Elementwise
Weight
PowOpConverter
);
Elementwise
Tensor
PowOpConverter
);
REGISTER_TRT_OP_CONVERTER
(
elementwise_add_tensor
,
ElementwiseTensorAddOpConverter
);
...
...
paddle/fluid/inference/tensorrt/op_teller.cc
浏览文件 @
d77c4955
...
...
@@ -1256,6 +1256,28 @@ bool OpTeller::Tell(const framework::ir::Node* node,
}
}
}
// not support following four inputs for slice in paddle-trt
auto
slice_inputs
=
desc
.
Inputs
();
// its size == 5
if
(
slice_inputs
.
find
(
"StartsTensor"
)
!=
slice_inputs
.
end
())
{
if
(
desc
.
Input
(
"StartsTensor"
).
size
())
{
return
false
;
}
}
if
(
slice_inputs
.
find
(
"EndsTensor"
)
!=
slice_inputs
.
end
())
{
if
(
desc
.
Input
(
"EndsTensor"
).
size
())
{
return
false
;
}
}
if
(
slice_inputs
.
find
(
"StartsTensorList"
)
!=
slice_inputs
.
end
())
{
if
(
desc
.
Input
(
"StartsTensorList"
).
size
())
{
return
false
;
}
}
if
(
slice_inputs
.
find
(
"EndsTensorList"
)
!=
slice_inputs
.
end
())
{
if
(
desc
.
Input
(
"EndsTensorList"
).
size
())
{
return
false
;
}
}
}
if
(
op_type
==
"elementwise_add"
||
op_type
==
"elementwise_mul"
||
...
...
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