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
940f5dbc
P
PaddleDetection
项目概览
PaddlePaddle
/
PaddleDetection
大约 1 年 前同步成功
通知
695
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
提交
940f5dbc
编写于
7月 23, 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
modify the tensorrt engine op to adapt to chage
上级
82527696
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
23 addition
and
24 deletion
+23
-24
paddle/fluid/operators/tensorrt_engine_op.cc
paddle/fluid/operators/tensorrt_engine_op.cc
+4
-3
paddle/fluid/operators/tensorrt_engine_op.h
paddle/fluid/operators/tensorrt_engine_op.h
+3
-5
paddle/fluid/operators/tensorrt_engine_op_test.cc
paddle/fluid/operators/tensorrt_engine_op_test.cc
+16
-16
未找到文件。
paddle/fluid/operators/tensorrt_engine_op.cc
浏览文件 @
940f5dbc
...
...
@@ -53,13 +53,14 @@ nvinfer1::Dims Vec2TRT_Dims(const std::vector<int64_t> &shape) {
PADDLE_ENFORCE_LE
(
shape
.
size
(),
4UL
,
"TensorRT' tensor input requires at most 4 dimensions"
);
// We should delete the batch size here.
switch
(
shape
.
size
())
{
case
2
:
return
nvinfer1
::
Dims2
(
shape
[
0
]
,
shape
[
1
]);
return
nvinfer1
::
Dims2
(
1
,
shape
[
1
]);
case
3
:
return
nvinfer1
::
Dims3
(
shape
[
0
]
,
shape
[
1
],
shape
[
2
]);
return
nvinfer1
::
Dims3
(
1
,
shape
[
1
],
shape
[
2
]);
case
4
:
return
nvinfer1
::
Dims4
(
shape
[
0
]
,
shape
[
1
],
shape
[
2
],
shape
[
3
]);
return
nvinfer1
::
Dims4
(
1
,
shape
[
1
],
shape
[
2
],
shape
[
3
]);
default:
return
nvinfer1
::
Dims
();
}
...
...
paddle/fluid/operators/tensorrt_engine_op.h
浏览文件 @
940f5dbc
...
...
@@ -95,16 +95,14 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
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
()))
{
// TODO(Superjomn) change this float to dtype size.
engine
->
GetOutputInCPU
(
y
,
fluid_t
->
mutable_data
<
float
>
(
platform
::
CPUPlace
()),
size
*
sizeof
(
float
));
y
,
fluid_t
->
mutable_data
<
float
>
(
platform
::
CPUPlace
()));
}
else
{
engine
->
GetOutputInGPU
(
y
,
fluid_t
->
mutable_data
<
float
>
(
platform
::
CUDAPlace
()),
size
*
sizeof
(
float
));
y
,
fluid_t
->
mutable_data
<
float
>
(
platform
::
CUDAPlace
()));
}
}
...
...
paddle/fluid/operators/tensorrt_engine_op_test.cc
浏览文件 @
940f5dbc
...
...
@@ -64,36 +64,37 @@ TEST(TensorRTEngineOp, manual) {
LOG
(
INFO
)
<<
"create block desc"
;
framework
::
BlockDesc
block_desc
(
&
program
,
block_
);
LOG
(
INFO
)
<<
"create
mul
op"
;
auto
*
mul
=
block_desc
.
AppendOp
();
mul
->
SetType
(
"mul"
);
mul
->
SetInput
(
"X"
,
std
::
vector
<
std
::
string
>
({
"x"
}));
// 2 x 4
mul
->
SetInput
(
"Y"
,
std
::
vector
<
std
::
string
>
({
"y"
}));
// 4 x 6
mul
->
SetOutput
(
"Out"
,
std
::
vector
<
std
::
string
>
({
"z"
}));
// 2 x 6
LOG
(
INFO
)
<<
"create
fc
op"
;
auto
*
fc0
=
block_desc
.
AppendOp
();
fc0
->
SetType
(
"mul"
);
fc0
->
SetInput
(
"X"
,
std
::
vector
<
std
::
string
>
({
"x"
}));
// 4 x 1 x 1
fc0
->
SetInput
(
"Y"
,
std
::
vector
<
std
::
string
>
({
"y"
}));
// 4 x 6
fc0
->
SetOutput
(
"Out"
,
std
::
vector
<
std
::
string
>
({
"z"
}));
// 6 x 1 x 1
LOG
(
INFO
)
<<
"create fc op"
;
auto
*
fc
=
block_desc
.
AppendOp
();
fc
->
SetType
(
"mul"
);
fc
->
SetInput
(
"X"
,
std
::
vector
<
std
::
string
>
({
"z"
}));
fc
->
SetInput
(
"Y"
,
std
::
vector
<
std
::
string
>
({
"y0"
}));
// 6 x 8
fc
->
SetOutput
(
"Out"
,
std
::
vector
<
std
::
string
>
({
"z0"
}));
// 2 x 8
auto
*
fc
1
=
block_desc
.
AppendOp
();
fc
1
->
SetType
(
"mul"
);
fc
1
->
SetInput
(
"X"
,
std
::
vector
<
std
::
string
>
({
"z"
}));
fc
1
->
SetInput
(
"Y"
,
std
::
vector
<
std
::
string
>
({
"y0"
}));
// 6 x 8
fc
1
->
SetOutput
(
"Out"
,
std
::
vector
<
std
::
string
>
({
"z0"
}));
// 8 x 1 x 1
// Set inputs' variable shape in BlockDesc
AddTensorToBlockDesc
(
block_
,
"x"
,
std
::
vector
<
int64_t
>
({
2
,
4
}));
// the batch size is 2, so the dims of 'x' is {2, 4, 1, 1}
AddTensorToBlockDesc
(
block_
,
"x"
,
std
::
vector
<
int64_t
>
({
2
,
4
,
1
,
1
}));
AddTensorToBlockDesc
(
block_
,
"y"
,
std
::
vector
<
int64_t
>
({
4
,
6
}));
AddTensorToBlockDesc
(
block_
,
"y0"
,
std
::
vector
<
int64_t
>
({
6
,
8
}));
AddTensorToBlockDesc
(
block_
,
"z"
,
std
::
vector
<
int64_t
>
({
2
,
6
}));
// It is wired, need to copy manually.
*
block_
->
add_ops
()
=
*
mul
->
Proto
();
*
block_
->
add_ops
()
=
*
fc
->
Proto
();
*
block_
->
add_ops
()
=
*
fc0
->
Proto
();
*
block_
->
add_ops
()
=
*
fc
1
->
Proto
();
ASSERT_EQ
(
block_
->
ops_size
(),
2
);
LOG
(
INFO
)
<<
"create tensorrt desc"
;
framework
::
OpDesc
engine_op_desc
(
nullptr
);
engine_op_desc
.
SetType
(
"tensorrt_engine"
);
engine_op_desc
.
SetInput
(
"Xs"
,
std
::
vector
<
std
::
string
>
({
"x"
,
"y"
,
"y0"
}));
engine_op_desc
.
SetInput
(
"Xs"
,
std
::
vector
<
std
::
string
>
({
"x"
}));
engine_op_desc
.
SetOutput
(
"Ys"
,
std
::
vector
<
std
::
string
>
({
"z0"
}));
SetAttr
<
std
::
string
>
(
engine_op_desc
.
Proto
(),
"subgraph"
,
block_
->
SerializeAsString
());
...
...
@@ -208,4 +209,3 @@ TEST(TensorRTEngineOp, fc) { Execute(40, 28, 28); }
}
// namespace paddle
USE_TRT_CONVERTER
(
mul
)
USE_TRT_CONVERTER
(
fc
)
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