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
OPTHREE
Paddle
提交
55334007
P
Paddle
项目概览
OPTHREE
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
1
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1
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
1
Issue
1
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
55334007
编写于
7月 25, 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
fix comments
上级
0dcbeda2
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
35 addition
and
32 deletion
+35
-32
paddle/fluid/inference/tensorrt/convert/pool2d_op.cc
paddle/fluid/inference/tensorrt/convert/pool2d_op.cc
+5
-4
paddle/fluid/inference/tensorrt/convert/test_activation_op.cc
...le/fluid/inference/tensorrt/convert/test_activation_op.cc
+2
-3
paddle/fluid/inference/tensorrt/convert/test_fc_op.cc
paddle/fluid/inference/tensorrt/convert/test_fc_op.cc
+2
-3
paddle/fluid/inference/tensorrt/convert/test_mul_op.cc
paddle/fluid/inference/tensorrt/convert/test_mul_op.cc
+2
-3
paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc
paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc
+4
-6
paddle/fluid/inference/tensorrt/convert/ut_helper.h
paddle/fluid/inference/tensorrt/convert/ut_helper.h
+20
-13
未找到文件。
paddle/fluid/inference/tensorrt/convert/pool2d_op.cc
浏览文件 @
55334007
...
@@ -30,6 +30,7 @@ class Pool2dOpConverter : public OpConverter {
...
@@ -30,6 +30,7 @@ class Pool2dOpConverter : public OpConverter {
framework
::
OpDesc
op_desc
(
op
,
nullptr
);
framework
::
OpDesc
op_desc
(
op
,
nullptr
);
// Declare inputs
// Declare inputs
auto
*
input1
=
engine_
->
GetITensor
(
op_desc
.
Input
(
"X"
)[
0
]);
auto
*
input1
=
engine_
->
GetITensor
(
op_desc
.
Input
(
"X"
)[
0
]);
std
::
string
pool_type
=
std
::
string
pool_type
=
boost
::
get
<
std
::
string
>
(
op_desc
.
GetAttr
(
"pooling_type"
));
boost
::
get
<
std
::
string
>
(
op_desc
.
GetAttr
(
"pooling_type"
));
std
::
vector
<
int
>
ksize
=
std
::
vector
<
int
>
ksize
=
...
@@ -45,18 +46,18 @@ class Pool2dOpConverter : public OpConverter {
...
@@ -45,18 +46,18 @@ class Pool2dOpConverter : public OpConverter {
PADDLE_ENFORCE_EQ
(
input1
->
getDimensions
().
nbDims
,
3UL
);
PADDLE_ENFORCE_EQ
(
input1
->
getDimensions
().
nbDims
,
3UL
);
nvinfer1
::
PoolingType
pool_t
=
nvinfer1
::
PoolingType
::
kMAX
;
nvinfer1
::
PoolingType
nv_pool_type
=
nvinfer1
::
PoolingType
::
kMAX
;
if
(
pool_type
==
"max"
)
{
if
(
pool_type
==
"max"
)
{
pool_t
=
nvinfer1
::
PoolingType
::
kMAX
;
nv_pool_type
=
nvinfer1
::
PoolingType
::
kMAX
;
}
else
if
(
pool_type
==
"avg"
)
{
}
else
if
(
pool_type
==
"avg"
)
{
pool_t
=
nvinfer1
::
PoolingType
::
kAVERAGE
;
nv_pool_type
=
nvinfer1
::
PoolingType
::
kAVERAGE
;
}
else
{
}
else
{
PADDLE_THROW
(
"TensorRT unsupported pooling type!"
);
PADDLE_THROW
(
"TensorRT unsupported pooling type!"
);
}
}
auto
*
layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Pooling
,
auto
*
layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Pooling
,
*
const_cast
<
nvinfer1
::
ITensor
*>
(
input1
),
*
const_cast
<
nvinfer1
::
ITensor
*>
(
input1
),
pool_t
,
nv_ksize
);
nv_pool_type
,
nv_ksize
);
PADDLE_ENFORCE_NOT_NULL
(
layer
,
"pool layer could not be created."
);
PADDLE_ENFORCE_NOT_NULL
(
layer
,
"pool layer could not be created."
);
layer
->
setStride
(
nv_strides
);
layer
->
setStride
(
nv_strides
);
layer
->
setPadding
(
nv_paddings
);
layer
->
setPadding
(
nv_paddings
);
...
...
paddle/fluid/inference/tensorrt/convert/test_activation_op.cc
浏览文件 @
55334007
...
@@ -23,8 +23,7 @@ namespace tensorrt {
...
@@ -23,8 +23,7 @@ namespace tensorrt {
TEST
(
ReluOpConverter
,
main
)
{
TEST
(
ReluOpConverter
,
main
)
{
framework
::
Scope
scope
;
framework
::
Scope
scope
;
std
::
unordered_set
<
std
::
string
>
parameters
;
std
::
unordered_set
<
std
::
string
>
parameters
;
int
runtime_batch
=
3
;
TRTConvertValidation
validator
(
10
,
parameters
,
scope
,
1000
);
TRTConvertValidation
validator
(
10
,
parameters
,
scope
,
1000
,
runtime_batch
);
validator
.
DeclInputVar
(
"relu-X"
,
nvinfer1
::
Dims2
(
10
,
6
));
validator
.
DeclInputVar
(
"relu-X"
,
nvinfer1
::
Dims2
(
10
,
6
));
validator
.
DeclOutputVar
(
"relu-Out"
,
nvinfer1
::
Dims2
(
10
,
6
));
validator
.
DeclOutputVar
(
"relu-Out"
,
nvinfer1
::
Dims2
(
10
,
6
));
...
@@ -38,7 +37,7 @@ TEST(ReluOpConverter, main) {
...
@@ -38,7 +37,7 @@ TEST(ReluOpConverter, main) {
validator
.
SetOp
(
*
desc
.
Proto
());
validator
.
SetOp
(
*
desc
.
Proto
());
LOG
(
INFO
)
<<
"execute"
;
LOG
(
INFO
)
<<
"execute"
;
validator
.
Execute
(
runtime_batch
);
validator
.
Execute
(
5
);
}
}
}
// namespace tensorrt
}
// namespace tensorrt
...
...
paddle/fluid/inference/tensorrt/convert/test_fc_op.cc
浏览文件 @
55334007
...
@@ -23,8 +23,7 @@ namespace tensorrt {
...
@@ -23,8 +23,7 @@ namespace tensorrt {
TEST
(
fc_op
,
test
)
{
TEST
(
fc_op
,
test
)
{
std
::
unordered_set
<
std
::
string
>
parameters
({
"mul-Y"
});
std
::
unordered_set
<
std
::
string
>
parameters
({
"mul-Y"
});
framework
::
Scope
scope
;
framework
::
Scope
scope
;
int
runtime_batch
=
2
;
TRTConvertValidation
validator
(
10
,
parameters
,
scope
,
1000
);
TRTConvertValidation
validator
(
10
,
parameters
,
scope
,
1000
,
runtime_batch
);
validator
.
DeclInputVar
(
"mul-X"
,
nvinfer1
::
Dims3
(
10
,
1
,
1
));
validator
.
DeclInputVar
(
"mul-X"
,
nvinfer1
::
Dims3
(
10
,
1
,
1
));
validator
.
DeclParamVar
(
"mul-Y"
,
nvinfer1
::
Dims2
(
10
,
2
));
validator
.
DeclParamVar
(
"mul-Y"
,
nvinfer1
::
Dims2
(
10
,
2
));
validator
.
DeclOutputVar
(
"mul-Out"
,
nvinfer1
::
Dims2
(
1
,
2
));
validator
.
DeclOutputVar
(
"mul-Out"
,
nvinfer1
::
Dims2
(
1
,
2
));
...
@@ -38,7 +37,7 @@ TEST(fc_op, test) {
...
@@ -38,7 +37,7 @@ TEST(fc_op, test) {
validator
.
SetOp
(
*
desc
.
Proto
());
validator
.
SetOp
(
*
desc
.
Proto
());
validator
.
Execute
(
runtime_batch
);
validator
.
Execute
(
10
);
}
}
}
// namespace tensorrt
}
// namespace tensorrt
...
...
paddle/fluid/inference/tensorrt/convert/test_mul_op.cc
浏览文件 @
55334007
...
@@ -23,8 +23,7 @@ namespace tensorrt {
...
@@ -23,8 +23,7 @@ namespace tensorrt {
TEST
(
MulOpConverter
,
main
)
{
TEST
(
MulOpConverter
,
main
)
{
framework
::
Scope
scope
;
framework
::
Scope
scope
;
std
::
unordered_set
<
std
::
string
>
parameters
;
std
::
unordered_set
<
std
::
string
>
parameters
;
int
runtime_batch
=
0
;
TRTConvertValidation
validator
(
10
,
parameters
,
scope
,
1000
,
false
);
TRTConvertValidation
validator
(
10
,
parameters
,
scope
,
1000
,
runtime_batch
);
validator
.
DeclInputVar
(
"mul-X"
,
nvinfer1
::
Dims2
(
10
,
6
));
validator
.
DeclInputVar
(
"mul-X"
,
nvinfer1
::
Dims2
(
10
,
6
));
validator
.
DeclInputVar
(
"mul-Y"
,
nvinfer1
::
Dims2
(
6
,
10
));
validator
.
DeclInputVar
(
"mul-Y"
,
nvinfer1
::
Dims2
(
6
,
10
));
validator
.
DeclOutputVar
(
"mul-Out"
,
nvinfer1
::
Dims2
(
10
,
10
));
validator
.
DeclOutputVar
(
"mul-Out"
,
nvinfer1
::
Dims2
(
10
,
10
));
...
@@ -40,7 +39,7 @@ TEST(MulOpConverter, main) {
...
@@ -40,7 +39,7 @@ TEST(MulOpConverter, main) {
validator
.
SetOp
(
*
desc
.
Proto
());
validator
.
SetOp
(
*
desc
.
Proto
());
LOG
(
INFO
)
<<
"execute"
;
LOG
(
INFO
)
<<
"execute"
;
validator
.
Execute
(
1
);
validator
.
Execute
(
2
);
}
}
}
// namespace tensorrt
}
// namespace tensorrt
...
...
paddle/fluid/inference/tensorrt/convert/test_pool2d_op.cc
浏览文件 @
55334007
...
@@ -23,12 +23,10 @@ namespace tensorrt {
...
@@ -23,12 +23,10 @@ namespace tensorrt {
TEST
(
Pool2dOpConverter
,
main
)
{
TEST
(
Pool2dOpConverter
,
main
)
{
framework
::
Scope
scope
;
framework
::
Scope
scope
;
std
::
unordered_set
<
std
::
string
>
parameters
;
std
::
unordered_set
<
std
::
string
>
parameters
;
int
runtime_batch
=
3
;
TRTConvertValidation
validator
(
5
,
parameters
,
scope
,
1
<<
15
);
TRTConvertValidation
validator
(
5
,
parameters
,
scope
,
1
<<
15
,
runtime_batch
);
// We have already set the runtime batchsize, so the
// The ITensor's Dims should not contain the batch size.
// Dims should not contain the batch size.
// So, the ITensor's Dims of input and output should be C * H * W.
// The ITensor's Dims of input and output should be C * H * W.
validator
.
DeclInputVar
(
"pool2d-X"
,
nvinfer1
::
Dims3
(
3
,
4
,
4
));
validator
.
DeclInputVar
(
"pool2d-X"
,
nvinfer1
::
Dims3
(
3
,
4
,
4
));
validator
.
DeclOutputVar
(
"pool2d-Out"
,
nvinfer1
::
Dims3
(
3
,
2
,
2
));
validator
.
DeclOutputVar
(
"pool2d-Out"
,
nvinfer1
::
Dims3
(
3
,
2
,
2
));
...
@@ -52,7 +50,7 @@ TEST(Pool2dOpConverter, main) {
...
@@ -52,7 +50,7 @@ TEST(Pool2dOpConverter, main) {
validator
.
SetOp
(
*
desc
.
Proto
());
validator
.
SetOp
(
*
desc
.
Proto
());
LOG
(
INFO
)
<<
"execute"
;
LOG
(
INFO
)
<<
"execute"
;
validator
.
Execute
(
runtime_batch
);
validator
.
Execute
(
3
);
}
}
}
// namespace tensorrt
}
// namespace tensorrt
...
...
paddle/fluid/inference/tensorrt/convert/ut_helper.h
浏览文件 @
55334007
...
@@ -66,10 +66,11 @@ class TRTConvertValidation {
...
@@ -66,10 +66,11 @@ class TRTConvertValidation {
TRTConvertValidation
(
int
max_batch_size
,
TRTConvertValidation
(
int
max_batch_size
,
const
std
::
unordered_set
<
std
::
string
>&
parameters
,
const
std
::
unordered_set
<
std
::
string
>&
parameters
,
framework
::
Scope
&
scope
,
// NOLINT
framework
::
Scope
&
scope
,
// NOLINT
int
workspace_size
=
1
<<
10
,
int
runtime_batch_size
=
1
)
int
workspace_size
=
1
<<
10
,
bool
if_add_batch
=
true
)
:
parameters_
(
parameters
),
:
parameters_
(
parameters
),
scope_
(
scope
),
scope_
(
scope
),
runtime_batch_size_
(
runtime_batch_size
)
{
if_add_batch_
(
if_add_batch
),
max_batch_size_
(
max_batch_size
)
{
// create engine.
// create engine.
engine_
.
reset
(
new
TensorRTEngine
(
max_batch_size
,
workspace_size
,
&
stream_
));
engine_
.
reset
(
new
TensorRTEngine
(
max_batch_size
,
workspace_size
,
&
stream_
));
engine_
->
InitNetwork
();
engine_
->
InitNetwork
();
...
@@ -102,12 +103,10 @@ class TRTConvertValidation {
...
@@ -102,12 +103,10 @@ class TRTConvertValidation {
// Init Fluid tensor.
// Init Fluid tensor.
std
::
vector
<
int
>
dim_vec
(
dims
.
d
,
dims
.
d
+
dims
.
nbDims
);
std
::
vector
<
int
>
dim_vec
(
dims
.
d
,
dims
.
d
+
dims
.
nbDims
);
// There is no batchsize in ITensor's shape, but We should add it to
// There is no batchsize in ITensor's shape, but We should add it to
// tensor's
// tensor's shape of fluid. If the variable is not parameter and the
// shape of fluid. If the variable is not parameter and the batch size
// if_add_batch_ flag is true, add the max batchsize to dim_vec.
// greater than 0,
if
(
is_param
!=
true
&&
if_add_batch_
==
true
)
// add the batchsize to dim_vec.
dim_vec
.
insert
(
dim_vec
.
begin
(),
max_batch_size_
);
if
(
is_param
!=
true
&&
runtime_batch_size_
>
0
)
dim_vec
.
insert
(
dim_vec
.
begin
(),
runtime_batch_size_
);
auto
*
x
=
scope_
.
Var
(
name
);
auto
*
x
=
scope_
.
Var
(
name
);
auto
*
x_tensor
=
x
->
GetMutable
<
framework
::
LoDTensor
>
();
auto
*
x_tensor
=
x
->
GetMutable
<
framework
::
LoDTensor
>
();
x_tensor
->
Resize
(
framework
::
make_ddim
(
dim_vec
));
x_tensor
->
Resize
(
framework
::
make_ddim
(
dim_vec
));
...
@@ -141,6 +140,7 @@ class TRTConvertValidation {
...
@@ -141,6 +140,7 @@ class TRTConvertValidation {
void
Execute
(
int
batch_size
)
{
void
Execute
(
int
batch_size
)
{
// Execute Fluid Op
// Execute Fluid Op
PADDLE_ENFORCE_LE
(
batch_size
,
max_batch_size_
);
platform
::
CPUPlace
place
;
platform
::
CPUPlace
place
;
platform
::
CPUDeviceContext
ctx
(
place
);
platform
::
CPUDeviceContext
ctx
(
place
);
op_
->
Run
(
scope_
,
place
);
op_
->
Run
(
scope_
,
place
);
...
@@ -159,9 +159,14 @@ class TRTConvertValidation {
...
@@ -159,9 +159,14 @@ class TRTConvertValidation {
auto
*
var
=
scope_
.
FindVar
(
output
);
auto
*
var
=
scope_
.
FindVar
(
output
);
auto
tensor
=
var
->
GetMutable
<
framework
::
LoDTensor
>
();
auto
tensor
=
var
->
GetMutable
<
framework
::
LoDTensor
>
();
framework
::
TensorToVector
(
*
tensor
,
ctx
,
&
fluid_out
);
framework
::
TensorToVector
(
*
tensor
,
ctx
,
&
fluid_out
);
size_t
fluid_out_size
=
fluid_out
.
size
();
if
(
if_add_batch_
==
true
)
{
fluid_out_size
=
batch_size
*
(
tensor
->
dims
().
size
()
/
max_batch_size_
);
}
// Compare two output
// Compare two output
ASSERT_FALSE
(
fluid_out
.
empty
());
ASSERT_FALSE
(
fluid_out
.
empty
());
for
(
size_t
i
=
0
;
i
<
fluid_out
.
size
()
;
i
++
)
{
for
(
size_t
i
=
0
;
i
<
fluid_out
_size
;
i
++
)
{
// Loose the threshold for CI in different machine model.
// Loose the threshold for CI in different machine model.
EXPECT_LT
(
std
::
abs
(
fluid_out
[
i
]
-
trt_out
[
i
]),
2e-5
);
EXPECT_LT
(
std
::
abs
(
fluid_out
[
i
]
-
trt_out
[
i
]),
2e-5
);
}
}
...
@@ -177,10 +182,12 @@ class TRTConvertValidation {
...
@@ -177,10 +182,12 @@ class TRTConvertValidation {
std
::
unique_ptr
<
framework
::
OpDesc
>
op_desc_
;
std
::
unique_ptr
<
framework
::
OpDesc
>
op_desc_
;
const
std
::
unordered_set
<
std
::
string
>&
parameters_
;
const
std
::
unordered_set
<
std
::
string
>&
parameters_
;
framework
::
Scope
&
scope_
;
framework
::
Scope
&
scope_
;
// It represents the runtime batchsize when we test.
// The ITensor of trt does not cotain the batch size,
// If the value greater than 0, we add this to
// bug, in most cases, we need to set batch size for
// the first dimension of tensor's shape of fluid.
// fluid's tensor shape. This variable indicates
int
runtime_batch_size_
;
// whether to add batch size to tensor shape of fluid.
bool
if_add_batch_
;
int
max_batch_size_
;
};
};
}
// namespace tensorrt
}
// namespace tensorrt
...
...
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