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
wzpei9
Paddle
提交
6e13c86d
P
Paddle
项目概览
wzpei9
/
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
6e13c86d
编写于
5月 23, 2018
作者:
Y
Yibing Liu
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Enable multiple groups for cudnn conv transpose
上级
669c0df6
变更
2
显示空白变更内容
内联
并排
Showing
2 changed file
with
54 addition
and
19 deletion
+54
-19
paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc
paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc
+39
-19
python/paddle/fluid/tests/unittests/test_conv2d_transpose_op.py
.../paddle/fluid/tests/unittests/test_conv2d_transpose_op.py
+15
-0
未找到文件。
paddle/fluid/operators/conv_transpose_cudnn_op.cu.cc
浏览文件 @
6e13c86d
...
...
@@ -44,6 +44,7 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel<T> {
std
::
vector
<
int
>
paddings
=
ctx
.
Attr
<
std
::
vector
<
int
>>
(
"paddings"
);
// cudnn v5 does not support dilations
std
::
vector
<
int
>
dilations
=
ctx
.
Attr
<
std
::
vector
<
int
>>
(
"dilations"
);
int
groups
=
ctx
.
Attr
<
int
>
(
"groups"
);
int
user_workspace_size
=
ctx
.
Attr
<
int
>
(
"workspace_size_MB"
);
const
T
*
input_data
=
input
->
data
<
T
>
();
...
...
@@ -64,13 +65,13 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel<T> {
// (N, M, H, W) or (N, M, D, H, W)
cudnnTensorDescriptor_t
cudnn_input_desc
=
input_desc
.
descriptor
<
T
>
(
layout
,
framework
::
vectorize2int
(
input
->
dims
()));
layout
,
framework
::
vectorize2int
(
input
->
dims
())
,
groups
);
// (N, C, O_h, O_w) or (N, C, O_d, O_h, O_w)
cudnnTensorDescriptor_t
cudnn_output_desc
=
output_desc
.
descriptor
<
T
>
(
layout
,
framework
::
vectorize2int
(
output
->
dims
()));
layout
,
framework
::
vectorize2int
(
output
->
dims
())
,
groups
);
// (M, C, K_h, K_w) or (M, C, K_d, K_h, K_w)
cudnnFilterDescriptor_t
cudnn_filter_desc
=
filter_desc
.
descriptor
<
T
>
(
layout
,
framework
::
vectorize2int
(
filter
->
dims
()));
layout
,
framework
::
vectorize2int
(
filter
->
dims
())
,
groups
);
cudnnConvolutionDescriptor_t
cudnn_conv_desc
=
conv_desc
.
descriptor
<
T
>
(
paddings
,
strides
,
dilations
);
...
...
@@ -104,11 +105,17 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel<T> {
cudnn_workspace
=
paddle
::
memory
::
Alloc
(
gpu
,
workspace_size_in_bytes
);
// ------------------- cudnn conv transpose forward ---------------------
int
input_offset
=
input
->
numel
()
/
input
->
dims
()[
0
]
/
groups
;
int
output_offset
=
output
->
numel
()
/
output
->
dims
()[
0
]
/
groups
;
int
filter_offset
=
filter
->
numel
()
/
groups
;
T
alpha
=
1.0
f
,
beta
=
0.0
f
;
for
(
int
g
=
0
;
g
<
groups
;
g
++
)
{
PADDLE_ENFORCE
(
platform
::
dynload
::
cudnnConvolutionBackwardData
(
handle
,
&
alpha
,
cudnn_filter_desc
,
filter_data
,
cudnn_input_desc
,
input_data
,
cudnn_conv_desc
,
algo
,
cudnn_workspace
,
workspace_size_in_bytes
,
&
beta
,
cudnn_output_desc
,
output_data
));
handle
,
&
alpha
,
cudnn_filter_desc
,
filter_data
+
filter_offset
*
g
,
cudnn_input_desc
,
input_data
+
input_offset
*
g
,
cudnn_conv_desc
,
algo
,
cudnn_workspace
,
workspace_size_in_bytes
,
&
beta
,
cudnn_output_desc
,
output_data
+
output_offset
*
g
));
}
// Release the cudnn workspace
paddle
::
memory
::
Free
(
gpu
,
cudnn_workspace
);
...
...
@@ -134,6 +141,7 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel<T> {
std
::
vector
<
int
>
paddings
=
ctx
.
Attr
<
std
::
vector
<
int
>>
(
"paddings"
);
// cudnn v5 does not support dilations
std
::
vector
<
int
>
dilations
=
ctx
.
Attr
<
std
::
vector
<
int
>>
(
"dilations"
);
int
groups
=
ctx
.
Attr
<
int
>
(
"groups"
);
int
user_workspace_size
=
ctx
.
Attr
<
int
>
(
"workspace_size_MB"
);
// ------------------- cudnn descriptors ---------------------
...
...
@@ -145,13 +153,13 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel<T> {
// Input: (N, M, H, W) or (N, M, D, H, W)
cudnnTensorDescriptor_t
cudnn_input_desc
=
input_desc
.
descriptor
<
T
>
(
layout
,
framework
::
vectorize2int
(
input
->
dims
()));
layout
,
framework
::
vectorize2int
(
input
->
dims
())
,
groups
);
// Output: (N, C, O_h, O_w) or (N, C, O_d, O_h, O_w)
cudnnTensorDescriptor_t
cudnn_output_desc
=
output_desc
.
descriptor
<
T
>
(
layout
,
framework
::
vectorize2int
(
output_grad
->
dims
()));
layout
,
framework
::
vectorize2int
(
output_grad
->
dims
())
,
groups
);
// Filter (M, C, K_h, K_w) or (M, C, K_d K_h, K_w)
cudnnFilterDescriptor_t
cudnn_filter_desc
=
filter_desc
.
descriptor
<
T
>
(
layout
,
framework
::
vectorize2int
(
filter
->
dims
()));
layout
,
framework
::
vectorize2int
(
filter
->
dims
())
,
groups
);
cudnnConvolutionDescriptor_t
cudnn_conv_desc
=
conv_desc
.
descriptor
<
T
>
(
paddings
,
strides
,
dilations
);
...
...
@@ -205,15 +213,22 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel<T> {
cudnn_workspace
=
paddle
::
memory
::
Alloc
(
gpu
,
workspace_size_in_bytes
);
// ------------------- cudnn conv backward data ---------------------
// FIXME(typhoonzero): template type T may not be the same as cudnn call.
int
input_offset
=
input
->
numel
()
/
input
->
dims
()[
0
]
/
groups
;
int
output_grad_offset
=
output_grad
->
numel
()
/
output_grad
->
dims
()[
0
]
/
groups
;
int
filter_offset
=
filter
->
numel
()
/
groups
;
T
alpha
=
1.0
f
,
beta
=
0.0
f
;
if
(
input_grad
)
{
T
*
input_grad_data
=
input_grad
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
// Because beta is zero, it is unnecessary to reset input_grad.
for
(
int
g
=
0
;
g
<
groups
;
g
++
)
{
PADDLE_ENFORCE
(
platform
::
dynload
::
cudnnConvolutionForward
(
handle
,
&
alpha
,
cudnn_output_desc
,
output_grad_data
,
cudnn_filter_desc
,
filter_data
,
cudnn_conv_desc
,
data_algo
,
handle
,
&
alpha
,
cudnn_output_desc
,
output_grad_data
+
output_grad_offset
*
g
,
cudnn_filter_desc
,
filter_data
+
filter_offset
*
g
,
cudnn_conv_desc
,
data_algo
,
cudnn_workspace
,
workspace_size_in_bytes
,
&
beta
,
cudnn_input_desc
,
input_grad_data
));
input_grad_data
+
input_offset
*
g
));
}
}
// ------------------- cudnn conv backward filter ---------------------
...
...
@@ -221,11 +236,16 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel<T> {
T
*
filter_grad_data
=
filter_grad
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
// Because beta is zero, it is unnecessary to reset filter_grad.
// Gradient with respect to the filter
for
(
int
g
=
0
;
g
<
groups
;
g
++
)
{
PADDLE_ENFORCE
(
platform
::
dynload
::
cudnnConvolutionBackwardFilter
(
handle
,
&
alpha
,
cudnn_output_desc
,
output_grad_data
,
cudnn_input_desc
,
input_data
,
cudnn_conv_desc
,
filter_algo
,
cudnn_workspace
,
workspace_size_in_bytes
,
&
beta
,
cudnn_filter_desc
,
filter_grad_data
));
handle
,
&
alpha
,
cudnn_output_desc
,
output_grad_data
+
output_grad_offset
*
g
,
cudnn_input_desc
,
input_data
+
input_offset
*
g
,
cudnn_conv_desc
,
filter_algo
,
cudnn_workspace
,
workspace_size_in_bytes
,
&
beta
,
cudnn_filter_desc
,
filter_grad_data
+
filter_offset
*
g
));
}
}
// Release the cudnn workspace
paddle
::
memory
::
Free
(
gpu
,
cudnn_workspace
);
}
...
...
python/paddle/fluid/tests/unittests/test_conv2d_transpose_op.py
浏览文件 @
6e13c86d
...
...
@@ -227,6 +227,21 @@ class TestCUDNNWithStride(TestWithStride):
self
.
op_type
=
"conv2d_transpose"
class
TestCUDNNWithGroups
(
TestWithGroups
):
def
init_test_case
(
self
):
self
.
pad
=
[
1
,
1
]
self
.
stride
=
[
1
,
1
]
self
.
dilations
=
[
1
,
1
]
self
.
groups
=
2
self
.
input_size
=
[
2
,
4
,
5
,
5
]
# NCHW
f_c
=
self
.
input_size
[
1
]
self
.
filter_size
=
[
f_c
,
3
,
3
,
3
]
def
init_op_type
(
self
):
self
.
use_cudnn
=
True
self
.
op_type
=
"conv2d_transpose"
# Please Don't remove the following code.
# Currently, CI use cudnn V5.0 which not support dilation conv.
# class TestCUDNNWithDilation(TestWithDilation):
...
...
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