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
d865b047
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 个月 前同步成功
通知
2292
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 >>
提交
d865b047
编写于
9月 20, 2017
作者:
Q
qingqing01
提交者:
GitHub
9月 20, 2017
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #4201 from qingqing01/fix_prelu
Refine platform::Transform function and fix prelu_op testing.
上级
a9202e89
2aa4d326
变更
4
显示空白变更内容
内联
并排
Showing
4 changed file
with
74 addition
and
40 deletion
+74
-40
paddle/operators/prelu_op.h
paddle/operators/prelu_op.h
+6
-4
paddle/platform/transform.h
paddle/platform/transform.h
+55
-29
paddle/platform/transform_test.cu
paddle/platform/transform_test.cu
+11
-5
python/paddle/v2/framework/tests/test_prelu_op.py
python/paddle/v2/framework/tests/test_prelu_op.py
+2
-2
未找到文件。
paddle/operators/prelu_op.h
浏览文件 @
d865b047
...
...
@@ -54,7 +54,8 @@ class PReluKernel : public framework::OpKernel {
int
numel
=
x
->
numel
();
Transform
(
context
.
device_context
(),
x_ptr
,
x_ptr
+
numel
,
o_ptr
,
Transform
<
Place
>
trans
;
trans
(
context
.
device_context
(),
x_ptr
,
x_ptr
+
numel
,
o_ptr
,
PReluFunctor
<
T
>
(
alpha_ptr
));
}
};
...
...
@@ -91,8 +92,9 @@ class PReluGradKernel : public framework::OpKernel {
const
T
*
out_ptr
=
out
->
data
<
T
>
();
int
numel
=
dx
->
numel
();
Transform
(
context
.
device_context
(),
out_ptr
,
out_ptr
+
numel
,
dout_ptr
,
dx_ptr
,
PReluGradFunctor
<
T
>
(
alpha_ptr
));
Transform
<
Place
>
trans
;
trans
(
context
.
device_context
(),
out_ptr
,
out_ptr
+
numel
,
dout_ptr
,
dx_ptr
,
PReluGradFunctor
<
T
>
(
alpha_ptr
));
// TODO (Zhuoyuan): add dalpha upgrade when GPU kernels ready
}
...
...
paddle/platform/transform.h
浏览文件 @
d865b047
...
...
@@ -29,45 +29,71 @@
namespace
paddle
{
namespace
platform
{
// Transform on host or device. It provides the same API in std library.
template
<
typename
InputIter
,
typename
OutputIter
,
typename
UnaryOperation
>
void
Transform
(
const
DeviceContext
&
context
,
InputIter
first
,
InputIter
last
,
template
<
typename
Place
>
struct
Transform
{
template
<
typename
InputIter
,
typename
OutputIter
,
typename
UnaryOperation
>
void
operator
()(
const
DeviceContext
&
context
,
InputIter
first
,
InputIter
last
,
OutputIter
result
,
UnaryOperation
op
);
template
<
typename
InputIter1
,
typename
InputIter2
,
typename
OutputIter
,
typename
BinaryOperation
>
void
operator
()(
const
DeviceContext
&
context
,
InputIter1
first1
,
InputIter1
last1
,
InputIter2
first2
,
OutputIter
result
,
BinaryOperation
op
);
};
template
<
>
struct
Transform
<
platform
::
CPUPlace
>
{
template
<
typename
InputIter
,
typename
OutputIter
,
typename
UnaryOperation
>
void
operator
()(
const
DeviceContext
&
context
,
InputIter
first
,
InputIter
last
,
OutputIter
result
,
UnaryOperation
op
)
{
auto
place
=
context
.
GetPlace
();
if
(
is_cpu_place
(
place
))
{
PADDLE_ENFORCE
(
is_cpu_place
(
place
),
"It must use CPU place."
);
std
::
transform
(
first
,
last
,
result
,
op
);
}
else
{
#ifdef __NVCC__
auto
&
ctx
=
reinterpret_cast
<
const
CUDADeviceContext
&>
(
context
);
using
namespace
details
;
thrust
::
transform
(
thrust
::
cuda
::
par
.
on
(
ctx
.
stream
()),
DevPtrCast
(
first
),
DevPtrCast
(
last
),
DevPtrCast
(
result
),
op
);
#else
PADDLE_THROW
(
"Do not invoke `Transform<GPUPlace>` in .cc file"
);
#endif
}
}
template
<
typename
InputIter1
,
typename
InputIter2
,
typename
OutputIter
,
template
<
typename
InputIter1
,
typename
InputIter2
,
typename
OutputIter
,
typename
BinaryOperation
>
void
Transform
(
const
DeviceContext
&
context
,
InputIter1
first1
,
void
operator
()
(
const
DeviceContext
&
context
,
InputIter1
first1
,
InputIter1
last1
,
InputIter2
first2
,
OutputIter
result
,
BinaryOperation
op
)
{
auto
place
=
context
.
GetPlace
();
if
(
is_cpu_place
(
place
))
{
PADDLE_ENFORCE
(
is_cpu_place
(
place
),
"It must use CPU place."
);
std
::
transform
(
first1
,
last1
,
first2
,
result
,
op
);
}
else
{
}
};
#ifdef __NVCC__
template
<
>
struct
Transform
<
platform
::
GPUPlace
>
{
template
<
typename
InputIter
,
typename
OutputIter
,
typename
UnaryOperation
>
void
operator
()(
const
DeviceContext
&
context
,
InputIter
first
,
InputIter
last
,
OutputIter
result
,
UnaryOperation
op
)
{
auto
place
=
context
.
GetPlace
();
PADDLE_ENFORCE
(
is_gpu_place
(
place
),
"It must use GPU place."
);
auto
&
ctx
=
reinterpret_cast
<
const
CUDADeviceContext
&>
(
context
);
using
namespace
details
;
thrust
::
transform
(
thrust
::
cuda
::
par
.
on
(
ctx
.
stream
()),
DevPtrCast
(
first1
),
DevPtrCast
(
last1
),
DevPtrCast
(
first2
),
DevPtrCast
(
result
),
thrust
::
transform
(
thrust
::
cuda
::
par
.
on
(
ctx
.
stream
()),
details
::
DevPtrCast
(
first
),
details
::
DevPtrCast
(
last
),
details
::
DevPtrCast
(
result
),
op
);
}
template
<
typename
InputIter1
,
typename
InputIter2
,
typename
OutputIter
,
typename
BinaryOperation
>
void
operator
()(
const
DeviceContext
&
context
,
InputIter1
first1
,
InputIter1
last1
,
InputIter2
first2
,
OutputIter
result
,
BinaryOperation
op
)
{
auto
place
=
context
.
GetPlace
();
PADDLE_ENFORCE
(
is_gpu_place
(
place
),
"It must use GPU place."
);
auto
&
ctx
=
reinterpret_cast
<
const
CUDADeviceContext
&>
(
context
);
thrust
::
transform
(
thrust
::
cuda
::
par
.
on
(
ctx
.
stream
()),
details
::
DevPtrCast
(
first1
),
details
::
DevPtrCast
(
last1
),
details
::
DevPtrCast
(
first2
),
details
::
DevPtrCast
(
result
),
op
);
#else
PADDLE_THROW
(
"Do not invoke `Transform<GPUPlace>` in .cc file"
);
#endif
}
};
#endif
}
// namespace platform
}
// namespace paddle
paddle/platform/transform_test.cu
浏览文件 @
d865b047
...
...
@@ -15,6 +15,7 @@
#include <gtest/gtest.h>
#include "paddle/memory/memcpy.h"
#include "paddle/memory/memory.h"
#include "paddle/platform/hostdevice.h"
#include "paddle/platform/transform.h"
template
<
typename
T
>
...
...
@@ -38,7 +39,8 @@ TEST(Transform, CPUUnary) {
using
namespace
paddle
::
platform
;
CPUDeviceContext
ctx
;
float
buf
[
4
]
=
{
0.1
,
0.2
,
0.3
,
0.4
};
Transform
(
ctx
,
buf
,
buf
+
4
,
buf
,
Scale
<
float
>
(
10
));
Transform
<
paddle
::
platform
::
CPUPlace
>
trans
;
trans
(
ctx
,
buf
,
buf
+
4
,
buf
,
Scale
<
float
>
(
10
));
for
(
int
i
=
0
;
i
<
4
;
++
i
)
{
ASSERT_NEAR
(
buf
[
i
],
static_cast
<
float
>
(
i
+
1
),
1e-5
);
}
...
...
@@ -52,7 +54,8 @@ TEST(Transform, GPUUnary) {
float
cpu_buf
[
4
]
=
{
0.1
,
0.2
,
0.3
,
0.4
};
float
*
gpu_buf
=
static_cast
<
float
*>
(
Alloc
(
gpu0
,
sizeof
(
float
)
*
4
));
Copy
(
gpu0
,
gpu_buf
,
CPUPlace
(),
cpu_buf
,
sizeof
(
cpu_buf
));
Transform
(
ctx
,
gpu_buf
,
gpu_buf
+
4
,
gpu_buf
,
Scale
<
float
>
(
10
));
Transform
<
paddle
::
platform
::
GPUPlace
>
trans
;
trans
(
ctx
,
gpu_buf
,
gpu_buf
+
4
,
gpu_buf
,
Scale
<
float
>
(
10
));
ctx
.
Wait
();
Copy
(
CPUPlace
(),
cpu_buf
,
gpu0
,
gpu_buf
,
sizeof
(
cpu_buf
));
Free
(
gpu0
,
gpu_buf
);
...
...
@@ -65,7 +68,9 @@ TEST(Transform, CPUBinary) {
using
namespace
paddle
::
platform
;
using
namespace
paddle
::
memory
;
int
buf
[
4
]
=
{
1
,
2
,
3
,
4
};
Transform
(
CPUDeviceContext
(),
buf
,
buf
+
4
,
buf
,
buf
,
Multiply
<
int
>
());
Transform
<
paddle
::
platform
::
CPUPlace
>
trans
;
CPUDeviceContext
ctx
;
trans
(
ctx
,
buf
,
buf
+
4
,
buf
,
buf
,
Multiply
<
int
>
());
for
(
int
i
=
0
;
i
<
4
;
++
i
)
{
ASSERT_EQ
((
i
+
1
)
*
(
i
+
1
),
buf
[
i
]);
}
...
...
@@ -79,7 +84,8 @@ TEST(Transform, GPUBinary) {
CUDADeviceContext
ctx
(
gpu0
);
int
*
gpu_buf
=
static_cast
<
int
*>
(
Alloc
(
gpu0
,
sizeof
(
buf
)));
Copy
(
gpu0
,
gpu_buf
,
CPUPlace
(),
buf
,
sizeof
(
buf
));
Transform
(
ctx
,
gpu_buf
,
gpu_buf
+
4
,
gpu_buf
,
gpu_buf
,
Multiply
<
int
>
());
Transform
<
paddle
::
platform
::
GPUPlace
>
trans
;
trans
(
ctx
,
gpu_buf
,
gpu_buf
+
4
,
gpu_buf
,
gpu_buf
,
Multiply
<
int
>
());
ctx
.
Wait
();
Copy
(
CPUPlace
(),
buf
,
gpu0
,
gpu_buf
,
sizeof
(
buf
));
Free
(
gpu0
,
gpu_buf
);
...
...
python/paddle/v2/framework/tests/test_prelu_op.py
浏览文件 @
d865b047
...
...
@@ -17,10 +17,10 @@ class PReluTest(OpTest):
assert
out_np
is
not
self
.
inputs
[
'X'
]
self
.
outputs
=
{
'Out'
:
out_np
}
def
not_
test_check_output
(
self
):
def
test_check_output
(
self
):
self
.
check_output
()
def
not_
test_check_grad
(
self
):
def
test_check_grad
(
self
):
self
.
check_grad
([
'X'
],
'Out'
)
...
...
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