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
sven357
Paddle
提交
a8288392
P
Paddle
项目概览
sven357
/
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
a8288392
编写于
2月 24, 2018
作者:
C
chengduoZH
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
follow comments
上级
22b9ab05
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
52 addition
and
58 deletion
+52
-58
paddle/fluid/operators/elementwise_op_function.h
paddle/fluid/operators/elementwise_op_function.h
+13
-58
paddle/fluid/platform/cuda_helper.h
paddle/fluid/platform/cuda_helper.h
+39
-0
未找到文件。
paddle/fluid/operators/elementwise_op_function.h
浏览文件 @
a8288392
...
...
@@ -20,6 +20,7 @@ limitations under the License. */
#ifdef __NVCC__
#include <thrust/iterator/iterator_adaptor.h>
#include "paddle/fluid/platform/cuda_helper.h"
constexpr
int
ELEMWISE_MAX_BLOCK_DIM
=
1024
;
#endif
...
...
@@ -357,25 +358,14 @@ static void ElemwiseGradBroadcast1CPU(const T* x, const T* y, const T* out,
}
}
#ifdef __NVCC__
// __shfl_down has been deprecated as of CUDA 9.0
#if CUDA_VERSION < 9000
template
<
typename
T
>
__forceinline__
__device__
T
__shfl_down_sync
(
unsigned
,
T
val
,
int
delta
)
{
return
__shfl_down
(
val
,
delta
);
}
#endif
template
<
typename
T
,
typename
DX_OP
,
typename
DY_OP
>
static
__global__
void
ElemwiseGradBroadcast1CUDAKernel
(
const
T
*
x
,
const
T
*
y
,
const
T
*
out
,
const
T
*
dout
,
int
h
,
int
w
,
DX_OP
dx_op
,
DY_OP
dy_op
,
T
*
dx
,
T
*
dy
)
{
extern
__shared__
char
shm_buffer
[];
T
*
shm
=
reinterpret_cast
<
T
*>
(
shm_buffer
);
int
j
=
blockIdx
.
x
;
int
i
=
threadIdx
.
x
;
int
tid
=
threadIdx
.
x
;
shm
[
tid
]
=
0
;
T
val
=
0
;
do
{
int
x_offset
=
i
*
w
+
j
;
...
...
@@ -383,29 +373,15 @@ static __global__ void ElemwiseGradBroadcast1CUDAKernel(
dx
[
x_offset
]
=
dx_op
(
x
[
x_offset
],
y
[
j
],
out
[
x_offset
],
dout
[
x_offset
]);
}
if
(
dy
)
{
shm
[
tid
]
+=
dy_op
(
x
[
x_offset
],
y
[
j
],
out
[
x_offset
],
dout
[
x_offset
]);
val
+=
dy_op
(
x
[
x_offset
],
y
[
j
],
out
[
x_offset
],
dout
[
x_offset
]);
}
i
+=
ELEMWISE_MAX_BLOCK_DIM
;
}
while
(
i
<
h
);
if
(
dy
)
{
T
val
=
shm
[
threadIdx
.
x
];
int
warpSize
=
32
;
for
(
int
offset
=
warpSize
/
2
;
offset
>
0
;
offset
/=
2
)
val
+=
__shfl_down_sync
(
0
,
val
,
offset
);
__syncthreads
();
shm
[
tid
]
=
0
;
if
(
threadIdx
.
x
%
32
==
0
)
{
shm
[
threadIdx
.
x
/
32
]
=
val
;
}
val
=
shm
[
threadIdx
.
x
];
for
(
int
offset
=
warpSize
/
2
;
offset
>
0
;
offset
/=
2
)
val
+=
__shfl_down_sync
(
0
,
val
,
offset
);
val
=
platform
::
ReduceSum
(
val
,
tid
);
if
(
threadIdx
.
x
==
0
)
{
dy
[
j
]
=
shm
[
0
]
;
dy
[
j
]
=
val
;
}
}
}
...
...
@@ -417,10 +393,8 @@ static void ElemwiseGradBroadcast1CUDA(cudaStream_t stream, const T* x,
T
*
dx
,
T
*
dy
)
{
int
block_size
=
std
::
min
(
ELEMWISE_MAX_BLOCK_DIM
,
h
);
int
gird_size
=
w
;
int
shared_mem_size
=
block_size
*
sizeof
(
T
);
ElemwiseGradBroadcast1CUDAKernel
<<<
gird_size
,
block_size
,
shared_mem_size
,
stream
>>>
(
x
,
y
,
out
,
dout
,
h
,
w
,
dx_op
,
dy_op
,
dx
,
dy
);
ElemwiseGradBroadcast1CUDAKernel
<<<
gird_size
,
block_size
,
0
,
stream
>>>
(
x
,
y
,
out
,
dout
,
h
,
w
,
dx_op
,
dy_op
,
dx
,
dy
);
}
#endif
...
...
@@ -451,7 +425,6 @@ static void ElemwiseGradBroadcast2CPU(const T* x, const T* y, const T* out,
}
#ifdef __NVCC__
template
<
typename
T
,
typename
DX_OP
,
typename
DY_OP
>
static
__global__
void
ElemwiseGradBroadcast2CUDAKernel
(
const
T
*
x
,
const
T
*
y
,
const
T
*
out
,
const
T
*
dout
,
int
pre
,
int
n
,
...
...
@@ -459,9 +432,7 @@ static __global__ void ElemwiseGradBroadcast2CUDAKernel(
int
tid
=
threadIdx
.
x
;
int
j
=
blockIdx
.
x
;
extern
__shared__
char
shm_buffer
[];
T
*
shm
=
reinterpret_cast
<
T
*>
(
shm_buffer
);
shm
[
tid
]
=
0
;
T
val
=
0
;
int
ttid
=
tid
;
while
(
true
)
{
...
...
@@ -476,30 +447,16 @@ static __global__ void ElemwiseGradBroadcast2CUDAKernel(
}
if
(
dy
!=
nullptr
)
{
shm
[
tid
]
+=
dy_op
(
x
[
x_offset
],
y
[
j
],
out
[
x_offset
],
dout
[
x_offset
]);
val
+=
dy_op
(
x
[
x_offset
],
y
[
j
],
out
[
x_offset
],
dout
[
x_offset
]);
}
ttid
+=
ELEMWISE_MAX_BLOCK_DIM
;
}
if
(
dy
)
{
T
val
=
shm
[
threadIdx
.
x
];
int
warpSize
=
32
;
for
(
int
offset
=
warpSize
/
2
;
offset
>
0
;
offset
/=
2
)
val
+=
__shfl_down_sync
(
0
,
val
,
offset
);
__syncthreads
();
shm
[
tid
]
=
0
;
if
(
threadIdx
.
x
%
32
==
0
)
{
shm
[
threadIdx
.
x
/
32
]
=
val
;
}
val
=
shm
[
threadIdx
.
x
];
for
(
int
offset
=
warpSize
/
2
;
offset
>
0
;
offset
/=
2
)
val
+=
__shfl_down_sync
(
0
,
val
,
offset
);
val
=
platform
::
ReduceSum
(
val
,
threadIdx
.
x
);
if
(
threadIdx
.
x
==
0
)
{
dy
[
j
]
=
shm
[
0
]
;
dy
[
j
]
=
val
;
}
}
}
...
...
@@ -511,10 +468,8 @@ static void ElemwiseGradBroadcast2CUDA(cudaStream_t stream, const T* x,
DY_OP
dy_op
,
T
*
dx
,
T
*
dy
)
{
int
block_size
=
std
::
min
(
ELEMWISE_MAX_BLOCK_DIM
,
pre
*
post
);
int
gird_size
=
n
;
int
shared_mem_size
=
block_size
*
sizeof
(
T
);
ElemwiseGradBroadcast2CUDAKernel
<<<
gird_size
,
block_size
,
shared_mem_size
,
stream
>>>
(
x
,
y
,
out
,
dout
,
pre
,
n
,
post
,
dx_op
,
dy_op
,
dx
,
dy
);
ElemwiseGradBroadcast2CUDAKernel
<<<
gird_size
,
block_size
,
0
,
stream
>>>
(
x
,
y
,
out
,
dout
,
pre
,
n
,
post
,
dx_op
,
dy_op
,
dx
,
dy
);
}
#endif
...
...
paddle/fluid/platform/cuda_helper.h
浏览文件 @
a8288392
...
...
@@ -62,5 +62,44 @@ CUDA_ATOMIC_WRAPPER(Add, double) {
}
#endif
// __shfl_down has been deprecated as of CUDA 9.0.
#if CUDA_VERSION < 9000
template
<
typename
T
>
__forceinline__
__device__
T
__shfl_down_sync
(
unsigned
,
T
val
,
int
delta
)
{
return
__shfl_down
(
val
,
delta
);
}
#define CREATE_SHFL_MASK(mask, predicate) unsigned mask = 0u;
#else
#define FULL_WARP_MASK 0xFFFFFFFF
#define CREATE_SHFL_MASK(mask, predicate) \
unsigned mask = __ballot_sync(FULL_WARP_MASK, (predicate))
#endif
template
<
typename
T
>
__device__
T
ReduceSum
(
T
val
,
int
tid
)
{
__shared__
T
shm
[
32
];
const
int
warpSize
=
32
;
for
(
int
offset
=
warpSize
/
2
;
offset
>
0
;
offset
/=
2
)
val
+=
__shfl_down_sync
(
-
1U
,
val
,
offset
);
if
(
tid
<
warpSize
)
shm
[
tid
]
=
0
;
__syncthreads
();
if
(
tid
%
warpSize
==
0
)
{
shm
[
tid
/
warpSize
]
=
val
;
}
CREATE_SHFL_MASK
(
mask
,
tid
<
warpSize
);
if
(
tid
<
warpSize
)
{
val
=
shm
[
tid
];
for
(
int
offset
=
warpSize
/
2
;
offset
>
0
;
offset
/=
2
)
val
+=
__shfl_down_sync
(
mask
,
val
,
offset
);
}
return
val
;
}
}
// namespace platform
}
// namespace paddle
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