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
兽拳
Paddle
提交
78808b20
P
Paddle
项目概览
兽拳
/
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
78808b20
编写于
9月 28, 2017
作者:
Z
zchen0211
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
1 api
上级
b851515b
变更
10
隐藏空白更改
内联
并排
Showing
10 changed file
with
55 addition
and
87 deletion
+55
-87
paddle/operators/cond_op.cc
paddle/operators/cond_op.cc
+2
-2
paddle/operators/gather.cu.h
paddle/operators/gather.cu.h
+11
-19
paddle/operators/gather.h
paddle/operators/gather.h
+15
-23
paddle/operators/gather_op.cu
paddle/operators/gather_op.cu
+2
-2
paddle/operators/gather_op.h
paddle/operators/gather_op.h
+1
-1
paddle/operators/gather_test.cc
paddle/operators/gather_test.cc
+1
-1
paddle/operators/scatter.cu.h
paddle/operators/scatter.cu.h
+14
-22
paddle/operators/scatter.h
paddle/operators/scatter.h
+6
-14
paddle/operators/scatter_op.cu
paddle/operators/scatter_op.cu
+2
-2
paddle/operators/scatter_op.h
paddle/operators/scatter_op.h
+1
-1
未找到文件。
paddle/operators/cond_op.cc
浏览文件 @
78808b20
...
...
@@ -169,8 +169,8 @@ void CondOp::Run(const Scope& scope,
tensor_child
->
Resize
(
dim
);
tensor_child
->
mutable_data
<
float
>
(
dim
,
platform
::
CPUPlace
());
CPU
T
Gather
<
float
>
(
dev_ctx
.
GetPlace
(),
tensor_parent
,
&
index_tensors
[
i
],
tensor_child
);
CPUGather
<
float
>
(
dev_ctx
.
GetPlace
(),
tensor_parent
,
&
index_tensors
[
i
],
tensor_child
);
}
}
...
...
paddle/operators/gather.cu.h
浏览文件 @
78808b20
...
...
@@ -38,19 +38,6 @@ __global__ void GatherCUDAKernel(const T* params, const int* indices, T* output,
}
}
// Implementation of GPU copy:
template
<
typename
T
>
struct
GPUGather
{
void
operator
()(
const
T
*
src
,
const
int
*
index
,
const
int
slice_size
,
const
int
index_size
,
T
*
output
)
{
int
block
=
512
;
int
n
=
slice_size
*
index_size
;
int
grid
=
(
n
+
block
-
1
)
/
block
;
GatherCUDAKernel
<
T
><<<
grid
,
block
>>>
(
src
,
index
,
output
,
index_size
,
slice_size
);
}
};
/**
* A thin wrapper on gpu tensor
* Return a new tensor from source tensor, gathered according to index
...
...
@@ -59,8 +46,8 @@ struct GPUGather {
* return: output tensor
*/
template
<
typename
T
>
void
GPU
T
Gather
(
const
Place
&
place
,
const
Tensor
*
src
,
const
Tensor
*
index
,
Tensor
*
output
)
{
void
GPUGather
(
const
Place
&
place
,
const
Tensor
*
src
,
const
Tensor
*
index
,
Tensor
*
output
)
{
PADDLE_ENFORCE
(
platform
::
is_gpu_place
(
place
));
// check index of shape 1-D
PADDLE_ENFORCE
(
index
->
dims
().
size
()
==
1
);
...
...
@@ -74,10 +61,15 @@ void GPUTGather(const Place& place, const Tensor* src, const Tensor* index,
int
slice_size
=
1
;
for
(
int
i
=
1
;
i
<
src_dims
.
size
();
++
i
)
slice_size
*=
src_dims
[
i
];
// Gathering
GPUGather
<
T
>
gather_functor
;
gather_functor
(
src
->
data
<
T
>
(),
index
->
data
<
int
>
(),
slice_size
,
index_size
,
output
->
data
<
T
>
());
const
T
*
p_src
=
src
->
data
<
T
>
();
const
int
*
p_index
=
index
->
data
<
int
>
();
T
*
p_output
=
output
->
data
<
T
>
();
int
block
=
512
;
int
n
=
slice_size
*
index_size
;
int
grid
=
(
n
+
block
-
1
)
/
block
;
GatherCUDAKernel
<
T
><<<
grid
,
block
>>>
(
p_src
,
p_index
,
p_output
,
index_size
,
slice_size
);
}
}
// namespace operators
...
...
paddle/operators/gather.h
浏览文件 @
78808b20
...
...
@@ -24,32 +24,18 @@ limitations under the License. */
namespace
paddle
{
namespace
operators
{
// Implementation of CPU copy
template
<
typename
T
>
struct
CPUGather
{
void
operator
()(
const
T
*
src
,
const
int
*
indices
,
const
int
slice_size
,
const
int
index_size
,
T
*
output
)
{
const
size_t
slice_bytes
=
slice_size
*
sizeof
(
T
);
for
(
int
i
=
0
;
i
<
index_size
;
++
i
)
{
int
index_
=
indices
[
i
];
memcpy
(
output
+
i
*
slice_size
,
src
+
index_
*
slice_size
,
slice_bytes
);
}
}
};
/**
* A thin wrapper on cpu tensor
* A thin wrapper
for gathering
on cpu tensor
* Return a new tensor from source tensor, gathered according to index
* input[src]: type-T source Tensor
* input[index]: type-int index Tensor (1-D)
* return: output tensor
*/
template
<
typename
T
>
void
CPU
T
Gather
(
const
platform
::
Place
&
place
,
const
paddle
::
framework
::
Tensor
*
src
,
const
paddle
::
framework
::
Tensor
*
index
,
paddle
::
framework
::
Tensor
*
output
)
{
void
CPUGather
(
const
platform
::
Place
&
place
,
const
paddle
::
framework
::
Tensor
*
src
,
const
paddle
::
framework
::
Tensor
*
index
,
paddle
::
framework
::
Tensor
*
output
)
{
PADDLE_ENFORCE
(
platform
::
is_cpu_place
(
place
));
// check index of shape 1-D
PADDLE_ENFORCE
(
index
->
dims
().
size
()
==
1
);
...
...
@@ -59,14 +45,20 @@ void CPUTGather(const platform::Place& place,
framework
::
DDim
output_dims
(
src_dims
);
output_dims
[
0
]
=
index_size
;
const
T
*
p_src
=
src
->
data
<
T
>
();
const
int
*
p_index
=
index
->
data
<
int
>
();
T
*
p_output
=
output
->
data
<
T
>
();
// slice size
int
slice_size
=
1
;
for
(
int
i
=
1
;
i
<
src_dims
.
size
();
++
i
)
slice_size
*=
src_dims
[
i
];
// Gathering
CPUGather
<
T
>
gather_functor
;
gather_functor
(
src
->
data
<
T
>
(),
index
->
data
<
int
>
(),
slice_size
,
index_size
,
output
->
data
<
T
>
());
const
size_t
slice_bytes
=
slice_size
*
sizeof
(
T
);
for
(
int
i
=
0
;
i
<
index_size
;
++
i
)
{
int
index_
=
p_index
[
i
];
memcpy
(
p_output
+
i
*
slice_size
,
p_src
+
index_
*
slice_size
,
slice_bytes
);
}
}
}
// namespace operators
...
...
paddle/operators/gather_op.cu
浏览文件 @
78808b20
...
...
@@ -32,7 +32,7 @@ class GatherOpCUDAKernel : public framework::OpKernel<T> {
output
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
GPU
T
Gather
<
T
>
(
ctx
.
GetPlace
(),
x
,
index
,
output
);
GPUGather
<
T
>
(
ctx
.
GetPlace
(),
x
,
index
,
output
);
}
};
...
...
@@ -53,7 +53,7 @@ class GatherGradOpCUDAKernel : public framework::OpKernel<T> {
auto
place
=
ctx
.
GetEigenDevice
<
platform
::
GPUPlace
>
();
dxt
.
device
(
place
)
=
dxt
.
constant
(
static_cast
<
T
>
(
0
));
GPU
TScatter
<
T
>
(
ctx
.
GetPlace
(),
dO
,
Index
,
dX
);
GPU
ScatterAssign
<
T
>
(
ctx
.
GetPlace
(),
dO
,
Index
,
dX
);
}
};
...
...
paddle/operators/gather_op.h
浏览文件 @
78808b20
...
...
@@ -36,7 +36,7 @@ class GatherOpKernel : public framework::OpKernel<T> {
output
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
CPU
T
Gather
<
T
>
(
ctx
.
GetPlace
(),
x
,
index
,
output
);
CPUGather
<
T
>
(
ctx
.
GetPlace
(),
x
,
index
,
output
);
}
};
...
...
paddle/operators/gather_test.cc
浏览文件 @
78808b20
...
...
@@ -41,7 +41,7 @@ TEST(Gather, GatherData) {
int
*
p_output
=
output
->
mutable_data
<
int
>
(
make_ddim
({
2
,
4
}),
CPUPlace
());
CPU
T
Gather
<
int
>
(
CPUPlace
(),
src
,
index
,
output
);
CPUGather
<
int
>
(
CPUPlace
(),
src
,
index
,
output
);
for
(
int
i
=
0
;
i
<
4
;
++
i
)
EXPECT_EQ
(
p_output
[
i
],
i
+
4
);
for
(
int
i
=
4
;
i
<
8
;
++
i
)
EXPECT_EQ
(
p_output
[
i
],
i
-
4
);
...
...
paddle/operators/scatter.cu.h
浏览文件 @
78808b20
...
...
@@ -36,20 +36,6 @@ __global__ void ScatterCUDAKernel(const T* params, const int* indices,
}
}
// Implementation of GPU copy:
template
<
typename
T
>
struct
GPUScatterAssign
{
void
operator
()(
const
T
*
src
,
const
int
*
index
,
const
int
slice_size
,
const
int
index_size
,
T
*
output
)
{
int
block
=
512
;
int
n
=
slice_size
*
index_size
;
int
grid
=
(
n
+
block
-
1
)
/
block
;
// printf("grid, block: %d %d\n", grid, block);
ScatterCUDAKernel
<
T
><<<
grid
,
block
>>>
(
src
,
index
,
output
,
index_size
,
slice_size
);
}
};
/**
* A thin wrapper on gpu tensor
* Return a new updated tensor from source tensor, scatter-assigned according to
...
...
@@ -59,10 +45,10 @@ struct GPUScatterAssign {
* return: output tensor
*/
template
<
typename
T
>
void
GPU
TScatter
(
const
platform
::
Place
&
place
,
const
paddle
::
framework
::
Tensor
*
src
,
const
paddle
::
framework
::
Tensor
*
index
,
paddle
::
framework
::
Tensor
*
output
)
{
void
GPU
ScatterAssign
(
const
platform
::
Place
&
place
,
const
paddle
::
framework
::
Tensor
*
src
,
const
paddle
::
framework
::
Tensor
*
index
,
paddle
::
framework
::
Tensor
*
output
)
{
PADDLE_ENFORCE
(
platform
::
is_gpu_place
(
place
));
// check index of shape 1-D
PADDLE_ENFORCE
(
index
->
dims
().
size
()
==
1
);
...
...
@@ -76,10 +62,16 @@ void GPUTScatter(const platform::Place& place,
int
slice_size
=
1
;
for
(
int
i
=
1
;
i
<
src_dims
.
size
();
++
i
)
slice_size
*=
src_dims
[
i
];
// Scatter Assign
GPUScatterAssign
<
T
>
scatter_functor
;
scatter_functor
(
src
->
data
<
T
>
(),
index
->
data
<
int
>
(),
slice_size
,
index_size
,
output
->
data
<
T
>
());
const
T
*
p_src
=
src
->
data
<
T
>
();
const
int
*
p_index
=
index
->
data
<
int
>
();
T
*
p_output
=
output
->
data
<
T
>
();
int
block
=
512
;
int
n
=
slice_size
*
index_size
;
int
grid
=
(
n
+
block
-
1
)
/
block
;
ScatterCUDAKernel
<
T
><<<
grid
,
block
>>>
(
p_src
,
p_index
,
p_output
,
index_size
,
slice_size
);
}
}
// namespace operators
...
...
paddle/operators/scatter.h
浏览文件 @
78808b20
...
...
@@ -25,19 +25,6 @@ namespace operators {
using
Tensor
=
framework
::
Tensor
;
// Implementation of CPU copy
template
<
typename
T
>
void
CPUScatterAssign
(
const
T
*
src
,
const
int
*
index
,
const
int
slice_size
,
const
int
index_size
,
T
*
output
)
{
// paddle::framework::DDim output_dims = output->dims();
const
size_t
slice_bytes
=
slice_size
*
sizeof
(
T
);
for
(
int
i
=
0
;
i
<
index_size
;
++
i
)
{
int
index_
=
index
[
i
];
memcpy
(
output
+
index_
*
slice_size
,
src
+
i
*
slice_size
,
slice_bytes
);
}
}
/**
* Return a updated tensor from source tensor, scattered according to index:
* dst[i] = src[index[i]]
...
...
@@ -70,7 +57,12 @@ void ScatterAssign(const platform::Place& place,
size_t
slice_size
=
1
;
for
(
int
i
=
1
;
i
<
src_dims
.
size
();
++
i
)
slice_size
*=
src_dims
[
i
];
CPUScatterAssign
<
T
>
(
p_src
,
p_index
,
slice_size
,
index_size
,
p_output
);
const
size_t
slice_bytes
=
slice_size
*
sizeof
(
T
);
for
(
int
i
=
0
;
i
<
index_size
;
++
i
)
{
int
index_
=
p_index
[
i
];
memcpy
(
p_output
+
index_
*
slice_size
,
p_src
+
i
*
slice_size
,
slice_bytes
);
}
}
}
// namespace operators
...
...
paddle/operators/scatter_op.cu
浏览文件 @
78808b20
...
...
@@ -32,7 +32,7 @@ class ScatterOpCUDAKernel : public framework::OpKernel<T> {
Out
->
ShareDataWith
<
T
>
(
*
Ref
);
GPU
TScatter
<
T
>
(
ctx
.
GetPlace
(),
Updates
,
Index
,
Out
);
GPU
ScatterAssign
<
T
>
(
ctx
.
GetPlace
(),
Updates
,
Index
,
Out
);
}
};
...
...
@@ -51,7 +51,7 @@ class ScatterGradOpCUDAKernel : public framework::OpKernel<T> {
dRef
->
ShareDataWith
<
T
>
(
*
dOut
);
dUpdates
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
// Gradient by Gather: dUpdates = dO[Index]
GPU
T
Gather
<
T
>
(
ctx
.
GetPlace
(),
dOut
,
Index
,
dUpdates
);
GPUGather
<
T
>
(
ctx
.
GetPlace
(),
dOut
,
Index
,
dUpdates
);
}
};
...
...
paddle/operators/scatter_op.h
浏览文件 @
78808b20
...
...
@@ -56,7 +56,7 @@ class ScatterGradientOpKernel : public framework::OpKernel<T> {
dRef
->
ShareDataWith
<
T
>
(
*
dOut
);
dUpdates
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
// Gradient by Gather: dUpdates += dO[Index]
CPU
T
Gather
<
T
>
(
ctx
.
GetPlace
(),
dOut
,
Index
,
dUpdates
);
CPUGather
<
T
>
(
ctx
.
GetPlace
(),
dOut
,
Index
,
dUpdates
);
}
};
...
...
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