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
PaddleDetection
提交
7e374713
P
PaddleDetection
项目概览
PaddlePaddle
/
PaddleDetection
大约 1 年 前同步成功
通知
695
Star
11112
Fork
2696
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
184
列表
看板
标记
里程碑
合并请求
40
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
PaddleDetection
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
184
Issue
184
列表
看板
标记
里程碑
合并请求
40
合并请求
40
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
提交
7e374713
编写于
8月 07, 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
modify
上级
6b3e0b78
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
3 addition
and
141 deletion
+3
-141
paddle/operators/CMakeLists.txt
paddle/operators/CMakeLists.txt
+1
-4
paddle/operators/gather.h
paddle/operators/gather.h
+0
-0
paddle/operators/gather_func.cc
paddle/operators/gather_func.cc
+0
-19
paddle/operators/gather_test.cc
paddle/operators/gather_test.cc
+2
-2
paddle/operators/scatter_func.h
paddle/operators/scatter_func.h
+0
-116
未找到文件。
paddle/operators/CMakeLists.txt
浏览文件 @
7e374713
...
...
@@ -41,10 +41,7 @@ function(op_library TARGET)
endif
()
endfunction
()
op_library
(
gather SRCS gather_func.cc
)
cc_test
(
gather_test SRCS gather_test.cc DEPS gather
)
op_library
(
scatter SRCS scatter_func.cc
)
cc_test
(
gather_test SRCS gather_test.cc DEPS tensor
)
op_library
(
add_op SRCS add_op.cc add_op.cu
)
cc_test
(
add_op_test SRCS add_op_test.cc DEPS add_op
)
...
...
paddle/operators/gather
_func
.h
→
paddle/operators/gather.h
浏览文件 @
7e374713
文件已移动
paddle/operators/gather_func.cc
已删除
100644 → 0
浏览文件 @
6b3e0b78
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/operators/gather_func.h"
#include <cstring>
#include "paddle/framework/ddim.h"
#include "paddle/framework/tensor.h"
#include "paddle/platform/place.h"
paddle/operators/gather_test.cc
浏览文件 @
7e374713
...
...
@@ -12,16 +12,16 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/operators/gather.h"
#include "paddle/framework/ddim.h"
#include "paddle/framework/tensor.h"
#include "paddle/operators/gather_func.h"
#include "paddle/platform/place.h"
#include <gtest/gtest.h>
#include <iostream>
#include <string>
TEST
(
_abc_
,
GatherData
)
{
TEST
(
Gather
,
GatherData
)
{
using
namespace
paddle
::
framework
;
using
namespace
paddle
::
platform
;
using
namespace
paddle
::
operators
;
...
...
paddle/operators/scatter_func.h
已删除
100644 → 0
浏览文件 @
6b3e0b78
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <cstring>
#include "paddle/framework/ddim.h"
#include "paddle/framework/tensor.h"
#include "paddle/platform/place.h"
/**
* Return a updated tensor from source tensor, scattered according to index:
* dst[i] += src[index[i]]
* input[src]: type-T source Tensor
* input[index]: type-int index Tensor (1-D)
* return: output tensor
*/
template
<
typename
Place
,
typename
T
>
void
ScatterUpdate
(
Tensor
*
src
,
Tensor
*
dst
,
Tensor
*
index
)
{
// Source shape
auto
src_dims
=
src
->
dims
();
auto
dst_dims
=
dst
->
dims
();
DDim
output_dims
(
dims_src
);
// check src shape and dst shape should match
for
(
size_t
i
=
1
;
i
<
src_dims
.
size
();
i
++
)
PADDLE_ENFORCE
(
src_dims
[
i
]
==
dst_dims
[
i
]);
int
index_size
=
index
->
dims
()[
0
];
/* slice size */
int
slice_size
=
1
;
for
(
size_t
i
=
0
;
i
<
src_dims
.
size
();
++
i
)
slice_size
*=
src_dims
[
i
];
if
(
place
==
CPUPlace
())
{
// init
output
=
new_tensor
.
mutable_data
<
T
>
(
output_dims
,
CPUPlace
());
CPUScatterUpdate
(
src
->
data
(),
index
->
data
(),
slice_size
,
new_tensor
->
mutable_data
());
}
else
{
// GPU
// init
output
=
new_tensor
.
mutable_data
<
T
>
(
output_dims
,
GPUPlace
());
/* how to specialize device??*/
GPUScatterUpdate
(
d
,
src
->
data
(),
index
->
data
(),
slice_size
,
new_tensor
->
mutable_data
());
}
}
/* Implementation of CPU copy */
template
<
typename
T
>
void
CPUScatterUpdate
(
const
T
*
src
,
const
int
*
index
,
const
int
slice_size
,
const
int
index_size
,
T
*
output
)
{
// const size_t slice_bytes = slice_size * sizeof(T);
for
(
size_t
i
=
0
;
i
<
index_size
;
++
i
)
{
int
index_
=
index
[
i
];
math
::
vAdd
<
T
>
(
slice_size
,
src
+
index_
*
slice_bytes
,
output
+
i
*
slice_bytes
,
output
+
i
*
slice_bytes
);
}
}
/* Implementation of GPU scatter:
I suppose the GPUDevice& d, contains gpu_id and thread_id
d = cuda_stream(gpu_id_, stream_id_);
*/
template
<
typename
T
>
void
GPUScatterUpdate
(
const
GPUDevice
&
d
,
const
T
*
src
,
const
int
*
index
,
const
int
slice_size
,
const
int
index_size
,
T
*
output
)
{
int
block_count
=
slice_size
*
index_size
;
int
thread_per_block
=
1024
;
ScatterOpKernel
<
T
><<<
block_count
,
thread_per_block
,
0
,
d
.
stream
()
>>>
(
src
,
index
,
output
,
slice_size
,
indices_size
,
slice_size
,
out_size
);
}
template
<
typename
T
>
__global__
void
ScatterOpKernel
(
const
T
*
params
,
const
int
*
indices
,
T
*
out
,
int64
indices_size
,
int64
slice_size
,
int64
out_size
)
{
/* I suppose we have the following macro,
which I strongly suggest that we should put in cuda:
#define CUDA_1D_KERNEL_LOOP(i, n) \
for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < n; \
i += blockDim.x * gridDim.x)
*/
CUDA_1D_KERNEL_LOOP
(
i
,
out_size
)
{
int
indices_i
=
i
/
slice_size
;
int
slice_i
=
i
-
indices_i
*
slice_size
;
// offset inside the slice
int
scatter_i
=
indices
[
indices_i
];
int
params_i
=
scatter_i
*
slice_size
+
slice_i
;
out
[
i
]
+=
*
(
params
+
params_i
);
}
}
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