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
机器未来
Paddle
提交
6b3e0b78
P
Paddle
项目概览
机器未来
/
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
6b3e0b78
编写于
8月 05, 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
gather function with test passed
上级
eef55ca7
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
126 addition
and
72 deletion
+126
-72
paddle/operators/CMakeLists.txt
paddle/operators/CMakeLists.txt
+5
-0
paddle/operators/gather_func.cc
paddle/operators/gather_func.cc
+19
-0
paddle/operators/gather_func.h
paddle/operators/gather_func.h
+52
-72
paddle/operators/gather_test.cc
paddle/operators/gather_test.cc
+50
-0
未找到文件。
paddle/operators/CMakeLists.txt
浏览文件 @
6b3e0b78
...
...
@@ -41,6 +41,11 @@ 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
)
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.cc
0 → 100644
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_func.h
浏览文件 @
6b3e0b78
...
...
@@ -4,7 +4,7 @@ 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
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,
...
...
@@ -13,51 +13,18 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <memory.h>
#include <cstring>
#include "paddle/framework/ddim.h"
#include "paddle/framework/tensor.h"
#include "paddle/platform/place.h"
/**
* 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
Place
,
typename
T
>
Tensor
*
Gather
(
Tensor
*
src
,
Tensor
*
index
)
{
// check index of shape 1-D
PADDLE_ENFORCE
(
index
->
dims
().
size
()
==
1
);
int
index_size
=
index
->
dims
()[
0
];
// Source shape
auto
src_dims
=
src
->
dims
();
DDim
output_dims
(
dims_src
);
// Create a tensor of shape [index_size, dim_src[1:]]
output_dims
[
0
]
=
index_size
;
Tensor
*
New_tensor
;
float
*
output
=
nullptr
;
/* slice size */
int
slice_size
=
1
;
for
(
size_t
i
=
0
;
i
<
src_dims
.
size
();
++
i
)
slice_size
*=
src_dims
[
i
];
using
paddle
::
framework
::
Tensor
;
using
paddle
::
framework
::
DDim
;
/* Gathering */
if
(
place
==
CPUPlace
())
{
// init for CPU
output
=
New_tensor
.
mutable_data
<
T
>
(
output_dims
,
CPUPlace
());
CPUGather
(
src
->
data
(),
index
->
data
(),
slice_size
,
new_tensor
->
mutable_data
());
}
else
{
// GPU
// init for GPU
output
=
New_tensor
.
mutable_data
<
T
>
(
output_dims
,
GPUPlace
());
/* how to specialize device??*/
GPUGather
(
d
,
src
->
data
(),
index
->
data
(),
slice_size
,
new_tensor
->
mutable_data
());
}
return
New_tensor
;
}
namespace
paddle
{
namespace
operators
{
/* Implementation of CPU copy */
template
<
typename
T
>
...
...
@@ -70,48 +37,61 @@ void CPUGather(const T* params,
for
(
size_t
i
=
0
;
i
<
index_size
;
++
i
)
{
int
index_
=
indices
[
i
];
/* copy src[index_] to output[i] */
memcpy
(
output
+
i
*
slice_bytes
,
params
+
index_
*
slice_bytes
,
slice_bytes
);
// copy src[index_] to output[i]
memcpy
(
output
+
i
*
slice_size
,
params
+
index_
*
slice_size
,
slice_bytes
);
}
}
/* Implementation of GPU copy:
I suppose the GPUDevice& d, contains gpu_id and thread_id
d = cuda_stream(gpu_id_, stream_id_);
I suppose the GPUDevice& d, contains gpu_id and thread_id
d = cuda_stream(gpu_id_, stream_id_);
*/
template
<
typename
T
>
void
GPUGather
(
const
GPUDevice
&
d
,
const
T
*
src
,
void
GPUGather
(
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
;
GatherOpKernel
<
T
><<<
block_count
,
thread_per_block
,
0
,
d
.
stream
()
>>>
(
src
,
index
,
output
,
slice_size
,
indices_size
,
slice_size
,
out_size
);
}
T
*
output
);
/**
* 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
>
__global__
void
GatherOpKernel
(
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
gather_i
=
indices
[
indices_i
];
int
params_i
=
gather_i
*
slice_size
+
slice_i
;
out
[
i
]
=
*
(
params
+
params_i
);
void
Gather
(
const
platform
::
Place
&
place
,
const
paddle
::
framework
::
Tensor
*
src
,
const
paddle
::
framework
::
Tensor
*
index
,
paddle
::
framework
::
Tensor
*
output
)
{
// check index of shape 1-D
PADDLE_ENFORCE
(
index
->
dims
().
size
()
==
1
);
int
index_size
=
index
->
dims
()[
0
];
auto
src_dims
=
src
->
dims
();
DDim
output_dims
(
src_dims
);
output_dims
[
0
]
=
index_size
;
// slice size
int
slice_size
=
1
;
for
(
size_t
i
=
1
;
i
<
src_dims
.
size
();
++
i
)
slice_size
*=
src_dims
[
i
];
// Gathering
if
(
platform
::
is_cpu_place
(
place
))
{
CPUGather
<
T
>
(
src
->
data
<
T
>
(),
index
->
data
<
int
>
(),
slice_size
,
index_size
,
output
->
data
<
T
>
());
}
else
{
// init for GPU
// output_arr = output->mutable_data<T>(output_dims, platform::GPUPlace());
// how to specialize device??
// GPUGather(
// d, src->data(), index->data(), slice_size,
// new_tensor->mutable_data());
}
}
}
// namespace operators
}
// namespace paddle
paddle/operators/gather_test.cc
0 → 100644
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/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
)
{
using
namespace
paddle
::
framework
;
using
namespace
paddle
::
platform
;
using
namespace
paddle
::
operators
;
Tensor
*
src
=
new
Tensor
();
Tensor
*
index
=
new
Tensor
();
Tensor
*
output
=
new
Tensor
();
// src.Resize(make_ddim({3, 4}));
int
*
p_src
=
nullptr
;
int
*
p_index
=
nullptr
;
p_src
=
src
->
mutable_data
<
int
>
(
make_ddim
({
3
,
4
}),
CPUPlace
());
p_index
=
index
->
mutable_data
<
int
>
(
make_ddim
({
2
}),
CPUPlace
());
for
(
size_t
i
=
0
;
i
<
12
;
++
i
)
p_src
[
i
]
=
i
;
p_index
[
0
]
=
1
;
p_index
[
1
]
=
0
;
// gather
int
*
p_output
=
output
->
mutable_data
<
int
>
(
make_ddim
({
2
,
4
}),
CPUPlace
());
Gather
<
int
>
(
CPUPlace
(),
src
,
index
,
output
);
for
(
size_t
i
=
0
;
i
<
4
;
++
i
)
EXPECT_EQ
(
p_output
[
i
],
i
+
4
);
for
(
size_t
i
=
4
;
i
<
8
;
++
i
)
EXPECT_EQ
(
p_output
[
i
],
i
-
4
);
}
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