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
47609ab2
P
Paddle
项目概览
PaddlePaddle
/
Paddle
1 年多 前同步成功
通知
2302
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
未验证
提交
47609ab2
编写于
4月 13, 2018
作者:
Y
Yi Wang
提交者:
GitHub
4月 13, 2018
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Document transform.h and fix cpplint errors (#9913)
上级
b43d87c9
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
50 addition
and
18 deletion
+50
-18
paddle/fluid/platform/details/cuda_transform_iterator_cast.h
paddle/fluid/platform/details/cuda_transform_iterator_cast.h
+24
-9
paddle/fluid/platform/transform.h
paddle/fluid/platform/transform.h
+26
-9
未找到文件。
paddle/fluid/platform/details/
device_pt
r_cast.h
→
paddle/fluid/platform/details/
cuda_transform_iterato
r_cast.h
浏览文件 @
47609ab2
...
...
@@ -18,16 +18,22 @@ limitations under the License. */
#error device_ptr_cast must be include by .cu file
#endif
#include <thrust/device_ptr.h>
#include <type_traits> // For std::remove_pointer and std::is_pointer.
#include "thrust/device_ptr.h"
namespace
paddle
{
namespace
platform
{
namespace
details
{
// PointerToThrustDevicePtr has two speicalizations, one casts a (CUDA
// device) pointer into thrust::device_ptr, the other keeps rest types
// un-casted.
template
<
typename
T
,
bool
is_ptr
>
struct
DevicePtrCast
;
struct
PointerToThrustDevicePtr
;
template
<
typename
T
>
struct
DevicePtrCast
<
T
,
true
>
{
struct
PointerToThrustDevicePtr
<
T
,
true
>
{
using
ELEM
=
typename
std
::
remove_pointer
<
T
>::
type
;
using
RTYPE
=
thrust
::
device_ptr
<
ELEM
>
;
...
...
@@ -37,17 +43,26 @@ struct DevicePtrCast<T, true> {
};
template
<
typename
T
>
struct
DevicePtrCast
<
T
,
false
>
{
struct
PointerToThrustDevicePtr
<
T
,
false
>
{
using
RTYPE
=
T
;
inline
RTYPE
operator
()(
RTYPE
it
)
const
{
return
it
;
}
};
// Cast T to thrust::device_ptr if T is a pointer.
// Otherwise, e.g., T is a iterator, return T itself.
// CastToCUDATransformIterator casts a pointer to thrust::device_ptr
// so it could be used as the iterator of thrust::transform. It
// doesn't cast other types.
//
// We need CastToCUDATransformIterator because it is often that we
// want to use device memory pointers as transform iterators, e.g., to
// transform a block of float32 to float16. In this case, we want
// CastToCUDATransformIterator to cast float16/32 pointers to
// thrust::device_ptr, otherwise they cannot work as the iterator
// required by thrust::transform. At the same time, we don't want to
// cast thrust::device_ptr to thrust::device_ptr repeatedly.
template
<
typename
T
>
auto
DevPtrCast
(
T
t
)
->
typename
DevicePtrCast
<
T
,
std
::
is_pointer
<
T
>::
value
>::
RTYPE
{
DevicePtrCast
<
T
,
std
::
is_pointer
<
T
>::
value
>
cast
;
auto
CastToCUDATransformIterator
(
T
t
)
->
typename
PointerToThrustDevicePtr
<
T
,
std
::
is_pointer
<
T
>::
value
>::
RTYPE
{
PointerToThrustDevicePtr
<
T
,
std
::
is_pointer
<
T
>::
value
>
cast
;
return
cast
(
t
);
}
...
...
paddle/fluid/platform/transform.h
浏览文件 @
47609ab2
...
...
@@ -14,29 +14,44 @@ limitations under the License. */
#pragma once
#include <algorithm>
#include <type_traits>
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/enforce.h"
#include "paddle/fluid/platform/hostdevice.h"
#include "paddle/fluid/platform/place.h"
#include <algorithm>
#include <type_traits>
#ifdef __NVCC__
#include <thrust/execution_policy.h>
#include <thrust/transform.h>
#include "paddle/fluid/platform/details/
device_pt
r_cast.h"
#include "paddle/fluid/platform/details/
cuda_transform_iterato
r_cast.h"
#endif
namespace
paddle
{
namespace
platform
{
// Transform on host or device. It provides the same API in std library.
// Transform applys a unary or a binary functor on each element in a
// range defined by a pair of iterators.
//
// - The specialization for CPU calls std::transform.
// - The specialization for CUDA calls thrust::tranform.
//
// NOTE: We need to define InputIter and OutputIter defined as
// different types, because the InputIter points op's inputs and
// OutputIter pints to op's outputs.
//
// NOTE: We don't assume that InputIter to be const InputType* and
// OutputIter to be OutputType*, because we might use a iterator
// class, paddle::fluid::operators::RowwiseTRansformIterator.
template
<
typename
DeviceContext
>
struct
Transform
{
// The unary version.
template
<
typename
InputIter
,
typename
OutputIter
,
typename
UnaryOperation
>
void
operator
()(
const
DeviceContext
&
context
,
InputIter
first
,
InputIter
last
,
OutputIter
result
,
UnaryOperation
op
);
// The binary version.
template
<
typename
InputIter1
,
typename
InputIter2
,
typename
OutputIter
,
typename
BinaryOperation
>
void
operator
()(
const
DeviceContext
&
context
,
InputIter1
first1
,
...
...
@@ -70,8 +85,9 @@ struct Transform<platform::CUDADeviceContext> {
auto
place
=
context
.
GetPlace
();
PADDLE_ENFORCE
(
is_gpu_place
(
place
),
"It must use GPU place."
);
thrust
::
transform
(
thrust
::
cuda
::
par
.
on
(
context
.
stream
()),
details
::
DevPtrCast
(
first
),
details
::
DevPtrCast
(
last
),
details
::
DevPtrCast
(
result
),
op
);
details
::
CastToCUDATransformIterator
(
first
),
details
::
CastToCUDATransformIterator
(
last
),
details
::
CastToCUDATransformIterator
(
result
),
op
);
}
template
<
typename
InputIter1
,
typename
InputIter2
,
typename
OutputIter
,
...
...
@@ -82,9 +98,10 @@ struct Transform<platform::CUDADeviceContext> {
auto
place
=
context
.
GetPlace
();
PADDLE_ENFORCE
(
is_gpu_place
(
place
),
"It must use GPU place."
);
thrust
::
transform
(
thrust
::
cuda
::
par
.
on
(
context
.
stream
()),
details
::
DevPtrCast
(
first1
),
details
::
DevPtrCast
(
last1
),
details
::
DevPtrCast
(
first2
),
details
::
DevPtrCast
(
result
),
op
);
details
::
CastToCUDATransformIterator
(
first1
),
details
::
CastToCUDATransformIterator
(
last1
),
details
::
CastToCUDATransformIterator
(
first2
),
details
::
CastToCUDATransformIterator
(
result
),
op
);
}
};
#endif
...
...
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