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
OPTHREE
Paddle
提交
553afc07
P
Paddle
项目概览
OPTHREE
/
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 >>
未验证
提交
553afc07
编写于
2月 20, 2022
作者:
Y
Yiqun Liu
提交者:
GitHub
2月 20, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Rename the general elementwise and broadcast functions. (#39623)
上级
267275d9
变更
14
展开全部
显示空白变更内容
内联
并排
Showing
14 changed file
with
612 addition
and
669 deletion
+612
-669
paddle/fluid/operators/dropout_impl.cu.h
paddle/fluid/operators/dropout_impl.cu.h
+0
-1
paddle/fluid/operators/elementwise/elementwise_op_broadcast.cu.h
...fluid/operators/elementwise/elementwise_op_broadcast.cu.h
+4
-36
paddle/fluid/operators/elementwise/elementwise_op_impl.cu.h
paddle/fluid/operators/elementwise/elementwise_op_impl.cu.h
+3
-3
paddle/pten/kernels/funcs/broadcast_function.h
paddle/pten/kernels/funcs/broadcast_function.h
+585
-0
paddle/pten/kernels/funcs/cuda_kernel_config.h
paddle/pten/kernels/funcs/cuda_kernel_config.h
+0
-57
paddle/pten/kernels/funcs/elementwise_base.h
paddle/pten/kernels/funcs/elementwise_base.h
+4
-5
paddle/pten/kernels/gpu/abs_kernel.cu
paddle/pten/kernels/gpu/abs_kernel.cu
+1
-2
paddle/pten/kernels/gpu/cast_kernel.cu
paddle/pten/kernels/gpu/cast_kernel.cu
+1
-1
paddle/pten/kernels/gpu/elementwise.h
paddle/pten/kernels/gpu/elementwise.h
+3
-553
paddle/pten/kernels/gpu/full_kernel.cu
paddle/pten/kernels/gpu/full_kernel.cu
+2
-2
paddle/pten/kernels/gpu/math_kernel.cu
paddle/pten/kernels/gpu/math_kernel.cu
+1
-1
paddle/pten/kernels/gpu/reduce.h
paddle/pten/kernels/gpu/reduce.h
+1
-2
paddle/pten/kernels/gpu/reduce_grad.h
paddle/pten/kernels/gpu/reduce_grad.h
+6
-5
paddle/pten/kernels/gpu/scale_kernel.cu
paddle/pten/kernels/gpu/scale_kernel.cu
+1
-1
未找到文件。
paddle/fluid/operators/dropout_impl.cu.h
浏览文件 @
553afc07
...
...
@@ -36,7 +36,6 @@ limitations under the License. */
#include "paddle/fluid/operators/elementwise/elementwise_op_impl.cu.h"
#include "paddle/fluid/platform/aligned_vector.h"
#include "paddle/fluid/platform/device/gpu/gpu_launch_config.h"
#include "paddle/pten/kernels/funcs/cuda_kernel_config.h"
namespace
paddle
{
namespace
operators
{
...
...
paddle/fluid/operators/elementwise/elementwise_op_broadcast.cu.h
浏览文件 @
553afc07
...
...
@@ -15,45 +15,13 @@
#pragma once
#include "paddle/fluid/operators/elementwise/elementwise_op_impl.cu.h"
#include "paddle/fluid/operators/kernel_primitives/kernel_primitives.h"
// only can include the headers in paddle/top/api dirs
#include "paddle/pten/kernels/gpu/elementwise.h"
namespace
paddle
{
namespace
operators
{
namespace
kps
=
paddle
::
operators
::
kernel_primitives
;
template
<
ElementwiseType
ET
,
typename
InT
,
typename
OutT
,
typename
Functor
,
int
NumOuts
=
1
>
void
LaunchBroadcastElementwiseCudaKernel
(
const
KPDevice
&
ctx
,
const
std
::
vector
<
const
framework
::
Tensor
*>
&
ins
,
std
::
vector
<
framework
::
Tensor
*>
*
outs
,
int
axis
,
Functor
func
)
{
std
::
vector
<
const
pten
::
DenseTensor
*>
pt_inputs
;
std
::
vector
<
pten
::
DenseTensor
*>
pt_outputs
;
// TODO(YuanRisheng) *_tmp for cache DenseTensor, because the temporary
// DenseTensor obj
// generated by MakePtenDenseTensor can be destroyed when exits loop. *_tmp
// can be deleted
// when DenseTensor support copy constructor.
std
::
vector
<
std
::
unique_ptr
<
pten
::
DenseTensor
>>
pt_inputs_tmp
;
std
::
vector
<
std
::
unique_ptr
<
pten
::
DenseTensor
>>
pt_outputs_tmp
;
for
(
auto
in
:
ins
)
{
pt_inputs_tmp
.
emplace_back
(
std
::
move
(
paddle
::
experimental
::
MakePtenDenseTensor
(
*
in
)));
}
for
(
auto
out
:
*
outs
)
{
pt_outputs_tmp
.
emplace_back
(
std
::
move
(
paddle
::
experimental
::
MakePtenDenseTensor
(
*
out
)));
}
for
(
int
i
=
0
;
i
<
pt_inputs_tmp
.
size
();
i
++
)
{
pt_inputs
.
push_back
(
pt_inputs_tmp
[
i
].
get
());
}
for
(
int
i
=
0
;
i
<
pt_outputs_tmp
.
size
();
i
++
)
{
pt_outputs
.
push_back
(
pt_outputs_tmp
[
i
].
get
());
}
pten
::
LaunchBroadcastElementwiseCudaKernel
<
ET
,
InT
,
OutT
,
Functor
,
NumOuts
>
(
ctx
,
pt_inputs
,
&
pt_outputs
,
axis
,
func
);
}
template
<
ElementwiseType
ET
,
typename
InT
,
typename
OutT
,
typename
Functor
,
int
NumOuts
=
1
>
void
LaunchElementwiseCudaKernel
(
...
...
@@ -82,7 +50,7 @@ void LaunchElementwiseCudaKernel(
for
(
int
i
=
0
;
i
<
pt_outputs_tmp
.
size
();
i
++
)
{
pt_outputs
.
push_back
(
pt_outputs_tmp
[
i
].
get
());
}
pten
::
LaunchElementwiseCuda
Kernel
<
ET
,
InT
,
OutT
,
Functor
,
NumOuts
>
(
pten
::
funcs
::
Broadcast
Kernel
<
ET
,
InT
,
OutT
,
Functor
,
NumOuts
>
(
ctx
,
pt_inputs
,
&
pt_outputs
,
axis
,
func
);
}
...
...
paddle/fluid/operators/elementwise/elementwise_op_impl.cu.h
浏览文件 @
553afc07
...
...
@@ -19,7 +19,7 @@ limitations under the License. */
// only can include the headers in paddle/top/api dirs
#include "paddle/pten/api/lib/utils/tensor_utils.h"
#include "paddle/pten/kernels/
gpu/elementwi
se.h"
#include "paddle/pten/kernels/
funcs/elementwise_ba
se.h"
namespace
paddle
{
namespace
operators
{
...
...
@@ -53,8 +53,8 @@ void LaunchSameDimsElementwiseCudaKernel(
for
(
int
i
=
0
;
i
<
pt_outputs_tmp
.
size
();
i
++
)
{
pt_outputs
.
push_back
(
pt_outputs_tmp
[
i
].
get
());
}
pten
::
funcs
::
LaunchSameDimsElementwiseCudaKernel
<
OutT
,
Functor
,
NumOuts
>
(
ctx
,
pt_inputs
,
&
pt_outputs
,
func
);
pten
::
funcs
::
ElementwiseKernel
<
OutT
,
Functor
,
NumOuts
>
(
ctx
,
pt_inputs
,
&
pt_outputs
,
func
);
}
}
// namespace operators
...
...
paddle/pten/kernels/funcs/broadcast_function.h
0 → 100644
浏览文件 @
553afc07
此差异已折叠。
点击以展开。
paddle/pten/kernels/funcs/cuda_kernel_config.h
已删除
100644 → 0
浏览文件 @
267275d9
/* Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
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 "paddle/fluid/platform/device/gpu/gpu_device_function.h"
#include "paddle/fluid/platform/device/gpu/gpu_launch_config.h"
#include "paddle/pten/backends/gpu/gpu_context.h"
#ifdef __HIPCC__
#define ELEMENTWISE_BLOCK_SIZE 256
#else
#define ELEMENTWISE_BLOCK_SIZE 512
#endif
namespace
pten
{
namespace
funcs
{
/*
* According to NVIDIA, if number of threads per block is 64/128/256/512,
* cuda performs better. And number of blocks should be greater (at least
* 2x~4x) than number of SMs. Hence, SM count is took into account within
* this function to determine the right number of threads per block.
*/
inline
int
GetThreadsConfig
(
const
pten
::
GPUContext
&
ctx
,
int64_t
numel
,
int
vec_size
)
{
int
threads
=
ELEMENTWISE_BLOCK_SIZE
;
int
sm_count
=
ctx
.
GetSMCount
();
int
active_threads_num
=
numel
/
vec_size
;
if
(
active_threads_num
/
(
sm_count
<<
1
)
<
ELEMENTWISE_BLOCK_SIZE
)
{
// Round up threads number into an exponential multiple of 2, while number
// of acitve blocks is about twice of SM, to acquire better performance.
threads
=
paddle
::
platform
::
RoundToPowerOfTwo
(
active_threads_num
/
(
sm_count
<<
1
));
}
else
if
(
active_threads_num
/
(
sm_count
<<
2
)
<
ELEMENTWISE_BLOCK_SIZE
)
{
// Round up threads number into an exponential multiple of 2, while number
// of acitve blocks is about 4 times of SM, to acquire better performance.
threads
=
paddle
::
platform
::
RoundToPowerOfTwo
(
active_threads_num
/
(
sm_count
<<
2
));
}
// Number of threads per block shall be larger than 64.
return
std
::
max
(
64
,
threads
);
}
}
// namespace funcs
}
// namespace pten
paddle/pten/kernels/funcs/elementwise_base.h
浏览文件 @
553afc07
...
...
@@ -746,8 +746,7 @@ void ElementwiseCudaKernel(const KPDevice &ctx,
}
template
<
typename
OutT
,
typename
Functor
,
int
NumOuts
=
1
>
void
LaunchSameDimsElementwiseCudaKernel
(
const
KPDevice
&
ctx
,
void
ElementwiseKernel
(
const
KPDevice
&
ctx
,
const
std
::
vector
<
const
DenseTensor
*>
&
ins
,
std
::
vector
<
DenseTensor
*>
*
outs
,
Functor
func
)
{
...
...
paddle/pten/kernels/gpu/abs_kernel.cu
浏览文件 @
553afc07
...
...
@@ -47,8 +47,7 @@ void AbsKernel(const Context& ctx, const DenseTensor& x, DenseTensor* out) {
std
::
vector
<
DenseTensor
*>
outs
=
{
out
};
auto
functor
=
CudaAbsFunctor
<
T
>
();
funcs
::
LaunchSameDimsElementwiseCudaKernel
<
pten
::
funcs
::
Real
<
T
>>
(
ctx
,
ins
,
&
outs
,
functor
);
funcs
::
ElementwiseKernel
<
pten
::
funcs
::
Real
<
T
>>
(
ctx
,
ins
,
&
outs
,
functor
);
}
}
// namespace pten
...
...
paddle/pten/kernels/gpu/cast_kernel.cu
浏览文件 @
553afc07
...
...
@@ -44,7 +44,7 @@ void CastCUDAKernelImpl(const GPUContext& dev_ctx,
inputs
.
emplace_back
(
&
x
);
outputs
.
emplace_back
(
out
);
dev_ctx
.
Alloc
<
OutT
>
(
out
);
pten
::
funcs
::
LaunchSameDimsElementwiseCuda
Kernel
<
OutT
>
(
pten
::
funcs
::
Elementwise
Kernel
<
OutT
>
(
dev_ctx
,
inputs
,
&
outputs
,
CastFuctor
<
InT
,
OutT
>
());
}
...
...
paddle/pten/kernels/gpu/elementwise.h
浏览文件 @
553afc07
此差异已折叠。
点击以展开。
paddle/pten/kernels/gpu/full_kernel.cu
浏览文件 @
553afc07
...
...
@@ -49,7 +49,7 @@ void FullKernel(const Context& dev_ctx,
// This function has no input, so the inputs.size() == 0. Use kUnary, but
// the data will not be loaded in the kernel because the number of
// parameters in the operator is 0
pten
::
funcs
::
LaunchSameDimsElementwiseCuda
Kernel
<
T
>
(
pten
::
funcs
::
Elementwise
Kernel
<
T
>
(
dev_ctx
,
inputs
,
&
outputs
,
FullFuctor
<
T
>
(
val
.
to
<
T
>
()));
}
}
...
...
@@ -91,7 +91,7 @@ void FullLikeKernel(const Context& dev_ctx,
// the operator is 0
int
numel
=
out
->
numel
();
if
(
numel
>
0
)
{
pten
::
funcs
::
LaunchSameDimsElementwiseCuda
Kernel
<
T
>
(
pten
::
funcs
::
Elementwise
Kernel
<
T
>
(
dev_ctx
,
inputs
,
&
outputs
,
FullFuctor
<
T
>
(
value
));
}
}
...
...
paddle/pten/kernels/gpu/math_kernel.cu
浏览文件 @
553afc07
...
...
@@ -48,7 +48,7 @@ namespace pten {
inputs.emplace_back(&y); \
outputs.emplace_back(out); \
dev_ctx.template Alloc<T>(out); \
LaunchElementwiseCudaKernel<ElementwiseType::kBinary, T, T>(
\
funcs::BroadcastKernel<ElementwiseType::kBinary, T, T>(
\
dev_ctx, inputs, &outputs, axis, funcs::name##Functor<T>()); \
}
...
...
paddle/pten/kernels/gpu/reduce.h
浏览文件 @
553afc07
...
...
@@ -1091,8 +1091,7 @@ void TensorReduceImpl(const pten::GPUContext& dev_ctx,
if
(
config
.
reduce_num
==
1
)
{
std
::
vector
<
const
DenseTensor
*>
inputs
=
{
&
x
};
std
::
vector
<
DenseTensor
*>
outputs
=
{
y
};
funcs
::
LaunchSameDimsElementwiseCudaKernel
<
Ty
>
(
dev_ctx
,
inputs
,
&
outputs
,
transform
);
funcs
::
ElementwiseKernel
<
Ty
>
(
dev_ctx
,
inputs
,
&
outputs
,
transform
);
return
;
}
...
...
paddle/pten/kernels/gpu/reduce_grad.h
浏览文件 @
553afc07
...
...
@@ -22,8 +22,10 @@
#include <numeric>
#include <set>
#include <vector>
#include "paddle/pten/kernels/gpu/elementwise.h"
#include "paddle/pten/kernels/funcs/broadcast_function.h"
namespace
pten
{
template
<
typename
InT
,
typename
Functor
>
void
ReduceGrad
(
const
GPUContext
&
dev_ctx
,
DenseTensor
*
d_out
,
...
...
@@ -33,12 +35,11 @@ void ReduceGrad(const GPUContext& dev_ctx,
std
::
vector
<
const
DenseTensor
*>
inputs
=
{
d_out
};
std
::
vector
<
DenseTensor
*>
outputs
=
{
d_x
};
PD_VISIT_ALL_TYPES
(
out_dtype
,
"LaunchBroadcastElementwiseCudaKernel"
,
([
&
]
{
LaunchBroadcastElementwiseCudaKernel
<
pten
::
ElementwiseType
::
kUnary
,
InT
,
data_t
>
(
out_dtype
,
"BroadcastKernel"
,
([
&
]
{
funcs
::
BroadcastKernel
<
pten
::
ElementwiseType
::
kUnary
,
InT
,
data_t
>
(
dev_ctx
,
inputs
,
&
outputs
,
0
,
functor
);
}));
}
}
// namespace pten
#endif
paddle/pten/kernels/gpu/scale_kernel.cu
浏览文件 @
553afc07
...
...
@@ -54,7 +54,7 @@ void ScaleKernel(const Context& dev_ctx,
inputs
.
emplace_back
(
&
x
);
outputs
.
emplace_back
(
out
);
dev_ctx
.
template
Alloc
<
T
>(
out
);
pten
::
funcs
::
LaunchSameDimsElementwiseCuda
Kernel
<
T
>
(
pten
::
funcs
::
Elementwise
Kernel
<
T
>
(
dev_ctx
,
inputs
,
&
outputs
,
...
...
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