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
9e3433bd
P
Paddle
项目概览
PaddlePaddle
/
Paddle
1 年多 前同步成功
通知
2305
Star
20932
Fork
5423
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
9e3433bd
编写于
7月 04, 2022
作者:
B
Bo Zhang
提交者:
GitHub
7月 04, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Merge dimensions && OP performance optimization (#43931)
上级
cf8e86df
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
67 addition
and
31 deletion
+67
-31
paddle/phi/kernels/gpu/cross_grad_kernel.cu
paddle/phi/kernels/gpu/cross_grad_kernel.cu
+32
-14
paddle/phi/kernels/gpu/cross_kernel.cu
paddle/phi/kernels/gpu/cross_kernel.cu
+35
-17
未找到文件。
paddle/phi/kernels/gpu/cross_grad_kernel.cu
浏览文件 @
9e3433bd
...
...
@@ -22,8 +22,6 @@
namespace
phi
{
using
funcs
::
IndexCalculator
;
template
<
typename
T
>
__global__
void
CrossGrad
(
const
T
*
x
,
const
T
*
y
,
...
...
@@ -32,7 +30,7 @@ __global__ void CrossGrad(const T* x,
T
*
out_dy
,
const
int
stride
,
const
int
N
,
IndexCalculator
index_calculator
)
{
phi
::
funcs
::
IndexCalculator
index_calculator
)
{
CUDA_KERNEL_LOOP
(
i
,
N
)
{
int
offset
=
index_calculator
(
i
);
...
...
@@ -107,32 +105,52 @@ void CrossGradKernel(const Context& dev_ctx,
std
::
vector
<
int
>
cal_dims
;
std
::
vector
<
int
>
left_strides
;
std
::
vector
<
int
>
full_strides
;
std
::
vector
<
int
>
merged_dims
;
for
(
int
i
=
0
;
i
<
dim
;
i
++
)
{
if
(
i
==
0
)
{
merged_dims
.
push_back
(
input_x_dims
[
i
]);
}
else
{
merged_dims
[
0
]
*=
input_x_dims
[
i
];
}
}
int
merge_axis
=
merged_dims
.
size
();
merged_dims
.
push_back
(
input_x_dims
[
dim
]);
for
(
int
i
=
dim
+
1
;
i
<
input_x_dims
.
size
();
i
++
)
{
if
(
i
==
dim
+
1
)
{
merged_dims
.
push_back
(
input_x_dims
[
i
]);
}
else
{
merged_dims
[
merge_axis
+
1
]
*=
input_x_dims
[
i
];
}
}
int
full_dim
=
1
;
int
left_dim
=
1
;
for
(
auto
i
=
0
;
i
<
input_x_dims
.
size
();
i
++
)
{
for
(
int
i
=
0
;
i
<
merged_dims
.
size
();
i
++
)
{
full_strides
.
insert
(
full_strides
.
begin
(),
full_dim
);
full_dim
*=
input_x_dims
[
input_x
_dims
.
size
()
-
i
-
1
];
if
(
i
==
dim
)
{
full_dim
*=
merged_dims
[
merged
_dims
.
size
()
-
i
-
1
];
if
(
i
==
merge_axis
)
{
continue
;
}
cal_dims
.
push_back
(
i
);
}
int
left_dim
=
1
;
for
(
int
i
=
merged_dims
.
size
()
-
1
;
i
>=
0
;
i
--
)
{
if
(
i
==
merge_axis
)
{
continue
;
}
left_strides
.
insert
(
left_strides
.
begin
(),
left_dim
);
left_dim
*=
input_x_dims
[
input_x_dims
.
size
()
-
i
-
1
];
left_dim
*=
merged_dims
[
i
];
}
const
auto
*
input_x_data
=
input_x
.
data
<
T
>
();
const
auto
*
input_y_data
=
input_y
.
data
<
T
>
();
const
auto
*
input_out_grad_data
=
input_out_grad
.
data
<
T
>
();
auto
*
output_x_grad_data
=
dev_ctx
.
template
Alloc
<
T
>(
x_grad
);
auto
*
output_y_grad_data
=
dev_ctx
.
template
Alloc
<
T
>(
y_grad
);
auto
index_calculator
=
IndexCalculator
(
input_x_dims
.
size
()
-
1
,
cal_dims
,
left_strides
,
full_strides
);
auto
index_calculator
=
phi
::
funcs
::
IndexCalculator
(
merged_dims
.
size
()
-
1
,
cal_dims
,
left_strides
,
full_strides
);
int64_t
numel
=
x
.
numel
();
backends
::
gpu
::
GpuLaunchConfig
config
=
backends
::
gpu
::
GetGpuLaunchConfig1D
(
dev_ctx
,
numel
/
3
);
...
...
@@ -144,7 +162,7 @@ void CrossGradKernel(const Context& dev_ctx,
input_out_grad_data
,
output_x_grad_data
,
output_y_grad_data
,
full_strides
[
dim
],
full_strides
[
merge_axis
],
numel
/
3
,
index_calculator
);
}
...
...
paddle/phi/kernels/gpu/cross_kernel.cu
浏览文件 @
9e3433bd
...
...
@@ -22,15 +22,13 @@
namespace
phi
{
using
funcs
::
IndexCalculator
;
template
<
typename
T
>
__global__
void
Cross
(
const
T
*
x
,
const
T
*
y
,
T
*
out
,
const
int
stride
,
const
int
N
,
IndexCalculator
index_calculator
)
{
phi
::
funcs
::
IndexCalculator
index_calculator
)
{
CUDA_KERNEL_LOOP
(
i
,
N
)
{
int
offset
=
index_calculator
(
i
);
...
...
@@ -96,30 +94,50 @@ void CrossKernel(const Context& dev_ctx,
std
::
vector
<
int
>
cal_dims
;
std
::
vector
<
int
>
left_strides
;
std
::
vector
<
int
>
full_strides
;
std
::
vector
<
int
>
merged_dims
;
for
(
int
i
=
0
;
i
<
dim
;
i
++
)
{
if
(
i
==
0
)
{
merged_dims
.
push_back
(
input_x_dims
[
i
]);
}
else
{
merged_dims
[
0
]
*=
input_x_dims
[
i
];
}
}
int
merge_axis
=
merged_dims
.
size
();
merged_dims
.
push_back
(
input_x_dims
[
dim
]);
for
(
int
i
=
dim
+
1
;
i
<
input_x_dims
.
size
();
i
++
)
{
if
(
i
==
dim
+
1
)
{
merged_dims
.
push_back
(
input_x_dims
[
i
]);
}
else
{
merged_dims
[
merge_axis
+
1
]
*=
input_x_dims
[
i
];
}
}
int
dims0
=
1
;
int
dims1
=
1
;
for
(
auto
i
=
0
;
i
<
input_x_dims
.
size
();
i
++
)
{
full_strides
.
insert
(
full_strides
.
begin
(),
dims0
);
dims0
*=
input_x_dims
[
input_x_dims
.
size
()
-
i
-
1
];
if
(
i
==
dim
)
{
int
full_dim
=
1
;
for
(
int
i
=
0
;
i
<
merged_dims
.
size
();
i
++
)
{
full_strides
.
insert
(
full_strides
.
begin
(),
full_dim
);
full_dim
*=
merged_dims
[
merged_dims
.
size
()
-
i
-
1
];
if
(
i
==
merge_axis
)
{
continue
;
}
cal_dims
.
push_back
(
i
);
left_strides
.
insert
(
left_strides
.
begin
(),
dims1
);
dims1
*=
input_x_dims
[
input_x_dims
.
size
()
-
i
-
1
];
}
int
left_dim
=
1
;
for
(
int
i
=
merged_dims
.
size
()
-
1
;
i
>=
0
;
i
--
)
{
if
(
i
==
merge_axis
)
{
continue
;
}
left_strides
.
insert
(
left_strides
.
begin
(),
left_dim
);
left_dim
*=
merged_dims
[
i
];
}
const
auto
*
input_x_data
=
input_x
.
data
<
T
>
();
const
auto
*
input_y_data
=
input_y
.
data
<
T
>
();
auto
*
out_data
=
dev_ctx
.
template
Alloc
<
T
>(
out
);
auto
index_calculator
=
IndexCalculator
(
input_x_dims
.
size
()
-
1
,
cal_dims
,
left_strides
,
full_strides
);
auto
index_calculator
=
phi
::
funcs
::
IndexCalculator
(
merged_dims
.
size
()
-
1
,
cal_dims
,
left_strides
,
full_strides
);
int64_t
numel
=
x
.
numel
();
backends
::
gpu
::
GpuLaunchConfig
config
=
backends
::
gpu
::
GetGpuLaunchConfig1D
(
dev_ctx
,
numel
/
3
);
...
...
@@ -129,7 +147,7 @@ void CrossKernel(const Context& dev_ctx,
dev_ctx
.
stream
()
>>>
(
input_x_data
,
input_y_data
,
out_data
,
full_strides
[
dim
],
full_strides
[
merge_axis
],
numel
/
3
,
index_calculator
);
}
...
...
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