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
c7acad9f
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 个月 前同步成功
通知
2292
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
体验新版 GitCode,发现更多精彩内容 >>
未验证
提交
c7acad9f
编写于
12月 25, 2020
作者:
T
taixiurong
提交者:
GitHub
12月 25, 2020
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
support some shape for matmul and cast in xpu place (#29900)
* support some shape in matmul and cast * modify matmul
上级
80eb7778
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
34 addition
and
0 deletion
+34
-0
paddle/fluid/operators/cast_op_xpu.cc
paddle/fluid/operators/cast_op_xpu.cc
+6
-0
paddle/fluid/operators/matmul_op_xpu.cc
paddle/fluid/operators/matmul_op_xpu.cc
+28
-0
未找到文件。
paddle/fluid/operators/cast_op_xpu.cc
浏览文件 @
c7acad9f
...
...
@@ -49,6 +49,12 @@ class CastXPUKernel : public framework::OpKernel<InT> {
auto
*
out_data
=
out
->
mutable_data
<
int64_t
>
(
context
.
GetPlace
());
r
=
xpu
::
cast_v2
<
InT
,
int64_t
>
(
dev_ctx
.
x_context
(),
in_data
,
out_data
,
numel
);
}
else
if
((
out_type
==
framework
::
proto
::
VarType
::
BOOL
)
&&
(
in_type
==
framework
::
proto
::
VarType
::
FP32
))
{
auto
*
out_data
=
out
->
mutable_data
<
bool
>
(
context
.
GetPlace
());
r
=
xpu
::
cast_v2
<
float
,
int8_t
>
(
dev_ctx
.
x_context
(),
(
const
float
*
)
in_data
,
reinterpret_cast
<
int8_t
*>
(
out_data
),
numel
);
}
else
{
PADDLE_THROW
(
platform
::
errors
::
Unavailable
(
"Not supported cast %d -> %d"
,
in_type
,
out_type
));
...
...
paddle/fluid/operators/matmul_op_xpu.cc
浏览文件 @
c7acad9f
...
...
@@ -111,6 +111,20 @@ class MatMulXPUKernel : public framework::OpKernel<T> {
auto
mat_dim_b
=
math
::
CreateMatrixDescriptor
(
ColumnMatrixFromVector
(
y
->
dims
()),
0
,
context
.
Attr
<
bool
>
(
"transpose_Y"
));
const
auto
&
x_dims
=
x
->
dims
();
const
auto
&
y_dims
=
y
->
dims
();
if
(
x_dims
.
size
()
==
3
&&
y_dims
.
size
()
<=
2
)
{
// if transpose_X is true, the transpose cost much time
if
(
!
context
.
Attr
<
bool
>
(
"transpose_X"
))
{
mat_dim_a
.
height_
*=
mat_dim_a
.
batch_size_
;
mat_dim_a
.
batch_size_
=
0
;
}
else
{
mat_dim_b
.
batch_size_
=
mat_dim_a
.
batch_size_
;
mat_dim_b
.
height_
=
mat_dim_b
.
height_
/
mat_dim_b
.
batch_size_
;
}
}
PADDLE_ENFORCE_EQ
(
mat_dim_a
.
width_
,
mat_dim_b
.
height_
,
platform
::
errors
::
InvalidArgument
(
"Shape mistake in matmul_op"
));
...
...
@@ -224,12 +238,26 @@ class MatMulGradXPUKernel : public framework::OpKernel<T> {
out
->
mutable_data
<
T
>
(
context
.
GetPlace
());
auto
mat_dim_a
=
math
::
CreateMatrixDescriptor
(
a
.
dims
(),
0
,
trans_a
);
auto
mat_dim_b
=
math
::
CreateMatrixDescriptor
(
b
.
dims
(),
0
,
trans_b
);
const
auto
&
a_dims
=
a
.
dims
();
const
auto
&
b_dims
=
b
.
dims
();
if
(
a_dims
.
size
()
==
3
&&
b_dims
.
size
()
<=
2
)
{
// if transpose_X is true, the transpose cost much time
if
(
!
context
.
Attr
<
bool
>
(
"transpose_X"
))
{
mat_dim_a
.
height_
*=
mat_dim_a
.
batch_size_
;
mat_dim_a
.
batch_size_
=
0
;
}
else
{
mat_dim_b
.
batch_size_
=
mat_dim_a
.
batch_size_
;
mat_dim_b
.
height_
=
mat_dim_b
.
height_
/
mat_dim_b
.
batch_size_
;
}
}
PADDLE_ENFORCE_EQ
(
mat_dim_a
.
width_
,
mat_dim_b
.
height_
,
platform
::
errors
::
InvalidArgument
(
"Shape mistake in matmul_grad_op"
));
PADDLE_ENFORCE_EQ
(
mat_dim_a
.
batch_size_
,
mat_dim_b
.
batch_size_
,
platform
::
errors
::
InvalidArgument
(
"Shape mistake in matmul_grad_op"
));
T
alpha
=
static_cast
<
T
>
(
context
.
Attr
<
float
>
(
"alpha"
));
auto
&
dev_ctx
=
context
.
template
device_context
<
DeviceContext
>();
...
...
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