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
afeixing77
Paddle
提交
1efef8ba
P
Paddle
项目概览
afeixing77
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
0
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
0
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
0
Issue
0
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
1efef8ba
编写于
12月 15, 2020
作者:
S
ShenLiang
提交者:
GitHub
12月 15, 2020
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Fix bug of matmul_v2 for broadcast case (#29599)
* fix bug of matmul_v2 for broadcast
上级
a9082082
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
55 addition
and
21 deletion
+55
-21
paddle/fluid/operators/matmul_v2_op.h
paddle/fluid/operators/matmul_v2_op.h
+31
-21
python/paddle/fluid/tests/unittests/test_matmul_v2_op.py
python/paddle/fluid/tests/unittests/test_matmul_v2_op.py
+24
-0
未找到文件。
paddle/fluid/operators/matmul_v2_op.h
浏览文件 @
1efef8ba
...
...
@@ -44,7 +44,6 @@ template <typename DeviceContext, typename T>
void
ReduceSumForMatmulGrad
(
const
Tensor
*
input
,
Tensor
*
output
,
const
std
::
vector
<
int
>&
reduce_dims
,
const
paddle
::
framework
::
ExecutionContext
&
ctx
)
{
if
(
reduce_dims
.
empty
())
return
;
#ifdef __NVCC__
auto
stream
=
ctx
.
cuda_device_context
().
stream
();
TensorReduce
<
T
,
T
,
cub
::
Sum
,
IdentityFunctor
<
T
>>
(
...
...
@@ -602,47 +601,48 @@ class MatMulV2GradKernel : public framework::OpKernel<T> {
// So we should avoid the case in reality.
VLOG
(
3
)
<<
"It need cost much time to reduce sum for the broadcast and "
"wastes the memory. So we should avoid the case in reality"
;
Tensor
dx_help
,
dy_help
;
if
(
transpose_x
)
{
if
(
transpose_y
)
{
// X'Y': dA = Y'G', dB = G'X'
if
(
dx
)
MatMulFunction
<
DeviceContext
,
T
>
(
&
y
,
&
dout
,
y_dims
,
dout_dims
,
dx
,
true
,
true
,
ctx
);
MatMulFunction
<
DeviceContext
,
T
>
(
&
y
,
&
dout
,
y_dims
,
dout_dims
,
&
dx_help
,
true
,
true
,
ctx
);
if
(
dy
)
MatMulFunction
<
DeviceContext
,
T
>
(
&
dout
,
&
x
,
dout_dims
,
x_dims
,
dy
,
true
,
true
,
ctx
);
MatMulFunction
<
DeviceContext
,
T
>
(
&
dout
,
&
x
,
dout_dims
,
x_dims
,
&
dy_help
,
true
,
true
,
ctx
);
}
else
{
// X'Y: dX = YG', dY = XG
if
(
dx
)
MatMulFunction
<
DeviceContext
,
T
>
(
&
y
,
&
dout
,
y_dims
,
dout_dims
,
dx
,
false
,
true
,
ctx
);
MatMulFunction
<
DeviceContext
,
T
>
(
&
y
,
&
dout
,
y_dims
,
dout_dims
,
&
dx_help
,
false
,
true
,
ctx
);
if
(
dy
)
MatMulFunction
<
DeviceContext
,
T
>
(
&
x
,
&
dout
,
x_dims
,
dout_dims
,
dy
,
false
,
false
,
ctx
);
MatMulFunction
<
DeviceContext
,
T
>
(
&
x
,
&
dout
,
x_dims
,
dout_dims
,
&
dy_help
,
false
,
false
,
ctx
);
}
}
else
{
if
(
transpose_y
)
{
// XY': dX = GY, dY = G'X
if
(
dx
)
MatMulFunction
<
DeviceContext
,
T
>
(
&
dout
,
&
y
,
dout_dims
,
y_dims
,
dx
,
false
,
false
,
ctx
);
MatMulFunction
<
DeviceContext
,
T
>
(
&
dout
,
&
y
,
dout_dims
,
y_dims
,
&
dx_help
,
false
,
false
,
ctx
);
if
(
dy
)
MatMulFunction
<
DeviceContext
,
T
>
(
&
dout
,
&
x
,
dout_dims
,
x_dims
,
dy
,
true
,
false
,
ctx
);
MatMulFunction
<
DeviceContext
,
T
>
(
&
dout
,
&
x
,
dout_dims
,
x_dims
,
&
dy_help
,
true
,
false
,
ctx
);
}
else
{
// XY: dX = GY', dY = X'G
if
(
dx
)
MatMulFunction
<
DeviceContext
,
T
>
(
&
dout
,
&
y
,
dout_dims
,
y_dims
,
dx
,
false
,
true
,
ctx
);
MatMulFunction
<
DeviceContext
,
T
>
(
&
dout
,
&
y
,
dout_dims
,
y_dims
,
&
dx_help
,
false
,
true
,
ctx
);
if
(
dy
)
MatMulFunction
<
DeviceContext
,
T
>
(
&
x
,
&
dout
,
x_dims
,
dout_dims
,
dy
,
true
,
false
,
ctx
);
MatMulFunction
<
DeviceContext
,
T
>
(
&
x
,
&
dout
,
x_dims
,
dout_dims
,
&
dy_help
,
true
,
false
,
ctx
);
}
}
// get help dims
const
std
::
vector
<
std
::
int64_t
>
dx_help_dims
=
vectorize
(
dx
->
dims
());
const
std
::
vector
<
std
::
int64_t
>
dy_help_dims
=
vectorize
(
dy
->
dims
());
const
std
::
vector
<
std
::
int64_t
>
dx_help_dims
=
vectorize
(
dx
_help
.
dims
());
const
std
::
vector
<
std
::
int64_t
>
dy_help_dims
=
vectorize
(
dy
_help
.
dims
());
std
::
vector
<
std
::
int64_t
>
dx_broadcast_dims
(
ndim
);
std
::
vector
<
std
::
int64_t
>
dy_broadcast_dims
(
ndim
);
...
...
@@ -668,11 +668,21 @@ class MatMulV2GradKernel : public framework::OpKernel<T> {
}
// reduce sum to get grad by ReduceSum
if
(
dx
)
{
ReduceSumForMatmulGrad
<
DeviceContext
,
T
>
(
dx
,
dx
,
dx_reduce_dims
,
ctx
);
if
(
dx_reduce_dims
.
empty
())
{
*
dx
=
std
::
move
(
dx_help
);
}
else
{
ReduceSumForMatmulGrad
<
DeviceContext
,
T
>
(
&
dx_help
,
dx
,
dx_reduce_dims
,
ctx
);
}
dx
->
Resize
(
x
.
dims
());
}
if
(
dy
)
{
ReduceSumForMatmulGrad
<
DeviceContext
,
T
>
(
dy
,
dy
,
dy_reduce_dims
,
ctx
);
if
(
dy_reduce_dims
.
empty
())
{
*
dy
=
std
::
move
(
dy_help
);
}
else
{
ReduceSumForMatmulGrad
<
DeviceContext
,
T
>
(
&
dy_help
,
dy
,
dy_reduce_dims
,
ctx
);
}
dy
->
Resize
(
y
.
dims
());
}
}
...
...
python/paddle/fluid/tests/unittests/test_matmul_v2_op.py
浏览文件 @
1efef8ba
...
...
@@ -286,6 +286,30 @@ class TestMatMuklOp17(TestMatMulV2Op):
self
.
trans_y
=
False
class
TestMatMuklOpBroadcast1
(
TestMatMulV2Op
):
"""
case 14_3
"""
def
config
(
self
):
self
.
x_shape
=
(
3
,
1
,
10
,
10
)
self
.
y_shape
=
(
1
,
2
,
10
,
10
)
self
.
trans_x
=
True
self
.
trans_y
=
True
class
TestMatMuklOpBroadcast2
(
TestMatMulV2Op
):
"""
case 14_4
"""
def
config
(
self
):
self
.
x_shape
=
(
3
,
1
,
10
,
10
)
self
.
y_shape
=
(
1
,
2
,
10
,
10
)
self
.
trans_x
=
False
self
.
trans_y
=
True
#--------------------test matmul fp16--------------------
...
...
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