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
BaiXuePrincess
Paddle
提交
c95c35a2
P
Paddle
项目概览
BaiXuePrincess
/
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
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
未验证
提交
c95c35a2
编写于
1月 12, 2023
作者:
G
gem5
提交者:
GitHub
1月 12, 2023
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
conv2d_fusion(cudnn or cutlass) (#49707)
上级
3fb4a08c
变更
1
显示空白变更内容
内联
并排
Showing
1 changed file
with
24 addition
and
21 deletion
+24
-21
paddle/fluid/framework/ir/conv2d_fusion_layout_transfer_pass.cc
.../fluid/framework/ir/conv2d_fusion_layout_transfer_pass.cc
+24
-21
未找到文件。
paddle/fluid/framework/ir/conv2d_fusion_layout_transfer_pass.cc
浏览文件 @
c95c35a2
...
...
@@ -155,7 +155,7 @@ void Conv2dFusionLayoutTransferPass::ApplyImpl(ir::Graph *graph) const {
}
#endif
if
(
!
(
is_fp16_precision
&&
cutlass_enable
)
)
return
;
if
(
!
is_fp16_precision
)
return
;
PADDLE_ENFORCE_EQ
(
graph
->
IsMainGraph
(),
true
,
...
...
@@ -180,16 +180,31 @@ void Conv2dFusionLayoutTransferPass::ApplyImpl(ir::Graph *graph) const {
std
::
string
target_op_type
=
"conv2d_fusion"
;
std
::
unordered_set
<
ir
::
Node
*>
valid_ops
;
auto
Op
IsValid
=
[
&
](
ir
::
Node
*
op_node
)
->
bool
{
auto
cuDNN
IsValid
=
[
&
](
ir
::
Node
*
op_node
)
->
bool
{
if
(
op_node
->
Op
()
->
Type
()
!=
target_op_type
)
return
false
;
auto
data_format
=
op_node
->
Op
()
->
GetAttrIfExists
<
std
::
string
>
(
"data_format"
);
if
(
data_format
!=
"NCHW"
)
return
false
;
auto
filter_names
=
op_node
->
Op
()
->
Input
(
"Filter"
);
auto
act_type
=
op_node
->
Op
()
->
GetAttrIfExists
<
std
::
string
>
(
"activation"
);
constexpr
int
CUTLASS_NHWC_ALIGNMENT
=
8
;
// If filter's channel is not multiple of 8, conv2d_fusion not run at nhwc.
for
(
const
auto
&
filter_name
:
filter_names
)
{
auto
*
filter_var
=
scope
->
FindLocalVar
(
filter_name
);
const
auto
&
filter_tensor
=
filter_var
->
Get
<
phi
::
DenseTensor
>
();
CHECK_EQ
(
filter_tensor
.
dims
().
size
()
==
4UL
,
true
);
int
oc
=
filter_tensor
.
dims
()[
0
];
int
ic
=
filter_tensor
.
dims
()[
1
];
bool
cutlass_can_support
=
oc
%
CUTLASS_NHWC_ALIGNMENT
==
0
&&
ic
%
CUTLASS_NHWC_ALIGNMENT
==
0
;
if
(
!
cutlass_can_support
)
{
return
false
;
}
}
return
true
;
};
auto
CutlassIsValid
=
[
&
](
ir
::
Node
*
op_node
)
->
bool
{
auto
act_type
=
op_node
->
Op
()
->
GetAttrIfExists
<
std
::
string
>
(
"activation"
);
// conv2d_fusion has two forms: conv + bias + act, conv + bias +
// elmentwise_add + act.
std
::
unordered_set
<
std
::
string
>
cutlass_cba_act_set
=
{
...
...
@@ -206,31 +221,19 @@ void Conv2dFusionLayoutTransferPass::ApplyImpl(ir::Graph *graph) const {
return
false
;
}
}
// If filter's channel is not multiple of 8, conv2d_fusion not run at nhwc.
for
(
const
auto
&
filter_name
:
filter_names
)
{
auto
*
filter_var
=
scope
->
FindLocalVar
(
filter_name
);
const
auto
&
filter_tensor
=
filter_var
->
Get
<
phi
::
DenseTensor
>
();
CHECK_EQ
(
filter_tensor
.
dims
().
size
()
==
4UL
,
true
);
int
oc
=
filter_tensor
.
dims
()[
0
];
int
ic
=
filter_tensor
.
dims
()[
1
];
bool
cutlass_can_support
=
oc
%
CUTLASS_NHWC_ALIGNMENT
==
0
&&
ic
%
CUTLASS_NHWC_ALIGNMENT
==
0
;
if
(
!
cutlass_can_support
)
{
return
false
;
}
}
return
true
;
};
for
(
auto
*
op_node
:
op_nodes
)
{
CHECK_EQ
(
op_node
->
IsOp
(),
true
);
if
(
Op
IsValid
(
op_node
))
{
if
(
cuDNN
IsValid
(
op_node
))
{
valid_ops
.
insert
(
op_node
);
auto
*
op_desc
=
op_node
->
Op
();
auto
nhwc_attr
=
framework
::
Attribute
(
std
::
string
(
"NHWC"
));
op_desc
->
SetAttr
(
"data_format"
,
nhwc_attr
);
if
(
cutlass_enable
&&
CutlassIsValid
(
op_node
))
{
op_desc
->
SetType
(
"conv2d_fusion_cutlass"
);
}
op_desc
->
Flush
();
// transfer weights
...
...
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