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
eb097d64
P
Paddle
项目概览
PaddlePaddle
/
Paddle
11 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
eb097d64
编写于
8月 28, 2020
作者:
J
joanna.wozna.intel
提交者:
GitHub
8月 28, 2020
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Fix int8 performace drop cpu_quantize_placement_pass (#26715)
* Fix cpu quantize placement pass * Include string lib
上级
02083bda
变更
5
显示空白变更内容
内联
并排
Showing
5 changed file
with
56 addition
and
27 deletion
+56
-27
paddle/fluid/framework/ir/graph_pattern_detector.cc
paddle/fluid/framework/ir/graph_pattern_detector.cc
+13
-0
paddle/fluid/framework/ir/graph_pattern_detector.h
paddle/fluid/framework/ir/graph_pattern_detector.h
+9
-0
paddle/fluid/framework/ir/mkldnn/cpu_quantize_placement_pass.cc
.../fluid/framework/ir/mkldnn/cpu_quantize_placement_pass.cc
+26
-23
paddle/fluid/framework/ir/mkldnn/cpu_quantize_placement_pass.h
...e/fluid/framework/ir/mkldnn/cpu_quantize_placement_pass.h
+6
-2
paddle/fluid/framework/ir/mkldnn/cpu_quantize_placement_pass_tester.cc
...framework/ir/mkldnn/cpu_quantize_placement_pass_tester.cc
+2
-2
未找到文件。
paddle/fluid/framework/ir/graph_pattern_detector.cc
浏览文件 @
eb097d64
...
...
@@ -1879,6 +1879,19 @@ PDNode *patterns::MultipleQuantize::operator()() {
return
prev_out
;
}
PDNode
*
patterns
::
QuantizePlacement
::
operator
()(
const
std
::
unordered_set
<
std
::
string
>
&
quantize_enabled_op_types
)
{
std
::
unordered_set
<
std
::
string
>
supported_op_types
=
std
::
unordered_set
<
std
::
string
>
({
"concat"
,
"conv2d"
,
"elementwise_add"
,
"fc"
,
"matmul"
,
"pool2d"
,
"prior_box"
,
"relu"
,
"reshape2"
,
"transpose2"
});
if
(
!
quantize_enabled_op_types
.
empty
())
{
supported_op_types
=
quantize_enabled_op_types
;
}
auto
*
op
=
pattern
->
NewNode
(
op_repr
())
->
assert_is_ops
(
supported_op_types
);
return
op
;
}
PDNode
*
patterns
::
MKLDNNInPlace
::
operator
()()
{
const
std
::
unordered_set
<
std
::
string
>
&
supported_op_types
=
{
"abs"
,
...
...
paddle/fluid/framework/ir/graph_pattern_detector.h
浏览文件 @
eb097d64
...
...
@@ -1120,6 +1120,15 @@ struct MultipleQuantize : public PatternBase {
PATTERN_DECL_NODE
(
prev_out
);
};
struct
QuantizePlacement
:
public
PatternBase
{
QuantizePlacement
(
PDPattern
*
pattern
,
const
std
::
string
&
name_scope
)
:
PatternBase
(
pattern
,
name_scope
,
"quantize_placement"
)
{}
PDNode
*
operator
()(
const
std
::
unordered_set
<
std
::
string
>&
quantize_enabled_op_types
);
PATTERN_DECL_NODE
(
op
);
};
// Pattern used for enforcing inplace computation for in-place computation
// supporting DNNL ops. softmax, batch_norm and layer_norm
struct
MKLDNNInPlace
:
public
PatternBase
{
...
...
paddle/fluid/framework/ir/mkldnn/cpu_quantize_placement_pass.cc
浏览文件 @
eb097d64
...
...
@@ -26,27 +26,33 @@ void CPUQuantizePlacementPass::ApplyImpl(ir::Graph* graph) const {
Get
<
std
::
unordered_set
<
int
>>
(
"quantize_excluded_op_ids"
);
const
auto
&
op_types_list
=
Get
<
std
::
unordered_set
<
std
::
string
>>
(
"quantize_enabled_op_types"
);
for
(
const
Node
*
n
:
graph
->
Nodes
())
{
if
(
n
->
IsOp
())
{
Init
(
name_scope_
,
graph
);
GraphPatternDetector
gpd
;
patterns
::
QuantizePlacement
quantize_placement_pattern
{
gpd
.
mutable_pattern
(),
"quantize_placement"
};
quantize_placement_pattern
(
op_types_list
);
auto
handler
=
[
&
](
const
GraphPatternDetector
::
subgraph_t
&
subgraph
,
Graph
*
g
)
{
GET_IR_NODE_FROM_SUBGRAPH
(
op
,
op
,
quantize_placement_pattern
);
if
(
std
::
find
(
excluded_ids_list
.
begin
(),
excluded_ids_list
.
end
(),
n
->
id
())
!=
excluded_ids_list
.
end
())
continue
;
auto
*
op
=
n
->
Op
();
if
(
op
->
HasAttr
(
"mkldnn_data_type"
)
||
op
->
HasProtoAttr
(
"mkldnn_data_type"
))
{
op
->
id
())
!=
excluded_ids_list
.
end
())
{
return
;
}
if
(
op
->
Op
()
->
HasAttr
(
"mkldnn_data_type"
)
||
op
->
Op
()
->
HasProtoAttr
(
"mkldnn_data_type"
))
{
// use_quantizer is no longer used
// assign value for compatibility
if
(
op
->
GetAttrIfExists
<
bool
>
(
"use_quantizer"
))
{
op
->
SetAttr
(
"mkldnn_data_type"
,
std
::
string
(
"int8"
));
}
if
(
std
::
find
(
op_types_list
.
begin
(),
op_types_list
.
end
(),
op
->
Type
())
!=
op_types_list
.
end
())
{
op
->
SetAttr
(
"mkldnn_data_type"
,
std
::
string
(
"int8"
));
op
->
SetAttr
(
"use_quantizer"
,
true
);
}
}
if
(
op
->
Op
()
->
GetAttrIfExists
<
bool
>
(
"use_quantizer"
))
{
op
->
Op
()
->
SetAttr
(
"mkldnn_data_type"
,
std
::
string
(
"int8"
));
}
op
->
Op
()
->
SetAttr
(
"mkldnn_data_type"
,
std
::
string
(
"int8"
));
op
->
Op
()
->
SetAttr
(
"use_quantizer"
,
true
);
}
};
gpd
(
graph
,
handler
);
}
}
// namespace ir
...
...
@@ -58,10 +64,7 @@ REGISTER_PASS(cpu_quantize_placement_pass,
// a vector of operator type names to be quantized ("conv2d" etc.)
// the second param is the default value for this vector
.
DefaultPassAttr
(
"quantize_enabled_op_types"
,
new
std
::
unordered_set
<
std
::
string
>
(
{
"concat"
,
"conv2d"
,
"elementwise_add"
,
"fc"
,
"matmul"
,
"pool2d"
,
"prior_box"
,
"relu"
,
"reshape2"
,
"transpose2"
}))
new
std
::
unordered_set
<
std
::
string
>
())
// a vector of operator ids that are to be excluded from quantization
// the second param is the default value for this vector
.
DefaultPassAttr
(
"quantize_excluded_op_ids"
,
new
std
::
unordered_set
<
int
>
());
paddle/fluid/framework/ir/mkldnn/cpu_quantize_placement_pass.h
浏览文件 @
eb097d64
...
...
@@ -15,7 +15,10 @@ limitations under the License. */
#pragma once
#include <memory>
#include "paddle/fluid/framework/ir/pass.h"
#include <string>
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_pattern_detector.h"
namespace
paddle
{
namespace
framework
{
...
...
@@ -23,9 +26,10 @@ namespace ir {
/*
* Specifies which operators should be quantized.
*/
class
CPUQuantizePlacementPass
:
public
Pass
{
class
CPUQuantizePlacementPass
:
public
FusePassBase
{
protected:
void
ApplyImpl
(
ir
::
Graph
*
graph
)
const
override
;
const
std
::
string
name_scope_
{
"cpu_quantize_placement_pass"
};
};
}
// namespace ir
...
...
paddle/fluid/framework/ir/mkldnn/cpu_quantize_placement_pass_tester.cc
浏览文件 @
eb097d64
...
...
@@ -131,8 +131,8 @@ TEST(QuantizerPlacementPass, enabled_conv_excluded_one) {
}
TEST
(
QuantizerPlacementPass
,
empty_list
)
{
//
no operator
quantized
MainTest
({},
{},
0
);
//
all operators
quantized
MainTest
({},
{},
6
);
}
TEST
(
QuantizerPlacementPass
,
default_attr_value
)
{
...
...
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