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
b0941102
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
b0941102
编写于
11月 02, 2021
作者:
W
wanghuancoder
提交者:
GitHub
11月 02, 2021
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
fix some bug, test=develop (#36888)
上级
093c4ec5
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
57 addition
and
10 deletion
+57
-10
paddle/fluid/framework/new_executor/interpretercore.cc
paddle/fluid/framework/new_executor/interpretercore.cc
+4
-3
paddle/fluid/framework/new_executor/interpretercore_util.cc
paddle/fluid/framework/new_executor/interpretercore_util.cc
+18
-5
paddle/fluid/framework/new_executor/new_executor_defs.h
paddle/fluid/framework/new_executor/new_executor_defs.h
+9
-0
paddle/fluid/operators/controlflow/fetch_v2_op.cc
paddle/fluid/operators/controlflow/fetch_v2_op.cc
+26
-2
未找到文件。
paddle/fluid/framework/new_executor/interpretercore.cc
浏览文件 @
b0941102
...
...
@@ -241,13 +241,14 @@ void InterpreterCore::BuildInplace() {
auto
&
outputs
=
instr
.
Outputs
();
for
(
auto
&
pair
:
in_to_outs
)
{
auto
iter
=
inputs
.
find
(
pair
.
first
);
if
(
iter
!=
inputs
.
end
())
{
if
(
iter
!=
inputs
.
end
()
&&
!
iter
->
second
.
empty
()
)
{
if
(
BuildInplaceCheckVarIsOnlyInput
(
iter
->
second
[
0
]))
{
auto
iterout
=
outputs
.
find
(
pair
.
second
);
if
(
iterout
!=
outputs
.
end
())
{
if
(
iterout
!=
outputs
.
end
()
&&
!
iterout
->
second
.
empty
()
)
{
auto
invar
=
global_scope_
->
Var
(
iter
->
second
[
0
]);
auto
outvar
=
global_scope_
->
Var
(
iterout
->
second
[
0
]);
if
(
invar
&&
outvar
)
{
if
(
invar
&&
outvar
&&
invar
->
IsType
<
LoDTensor
>
()
&&
outvar
->
IsType
<
LoDTensor
>
())
{
instr
.
AddInplace
(
invar
,
outvar
);
VLOG
(
3
)
<<
"inplace "
<<
vec_instruction_
[
i
].
OpBase
()
->
Type
()
<<
" "
<<
global_scope_
->
GetNameById
(
iter
->
second
[
0
])
...
...
paddle/fluid/framework/new_executor/interpretercore_util.cc
浏览文件 @
b0941102
...
...
@@ -142,8 +142,8 @@ void build_variable_scope(const framework::ProgramDesc& pdesc,
if
(
nullptr
==
var_scope
->
FindVar
(
var_name
))
{
var_scope
->
AddVar
(
var_desc
->
Name
(),
var_desc
);
}
else
{
auto
*
var_desc
=
var_scope
->
VarDesc
(
var_name
);
if
(
nullptr
==
var_desc
)
{
auto
*
var_desc
_tmp
=
var_scope
->
VarDesc
(
var_name
);
if
(
nullptr
==
var_desc
_tmp
)
{
VLOG
(
3
)
<<
"update var:"
<<
var_name
<<
" desc from nullptr into "
<<
var_desc
;
var_scope
->
VarMetaInfo
(
var_name
).
vardesc_
=
var_desc
;
...
...
@@ -206,9 +206,22 @@ void apply_device_guard(const OperatorBase* op_base,
VLOG
(
3
)
<<
"Switch into CPUPlace by device_guard."
;
expected_kernel_key
->
place_
=
platform
::
CPUPlace
();
}
else
if
(
op_device
.
find
(
"gpu"
)
!=
std
::
string
::
npos
&&
platform
::
is_gpu_place
(
place
))
{
VLOG
(
3
)
<<
"Switch into "
<<
place
<<
" by device_guard."
;
expected_kernel_key
->
place_
=
place
;
(
platform
::
is_gpu_place
(
place
)
||
platform
::
is_npu_place
(
place
)))
{
// when the Op that only has CPUKernel is assigned to GPU, the CPUKernel
// will be executed and a warning will be given at the same time.
if
(
op_base
->
SupportGPU
())
{
expected_kernel_key
->
place_
=
place
;
}
else
if
(
op_base
->
SupportNPU
())
{
expected_kernel_key
->
place_
=
place
;
}
else
{
expected_kernel_key
->
place_
=
platform
::
CPUPlace
();
LOG_FIRST_N
(
WARNING
,
1
)
<<
"Op("
<<
op_base
->
Type
()
<<
") has no CUDA implementation. It will be assigned to CPUPlace."
;
}
VLOG
(
3
)
<<
"Switch into "
<<
expected_kernel_key
->
place_
<<
" by device_guard."
;
}
else
{
PADDLE_THROW
(
platform
::
errors
::
Fatal
(
"Unsupported current place %s"
,
op_device
));
...
...
paddle/fluid/framework/new_executor/new_executor_defs.h
浏览文件 @
b0941102
...
...
@@ -474,6 +474,15 @@ struct VariableMetaInfo {
// TODO(zhiqiu): Maybe we need to add rwlock for VariableScope?
class
VariableScope
:
public
ScopeBase
{
public:
VariableScope
()
{
// for @EMPTY@ variable
var_list_
.
push_back
(
nullptr
);
name2id_
[
kEmptyVarName
]
=
0
;
VariableMetaInfo
info
;
info
.
var_ref_count_
=
0
;
info
.
vardesc_
=
nullptr
;
vec_meta_info_
.
push_back
(
info
);
}
Variable
*
FindVar
(
const
std
::
string
&
name
)
const
{
auto
it
=
name2id_
.
find
(
name
);
if
(
it
!=
name2id_
.
end
())
{
...
...
paddle/fluid/operators/controlflow/fetch_v2_op.cc
浏览文件 @
b0941102
...
...
@@ -77,12 +77,35 @@ class FetchV2Op : public framework::OperatorWithKernel {
framework
::
OpKernelType
GetKernelTypeForVar
(
const
std
::
string
&
var_name
,
const
framework
::
Tensor
&
tensor
,
const
framework
::
OpKernelType
&
expected_kernel_type
)
const
override
{
if
(
!
tensor
.
IsInitialized
())
{
return
expected_kernel_type
;
}
return
framework
::
OpKernelType
(
expected_kernel_type
.
data_type_
,
tensor
.
place
(),
tensor
.
layout
());
}
framework
::
OpKernelType
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
*
fetch_var
=
ctx
.
InputVar
(
"X"
);
if
(
fetch_var
==
nullptr
)
{
return
framework
::
OpKernelType
(
framework
::
proto
::
VarType
::
FP32
,
platform
::
CPUPlace
());
}
if
(
fetch_var
->
IsType
<
framework
::
LoDTensor
>
())
{
auto
&
src_item
=
fetch_var
->
Get
<
framework
::
LoDTensor
>
();
if
(
!
src_item
.
IsInitialized
())
{
return
framework
::
OpKernelType
(
framework
::
proto
::
VarType
::
FP32
,
platform
::
CPUPlace
());
}
}
else
{
auto
&
src_item
=
fetch_var
->
Get
<
framework
::
LoDTensorArray
>
();
if
(
src_item
.
empty
()
||
!
src_item
[
0
].
IsInitialized
())
{
return
framework
::
OpKernelType
(
framework
::
proto
::
VarType
::
FP32
,
platform
::
CPUPlace
());
}
}
return
framework
::
OpKernelType
(
OperatorWithKernel
::
IndicateVarDataType
(
ctx
,
"X"
),
platform
::
CPUPlace
());
...
...
@@ -127,6 +150,9 @@ class FetchV2Kernel {
if
(
fetch_var
->
IsType
<
framework
::
LoDTensor
>
())
{
auto
&
src_item
=
fetch_var
->
Get
<
framework
::
LoDTensor
>
();
if
(
!
src_item
.
IsInitialized
())
{
return
;
}
auto
*
dst_item
=
&
(
BOOST_GET
(
framework
::
LoDTensor
,
fetch_list
->
at
(
col
)));
bool
check_place
=
platform
::
is_cpu_place
(
src_item
.
place
())
||
platform
::
is_cuda_pinned_place
(
src_item
.
place
());
...
...
@@ -173,9 +199,7 @@ class FetchV2OpProtoMaker : public framework::OpProtoAndCheckerMaker {
.
SetDefault
(
true
);
AddComment
(
R"DOC(
FetchV2 Operator.
It should not be configured by users directly.
)DOC"
);
}
};
...
...
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