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
wmsofts
Paddle
提交
ec814cf5
P
Paddle
项目概览
wmsofts
/
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
ec814cf5
编写于
2月 27, 2023
作者:
C
csy0225
提交者:
GitHub
2月 27, 2023
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
revert operator.cc (#50895)
上级
cf209204
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
27 addition
and
63 deletion
+27
-63
paddle/fluid/framework/operator.cc
paddle/fluid/framework/operator.cc
+26
-61
paddle/fluid/framework/operator.h
paddle/fluid/framework/operator.h
+1
-2
未找到文件。
paddle/fluid/framework/operator.cc
浏览文件 @
ec814cf5
...
...
@@ -1618,57 +1618,6 @@ void OperatorWithKernel::CheckWhetherPreparePhiData(
}
}
// When do we need to reset runtime context?
// 1. When enable cache runtime context, if the program runs for the first time,
// runtime_ctx_.get() == nullptr, we need to create a new runtime context.
// 2. When enable cache runtime context, if the program is not running for the
// first time,
// but the input shape or tensor layout of the operator has changed, we cannot
// use the runtime context stored in the cache at this time, and need to
// create a new one.
bool
OperatorWithKernel
::
NeedResetRuntimeContext
(
const
Scope
&
scope
)
const
{
if
(
runtime_ctx_
.
get
()
==
nullptr
)
return
true
;
const
auto
&
name_map
=
Inputs
();
for
(
auto
&
var_name_item
:
name_map
)
{
auto
&
name_vec
=
var_name_item
.
second
;
std
::
vector
<
Variable
*>&
cache_input_vars
=
runtime_ctx_
->
inputs
[
var_name_item
.
first
];
PADDLE_ENFORCE_EQ
(
name_vec
.
size
(),
cache_input_vars
.
size
(),
platform
::
errors
::
InvalidArgument
(
"The size of input variable names (%d) must be equal to "
"the size of cache input variable ptrs (%d)."
,
name_vec
.
size
(),
cache_input_vars
.
size
()));
for
(
size_t
i
=
0
;
i
<
name_vec
.
size
();
i
++
)
{
auto
var_name
=
name_vec
[
i
];
auto
*
cache_input_var
=
cache_input_vars
[
i
];
if
(
!
VarIsTensor
(
*
cache_input_var
))
continue
;
auto
*
cache_input_tensor
=
GetMutableLoDTensorOrSelectedRowsValueFromVar
(
cache_input_var
);
auto
cache_input_tensor_dims
=
cache_input_tensor
->
dims
();
auto
*
current_input_var
=
scope
.
FindVar
(
var_name
);
PADDLE_ENFORCE_NOT_NULL
(
current_input_var
,
platform
::
errors
::
NotFound
(
"The variable %s is not found when "
"enable_cache_runtime_context_cache in origin scope."
,
var_name
));
auto
*
current_input_tensor
=
GetMutableLoDTensorOrSelectedRowsValueFromVar
(
current_input_var
);
auto
current_input_tensor_dims
=
current_input_tensor
->
dims
();
if
(
cache_input_tensor_dims
!=
current_input_tensor_dims
||
NeedTransformLayout
(
current_input_tensor
->
layout
(),
cache_input_tensor
->
layout
()))
{
need_prepare_data_
=
true
;
return
true
;
}
}
}
return
false
;
}
void
OperatorWithKernel
::
RunImpl
(
const
Scope
&
scope
,
const
platform
::
Place
&
place
)
const
{
// To reduce the elapsed time of HasAttr, we use bool variable to record the
...
...
@@ -1678,6 +1627,7 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
if
(
!
all_kernels_must_compute_runtime_shape_
&&
HasAttr
(
kAllKernelsMustComputeRuntimeShape
))
all_kernels_must_compute_runtime_shape_
=
true
;
const
Scope
*
cur_scope
=
&
scope
;
CheckWhetherPreparePhiData
(
Inputs
(),
Outputs
(),
scope
);
if
(
!
enable_cache_runtime_context_
)
{
RuntimeContext
ctx
(
Inputs
(),
Outputs
(),
scope
);
...
...
@@ -1689,9 +1639,12 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
}
(
*
phi_kernel_
)(
impl_
->
getKernelContext
());
}
else
{
if
(
NeedResetRuntimeContext
(
scope
)
)
{
if
(
runtime_ctx_
.
get
()
==
nullptr
||
pre_scope_
!=
cur_scope
)
{
std
::
lock_guard
<
std
::
mutex
>
lock
(
cache_update_mutex_
);
runtime_ctx_
.
reset
(
new
RuntimeContext
(
Inputs
(),
Outputs
(),
scope
));
if
(
runtime_ctx_
.
get
()
==
nullptr
||
pre_scope_
!=
cur_scope
)
{
runtime_ctx_
.
reset
(
new
RuntimeContext
(
Inputs
(),
Outputs
(),
scope
));
pre_scope_
=
cur_scope
;
}
}
RunImpl
(
scope
,
place
,
runtime_ctx_
.
get
());
}
...
...
@@ -2086,9 +2039,8 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
// To solve issue #15032, have a discussion with @Luotao for cpu inference,
// do not cache transfer scope, hence in this case delete transfer scope
// after run to avoid memory leak
if
(
cache_transfer_scope_
&&
!
run_by_executor_
&&
!
enable_cache_transfer_scope_
)
{
scope
.
DeleteScope
(
cache_transfer_scope_
);
if
(
transfer_scope
&&
!
run_by_executor_
&&
!
enable_cache_transfer_scope_
)
{
scope
.
DeleteScope
(
transfer_scope
);
}
}
...
...
@@ -2623,25 +2575,33 @@ Scope* OperatorWithKernel::PrepareData(
kernel_type_for_var
.
backend
()
==
phi
::
Backend
::
GPUDNN
||
new_expected_kernel_key
->
backend
()
==
phi
::
Backend
::
GPU
||
new_expected_kernel_key
->
backend
()
==
phi
::
Backend
::
GPUDNN
)
{
cache_transfer_scope_
=
TryCreateTransferScope
(
new_scope
=
TryCreateTransferScope
(
kernel_type_for_var
,
*
new_expected_kernel_key
,
&
scope
);
enable_cache_transfer_scope_
=
true
;
new_scope
=
cache_transfer_scope_
;
}
}
else
if
(
kernel_type_for_var
.
backend
()
==
phi
::
Backend
::
GPU
||
kernel_type_for_var
.
backend
()
==
phi
::
Backend
::
GPUDNN
||
expected_kernel_key
.
backend
()
==
phi
::
Backend
::
GPU
||
expected_kernel_key
.
backend
()
==
phi
::
Backend
::
GPUDNN
)
{
cache_transfer_scope_
=
TryCreateTransferScope
(
new_scope
=
TryCreateTransferScope
(
kernel_type_for_var
,
expected_kernel_key
,
&
scope
);
enable_cache_transfer_scope_
=
true
;
new_scope
=
cache_transfer_scope_
;
}
}
if
(
!
new_scope
)
{
new_scope
=
&
scope
.
NewScope
();
}
// For inference, if a gpu model has an op which could only run on CPU,
// each result of different input will be the same with the first one.
// The reason is that if a gpu tensor is the input of a cpu kernel,
// we will create a new cpu tensor in new scope.
// However, if enable_cache_runtime_context_, we get the cpu tensor each
// time, not the gpu tensor. Thus, we set pre_scope_ = nullptr
// to trigger `new RuntimeContext()` in RunImpl().
if
(
enable_cache_runtime_context_
)
{
pre_scope_
=
nullptr
;
}
// Create new var with the same name in transfer scopes
auto
*
trans_var
=
new_scope
->
Var
(
var_name
);
...
...
@@ -2727,13 +2687,18 @@ Scope* OperatorWithKernel::PrepareData(
}
}
// If pre_scope = &scope, it means that scope is cached and the op is not in
// while block. If new_scope = nullptr, it means that for each input of this
// Op, there is no need to do PrepareData. So PrepareData could be skipped at
// the rest iterations to save the elapsed time.
// We do not support skipping PrepareData in while block, because the Op's
// input may be changed by subsequent Ops, which may cause an error.
// For inference, ops that behind conditional branch aren't supported well,
// so disable prepare optimization conservatively.
bool
force_prepare_data
=
HasAttr
(
"inference_force_prepare_data"
)
&&
Attr
<
bool
>
(
"inference_force_prepare_data"
);
if
(
enable_cache_runtime_context_
&&
!
force_prepare_data
)
{
if
(
pre_scope_
==
&
scope
&&
new_scope
==
nullptr
&&
!
force_prepare_data
)
{
need_prepare_data_
=
false
;
}
...
...
paddle/fluid/framework/operator.h
浏览文件 @
ec814cf5
...
...
@@ -781,19 +781,18 @@ class OperatorWithKernel : public OperatorBase {
// used for IndicateOrPromoteVarDataTypes
phi
::
DenseTensor
*
GetTensorFormInputSafely
(
const
ExecutionContext
&
ctx
,
const
std
::
string
&
name
)
const
;
bool
NeedResetRuntimeContext
(
const
Scope
&
scope
)
const
;
protected:
mutable
std
::
unique_ptr
<
OpKernelType
>
kernel_type_
;
mutable
std
::
unique_ptr
<
OpKernelFunc
>
kernel_func_
;
mutable
std
::
unique_ptr
<
RuntimeContext
>
runtime_ctx_
;
mutable
const
Scope
*
pre_scope_
=
nullptr
;
mutable
bool
need_prepare_data_
=
true
;
mutable
bool
need_prepare_phi_data_
=
false
;
mutable
bool
enable_cache_runtime_context_
=
false
;
mutable
bool
all_kernels_must_compute_runtime_shape_
=
false
;
mutable
std
::
mutex
cache_update_mutex_
;
mutable
bool
enable_cache_transfer_scope_
=
false
;
mutable
Scope
*
cache_transfer_scope_
=
nullptr
;
// NOTE(jiahongyu): Whether fallback to plain kernel after calling
// GetExpectedKernelType, use this bool flag to solve mkldnn and cudnn hard
// code
...
...
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