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
b849157e
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 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
b849157e
编写于
12月 19, 2018
作者:
G
gongweibao
提交者:
GitHub
12月 19, 2018
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Add size enforce (#14919)
上级
aa6e9c30
变更
4
显示空白变更内容
内联
并排
Showing
4 changed file
with
33 addition
and
9 deletion
+33
-9
paddle/fluid/operators/distributed/brpc_sendrecvop_utils.cc
paddle/fluid/operators/distributed/brpc_sendrecvop_utils.cc
+17
-6
paddle/fluid/operators/distributed/grpc_serde.cc
paddle/fluid/operators/distributed/grpc_serde.cc
+8
-0
paddle/fluid/operators/distributed/sendrecvop_utils.h
paddle/fluid/operators/distributed/sendrecvop_utils.h
+7
-2
paddle/fluid/operators/distributed/variable_response.cc
paddle/fluid/operators/distributed/variable_response.cc
+1
-1
未找到文件。
paddle/fluid/operators/distributed/brpc_sendrecvop_utils.cc
浏览文件 @
b849157e
...
...
@@ -16,6 +16,7 @@ limitations under the License. */
#include <nccl.h>
#endif
#include <sys/time.h>
#include <limits>
#include <thread> // NOLINT
#include "paddle/fluid/framework/data_type.h"
...
...
@@ -31,7 +32,12 @@ namespace distributed {
class
IOBufWriter
{
public:
static
void
Append
(
butil
::
IOBuf
*
iobuf
,
int
k
,
const
char
*
v
,
int64_t
vlen
)
{
static
void
Append
(
const
std
::
string
&
varname
,
butil
::
IOBuf
*
iobuf
,
int
k
,
const
char
*
v
,
int64_t
vlen
)
{
if
(
vlen
>=
std
::
numeric_limits
<
int
>::
max
()
||
vlen
<
0
)
{
LOG
(
FATAL
)
<<
"AppendZeroCopy varname:"
<<
varname
<<
", vlen:"
<<
vlen
;
}
iobuf
->
append
(
reinterpret_cast
<
char
*>
(
&
k
),
4
);
iobuf
->
append
(
reinterpret_cast
<
char
*>
(
&
vlen
),
8
);
iobuf
->
append
(
v
,
vlen
);
...
...
@@ -87,6 +93,10 @@ class IOBufWriter {
int
k
,
const
char
*
v
,
int64_t
vlen
,
bool
in_cuda_pinned
,
void
(
*
destroy
)(
void
*
),
void
*
user_data
)
{
if
(
vlen
>=
std
::
numeric_limits
<
int
>::
max
()
||
vlen
<
0
)
{
LOG
(
FATAL
)
<<
"AppendZeroCopy varname:"
<<
varname
<<
", vlen:"
<<
vlen
;
}
#ifdef PADDLE_WITH_BRPC_RDMA
IOBufWriter
::
AppendRdmaZeroCopy
(
varname
,
iobuf
,
k
,
v
,
vlen
,
in_cuda_pinned
,
destroy
,
user_data
);
...
...
@@ -134,7 +144,7 @@ void SerializeToIOBuf(const std::string& name, framework::Variable* var,
request
->
set_type
(
::
sendrecv
::
NCCL_ID
);
const
ncclUniqueId
&
uid
=
var
->
Get
<
ncclUniqueId
>
();
// TODO(gongwb): use append_zero to avoid data copy.
IOBufWriter
::
Append
(
iobuf
,
IOBufWriter
::
Append
(
name
,
iobuf
,
sendrecv
::
VariableMessage
::
kSerializedFieldNumber
,
uid
.
internal
,
NCCL_UNIQUE_ID_BYTES
);
return
;
...
...
@@ -149,7 +159,7 @@ void SerializeToIOBuf(const std::string& name, framework::Variable* var,
// FIXME(gongwb): it seems that can use zero copy.
if
(
var_is_not_stable
)
{
IOBufWriter
::
Append
(
iobuf
,
::
sendrecv
::
VariableMessage
::
kSerializedFieldNumber
,
name
,
iobuf
,
::
sendrecv
::
VariableMessage
::
kSerializedFieldNumber
,
static_cast
<
const
char
*>
(
payload
->
ptr
()),
payload
->
memory_size
());
}
else
{
if
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()))
{
...
...
@@ -171,10 +181,11 @@ void SerializeToIOBuf(const std::string& name, framework::Variable* var,
if
(
var
->
IsType
<
framework
::
SelectedRows
>
())
{
auto
*
slr
=
var
->
GetMutable
<
framework
::
SelectedRows
>
();
size_t
rows_memory_size
=
slr
->
rows
().
size
()
*
framework
::
SizeOfType
(
typeid
(
int64_t
)
);
PADDLE_ENFORCE
(
VectorElemName
(
slr
->
rows
())
==
typeid
(
int64_t
).
name
());
size_t
rows_memory_size
=
slr
->
rows
().
size
()
*
sizeof
(
int64_t
);
IOBufWriter
::
Append
(
iobuf
,
::
sendrecv
::
VariableMessage
::
kRowsFieldNumber
,
IOBufWriter
::
Append
(
name
,
iobuf
,
::
sendrecv
::
VariableMessage
::
kRowsFieldNumber
,
reinterpret_cast
<
const
char
*>
(
slr
->
rows
().
data
()),
static_cast
<
int64_t
>
(
rows_memory_size
));
}
...
...
paddle/fluid/operators/distributed/grpc_serde.cc
浏览文件 @
b849157e
...
...
@@ -15,6 +15,7 @@ limitations under the License. */
#ifdef PADDLE_WITH_CUDA
#include <nccl.h>
#endif
#include <limits>
#include <thread> // NOLINT
#include "google/protobuf/io/coded_stream.h"
...
...
@@ -102,6 +103,10 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
e
.
WriteVarlengthBeginning
(
VarMsg
::
kSerializedFieldNumber
,
payload
->
memory_size
());
if
(
payload
->
memory_size
()
>=
std
::
numeric_limits
<
int
>::
max
())
{
LOG
(
FATAL
)
<<
"AppendZeroCopy varname:"
<<
name
<<
", vlen:"
<<
payload
->
memory_size
();
}
// steal reference of tensor data
::
grpc
::
Slice
slices
[
4
];
// metadata, tensor, rows meta, rows
int
num_slices
=
2
;
// only SelectedRows have rows buffer
...
...
@@ -115,7 +120,10 @@ void SerializeToByteBuffer(const std::string& name, framework::Variable* var,
if
(
var
->
IsType
<
framework
::
SelectedRows
>
())
{
auto
*
slr
=
var
->
GetMutable
<
framework
::
SelectedRows
>
();
ProtoEncodeHelper
e2
(
static_cast
<
char
*>
(
buf
),
128
);
PADDLE_ENFORCE
(
VectorElemName
(
slr
->
rows
())
==
typeid
(
int64_t
).
name
());
size_t
rows_memory_size
=
slr
->
rows
().
size
()
*
sizeof
(
int64_t
);
e2
.
WriteVarlengthBeginning
(
VarMsg
::
kRowsFieldNumber
,
rows_memory_size
);
slices
[
2
]
=
::
grpc
::
Slice
(
e2
.
size
());
memcpy
(
const_cast
<
uint8_t
*>
(
slices
[
2
].
begin
()),
e2
.
data
(),
e2
.
size
());
...
...
paddle/fluid/operators/distributed/sendrecvop_utils.h
浏览文件 @
b849157e
...
...
@@ -15,6 +15,7 @@ limitations under the License. */
#pragma once
#include <iostream>
#include <string>
#include <typeindex>
#include <vector>
#include "paddle/fluid/framework/data_type.h"
...
...
@@ -23,9 +24,8 @@ limitations under the License. */
#include "paddle/fluid/framework/selected_rows.h"
#include "paddle/fluid/framework/tensor_util.h"
#include "paddle/fluid/framework/var_type.h"
#include "paddle/fluid/platform/port.h"
#include "paddle/fluid/operators/distributed/send_recv.pb.h"
#include "paddle/fluid/platform/port.h"
namespace
paddle
{
namespace
operators
{
...
...
@@ -83,6 +83,11 @@ inline framework::proto::VarType::Type ToVarType(
}
}
template
<
template
<
typename
>
class
T
,
typename
Elem
>
std
::
string
VectorElemName
(
const
T
<
Elem
>&
arg
)
{
return
typeid
(
Elem
).
name
();
}
}
// namespace distributed
}
// namespace operators
}
// namespace paddle
paddle/fluid/operators/distributed/variable_response.cc
浏览文件 @
b849157e
...
...
@@ -118,7 +118,7 @@ bool VariableResponse::CopyLodTensorData(
VLOG
(
6
)
<<
"Tensor.memory_size = "
<<
tensor
->
memory_size
()
<<
", Buffer Size = "
<<
length
;
PADDLE_ENFORCE_EQ
(
tensor
->
memory_size
(),
length
);
PADDLE_ENFORCE_EQ
(
tensor
->
memory_size
(),
static_cast
<
unsigned
int
>
(
length
)
);
return
ReadRaw
(
input
,
ctx
,
tensor
->
place
(),
tensor_data
,
length
);
}
...
...
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