Skip to content
项目
群组
代码片段
帮助
当前项目
正在载入...
登录 / 注册
切换导航面板
N
NiuTrans.Tensor
概览
Overview
Details
Activity
Cycle Analytics
版本库
Repository
Files
Commits
Branches
Tags
Contributors
Graph
Compare
Charts
问题
8
Issues
8
列表
Board
标记
里程碑
合并请求
0
Merge Requests
0
CI / CD
CI / CD
流水线
作业
日程表
图表
维基
Wiki
代码片段
Snippets
成员
Collapse sidebar
Close sidebar
活动
图像
聊天
创建新问题
作业
提交
Issue Boards
Open sidebar
NiuTrans
NiuTrans.Tensor
Commits
ecc8c6ed
Commit
ecc8c6ed
authored
Mar 31, 2021
by
xiaotong
Browse files
Options
Browse Files
Download
Email Patches
Plain Diff
define the functions of all-reduce collection
parent
6a0f0557
显示空白字符变更
内嵌
并排
正在显示
4 个修改的文件
包含
132 行增加
和
38 行删除
+132
-38
source/train/XLeaderAllReduce.cpp
+4
-4
source/train/XLeaderPS.cpp
+1
-1
source/train/XWorkerCollect.cpp
+102
-25
source/train/XWorkerCollect.h
+25
-8
没有找到文件。
source/train/XLeaderAllReduce.cpp
查看文件 @
ecc8c6ed
...
...
@@ -290,15 +290,15 @@ void XLeaderAllReduce::RunUpdate(XConfig* config, XOptimizer* optimizer, const i
if
(
finishedCount
[
j
]
==
activeModelCount
)
{
paramServer
.
flag
=
PARAM_STATE_COLLECTED
;
/*
call the all-reduce method
to collect the gradient and share
/*
run the all-reduce procedure
to collect the gradient and share
the gradient sum across models */
/* update on every model. NOTE THAT we do not worry about the
inconsistence issue of updated parameters across models because
the all-reduce method can g
arantee that the model shared the sam
e
copy of the gradient. */
the all-reduce method can g
uarantee that all the models shar
e
the same
copy of the gradient. */
for
(
int
k
=
0
;
k
<
modelNum
;
k
++
)
{
XWorkerUpdate
*
updater
=
(
XWorkerUpdate
*
)
uworkers
[
k
];
XWorkerUpdate
*
updater
=
(
XWorkerUpdate
*
)
uworkers
[
k
];
updater
->
AddJobUpdate
(
NULL
,
&
paramServer
,
optimizer
);
updater
->
AddJobEnqueueFinished
();
}
...
...
source/train/XLeaderPS.cpp
查看文件 @
ecc8c6ed
...
...
@@ -298,7 +298,7 @@ void XLeaderPS::RunUpdate(XConfig* config, XOptimizer* optimizer, const int* act
XQueue
*
jobQueue
=
(
XQueue
*
)
jobQueues
.
GetItem
(
j
);
/* data transmit */
collecter
->
AddJobCollect
DataP2P
(
jobQueue
,
paramWorker
.
grad
,
paramServer
.
grad
);
collecter
->
AddJobCollect
GradP2P
(
jobQueue
,
&
paramWorker
,
&
paramServer
);
collecter
->
AddJobEnqueueFinished
(
jobQueue
);
/* We keep the worker parameter in a list. It would be used when we broadcast
...
...
source/train/XWorkerCollect.cpp
查看文件 @
ecc8c6ed
...
...
@@ -78,33 +78,22 @@ void XWorkerCollect::CollectP2P(XTensor * source, XTensor * target)
}
/*
sum-reduce for given tensors
target += source_0
target += source_1
...
target += source_n
P2P data collection
target += source
>> source - the source tensor
>> target - the target tensor
>> isGrad - indicates whether we want to collect gradient
*/
void
XWorkerCollect
::
Collect
ReduceSum
(
XList
*
source
,
XTensor
*
target
)
void
XWorkerCollect
::
Collect
P2P
(
XTensorKeeper
*
source
,
XTensorKeeper
*
target
,
const
bool
isGrad
)
{
for
(
int
i
=
0
;
i
<
source
->
count
;
i
++
)
{
XTensor
*
s
=
(
XTensor
*
)
source
->
GetItem
(
i
);
CollectP2P
(
s
,
target
);
}
}
/*
all-reduce: the well-known all-reduce method
every tensor is involved in every data transmition. The final outcome
is that all input tensors share the same value (i.e., the sum of them).
CheckNTErrors
(
source
!=
NULL
,
"The source tensor keeper should not be NULL!"
);
CheckNTErrors
(
target
!=
NULL
,
"The target tensor keeper should not be NULL!"
);
>> all - the tensors for sum
*/
void
XWorkerCollect
::
CollectAllReduce
(
XList
*
all
)
{
ShowNTErrors
(
"TODO!"
);
if
(
isGrad
)
CollectP2P
(
source
->
grad
,
target
->
grad
);
else
CollectP2P
(
source
->
tensor
,
target
->
tensor
);
}
/* wrapper of Collect */
...
...
@@ -113,11 +102,12 @@ void XWorkerCollect::CollectDataP2P(XList * args)
int
paramCount
=
0
;
XWorkerCollect
*
collecter
=
(
XWorkerCollect
*
)
args
->
GetItem
(
paramCount
++
);
XTensor
*
source
=
(
XTensor
*
)
args
->
GetItem
(
paramCount
++
);
XTensor
*
target
=
(
XTensor
*
)
args
->
GetItem
(
paramCount
++
);
XTensorKeeper
*
source
=
(
XTensorKeeper
*
)
args
->
GetItem
(
paramCount
++
);
XTensorKeeper
*
target
=
(
XTensorKeeper
*
)
args
->
GetItem
(
paramCount
++
);
bool
isGrad
=
(
bool
)
args
->
GetInt
(
paramCount
++
);
if
(
collecter
!=
NULL
)
collecter
->
CollectP2P
(
source
,
target
);
collecter
->
CollectP2P
(
source
,
target
,
isGrad
);
}
/*
...
...
@@ -125,8 +115,9 @@ add a new job of collecting data
>> jobQueue - the queue where we run the job
>> source - where we collect the data from
>> target - where we place the data (on the server end)
>> isGrad - indicates whether we want to collect gradient
*/
bool
XWorkerCollect
::
AddJobCollectDataP2P
(
XQueue
*
jobQueue
,
XTensor
*
source
,
XTensor
*
target
)
bool
XWorkerCollect
::
AddJobCollectDataP2P
(
XQueue
*
jobQueue
,
XTensor
Keeper
*
source
,
XTensorKeeper
*
target
,
const
bool
isGrad
)
{
CheckNTErrors
(
source
!=
NULL
,
"No input soure tensor!"
);
CheckNTErrors
(
target
!=
NULL
,
"No input target tensor!"
);
...
...
@@ -135,6 +126,7 @@ bool XWorkerCollect::AddJobCollectDataP2P(XQueue * jobQueue, XTensor * source, X
args
.
Add
(
this
);
args
.
Add
(
source
);
args
.
Add
(
target
);
args
.
AddInt
((
int
)
isGrad
);
XQueue
&
queueRun
=
jobQueue
!=
NULL
?
*
jobQueue
:
queue
;
...
...
@@ -146,4 +138,89 @@ bool XWorkerCollect::AddJobCollectDataP2P(XQueue * jobQueue, XTensor * source, X
return
true
;
}
/*
add a new job of collecting gradient
>> jobQueue - the queue where we run the job
>> source - where we collect the data from
>> target - where we place the data (on the server end)
*/
bool
XWorkerCollect
::
AddJobCollectGradP2P
(
XQueue
*
jobQueue
,
XTensorKeeper
*
source
,
XTensorKeeper
*
target
)
{
return
AddJobCollectDataP2P
(
jobQueue
,
source
,
target
,
true
);
}
/*
add a new job of collecting data in standard tensors
>> jobQueue - the queue where we run the job
>> source - where we collect the data from
>> target - where we place the data (on the server end)
*/
bool
XWorkerCollect
::
AddJobCollectTensorP2P
(
XQueue
*
jobQueue
,
XTensorKeeper
*
source
,
XTensorKeeper
*
target
)
{
return
AddJobCollectDataP2P
(
jobQueue
,
source
,
target
,
false
);
}
/*
all-reduce: the well-known all-reduce method
every tensor is involved in every data transmition. The final outcome
is that all input tensors share the same value (i.e., the sum of them).
>> all - the tensors for sum
>> isGrad - indicates whether we collect gradient
*/
void
XWorkerCollect
::
CollectAllReduce
(
XList
*
all
,
const
bool
isGrad
)
{
ShowNTErrors
(
"TODO!"
);
}
/* wrapper of CollectAllReduce via all-reduce */
void
XWorkerCollect
::
CollectDataAllReduce
(
XList
*
args
)
{
int
paramCount
=
0
;
XWorkerCollect
*
collecter
=
(
XWorkerCollect
*
)
args
->
GetItem
(
paramCount
++
);
XList
*
all
=
(
XList
*
)
args
->
GetItem
(
paramCount
++
);
bool
isGrad
=
(
bool
)
args
->
GetInt
(
paramCount
++
);
if
(
collecter
!=
NULL
)
collecter
->
CollectAllReduce
(
all
,
isGrad
);
}
/*
add a new job of collecting data via all-reduce
>> jobQueue - the queue where we run the job
>> all - the tensors for sum
>> isGrad - indicates whether we collect gradient
*/
bool
XWorkerCollect
::
AddJobCollectDataAllReduce
(
XQueue
*
jobQueue
,
XList
*
all
,
const
bool
isGrad
)
{
CheckNTErrors
(
all
!=
NULL
,
"No input tensor keeper list!"
);
XList
args
;
args
.
Add
(
this
);
args
.
Add
(
all
);
args
.
AddInt
((
int
)
isGrad
);
XQueue
&
queueRun
=
jobQueue
!=
NULL
?
*
jobQueue
:
queue
;
if
(
isInstantRun
)
XWorkerCollect
::
CollectDataAllReduce
(
&
args
);
else
queueRun
.
EnqueueJob
((
void
*
)(
char
*
)
XWorkerCollect
::
CollectDataAllReduce
,
&
args
);
return
true
;
}
/* add a new job of collecting gradient via all-reduce */
bool
XWorkerCollect
::
AddJobCollectGradAllReduce
(
XQueue
*
jobQueue
,
XList
*
all
)
{
return
AddJobCollectDataAllReduce
(
jobQueue
,
all
,
true
);
}
/* add a new job of collecting data in standard tensors via all-reduce */
bool
XWorkerCollect
::
AddJobCollectTensorAllReduce
(
XQueue
*
jobQueue
,
XList
*
all
)
{
return
AddJobCollectDataAllReduce
(
jobQueue
,
all
,
false
);
}
}
source/train/XWorkerCollect.h
查看文件 @
ecc8c6ed
...
...
@@ -68,20 +68,37 @@ public:
/* P2P data collection */
void
CollectP2P
(
XTensor
*
source
,
XTensor
*
target
);
/* sum-reduce for given tensors */
void
CollectReduceSum
(
XList
*
source
,
XTensor
*
target
);
/* all-reduce */
void
CollectAllReduce
(
XList
*
all
);
/* P2P data collection */
void
CollectP2P
(
XTensorKeeper
*
source
,
XTensorKeeper
*
target
,
const
bool
isGrad
);
/* wrapper of Collect */
/* wrapper of Collect
P2P
*/
static
void
CollectDataP2P
(
XList
*
args
);
/* add a new job of collecting data */
bool
AddJobCollectDataP2P
(
XQueue
*
jobQueue
,
XTensor
*
source
,
XTensor
*
target
);
/* add a new job of collecting data via p2p data transmission */
bool
AddJobCollectDataP2P
(
XQueue
*
jobQueue
,
XTensorKeeper
*
source
,
XTensorKeeper
*
target
,
const
bool
isGrad
);
/* add a new job of collecting gradient via p2p data transmission */
bool
AddJobCollectGradP2P
(
XQueue
*
jobQueue
,
XTensorKeeper
*
source
,
XTensorKeeper
*
target
);
/* add a new job of collecting data in standard tensors via p2p data transmission */
bool
AddJobCollectTensorP2P
(
XQueue
*
jobQueue
,
XTensorKeeper
*
source
,
XTensorKeeper
*
target
);
/* all-reduce data collection via all-reduce */
void
CollectAllReduce
(
XList
*
all
,
const
bool
isGrad
);
/* wrapper of CollectAllReduce via all-reduce */
static
void
CollectDataAllReduce
(
XList
*
args
);
/* add a new job of collecting data via all-reduce */
bool
AddJobCollectDataAllReduce
(
XQueue
*
jobQueue
,
XList
*
all
,
const
bool
isGrad
);
/* add a new job of collecting gradient via all-reduce */
bool
AddJobCollectGradAllReduce
(
XQueue
*
jobQueue
,
XList
*
all
);
/* add a new job of collecting data in standard tensors via all-reduce */
bool
AddJobCollectTensorAllReduce
(
XQueue
*
jobQueue
,
XList
*
all
);
};
...
...
编写
预览
Markdown
格式
0%
重试
或
添加新文件
添加附件
取消
您添加了
0
人
到此讨论。请谨慎行事。
请先完成此评论的编辑!
取消
请
注册
或者
登录
后发表评论