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
14fe40aa
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2298
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
14fe40aa
编写于
3月 14, 2018
作者:
D
dzhwinter
提交者:
GitHub
3月 14, 2018
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Refine/nccl (#9009)
* "Refine nccl op" * "refine code " * "refine nccl code"
上级
788c600e
变更
2
显示空白变更内容
内联
并排
Showing
2 changed file
with
89 addition
and
142 deletion
+89
-142
paddle/fluid/operators/nccl_op.cc
paddle/fluid/operators/nccl_op.cc
+45
-47
paddle/fluid/operators/nccl_op.cu.cc
paddle/fluid/operators/nccl_op.cu.cc
+44
-95
未找到文件。
paddle/fluid/operators/nccl_op.cc
浏览文件 @
14fe40aa
...
...
@@ -104,19 +104,38 @@ class NCCLAllReduceOp : public framework::OperatorWithKernel {
" Input(Communicator) of AllReduce op input should not be NULL"
);
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"Out"
),
" Output(Out) of AllReduce op output should not be NULL"
);
auto
x_dims
=
ctx
->
GetInputsDim
(
"X"
);
std
::
string
reduction
=
ctx
->
Attrs
().
Get
<
std
::
string
>
(
"reduction"
);
PADDLE_ENFORCE
((
reduction
==
"ncclSum"
||
reduction
==
"ncclProd"
||
reduction
==
"ncclMin"
||
reduction
==
"ncclMax"
),
"invalid reduction."
);
auto
x_dims
=
ctx
->
GetInputsDim
(
"X"
);
ctx
->
SetOutputsDim
(
"Out"
,
x_dims
);
ctx
->
ShareLoD
(
"X"
,
/*->*/
"Out"
);
}
};
// AllReduceOp
class
NCCLAllReduceOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
NCCLAllReduceOpMaker
(
OpProto
*
proto
,
OpAttrChecker
*
op_checker
)
:
OpProtoAndCheckerMaker
(
proto
,
op_checker
)
{
AddInput
(
"X"
,
"The input of AllReduce op"
);
AddInput
(
"Communicator"
,
"Communicator for communicating between gpus"
);
AddOutput
(
"Out"
,
"The output of AllReduce op"
);
AddAttr
<
std
::
string
>
(
"reduction"
,
"(string, default 'ncclSum') "
"{'ncclMin', 'ncclMax', 'ncclProd', 'ncclSum'}."
)
.
SetDefault
(
"ncclSum"
);
AddComment
(
R"DOC(
NCCLAllReduce Operator.
AllReduce the input tensors.
)DOC"
);
}
};
// ReduceOp
class
NCCLReduceOp
:
public
framework
::
OperatorWithKernel
{
public:
...
...
@@ -143,50 +162,6 @@ class NCCLReduceOp : public framework::OperatorWithKernel {
}
};
// BcastOp
class
NCCLBcastOp
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
protected:
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"X"
),
" Input(X) of Bcast op input should not be NULL"
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Communicator"
),
" Input(Communicator) of Bcast op input should not be NULL"
);
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"Out"
),
" Output(Out) of Bcast op output should not be NULL"
);
int
root
=
ctx
->
Attrs
().
Get
<
int
>
(
"root"
);
PADDLE_ENFORCE
(
root
!=
platform
::
kInvalidGPUId
,
"Bcast root must be set."
);
auto
x_dims
=
ctx
->
GetInputsDim
(
"X"
);
ctx
->
SetOutputsDim
(
"Out"
,
x_dims
);
ctx
->
ShareLoD
(
"X"
,
/*->*/
"Out"
);
}
};
// AllreduceOp
class
NCCLAllReduceOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
NCCLAllReduceOpMaker
(
OpProto
*
proto
,
OpAttrChecker
*
op_checker
)
:
OpProtoAndCheckerMaker
(
proto
,
op_checker
)
{
AddInput
(
"X"
,
"The input of AllReduce op"
);
AddInput
(
"Communicator"
,
"Communicator for communicating between gpus"
);
AddOutput
(
"Out"
,
"The output of AllReduce op"
);
AddAttr
<
std
::
string
>
(
"reduction"
,
"(string, default 'ncclSum') "
"{'ncclMin', 'ncclMax', 'ncclProd', 'ncclSum'}."
)
.
SetDefault
(
"ncclSum"
);
AddComment
(
R"DOC(
NCCLAllReduce Operator.
AllReduce the input tensors.
)DOC"
);
}
};
// ReduceOp
class
NCCLReduceOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
...
...
@@ -213,6 +188,29 @@ Reduce the tensors.
}
};
// BcastOp
class
NCCLBcastOp
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
protected:
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"X"
),
" Input(X) of Bcast op input should not be NULL"
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Communicator"
),
" Input(Communicator) of Bcast op input should not be NULL"
);
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"Out"
),
" Output(Out) of Bcast op output should not be NULL"
);
int
root
=
ctx
->
Attrs
().
Get
<
int
>
(
"root"
);
PADDLE_ENFORCE
(
root
!=
platform
::
kInvalidGPUId
,
"Bcast root must be set."
);
auto
x_dims
=
ctx
->
GetInputsDim
(
"X"
);
ctx
->
SetOutputsDim
(
"Out"
,
x_dims
);
ctx
->
ShareLoD
(
"X"
,
/*->*/
"Out"
);
}
};
// BcastOp
class
NCCLBcastOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
...
...
paddle/fluid/operators/nccl_op.cu.cc
浏览文件 @
14fe40aa
...
...
@@ -43,13 +43,12 @@ class NCCLAllReduceKernel : public framework::OpKernel<T> {
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
PADDLE_ENFORCE
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()),
"This kernel only runs on GPU device."
);
auto
ins
=
ctx
.
MultiInput
<
LoDTensor
>
(
"X"
);
auto
outs
=
ctx
.
MultiOutput
<
LoDTensor
>
(
"Out"
);
auto
*
x
=
ctx
.
Input
<
LoDTensor
>
(
"X"
);
auto
*
out
=
ctx
.
Output
<
LoDTensor
>
(
"Out"
);
auto
*
comm
=
ctx
.
Input
<
Communicator
>
(
"Communicator"
);
std
::
string
reduction
=
ctx
.
Attr
<
std
::
string
>
(
"reduction"
);
ncclRedOp_t
reduction_op_
=
ncclSum
;
ncclRedOp_t
reduction_op_
=
ncclSum
;
if
(
reduction
==
"ncclMin"
)
{
reduction_op_
=
ncclMin
;
}
else
if
(
reduction
==
"ncclMax"
)
{
...
...
@@ -61,30 +60,19 @@ class NCCLAllReduceKernel : public framework::OpKernel<T> {
}
else
{
PADDLE_THROW
(
"Invalid reduction. default ncclSum."
);
}
auto
*
comm
=
ctx
.
Input
<
Communicator
>
(
"Communicator"
);
auto
stream
=
ctx
.
cuda_device_context
().
stream
();
// device id
int
gpu_id
=
boost
::
get
<
platform
::
CUDAPlace
>
(
ctx
.
GetPlace
()).
GetDeviceId
();
int
idx
=
comm
->
GetCommId
(
gpu_id
);
for
(
size_t
i
=
0
;
i
<
ins
.
size
();
++
i
)
{
VLOG
(
1
)
<<
"gpu : "
<<
" invoke allreduce. send "
<<
ins
[
i
]
->
numel
()
<<
" recv "
<<
outs
[
i
]
->
numel
();
VLOG
(
3
)
<<
"gpu : "
<<
" invoke allreduce. send "
<<
x
->
numel
()
<<
" recv "
<<
out
->
numel
();
PADDLE_ENFORCE
(
platform
::
dynload
::
ncclAllReduce
(
ins
[
i
]
->
data
<
T
>
(),
outs
[
i
]
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()),
outs
[
i
]
->
numel
(),
NCCLTypeWrapper
<
T
>::
type
,
reduction_op_
,
comm
->
comms
().
at
(
idx
),
stream
));
PADDLE_ENFORCE
(
cudaStreamSynchronize
(
stream
));
VLOG
(
1
)
<<
"gpu : "
<<
" finished allreduce. send "
<<
ins
[
i
]
->
numel
()
<<
" recv "
<<
outs
[
i
]
->
numel
();
}
x
->
data
<
T
>
(),
out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()),
out
->
numel
(),
NCCLTypeWrapper
<
T
>::
type
,
reduction_op_
,
comm
->
comms
().
at
(
idx
),
ctx
.
cuda_device_context
().
stream
()));
VLOG
(
3
)
<<
"gpu : "
<<
" finished allreduce. send "
<<
x
->
numel
()
<<
" recv "
<<
out
->
numel
();
}
};
...
...
@@ -94,13 +82,13 @@ class NCCLReduceKernel : public framework::OpKernel<T> {
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
PADDLE_ENFORCE
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()),
"This kernel only runs on GPU device."
);
auto
ins
=
ctx
.
MultiInput
<
LoDTensor
>
(
"X"
);
// x0, x1, x2
auto
outs
=
ctx
.
MultiOutput
<
LoDTensor
>
(
"Out
"
);
auto
x
=
ctx
.
Input
<
LoDTensor
>
(
"X"
);
// x0, x1, x2
auto
out
=
ctx
.
Output
<
LoDTensor
>
(
"Out"
);
auto
*
comm
=
ctx
.
Input
<
Communicator
>
(
"Communicator
"
);
int
root
=
ctx
.
Attr
<
int
>
(
"root"
);
std
::
string
reduction
=
ctx
.
Attr
<
std
::
string
>
(
"reduction"
);
ncclRedOp_t
reduction_op_
=
ncclSum
;
ncclRedOp_t
reduction_op_
=
ncclSum
;
if
(
reduction
==
"ncclMin"
)
{
reduction_op_
=
ncclMin
;
}
else
if
(
reduction
==
"ncclMax"
)
{
...
...
@@ -112,40 +100,21 @@ class NCCLReduceKernel : public framework::OpKernel<T> {
}
else
{
PADDLE_THROW
(
"Invalid reduction. default ncclSum."
);
}
int
root
=
ctx
.
Attr
<
int
>
(
"root"
);
auto
*
comm
=
ctx
.
Input
<
Communicator
>
(
"Communicator"
);
auto
stream
=
reinterpret_cast
<
const
platform
::
CUDADeviceContext
&>
(
ctx
.
device_context
())
.
stream
();
// device id
int
gpu_id
=
boost
::
get
<
platform
::
CUDAPlace
>
(
ctx
.
GetPlace
()).
GetDeviceId
();
int
idx
=
comm
->
GetCommId
(
gpu_id
);
auto
ins_names
=
ctx
.
Inputs
(
"X"
);
std
::
hash
<
std
::
string
>
hasher
;
for
(
size_t
i
=
0
;
i
<
ins
.
size
();
++
i
)
{
if
(
root
==
platform
::
kInvalidGPUId
)
{
root
=
hasher
(
ins_names
[
i
])
%
comm
->
comms
().
size
();
}
T
*
recvbuffer
=
nullptr
;
if
(
root
==
gpu_id
)
{
recvbuffer
=
outs
[
i
]
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
recvbuffer
=
out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
}
VLOG
(
1
)
<<
"gpu : "
<<
gpu_id
<<
" invoke reduce. send "
<<
ins
[
i
]
->
numel
()
<<
" recv "
<<
outs
[
i
]
->
numel
();
VLOG
(
3
)
<<
"gpu : "
<<
gpu_id
<<
" invoke reduce. send "
<<
x
->
numel
()
<<
" recv "
<<
out
->
numel
();
PADDLE_ENFORCE
(
platform
::
dynload
::
ncclReduce
(
ins
[
i
]
->
data
<
T
>
(),
recvbuffer
,
ins
[
i
]
->
numel
(),
NCCLTypeWrapper
<
T
>::
type
,
reduction_op_
,
root
,
comm
->
comms
().
at
(
idx
),
stream
));
PADDLE_ENFORCE
(
cudaStreamSynchronize
(
stream
));
VLOG
(
1
)
<<
"gpu : "
<<
gpu_id
<<
" finished reduce. send "
<<
ins
[
i
]
->
numel
()
<<
" recv "
<<
outs
[
i
]
->
numel
();
}
x
->
data
<
T
>
(),
recvbuffer
,
x
->
numel
(),
NCCLTypeWrapper
<
T
>::
type
,
reduction_op_
,
root
,
comm
->
comms
().
at
(
idx
),
ctx
.
cuda_device_context
().
stream
()));
VLOG
(
3
)
<<
"gpu : "
<<
gpu_id
<<
" finished reduce. send "
<<
x
->
numel
()
<<
" recv "
<<
out
->
numel
();
}
};
...
...
@@ -155,47 +124,27 @@ class NCCLBcastKernel : public framework::OpKernel<T> {
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
PADDLE_ENFORCE
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()),
"This kernel only runs on GPU device."
);
int
root
=
ctx
.
Attr
<
int
>
(
"root"
);
auto
*
comm
=
ctx
.
Input
<
Communicator
>
(
"Communicator"
);
auto
stream
=
reinterpret_cast
<
const
platform
::
CUDADeviceContext
&>
(
ctx
.
device_context
())
.
stream
();
// device id
int
gpu_id
=
boost
::
get
<
platform
::
CUDAPlace
>
(
ctx
.
GetPlace
()).
GetDeviceId
();
int
idx
=
comm
->
GetCommId
(
gpu_id
);
if
(
idx
==
root
)
{
auto
ins
=
ctx
.
MultiInput
<
LoDTensor
>
(
"X"
);
for
(
size_t
i
=
0
;
i
<
ins
.
size
();
++
i
)
{
VLOG
(
1
)
<<
"gpu : "
<<
gpu_id
<<
" invoke Bcast. send "
<<
ins
[
i
]
->
numel
();
VLOG
(
1
)
<<
" before ncclBcast"
;
auto
*
x
=
ctx
.
Input
<
LoDTensor
>
(
"X"
);
VLOG
(
3
)
<<
"gpu : "
<<
gpu_id
<<
" invoke Bcast. send "
<<
x
->
numel
();
PADDLE_ENFORCE
(
platform
::
dynload
::
ncclBcast
(
(
void
*
)
ins
[
i
]
->
data
<
T
>
(),
ins
[
i
]
->
numel
(),
NCCLTypeWrapper
<
T
>::
type
,
root
,
comm
->
comms
().
at
(
idx
),
stream
));
VLOG
(
1
)
<<
" after ncclBcast"
;
PADDLE_ENFORCE
(
cudaStreamSynchronize
(
stream
));
VLOG
(
1
)
<<
"gpu : "
<<
gpu_id
<<
" finished Bcast."
;
}
(
void
*
)
x
->
data
<
T
>
(),
x
->
numel
(),
NCCLTypeWrapper
<
T
>::
type
,
root
,
comm
->
comms
().
at
(
idx
),
ctx
.
cuda_device_context
().
stream
()));
VLOG
(
3
)
<<
"gpu : "
<<
gpu_id
<<
" finished Bcast."
;
}
else
{
auto
outs
=
ctx
.
MultiOutput
<
LoDTensor
>
(
"Out"
);
for
(
size_t
i
=
0
;
i
<
outs
.
size
();
++
i
)
{
VLOG
(
1
)
<<
"gpu : "
<<
gpu_id
<<
" invoke Bcast. recv buffer "
<<
framework
::
product
(
outs
[
i
]
->
dims
());
auto
*
out
=
ctx
.
Output
<
LoDTensor
>
(
"Out"
);
VLOG
(
3
)
<<
"gpu : "
<<
gpu_id
<<
" invoke Bcast. recv buffer "
<<
framework
::
product
(
out
->
dims
());
PADDLE_ENFORCE
(
platform
::
dynload
::
ncclBcast
(
outs
[
i
]
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()),
outs
[
i
]
->
numel
(),
NCCLTypeWrapper
<
T
>::
type
,
root
,
comm
->
comms
().
at
(
idx
),
stream
));
PADDLE_ENFORCE
(
cudaStreamSynchronize
(
stream
));
VLOG
(
1
)
<<
"gpu : "
<<
gpu_id
<<
" finished Bcast. recv "
<<
outs
[
i
]
->
numel
();
}
out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()),
out
->
numel
(),
NCCLTypeWrapper
<
T
>::
type
,
root
,
comm
->
comms
().
at
(
idx
),
ctx
.
cuda_device_context
().
stream
()));
VLOG
(
3
)
<<
"gpu : "
<<
gpu_id
<<
" finished Bcast. recv "
<<
out
->
numel
();
}
}
};
...
...
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