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
wzpei9
Paddle
提交
3e45a5a5
P
Paddle
项目概览
wzpei9
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
1
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1
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
1
Issue
1
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 >>
提交
3e45a5a5
编写于
11月 29, 2018
作者:
Q
Qiao Longfei
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
lookup_table gpu kernel support prefetch
test=develop
上级
3a3cfc2d
变更
3
显示空白变更内容
内联
并排
Showing
3 changed file
with
56 addition
and
29 deletion
+56
-29
paddle/fluid/operators/distributed/CMakeLists.txt
paddle/fluid/operators/distributed/CMakeLists.txt
+2
-2
paddle/fluid/operators/distributed/parameter_prefetch.cc
paddle/fluid/operators/distributed/parameter_prefetch.cc
+12
-6
paddle/fluid/operators/lookup_table_op.cu
paddle/fluid/operators/lookup_table_op.cu
+42
-21
未找到文件。
paddle/fluid/operators/distributed/CMakeLists.txt
浏览文件 @
3e45a5a5
...
...
@@ -23,7 +23,7 @@ if(WITH_GRPC)
cc_test
(
rpc_server_test SRCS rpc_server_test.cc
DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor proto_desc lookup_sparse_table_op SERIAL
)
cc_test
(
varhandle_test SRCS varhandle_test.cc DEPS profiler
)
cc_library
(
parameter_prefetch SRCS parameter_prefetch.cc DEPS sendrecvop_grpc
)
cc_library
(
parameter_prefetch SRCS parameter_prefetch.cc DEPS sendrecvop_grpc
memory
)
else
()
set_source_files_properties
(
brpc_server.cc brpc_client.cc rpc_server_test.cc brpc_serde_test.cc
brpc_variable_response.cc brpc_sendrecvop_utils.cc brpc_rdma_pool.cc PROPERTIES COMPILE_FLAGS
${
DISTRIBUTE_COMPILE_FLAGS
}
)
...
...
@@ -33,7 +33,7 @@ else()
PROTO send_recv.proto
DEPS lod_tensor selected_rows memory
)
cc_library
(
parameter_prefetch SRCS parameter_prefetch.cc DEPS sendrecvop_brpc
)
cc_library
(
parameter_prefetch SRCS parameter_prefetch.cc DEPS sendrecvop_brpc
memory
)
set
(
brpc_test_depends sendrecvop_brpc brpc ssl crypto protobuf leveldb gflags glog executor proto_desc lookup_table_op snappystream snappy
)
...
...
paddle/fluid/operators/distributed/parameter_prefetch.cc
浏览文件 @
3e45a5a5
...
...
@@ -102,7 +102,8 @@ static void MergeMultipleVarsIntoOneBySection(
const
std
::
string
&
out_name
,
const
std
::
vector
<
std
::
string
>&
out_var_names
,
const
std
::
vector
<
int
>&
height_section
,
const
std
::
vector
<
std
::
vector
<
int64_t
>>&
splited_ids
,
const
framework
::
ExecutionContext
&
context
,
framework
::
Scope
*
scope
)
{
const
framework
::
ExecutionContext
&
context
,
framework
::
Scope
*
scope
,
platform
::
DeviceContext
*
actual_ctx
)
{
PADDLE_ENFORCE_EQ
(
out_var_names
.
size
(),
height_section
.
size
(),
""
);
auto
cpu_place
=
platform
::
CPUPlace
();
...
...
@@ -151,10 +152,12 @@ static void MergeMultipleVarsIntoOneBySection(
#ifndef PADDLE_WITH_CUDA
PADDLE_THROW
(
"paddle is not compiled with CUDA!"
);
#else
auto
stream
=
static_cast
<
platform
::
CUDADeviceContext
*>
(
actual_ctx
)
->
stream
();
memory
::
Copy
(
boost
::
get
<
platform
::
CUDAPlace
>
(
id_tensor
.
place
()),
out_tensor_data
+
offset
*
row_numel
,
cpu_place
,
out_var_data
+
i
*
row_numel
,
sizeof
(
float
)
*
row_numel
);
sizeof
(
float
)
*
row_numel
,
stream
);
#endif
}
}
...
...
@@ -174,6 +177,7 @@ void prefetch(const std::string& id_name, const std::string& out_name,
platform
::
DeviceContextPool
&
pool
=
platform
::
DeviceContextPool
::
Instance
();
auto
&
cpu_ctx
=
*
pool
.
Get
(
platform
::
CPUPlace
());
auto
&
actual_ctx
=
*
pool
.
Get
(
context
.
GetPlace
());
distributed
::
RPCClient
*
rpc_client
=
distributed
::
RPCClient
::
GetInstance
<
RPCCLIENT_T
>
(
...
...
@@ -201,11 +205,13 @@ void prefetch(const std::string& id_name, const std::string& out_name,
framework
::
Tensor
cpu_tensor
;
auto
*
cpu_tensor_data
=
cpu_tensor
.
mutable_data
<
int64_t
>
(
id_tensor
.
dims
(),
cpu_place
);
auto
stream
=
static_cast
<
platform
::
CUDADeviceContext
*>
(
&
actual_ctx
)
->
stream
();
memory
::
Copy
(
cpu_place
,
cpu_tensor_data
,
boost
::
get
<
platform
::
CUDAPlace
>
(
id_tensor
.
place
()),
id_tensor
.
data
<
int64_t
>
(),
s
izeof
(
int64_t
)
*
id_tensor
.
numel
()
);
for
(
size_t
i
=
0
;
i
<
id
_tensor
.
numel
();
++
i
)
{
id_tensor
.
data
<
int64_t
>
(),
sizeof
(
int64_t
)
*
id_tensor
.
numel
(),
s
tream
);
for
(
size_t
i
=
0
;
i
<
cpu
_tensor
.
numel
();
++
i
)
{
ids_vector
.
push_back
(
cpu_tensor_data
[
i
]);
}
#endif
...
...
@@ -239,7 +245,7 @@ void prefetch(const std::string& id_name, const std::string& out_name,
MergeMultipleVarsIntoOneBySection
(
id_name
,
ids_vector
,
out_name
,
out_var_names
,
height_sections
,
splited_ids
,
context
,
&
local_scope
);
context
,
&
local_scope
,
&
actual_ctx
);
context
.
scope
().
DeleteScope
(
&
local_scope
);
}
...
...
paddle/fluid/operators/lookup_table_op.cu
浏览文件 @
3e45a5a5
...
...
@@ -78,6 +78,27 @@ class LookupTableCUDAKernel : public framework::OpKernel<T> {
auto
*
output_t
=
context
.
Output
<
LoDTensor
>
(
"Out"
);
int64_t
padding_idx
=
context
.
Attr
<
int64_t
>
(
"padding_idx"
);
auto
id_name
=
context
.
Inputs
(
"Ids"
).
front
();
auto
out_name
=
context
.
Outputs
(
"Out"
).
front
();
// for remote prefetch
auto
epmap
=
context
.
Attr
<
std
::
vector
<
std
::
string
>>
(
"epmap"
);
auto
height_sections
=
context
.
Attr
<
std
::
vector
<
int
>>
(
"height_sections"
);
auto
table_names
=
context
.
Attr
<
std
::
vector
<
std
::
string
>>
(
"table_names"
);
if
(
!
epmap
.
empty
())
{
// if epmap is not empty, then the parameter will be fetched from remote
// parameter
// server
#ifdef PADDLE_WITH_DISTRIBUTE
operators
::
distributed
::
prefetch
(
id_name
,
out_name
,
table_names
,
epmap
,
height_sections
,
context
);
#else
PADDLE_THROW
(
"paddle is not compiled with distribute support, can not do "
"parameter prefetch!"
);
#endif
}
else
{
size_t
N
=
table_t
->
dims
()[
0
];
size_t
D
=
table_t
->
dims
()[
1
];
size_t
K
=
ids_t
->
numel
();
...
...
@@ -90,16 +111,15 @@ class LookupTableCUDAKernel : public framework::OpKernel<T> {
dim3
grids
(
8
,
1
);
if
(
padding_idx
==
-
1
)
LookupTable
<
T
,
128
,
8
,
8
,
false
><<<
grids
,
threads
,
0
,
context
.
cuda_device_context
().
stream
()
>>>
(
LookupTable
<
T
,
128
,
8
,
8
,
false
><<<
grids
,
threads
,
0
,
context
.
cuda_device_context
().
stream
()
>>>
(
output
,
table
,
ids
,
N
,
K
,
D
,
padding_idx
);
else
LookupTable
<
T
,
128
,
8
,
8
,
true
><<<
grids
,
threads
,
0
,
context
.
cuda_device_context
().
stream
()
>>>
(
LookupTable
<
T
,
128
,
8
,
8
,
true
><<<
grids
,
threads
,
0
,
context
.
cuda_device_context
().
stream
()
>>>
(
output
,
table
,
ids
,
N
,
K
,
D
,
padding_idx
);
}
}
};
template
<
typename
T
>
...
...
@@ -109,6 +129,7 @@ class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
auto
&
dev_ctx
=
context
.
template
device_context
<
platform
::
CUDADeviceContext
>();
bool
is_sparse
=
context
.
Attr
<
bool
>
(
"is_sparse"
);
// Since paddings are not trainable and fixed in forward, the gradient of
// paddings makes no sense and we don't deal with it in backward.
if
(
is_sparse
)
{
...
...
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