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
csdn_franckjun
Paddle
提交
4f5e4540
P
Paddle
项目概览
csdn_franckjun
/
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 >>
未验证
提交
4f5e4540
编写于
3月 23, 2020
作者:
A
Adam
提交者:
GitHub
3月 23, 2020
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Improve SGD jit code to work with large data (#23120)
上级
6fc6bb3e
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
43 addition
and
41 deletion
+43
-41
paddle/fluid/operators/jit/gen/sgd.cc
paddle/fluid/operators/jit/gen/sgd.cc
+42
-41
paddle/fluid/operators/jit/gen/sgd.h
paddle/fluid/operators/jit/gen/sgd.h
+1
-0
未找到文件。
paddle/fluid/operators/jit/gen/sgd.cc
浏览文件 @
4f5e4540
...
...
@@ -24,22 +24,41 @@ namespace operators {
namespace
jit
{
namespace
gen
{
void
SgdJitCode
::
mainCode
(
int
num_regs
)
{
constexpr
size_t
block_size
=
sizeof
(
float
)
*
YMM_FLOAT_BLOCK
;
// load grad
for
(
int
reg_i
=
0
;
reg_i
<
num_regs
;
++
reg_i
)
{
vmovups
(
ymm_t
(
reg_i
),
ptr
[
reg_ptr_grad_i
]);
add
(
reg_ptr_grad_i
,
block_size
);
}
// load param
for
(
int
reg_i
=
0
;
reg_i
<
num_regs
;
++
reg_i
)
{
vmovups
(
ymm_t
(
reg_i
+
num_regs
),
ptr
[
reg_ptr_param_i
]);
add
(
reg_ptr_param_i
,
block_size
);
}
// compute out
for
(
int
reg_i
=
0
;
reg_i
<
num_regs
;
++
reg_i
)
{
vmulps
(
ymm_t
(
reg_i
),
ymm_t
(
reg_i
),
ymm_lr
);
vsubps
(
ymm_t
(
reg_i
+
num_regs
),
ymm_t
(
reg_i
+
num_regs
),
ymm_t
(
reg_i
));
}
// save out
for
(
int
reg_i
=
0
;
reg_i
<
num_regs
;
++
reg_i
)
{
vmovups
(
ptr
[
reg_ptr_out_i
],
ymm_t
(
reg_i
+
num_regs
));
add
(
reg_ptr_out_i
,
block_size
);
}
}
void
SgdJitCode
::
genCode
()
{
preCode
();
constexpr
int
block
=
YMM_FLOAT_BLOCK
;
constexpr
int
max_num_regs
=
7
;
const
int
num_block
=
w_
/
block
;
const
int
num_groups
=
num_block
/
max_num_regs
;
const
size_t
block_size
=
sizeof
(
float
)
*
block
;
const
size_t
width_size
=
w_
*
sizeof
(
float
);
std
::
vector
<
int
>
groups
(
num_groups
,
max_num_regs
);
int
rest_num_regs
=
num_block
%
max_num_regs
;
if
(
rest_num_regs
>
0
)
{
groups
.
push_back
(
rest_num_regs
);
}
const
size_t
width_size
=
w_
*
sizeof
(
float
);
vbroadcastss
(
ymm_lr
,
ptr
[
param_lr
]);
// protect rdx
mov
(
reg_ptr_grad_i
,
param_grad
);
mov
(
reg_ptr_rows_i
,
param_rows
);
...
...
@@ -63,43 +82,27 @@ void SgdJitCode::genCode() {
add
(
reg_ptr_param_i
,
reg_row
);
add
(
reg_ptr_out_i
,
reg_row
);
size_t
w_offset
=
0
;
for
(
int
num_regs
:
groups
)
{
// load grad
size_t
inner_offfset
=
w_offset
;
for
(
int
reg_i
=
0
;
reg_i
<
num_regs
;
++
reg_i
)
{
vmovups
(
ymm_t
(
reg_i
),
ptr
[
reg_ptr_grad_i
+
inner_offfset
]);
inner_offfset
+=
block_size
;
}
// load param
inner_offfset
=
w_offset
;
for
(
int
reg_i
=
0
;
reg_i
<
num_regs
;
++
reg_i
)
{
vmovups
(
ymm_t
(
reg_i
+
num_regs
),
ptr
[
reg_ptr_param_i
+
inner_offfset
]);
inner_offfset
+=
block_size
;
}
// compute out
for
(
int
reg_i
=
0
;
reg_i
<
num_regs
;
++
reg_i
)
{
vmulps
(
ymm_t
(
reg_i
),
ymm_t
(
reg_i
),
ymm_lr
);
vsubps
(
ymm_t
(
reg_i
+
num_regs
),
ymm_t
(
reg_i
+
num_regs
),
ymm_t
(
reg_i
));
}
// save out
inner_offfset
=
w_offset
;
for
(
int
reg_i
=
0
;
reg_i
<
num_regs
;
++
reg_i
)
{
vmovups
(
ptr
[
reg_ptr_out_i
+
inner_offfset
],
ymm_t
(
reg_i
+
num_regs
));
inner_offfset
+=
block_size
;
}
w_offset
+=
(
block_size
*
num_regs
);
Label
inner_loop
;
Label
escape_loop
;
mov
(
rax
,
0
);
L
(
inner_loop
);
{
cmp
(
rax
,
num_groups
);
jnb
(
escape_loop
,
T_NEAR
);
mainCode
(
max_num_regs
);
inc
(
rax
);
jmp
(
inner_loop
,
T_NEAR
);
}
L
(
escape_loop
);
mainCode
(
rest_num_regs
);
add
(
reg_ptr_grad_i
,
width_size
);
add
(
reg_ptr_rows_i
,
sizeof
(
int64_t
));
cmp
(
reg_ptr_rows_i
,
reg_rows_size_in_byte
);
jl
(
l_next_row
,
T_NEAR
);
}
postCode
();
}
...
...
@@ -109,9 +112,7 @@ class SgdCreator : public JitCodeCreator<sgd_attr_t> {
return
platform
::
MayIUse
(
platform
::
avx
)
&&
attr
.
grad_width
%
YMM_FLOAT_BLOCK
==
0
;
}
size_t
CodeSize
(
const
sgd_attr_t
&
attr
)
const
override
{
return
96
+
(
attr
.
grad_width
/
YMM_FLOAT_BLOCK
)
*
32
*
8
;
}
size_t
CodeSize
(
const
sgd_attr_t
&
attr
)
const
override
{
return
96
+
32
*
8
;
}
std
::
unique_ptr
<
GenBase
>
CreateJitCode
(
const
sgd_attr_t
&
attr
)
const
override
{
PADDLE_ENFORCE_EQ
(
attr
.
param_width
,
attr
.
grad_width
);
...
...
paddle/fluid/operators/jit/gen/sgd.h
浏览文件 @
4f5e4540
...
...
@@ -34,6 +34,7 @@ class SgdJitCode : public JitCode {
DECLARE_JIT_CODE
(
SgdJitCode
);
void
genCode
()
override
;
void
mainCode
(
int
num_regs
);
private:
int
w_
;
...
...
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