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
TonyTonyFun
Paddle
提交
68bfa0cd
P
Paddle
项目概览
TonyTonyFun
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
0
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
0
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
0
Issue
0
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
68bfa0cd
编写于
8月 31, 2022
作者:
H
HongyuJia
提交者:
GitHub
8月 31, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
migrate update_loss_scaling and check_finite_and_upscale xpu to phi, test=kunlun (#45569)
上级
077aa382
变更
3
显示空白变更内容
内联
并排
Showing
3 changed file
with
288 addition
and
331 deletion
+288
-331
paddle/fluid/operators/amp/check_finite_and_unscale_op_xpu.cc
...le/fluid/operators/amp/check_finite_and_unscale_op_xpu.cc
+0
-142
paddle/fluid/operators/amp/update_loss_scaling_op_xpu.cc
paddle/fluid/operators/amp/update_loss_scaling_op_xpu.cc
+0
-189
paddle/phi/kernels/xpu/amp_kernel.cc
paddle/phi/kernels/xpu/amp_kernel.cc
+288
-0
未找到文件。
paddle/fluid/operators/amp/check_finite_and_unscale_op_xpu.cc
已删除
100644 → 0
浏览文件 @
077aa382
/* Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#ifdef PADDLE_WITH_XPU
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/amp/fp16_type_traits.h"
#include "paddle/fluid/platform/device/device_wrapper.h"
#include "paddle/fluid/platform/float16.h"
namespace
paddle
{
namespace
operators
{
template
<
typename
T
>
class
CheckFiniteAndUnscaleXPUKernel
:
public
framework
::
OpKernel
<
T
>
{
using
MPDType
=
typename
details
::
MPTypeTrait
<
T
>::
Type
;
using
XPUTyp
=
typename
XPUTypeTrait
<
T
>::
Type
;
using
float16
=
typename
XPUTypeTrait
<
paddle
::
platform
::
float16
>::
Type
;
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
{
auto
&
dev_ctx
=
ctx
.
template
device_context
<
platform
::
XPUDeviceContext
>();
const
auto
xs
=
ctx
.
MultiInput
<
framework
::
Tensor
>
(
"X"
);
const
auto
*
scale
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Scale"
);
auto
outs
=
ctx
.
MultiOutput
<
framework
::
Tensor
>
(
"Out"
);
auto
*
found_inf
=
ctx
.
Output
<
framework
::
Tensor
>
(
"FoundInfinite"
);
const
MPDType
*
scale_data
=
scale
->
data
<
MPDType
>
();
bool
*
found_inf_data
=
found_inf
->
mutable_data
<
bool
>
(
dev_ctx
.
GetPlace
());
// cpy to cpu
bool
cpu_found_inf_data
=
false
;
// number of inf and nans
int
nums_inf_nans
=
0
;
MPDType
cpu_scale_data
;
if
(
platform
::
is_xpu_place
(
scale
->
place
()))
{
memory
::
Copy
(
platform
::
CPUPlace
(),
static_cast
<
void
*>
(
&
cpu_scale_data
),
scale
->
place
(),
static_cast
<
const
void
*>
(
scale_data
),
sizeof
(
MPDType
));
}
else
{
cpu_scale_data
=
(
*
scale_data
);
}
MPDType
inverse_scale
=
1.0
/
cpu_scale_data
;
for
(
size_t
i
=
0
;
i
<
xs
.
size
();
++
i
)
{
const
auto
*
x
=
xs
[
i
];
auto
*
out
=
outs
[
i
];
out
->
mutable_data
<
T
>
(
dev_ctx
.
GetPlace
());
framework
::
Tensor
inf_nan_count
=
ctx
.
AllocateTmpTensor
<
int
,
platform
::
XPUDeviceContext
>
(
found_inf
->
dims
(),
dev_ctx
);
if
(
nums_inf_nans
==
0
)
{
int
r
=
xpu
::
count_nan_or_inf
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
XPUTyp
*>
(
x
->
data
<
T
>
()),
inf_nan_count
.
data
<
int
>
(),
x
->
numel
());
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"count_nan_or_inf"
);
memory
::
Copy
(
platform
::
CPUPlace
(),
&
nums_inf_nans
,
dev_ctx
.
GetPlace
(),
inf_nan_count
.
data
<
int
>
(),
sizeof
(
int
));
}
if
(
nums_inf_nans
>
0
)
{
cpu_found_inf_data
=
true
;
inverse_scale
=
0.0
;
}
auto
version
=
platform
::
get_xpu_version
(
ctx
.
GetPlace
().
GetDeviceId
());
framework
::
Tensor
float_x
;
framework
::
Tensor
float_out
;
if
(
std
::
is_same
<
T
,
paddle
::
platform
::
float16
>::
value
&&
(
version
==
phi
::
backends
::
xpu
::
XPUVersion
::
XPU1
))
{
float_x
.
mutable_data
<
MPDType
>
(
dev_ctx
.
GetPlace
(),
x
->
numel
()
*
sizeof
(
MPDType
));
float_out
.
mutable_data
<
MPDType
>
(
dev_ctx
.
GetPlace
(),
out
->
numel
()
*
sizeof
(
MPDType
));
int
r
=
xpu
::
cast_v2
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
float16
*>
(
x
->
data
<
T
>
()),
float_x
.
data
<
MPDType
>
(),
x
->
numel
());
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"cast_v2"
);
r
=
xpu
::
scale
(
dev_ctx
.
x_context
(),
float_x
.
data
<
MPDType
>
(),
float_out
.
data
<
MPDType
>
(),
x
->
numel
(),
false
,
inverse_scale
,
0.0
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"scale"
);
r
=
xpu
::
cast_v2
(
dev_ctx
.
x_context
(),
float_out
.
data
<
MPDType
>
(),
reinterpret_cast
<
float16
*>
(
out
->
data
<
T
>
()),
out
->
numel
());
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"cast_v2"
);
}
else
{
int
r
=
xpu
::
scale
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
XPUTyp
*>
(
x
->
data
<
T
>
()),
reinterpret_cast
<
XPUTyp
*>
(
out
->
data
<
T
>
()),
x
->
numel
(),
false
,
inverse_scale
,
0.0
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"scale"
);
}
}
memory
::
Copy
(
dev_ctx
.
GetPlace
(),
found_inf_data
,
platform
::
CPUPlace
(),
&
cpu_found_inf_data
,
sizeof
(
bool
));
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
namespace
plat
=
paddle
::
platform
;
REGISTER_OP_XPU_KERNEL
(
check_finite_and_unscale
,
ops
::
CheckFiniteAndUnscaleXPUKernel
<
float
>
,
ops
::
CheckFiniteAndUnscaleXPUKernel
<
plat
::
float16
>
);
#endif
paddle/fluid/operators/amp/update_loss_scaling_op_xpu.cc
已删除
100644 → 0
浏览文件 @
077aa382
/* Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#ifdef PADDLE_WITH_XPU
#include <cstring>
#include <string>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/amp/fp16_type_traits.h"
#include "paddle/fluid/platform/float16.h"
namespace
paddle
{
namespace
operators
{
using
Tensor
=
framework
::
Tensor
;
template
<
typename
T
>
class
UpdateLossScalingXPUKernel
:
public
framework
::
OpKernel
<
T
>
{
using
MPDType
=
typename
details
::
MPTypeTrait
<
T
>::
Type
;
using
XPUTyp
=
typename
XPUTypeTrait
<
T
>::
Type
;
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
&
dev_ctx
=
ctx
.
template
device_context
<
platform
::
XPUDeviceContext
>();
const
auto
xs
=
ctx
.
MultiInput
<
framework
::
Tensor
>
(
"X"
);
auto
outs
=
ctx
.
MultiOutput
<
framework
::
Tensor
>
(
"Out"
);
const
auto
*
found_inf
=
ctx
.
Input
<
Tensor
>
(
"FoundInfinite"
);
PADDLE_ENFORCE_EQ
(
found_inf
->
numel
(),
1
,
platform
::
errors
::
InvalidArgument
(
"FoundInfinite must has only one element."
));
const
bool
*
found_inf_data
=
found_inf
->
data
<
bool
>
();
bool
cpu_found_inf_data
=
false
;
if
(
platform
::
is_xpu_place
(
found_inf
->
place
()))
{
memory
::
Copy
(
platform
::
CPUPlace
(),
static_cast
<
void
*>
(
&
cpu_found_inf_data
),
found_inf
->
place
(),
static_cast
<
const
void
*>
(
found_inf_data
),
sizeof
(
bool
));
}
else
{
cpu_found_inf_data
=
(
*
found_inf_data
);
}
for
(
size_t
i
=
0
;
i
<
xs
.
size
();
++
i
)
{
auto
*
out
=
outs
[
i
];
T
*
out_data
=
out
->
mutable_data
<
T
>
(
dev_ctx
.
GetPlace
());
int
num
=
out
->
numel
();
if
(
cpu_found_inf_data
)
{
VLOG
(
1
)
<<
"-- UpdateLossScaling: Find infinite grads. --"
;
int
r
=
0
;
r
=
xpu
::
constant
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
XPUTyp
*>
(
out_data
),
num
,
XPUTyp
(
0.0
));
PADDLE_ENFORCE_EQ
(
r
,
XPU_SUCCESS
,
platform
::
errors
::
External
(
"XPU API(constant) return wrong "
"value[%d %s]"
,
r
,
XPUAPIErrorMsg
[
r
]));
}
}
const
bool
stop_update
=
ctx
.
Attr
<
bool
>
(
"stop_update"
);
if
(
stop_update
)
{
return
;
}
const
auto
*
pre_loss_scaling
=
ctx
.
Input
<
Tensor
>
(
"PrevLossScaling"
);
const
auto
*
good_in
=
ctx
.
Input
<
Tensor
>
(
"InGoodSteps"
);
const
auto
*
bad_in
=
ctx
.
Input
<
Tensor
>
(
"InBadSteps"
);
auto
*
updated_loss_scaling
=
ctx
.
Output
<
Tensor
>
(
"LossScaling"
);
auto
*
good_out
=
ctx
.
Output
<
Tensor
>
(
"OutGoodSteps"
);
auto
*
bad_out
=
ctx
.
Output
<
Tensor
>
(
"OutBadSteps"
);
const
MPDType
*
pre_loss_scaling_data
=
pre_loss_scaling
->
data
<
MPDType
>
();
const
int
*
good_in_data
=
good_in
->
data
<
int
>
();
const
int
*
bad_in_data
=
bad_in
->
data
<
int
>
();
MPDType
*
updated_loss_scaling_data
=
updated_loss_scaling
->
mutable_data
<
MPDType
>
(
dev_ctx
.
GetPlace
());
int
*
good_out_data
=
good_out
->
mutable_data
<
int
>
(
dev_ctx
.
GetPlace
());
int
*
bad_out_data
=
bad_out
->
mutable_data
<
int
>
(
dev_ctx
.
GetPlace
());
const
int
incr_every_n_steps
=
ctx
.
Attr
<
int
>
(
"incr_every_n_steps"
);
const
int
decr_every_n_nan_or_inf
=
ctx
.
Attr
<
int
>
(
"decr_every_n_nan_or_inf"
);
const
float
incr_ratio
=
ctx
.
Attr
<
float
>
(
"incr_ratio"
);
const
float
decr_ratio
=
ctx
.
Attr
<
float
>
(
"decr_ratio"
);
int
cpu_bad_in_data
;
int
cpu_good_in_data
;
MPDType
cpu_pre_loss_scaling_data
;
if
(
platform
::
is_xpu_place
(
bad_in
->
place
()))
{
memory
::
Copy
(
platform
::
CPUPlace
(),
static_cast
<
void
*>
(
&
cpu_bad_in_data
),
bad_in
->
place
(),
static_cast
<
const
void
*>
(
bad_in_data
),
sizeof
(
int
));
}
else
{
cpu_bad_in_data
=
(
*
bad_in_data
);
}
if
(
platform
::
is_xpu_place
(
good_in
->
place
()))
{
memory
::
Copy
(
platform
::
CPUPlace
(),
static_cast
<
void
*>
(
&
cpu_good_in_data
),
good_in
->
place
(),
static_cast
<
const
void
*>
(
good_in_data
),
sizeof
(
int
));
}
else
{
cpu_good_in_data
=
(
*
good_in_data
);
}
if
(
platform
::
is_xpu_place
(
pre_loss_scaling
->
place
()))
{
memory
::
Copy
(
platform
::
CPUPlace
(),
static_cast
<
void
*>
(
&
cpu_pre_loss_scaling_data
),
pre_loss_scaling
->
place
(),
static_cast
<
const
void
*>
(
pre_loss_scaling_data
),
sizeof
(
MPDType
));
}
else
{
cpu_pre_loss_scaling_data
=
(
*
pre_loss_scaling_data
);
}
int
cpu_good_out_data
=
0
;
int
cpu_bad_out_data
=
0
;
MPDType
cpu_updated_loss_scaling_data
=
cpu_pre_loss_scaling_data
;
if
(
cpu_found_inf_data
)
{
cpu_good_out_data
=
0
;
cpu_bad_out_data
=
cpu_bad_in_data
+
1
;
if
(
cpu_bad_out_data
==
decr_every_n_nan_or_inf
)
{
MPDType
new_loss_scaling
=
cpu_pre_loss_scaling_data
*
decr_ratio
;
cpu_updated_loss_scaling_data
=
(
new_loss_scaling
<
static_cast
<
MPDType
>
(
1
))
?
(
static_cast
<
MPDType
>
(
1
))
:
(
new_loss_scaling
);
cpu_bad_out_data
=
0
;
}
}
else
{
cpu_bad_out_data
=
0
;
cpu_good_out_data
=
cpu_good_in_data
+
1
;
if
(
cpu_good_out_data
==
incr_every_n_steps
)
{
MPDType
new_loss_scaling
=
cpu_pre_loss_scaling_data
*
incr_ratio
;
cpu_updated_loss_scaling_data
=
(
std
::
isfinite
(
new_loss_scaling
))
?
new_loss_scaling
:
cpu_pre_loss_scaling_data
;
cpu_good_out_data
=
0
;
}
}
// copy to device
memory
::
Copy
(
dev_ctx
.
GetPlace
(),
bad_out_data
,
platform
::
CPUPlace
(),
&
cpu_bad_out_data
,
sizeof
(
int
));
memory
::
Copy
(
dev_ctx
.
GetPlace
(),
good_out_data
,
platform
::
CPUPlace
(),
&
cpu_good_out_data
,
sizeof
(
int
));
memory
::
Copy
(
dev_ctx
.
GetPlace
(),
updated_loss_scaling_data
,
platform
::
CPUPlace
(),
&
cpu_updated_loss_scaling_data
,
sizeof
(
MPDType
));
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
namespace
plat
=
paddle
::
platform
;
REGISTER_OP_XPU_KERNEL
(
update_loss_scaling
,
ops
::
UpdateLossScalingXPUKernel
<
float
>
,
ops
::
UpdateLossScalingXPUKernel
<
plat
::
float16
>
);
#endif
paddle/phi/kernels/xpu/amp_kernel.cc
0 → 100644
浏览文件 @
68bfa0cd
/* Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/phi/kernels/amp_kernel.h"
#include <cstring>
#include <string>
#include <vector>
#include "paddle/fluid/memory/memcpy.h"
#include "paddle/phi/backends/xpu/enforce_xpu.h"
#include "paddle/phi/backends/xpu/xpu_context.h"
#include "paddle/phi/common/amp_type_traits.h"
#include "paddle/phi/common/float16.h"
#include "paddle/phi/core/kernel_registry.h"
namespace
phi
{
template
<
typename
T
,
typename
Context
>
void
UpdateLossScalingKernel
(
const
Context
&
dev_ctx
,
const
std
::
vector
<
const
DenseTensor
*>&
xs
,
const
DenseTensor
&
found_infinite
,
const
DenseTensor
&
prev_loss_scaling
,
const
DenseTensor
&
in_good_steps
,
const
DenseTensor
&
in_bad_steps
,
int
incr_every_n_steps
,
int
decr_every_n_nan_or_inf
,
float
incr_ratio
,
float
decr_ratio
,
const
Scalar
&
stop_update
,
std
::
vector
<
DenseTensor
*>
outs
,
DenseTensor
*
loss_scaling
,
DenseTensor
*
out_good_steps
,
DenseTensor
*
out_bad_steps
)
{
using
MPDType
=
typename
phi
::
dtype
::
MPTypeTrait
<
T
>::
Type
;
using
XPUTyp
=
typename
XPUTypeTrait
<
T
>::
Type
;
PADDLE_ENFORCE_EQ
(
found_infinite
.
numel
(),
1
,
phi
::
errors
::
InvalidArgument
(
"FoundInfinite must has only one element."
));
const
bool
*
found_inf_data
=
found_infinite
.
data
<
bool
>
();
bool
cpu_found_inf_data
=
false
;
if
(
found_infinite
.
place
().
GetType
()
==
phi
::
AllocationType
::
XPU
)
{
paddle
::
memory
::
Copy
(
phi
::
CPUPlace
(),
static_cast
<
void
*>
(
&
cpu_found_inf_data
),
found_infinite
.
place
(),
static_cast
<
const
void
*>
(
found_inf_data
),
sizeof
(
bool
));
}
else
{
cpu_found_inf_data
=
(
*
found_inf_data
);
}
for
(
size_t
i
=
0
;
i
<
xs
.
size
();
++
i
)
{
auto
*
out
=
outs
[
i
];
T
*
out_data
=
dev_ctx
.
template
Alloc
<
T
>(
out
);
int
num
=
out
->
numel
();
if
(
cpu_found_inf_data
)
{
VLOG
(
1
)
<<
"-- UpdateLossScaling: Find infinite grads. --"
;
int
r
=
0
;
r
=
xpu
::
constant
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
XPUTyp
*>
(
out_data
),
num
,
XPUTyp
(
0.0
));
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"constant"
);
}
}
if
(
stop_update
.
to
<
bool
>
())
{
return
;
}
const
MPDType
*
pre_loss_scaling_data
=
prev_loss_scaling
.
data
<
MPDType
>
();
const
int
*
good_in_data
=
in_good_steps
.
data
<
int
>
();
const
int
*
bad_in_data
=
in_bad_steps
.
data
<
int
>
();
MPDType
*
updated_loss_scaling_data
=
dev_ctx
.
template
Alloc
<
MPDType
>(
loss_scaling
);
int
*
good_out_data
=
dev_ctx
.
template
Alloc
<
int
>(
out_good_steps
);
int
*
bad_out_data
=
dev_ctx
.
template
Alloc
<
int
>(
out_bad_steps
);
int
cpu_bad_in_data
;
int
cpu_good_in_data
;
MPDType
cpu_pre_loss_scaling_data
;
if
(
in_bad_steps
.
place
().
GetType
()
==
phi
::
AllocationType
::
XPU
)
{
paddle
::
memory
::
Copy
(
phi
::
CPUPlace
(),
static_cast
<
void
*>
(
&
cpu_bad_in_data
),
in_bad_steps
.
place
(),
static_cast
<
const
void
*>
(
bad_in_data
),
sizeof
(
int
));
}
else
{
cpu_bad_in_data
=
(
*
bad_in_data
);
}
if
(
in_good_steps
.
place
().
GetType
()
==
phi
::
AllocationType
::
XPU
)
{
paddle
::
memory
::
Copy
(
phi
::
CPUPlace
(),
static_cast
<
void
*>
(
&
cpu_good_in_data
),
in_good_steps
.
place
(),
static_cast
<
const
void
*>
(
good_in_data
),
sizeof
(
int
));
}
else
{
cpu_good_in_data
=
(
*
good_in_data
);
}
if
(
prev_loss_scaling
.
place
().
GetType
()
==
phi
::
AllocationType
::
XPU
)
{
paddle
::
memory
::
Copy
(
phi
::
CPUPlace
(),
static_cast
<
void
*>
(
&
cpu_pre_loss_scaling_data
),
prev_loss_scaling
.
place
(),
static_cast
<
const
void
*>
(
pre_loss_scaling_data
),
sizeof
(
MPDType
));
}
else
{
cpu_pre_loss_scaling_data
=
(
*
pre_loss_scaling_data
);
}
int
cpu_good_out_data
=
0
;
int
cpu_bad_out_data
=
0
;
MPDType
cpu_updated_loss_scaling_data
=
cpu_pre_loss_scaling_data
;
if
(
cpu_found_inf_data
)
{
cpu_good_out_data
=
0
;
cpu_bad_out_data
=
cpu_bad_in_data
+
1
;
if
(
cpu_bad_out_data
==
decr_every_n_nan_or_inf
)
{
MPDType
new_loss_scaling
=
cpu_pre_loss_scaling_data
*
decr_ratio
;
cpu_updated_loss_scaling_data
=
(
new_loss_scaling
<
static_cast
<
MPDType
>
(
1
))
?
(
static_cast
<
MPDType
>
(
1
))
:
(
new_loss_scaling
);
cpu_bad_out_data
=
0
;
}
}
else
{
cpu_bad_out_data
=
0
;
cpu_good_out_data
=
cpu_good_in_data
+
1
;
if
(
cpu_good_out_data
==
incr_every_n_steps
)
{
MPDType
new_loss_scaling
=
cpu_pre_loss_scaling_data
*
incr_ratio
;
cpu_updated_loss_scaling_data
=
(
std
::
isfinite
(
new_loss_scaling
))
?
new_loss_scaling
:
cpu_pre_loss_scaling_data
;
cpu_good_out_data
=
0
;
}
}
// copy to device
paddle
::
memory
::
Copy
(
dev_ctx
.
GetPlace
(),
bad_out_data
,
phi
::
CPUPlace
(),
&
cpu_bad_out_data
,
sizeof
(
int
));
paddle
::
memory
::
Copy
(
dev_ctx
.
GetPlace
(),
good_out_data
,
phi
::
CPUPlace
(),
&
cpu_good_out_data
,
sizeof
(
int
));
paddle
::
memory
::
Copy
(
dev_ctx
.
GetPlace
(),
updated_loss_scaling_data
,
phi
::
CPUPlace
(),
&
cpu_updated_loss_scaling_data
,
sizeof
(
MPDType
));
}
template
<
typename
T
,
typename
Context
>
void
CheckFiniteAndUnscaleKernel
(
const
Context
&
dev_ctx
,
const
std
::
vector
<
const
DenseTensor
*>&
xs
,
const
DenseTensor
&
scale
,
std
::
vector
<
DenseTensor
*>
outs
,
DenseTensor
*
found_infinite
)
{
using
MPDType
=
typename
phi
::
dtype
::
MPTypeTrait
<
T
>::
Type
;
using
XPUType
=
typename
XPUTypeTrait
<
T
>::
Type
;
using
float16
=
typename
XPUTypeTrait
<
phi
::
dtype
::
float16
>::
Type
;
const
MPDType
*
scale_data
=
scale
.
data
<
MPDType
>
();
bool
*
found_inf_data
=
dev_ctx
.
template
Alloc
<
bool
>(
found_infinite
);
// cpy to cpu
bool
cpu_found_inf_data
=
false
;
// number of inf and nans
int
nums_inf_nans
=
0
;
MPDType
cpu_scale_data
;
if
(
scale
.
place
().
GetType
()
==
phi
::
AllocationType
::
XPU
)
{
paddle
::
memory
::
Copy
(
phi
::
CPUPlace
(),
static_cast
<
void
*>
(
&
cpu_scale_data
),
scale
.
place
(),
static_cast
<
const
void
*>
(
scale_data
),
sizeof
(
MPDType
));
}
else
{
cpu_scale_data
=
(
*
scale_data
);
}
MPDType
inverse_scale
=
1.0
/
cpu_scale_data
;
for
(
size_t
i
=
0
;
i
<
xs
.
size
();
++
i
)
{
const
auto
*
x
=
xs
[
i
];
auto
*
out
=
outs
[
i
];
dev_ctx
.
template
Alloc
<
T
>(
out
);
DenseTensor
inf_nan_count
;
inf_nan_count
.
Resize
(
found_infinite
->
dims
());
dev_ctx
.
template
Alloc
<
int
>(
&
inf_nan_count
);
if
(
nums_inf_nans
==
0
)
{
int
r
=
xpu
::
count_nan_or_inf
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
XPUType
*>
(
x
->
data
<
T
>
()),
inf_nan_count
.
data
<
int
>
(),
x
->
numel
());
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"count_nan_or_inf"
);
paddle
::
memory
::
Copy
(
phi
::
CPUPlace
(),
&
nums_inf_nans
,
dev_ctx
.
GetPlace
(),
inf_nan_count
.
data
<
int
>
(),
sizeof
(
int
));
}
if
(
nums_inf_nans
>
0
)
{
cpu_found_inf_data
=
true
;
inverse_scale
=
0.0
;
}
auto
version
=
phi
::
backends
::
xpu
::
get_xpu_version
(
dev_ctx
.
GetPlace
().
GetDeviceId
());
DenseTensor
float_x
;
DenseTensor
float_out
;
if
(
std
::
is_same
<
T
,
phi
::
dtype
::
float16
>::
value
&&
(
version
==
phi
::
backends
::
xpu
::
XPUVersion
::
XPU1
))
{
dev_ctx
.
template
Alloc
<
MPDType
>(
&
float_x
,
x
->
numel
()
*
sizeof
(
MPDType
));
dev_ctx
.
template
Alloc
<
MPDType
>(
&
float_out
,
out
->
numel
()
*
sizeof
(
MPDType
));
int
r
=
xpu
::
cast_v2
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
float16
*>
(
x
->
data
<
T
>
()),
float_x
.
data
<
MPDType
>
(),
x
->
numel
());
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"cast_v2"
);
r
=
xpu
::
scale
(
dev_ctx
.
x_context
(),
float_x
.
data
<
MPDType
>
(),
float_out
.
data
<
MPDType
>
(),
x
->
numel
(),
false
,
inverse_scale
,
0.0
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"scale"
);
r
=
xpu
::
cast_v2
(
dev_ctx
.
x_context
(),
float_out
.
data
<
MPDType
>
(),
reinterpret_cast
<
float16
*>
(
out
->
data
<
T
>
()),
out
->
numel
());
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"cast_v2"
);
}
else
{
int
r
=
xpu
::
scale
(
dev_ctx
.
x_context
(),
reinterpret_cast
<
const
XPUType
*>
(
x
->
data
<
T
>
()),
reinterpret_cast
<
XPUType
*>
(
out
->
data
<
T
>
()),
x
->
numel
(),
false
,
inverse_scale
,
0.0
);
PADDLE_ENFORCE_XDNN_SUCCESS
(
r
,
"scale"
);
}
}
paddle
::
memory
::
Copy
(
dev_ctx
.
GetPlace
(),
found_inf_data
,
phi
::
CPUPlace
(),
&
cpu_found_inf_data
,
sizeof
(
bool
));
}
}
// namespace phi
PD_REGISTER_KERNEL
(
update_loss_scaling
,
XPU
,
ALL_LAYOUT
,
phi
::
UpdateLossScalingKernel
,
float
,
phi
::
dtype
::
float16
)
{}
PD_REGISTER_KERNEL
(
check_finite_and_unscale
,
XPU
,
ALL_LAYOUT
,
phi
::
CheckFiniteAndUnscaleKernel
,
float
,
phi
::
dtype
::
float16
)
{}
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