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
44bdbe93
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
44bdbe93
编写于
7月 16, 2021
作者:
Y
Yuang Liu
提交者:
GitHub
7月 16, 2021
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
softmax mask fuse op, test=develop (#33841)
上级
380bc4e6
变更
7
隐藏空白更改
内联
并排
Showing
7 changed file
with
843 addition
and
1 deletion
+843
-1
paddle/fluid/operators/fused_softmax_mask_op.cc
paddle/fluid/operators/fused_softmax_mask_op.cc
+117
-0
paddle/fluid/operators/fused_softmax_mask_op.cu
paddle/fluid/operators/fused_softmax_mask_op.cu
+538
-0
paddle/fluid/operators/fused_softmax_mask_op.h
paddle/fluid/operators/fused_softmax_mask_op.h
+32
-0
python/paddle/fluid/tests/unittests/test_softmax_mask_fuse_op.py
...paddle/fluid/tests/unittests/test_softmax_mask_fuse_op.py
+120
-0
python/paddle/incubate/__init__.py
python/paddle/incubate/__init__.py
+2
-1
python/paddle/incubate/operators/__init__.py
python/paddle/incubate/operators/__init__.py
+1
-0
python/paddle/incubate/operators/softmax_mask_fuse.py
python/paddle/incubate/operators/softmax_mask_fuse.py
+33
-0
未找到文件。
paddle/fluid/operators/fused_softmax_mask_op.cc
0 → 100644
浏览文件 @
44bdbe93
/* Copyright (c) 2021 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/fluid/operators/fused_softmax_mask_op.h"
#include "paddle/fluid/framework/generator.h"
#include "paddle/fluid/framework/op_registry.h"
namespace
paddle
{
namespace
operators
{
using
framework
::
Tensor
;
class
SoftmaxMaskFuseOp
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
OP_INOUT_CHECK
(
ctx
->
HasInput
(
"X"
),
"Input"
,
"X"
,
"SoftmaxMaskFuse"
);
OP_INOUT_CHECK
(
ctx
->
HasInput
(
"Mask"
),
"Input"
,
"Mask"
,
"SoftmaxMaskFuse"
);
OP_INOUT_CHECK
(
ctx
->
HasOutput
(
"Out"
),
"Output"
,
"Out"
,
"SoftmaxMaskFuse"
);
auto
x_dims
=
ctx
->
GetInputDim
(
"X"
);
auto
mask_dims
=
ctx
->
GetInputDim
(
"Mask"
);
PADDLE_ENFORCE_EQ
(
x_dims
.
size
(),
4
,
platform
::
errors
::
InvalidArgument
(
"Input x must be in 4D dimension but "
"received the dimension of X is %d"
,
x_dims
.
size
()));
PADDLE_ENFORCE_EQ
(
mask_dims
.
size
(),
4
,
platform
::
errors
::
InvalidArgument
(
"Input mask must be in 4D dimension but "
"received the dimension of mask is %d"
,
mask_dims
.
size
()));
ctx
->
SetOutputDim
(
"Out"
,
x_dims
);
ctx
->
ShareLoD
(
"X"
,
"Out"
);
}
};
class
SoftmaxMaskFuseOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
void
Make
()
override
{
AddInput
(
"X"
,
"The input of softmax_mask_fuse op, "
"which is the result of matmul(QK)/sqrt(dk)."
);
AddInput
(
"Mask"
,
"The mask attr of the op, multi-head attention's mask"
);
AddOutput
(
"Out"
,
"The result of softmax_mask_fuse op."
);
AddComment
(
R"DOC(
Softmax Mask Fuse Operator.
In general, the compute pass is:
product = matmul(QK)/sqrt(dk)
pre_softmax = product + attn_mask
output = softmax(pre_softmax)
To reduce the launch op time and reduce the number of forward and backward,
and to reduce the memory cost for the pre_softmax var during the compute
this op fuse last two operations into one, so users can simply call
product = matmul(QK)/sqrt(dk)
output = softmax_mask_fuse(product, attn_mask)
to get the final output.
By doing this fusion, we can optimize the training by
1. saving one launch cost, one forward and one backward cost
2. saving the memory cost used to save the tmp var
)DOC"
);
}
};
class
SoftmaxMaskFuseOpGrad
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
OP_INOUT_CHECK
(
ctx
->
HasInput
(
framework
::
GradVarName
(
"Out"
)),
"Input"
,
framework
::
GradVarName
(
"Out"
),
"SoftmaxMaskFuseGrad"
);
auto
out_dims
=
ctx
->
GetInputDim
(
framework
::
GradVarName
(
"Out"
));
ctx
->
SetOutputDim
(
framework
::
GradVarName
(
"X"
),
out_dims
);
ctx
->
ShareLoD
(
framework
::
GradVarName
(
"Out"
),
framework
::
GradVarName
(
"X"
));
}
};
template
<
typename
T
>
class
SoftmaxMaskFuseGradOpMaker
:
public
framework
::
SingleGradOpMaker
<
T
>
{
public:
using
framework
::
SingleGradOpMaker
<
T
>::
SingleGradOpMaker
;
protected:
void
Apply
(
GradOpPtr
<
T
>
op
)
const
override
{
op
->
SetType
(
"fused_softmax_mask_grad"
);
op
->
SetInput
(
"Softmax"
,
this
->
Output
(
"Out"
));
op
->
SetInput
(
framework
::
GradVarName
(
"Out"
),
this
->
OutputGrad
(
"Out"
));
op
->
SetOutput
(
framework
::
GradVarName
(
"X"
),
this
->
InputGrad
(
"X"
));
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OPERATOR
(
fused_softmax_mask
,
ops
::
SoftmaxMaskFuseOp
,
ops
::
SoftmaxMaskFuseOpMaker
,
ops
::
SoftmaxMaskFuseGradOpMaker
<
paddle
::
framework
::
OpDesc
>
,
ops
::
SoftmaxMaskFuseGradOpMaker
<
paddle
::
imperative
::
OpBase
>
);
REGISTER_OPERATOR
(
fused_softmax_mask_grad
,
ops
::
SoftmaxMaskFuseOpGrad
);
REGISTER_OP_CPU_KERNEL
(
fused_softmax_mask
,
ops
::
SoftmaxMaskFuseCPUKernel
<
paddle
::
platform
::
CPUDeviceContext
,
float
>
,
ops
::
SoftmaxMaskFuseCPUKernel
<
paddle
::
platform
::
CPUDeviceContext
,
double
>
);
paddle/fluid/operators/fused_softmax_mask_op.cu
0 → 100644
浏览文件 @
44bdbe93
/* Copyright (c) 2021 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. */
// this file is inspired by:
// https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/fused_kernels/scaled_masked_softmax.h
#ifdef PADDLE_WITH_CUDA
#include <cuda.h>
#include <curand_kernel.h>
#endif
#ifdef PADDLE_WITH_HIP
#include <hip/hip_runtime.h>
#include <hiprand_kernel.h>
#endif
#include <stdint.h>
#include <thrust/device_ptr.h>
#include <thrust/iterator/counting_iterator.h>
#include <thrust/transform.h>
#include <algorithm>
#include <string>
#include "paddle/fluid/framework/generator.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/memory/memcpy.h"
#include "paddle/fluid/operators/fused_softmax_mask_op.h"
#include "paddle/fluid/platform/float16.h"
namespace
paddle
{
namespace
operators
{
using
framework
::
Tensor
;
#ifdef PADDLE_WITH_HIP
#define WARP_SIZE 64
#else
#define WARP_SIZE 32
#endif
#define MASK 0xffffffff
namespace
plat
=
paddle
::
platform
;
__device__
__inline__
void
load_data
(
plat
::
float16
*
dst
,
const
plat
::
float16
*
src
)
{
*
(
reinterpret_cast
<
float2
*>
(
dst
))
=
*
(
reinterpret_cast
<
const
float2
*>
(
src
));
}
__device__
__inline__
void
load_data
(
float
*
dst
,
const
float
*
src
)
{
*
(
reinterpret_cast
<
float4
*>
(
dst
))
=
*
(
reinterpret_cast
<
const
float4
*>
(
src
));
}
int
get_pow2
(
int
value
)
{
// get next pow2 index
int
pow2_index
=
0
;
while
((
1
<<
pow2_index
)
<
value
)
{
++
pow2_index
;
}
return
pow2_index
;
}
template
<
typename
T
>
struct
AddOP
{
__device__
__forceinline__
T
operator
()(
T
a
,
T
b
)
const
{
return
a
+
b
;
}
};
template
<
typename
T
>
struct
MaxOP
{
__device__
__forceinline__
T
operator
()(
T
a
,
T
b
)
const
{
return
a
<
b
?
b
:
a
;
}
};
template
<
typename
T
>
__device__
__forceinline__
T
warp_shfl_xor
(
T
value
,
int
laneMask
,
int
width
,
unsigned
int
mask
=
MASK
)
{
#if CUDA_VERSION >= 9000
return
__shfl_xor_sync
(
mask
,
value
,
laneMask
,
width
);
#else
return
__shfl_xor
(
value
,
laneMask
,
width
);
#endif
}
template
<
typename
T
,
int
batch
,
int
width
,
template
<
typename
>
class
ReduceOp
>
__device__
__forceinline__
void
warp_reduce
(
T
*
sum
)
{
ReduceOp
<
T
>
r
;
#pragma unroll
for
(
int
offset
=
width
/
2
;
offset
>
0
;
offset
/=
2
)
{
#pragma unroll
for
(
int
i
=
0
;
i
<
batch
;
++
i
)
{
T
b
=
warp_shfl_xor
(
sum
[
i
],
offset
,
width
);
sum
[
i
]
=
r
(
sum
[
i
],
b
);
}
}
}
// T == fp16
template
<
typename
T
,
int
pow2_index
>
__global__
void
SoftmaxMaskFuseGPUKernel
(
const
T
*
x_data
,
const
T
*
mask_data
,
T
*
y_data
,
int
batch_count
,
int
key_seq_len
)
{
// the forward gpu kernel
constexpr
int
next_pow2
=
1
<<
pow2_index
;
constexpr
int
warp_size
=
(
next_pow2
<
WARP_SIZE
)
?
next_pow2
:
WARP_SIZE
;
constexpr
int
kLocalIterations
=
std
::
max
(
next_pow2
/
warp_size
,
4
);
constexpr
int
kLocalBatchSize
=
(
next_pow2
<=
128
)
?
2
:
1
;
constexpr
int
kOneLoadingCounts
=
4
;
int
data_first_idx
=
(
blockDim
.
y
*
(
blockIdx
.
x
+
gridDim
.
x
*
(
blockIdx
.
y
+
gridDim
.
y
*
blockIdx
.
z
))
+
threadIdx
.
y
)
*
kLocalBatchSize
;
int
mask_fist_idx
=
(
blockDim
.
y
*
(
blockIdx
.
x
+
gridDim
.
x
*
blockIdx
.
z
)
+
threadIdx
.
y
)
*
kLocalBatchSize
;
// batch_count might not be a multiple of kLocalBatchSize. Check how
// many batches have to computed within this WARP.
int
local_batches
=
batch_count
-
data_first_idx
;
if
(
local_batches
>
kLocalBatchSize
)
local_batches
=
kLocalBatchSize
;
// might be many batches per warp. compute the index within the batch
int
local_idx
=
threadIdx
.
x
;
int
x_offset
=
data_first_idx
*
key_seq_len
+
kOneLoadingCounts
*
local_idx
;
int
mask_offset
=
mask_fist_idx
*
key_seq_len
+
kOneLoadingCounts
*
local_idx
;
x_data
+=
x_offset
;
mask_data
+=
mask_offset
;
y_data
+=
x_offset
;
// using float for all inter compute
float
data
[
kLocalBatchSize
][
kLocalIterations
];
T
temp_data
[
kOneLoadingCounts
];
T
temp_mask
[
kOneLoadingCounts
];
#pragma unroll
for
(
int
i
=
0
;
i
<
kLocalBatchSize
;
++
i
)
{
int
batch_data
=
(
i
>=
local_batches
)
?
0
:
key_seq_len
;
#pragma unroll
for
(
int
ii
=
0
;
ii
<
kLocalIterations
;
ii
+=
kOneLoadingCounts
)
{
int
data_index
=
kOneLoadingCounts
*
local_idx
+
ii
*
warp_size
;
if
(
data_index
<
batch_data
)
{
int
itr_idx
=
i
*
key_seq_len
+
ii
*
warp_size
;
// efficiently load data from global memory
load_data
(
temp_data
,
x_data
+
itr_idx
);
load_data
(
temp_mask
,
mask_data
+
itr_idx
);
#pragma unroll
for
(
int
counter
=
0
;
counter
<
kOneLoadingCounts
;
++
counter
)
{
data
[
i
][
ii
+
counter
]
=
static_cast
<
float
>
(
temp_data
[
counter
])
+
static_cast
<
float
>
(
temp_mask
[
counter
]);
}
}
else
{
#pragma unroll
for
(
int
counter
=
0
;
counter
<
kOneLoadingCounts
;
++
counter
)
{
data
[
i
][
ii
+
counter
]
=
-
std
::
numeric_limits
<
float
>::
infinity
();
}
}
}
}
// compute max_value
// max value for each batch for current warp
float
samples_max_value
[
kLocalBatchSize
];
#pragma unroll
for
(
int
i
=
0
;
i
<
kLocalBatchSize
;
++
i
)
{
samples_max_value
[
i
]
=
data
[
i
][
0
];
#pragma unroll
for
(
int
ii
=
1
;
ii
<
kLocalIterations
;
++
ii
)
{
samples_max_value
[
i
]
=
(
samples_max_value
[
i
]
>
data
[
i
][
ii
])
?
samples_max_value
[
i
]
:
data
[
i
][
ii
];
}
}
// max value for each batch for all warp
warp_reduce
<
float
,
kLocalBatchSize
,
warp_size
,
MaxOP
>
(
samples_max_value
);
// compute the sum for each batch for current warp
float
samples_sum
[
kLocalBatchSize
]{
0.0
f
};
#pragma unroll
for
(
int
i
=
0
;
i
<
kLocalBatchSize
;
++
i
)
{
#pragma unroll
for
(
int
ii
=
0
;
ii
<
kLocalIterations
;
++
ii
)
{
data
[
i
][
ii
]
=
std
::
exp
((
data
[
i
][
ii
]
-
samples_max_value
[
i
]));
samples_sum
[
i
]
+=
data
[
i
][
ii
];
}
}
// samples_sum for each batch for all warp
warp_reduce
<
float
,
kLocalBatchSize
,
warp_size
,
AddOP
>
(
samples_sum
);
// load the result from device back to host
T
samples_out
[
kOneLoadingCounts
];
#pragma unroll
for
(
int
i
=
0
;
i
<
kLocalBatchSize
;
++
i
)
{
if
(
i
>=
local_batches
)
break
;
#pragma unroll
for
(
int
ii
=
0
;
ii
<
kLocalIterations
;
ii
+=
kOneLoadingCounts
)
{
int
idx
=
kOneLoadingCounts
*
local_idx
+
ii
*
warp_size
;
if
(
idx
<
key_seq_len
)
{
#pragma unroll
for
(
int
counter
=
0
;
counter
<
kOneLoadingCounts
;
++
counter
)
{
samples_out
[
counter
]
=
data
[
i
][
ii
+
counter
]
/
samples_sum
[
i
];
}
load_data
(
y_data
+
i
*
key_seq_len
+
ii
*
warp_size
,
samples_out
);
}
else
{
break
;
}
}
}
}
template
<
typename
T
,
int
pow2_index
>
__global__
void
SoftmaxMaskFuseGradGPUKernel
(
const
T
*
grad_input
,
T
*
grad_output
,
const
T
*
softmax_rst
,
int
batch_count
,
int
key_seq_len
)
{
constexpr
int
next_pow2
=
1
<<
pow2_index
;
constexpr
int
warp_size
=
(
next_pow2
<
WARP_SIZE
)
?
next_pow2
:
WARP_SIZE
;
constexpr
int
kLocalIterations
=
std
::
max
(
next_pow2
/
warp_size
,
4
);
constexpr
int
kLocalBatchSize
=
(
next_pow2
<=
128
)
?
2
:
1
;
constexpr
int
kOneLoadingCounts
=
4
;
int
data_first_idx
=
(
blockDim
.
y
*
blockIdx
.
x
+
threadIdx
.
y
)
*
kLocalBatchSize
;
// batch_count might not be a multiple of kLocalBatchSize. Check how
// many batches have to computed within this WARP.
int
local_batches
=
batch_count
-
data_first_idx
;
if
(
local_batches
>
kLocalBatchSize
)
local_batches
=
kLocalBatchSize
;
// might be many batches per warp. compute the index within the batch
int
local_idx
=
threadIdx
.
x
;
// the first element to process by the current thread
int
offset
=
data_first_idx
*
key_seq_len
+
kOneLoadingCounts
*
local_idx
;
grad_input
+=
offset
;
grad_output
+=
offset
;
softmax_rst
+=
offset
;
// using float for all inter compute
float
grad_input_reg
[
kLocalBatchSize
][
kLocalIterations
]{
0.0
f
};
float
softmax_rst_reg
[
kLocalBatchSize
][
kLocalIterations
]{
0.0
f
};
T
temp_grad_input
[
kOneLoadingCounts
];
T
temp_softmax_rst
[
kOneLoadingCounts
];
#pragma unroll
for
(
int
i
=
0
;
i
<
kLocalBatchSize
;
++
i
)
{
int
batch_data
=
(
i
>=
local_batches
)
?
0
:
key_seq_len
;
#pragma unroll
for
(
int
ii
=
0
;
ii
<
kLocalIterations
;
ii
+=
kOneLoadingCounts
)
{
int
data_index
=
kOneLoadingCounts
*
local_idx
+
ii
*
WARP_SIZE
;
if
(
data_index
<
batch_data
)
{
load_data
(
temp_grad_input
,
grad_input
+
i
*
key_seq_len
+
ii
*
warp_size
);
load_data
(
temp_softmax_rst
,
softmax_rst
+
i
*
key_seq_len
+
ii
*
warp_size
);
#pragma unroll
for
(
int
counter
=
0
;
counter
<
kOneLoadingCounts
;
++
counter
)
{
softmax_rst_reg
[
i
][
ii
+
counter
]
=
static_cast
<
float
>
(
temp_softmax_rst
[
counter
]);
}
#pragma unroll
for
(
int
counter
=
0
;
counter
<
kOneLoadingCounts
;
++
counter
)
{
grad_input_reg
[
i
][
ii
+
counter
]
=
static_cast
<
float
>
(
temp_grad_input
[
counter
])
*
softmax_rst_reg
[
i
][
ii
+
counter
];
}
}
}
}
float
samples_sum
[
kLocalBatchSize
];
#pragma unroll
for
(
int
i
=
0
;
i
<
kLocalBatchSize
;
++
i
)
{
samples_sum
[
i
]
=
grad_input_reg
[
i
][
0
];
#pragma unroll
for
(
int
ii
=
1
;
ii
<
kLocalIterations
;
++
ii
)
{
samples_sum
[
i
]
+=
grad_input_reg
[
i
][
ii
];
}
}
warp_reduce
<
float
,
kLocalBatchSize
,
warp_size
,
AddOP
>
(
samples_sum
);
#pragma unroll
for
(
int
i
=
0
;
i
<
kLocalBatchSize
;
++
i
)
{
if
(
i
>=
local_batches
)
break
;
#pragma unroll
for
(
int
ii
=
0
;
ii
<
kLocalIterations
;
ii
+=
kOneLoadingCounts
)
{
int
data_index
=
kOneLoadingCounts
*
local_idx
+
ii
*
warp_size
;
if
(
data_index
<
key_seq_len
)
{
// compute gradients
T
samples_out
[
kOneLoadingCounts
];
#pragma unroll
for
(
int
counter
=
0
;
counter
<
kOneLoadingCounts
;
++
counter
)
{
samples_out
[
counter
]
=
grad_input_reg
[
i
][
ii
+
counter
]
-
softmax_rst_reg
[
i
][
ii
+
counter
]
*
samples_sum
[
i
];
}
load_data
(
grad_output
+
i
*
key_seq_len
+
ii
*
warp_size
,
samples_out
);
}
}
}
}
// T only supports fp16
// leave as template only for future update
template
<
typename
Place
,
typename
T
>
class
SoftmaxMaskFuseKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
auto
*
x
=
context
.
Input
<
Tensor
>
(
"X"
);
auto
*
mask
=
context
.
Input
<
Tensor
>
(
"Mask"
);
auto
*
y
=
context
.
Output
<
Tensor
>
(
"Out"
);
auto
*
x_data
=
x
->
data
<
T
>
();
auto
*
mask_data
=
mask
->
data
<
T
>
();
auto
*
y_data
=
y
->
mutable_data
<
T
>
(
context
.
GetPlace
());
auto
x_dim
=
x
->
dims
();
auto
mask_dim
=
mask
->
dims
();
auto
batches
=
x_dim
[
0
];
auto
attn_heads
=
x_dim
[
1
];
auto
query_seq_len
=
x_dim
[
2
];
auto
key_seq_len
=
x_dim
[
3
];
PADDLE_ENFORCE_GT
(
query_seq_len
,
1
,
platform
::
errors
::
InvalidArgument
(
"Input x's second last dim must be large than 1 but "
"received the second last dimension of x is %d"
,
query_seq_len
));
PADDLE_ENFORCE_EQ
(
key_seq_len
>=
32
&&
key_seq_len
<
8192
,
true
,
platform
::
errors
::
InvalidArgument
(
"Input x's last dim must be between [32, 8192) "
"received the last dimension of x is %d"
,
key_seq_len
));
PADDLE_ENFORCE_EQ
(
mask_dim
[
1
],
1
,
platform
::
errors
::
InvalidArgument
(
"Input mask's second dim must be 1 "
"received the second dimension of mask is %d"
,
mask_dim
[
1
]));
// dim of x and mask must be equal
for
(
size_t
idx
=
0
;
idx
<
4
;
++
idx
)
{
if
(
idx
==
1
)
continue
;
PADDLE_ENFORCE_EQ
(
x_dim
[
idx
],
mask_dim
[
idx
],
platform
::
errors
::
InvalidArgument
(
"Input x's %dth dim should be equal with input mask's %dth dim "
"but "
"received the %dth dimension of x and mask are not equal "
"the %dth dim of x is %d, while the %dth dim of mask is %d."
,
idx
,
idx
,
idx
,
idx
,
x_dim
[
idx
],
idx
,
mask_dim
[
idx
]));
}
auto
&
place
=
*
context
.
template
device_context
<
Place
>().
eigen_device
();
auto
stream
=
context
.
cuda_device_context
().
stream
();
int
pow2_index
=
get_pow2
(
key_seq_len
);
const
int
next_pow2
=
1
<<
pow2_index
;
int
batch_count
=
batches
*
attn_heads
*
query_seq_len
;
int
warp_size
=
(
next_pow2
<
WARP_SIZE
)
?
next_pow2
:
WARP_SIZE
;
int
batches_per_warp
=
(
next_pow2
<=
128
)
?
2
:
1
;
// use 128 threads per block to maximum gpu utilization
constexpr
int
threads_per_block
=
128
;
int
warps_per_block
=
(
threads_per_block
/
warp_size
);
int
batches_per_block
=
warps_per_block
*
batches_per_warp
;
PADDLE_ENFORCE_EQ
(
query_seq_len
%
batches_per_block
,
0
,
platform
::
errors
::
InvalidArgument
(
"The query seq len (third dim of input X) must can divide the "
"number of batches per block. The query seq len is %d, while "
"the number of batches per block is %d."
,
query_seq_len
,
batches_per_block
));
dim3
blocks
(
query_seq_len
/
batches_per_block
,
attn_heads
,
batches
);
dim3
threads
(
warp_size
,
warps_per_block
,
1
);
// launch the kernel based on the pow2_index
switch
(
pow2_index
)
{
case
5
:
// 32
SoftmaxMaskFuseGPUKernel
<
T
,
5
><<<
blocks
,
threads
,
0
,
stream
>>>
(
x_data
,
mask_data
,
y_data
,
batch_count
,
key_seq_len
);
break
;
case
6
:
// 64
SoftmaxMaskFuseGPUKernel
<
T
,
6
><<<
blocks
,
threads
,
0
,
stream
>>>
(
x_data
,
mask_data
,
y_data
,
batch_count
,
key_seq_len
);
break
;
case
7
:
// 128
SoftmaxMaskFuseGPUKernel
<
T
,
7
><<<
blocks
,
threads
,
0
,
stream
>>>
(
x_data
,
mask_data
,
y_data
,
batch_count
,
key_seq_len
);
break
;
case
8
:
// 256
SoftmaxMaskFuseGPUKernel
<
T
,
8
><<<
blocks
,
threads
,
0
,
stream
>>>
(
x_data
,
mask_data
,
y_data
,
batch_count
,
key_seq_len
);
break
;
case
9
:
// 512
SoftmaxMaskFuseGPUKernel
<
T
,
9
><<<
blocks
,
threads
,
0
,
stream
>>>
(
x_data
,
mask_data
,
y_data
,
batch_count
,
key_seq_len
);
break
;
case
10
:
// 1024
SoftmaxMaskFuseGPUKernel
<
T
,
10
><<<
blocks
,
threads
,
0
,
stream
>>>
(
x_data
,
mask_data
,
y_data
,
batch_count
,
key_seq_len
);
break
;
case
11
:
// 2048
SoftmaxMaskFuseGPUKernel
<
T
,
11
><<<
blocks
,
threads
,
0
,
stream
>>>
(
x_data
,
mask_data
,
y_data
,
batch_count
,
key_seq_len
);
break
;
case
12
:
// 4096
SoftmaxMaskFuseGPUKernel
<
T
,
12
><<<
blocks
,
threads
,
0
,
stream
>>>
(
x_data
,
mask_data
,
y_data
,
batch_count
,
key_seq_len
);
break
;
case
13
:
// 8192
SoftmaxMaskFuseGPUKernel
<
T
,
13
><<<
blocks
,
threads
,
0
,
stream
>>>
(
x_data
,
mask_data
,
y_data
,
batch_count
,
key_seq_len
);
break
;
default:
break
;
}
}
};
template
<
typename
Place
,
typename
T
>
class
SoftmaxMaskFuseGradKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
auto
*
grad_x
=
context
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"X"
));
auto
*
grad_y
=
context
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Out"
));
auto
*
softmax_rst
=
context
.
Input
<
Tensor
>
(
"Softmax"
);
auto
*
grad_x_data
=
grad_x
->
mutable_data
<
T
>
(
context
.
GetPlace
());
auto
*
grad_y_data
=
grad_y
->
data
<
T
>
();
auto
*
softmax_rst_data
=
softmax_rst
->
data
<
T
>
();
auto
y_dim
=
grad_y
->
dims
();
auto
batches
=
y_dim
[
0
];
auto
attn_heads
=
y_dim
[
1
];
auto
query_seq_len
=
y_dim
[
2
];
auto
key_seq_len
=
y_dim
[
3
];
auto
&
place
=
*
context
.
template
device_context
<
Place
>().
eigen_device
();
auto
stream
=
context
.
cuda_device_context
().
stream
();
int
pow2_index
=
get_pow2
(
key_seq_len
);
const
int
next_pow2
=
1
<<
pow2_index
;
int
batch_count
=
batches
*
attn_heads
*
query_seq_len
;
int
warp_size
=
(
next_pow2
<
WARP_SIZE
)
?
next_pow2
:
WARP_SIZE
;
int
batches_per_warp
=
(
next_pow2
<=
128
)
?
2
:
1
;
// use 128 threads per block to maximum gpu utilization
constexpr
int
threads_per_block
=
128
;
int
warps_per_block
=
(
threads_per_block
/
warp_size
);
int
batches_per_block
=
warps_per_block
*
batches_per_warp
;
int
blocks
=
batch_count
/
batches_per_block
;
dim3
threads
(
warp_size
,
warps_per_block
,
1
);
// launch the kernel based on the pow2_index
switch
(
pow2_index
)
{
case
5
:
// 32
SoftmaxMaskFuseGradGPUKernel
<
T
,
5
><<<
blocks
,
threads
,
0
,
stream
>>>
(
grad_y_data
,
grad_x_data
,
softmax_rst_data
,
batch_count
,
key_seq_len
);
break
;
case
6
:
// 64
SoftmaxMaskFuseGradGPUKernel
<
T
,
6
><<<
blocks
,
threads
,
0
,
stream
>>>
(
grad_y_data
,
grad_x_data
,
softmax_rst_data
,
batch_count
,
key_seq_len
);
break
;
case
7
:
// 128
SoftmaxMaskFuseGradGPUKernel
<
T
,
7
><<<
blocks
,
threads
,
0
,
stream
>>>
(
grad_y_data
,
grad_x_data
,
softmax_rst_data
,
batch_count
,
key_seq_len
);
break
;
case
8
:
// 256
SoftmaxMaskFuseGradGPUKernel
<
T
,
8
><<<
blocks
,
threads
,
0
,
stream
>>>
(
grad_y_data
,
grad_x_data
,
softmax_rst_data
,
batch_count
,
key_seq_len
);
break
;
case
9
:
// 512
SoftmaxMaskFuseGradGPUKernel
<
T
,
9
><<<
blocks
,
threads
,
0
,
stream
>>>
(
grad_y_data
,
grad_x_data
,
softmax_rst_data
,
batch_count
,
key_seq_len
);
break
;
case
10
:
// 1024
SoftmaxMaskFuseGradGPUKernel
<
T
,
10
><<<
blocks
,
threads
,
0
,
stream
>>>
(
grad_y_data
,
grad_x_data
,
softmax_rst_data
,
batch_count
,
key_seq_len
);
break
;
case
11
:
// 2048
SoftmaxMaskFuseGradGPUKernel
<
T
,
11
><<<
blocks
,
threads
,
0
,
stream
>>>
(
grad_y_data
,
grad_x_data
,
softmax_rst_data
,
batch_count
,
key_seq_len
);
break
;
case
12
:
// 4096
SoftmaxMaskFuseGradGPUKernel
<
T
,
12
><<<
blocks
,
threads
,
0
,
stream
>>>
(
grad_y_data
,
grad_x_data
,
softmax_rst_data
,
batch_count
,
key_seq_len
);
break
;
case
13
:
// 8192
SoftmaxMaskFuseGradGPUKernel
<
T
,
13
><<<
blocks
,
threads
,
0
,
stream
>>>
(
grad_y_data
,
grad_x_data
,
softmax_rst_data
,
batch_count
,
key_seq_len
);
break
;
default:
break
;
}
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
namespace
plat
=
paddle
::
platform
;
REGISTER_OP_CUDA_KERNEL
(
fused_softmax_mask
,
ops
::
SoftmaxMaskFuseKernel
<
plat
::
CUDADeviceContext
,
plat
::
float16
>
,
ops
::
SoftmaxMaskFuseKernel
<
plat
::
CUDADeviceContext
,
float
>
);
REGISTER_OP_CUDA_KERNEL
(
fused_softmax_mask_grad
,
ops
::
SoftmaxMaskFuseGradKernel
<
plat
::
CUDADeviceContext
,
plat
::
float16
>
,
ops
::
SoftmaxMaskFuseGradKernel
<
plat
::
CUDADeviceContext
,
float
>
);
paddle/fluid/operators/fused_softmax_mask_op.h
0 → 100644
浏览文件 @
44bdbe93
/* Copyright (c) 2021 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. */
#pragma once
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
namespace
paddle
{
namespace
operators
{
template
<
typename
DeviceContext
,
typename
T
>
class
SoftmaxMaskFuseCPUKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
PADDLE_ENFORCE_EQ
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()),
true
,
platform
::
errors
::
Unimplemented
(
"Softmax mask fuse op only supports GPU now."
));
}
};
}
// namespace operators
}
// namespace paddle
python/paddle/fluid/tests/unittests/test_softmax_mask_fuse_op.py
0 → 100644
浏览文件 @
44bdbe93
# Copyright (c) 2021 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.
from
__future__
import
print_function
import
unittest
import
numpy
as
np
import
paddle.fluid.core
as
core
from
op_test
import
OpTest
import
paddle
import
paddle.fluid
as
fluid
import
paddle.incubate
as
incubate
paddle
.
enable_static
()
def
_get_softmax
(
x
,
mask
,
fp16
=
True
):
masked_x
=
(
x
+
mask
).
astype
(
"float32"
)
max_value
=
np
.
max
(
masked_x
,
axis
=-
1
,
keepdims
=
True
)
before_exp
=
masked_x
-
max_value
exp
=
np
.
exp
(
before_exp
)
exp_sum
=
np
.
sum
(
exp
,
axis
=-
1
,
keepdims
=
True
)
rst
=
exp
/
exp_sum
if
fp16
:
rst
=
rst
.
astype
(
"float16"
)
return
rst
@
unittest
.
skipIf
(
not
core
.
is_compiled_with_cuda
(),
"core is not compiled with CUDA"
)
class
TestSoftmaxMaskFuseOp
(
OpTest
):
def
setUp
(
self
):
self
.
op_type
=
"fused_softmax_mask"
x
=
np
.
random
.
random
((
1
,
1
,
8
,
32
))
mask
=
np
.
random
.
randint
(
0
,
2
,
(
1
,
1
,
8
,
32
))
mask_input
=
np
.
where
(
mask
==
1
,
-
10000.0
,
mask
)
self
.
inputs
=
{
'X'
:
x
,
'Mask'
:
mask_input
}
rst
=
_get_softmax
(
x
,
mask_input
)
self
.
outputs
=
{
'Out'
:
rst
}
def
test_check_output
(
self
):
try
:
self
.
check_output_with_place
(
core
.
CPUPlace
())
except
NotImplementedError
:
pass
def
test_check_grad
(
self
):
try
:
self
.
check_grad_with_place
(
core
.
CPUPlace
(),
[
"X"
],
"Out"
)
except
NotImplementedError
:
pass
@
unittest
.
skipIf
(
not
core
.
is_compiled_with_cuda
(),
"core is not compiled with CUDA"
)
class
TestSoftmaxMaskFuseOp0
(
OpTest
):
def
setUp
(
self
):
self
.
op_type
=
"fused_softmax_mask"
x
=
np
.
random
.
random
((
1
,
1
,
8
,
32
)).
astype
(
"float16"
)
mask
=
np
.
random
.
randint
(
0
,
2
,
(
1
,
1
,
8
,
32
)).
astype
(
"float16"
)
mask_input
=
np
.
where
(
mask
==
1
,
-
10000.0
,
mask
)
self
.
inputs
=
{
'X'
:
x
,
'Mask'
:
mask_input
}
rst
=
_get_softmax
(
x
,
mask_input
)
self
.
outputs
=
{
'Out'
:
rst
}
def
test_check_output
(
self
):
self
.
check_output_with_place
(
core
.
CUDAPlace
(
0
))
def
test_check_grad
(
self
):
self
.
check_grad_with_place
(
core
.
CUDAPlace
(
0
),
[
"X"
],
"Out"
)
@
unittest
.
skipIf
(
not
core
.
is_compiled_with_cuda
(),
"core is not compiled with CUDA"
)
class
TestDropoutBiasFuseOp3
(
unittest
.
TestCase
):
def
test_static_result
(
self
):
with
fluid
.
program_guard
(
fluid
.
Program
(),
fluid
.
Program
()):
input_x
=
fluid
.
data
(
name
=
"x"
,
shape
=
[
1
,
1
,
8
,
32
],
dtype
=
"float32"
)
input_mask
=
fluid
.
data
(
name
=
"mask"
,
shape
=
[
1
,
1
,
8
,
32
],
dtype
=
"float32"
)
rst
=
incubate
.
softmax_mask_fuse
(
input_x
,
input_mask
)
x_in_np
=
np
.
random
.
random
((
1
,
1
,
8
,
32
)).
astype
(
"float32"
)
mask
=
np
.
random
.
randint
(
0
,
2
,
(
1
,
1
,
8
,
32
)).
astype
(
"float32"
)
mask_in_np
=
np
.
where
(
mask
==
1
,
-
10000.0
,
mask
)
rst_np
=
_get_softmax
(
x_in_np
,
mask_in_np
,
False
)
exe
=
fluid
.
Executor
(
fluid
.
CUDAPlace
(
0
))
fetches
=
exe
.
run
(
fluid
.
default_main_program
(),
feed
=
{
"x"
:
x_in_np
,
"mask"
:
mask_in_np
},
fetch_list
=
[
rst
])
self
.
assertTrue
(
np
.
allclose
(
fetches
[
0
],
rst_np
))
def
test_dygraph
(
self
):
with
fluid
.
dygraph
.
guard
(
fluid
.
CUDAPlace
(
0
)):
x_in_np
=
np
.
random
.
random
((
1
,
1
,
8
,
32
)).
astype
(
"float32"
)
mask
=
np
.
random
.
randint
(
0
,
2
,
(
1
,
1
,
8
,
32
)).
astype
(
"float32"
)
mask_in_np
=
np
.
where
(
mask
==
1
,
-
10000.0
,
mask
)
rst_np
=
_get_softmax
(
x_in_np
,
mask_in_np
,
False
)
input_x
=
fluid
.
dygraph
.
to_variable
(
x_in_np
)
input_mask
=
fluid
.
dygraph
.
to_variable
(
mask_in_np
)
rst
=
incubate
.
softmax_mask_fuse
(
input_x
,
input_mask
)
self
.
assertTrue
(
np
.
allclose
(
rst
,
rst_np
))
if
__name__
==
'__main__'
:
unittest
.
main
()
python/paddle/incubate/__init__.py
浏览文件 @
44bdbe93
...
...
@@ -17,7 +17,8 @@ from .optimizer import ModelAverage # noqa: F401
from
.checkpoint
import
auto_checkpoint
# noqa: F401
from
..fluid.layer_helper
import
LayerHelper
# noqa: F401
from
.operators
import
softmax_mask_fuse_upper_triangle
# noqa: F401
from
.operators
import
softmax_mask_fuse
# noqa: F401
__all__
=
[
# noqa
'LookAhead'
,
'ModelAverage'
,
'softmax_mask_fuse_upper_triangle'
'LookAhead'
,
'ModelAverage'
,
'softmax_mask_fuse_upper_triangle'
,
'softmax_mask_fuse'
]
python/paddle/incubate/operators/__init__.py
浏览文件 @
44bdbe93
...
...
@@ -13,3 +13,4 @@
# limitations under the License.
from
.softmax_mask_fuse_upper_triangle
import
softmax_mask_fuse_upper_triangle
# noqa: F401
from
.softmax_mask_fuse
import
softmax_mask_fuse
# noqa: F401
python/paddle/incubate/operators/softmax_mask_fuse.py
0 → 100644
浏览文件 @
44bdbe93
# Copyright (c) 2021 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.
from
__future__
import
print_function
from
paddle.fluid.layer_helper
import
LayerHelper
from
paddle.fluid.framework
import
in_dygraph_mode
from
paddle.fluid
import
core
def
softmax_mask_fuse
(
x
,
mask
,
name
=
None
):
if
in_dygraph_mode
():
out
=
core
.
ops
.
fused_softmax_mask
(
x
,
mask
)
return
out
helper
=
LayerHelper
(
'fused_softmax_mask'
,
**
locals
())
out
=
helper
.
create_variable_for_type_inference
(
dtype
=
x
.
dtype
)
helper
.
append_op
(
type
=
'fused_softmax_mask'
,
inputs
=
{
'X'
:
[
x
],
'Mask'
:
[
mask
]},
outputs
=
{
'Out'
:
[
out
]})
return
out
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