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
5e0614a1
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
5e0614a1
编写于
10月 10, 2022
作者:
R
Rayman
提交者:
GitHub
10月 10, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
【Hackathon No.56&38】deformable_conv_v1 算子实现 float16 数据类型支持&前向运行加速 (#46111)
support fp16 for deformable conv
上级
a7e1b9d2
变更
9
隐藏空白更改
内联
并排
Showing
9 changed file
with
275 addition
and
137 deletion
+275
-137
paddle/phi/kernels/cpu/deformable_conv_grad_kernel.cc
paddle/phi/kernels/cpu/deformable_conv_grad_kernel.cc
+17
-17
paddle/phi/kernels/funcs/deformable_conv_functor.cc
paddle/phi/kernels/funcs/deformable_conv_functor.cc
+3
-3
paddle/phi/kernels/funcs/deformable_conv_functor.cu
paddle/phi/kernels/funcs/deformable_conv_functor.cu
+30
-11
paddle/phi/kernels/funcs/deformable_conv_functor.h
paddle/phi/kernels/funcs/deformable_conv_functor.h
+29
-26
paddle/phi/kernels/gpu/deformable_conv_grad_kernel.cu
paddle/phi/kernels/gpu/deformable_conv_grad_kernel.cu
+47
-43
paddle/phi/kernels/gpu/deformable_conv_kernel.cu
paddle/phi/kernels/gpu/deformable_conv_kernel.cu
+2
-1
paddle/phi/kernels/impl/deformable_conv_grad_kernel_impl.h
paddle/phi/kernels/impl/deformable_conv_grad_kernel_impl.h
+33
-23
paddle/phi/kernels/impl/deformable_conv_kernel_impl.h
paddle/phi/kernels/impl/deformable_conv_kernel_impl.h
+41
-5
python/paddle/fluid/tests/unittests/test_deformable_conv_v1_op.py
...addle/fluid/tests/unittests/test_deformable_conv_v1_op.py
+73
-8
未找到文件。
paddle/phi/kernels/cpu/deformable_conv_grad_kernel.cc
浏览文件 @
5e0614a1
...
...
@@ -97,14 +97,14 @@ inline void ModulatedDeformableCol2imCPUKernel(
width
);
*
(
grad_im
+
cur_bottom_grad_pos
)
=
*
(
grad_im
+
cur_bottom_grad_pos
)
+
weight
*
cur_top_grad
;
*
(
grad_im
+
cur_bottom_grad_pos
)
+
(
weight
*
cur_top_grad
)
;
}
}
}
}
}
template
<
typename
T
,
typename
Context
>
template
<
typename
T
,
typename
MT
,
typename
Context
>
void
ModulatedDeformableCol2im
(
const
Context
&
dev_ctx
,
const
T
*
data_col
,
const
T
*
data_offset
,
...
...
@@ -116,7 +116,7 @@ void ModulatedDeformableCol2im(const Context& dev_ctx,
const
std
::
vector
<
int
>&
stride
,
const
std
::
vector
<
int
>&
dilation
,
const
int
deformable_group
,
T
*
grad_im
)
{
M
T
*
grad_im
)
{
int
channel_per_deformable_group
=
im_shape
[
0
]
/
deformable_group
;
int
num_kernels
=
col_shape
[
0
]
*
col_shape
[
1
]
*
col_shape
[
2
]
*
col_shape
[
3
];
...
...
@@ -222,22 +222,22 @@ void ModulatedDeformableCol2imCoordCPUKernel(
if
(
inv_h
<=
-
1
||
inv_w
<=
-
1
||
inv_h
>=
height
||
inv_w
>=
width
)
{
inv_h
=
inv_w
=
-
2
;
}
else
{
mval
+=
data_col_ptr
[
col_pos
]
*
funcs
::
DmcnIm2colBilinear
(
data_im_ptr
+
cnt
*
height
*
width
,
width
,
height
,
width
,
inv_h
,
inv_w
);
mval
+=
data_col_ptr
[
col_pos
]
*
funcs
::
DmcnIm2colBilinear
<
T
,
T
>
(
data_im_ptr
+
cnt
*
height
*
width
,
width
,
height
,
width
,
inv_h
,
inv_w
);
}
const
T
weight
=
DmcnGetCoordinateWeight
(
inv_h
,
inv_w
,
height
,
width
,
data_im_ptr
+
cnt
*
height
*
width
,
width
,
bp_dir
);
DmcnGetCoordinateWeight
<
T
,
T
>
(
inv_h
,
inv_w
,
height
,
width
,
data_im_ptr
+
cnt
*
height
*
width
,
width
,
bp_dir
);
if
(
data_mask_ptr
)
{
const
int
data_mask_hw_ptr
=
(((
i
*
kernel_w
+
j
)
*
height_col
+
h_out
)
*
width_col
+
w_out
);
...
...
paddle/phi/kernels/funcs/deformable_conv_functor.cc
浏览文件 @
5e0614a1
...
...
@@ -13,8 +13,8 @@
// limitations under the License.
#include "paddle/phi/kernels/funcs/deformable_conv_functor.h"
#include "paddle/phi/backends/cpu/cpu_context.h"
#include "paddle/phi/common/amp_type_traits.h"
namespace
phi
{
namespace
funcs
{
...
...
@@ -82,8 +82,8 @@ inline void ModulatedDeformableIm2colCPUKernel(
const
T
h_im
=
h_in
+
i
*
dilation_h
+
offset_h
;
const
T
w_im
=
w_in
+
j
*
dilation_w
+
offset_w
;
if
(
h_im
>
-
1
&&
w_im
>
-
1
&&
h_im
<
height
&&
w_im
<
width
)
{
val
=
DmcnIm2colBilinear
(
data_im_ptr
,
width
,
height
,
width
,
h_im
,
w_im
);
val
=
DmcnIm2colBilinear
<
T
,
T
>
(
data_im_ptr
,
width
,
height
,
width
,
h_im
,
w_im
);
}
*
data_col_ptr
=
val
;
if
(
data_mask_ptr
)
{
...
...
paddle/phi/kernels/funcs/deformable_conv_functor.cu
浏览文件 @
5e0614a1
...
...
@@ -12,8 +12,11 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/kernels/funcs/deformable_conv_functor.h"
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/common/amp_type_traits.h"
#include "paddle/phi/common/float16.h"
#include "paddle/phi/core/device_context.h"
namespace
phi
{
namespace
funcs
{
...
...
@@ -51,6 +54,8 @@ __global__ void ModulatedDeformableIm2colGpuKernel(
T
*
data_col
)
{
int
index
=
blockIdx
.
x
*
blockDim
.
x
+
threadIdx
.
x
;
int
offset
=
blockDim
.
x
*
gridDim
.
x
;
using
MT
=
typename
phi
::
dtype
::
MPTypeTrait
<
T
>::
Type
;
for
(
size_t
i
=
index
;
i
<
nthreads
;
i
+=
offset
)
{
const
int
w_col
=
i
%
width_col
;
const
int
h_col
=
(
i
/
width_col
)
%
height_col
;
...
...
@@ -85,22 +90,22 @@ __global__ void ModulatedDeformableIm2colGpuKernel(
((
2
*
(
i
*
kernel_w
+
j
)
+
1
)
*
height_col
+
h_col
)
*
width_col
+
w_col
;
const
T
offset_h
=
data_offset_ptr
[
data_offset_h_ptr
]
;
const
T
offset_w
=
data_offset_ptr
[
data_offset_w_ptr
]
;
T
val
=
static_cast
<
T
>
(
0
);
const
T
h_im
=
h_in
+
i
*
dilation_h
+
offset_h
;
const
T
w_im
=
w_in
+
j
*
dilation_w
+
offset_w
;
const
MT
offset_h
=
static_cast
<
MT
>
(
data_offset_ptr
[
data_offset_h_ptr
])
;
const
MT
offset_w
=
static_cast
<
MT
>
(
data_offset_ptr
[
data_offset_w_ptr
])
;
MT
val
=
static_cast
<
M
T
>
(
0
);
const
M
T
h_im
=
h_in
+
i
*
dilation_h
+
offset_h
;
const
M
T
w_im
=
w_in
+
j
*
dilation_w
+
offset_w
;
if
(
h_im
>
-
1
&&
w_im
>
-
1
&&
h_im
<
height
&&
w_im
<
width
)
{
val
=
DmcnIm2colBilinear
(
data_im_ptr
,
width
,
height
,
width
,
h_im
,
w_im
);
val
=
DmcnIm2colBilinear
<
T
,
MT
>
(
data_im_ptr
,
width
,
height
,
width
,
h_im
,
w_im
);
}
*
data_col_ptr
=
val
;
if
(
data_mask_ptr
)
{
const
int
data_mask_hw_ptr
=
((
i
*
kernel_w
+
j
)
*
height_col
+
h_col
)
*
width_col
+
w_col
;
const
T
mask
=
data_mask_ptr
[
data_mask_hw_ptr
]
;
*
data_col_ptr
*=
mask
;
const
MT
mask
=
static_cast
<
MT
>
(
data_mask_ptr
[
data_mask_hw_ptr
])
;
val
*=
mask
;
}
*
data_col_ptr
=
static_cast
<
T
>
(
val
);
data_col_ptr
+=
batch_size
*
height_col
*
width_col
;
}
}
...
...
@@ -164,6 +169,20 @@ template void ModulatedDeformableIm2col(
const
int
deformable_groups
,
float
*
data_col
);
template
void
ModulatedDeformableIm2col
(
const
phi
::
GPUContext
&
dev_ctx
,
const
phi
::
dtype
::
float16
*
data_im
,
const
phi
::
dtype
::
float16
*
data_offset
,
const
phi
::
dtype
::
float16
*
data_mask
,
const
std
::
vector
<
int64_t
>
&
im_shape
,
const
std
::
vector
<
int64_t
>&
col_shape
,
const
std
::
vector
<
int64_t
>&
filter_shape
,
const
std
::
vector
<
int
>&
paddings
,
const
std
::
vector
<
int
>&
strides
,
const
std
::
vector
<
int
>&
dilations
,
const
int
deformable_groups
,
phi
::
dtype
::
float16
*
data_col
);
template
void
ModulatedDeformableIm2col
(
const
phi
::
GPUContext
&
dev_ctx
,
const
double
*
data_im
,
...
...
paddle/phi/kernels/funcs/deformable_conv_functor.h
浏览文件 @
5e0614a1
...
...
@@ -14,44 +14,47 @@
#pragma once
#include "paddle/phi/common/amp_type_traits.h"
#include "paddle/phi/common/float16.h"
#include "paddle/phi/core/dense_tensor.h"
namespace
phi
{
namespace
funcs
{
template
<
typename
T
>
HOSTDEVICE
T
DmcnIm2colBilinear
(
const
T
*
bottom_data
,
const
int
data_width
,
const
int
height
,
const
int
width
,
T
h
,
T
w
)
{
template
<
typename
T
,
typename
MT
>
HOSTDEVICE
M
T
DmcnIm2colBilinear
(
const
T
*
bottom_data
,
const
int
data_width
,
const
int
height
,
const
int
width
,
M
T
h
,
M
T
w
)
{
int
h_low
=
floor
(
h
);
int
w_low
=
floor
(
w
);
int
h_high
=
h_low
+
1
;
int
w_high
=
w_low
+
1
;
T
lh
=
h
-
h_low
;
T
lw
=
w
-
w_low
;
T
hh
=
1
-
lh
;
T
hw
=
1
-
lw
;
M
T
lh
=
h
-
h_low
;
M
T
lw
=
w
-
w_low
;
M
T
hh
=
1
-
lh
;
M
T
hw
=
1
-
lw
;
T
v1
=
(
h_low
>=
0
&&
w_low
>=
0
)
?
bottom_data
[
h_low
*
data_width
+
w_low
]
:
0
;
T
v2
=
(
h_low
>=
0
&&
w_high
<=
width
-
1
)
?
bottom_data
[
h_low
*
data_width
+
w_high
]
:
0
;
T
v3
=
(
h_high
<=
height
-
1
&&
w_low
>=
0
)
?
bottom_data
[
h_high
*
data_width
+
w_low
]
:
0
;
T
v4
=
(
h_high
<=
height
-
1
&&
w_high
<=
width
-
1
)
?
bottom_data
[
h_high
*
data_width
+
w_high
]
:
0
;
MT
v1
=
(
h_low
>=
0
&&
w_low
>=
0
)
?
static_cast
<
MT
>
(
bottom_data
[
h_low
*
data_width
+
w_low
])
:
0
;
MT
v2
=
(
h_low
>=
0
&&
w_high
<=
width
-
1
)
?
static_cast
<
MT
>
(
bottom_data
[
h_low
*
data_width
+
w_high
])
:
0
;
MT
v3
=
(
h_high
<=
height
-
1
&&
w_low
>=
0
)
?
static_cast
<
MT
>
(
bottom_data
[
h_high
*
data_width
+
w_low
])
:
0
;
MT
v4
=
(
h_high
<=
height
-
1
&&
w_high
<=
width
-
1
)
?
static_cast
<
MT
>
(
bottom_data
[
h_high
*
data_width
+
w_high
])
:
0
;
T
w1
=
hh
*
hw
;
T
w2
=
hh
*
lw
;
T
w3
=
lh
*
hw
;
T
w4
=
lh
*
lw
;
M
T
w1
=
hh
*
hw
;
M
T
w2
=
hh
*
lw
;
M
T
w3
=
lh
*
hw
;
M
T
w4
=
lh
*
lw
;
return
w1
*
v1
+
w2
*
v2
+
w3
*
v3
+
w4
*
v4
;
}
...
...
paddle/phi/kernels/gpu/deformable_conv_grad_kernel.cu
浏览文件 @
5e0614a1
...
...
@@ -29,7 +29,7 @@ static inline int NumBlocks(const int N) {
kNumMaximumNumBlocks
);
}
template
<
typename
T
>
template
<
typename
T
,
typename
MT
>
__global__
void
ModulatedDeformableCol2imGpuKernel
(
const
int
nthreads
,
const
T
*
data_col
,
...
...
@@ -51,9 +51,10 @@ __global__ void ModulatedDeformableCol2imGpuKernel(
const
int
deformable_group
,
const
int
height_col
,
const
int
width_col
,
T
*
grad_im
)
{
M
T
*
grad_im
)
{
int
index
=
blockIdx
.
x
*
blockDim
.
x
+
threadIdx
.
x
;
int
offset
=
blockDim
.
x
*
gridDim
.
x
;
// using MT = typename phi::dtype::MPTypeTrait<T>::Type;
for
(
size_t
thread
=
index
;
thread
<
nthreads
;
thread
+=
offset
)
{
const
int
j
=
(
thread
/
width_col
/
height_col
/
batch_size
)
%
kernel_w
;
const
int
i
=
...
...
@@ -78,17 +79,17 @@ __global__ void ModulatedDeformableCol2imGpuKernel(
((
2
*
(
i
*
kernel_w
+
j
)
+
1
)
*
height_col
+
h_out
)
*
width_col
+
w_out
;
const
int
data_mask_hw_ptr
=
((
i
*
kernel_w
+
j
)
*
height_col
+
h_out
)
*
width_col
+
w_out
;
const
T
offset_h
=
data_offset_ptr
[
data_offset_h_ptr
]
;
const
T
offset_w
=
data_offset_ptr
[
data_offset_w_ptr
]
;
const
T
cur_inv_h_data
=
h_in
+
i
*
dilation_h
+
offset_h
;
const
T
cur_inv_w_data
=
w_in
+
j
*
dilation_w
+
offset_w
;
const
MT
offset_h
=
static_cast
<
MT
>
(
data_offset_ptr
[
data_offset_h_ptr
])
;
const
MT
offset_w
=
static_cast
<
MT
>
(
data_offset_ptr
[
data_offset_w_ptr
])
;
const
M
T
cur_inv_h_data
=
h_in
+
i
*
dilation_h
+
offset_h
;
const
M
T
cur_inv_w_data
=
w_in
+
j
*
dilation_w
+
offset_w
;
T
cur_top_grad
=
data_col
[
thread
]
;
MT
cur_top_grad
=
static_cast
<
MT
>
(
data_col
[
thread
])
;
if
(
data_mask
)
{
const
T
*
data_mask_ptr
=
data_mask
+
(
b
*
deformable_group
+
deformable_group_index
)
*
kernel_h
*
kernel_w
*
height_col
*
width_col
;
const
T
mask
=
data_mask_ptr
[
data_mask_hw_ptr
]
;
const
MT
mask
=
static_cast
<
MT
>
(
data_mask_ptr
[
data_mask_hw_ptr
])
;
cur_top_grad
*=
mask
;
}
const
int
cur_h
=
static_cast
<
int
>
(
cur_inv_h_data
);
...
...
@@ -100,13 +101,12 @@ __global__ void ModulatedDeformableCol2imGpuKernel(
abs
(
cur_inv_w_data
-
(
cur_w
+
dx
))
<
1
)
{
int
cur_bottom_grad_pos
=
((
b
*
channels
+
c
)
*
height
+
cur_h
+
dy
)
*
width
+
cur_w
+
dx
;
T
weight
=
DmcnGetGradientWeight
(
cur_inv_h_data
,
cur_inv_w_data
,
cur_h
+
dy
,
cur_w
+
dx
,
height
,
width
);
MT
weight
=
DmcnGetGradientWeight
(
cur_inv_h_data
,
cur_inv_w_data
,
cur_h
+
dy
,
cur_w
+
dx
,
height
,
width
);
paddle
::
platform
::
CudaAtomicAdd
(
grad_im
+
cur_bottom_grad_pos
,
weight
*
cur_top_grad
);
}
...
...
@@ -115,7 +115,7 @@ __global__ void ModulatedDeformableCol2imGpuKernel(
}
}
template
<
typename
T
,
typename
Context
>
template
<
typename
T
,
typename
MT
,
typename
Context
>
void
ModulatedDeformableCol2im
(
const
Context
&
dev_ctx
,
const
T
*
data_col
,
const
T
*
data_offset
,
...
...
@@ -127,13 +127,13 @@ void ModulatedDeformableCol2im(const Context& dev_ctx,
const
std
::
vector
<
int
>&
stride
,
const
std
::
vector
<
int
>&
dilation
,
const
int
deformable_group
,
T
*
grad_im
)
{
M
T
*
grad_im
)
{
int
channel_per_deformable_group
=
im_shape
[
0
]
/
deformable_group
;
int
num_kernels
=
col_shape
[
0
]
*
col_shape
[
1
]
*
col_shape
[
2
]
*
col_shape
[
3
];
int
blocks
=
NumBlocks
(
num_kernels
);
int
threads
=
kNumCUDAThreads
;
ModulatedDeformableCol2imGpuKernel
<
T
>
ModulatedDeformableCol2imGpuKernel
<
T
,
MT
>
<<<
blocks
,
threads
,
0
,
dev_ctx
.
stream
()
>>>
(
num_kernels
,
data_col
,
data_offset
,
...
...
@@ -185,8 +185,9 @@ __global__ void ModulatedDeformableCol2imCoordGpuKernel(
T
*
grad_mask
)
{
int
index
=
blockIdx
.
x
*
blockDim
.
x
+
threadIdx
.
x
;
int
offset
=
blockDim
.
x
*
gridDim
.
x
;
using
MT
=
typename
phi
::
dtype
::
MPTypeTrait
<
T
>::
Type
;
for
(
size_t
i
=
index
;
i
<
nthreads
;
i
+=
offset
)
{
T
val
=
0
,
mval
=
0
;
M
T
val
=
0
,
mval
=
0
;
const
int
w
=
i
%
width_col
;
const
int
h
=
(
i
/
width_col
)
%
height_col
;
const
int
c
=
(
i
/
width_col
/
height_col
)
%
offset_channels
;
...
...
@@ -231,40 +232,42 @@ __global__ void ModulatedDeformableCol2imCoordGpuKernel(
const
int
data_offset_w_ptr
=
(((
2
*
(
i
*
kernel_w
+
j
)
+
1
)
*
height_col
+
h_out
)
*
width_col
+
w_out
);
const
T
offset_h
=
data_offset_ptr
[
data_offset_h_ptr
];
const
T
offset_w
=
data_offset_ptr
[
data_offset_w_ptr
];
T
inv_h
=
h_in
+
i
*
dilation_h
+
offset_h
;
T
inv_w
=
w_in
+
j
*
dilation_w
+
offset_w
;
const
MT
offset_h
=
static_cast
<
MT
>
(
data_offset_ptr
[
data_offset_h_ptr
]);
const
MT
offset_w
=
static_cast
<
MT
>
(
data_offset_ptr
[
data_offset_w_ptr
]);
MT
inv_h
=
h_in
+
i
*
dilation_h
+
offset_h
;
MT
inv_w
=
w_in
+
j
*
dilation_w
+
offset_w
;
if
(
inv_h
<=
-
1
||
inv_w
<=
-
1
||
inv_h
>=
height
||
inv_w
>=
width
)
{
inv_h
=
inv_w
=
-
2
;
}
else
{
mval
+=
data_col_ptr
[
col_pos
]
*
funcs
::
DmcnIm2colBilinear
(
data_im_ptr
+
cnt
*
height
*
width
,
width
,
height
,
width
,
inv_h
,
inv_w
);
mval
+=
static_cast
<
MT
>
(
data_col_ptr
[
col_pos
])
*
funcs
::
DmcnIm2colBilinear
<
T
,
MT
>
(
data_im_ptr
+
cnt
*
height
*
width
,
width
,
height
,
width
,
inv_h
,
inv_w
);
}
const
T
weight
=
DmcnGetCoordinateWeight
(
inv_h
,
inv_w
,
height
,
width
,
data_im_ptr
+
cnt
*
height
*
width
,
width
,
bp_dir
);
const
M
T
weight
=
DmcnGetCoordinateWeight
<
T
,
MT
>
(
inv_h
,
inv_w
,
height
,
width
,
data_im_ptr
+
cnt
*
height
*
width
,
width
,
bp_dir
);
if
(
data_mask_ptr
)
{
const
int
data_mask_hw_ptr
=
(((
i
*
kernel_w
+
j
)
*
height_col
+
h_out
)
*
width_col
+
w_out
);
const
T
mask
=
data_mask_ptr
[
data_mask_hw_ptr
]
;
val
+=
weight
*
data_col_ptr
[
col_pos
]
*
mask
;
const
MT
mask
=
static_cast
<
MT
>
(
data_mask_ptr
[
data_mask_hw_ptr
])
;
val
+=
weight
*
static_cast
<
MT
>
(
data_col_ptr
[
col_pos
])
*
mask
;
}
else
{
val
+=
weight
*
data_col_ptr
[
col_pos
]
;
val
+=
weight
*
static_cast
<
MT
>
(
data_col_ptr
[
col_pos
])
;
}
cnt
+=
1
;
}
grad_offset
[
i
]
=
val
;
grad_offset
[
i
]
=
static_cast
<
T
>
(
val
)
;
if
(
grad_mask
&&
offset_c
%
2
==
0
)
grad_mask
[(((
b
*
deformable_group
+
deformable_group_index
)
*
kernel_h
*
kernel_w
+
...
...
@@ -359,4 +362,5 @@ PD_REGISTER_KERNEL(deformable_conv_grad,
ALL_LAYOUT
,
phi
::
DeformableConvGradKernel
,
float
,
double
)
{}
double
,
paddle
::
platform
::
float16
)
{}
paddle/phi/kernels/gpu/deformable_conv_kernel.cu
浏览文件 @
5e0614a1
...
...
@@ -23,4 +23,5 @@ PD_REGISTER_KERNEL(deformable_conv,
ALL_LAYOUT
,
phi
::
DeformableConvKernel
,
float
,
double
)
{}
double
,
phi
::
dtype
::
float16
)
{}
paddle/phi/kernels/impl/deformable_conv_grad_kernel_impl.h
浏览文件 @
5e0614a1
...
...
@@ -14,8 +14,10 @@
#pragma once
#include "paddle/phi/common/amp_type_traits.h"
#include "paddle/phi/core/dense_tensor.h"
#include "paddle/phi/core/hostdevice.h"
#include "paddle/phi/kernels/cast_kernel.h"
#include "paddle/phi/kernels/empty_kernel.h"
#include "paddle/phi/kernels/full_kernel.h"
#include "paddle/phi/kernels/funcs/blas/blas.h"
...
...
@@ -58,14 +60,14 @@ HOSTDEVICE T DmcnGetGradientWeight(T argmax_h,
return
weight
;
}
template
<
typename
T
>
HOSTDEVICE
T
DmcnGetCoordinateWeight
(
T
argmax_h
,
T
argmax_w
,
const
int
height
,
const
int
width
,
const
T
*
im_data
,
const
int
data_width
,
const
int
bp_dir
)
{
template
<
typename
T
,
typename
MT
>
HOSTDEVICE
MT
DmcnGetCoordinateWeight
(
M
T
argmax_h
,
M
T
argmax_w
,
const
int
height
,
const
int
width
,
const
T
*
im_data
,
const
int
data_width
,
const
int
bp_dir
)
{
if
(
argmax_h
<=
-
1
||
argmax_h
>=
height
||
argmax_w
<=
-
1
||
argmax_w
>=
width
)
{
return
0
;
...
...
@@ -76,43 +78,51 @@ HOSTDEVICE T DmcnGetCoordinateWeight(T argmax_h,
int
argmax_h_high
=
argmax_h_low
+
1
;
int
argmax_w_high
=
argmax_w_low
+
1
;
T
weight
=
0
;
M
T
weight
=
0
;
if
(
bp_dir
==
0
)
{
weight
+=
(
argmax_h_low
>=
0
&&
argmax_w_low
>=
0
)
?
-
1
*
(
argmax_w_low
+
1
-
argmax_w
)
*
im_data
[
argmax_h_low
*
data_width
+
argmax_w_low
]
static_cast
<
MT
>
(
im_data
[
argmax_h_low
*
data_width
+
argmax_w_low
])
:
0
;
weight
+=
(
argmax_h_low
>=
0
&&
argmax_w_high
<=
width
-
1
)
?
-
1
*
(
argmax_w
-
argmax_w_low
)
*
im_data
[
argmax_h_low
*
data_width
+
argmax_w_high
]
static_cast
<
MT
>
(
im_data
[
argmax_h_low
*
data_width
+
argmax_w_high
])
:
0
;
weight
+=
(
argmax_h_high
<=
height
-
1
&&
argmax_w_low
>=
0
)
?
(
argmax_w_low
+
1
-
argmax_w
)
*
im_data
[
argmax_h_high
*
data_width
+
argmax_w_low
]
static_cast
<
MT
>
(
im_data
[
argmax_h_high
*
data_width
+
argmax_w_low
])
:
0
;
weight
+=
(
argmax_h_high
<=
height
-
1
&&
argmax_w_high
<=
width
-
1
)
?
(
argmax_w
-
argmax_w_low
)
*
im_data
[
argmax_h_high
*
data_width
+
argmax_w_high
]
static_cast
<
MT
>
(
im_data
[
argmax_h_high
*
data_width
+
argmax_w_high
])
:
0
;
}
else
if
(
bp_dir
==
1
)
{
weight
+=
(
argmax_h_low
>=
0
&&
argmax_w_low
>=
0
)
?
-
1
*
(
argmax_h_low
+
1
-
argmax_h
)
*
im_data
[
argmax_h_low
*
data_width
+
argmax_w_low
]
static_cast
<
MT
>
(
im_data
[
argmax_h_low
*
data_width
+
argmax_w_low
])
:
0
;
weight
+=
(
argmax_h_low
>=
0
&&
argmax_w_high
<=
width
-
1
)
?
(
argmax_h_low
+
1
-
argmax_h
)
*
im_data
[
argmax_h_low
*
data_width
+
argmax_w_high
]
static_cast
<
MT
>
(
im_data
[
argmax_h_low
*
data_width
+
argmax_w_high
])
:
0
;
weight
+=
(
argmax_h_high
<=
height
-
1
&&
argmax_w_low
>=
0
)
?
-
1
*
(
argmax_h
-
argmax_h_low
)
*
im_data
[
argmax_h_high
*
data_width
+
argmax_w_low
]
static_cast
<
MT
>
(
im_data
[
argmax_h_high
*
data_width
+
argmax_w_low
])
:
0
;
weight
+=
(
argmax_h_high
<=
height
-
1
&&
argmax_w_high
<=
width
-
1
)
?
(
argmax_h
-
argmax_h_low
)
*
im_data
[
argmax_h_high
*
data_width
+
argmax_w_high
]
static_cast
<
MT
>
(
im_data
[
argmax_h_high
*
data_width
+
argmax_w_high
])
:
0
;
}
...
...
@@ -135,7 +145,7 @@ void ModulatedDeformableCol2imCoord(const Context& dev_ctx,
T
*
grad_offset
,
T
*
grad_mask
);
template
<
typename
T
,
typename
Context
>
template
<
typename
T
,
typename
MT
,
typename
Context
>
void
ModulatedDeformableCol2im
(
const
Context
&
dev_ctx
,
const
T
*
data_col
,
const
T
*
data_offset
,
...
...
@@ -147,7 +157,7 @@ void ModulatedDeformableCol2im(const Context& dev_ctx,
const
std
::
vector
<
int
>&
stride
,
const
std
::
vector
<
int
>&
dilation
,
const
int
deformable_group
,
T
*
grad_im
);
M
T
*
grad_im
);
template
<
typename
T
,
typename
Context
>
void
FilterGradAddup
(
const
Context
&
dev_ctx
,
...
...
@@ -176,7 +186,7 @@ void DeformableConvGradKernel(const Context& dev_ctx,
DenseTensor
*
filter_grad
,
DenseTensor
*
mask_grad
)
{
const
int
batch_size
=
static_cast
<
int
>
(
x
.
dims
()[
0
]);
using
MT
=
typename
phi
::
dtype
::
MPTypeTrait
<
T
>::
Type
;
DDim
input_shape
=
phi
::
slice_ddim
(
x
.
dims
(),
1
,
x
.
dims
().
size
());
std
::
vector
<
int64_t
>
input_shape_vec
=
phi
::
vectorize
(
input_shape
);
std
::
vector
<
int64_t
>
filter_shape_vec
(
phi
::
vectorize
(
filter
.
dims
()));
...
...
@@ -292,8 +302,8 @@ void DeformableConvGradKernel(const Context& dev_ctx,
mask_grad_data_ptr
);
}
if
(
dx
)
{
T
*
dx_ptr
=
dx
->
data
<
T
>
(
);
// get grad of input
MT
*
mt_dx_ptr
=
dev_ctx
.
template
Alloc
<
MT
>(
dx
);
ModulatedDeformableCol2im
(
dev_ctx
,
col_buffer_ptr
,
offset_ptr
+
i
*
im2col_step
*
input_offset_dim
,
...
...
@@ -305,7 +315,7 @@ void DeformableConvGradKernel(const Context& dev_ctx,
strides
,
dilations
,
deformable_groups
,
dx_ptr
+
i
*
im2col_step
*
input_dim
);
mt_
dx_ptr
+
i
*
im2col_step
*
input_dim
);
dx
->
Resize
(
x
.
dims
());
}
...
...
paddle/phi/kernels/impl/deformable_conv_kernel_impl.h
浏览文件 @
5e0614a1
...
...
@@ -14,11 +14,13 @@
#pragma once
#include "paddle/phi/common/amp_type_traits.h"
#include "paddle/phi/core/dense_tensor.h"
#include "paddle/phi/core/hostdevice.h"
#include "paddle/phi/kernels/empty_kernel.h"
#include "paddle/phi/kernels/funcs/blas/blas.h"
#include "paddle/phi/kernels/funcs/deformable_conv_functor.h"
#include "paddle/phi/kernels/transpose_kernel.h"
#include "paddle/utils/optional.h"
namespace
phi
{
...
...
@@ -38,6 +40,12 @@ void DeformableConvKernel(const Context& dev_ctx,
DenseTensor
*
out
)
{
const
int
batch_size
=
static_cast
<
int
>
(
x
.
dims
()[
0
]);
int
temp_step
=
std
::
min
(
64
,
batch_size
);
if
(
batch_size
%
temp_step
==
0
)
{
im2col_step
=
temp_step
;
}
using
MT
=
typename
phi
::
dtype
::
MPTypeTrait
<
T
>::
Type
;
std
::
vector
<
int64_t
>
filter_shape_vec
(
phi
::
vectorize
(
filter
.
dims
()));
std
::
vector
<
int64_t
>
output_shape_vec
(
phi
::
vectorize
(
out
->
dims
()));
...
...
@@ -101,8 +109,11 @@ void DeformableConvKernel(const Context& dev_ctx,
dilations
,
deformable_groups
,
col_buffer_ptr
);
DenseTensor
output_3d
=
output_4d
.
Slice
(
i
,
i
+
1
).
Resize
(
phi
::
slice_ddim
(
output_4d
.
dims
(),
1
,
output_4d
.
dims
().
size
()));
DenseTensor
output_3d
=
output_4d
.
Slice
(
i
,
i
+
1
).
Resize
(
phi
::
slice_ddim
(
output_4d
.
dims
(),
1
,
output_4d
.
dims
().
size
()));
// group * C/group * (im2step * H * W)
// get the product of pixel and weight
for
(
int
g
=
0
;
g
<
groups
;
++
g
)
{
DenseTensor
weight_3d_slice
=
weight_3d
.
Slice
(
g
,
g
+
1
).
Resize
(
...
...
@@ -110,8 +121,11 @@ void DeformableConvKernel(const Context& dev_ctx,
DenseTensor
col_buffer_3d_slice
=
col_buffer_3d
.
Slice
(
g
,
g
+
1
).
Resize
(
phi
::
slice_ddim
(
col_buffer_3d
.
dims
(),
1
,
col_buffer_3d
.
dims
().
size
()));
DenseTensor
output_3d_slice
=
output_3d
.
Slice
(
g
,
g
+
1
).
Resize
(
phi
::
slice_ddim
(
output_3d
.
dims
(),
1
,
output_3d
.
dims
().
size
()));
DenseTensor
output_3d_slice
=
output_3d
.
Slice
(
g
,
g
+
1
).
Resize
(
phi
::
slice_ddim
(
output_3d
.
dims
(),
1
,
output_3d
.
dims
().
size
()));
// C * ((im2col_step)*H*W))
blas
.
MatMul
(
weight_3d_slice
,
false
,
col_buffer_3d_slice
,
...
...
@@ -121,7 +135,29 @@ void DeformableConvKernel(const Context& dev_ctx,
T
(
0.0
));
}
}
out
->
ShareDataWith
(
output_buffer
).
Resize
(
phi
::
make_ddim
(
output_shape_vec
));
// swap axis to get the right result when im2col_step is greater than 1
if
(
im2col_step
>
1
)
{
std
::
vector
<
int
>
axis
(
4
);
axis
[
0
]
=
0
;
axis
[
1
]
=
2
;
axis
[
2
]
=
1
;
axis
[
3
]
=
3
;
DenseTensor
real_output_buffer
=
phi
::
Transpose
<
T
,
Context
>
(
dev_ctx
,
output_4d
.
Resize
(
phi
::
make_ddim
({
batch_size
/
im2col_step
,
output_shape_vec
[
1
],
im2col_step
,
output_shape_vec
[
2
]
*
output_shape_vec
[
3
]})),
axis
);
out
->
ShareDataWith
(
real_output_buffer
)
.
Resize
(
phi
::
make_ddim
(
output_shape_vec
));
}
else
{
out
->
ShareDataWith
(
output_buffer
).
Resize
(
phi
::
make_ddim
(
output_shape_vec
));
}
}
}
// namespace phi
python/paddle/fluid/tests/unittests/test_deformable_conv_v1_op.py
浏览文件 @
5e0614a1
...
...
@@ -17,9 +17,11 @@ import unittest
import
numpy
as
np
import
paddle.fluid
as
fluid
import
paddle.fluid.core
as
core
from
op_test
import
OpTest
from
op_test
import
OpTest
,
skip_check_grad_ci
from
paddle.fluid.framework
import
_test_eager_guard
paddle
.
enable_static
()
def
dmc_bilinear
(
data_im
,
height
,
width
,
h
,
w
):
h_low
=
int
(
np
.
floor
(
h
))
...
...
@@ -59,8 +61,8 @@ def dconv_im2col_gemm(input, offset, filter, group, conv_param):
assert
f_c
*
group
==
in_c
assert
np
.
mod
(
out_c
,
group
)
==
0
stride
,
pad
,
dilation
=
conv_param
[
'stride'
],
conv_param
[
'pad'
],
\
conv_param
[
'dilation'
]
stride
,
pad
,
dilation
=
conv_param
[
'stride'
],
conv_param
[
'pad'
],
\
conv_param
[
'dilation'
]
out_h
=
1
+
(
in_h
+
2
*
pad
[
0
]
-
(
dilation
[
0
]
*
(
f_h
-
1
)
+
1
))
//
stride
[
0
]
out_w
=
1
+
(
in_w
+
2
*
pad
[
1
]
-
(
dilation
[
1
]
*
(
f_w
-
1
)
+
1
))
//
stride
[
1
]
assert
out_h
==
in_h
...
...
@@ -74,18 +76,18 @@ def dconv_im2col_gemm(input, offset, filter, group, conv_param):
for
kh
in
range
(
f_h
):
for
kw
in
range
(
f_w
):
offset_h_table
=
\
offset
[
n
,
::
2
,
h
,
w
].
reshape
(
f_h
,
f_w
)
offset
[
n
,
::
2
,
h
,
w
].
reshape
(
f_h
,
f_w
)
offset_w_table
=
\
offset
[
n
,
1
::
2
,
h
,
w
].
reshape
(
f_h
,
f_w
)
offset
[
n
,
1
::
2
,
h
,
w
].
reshape
(
f_h
,
f_w
)
offset_h
=
offset_h_table
[
kh
,
kw
]
offset_w
=
offset_w_table
[
kh
,
kw
]
val
=
0
im_h
=
h
*
stride
[
0
]
+
kh
*
dilation
[
0
]
\
+
offset_h
-
pad
[
0
]
+
offset_h
-
pad
[
0
]
im_w
=
w
*
stride
[
0
]
+
kw
*
dilation
[
0
]
\
+
offset_w
-
pad
[
1
]
+
offset_w
-
pad
[
1
]
if
im_h
>
-
1
and
im_w
>
-
1
and
\
im_h
<
in_h
and
im_w
<
in_h
:
im_h
<
in_h
and
im_w
<
in_h
:
val
=
dmc_bilinear
(
input
[
n
,
c
],
in_h
,
in_w
,
im_h
,
im_w
)
val_out
=
val
...
...
@@ -284,6 +286,69 @@ class TestWithDouble(TestModulatedDeformableConvOp):
self
.
dtype
=
np
.
float64
class
TestFP16
(
unittest
.
TestCase
):
def
check_main
(
self
,
input_np
,
offset_np
,
filter_np
,
dtype
):
paddle
.
disable_static
()
input_np
=
input_np
.
astype
(
dtype
)
offset_np
=
offset_np
.
astype
(
dtype
)
filter_np
=
filter_np
.
astype
(
dtype
)
input
=
paddle
.
to_tensor
(
input_np
)
offset
=
paddle
.
to_tensor
(
offset_np
)
filter
=
paddle
.
to_tensor
(
filter_np
)
input
.
stop_gradient
=
False
offset
.
stop_gradient
=
False
filter
.
stop_gradient
=
False
y
=
paddle
.
vision
.
ops
.
deform_conv2d
(
input
,
offset
,
filter
)
input_grad
,
offset_grad
,
filter_grad
=
paddle
.
grad
(
y
,
[
input
,
offset
,
filter
])
y_np
=
y
.
numpy
().
astype
(
'float32'
)
input_grad_np
=
input_grad
.
numpy
().
astype
(
'float32'
)
offset_grad_np
=
offset_grad
.
numpy
().
astype
(
'float32'
)
filter_grad_np
=
filter_grad
.
numpy
().
astype
(
'float32'
)
paddle
.
enable_static
()
return
y_np
,
input_grad_np
,
offset_grad_np
,
filter_grad_np
def
test_main
(
self
):
if
not
paddle
.
is_compiled_with_cuda
():
return
self
.
pad
=
[
1
,
1
]
self
.
stride
=
[
1
,
1
]
self
.
dilations
=
[
1
,
1
]
self
.
groups
=
1
self
.
input_size
=
[
2
,
3
,
5
,
5
]
# NCHW
assert
np
.
mod
(
self
.
input_size
[
1
],
self
.
groups
)
==
0
f_c
=
self
.
input_size
[
1
]
//
self
.
groups
self
.
filter_size
=
[
40
,
f_c
,
1
,
1
]
self
.
im2col_step
=
1
self
.
deformable_groups
=
1
offset_c
=
2
*
self
.
deformable_groups
*
self
.
filter_size
[
2
]
*
self
.
filter_size
[
3
]
self
.
offset_size
=
[
self
.
input_size
[
0
],
offset_c
,
self
.
input_size
[
2
],
self
.
input_size
[
3
]
]
input
=
np
.
random
.
random
(
self
.
input_size
)
offset
=
10
*
np
.
random
.
random
(
self
.
offset_size
)
filter
=
np
.
random
.
random
(
self
.
filter_size
)
y_np_1
,
input_g_np_1
,
offset_g_np_1
,
filter_g_np_1
=
self
.
check_main
(
input
,
offset
,
filter
,
'float16'
)
y_np_2
,
input_g_np_2
,
offset_g_np_2
,
filter_g_np_2
=
self
.
check_main
(
input
,
offset
,
filter
,
'float32'
)
def
assert_equal
(
x
,
y
):
np
.
testing
.
assert_allclose
(
x
,
y
,
atol
=
3e-2
)
assert_equal
(
y_np_1
,
y_np_2
)
assert_equal
(
input_g_np_1
,
input_g_np_2
)
assert_equal
(
offset_g_np_1
,
offset_g_np_2
)
assert_equal
(
filter_g_np_1
,
filter_g_np_2
)
class
TestModulatedDeformableConvV1InvalidInput
(
unittest
.
TestCase
):
def
test_error
(
self
):
...
...
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