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
276017bb
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
276017bb
编写于
3月 21, 2022
作者:
Z
zhangyikun02
提交者:
GitHub
3月 21, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
conv2d support FP16 on xpu and update unittest for conv2d, test=kunlun (#40395)
上级
1eb96eec
变更
3
展开全部
隐藏空白更改
内联
并排
Showing
3 changed file
with
382 addition
and
338 deletion
+382
-338
paddle/fluid/operators/conv_op_xpu.cc
paddle/fluid/operators/conv_op_xpu.cc
+51
-26
paddle/fluid/platform/device/xpu/xpu2_op_list.h
paddle/fluid/platform/device/xpu/xpu2_op_list.h
+8
-4
python/paddle/fluid/tests/unittests/xpu/test_conv2d_op_xpu.py
...on/paddle/fluid/tests/unittests/xpu/test_conv2d_op_xpu.py
+323
-308
未找到文件。
paddle/fluid/operators/conv_op_xpu.cc
浏览文件 @
276017bb
...
...
@@ -19,14 +19,16 @@ namespace operators {
template
<
typename
DeviceContext
,
typename
T
>
class
GemmConvXPUKernel
:
public
framework
::
OpKernel
<
T
>
{
using
XPUT
=
typename
XPUTypeTrait
<
T
>::
Type
;
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
const
Tensor
*
input
=
context
.
Input
<
Tensor
>
(
"Input"
);
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
const
Tensor
*
input
=
context
.
Input
<
Tensor
>
(
"Input"
);
// The filter will be reshaped in the calculations,
// so here use an assignment operation,
// that avoids modifying the variable in the Scope.
Tensor
filter
=
*
context
.
Input
<
Tensor
>
(
"Filter"
);
Tensor
*
output
=
context
.
Output
<
Tensor
>
(
"Output"
);
Tensor
*
output
=
context
.
Output
<
Tensor
>
(
"Output"
);
output
->
mutable_data
<
T
>
(
context
.
GetPlace
());
int
groups
=
context
.
Attr
<
int
>
(
"groups"
);
std
::
vector
<
int
>
strides
=
context
.
Attr
<
std
::
vector
<
int
>>
(
"strides"
);
...
...
@@ -53,11 +55,16 @@ class GemmConvXPUKernel : public framework::OpKernel<T> {
const
int
img_h
=
static_cast
<
int
>
(
input
->
dims
()[
2
]);
const
int
img_w
=
static_cast
<
int
>
(
input
->
dims
()[
3
]);
const
int
f
=
static_cast
<
int
>
(
filter
.
dims
()[
0
]);
auto
&
dev_ctx
=
context
.
template
device_context
<
DeviceContext
>();
int
r
=
xpu
::
conv2d
<
float
,
float
,
float
,
int16_t
>
(
dev_ctx
.
x_context
(),
input
->
data
<
float
>
(),
filter
.
data
<
float
>
(),
output
->
data
<
float
>
(),
batch_size
,
img_c
,
img_h
,
img_w
,
f
,
ksize
,
strides
,
paddings
,
dilations
,
groups
,
nullptr
,
nullptr
,
nullptr
,
true
);
const
XPUT
*
input_data
=
reinterpret_cast
<
const
XPUT
*>
(
input
->
data
<
T
>
());
const
XPUT
*
filter_data
=
reinterpret_cast
<
const
XPUT
*>
(
filter
.
data
<
T
>
());
XPUT
*
output_data
=
reinterpret_cast
<
XPUT
*>
(
output
->
data
<
T
>
());
auto
&
dev_ctx
=
context
.
template
device_context
<
DeviceContext
>();
int
r
=
xpu
::
conv2d
<
XPUT
,
XPUT
,
XPUT
,
int16_t
>
(
dev_ctx
.
x_context
(),
input_data
,
filter_data
,
output_data
,
batch_size
,
img_c
,
img_h
,
img_w
,
f
,
ksize
,
strides
,
paddings
,
dilations
,
groups
,
nullptr
,
nullptr
,
nullptr
,
true
);
PADDLE_ENFORCE_EQ
(
r
,
XPU_SUCCESS
,
platform
::
errors
::
External
(
"XPU conv kernel return wrong value[%d %s]"
,
...
...
@@ -67,14 +74,16 @@ class GemmConvXPUKernel : public framework::OpKernel<T> {
template
<
typename
DeviceContext
,
typename
T
>
class
GemmConvGradXPUKernel
:
public
framework
::
OpKernel
<
T
>
{
using
XPUT
=
typename
XPUTypeTrait
<
T
>::
Type
;
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
const
Tensor
*
input
=
context
.
Input
<
Tensor
>
(
"Input"
);
const
Tensor
*
output_grad
=
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
const
Tensor
*
input
=
context
.
Input
<
Tensor
>
(
"Input"
);
const
Tensor
*
output_grad
=
context
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Output"
));
Tensor
*
input_grad
=
Tensor
*
input_grad
=
context
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"Input"
));
Tensor
*
filter_grad
=
Tensor
*
filter_grad
=
context
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"Filter"
));
// The filter and filter_grad will be reshaped in the calculations,
// so here use an assignment operation,
...
...
@@ -107,19 +116,27 @@ class GemmConvGradXPUKernel : public framework::OpKernel<T> {
const
int
img_h
=
static_cast
<
int
>
(
input
->
dims
()[
2
]);
const
int
img_w
=
static_cast
<
int
>
(
input
->
dims
()[
3
]);
const
int
f
=
static_cast
<
int
>
(
filter
.
dims
()[
0
]);
const
XPUT
*
input_data
=
reinterpret_cast
<
const
XPUT
*>
(
input
->
data
<
T
>
());
const
XPUT
*
filter_data
=
reinterpret_cast
<
const
XPUT
*>
(
filter
.
data
<
T
>
());
const
XPUT
*
output_grad_data
=
reinterpret_cast
<
const
XPUT
*>
(
output_grad
->
data
<
T
>
());
XPUT
*
input_grad_data
=
nullptr
;
if
(
input_grad
)
{
input_grad
->
mutable_data
<
T
>
(
context
.
GetPlace
());
input_grad_data
=
reinterpret_cast
<
XPUT
*>
(
input_grad
->
data
<
T
>
());
}
XPUT
*
filter_grad_data
=
nullptr
;
if
(
filter_grad
)
{
filter_grad
->
mutable_data
<
T
>
(
context
.
GetPlace
());
filter_grad_data
=
reinterpret_cast
<
XPUT
*>
(
filter_grad
->
data
<
T
>
());
}
auto
&
dev_ctx
=
context
.
template
device_context
<
DeviceContext
>();
int
r
=
xpu
::
conv2d_grad
<
float
,
float
,
float
,
int16_t
>
(
dev_ctx
.
x_context
(),
input
->
data
<
T
>
(),
filter
.
data
<
T
>
(),
output_grad
->
data
<
T
>
(),
input_grad
?
input_grad
->
data
<
T
>
()
:
nullptr
,
filter_grad
?
filter_grad
->
data
<
T
>
()
:
nullptr
,
batch_size
,
img_c
,
img_h
,
img_w
,
f
,
ksize
,
strides
,
paddings
,
dilations
,
groups
,
nullptr
,
nullptr
,
nullptr
,
nullptr
,
nullptr
,
true
);
auto
&
dev_ctx
=
context
.
template
device_context
<
DeviceContext
>();
int
r
=
xpu
::
conv2d_grad
<
XPUT
,
XPUT
,
XPUT
,
int16_t
>
(
dev_ctx
.
x_context
(),
input_data
,
filter_data
,
output_grad_data
,
input_grad_data
,
filter_grad_data
,
batch_size
,
img_c
,
img_h
,
img_w
,
f
,
ksize
,
strides
,
paddings
,
dilations
,
groups
,
nullptr
,
nullptr
,
nullptr
,
nullptr
,
nullptr
,
true
);
PADDLE_ENFORCE_EQ
(
r
,
XPU_SUCCESS
,
platform
::
errors
::
External
(
"XPU conv kernel return wrong value[%d %s]"
,
...
...
@@ -130,14 +147,22 @@ class GemmConvGradXPUKernel : public framework::OpKernel<T> {
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_XPU_KERNEL
(
depthwise_conv2d
,
ops
::
GemmConvXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
float
>
);
REGISTER_OP_XPU_KERNEL
(
conv2d
,
ops
::
GemmConvXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
float
>
);
conv2d
,
ops
::
GemmConvXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
float
>
,
ops
::
GemmConvXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
paddle
::
platform
::
float16
>
);
REGISTER_OP_XPU_KERNEL
(
conv2d_grad
,
ops
::
GemmConvGradXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
float
>
);
ops
::
GemmConvGradXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
float
>
,
ops
::
GemmConvGradXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
paddle
::
platform
::
float16
>
);
REGISTER_OP_XPU_KERNEL
(
depthwise_conv2d
,
ops
::
GemmConvXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
float
>
,
ops
::
GemmConvXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
paddle
::
platform
::
float16
>
);
REGISTER_OP_XPU_KERNEL
(
depthwise_conv2d_grad
,
ops
::
GemmConvGradXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
float
>
);
ops
::
GemmConvGradXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
float
>
,
ops
::
GemmConvGradXPUKernel
<
paddle
::
platform
::
XPUDeviceContext
,
paddle
::
platform
::
float16
>
);
#endif
paddle/fluid/platform/device/xpu/xpu2_op_list.h
浏览文件 @
276017bb
...
...
@@ -51,16 +51,20 @@ XPUOpMap& get_kl2_ops() {
{
"clip"
,
XPUKernelSet
({
pOpKernelType
(
vartype
::
FP32
,
XPUPlace
())})},
{
"concat_grad"
,
XPUKernelSet
({
pOpKernelType
(
vartype
::
FP32
,
XPUPlace
())})},
{
"concat"
,
XPUKernelSet
({
pOpKernelType
(
vartype
::
FP32
,
XPUPlace
())})},
{
"conv2d_grad"
,
XPUKernelSet
({
pOpKernelType
(
vartype
::
FP32
,
XPUPlace
())})},
{
"conv2d"
,
XPUKernelSet
({
pOpKernelType
(
vartype
::
FP32
,
XPUPlace
())})},
{
"conv2d_grad"
,
XPUKernelSet
({
pOpKernelType
(
vartype
::
FP32
,
XPUPlace
()),
pOpKernelType
(
vartype
::
FP16
,
XPUPlace
())})},
{
"conv2d"
,
XPUKernelSet
({
pOpKernelType
(
vartype
::
FP32
,
XPUPlace
()),
pOpKernelType
(
vartype
::
FP16
,
XPUPlace
())})},
{
"conv2d_transpose_grad"
,
XPUKernelSet
({
pOpKernelType
(
vartype
::
FP32
,
XPUPlace
())})},
{
"conv2d_transpose"
,
XPUKernelSet
({
pOpKernelType
(
vartype
::
FP32
,
XPUPlace
())})},
{
"depthwise_conv2d_grad"
,
XPUKernelSet
({
pOpKernelType
(
vartype
::
FP32
,
XPUPlace
())})},
XPUKernelSet
({
pOpKernelType
(
vartype
::
FP32
,
XPUPlace
()),
pOpKernelType
(
vartype
::
FP16
,
XPUPlace
())})},
{
"depthwise_conv2d"
,
XPUKernelSet
({
pOpKernelType
(
vartype
::
FP32
,
XPUPlace
())})},
XPUKernelSet
({
pOpKernelType
(
vartype
::
FP32
,
XPUPlace
()),
pOpKernelType
(
vartype
::
FP16
,
XPUPlace
())})},
{
"dropout_grad"
,
XPUKernelSet
({
pOpKernelType
(
vartype
::
FP32
,
XPUPlace
())})},
{
"dropout"
,
XPUKernelSet
({
pOpKernelType
(
vartype
::
FP32
,
XPUPlace
())})},
...
...
python/paddle/fluid/tests/unittests/xpu/test_conv2d_op_xpu.py
浏览文件 @
276017bb
此差异已折叠。
点击以展开。
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