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
OPTHREE
Paddle
提交
3b9db171
P
Paddle
项目概览
OPTHREE
/
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
3b9db171
编写于
3月 03, 2021
作者:
Q
Qi Li
提交者:
GitHub
3月 03, 2021
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[ROCM] update fluid operators for rocm (part7), test=develop (#31307)
上级
db50fb67
变更
19
隐藏空白更改
内联
并排
Showing
19 changed file
with
167 addition
and
37 deletion
+167
-37
paddle/fluid/operators/CMakeLists.txt
paddle/fluid/operators/CMakeLists.txt
+9
-4
paddle/fluid/operators/bmm_op.cu
paddle/fluid/operators/bmm_op.cu
+1
-1
paddle/fluid/operators/cholesky_op.cu
paddle/fluid/operators/cholesky_op.cu
+5
-0
paddle/fluid/operators/clip_op.h
paddle/fluid/operators/clip_op.h
+2
-2
paddle/fluid/operators/coalesce_tensor_op.cc
paddle/fluid/operators/coalesce_tensor_op.cc
+1
-1
paddle/fluid/operators/correlation_op.cu
paddle/fluid/operators/correlation_op.cu
+5
-0
paddle/fluid/operators/cudnn_lstm_op.cu.cc
paddle/fluid/operators/cudnn_lstm_op.cu.cc
+62
-11
paddle/fluid/operators/cumsum_op.cu
paddle/fluid/operators/cumsum_op.cu
+7
-1
paddle/fluid/operators/data_norm_op.cu
paddle/fluid/operators/data_norm_op.cu
+6
-2
paddle/fluid/operators/diag_embed_op.h
paddle/fluid/operators/diag_embed_op.h
+1
-1
paddle/fluid/operators/dot_op.h
paddle/fluid/operators/dot_op.h
+2
-2
paddle/fluid/operators/dropout_op.cu
paddle/fluid/operators/dropout_op.cu
+51
-3
paddle/fluid/operators/dropout_op.h
paddle/fluid/operators/dropout_op.h
+2
-2
paddle/fluid/operators/fake_quantize_op.cu
paddle/fluid/operators/fake_quantize_op.cu
+4
-0
paddle/fluid/operators/fill_constant_batch_size_like_op.h
paddle/fluid/operators/fill_constant_batch_size_like_op.h
+1
-1
paddle/fluid/operators/fill_constant_op.h
paddle/fluid/operators/fill_constant_op.h
+2
-2
paddle/fluid/operators/filter_by_instag_op.h
paddle/fluid/operators/filter_by_instag_op.h
+1
-1
paddle/fluid/operators/gelu_op.h
paddle/fluid/operators/gelu_op.h
+4
-2
paddle/fluid/operators/get_tensor_from_selected_rows_op.cc
paddle/fluid/operators/get_tensor_from_selected_rows_op.cc
+1
-1
未找到文件。
paddle/fluid/operators/CMakeLists.txt
浏览文件 @
3b9db171
...
...
@@ -73,9 +73,11 @@ register_operators(EXCLUDES py_func_op warpctc_op dgc_op lstm_op run_program_op
op_library
(
run_program_op SRCS run_program_op.cc run_program_op.cu.cc DEPS executor_cache
${
OP_HEADER_DEPS
}
)
if
(
WITH_GPU
)
if
(
WITH_GPU OR WITH_ROCM
)
if
(
WITH_ROCM
)
op_library
(
warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale SRCS warpctc_op.cc warpctc_op.cu.cc
)
# warpctc_op needs cudnn 7 above
if
(
${
CUDNN_MAJOR_VERSION
}
VERSION_LESS 7
)
elseif
(
${
CUDNN_MAJOR_VERSION
}
VERSION_LESS 7
)
op_library
(
warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale SRCS warpctc_op.cc warpctc_op.cu.cc
)
else
()
op_library
(
warpctc_op DEPS dynload_warpctc sequence_padding sequence_scale
)
...
...
@@ -108,7 +110,7 @@ set(COMMON_OP_DEPS ${COMMON_OP_DEPS} sequence_padding sequence_scale cos_sim_fun
set
(
COMMON_OP_DEPS
${
COMMON_OP_DEPS
}
sequence2batch lstm_compute matrix_bit_code gru_compute activation_functions beam_search fc matrix_inverse
)
set
(
COMMON_OP_DEPS
${
COMMON_OP_DEPS
}
box_wrapper boost ps_gpu_wrapper
)
set
(
COMMON_OP_DEPS
${
COMMON_OP_DEPS
}
common_infer_shape_functions
)
if
(
WITH_GPU
)
if
(
WITH_GPU
OR WITH_ROCM
)
set
(
COMMON_OP_DEPS
${
COMMON_OP_DEPS
}
depthwise_conv prelu bert_encoder_functor
)
endif
()
set
(
COMMON_OP_DEPS
${
COMMON_OP_DEPS
}
device_memory_aligment
)
...
...
@@ -139,9 +141,12 @@ cc_test(beam_search_decode_op_test SRCS beam_search_decode_op_test.cc DEPS lod_t
cc_test
(
strided_memcpy_test SRCS strided_memcpy_test.cc DEPS tensor memory
)
cc_test
(
save_load_op_test SRCS save_load_op_test.cc DEPS save_op load_op
)
cc_test
(
save_load_combine_op_test SRCS save_load_combine_op_test.cc DEPS save_combine_op load_combine_op
)
nv_test
(
dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor generator
)
if
(
WITH_GPU
)
nv_test
(
dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor generator
)
nv_test
(
test_leaky_relu_grad_grad_functor SRCS test_leaky_relu_grad_grad_functor.cc test_leaky_relu_grad_grad_functor.cu DEPS tensor device_context eigen3
)
elseif
(
WITH_ROCM
)
hip_test
(
dropout_op_test SRCS dropout_op_test.cc DEPS dropout_op tensor generator
)
hip_test
(
test_leaky_relu_grad_grad_functor SRCS test_leaky_relu_grad_grad_functor.cc test_leaky_relu_grad_grad_functor.cu DEPS tensor device_context eigen3
)
else
()
cc_test
(
test_leaky_relu_grad_grad_functor SRCS test_leaky_relu_grad_grad_functor.cc DEPS tensor device_context eigen3
)
endif
()
...
...
paddle/fluid/operators/bmm_op.cu
浏览文件 @
3b9db171
...
...
@@ -11,7 +11,7 @@
#include "paddle/fluid/operators/bmm_op.h"
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_CUDA_KERNEL
(
bmm
,
ops
::
BmmKernel
<
paddle
::
platform
::
CUDADeviceContext
,
float
>
,
...
...
paddle/fluid/operators/cholesky_op.cu
浏览文件 @
3b9db171
...
...
@@ -12,6 +12,9 @@ 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. */
#ifndef PADDLE_WITH_HIP
// HIP not support cusolver
#include <thrust/device_vector.h>
#include <algorithm>
#include <vector>
...
...
@@ -164,3 +167,5 @@ REGISTER_OP_CUDA_KERNEL(
cholesky_grad
,
ops
::
CholeskyGradKernel
<
paddle
::
platform
::
CUDADeviceContext
,
float
>
,
ops
::
CholeskyGradKernel
<
paddle
::
platform
::
CUDADeviceContext
,
double
>
);
#endif // not PADDLE_WITH_HIP
paddle/fluid/operators/clip_op.h
浏览文件 @
3b9db171
...
...
@@ -25,7 +25,7 @@ namespace operators {
using
framework
::
Tensor
;
using
platform
::
Transform
;
#if
def __NVCC__
#if
defined(__NVCC__) || defined(__HIPCC__)
template
<
typename
T
,
typename
UnaryOperation
>
__global__
void
ClipCudaKernel
(
const
T
*
input
,
T
*
out
,
int
num
,
UnaryOperation
op
)
{
...
...
@@ -105,7 +105,7 @@ class ClipKernel : public framework::OpKernel<T> {
const
T
*
x_data
=
x
->
data
<
T
>
();
int64_t
numel
=
x
->
numel
();
if
(
platform
::
is_gpu_place
(
context
.
GetPlace
()))
{
#if
def __NVCC__
#if
defined(__NVCC__) || defined(__HIPCC__)
int
threads
=
256
;
int
blocks
=
(
numel
+
threads
-
1
)
/
threads
;
ClipCudaKernel
<
T
,
ClipFunctor
<
T
>><<<
...
...
paddle/fluid/operators/coalesce_tensor_op.cc
浏览文件 @
3b9db171
...
...
@@ -289,7 +289,7 @@ REGISTER_OP_CPU_KERNEL(
ops
::
CoalesceTensorOpKernel
<
paddle
::
platform
::
CPUDeviceContext
,
float
>
,
ops
::
CoalesceTensorOpKernel
<
paddle
::
platform
::
CPUDeviceContext
,
double
>
);
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
REGISTER_OP_CUDA_KERNEL
(
coalesce_tensor
,
ops
::
CoalesceTensorOpKernel
<
paddle
::
platform
::
CUDADeviceContext
,
...
...
paddle/fluid/operators/correlation_op.cu
浏览文件 @
3b9db171
...
...
@@ -12,6 +12,9 @@ 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. */
#ifndef PADDLE_WITH_HIP
// HIP not supported yet
#include <algorithm>
#include <string>
#include "paddle/fluid/framework/op_registry.h"
...
...
@@ -480,3 +483,5 @@ REGISTER_OP_CUDA_KERNEL(correlation, ops::CorrelationCUDAKernel<float>,
ops
::
CorrelationCUDAKernel
<
double
>
);
REGISTER_OP_CUDA_KERNEL
(
correlation_grad
,
ops
::
CorrelationCUDAGradKernel
<
float
>
,
ops
::
CorrelationCUDAGradKernel
<
double
>
);
#endif // not PADDLE_WITH_HIP
paddle/fluid/operators/cudnn_lstm_op.cu.cc
浏览文件 @
3b9db171
...
...
@@ -14,9 +14,14 @@ limitations under the License. */
#include "paddle/fluid/framework/generator.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/cudnn_lstm_cache.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/operators/utils.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/operators/cudnn_lstm_cache.h"
#endif
#ifdef PADDLE_WITH_HIP
#include "paddle/fluid/operators/miopen_lstm_cache.h"
#endif
namespace
paddle
{
namespace
platform
{
...
...
@@ -54,7 +59,7 @@ int size_sum(const std::vector<const Tensor *> &weight_list) {
}
template
<
typename
T
>
void
weight_to_tensor
(
const
platform
::
Place
&
place
,
cuda
Stream_t
stream
,
void
weight_to_tensor
(
const
platform
::
Place
&
place
,
gpu
Stream_t
stream
,
const
std
::
vector
<
const
Tensor
*>
&
weight_list
,
Tensor
*
weight
)
{
auto
weight_data
=
weight
->
data
<
T
>
();
...
...
@@ -72,7 +77,7 @@ void weight_to_tensor(const platform::Place &place, cudaStream_t stream,
}
template
<
typename
T
>
void
weight_to_tensor_list
(
const
platform
::
Place
&
place
,
cuda
Stream_t
stream
,
void
weight_to_tensor_list
(
const
platform
::
Place
&
place
,
gpu
Stream_t
stream
,
std
::
vector
<
Tensor
*>
*
weight_grad
,
const
std
::
vector
<
const
Tensor
*>
&
weight_input
,
const
Tensor
*
weight
)
{
...
...
@@ -92,23 +97,36 @@ void weight_to_tensor_list(const platform::Place &place, cudaStream_t stream,
}
template
<
typename
T
>
#ifdef PADDLE_WITH_HIP
void
LSTMInferece
(
const
bool
&
has_seq_length
,
const
miopenHandle_t
&
handle
,
#else
void
LSTMInferece
(
const
bool
&
has_seq_length
,
const
cudnnHandle_t
&
handle
,
#endif
const
int
&
seq_length
,
ScopedRNNBase
*
rnn
,
const
T
*
x_data
,
const
T
*
init_h_data
,
const
T
*
init_c_data
,
const
T
*
w_data
,
T
*
out_data
,
T
*
last_h_data
,
T
*
last_c_data
,
framework
::
Tensor
*
workspace_data
,
const
size_t
&
workspace_size
)
{
if
(
!
has_seq_length
)
{
// for inference
// This interface is used when the input/output is unpadded.
// for inference
// This interface is used when the input/output is unpadded.
#ifdef PADDLE_WITH_HIP
PADDLE_ENFORCE_CUDA_SUCCESS
(
platform
::
dynload
::
miopenRNNForwardInference
(
handle
,
rnn
->
rnn_desc
(),
seq_length
,
rnn
->
x_descs
(),
x_data
,
rnn
->
init_h_desc
(),
init_h_data
,
rnn
->
init_c_desc
(),
init_c_data
,
rnn
->
weight_desc
(),
w_data
,
rnn
->
y_descs
(),
out_data
,
rnn
->
last_h_desc
(),
last_h_data
,
rnn
->
last_c_desc
(),
last_c_data
,
workspace_data
->
data
<
uint8_t
>
(),
workspace_size
));
#else
PADDLE_ENFORCE_CUDA_SUCCESS
(
platform
::
dynload
::
cudnnRNNForwardInference
(
handle
,
rnn
->
rnn_desc
(),
seq_length
,
rnn
->
x_descs
(),
x_data
,
rnn
->
init_h_desc
(),
init_h_data
,
rnn
->
init_c_desc
(),
init_c_data
,
rnn
->
weight_desc
(),
w_data
,
rnn
->
y_descs
(),
out_data
,
rnn
->
last_h_desc
(),
last_h_data
,
rnn
->
last_c_desc
(),
last_c_data
,
workspace_data
->
data
<
uint8_t
>
(),
workspace_size
));
#endif
}
else
{
#if CUDNN_VERSION >= 7201
#if
!defined(PADDLE_WITH_HIP) &&
CUDNN_VERSION >= 7201
// for inference
// This interface is used when the input/output is padded.
PADDLE_ENFORCE_CUDA_SUCCESS
(
platform
::
dynload
::
cudnnRNNForwardInferenceEx
(
...
...
@@ -256,8 +274,17 @@ class CudnnLSTMGPUKernel : public framework::OpKernel<T> {
last_c_data
,
&
workspace_data_
,
workspace_size
);
}
else
{
if
(
!
has_seq_length
)
{
// for train
// This interface is used when the input/output is unpadded.
// for train
// This interface is used when the input/output is unpadded.
#ifdef PADDLE_WITH_HIP
PADDLE_ENFORCE_CUDA_SUCCESS
(
platform
::
dynload
::
miopenRNNForwardTraining
(
handle
,
rnn
.
rnn_desc
(),
seq_length
,
rnn
.
x_descs
(),
x_data
,
rnn
.
init_h_desc
(),
init_h_data
,
rnn
.
init_c_desc
(),
init_c_data
,
rnn
.
weight_desc
(),
w_data
,
rnn
.
y_descs
(),
out_data
,
rnn
.
last_h_desc
(),
last_h_data
,
rnn
.
last_c_desc
(),
last_c_data
,
workspace_data_
.
data
<
uint8_t
>
(),
workspace_size
,
reserve_data
,
reserve_size
));
#else
PADDLE_ENFORCE_CUDA_SUCCESS
(
platform
::
dynload
::
cudnnRNNForwardTraining
(
handle
,
rnn
.
rnn_desc
(),
seq_length
,
rnn
.
x_descs
(),
x_data
,
rnn
.
init_h_desc
(),
init_h_data
,
rnn
.
init_c_desc
(),
init_c_data
,
...
...
@@ -265,8 +292,9 @@ class CudnnLSTMGPUKernel : public framework::OpKernel<T> {
rnn
.
last_h_desc
(),
last_h_data
,
rnn
.
last_c_desc
(),
last_c_data
,
workspace_data_
.
data
<
uint8_t
>
(),
workspace_size
,
reserve_data
,
reserve_size
));
#endif
}
else
{
#if CUDNN_VERSION >= 7201
#if
!defined(PADDLE_WITH_HIP) &&
CUDNN_VERSION >= 7201
// for train
// This interface is used when the input/output is padded.
PADDLE_ENFORCE_CUDA_SUCCESS
(
...
...
@@ -403,7 +431,23 @@ class CudnnLSTMGPUGradKernel : public framework::OpKernel<T> {
const
uint8_t
*
reserve_data
=
reserve
->
data
<
uint8_t
>
();
if
(
!
has_seq_length
)
{
// This interface is used when the input/output is unpadded.
// This interface is used when the input/output is unpadded.
#ifdef PADDLE_WITH_HIP
PADDLE_ENFORCE_CUDA_SUCCESS
(
platform
::
dynload
::
miopenRNNBackwardData
(
handle
,
rnn
.
rnn_desc
(),
seq_length
,
rnn
.
y_descs
(),
out_data
,
rnn
.
y_descs
(),
out_grad_data
,
rnn
.
last_h_desc
(),
last_h_grad_data
,
rnn
.
last_c_desc
(),
last_c_grad_data
,
rnn
.
weight_desc
(),
weight_data
,
rnn
.
init_h_desc
(),
init_h_data
,
rnn
.
init_c_desc
(),
init_c_data
,
rnn
.
x_descs
(),
in_grad_data
,
rnn
.
init_h_desc
(),
init_h_grad_data
,
rnn
.
init_c_desc
(),
init_c_grad_data
,
workspace_data_
.
data
<
uint8_t
>
(),
workspace_size
,
const_cast
<
uint8_t
*>
(
reserve_data
),
reserve_size
));
PADDLE_ENFORCE_CUDA_SUCCESS
(
platform
::
dynload
::
miopenRNNBackwardWeights
(
handle
,
rnn
.
rnn_desc
(),
seq_length
,
rnn
.
x_descs
(),
input
->
data
<
T
>
(),
rnn
.
init_h_desc
(),
init_h
->
data
<
T
>
(),
rnn
.
y_descs
(),
out
->
data
<
T
>
(),
rnn
.
weight_desc
(),
weight_grad_data
,
workspace_data_
.
data
<
uint8_t
>
(),
workspace_size
,
const_cast
<
uint8_t
*>
(
reserve_data
),
reserve_size
));
#else
PADDLE_ENFORCE_CUDA_SUCCESS
(
platform
::
dynload
::
cudnnRNNBackwardData
(
handle
,
rnn
.
rnn_desc
(),
seq_length
,
rnn
.
y_descs
(),
out_data
,
rnn
.
y_descs
(),
out_grad_data
,
rnn
.
last_h_desc
(),
last_h_grad_data
,
...
...
@@ -418,8 +462,9 @@ class CudnnLSTMGPUGradKernel : public framework::OpKernel<T> {
rnn
.
init_h_desc
(),
init_h
->
data
<
T
>
(),
rnn
.
y_descs
(),
out
->
data
<
T
>
(),
workspace_data_
.
data
<
uint8_t
>
(),
workspace_size
,
rnn
.
weight_desc
(),
weight_grad_data
,
const_cast
<
uint8_t
*>
(
reserve_data
),
reserve_size
));
#endif
}
else
{
#if CUDNN_VERSION >= 7201
#if
!defined(PADDLE_WITH_HIP) &&
CUDNN_VERSION >= 7201
// for train
// This interface is used when the input/output is padded.
PADDLE_ENFORCE_CUDA_SUCCESS
(
platform
::
dynload
::
cudnnRNNBackwardDataEx
(
...
...
@@ -452,7 +497,13 @@ class CudnnLSTMGPUGradKernel : public framework::OpKernel<T> {
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
#ifdef PADDLE_WITH_HIP
// MIOPEN do not support double
REGISTER_OP_CUDA_KERNEL
(
cudnn_lstm
,
ops
::
CudnnLSTMGPUKernel
<
float
>
);
REGISTER_OP_CUDA_KERNEL
(
cudnn_lstm_grad
,
ops
::
CudnnLSTMGPUGradKernel
<
float
>
);
#else
REGISTER_OP_CUDA_KERNEL
(
cudnn_lstm
,
ops
::
CudnnLSTMGPUKernel
<
float
>
,
ops
::
CudnnLSTMGPUKernel
<
double
>
);
REGISTER_OP_CUDA_KERNEL
(
cudnn_lstm_grad
,
ops
::
CudnnLSTMGPUGradKernel
<
float
>
,
ops
::
CudnnLSTMGPUGradKernel
<
double
>
);
#endif
paddle/fluid/operators/cumsum_op.cu
浏览文件 @
3b9db171
...
...
@@ -16,7 +16,13 @@ limitations under the License. */
#include <thrust/device_vector.h>
#include <thrust/reverse.h>
#include <thrust/scan.h>
#include "cub/cub.cuh"
#ifdef __NVCC__
#include <cub/cub.cuh>
#endif
#ifdef __HIPCC__
#include <hipcub/hipcub.hpp>
namespace
cub
=
hipcub
;
#endif
#include "paddle/fluid/operators/cum_op.h"
#include "paddle/fluid/platform/gpu_launch_config.h"
...
...
paddle/fluid/operators/data_norm_op.cu
浏览文件 @
3b9db171
...
...
@@ -17,7 +17,7 @@ limitations under the License. */
#include "paddle/fluid/framework/data_layout.h"
#include "paddle/fluid/operators/data_norm_op.h"
#include "paddle/fluid/platform/cuda_primitives.h"
#if defined(PADDLE_WITH_NCCL)
#if defined(PADDLE_WITH_NCCL)
|| defined(PADDLE_WITH_RCCL)
#include "paddle/fluid/platform/collective_helper.h"
#include "paddle/fluid/platform/nccl_helper.h"
#endif
...
...
@@ -174,7 +174,7 @@ class DataNormGradKernel<platform::CUDADeviceContext, T>
d_batch_sum
,
d_batch_square_sum
);
if
(
need_sync_stats
)
{
#if defined(PADDLE_WITH_NCCL)
#if defined(PADDLE_WITH_NCCL)
|| defined(PADDLE_WITH_RCCL)
auto
comm
=
platform
::
NCCLCommContext
::
Instance
().
Get
(
0
,
ctx
.
GetPlace
());
PADDLE_ENFORCE_CUDA_SUCCESS
(
platform
::
dynload
::
ncclAllReduce
(
reinterpret_cast
<
const
void
*>
(
d_batch_size
),
...
...
@@ -188,7 +188,11 @@ class DataNormGradKernel<platform::CUDADeviceContext, T>
reinterpret_cast
<
const
void
*>
(
d_batch_square_sum
),
reinterpret_cast
<
void
*>
(
d_batch_square_sum
),
C
,
platform
::
ToNCCLDataType
(
x
->
type
()),
ncclSum
,
comm
->
comm
(),
stream
));
#ifdef PADDLE_WITH_RCCL
PADDLE_ENFORCE_CUDA_SUCCESS
(
hipStreamSynchronize
(
stream
));
#else
PADDLE_ENFORCE_CUDA_SUCCESS
(
cudaStreamSynchronize
(
stream
));
#endif
#else
PADDLE_THROW
(
platform
::
errors
::
PreconditionNotMet
(
"PaddlePaddle should compile with GPU, and need_sync_stats connot be "
...
...
paddle/fluid/operators/diag_embed_op.h
浏览文件 @
3b9db171
...
...
@@ -100,7 +100,7 @@ class DiagEmbedKernel : public framework::OpKernel<T> {
strides
.
push_back
(
stride
[
dim1_
]
+
stride
[
dim2_
]);
const
auto
dims
=
vectorize
(
input
->
dims
());
#if
def __NVCC__
#if
defined(__NVCC__) || defined(__HIPCC__)
thrust
::
device_vector
<
int64_t
>
dims_vec
(
dims
);
const
int64_t
*
dims_arr
=
thrust
::
raw_pointer_cast
(
dims_vec
.
data
());
thrust
::
device_vector
<
int64_t
>
strides_vec
(
strides
);
...
...
paddle/fluid/operators/dot_op.h
浏览文件 @
3b9db171
...
...
@@ -45,7 +45,7 @@ struct DotGradFunction<DeviceContext, T, math::EnableComplex<T>> {
const
Tensor
*
tensor_dout
,
Tensor
*
tensor_dx
,
Tensor
*
tensor_dy
,
const
paddle
::
framework
::
ExecutionContext
&
ctx
)
{
#if
def __NVCC__
#if
defined(__NVCC__) || defined(__HIPCC__)
if
(
1
==
tensor_dout
->
dims
().
size
())
{
auto
dout
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
tensor_dout
);
...
...
@@ -249,7 +249,7 @@ class DotKernel : public framework::OpKernel<T> {
auto
*
tensor_out
=
ctx
.
Output
<
Tensor
>
(
"Out"
);
tensor_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
#if
def __NVCC__
#if
defined(__NVCC__) || defined(__HIPCC__)
if
(
1
==
tensor_out
->
dims
().
size
())
{
auto
out
=
framework
::
EigenScalar
<
T
>::
From
(
*
tensor_out
);
auto
x
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
tensor_x
);
...
...
paddle/fluid/operators/dropout_op.cu
浏览文件 @
3b9db171
...
...
@@ -11,8 +11,17 @@ 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_CUDA
#include <cuda.h>
#include <curand_kernel.h>
#include "paddle/fluid/platform/dynload/curand.h"
#endif
#ifdef PADDLE_WITH_HIP
#include <hip/hip_runtime.h>
#include <hiprand_kernel.h>
#include "paddle/fluid/platform/dynload/hiprand.h"
#endif
#include <thrust/device_ptr.h>
#include <thrust/iterator/counting_iterator.h>
#include <thrust/random.h>
...
...
@@ -21,7 +30,6 @@ limitations under the License. */
#include <string>
#include "paddle/fluid/memory/memcpy.h"
#include "paddle/fluid/operators/dropout_op.h"
#include "paddle/fluid/platform/dynload/curand.h"
#include "paddle/fluid/platform/float16.h"
namespace
paddle
{
...
...
@@ -32,15 +40,24 @@ __global__ void RandomGenerator(const size_t n, uint64_t seed,
const
float
dropout_prob
,
const
T
*
src
,
MaskType
*
mask_data
,
T
*
dst
,
bool
is_upscale_in_train
,
uint64_t
increment
)
{
curandStatePhilox4_32_10_t
state
;
int
idx
=
blockDim
.
x
*
blockIdx
.
x
+
threadIdx
.
x
;
#ifdef PADDLE_WITH_HIP
hiprandStatePhilox4_32_10_t
state
;
hiprand_init
(
seed
,
idx
,
increment
,
&
state
);
#else
curandStatePhilox4_32_10_t
state
;
curand_init
(
seed
,
idx
,
increment
,
&
state
);
#endif
MaskType
mask
;
T
dest
;
for
(;
idx
<
n
;
idx
+=
blockDim
.
x
*
gridDim
.
x
)
{
T
s
=
src
[
idx
];
#ifdef PADDLE_WITH_HIP
if
(
hiprand_uniform
(
&
state
)
<
dropout_prob
)
{
#else
if
(
curand_uniform
(
&
state
)
<
dropout_prob
)
{
#endif
mask
=
0
;
dest
=
0
;
}
else
{
...
...
@@ -62,9 +79,15 @@ __global__ void VectorizedRandomGenerator(const size_t n, uint64_t seed,
const
T
*
src
,
MaskType
*
mask_data
,
T
*
dst
,
bool
is_upscale_in_train
,
uint64_t
increment
)
{
#ifdef PADDLE_WITH_HIP
int64_t
idx
=
hipBlockDim_x
*
hipBlockIdx_x
+
hipThreadIdx_x
;
hiprandStatePhilox4_32_10_t
state
;
hiprand_init
(
seed
,
idx
,
increment
,
&
state
);
#else
int64_t
idx
=
blockDim
.
x
*
blockIdx
.
x
+
threadIdx
.
x
;
curandStatePhilox4_32_10_t
state
;
curand_init
(
seed
,
idx
,
increment
,
&
state
);
#endif
MaskType
mask
;
T
dest
;
...
...
@@ -75,7 +98,11 @@ __global__ void VectorizedRandomGenerator(const size_t n, uint64_t seed,
T
src_vec
[
VecSize
];
LoadT
*
value
=
reinterpret_cast
<
LoadT
*>
(
&
src_vec
);
*
value
=
*
reinterpret_cast
<
const
LoadT
*>
(
&
src
[
i
]);
#ifdef PADDLE_WITH_HIP
float4
rand
=
hiprand_uniform4
(
&
state
);
#else
float4
rand
=
curand_uniform4
(
&
state
);
#endif
T
dest_vec
[
VecSize
];
MaskType
mask_vec
[
VecSize
];
...
...
@@ -131,10 +158,17 @@ class GPUDropoutKernel : public framework::OpKernel<T> {
auto
*
x_data
=
x
->
data
<
T
>
();
auto
*
y_data
=
y
->
mutable_data
<
T
>
(
context
.
GetPlace
());
if
(
dropout_prob
==
1.0
f
)
{
#ifdef PADDLE_WITH_HIP
PADDLE_ENFORCE_CUDA_SUCCESS
(
hipMemsetAsync
(
y_data
,
0
,
x_numel
*
sizeof
(
T
),
stream
));
PADDLE_ENFORCE_CUDA_SUCCESS
(
hipMemsetAsync
(
mask_data
,
0
,
x_numel
*
sizeof
(
*
mask_data
),
stream
));
#else
PADDLE_ENFORCE_CUDA_SUCCESS
(
cudaMemsetAsync
(
y_data
,
0
,
x_numel
*
sizeof
(
T
),
stream
));
PADDLE_ENFORCE_CUDA_SUCCESS
(
cudaMemsetAsync
(
mask_data
,
0
,
x_numel
*
sizeof
(
*
mask_data
),
stream
));
#endif
return
;
}
...
...
@@ -180,6 +214,20 @@ class GPUDropoutKernel : public framework::OpKernel<T> {
increment
=
offset
;
}
#ifdef __HIPCC__
if
(
vec_size
==
4
&&
size
%
4
==
0
)
{
hipLaunchKernelGGL
(
HIP_KERNEL_NAME
(
VectorizedRandomGenerator
<
T
,
uint8_t
,
4
>
),
config
.
block_per_grid
,
config
.
thread_per_block
,
0
,
stream
,
size
,
seed_data
,
dropout_prob
,
x_data
,
mask_data
,
y_data
,
upscale_in_train
,
increment
);
}
else
{
hipLaunchKernelGGL
(
HIP_KERNEL_NAME
(
RandomGenerator
<
T
,
uint8_t
>
),
config
.
block_per_grid
,
config
.
thread_per_block
,
0
,
stream
,
size
,
seed_data
,
dropout_prob
,
x_data
,
mask_data
,
y_data
,
upscale_in_train
,
increment
);
}
#else
if
(
vec_size
==
4
&&
size
%
4
==
0
)
{
VectorizedRandomGenerator
<
T
,
uint8_t
,
...
...
@@ -192,7 +240,7 @@ class GPUDropoutKernel : public framework::OpKernel<T> {
size
,
seed_data
,
dropout_prob
,
x_data
,
mask_data
,
y_data
,
upscale_in_train
,
increment
);
}
#endif
}
else
{
auto
X
=
EigenMatrix
<
T
>::
Reshape
(
*
x
,
1
);
auto
Y
=
EigenMatrix
<
T
>::
Reshape
(
*
y
,
1
);
...
...
paddle/fluid/operators/dropout_op.h
浏览文件 @
3b9db171
...
...
@@ -42,7 +42,7 @@ inline int VectorizedSize(const T* pointer) {
return
1
;
}
#if
def __NVCC__
#if
defined(__NVCC__) || defined(__HIPCC__)
template
<
typename
T
,
typename
MaskType
,
int
VecSize
>
__global__
void
DropoutGradCUDAKernel
(
const
T
*
dout
,
const
MaskType
*
mask
,
const
T
factor
,
const
int64_t
size
,
...
...
@@ -186,7 +186,7 @@ class DropoutGradKernel : public framework::OpKernel<T> {
int
vec_size
=
VectorizedSize
<
T
>
(
grad_y
->
data
<
T
>
());
if
(
platform
::
is_gpu_place
(
context
.
GetPlace
())
&&
vec_size
==
4
&&
size
%
4
==
0
)
{
#if
def __NVCC__
#if
defined(__NVCC__) || defined(__HIPCC__)
auto
factor
=
static_cast
<
T
>
(
1.0
f
/
(
1.0
f
-
dropout_prob
));
auto
stream
=
context
.
cuda_device_context
().
stream
();
platform
::
GpuLaunchConfig
config
=
platform
::
GetGpuLaunchConfig1D
(
...
...
paddle/fluid/operators/fake_quantize_op.cu
浏览文件 @
3b9db171
...
...
@@ -162,7 +162,11 @@ struct FindChannelAbsMaxFunctor<platform::CUDADeviceContext, T> {
int
grid
=
cout
;
int
max_threads
=
1024
;
#ifdef PADDLE_WITH_HIP
hipMemset
(
out_abs_max
,
0
,
sizeof
(
T
)
*
cout
);
#else
cudaMemset
(
out_abs_max
,
0
,
sizeof
(
T
)
*
cout
);
#endif
for
(
int
i
=
0
;
i
<
cin
/
max_threads
;
i
++
)
{
int
block
=
max_threads
;
...
...
paddle/fluid/operators/fill_constant_batch_size_like_op.h
浏览文件 @
3b9db171
...
...
@@ -65,7 +65,7 @@ class FillConstantBatchSizeLikeOpKernel : public framework::OpKernel<T> {
functor
(
reinterpret_cast
<
const
platform
::
CPUDeviceContext
&>
(
dev_ctx
),
out
,
static_cast
<
T
>
(
value
));
}
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
if
(
!
cpu_place
)
{
math
::
SetConstant
<
platform
::
CUDADeviceContext
,
T
>
functor
;
out
->
mutable_data
(
ctx
.
GetPlace
(),
data_type
);
...
...
paddle/fluid/operators/fill_constant_op.h
浏览文件 @
3b9db171
...
...
@@ -121,7 +121,7 @@ class FillConstantKernel : public framework::OpKernel<T> {
functor
(
reinterpret_cast
<
const
platform
::
CPUDeviceContext
&>
(
dev_ctx
),
tensor
,
static_cast
<
T
>
(
value
));
}
else
if
(
actual_place
==
1
)
{
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
tensor
->
mutable_data
(
ctx
.
GetPlace
(),
data_type
);
math
::
SetConstant
<
platform
::
CUDADeviceContext
,
T
>
functor
;
functor
(
reinterpret_cast
<
const
platform
::
CUDADeviceContext
&>
(
dev_ctx
),
...
...
@@ -131,7 +131,7 @@ class FillConstantKernel : public framework::OpKernel<T> {
"PaddlePaddle should compile with GPU."
));
#endif
}
else
if
(
actual_place
==
2
)
{
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
tensor
->
mutable_data
(
platform
::
CUDAPinnedPlace
(),
data_type
);
math
::
SetConstant
<
platform
::
CPUDeviceContext
,
T
>
functor
;
functor
(
reinterpret_cast
<
const
platform
::
CPUDeviceContext
&>
(
dev_ctx
),
...
...
paddle/fluid/operators/filter_by_instag_op.h
浏览文件 @
3b9db171
...
...
@@ -31,7 +31,7 @@ namespace operators {
using
Tensor
=
framework
::
Tensor
;
using
SelectedRows
=
framework
::
SelectedRows
;
using
LoDTensor
=
framework
::
LoDTensor
;
#if defined(PADDLE_WITH_CUDA)
#if defined(PADDLE_WITH_CUDA)
|| defined(PADDLE_WITH_HIP)
template
<
typename
T
>
using
Vector
=
framework
::
Vector
<
T
>
;
#else
...
...
paddle/fluid/operators/gelu_op.h
浏览文件 @
3b9db171
...
...
@@ -54,7 +54,8 @@ struct GeluFunctor {
}
}
else
{
#if defined(PADDLE_WITH_MKLML) && !defined(_WIN32) && !defined(__APPLE__) && \
!defined(__OSX__) && !defined(PADDLE_WITH_CUDA)
!defined(__OSX__) && !defined(PADDLE_WITH_CUDA) && \
!defined(PADDLE_WITH_HIP)
auto
x_data
=
x
.
data
();
auto
out_data
=
out
.
data
();
int
n
=
std
::
min
(
x
.
size
(),
out
.
size
());
...
...
@@ -121,7 +122,8 @@ struct GeluGradFunctor {
}
}
else
{
#if defined(PADDLE_WITH_MKLML) && !defined(_WIN32) && !defined(__APPLE__) && \
!defined(__OSX__) && !defined(PADDLE_WITH_CUDA)
!defined(__OSX__) && !defined(PADDLE_WITH_CUDA) && \
!defined(PADDLE_WITH_HIP)
auto
x_data
=
x
.
data
();
auto
dx_data
=
dx
.
data
();
auto
dout_data
=
dout
.
data
();
...
...
paddle/fluid/operators/get_tensor_from_selected_rows_op.cc
浏览文件 @
3b9db171
...
...
@@ -107,7 +107,7 @@ REGISTER_OP_CPU_KERNEL_FUNCTOR(get_tensor_from_selected_rows, float,
ops
::
GetTensorFromSelectedRowsKernel
,
int64_t
,
ops
::
GetTensorFromSelectedRowsKernel
);
#if
def PADDLE_WITH_CUDA
#if
defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
REGISTER_OP_CUDA_KERNEL_FUNCTOR
(
get_tensor_from_selected_rows
,
float
,
ops
::
GetTensorFromSelectedRowsKernel
,
double
,
ops
::
GetTensorFromSelectedRowsKernel
,
int
,
...
...
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