Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Fix a bug when reduce_num = 1 in Reduce Op #38771

Merged
merged 3 commits into from
Jan 7, 2022
Merged

Fix a bug when reduce_num = 1 in Reduce Op #38771

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
0bbcc06
Replace Copy and Cast with Elementwise
AnnaTrainingG Jan 6, 2022
f76a0c3
delete cast_kernel.h and copy_kernel.h
AnnaTrainingG Jan 7, 2022
bf1a2e8
update
AnnaTrainingG Jan 7, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
31 changes: 5 additions & 26 deletions paddle/pten/kernels/gpu/reduce.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,8 +45,7 @@ namespace cub = hipcub;
#include "paddle/pten/api/ext/dispatch.h"
#include "paddle/pten/backends/gpu/gpu_context.h"
#include "paddle/pten/core/dense_tensor.h"
#include "paddle/pten/kernels/cast_kernel.h"
#include "paddle/pten/kernels/copy_kernel.h"
#include "paddle/pten/kernels/gpu/elementwise.h"

// Reduce split or not, Whether to use ReduceHigherDim
#define REDUCE_SPLIT_BOUNDARY 512
Expand Down Expand Up @@ -1062,23 +1061,6 @@ static
"Tx should not be float16 when using cub::DeviceReduce::Reduce()."));
}

static void AsyncCopy(const pten::DenseTensor& src, pten::DenseTensor* dst) {
paddle::platform::DeviceContextPool& pool =
paddle::platform::DeviceContextPool::Instance();
const paddle::platform::CUDADeviceContext* dev_ctx;
if (paddle::platform::is_gpu_place(dst->place()) ||
paddle::platform::is_npu_place(dst->place())) {
dev_ctx = static_cast<paddle::platform::CUDADeviceContext*>(
pool.Get(dst->place()));

} else {
dev_ctx = static_cast<paddle::platform::CUDADeviceContext*>(
pool.Get(src.place()));
}

pten::Copy(*dev_ctx, src, false, dst);
}

template <typename Tx,
typename Ty,
template <typename> class ReduceOp,
Expand Down Expand Up @@ -1111,13 +1093,10 @@ void TensorReduceFunctorImpl(const pten::DenseTensor& x,
auto* dev_ctx = static_cast<paddle::platform::CUDADeviceContext*>(
paddle::platform::DeviceContextPool::Instance().Get(x.place()));
if (config.reduce_num == 1) {
auto out_dims = y->dims();
if (x.dtype() == y->dtype()) {
AsyncCopy(x, y);
y->Resize(out_dims);
} else {
pten::CastKernel<Tx>(*dev_ctx, x, y->dtype(), y);
}
std::vector<const DenseTensor*> inputs = {&x};
std::vector<DenseTensor*> outputs = {y};
pten::LaunchSameDimsElementwiseCudaKernel<ElementwiseType::kUnary, Tx, Ty>(
*dev_ctx, inputs, &outputs, transform);
return;
}

Expand Down