Skip to content

Simplify copy kernel #28428

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

Closed
zasdfgbnm wants to merge 9 commits into from
Closed

Simplify copy kernel #28428

Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
96ff5ff
Simplify copy kernel
zasdfgbnm Oct 22, 2019
0cf634c
Update on "Simplify copy kernel"
zasdfgbnm Oct 22, 2019
e84ac97
Update on "Simplify copy kernel"
zasdfgbnm Oct 23, 2019
5fc3a07
Update on "Simplify copy kernel"
zasdfgbnm Oct 24, 2019
1a1c3f8
Update on "Simplify copy kernel"
zasdfgbnm Oct 24, 2019
44b398e
Update on "Simplify copy kernel"
zasdfgbnm Oct 25, 2019
dd612d0
Update on "Simplify copy kernel"
zasdfgbnm Oct 25, 2019
25e7b33
Update on "Simplify copy kernel"
zasdfgbnm Oct 26, 2019
27ecdf9
Update on "Simplify copy kernel"
zasdfgbnm Oct 26, 2019
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
18 changes: 5 additions & 13 deletions aten/src/ATen/native/cuda/Copy.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,20 +8,12 @@
#include <ATen/native/TensorIterator.h>
#include <ATen/native/cuda/Loops.cuh>
#include <THC/THC.h>
#include <c10/util/TypeCast.h>

namespace at {
namespace native {

using namespace at::cuda;

template <typename dst_t, typename src_t>
void copy_kernel_impl(TensorIterator& iter) {
gpu_kernel(iter, []GPU_LAMBDA(src_t x) -> dst_t {
return c10::static_cast_with_inter_type<dst_t>(x);
});
}

// device-to-device copy, does type conversion
static void copy_device_to_device(TensorIterator& iter, bool non_blocking) {
int64_t numel = iter.numel();
Expand Down Expand Up @@ -66,11 +58,11 @@ static void copy_device_to_device(TensorIterator& iter, bool non_blocking) {
cudaMemcpyDeviceToDevice,
copy_stream));
} else {
AT_DISPATCH_ALL_TYPES_AND2(kHalf, kBool, iter.dtype(0), "copy_", [&] {
using dst_t = scalar_t;
AT_DISPATCH_ALL_TYPES_AND2(kHalf, kBool, iter.dtype(1), "copy_", [&] {
copy_kernel_impl<dst_t, scalar_t>(iter);
});
// this is done intentionally done after build because copy has a "promotion"
// rule that always "promote" to target dtype.
iter.promote_common_dtype();
AT_DISPATCH_ALL_TYPES_AND3(kHalf, kBool, kBFloat16, iter.dtype(0), "copy_", [&] {
Copy link
Collaborator Author

@zasdfgbnm zasdfgbnm Oct 26, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I see a PR trying to enable BFloat16 for CUDA:
https://github.com/pytorch/pytorch/pull/27259/files#diff-6684cb81a1865b7d52d9f2f1789cd0ceR68-R70
So I include it in this PR also, so that I can benchmark

gpu_kernel(iter, []GPU_LAMBDA(scalar_t x) { return x; });
});
}

Expand Down