Skip to content

Add subgroup topk kernel for XPU (part1 of #3369) #3371

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

Merged
CuiYifeng merged 9 commits into from
Apr 29, 2026
+626 −11
Merged

Add subgroup topk kernel for XPU (part1 of #3369) #3371

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
faebb77
Add subgroup topk kernel for XPU
jianyizh Apr 17, 2026
c970ea5
Address review: kernel properties, integer safety, comment cleanup
jianyizh Apr 17, 2026
d4daf78
Address review: use IndexT for element indices, add bitonic sort comm…
jianyizh Apr 17, 2026
8c3f054
Remove benchmark source details from dispatch comment
jianyizh Apr 18, 2026
6aa63c9
Fix sentinel-value insert bug, add alignas and pointer alignment check
jianyizh Apr 18, 2026
ad407ac
Lower subgroup topk dispatch threshold from dim>=1024 to dim>=32
jianyizh Apr 20, 2026
3e536e0
Dispatch on smallest power-of-two K to reduce register pressure
jianyizh Apr 21, 2026
bde1bdb
Use overflow-safe check for IndexT dispatch instead of nsegments*nele…
jianyizh Apr 26, 2026
3a7de35
Address review: use PowerOf2Ceil, sycl_kernel_submit, simplify dispatch
jianyizh Apr 27, 2026
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: 20 additions & 11 deletions src/ATen/native/xpu/sycl/TensorTopKKernel.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,6 +16,7 @@
#include <ATen/native/xpu/sycl/SortingKernels.h>

#include <ATen/native/xpu/sycl/TensorTopKKernel.h>
#include <ATen/native/xpu/sycl/TensorTopKSbtopkKernel.h>

namespace at {
namespace native {
Expand Down Expand Up @@ -113,17 +114,25 @@ void topk_kernel(
const scalar_t* self_ptr = self_.const_data_ptr<scalar_t>();
scalar_t* values_ptr = values_.data_ptr<scalar_t>();
int64_t* indices_ptr = indices_.data_ptr<int64_t>();
segmented_group_select_pairs<scalar_t, int64_t>(
self_ptr,
values_ptr,
nullptr,
(int64_t*)indices_ptr,
nsegments,
nelements,
k,
largest);

if (sorted) {

SbtopkResult sbtopk_result = sbtopk_try_launch(
self_, nsegments, nelements, k, largest, values_, indices_);
if (sbtopk_result == SbtopkResult::FAILED) {
Comment thread
jianyizh marked this conversation as resolved.
Comment thread
jianyizh marked this conversation as resolved.
segmented_group_select_pairs<scalar_t, int64_t>(
self_ptr,
values_ptr,
nullptr,
(int64_t*)indices_ptr,
nsegments,
nelements,
k,
largest);
Comment thread
jianyizh marked this conversation as resolved.
}

// Only sort if the user asked for sorted output AND the optimized
// kernel did not already produce a sorted result. The subgroup topk
// kernel returns SORTED; the original kernel returns FAILED.
if (sorted && sbtopk_result != SbtopkResult::SORTED) {
segmented_sort_pairs<scalar_t, int64_t>(
values_ptr,
values_ptr,
Expand Down
Loading
Loading