Update comment

intel · Jul 22, 2024 · 177872c · 177872c
1 parent 215d1b5
commit 177872c
Showing 1 changed file with 3 additions and 2 deletions.
diff --git a/src/ATen/native/xpu/sycl/SortingKernels.h b/src/ATen/native/xpu/sycl/SortingKernels.h
@@ -421,8 +421,9 @@ void segmented_sort_pairs_(
     int num_elements) {
   constexpr int scaling_coef = sizeof(key_t) * sizeof(value_t) >= 64
       ? 2
-      : 1; // Attempt to reduce register pressure. For some reason, register
-           // spilling causes some incorrect results.
+      : 1; // Attempt to reduce register pressure. Excessive register usage can
+           // cause calculation errors.
+           // https://github.com/intel/torch-xpu-ops/issues/626
   if (num_elements > 4096 / scaling_coef) {
     // Considering register pressure, we use a problem size of 4096 to delineate
     // the boundary between single tile sort and group sort.