Fix typo in row_reduce_small

This commit is contained in:
Cheng 2025-05-13 19:48:11 +09:00
parent 659a51919f
commit fe8f4c8cf7

View File

@ -224,7 +224,7 @@ template <
if ((non_row_reductions < 32 && row_size <= 8) || non_row_reductions <= 8) {
// Simple loop over non_row_reductions and reduce the row in the thread.
IdxT out_idx = tid.x + tsize.y * IdxT(tid.y);
IdxT out_idx = tid.x + tsize.x * IdxT(tid.y);
in += elem_to_loc<IdxT>(out_idx, shape, strides, ndim);
for (uint r = 0; r < non_row_reductions; r++) {