Change of post shuffle thread descriptor

d3f2dbbd · rocking · 5490de18 · d3f2dbbd
Commit d3f2dbbd authored Dec 26, 2022 by rocking
Show whitespace changes
Inline Side-by-side

Showing with 2 additions and 2 deletions

include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp ...dwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp +2 -2

No files found.
--- a/include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp
@@ -755,7 +755,7 @@ struct GridwiseGemmMultipleDWelfordFirstHalf_xdl_cshuffle
            // To apply D0, D1, ... and Welford.
            // threadwise copy from LDS to VGPR
            constexpr auto post_shuffle_thread_cluster_desc =
-                make_cluster_descriptor(PostShuffleThreadClusterSize_M_N{}, Sequence<1, 0>{});
+                make_cluster_descriptor(PostShuffleThreadClusterSize_M_N{}, Sequence<0, 1>{});

            const auto post_shuffle_thread_cluster_idx =
                post_shuffle_thread_cluster_desc.CalculateBottomIndex(
@@ -861,7 +861,7 @@ struct GridwiseGemmMultipleDWelfordFirstHalf_xdl_cshuffle
            using BlockwiseWelford = BlockwiseWelford<AccDataType,
                                                      BlockSize,
                                                      PostShuffleThreadClusterSize_M_N,
-                                                      Sequence<1, 0>,
+                                                      Sequence<0, 1>,
                                                      false>;

            constexpr int num_shuffleM =