Do not use snakeCurved, it makes determination of padding for welford difficult

982c85a3 · rocking · 8749678a · 982c85a3
Commit 982c85a3 authored Nov 24, 2022 by rocking
Hide whitespace changes
Inline Side-by-side

Showing with 4 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 +4 -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
@@ -698,7 +698,8 @@ struct GridwiseGemmMultipleDWelfordFirstHalf_xdl_cshuffle
                                           M2,
                                           1,
                                           M4,
-                                           1>>{};
+                                           1>,
+                                  false>{};

            // space filling curve for shuffled blockwise C in global mem
            constexpr auto sfc_der_global =
@@ -707,7 +708,8 @@ struct GridwiseGemmMultipleDWelfordFirstHalf_xdl_cshuffle
                                  Sequence<1,
                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
                                           1,
-                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>,
+                                  false>{};

            // LDS c_reduce_block_desc_mperblock_nperblock
            constexpr auto c_reduce_block_desc_mperblock_nperblock = transform_tensor_descriptor(