Merge branch 'develop' into kylasa_1870

include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v3.hpp

@@ -76,6 +76,8 @@ struct GemmPipelineAgBgCrCompV3 : public BaseGemmPipelineAgBgCrCompV3<Problem>
     static constexpr bool kPadN = Problem::kPadN;
     static constexpr bool kPadK = Problem::kPadK;
 
+    static constexpr bool DoubleSmemBuffer = Problem::DoubleSmemBuffer;
+
     static constexpr bool HasHotLoop = Problem::HasHotLoop;
     static constexpr auto TailNum    = Problem::TailNum;
     static constexpr auto Scheduler  = Problem::Scheduler;

include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v4.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+#pragma once
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
+#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_base.hpp"
+#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v4_default_policy.hpp"
+
+namespace ck_tile {
+
+//  A Tile Window: global memory
+//  B Tile Window: global memory
+//  C Distributed tensor: register
+template <typename Problem>
+struct BaseGemmPipelineAgBgCrCompV4
+{
+    static constexpr index_t PrefetchStages  = 2;
+    static constexpr index_t PrefillStages   = 1;
+    static constexpr index_t GlobalBufferNum = 1;
+
+    CK_TILE_HOST static constexpr bool BlockHasHotloop(index_t num_loop)
+    {
+        return num_loop > PrefetchStages;
+    }
+
+    CK_TILE_HOST static constexpr TailNumber GetBlockLoopTailNum(index_t num_loop)
+    {
+        if(num_loop % PrefetchStages == 1)
+        {
+            return TailNumber::Three;
+        }
+        else
+        {
+            return TailNumber::Two;
+        }
+    }
+};
+
+/**
+ * @brief Compute optimized pipeline version 4
+ *
+ * This version introduces a dual LDS window mechanism using a ping-pong buffer approach
+ * for more efficient data handling from global memory. Unlike compute version 3, this method
+ * allows one LDS to fetch data from global memory while the other LDS executes warps for MFMA
+ * matrix multiplication. This dual operation helps in keeping the Warp unit continuously busy,
+ * thereby significantly reducing memory load times and enhancing overall performance.
+ *
+ * @note This version shows improved performance over Compute Version 3 with the same block tile.
+ * It is particularly more efficient for large matrices where M, N, and K are greater than 8K,
+ * even when Compute Version 3's block size is twice that of Compute Version 4.
+ */
+template <typename Problem, typename Policy = GemmPipelineAgBgCrCompV4DefaultPolicy>
+struct GemmPipelineAgBgCrCompV4 : public BaseGemmPipelineAgBgCrCompV4<Problem>
+{
+    using Base             = BaseGemmPipelineAgBgCrCompV4<Problem>;
+    using PipelineImplBase = GemmPipelineAgBgCrImplBase<Problem, Policy>;
+
+    using ADataType      = remove_cvref_t<typename Problem::ADataType>;
+    using BDataType      = remove_cvref_t<typename Problem::BDataType>;
+    using CDataType      = remove_cvref_t<typename Problem::CDataType>;
+    using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>;
+
+    using ALayout = remove_cvref_t<typename Problem::ALayout>;
+    using BLayout = remove_cvref_t<typename Problem::BLayout>;
+    using CLayout = remove_cvref_t<typename Problem::CLayout>;
+
+    using BlockGemm = remove_cvref_t<decltype(Policy::template GetBlockGemm<Problem>())>;
+    using I0        = number<0>;
+    using I1        = number<1>;
+    using I2        = number<2>;
+
+    static constexpr index_t BlockSize = Problem::kBlockSize;
+
+    static constexpr index_t MPerBlock = BlockGemmShape::kM;
+    static constexpr index_t NPerBlock = BlockGemmShape::kN;
+    static constexpr index_t KPerBlock = BlockGemmShape::kK;
+
+    static constexpr index_t GetVectorSizeA() { return Policy::template GetVectorSizeA<Problem>(); }
+    static constexpr index_t GetVectorSizeB() { return Policy::template GetVectorSizeB<Problem>(); }
+    static constexpr index_t GetVectorSizeC() { return Policy::template GetVectorSizeC<Problem>(); }
+
+    static constexpr bool kPadM = Problem::kPadM;
+    static constexpr bool kPadN = Problem::kPadN;
+    static constexpr bool kPadK = Problem::kPadK;
+
+    static constexpr bool DoubleSmemBuffer = Problem::DoubleSmemBuffer;
+
+    static constexpr bool HasHotLoop = Problem::HasHotLoop;
+    static constexpr auto TailNum    = Problem::TailNum;
+    static constexpr auto Scheduler  = Problem::Scheduler;
+
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
+    {
+        return Policy::template GetSmemSize<Problem>();
+    }
+
+    CK_TILE_HOST_DEVICE static constexpr auto IsTransposeC()
+    {
+        return Policy::template IsTransposeC<Problem>();
+    }
+
+    template <GemmPipelineScheduler Scheduler>
+    struct PipelineImpl : public PipelineImplBase
+    {
+    };
+
+    template <>
+    struct PipelineImpl<GemmPipelineScheduler::Intrawave> : public PipelineImplBase
+    {
+        using Base = PipelineImplBase;
+
+        CK_TILE_DEVICE static constexpr auto HotLoopScheduler()
+        {
+            constexpr index_t MPerXDL = BlockGemmShape::WarpTile::at(I0{});
+            constexpr index_t NPerXDL = BlockGemmShape::WarpTile::at(I1{});
+            constexpr index_t KPerXDL = BlockGemmShape::WarpTile::at(I2{});
+
+            constexpr index_t WaveSize = 64;
+            constexpr index_t WaveNumM = BlockGemmShape::BlockWarps::at(I0{});
+            constexpr index_t WaveNumN = BlockGemmShape::BlockWarps::at(I1{});
+
+            constexpr index_t A_LDS_Read_Width = KPerXDL;
+            constexpr index_t B_LDS_Read_Width = KPerXDL;
+
+            constexpr index_t A_Buffer_Load_Inst_Num =
+                MPerBlock * KPerBlock / (BlockSize * GetVectorSizeA());
+            constexpr index_t B_Buffer_Load_Inst_Num =
+                NPerBlock * KPerBlock / (BlockSize * GetVectorSizeB());
+
+            constexpr index_t A_LDS_Write_Inst_Num = MPerBlock * KPerBlock / (BlockSize * KPerXDL);
+            constexpr index_t B_LDS_Write_Inst_Num = NPerBlock * KPerBlock / (BlockSize * KPerXDL);
+
+            constexpr index_t A_LDS_Read_Inst_Num =
+                WaveNumN * MPerBlock * KPerBlock / (BlockSize * KPerXDL);
+            constexpr index_t B_LDS_Read_Inst_Num =
+                WaveNumM * MPerBlock * KPerBlock / (BlockSize * KPerXDL);
+
+            constexpr index_t C_MFMA_Inst_Num = MPerBlock * NPerBlock * KPerBlock /
+                                                (BlockSize / WaveSize) /
+                                                (MPerXDL * NPerXDL * KPerXDL);
+
+            constexpr auto num_ds_read_inst_a = A_LDS_Read_Width * sizeof(ADataType) == 16
+                                                    ? A_LDS_Read_Inst_Num
+                                                    : A_LDS_Read_Inst_Num / 2;
+            constexpr auto num_ds_read_inst_b = B_LDS_Read_Width * sizeof(BDataType) == 16
+                                                    ? B_LDS_Read_Inst_Num
+                                                    : B_LDS_Read_Inst_Num / 2;
+
+            constexpr auto num_ds_read_inst     = num_ds_read_inst_a + num_ds_read_inst_b;
+            constexpr auto num_ds_write_inst    = A_LDS_Write_Inst_Num + B_LDS_Write_Inst_Num;
+            constexpr auto num_buffer_load_inst = A_Buffer_Load_Inst_Num + B_Buffer_Load_Inst_Num;
+            constexpr auto num_issue            = num_buffer_load_inst;
+
+            static_for<0, num_buffer_load_inst, 1>{}([&](auto i) {
+                ignore = i;
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA : 1
+                __builtin_amdgcn_sched_group_barrier(
+                    0x100, num_ds_read_inst / num_issue, 0);       // DS read : 2
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA: 1
+                __builtin_amdgcn_sched_group_barrier(
+                    0x200, num_ds_write_inst / num_issue, 0);      // DS write : 1
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA : 1
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read :1
+                __builtin_amdgcn_sched_group_barrier(
+                    0x008, C_MFMA_Inst_Num / num_issue - 3, 0); // MFMA : 5
+            });
+            __builtin_amdgcn_sched_barrier(0);
+        }
+
+        template <bool HasHotLoop,
+                  TailNumber TailNum,
+                  typename ADramBlockWindowTmp,
+                  typename BDramBlockWindowTmp,
+                  typename AElementFunction,
+                  typename BElementFunction>
+        CK_TILE_DEVICE auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
+                                       const AElementFunction& a_element_func,
+                                       const BDramBlockWindowTmp& b_dram_block_window_tmp,
+                                       const BElementFunction& b_element_func,
+                                       index_t num_loop,
+                                       void* __restrict__ p_smem_0,
+                                       void* __restrict__ p_smem_1) const
+        {
+            static_assert(
+                std::is_same_v<ADataType, remove_cvref_t<typename ADramBlockWindowTmp::DataType>> &&
+                    std::is_same_v<BDataType,
+                                   remove_cvref_t<typename BDramBlockWindowTmp::DataType>>,
+                "Data Type conflict on A and B matrix input data type.");
+
+            constexpr bool is_a_col_major =
+                std::is_same_v<ALayout, tensor_layout::gemm::ColumnMajor>;
+            constexpr bool is_b_row_major = std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>;
+
+            static_assert(is_a_col_major
+                              ? (KPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
+                                 MPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I1{}])
+                              : (MPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
+                                 KPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I1{}]),
+                          "A block window has incorrect lengths for defined ALayout!");
+            static_assert(is_b_row_major
+                              ? (KPerBlock == BDramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
+                                 NPerBlock == BDramBlockWindowTmp{}.get_window_lengths()[I1{}])
+                              : (NPerBlock == BDramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
+                                 KPerBlock == BDramBlockWindowTmp{}.get_window_lengths()[I1{}]),
+                          "B block window has incorrect lengths for defined BLayout!");
+
+            ////////////// global window & register /////////////////
+            // A DRAM tile window for load
+            auto a_copy_dram_window =
+                make_tile_window_linear(a_dram_block_window_tmp.get_bottom_tensor_view(),
+                                        make_tuple(number<MPerBlock>{}, number<KPerBlock>{}),
+                                        a_dram_block_window_tmp.get_window_origin(),
+                                        Policy::template MakeADramTileDistribution<Problem>());
+
+            // B DRAM tile window for load
+            auto b_copy_dram_window =
+                make_tile_window_linear(b_dram_block_window_tmp.get_bottom_tensor_view(),
+                                        make_tuple(number<NPerBlock>{}, number<KPerBlock>{}),
+                                        b_dram_block_window_tmp.get_window_origin(),
+                                        Policy::template MakeBDramTileDistribution<Problem>());
+
+            // A register tile for global load
+            constexpr auto ABlockTileDistr = a_copy_dram_window.get_tile_distribution();
+            constexpr auto BBlockTileDistr = b_copy_dram_window.get_tile_distribution();
+            using ABlockTile = decltype(make_static_distributed_tensor<ADataType>(ABlockTileDistr));
+            using BBlockTile = decltype(make_static_distributed_tensor<BDataType>(BBlockTileDistr));
+            ABlockTile a_global_load_tile;
+            BBlockTile b_global_load_tile;
+
+            using ADramTileWindowStep = typename ADramBlockWindowTmp::BottomTensorIndex;
+            using BDramTileWindowStep = typename BDramBlockWindowTmp::BottomTensorIndex;
+
+            constexpr ADramTileWindowStep a_dram_tile_window_step =
+                is_a_col_major ? make_array(KPerBlock, 0) : make_array(0, KPerBlock);
+            constexpr BDramTileWindowStep b_dram_tile_window_step =
+                is_b_row_major ? make_array(KPerBlock, 0) : make_array(0, KPerBlock);
+
+            // global prefetch 0
+            // global read 0
+            Base::GlobalPrefetch(a_global_load_tile, a_copy_dram_window, a_dram_tile_window_step);
+            Base::GlobalPrefetch(b_global_load_tile, b_copy_dram_window, b_dram_tile_window_step);
+            ////////////// LDS desc, window & register /////////////////
+            auto&& [a_lds_block0, b_lds_block0] = Base::GetABLdsTensorViews(p_smem_0);
+            auto&& [a_lds_block1, b_lds_block1] = Base::GetABLdsTensorViews(p_smem_1);
+
+            auto a_copy_lds_window0 = make_tile_window(
+                a_lds_block0, make_tuple(number<MPerBlock>{}, number<KPerBlock>{}), {0, 0});
+
+            auto a_copy_lds_window1 = make_tile_window(
+                a_lds_block1, make_tuple(number<MPerBlock>{}, number<KPerBlock>{}), {0, 0});
+
+            auto b_copy_lds_window0 = make_tile_window(
+                b_lds_block0, make_tuple(number<NPerBlock>{}, number<KPerBlock>{}), {0, 0});
+
+            auto b_copy_lds_window1 = make_tile_window(
+                b_lds_block1, make_tuple(number<NPerBlock>{}, number<KPerBlock>{}), {0, 0});
+
+            // Block GEMM
+            auto block_gemm   = BlockGemm();
+            auto c_block_tile = block_gemm.MakeCBlockTile();
+
+            // initialize C
+            tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
+
+            // LDS write 0
+            if constexpr(is_a_col_major)
+            {
+                auto a_shuffle_tmp = make_static_distributed_tensor<ADataType>(
+                    Policy::template MakeShuffledARegTileDistribution<Problem>());
+                transpose_tile2d(a_shuffle_tmp, a_global_load_tile);
+                Base::LocalPrefill(a_copy_lds_window0, a_shuffle_tmp, a_element_func);
+            }
+            else
+            {
+                Base::LocalPrefill(a_copy_lds_window0, a_global_load_tile, a_element_func);
+            }
+            if constexpr(is_b_row_major)
+            {
+                auto b_shuffle_tmp = make_static_distributed_tensor<BDataType>(
+                    Policy::template MakeShuffledBRegTileDistribution<Problem>());
+                transpose_tile2d(b_shuffle_tmp, b_global_load_tile);
+                Base::LocalPrefill(b_copy_lds_window0, b_shuffle_tmp, b_element_func);
+            }
+            else
+            {
+                Base::LocalPrefill(b_copy_lds_window0, b_global_load_tile, b_element_func);
+            }
+
+            // global read 1
+            Base::GlobalPrefetch(a_global_load_tile, a_copy_dram_window, a_dram_tile_window_step);
+            Base::GlobalPrefetch(b_global_load_tile, b_copy_dram_window, b_dram_tile_window_step);
+
+            block_sync_lds();
+
+            constexpr auto ALdsTileDistr = decltype(make_static_tile_distribution(
+                BlockGemm::MakeABlockDistributionEncode())){};
+            constexpr auto BLdsTileDistr = decltype(make_static_tile_distribution(
+                BlockGemm::MakeBBlockDistributionEncode())){};
+
+            using ALdsTile = decltype(make_static_distributed_tensor<ADataType>(ALdsTileDistr));
+            using BLdsTile = decltype(make_static_distributed_tensor<BDataType>(BLdsTileDistr));
+
+            ALdsTile a_block_tile0;
+            ALdsTile a_block_tile1;
+
+            BLdsTile b_block_tile0;
+            BLdsTile b_block_tile1;
+
+            auto a_lds_ld_window0 =
+                make_tile_window_linear(a_lds_block0,
+                                        make_tuple(number<MPerBlock>{}, number<KPerBlock>{}),
+                                        {0, 0},
+                                        ALdsTileDistr);
+            auto a_lds_ld_window1 =
+                make_tile_window_linear(a_lds_block1,
+                                        make_tuple(number<MPerBlock>{}, number<KPerBlock>{}),
+                                        {0, 0},
+                                        ALdsTileDistr);
+            auto b_lds_ld_window0 =
+                make_tile_window_linear(b_lds_block0,
+                                        make_tuple(number<NPerBlock>{}, number<KPerBlock>{}),
+                                        {0, 0},
+                                        BLdsTileDistr);
+            auto b_lds_ld_window1 =
+                make_tile_window_linear(b_lds_block1,
+                                        make_tuple(number<NPerBlock>{}, number<KPerBlock>{}),
+                                        {0, 0},
+                                        BLdsTileDistr);
+
+            Base::LocalPrefetch(a_block_tile0, a_lds_ld_window0);
+            Base::LocalPrefetch(b_block_tile0, b_lds_ld_window0);
+
+            if constexpr(is_a_col_major)
+            {
+                auto a_shuffle_tmp = make_static_distributed_tensor<ADataType>(
+                    Policy::template MakeShuffledARegTileDistribution<Problem>());
+                transpose_tile2d(a_shuffle_tmp, a_global_load_tile);
+                Base::LocalPrefill(a_copy_lds_window1, a_shuffle_tmp, a_element_func);
+            }
+            else
+            {
+                Base::LocalPrefill(a_copy_lds_window1, a_global_load_tile, a_element_func);
+            }
+            if constexpr(is_b_row_major)
+            {
+                auto b_shuffle_tmp = make_static_distributed_tensor<BDataType>(
+                    Policy::template MakeShuffledBRegTileDistribution<Problem>());
+                transpose_tile2d(b_shuffle_tmp, b_global_load_tile);
+                Base::LocalPrefill(b_copy_lds_window1, b_shuffle_tmp, b_element_func);
+            }
+            else
+            {
+                Base::LocalPrefill(b_copy_lds_window1, b_global_load_tile, b_element_func);
+            }
+
+            Base::GlobalPrefetch(a_global_load_tile, a_copy_dram_window, a_dram_tile_window_step);
+            Base::GlobalPrefetch(b_global_load_tile, b_copy_dram_window, b_dram_tile_window_step);
+
+            if(HasHotLoop)
+            {
+                // minus 2 because we have ping-pong double buffer.
+                index_t iCounter = __builtin_amdgcn_readfirstlane(num_loop - 2);
+                do
+                {
+                    // ping
+                    {
+                        block_sync_lds();
+                        Base::LocalPrefetch(a_block_tile1, a_lds_ld_window1);
+                        Base::LocalPrefetch(b_block_tile1, b_lds_ld_window1);
+
+                        if constexpr(is_a_col_major)
+                        {
+                            auto a_shuffle_tmp = make_static_distributed_tensor<ADataType>(
+                                Policy::template MakeShuffledARegTileDistribution<Problem>());
+                            transpose_tile2d(a_shuffle_tmp, a_global_load_tile);
+                            Base::LocalPrefill(a_copy_lds_window0, a_shuffle_tmp, a_element_func);
+                        }
+                        else
+                        {
+                            Base::LocalPrefill(
+                                a_copy_lds_window0, a_global_load_tile, a_element_func);
+                        }
+                        if constexpr(is_b_row_major)
+                        {
+                            auto b_shuffle_tmp = make_static_distributed_tensor<BDataType>(
+                                Policy::template MakeShuffledBRegTileDistribution<Problem>());
+                            transpose_tile2d(b_shuffle_tmp, b_global_load_tile);
+                            Base::LocalPrefill(b_copy_lds_window0, b_shuffle_tmp, b_element_func);
+                        }
+                        else
+                        {
+                            Base::LocalPrefill(
+                                b_copy_lds_window0, b_global_load_tile, b_element_func);
+                        }
+
+                        Base::GlobalPrefetch(
+                            a_global_load_tile, a_copy_dram_window, a_dram_tile_window_step);
+                        Base::GlobalPrefetch(
+                            b_global_load_tile, b_copy_dram_window, b_dram_tile_window_step);
+                        // gemm
+                        block_gemm(c_block_tile, a_block_tile0, b_block_tile0);
+                        HotLoopScheduler();
+                        __builtin_amdgcn_sched_barrier(0);
+                    }
+                    // pong
+                    {
+                        block_sync_lds();
+                        Base::LocalPrefetch(a_block_tile0, a_lds_ld_window0);
+                        Base::LocalPrefetch(b_block_tile0, b_lds_ld_window0);
+
+                        if constexpr(is_a_col_major)
+                        {
+                            auto a_shuffle_tmp = make_static_distributed_tensor<ADataType>(
+                                Policy::template MakeShuffledARegTileDistribution<Problem>());
+                            transpose_tile2d(a_shuffle_tmp, a_global_load_tile);
+                            Base::LocalPrefill(a_copy_lds_window1, a_shuffle_tmp, a_element_func);
+                        }
+                        else
+                        {
+                            Base::LocalPrefill(
+                                a_copy_lds_window1, a_global_load_tile, a_element_func);
+                        }
+                        if constexpr(is_b_row_major)
+                        {
+                            auto b_shuffle_tmp = make_static_distributed_tensor<BDataType>(
+                                Policy::template MakeShuffledBRegTileDistribution<Problem>());
+                            transpose_tile2d(b_shuffle_tmp, b_global_load_tile);
+                            Base::LocalPrefill(b_copy_lds_window1, b_shuffle_tmp, b_element_func);
+                        }
+                        else
+                        {
+                            Base::LocalPrefill(
+                                b_copy_lds_window1, b_global_load_tile, b_element_func);
+                        }
+
+                        Base::GlobalPrefetch(
+                            a_global_load_tile, a_copy_dram_window, a_dram_tile_window_step);
+                        Base::GlobalPrefetch(
+                            b_global_load_tile, b_copy_dram_window, b_dram_tile_window_step);
+                        // gemm
+                        block_gemm(c_block_tile, a_block_tile1, b_block_tile1);
+                        HotLoopScheduler();
+                        __builtin_amdgcn_sched_barrier(0);
+                    }
+                    iCounter -= 2;
+                } while(iCounter > 1);
+            }
+
+            // tail 3
+            if(TailNum == TailNumber::Three)
+            {
+                // 3
+                {
+                    block_sync_lds();
+                    Base::LocalPrefetch(a_block_tile1, a_lds_ld_window1);
+                    Base::LocalPrefetch(b_block_tile1, b_lds_ld_window1);
+                    if constexpr(is_a_col_major)
+                    {
+                        auto a_shuffle_tmp = make_static_distributed_tensor<ADataType>(
+                            Policy::template MakeShuffledARegTileDistribution<Problem>());
+                        transpose_tile2d(a_shuffle_tmp, a_global_load_tile);
+                        Base::LocalPrefill(a_copy_lds_window0, a_shuffle_tmp, a_element_func);
+                    }
+                    else
+                    {
+                        Base::LocalPrefill(a_copy_lds_window0, a_global_load_tile, a_element_func);
+                    }
+                    if constexpr(is_b_row_major)
+                    {
+                        auto b_shuffle_tmp = make_static_distributed_tensor<BDataType>(
+                            Policy::template MakeShuffledBRegTileDistribution<Problem>());
+                        transpose_tile2d(b_shuffle_tmp, b_global_load_tile);
+                        Base::LocalPrefill(b_copy_lds_window0, b_shuffle_tmp, b_element_func);
+                    }
+                    else
+                    {
+                        Base::LocalPrefill(b_copy_lds_window0, b_global_load_tile, b_element_func);
+                    }
+                    block_gemm(c_block_tile, a_block_tile0, b_block_tile0);
+                }
+                // 2
+                {
+                    block_sync_lds();
+                    Base::LocalPrefetch(a_block_tile0, a_lds_ld_window0);
+                    Base::LocalPrefetch(a_block_tile0, a_lds_ld_window0);
+                    block_gemm(c_block_tile, a_block_tile1, b_block_tile1);
+                }
+                // 1
+                {
+                    block_gemm(c_block_tile, a_block_tile0, b_block_tile0);
+                    __builtin_amdgcn_sched_barrier(0);
+                }
+            }
+            else
+            {
+                // 2
+                {
+                    block_sync_lds();
+                    Base::LocalPrefetch(a_block_tile1, a_lds_ld_window1);
+                    Base::LocalPrefetch(b_block_tile1, b_lds_ld_window1);
+                    block_gemm(c_block_tile, a_block_tile0, b_block_tile0);
+                    static_for<0, 8, 1>{}([&](auto i) {
+                        ignore = i;
+                        __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                        __builtin_amdgcn_sched_group_barrier(0x008, 8, 0); // MFMA
+                    });
+                    __builtin_amdgcn_sched_barrier(0);
+                }
+                // 1
+                {
+                    block_gemm(c_block_tile, a_block_tile1, b_block_tile1);
+                    __builtin_amdgcn_sched_barrier(0);
+                }
+            }
+            return c_block_tile;
+        }
+    };
+
+    template <typename ADramBlockWindowTmp,
+              typename BDramBlockWindowTmp,
+              typename AElementFunction,
+              typename BElementFunction>
+    CK_TILE_DEVICE auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
+                                   const AElementFunction& a_element_func,
+                                   const BDramBlockWindowTmp& b_dram_block_window_tmp,
+                                   const BElementFunction& b_element_func,
+                                   index_t num_loop,
+                                   void* p_smem_0,
+                                   void* p_smem_1) const
+    {
+        return PipelineImpl<Scheduler>{}.template operator()<HasHotLoop, TailNum>(
+            a_dram_block_window_tmp,
+            a_element_func,
+            b_dram_block_window_tmp,
+            b_element_func,
+            num_loop,
+            p_smem_0,
+            p_smem_1);
+    }
+
+    public:
+    template <typename ADramBlockWindowTmp, typename BDramBlockWindowTmp>
+    CK_TILE_DEVICE auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
+                                   const BDramBlockWindowTmp& b_dram_block_window_tmp,
+                                   const index_t num_loop,
+                                   void* __restrict__ p_smem_0,
+                                   void* __restrict__ p_smem_1) const
+    {
+        return PipelineImpl<Scheduler>{}.template operator()<HasHotLoop, TailNum>(
+            a_dram_block_window_tmp,
+            [](const ADataType& a) { return a; },
+            b_dram_block_window_tmp,
+            [](const BDataType& b) { return b; },
+            num_loop,
+            p_smem_0,
+            p_smem_1);
+    }
+};
+} // namespace ck_tile

include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_comp_v4_default_policy.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp"
+#include "ck_tile/ops/common/tensor_layout.hpp"
+#include "ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp"
+
+namespace ck_tile {
+// Default policy for GemmPipelineAGmemBGmemCregComputeV4, except the block gemm method, it shares
+// the same vector size implementation, SmemSize, Global memory tile distiribution as the
+// UniversalGemm Pipeline Policy.
+// Default policy class should not be templated, put template on
+// member functions instead.
+struct GemmPipelineAgBgCrCompV4DefaultPolicy
+    : public UniversalGemmBasePolicy<GemmPipelineAgBgCrCompV4DefaultPolicy>
+{
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeALdsBlockDescriptor()
+    {
+        using namespace ck_tile;
+
+        constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+        constexpr index_t KPack      = GetSmemPackA<Problem>();
+
+        constexpr auto a_lds_block_desc_0 = make_naive_tensor_descriptor(
+            make_tuple(number<kKPerBlock / KPack>{}, number<kMPerBlock>{}, number<KPack>{}),
+            make_tuple(number<kMPerBlock * KPack>{}, number<KPack>{}, number<1>{}),
+            number<KPack>{},
+            number<1>{});
+
+        constexpr auto a_lds_block_desc = transform_tensor_descriptor(
+            a_lds_block_desc_0,
+            make_tuple(
+                make_pass_through_transform(number<kMPerBlock>{}),
+                make_merge_transform(make_tuple(number<kKPerBlock>{} / KPack, number<KPack>{}))),
+            make_tuple(sequence<1>{}, sequence<0, 2>{}),
+            make_tuple(sequence<0>{}, sequence<1>{}));
+
+        return a_lds_block_desc;
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeBLdsBlockDescriptor()
+    {
+        constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+        constexpr index_t KPack      = GetSmemPackB<Problem>();
+
+        constexpr auto b_lds_block_desc_0 = make_naive_tensor_descriptor(
+            make_tuple(number<kKPerBlock / KPack>{}, number<kNPerBlock>{}, number<KPack>{}),
+            make_tuple(number<(kNPerBlock)*KPack>{}, number<KPack>{}, number<1>{}),
+            number<KPack>{},
+            number<1>{});
+
+        constexpr auto b_lds_block_desc = transform_tensor_descriptor(
+            b_lds_block_desc_0,
+            make_tuple(
+                make_pass_through_transform(number<kNPerBlock>{}),
+                make_merge_transform(make_tuple(number<kKPerBlock / KPack>{}, number<KPack>{}))),
+            make_tuple(sequence<1>{}, sequence<0, 2>{}),
+            make_tuple(sequence<0>{}, sequence<1>{}));
+
+        return b_lds_block_desc;
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetBlockGemm()
+    {
+        using AccDataType     = float;
+        using BlockWarps      = typename Problem::BlockGemmShape::BlockWarps;
+        using WarpTile        = typename Problem::BlockGemmShape::WarpTile;
+        using WarpGemm        = WarpGemmMfmaDispatcher<typename Problem::ADataType,
+                                                typename Problem::BDataType,
+                                                AccDataType,
+                                                WarpTile::at(I0),
+                                                WarpTile::at(I1),
+                                                WarpTile::at(I2),
+                                                Problem::TransposeC>;
+        using BlockGemmPolicy = BlockGemmARegBRegCRegV1CustomPolicy<typename Problem::ADataType,
+                                                                    typename Problem::BDataType,
+                                                                    typename Problem::CDataType,
+                                                                    BlockWarps,
+                                                                    WarpGemm>;
+
+        return BlockGemmARegBRegCRegV1<Problem, BlockGemmPolicy>{};
+    }
+};
+} // namespace ck_tile

include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_mem.hpp

@@ -124,6 +124,8 @@ struct GemmPipelineAgBgCrMem : public BaseGemmPipelineAgBgCrMem<Problem>
     static constexpr bool kPadN = Problem::kPadN;
     static constexpr bool kPadK = Problem::kPadK;
 
+    static constexpr bool DoubleSmemBuffer = Problem::DoubleSmemBuffer;
+
     // Where is the right place for HasHotLoop and TailNum ???
     static constexpr bool HasHotLoop = Problem::HasHotLoop;
     static constexpr auto TailNum    = Problem::TailNum;

include/ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1.hpp

@@ -52,6 +52,9 @@ struct GemmPipelineAGmemBGmemCRegV1
         // clang-format on
     }
 
+    // For the basic gemm pipelien DoubleSmemBuffer set to be false naturally.
+    static constexpr bool DoubleSmemBuffer = false;
+
     CK_TILE_HOST_DEVICE static constexpr auto TransposeC() { return Problem::TransposeC; }
 
     CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()

include/ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp

@@ -338,7 +338,7 @@ struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
     {
         using ALayout   = remove_cvref_t<typename Problem::ALayout>;
         using ADataType = remove_cvref_t<typename Problem::ADataType>;
-        static_assert(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::RowMajor>);
+        static_assert(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>);
         constexpr index_t kBlockSize = Problem::kBlockSize;
         constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
         constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;

include/ck_tile/ops/gemm/pipeline/gemm_pipeline_problem.hpp

@@ -36,6 +36,8 @@ struct GemmPipelineProblemBase
     static constexpr bool kPadN = Traits::kPadN;
     static constexpr bool kPadK = Traits::kPadK;
 
+    static constexpr bool DoubleSmemBuffer = Traits::DoubleSmemBuffer;
+
     static constexpr auto Scheduler         = GemmPipelineScheduler::Default;
     static constexpr index_t VectorLoadSize = Traits::_VectorSize;
 
@@ -173,6 +175,8 @@ struct UniversalGemmPipelineProblem
     static constexpr bool kPadN = Traits::kPadN;
     static constexpr bool kPadK = Traits::kPadK;
 
+    static constexpr bool DoubleSmemBuffer = Traits::DoubleSmemBuffer;
+
     static constexpr auto Scheduler  = Scheduler_;
     static constexpr auto HasHotLoop = HasHotLoop_;
     static constexpr auto TailNum    = TailNum_;

include/ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp

@@ -9,8 +9,8 @@
 
 namespace ck_tile {
 
-// UniversalGemm Policy
-struct UniversalGemmPipelineAgBgCrPolicy
+template <typename Derived>
+struct UniversalGemmBasePolicy
 {
     static constexpr auto I0 = number<0>{};
     static constexpr auto I1 = number<1>{};
@@ -113,7 +113,7 @@ struct UniversalGemmPipelineAgBgCrPolicy
     template <typename Problem>
     CK_TILE_HOST_DEVICE static constexpr auto GetVectorSizeC()
     {
-        using BlockGemm = remove_cvref_t<decltype(GetBlockGemm<Problem>())>;
+        using BlockGemm = remove_cvref_t<decltype(Derived::template GetBlockGemm<Problem>())>;
         using WG        = typename BlockGemm::WarpGemm;
 
         constexpr bool TransposeC = Problem::TransposeC;
@@ -166,10 +166,116 @@ struct UniversalGemmPipelineAgBgCrPolicy
         }
     }
 
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto IsTransposeC()
+    {
+        return Problem::TransposeC;
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeADramTileDistribution()
+    {
+        using ALayout = remove_cvref_t<typename Problem::ALayout>;
+
+        constexpr index_t BlockSize   = Problem::kBlockSize;
+        constexpr index_t MPerBlock   = Problem::BlockGemmShape::kM;
+        constexpr index_t KPerBlock   = Problem::BlockGemmShape::kK;
+        constexpr index_t VecLoadSize = GetVectorSizeA<Problem>();
+
+        // Tile: MPerBlock X KPerBlock
+        if constexpr(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::RowMajor>)
+        {
+            using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
+                                                                          MPerBlock,
+                                                                          KPerBlock,
+                                                                          VecLoadSize,
+                                                                          ATileAccessPattern>;
+            return TileEncodingPattern::Make2DStaticTileDistribution();
+        }
+        // Tile: KPerBlock X MPerBlock
+        else
+        {
+            using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
+                                                                          KPerBlock,
+                                                                          MPerBlock,
+                                                                          VecLoadSize,
+                                                                          ATileAccessPattern>;
+            return TileEncodingPattern::Make2DStaticTileDistribution();
+        }
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeBDramTileDistribution()
+    {
+        using BLayout = remove_cvref_t<typename Problem::BLayout>;
+
+        constexpr index_t BlockSize   = Problem::kBlockSize;
+        constexpr index_t NPerBlock   = Problem::BlockGemmShape::kN;
+        constexpr index_t KPerBlock   = Problem::BlockGemmShape::kK;
+        constexpr index_t VecLoadSize = GetVectorSizeB<Problem>();
+
+        // Tile: KPerBlock X NPerBlock
+        if constexpr(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>)
+        {
+            using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
+                                                                          KPerBlock,
+                                                                          NPerBlock,
+                                                                          VecLoadSize,
+                                                                          BTileAccessPattern>;
+            return TileEncodingPattern::Make2DStaticTileDistribution();
+        }
+        // Tile: NPerBlock X KPerBlock
+        else
+        {
+            using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
+                                                                          NPerBlock,
+                                                                          KPerBlock,
+                                                                          VecLoadSize,
+                                                                          BTileAccessPattern>;
+            return TileEncodingPattern::Make2DStaticTileDistribution();
+        }
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledARegTileDistribution()
+    {
+        using ALayout = remove_cvref_t<typename Problem::ALayout>;
+        static_assert(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>);
+        constexpr index_t BlockSize   = Problem::kBlockSize;
+        constexpr index_t MPerBlock   = Problem::BlockGemmShape::kN;
+        constexpr index_t KPerBlock   = Problem::BlockGemmShape::kK;
+        constexpr index_t VecLoadSize = GetVectorSizeA<Problem>();
+
+        using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
+                                                                      KPerBlock,
+                                                                      MPerBlock,
+                                                                      VecLoadSize,
+                                                                      ATileAccessPattern>;
+        return TileEncodingPattern::MakeShuffled2DStaticTileDistribution();
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledBRegTileDistribution()
+    {
+        using BLayout = remove_cvref_t<typename Problem::BLayout>;
+        static_assert(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>);
+        constexpr index_t BlockSize   = Problem::kBlockSize;
+        constexpr index_t NPerBlock   = Problem::BlockGemmShape::kN;
+        constexpr index_t KPerBlock   = Problem::BlockGemmShape::kK;
+        constexpr index_t VecLoadSize = GetVectorSizeB<Problem>();
+
+        using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
+                                                                      KPerBlock,
+                                                                      NPerBlock,
+                                                                      VecLoadSize,
+                                                                      BTileAccessPattern>;
+        return TileEncodingPattern::MakeShuffled2DStaticTileDistribution();
+    }
+
     template <typename Problem>
     CK_TILE_HOST_DEVICE static constexpr auto GetSmemPackA()
     {
-        using BlockGemm         = decltype(GetBlockGemm<Problem>());
+        using BlockGemm = remove_cvref_t<decltype(Derived::template GetBlockGemm<Problem>())>;
         constexpr index_t KPack = BlockGemm::Traits::KPack;
         return KPack;
     }
@@ -177,11 +283,43 @@ struct UniversalGemmPipelineAgBgCrPolicy
     template <typename Problem>
     CK_TILE_HOST_DEVICE static constexpr auto GetSmemPackB()
     {
-        using BlockGemm         = decltype(GetBlockGemm<Problem>());
+        using BlockGemm = remove_cvref_t<decltype(Derived::template GetBlockGemm<Problem>())>;
         constexpr index_t KPack = BlockGemm::Traits::KPack;
         return KPack;
     }
 
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSizeA()
+    {
+        constexpr auto a_lds_desc     = Derived::template MakeALdsBlockDescriptor<Problem>();
+        constexpr index_t smem_size_a = integer_least_multiple(
+            sizeof(typename Problem::ADataType) * a_lds_desc.get_element_space_size(), 16);
+        return smem_size_a;
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSizeB()
+    {
+        constexpr auto b_lds_desc     = Derived::template MakeBLdsBlockDescriptor<Problem>();
+        constexpr index_t smem_size_b = integer_least_multiple(
+            sizeof(typename Problem::BDataType) * b_lds_desc.get_element_space_size(), 16);
+        return smem_size_b;
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
+    {
+        constexpr index_t smem_size_a = GetSmemSizeA<Problem>();
+        constexpr index_t smem_size_b = GetSmemSizeB<Problem>();
+
+        return smem_size_a + smem_size_b;
+    }
+};
+
+// UniversalGemm Policy
+struct UniversalGemmPipelineAgBgCrPolicy
+    : public UniversalGemmBasePolicy<UniversalGemmPipelineAgBgCrPolicy>
+{
     template <typename Problem>
     CK_TILE_HOST_DEVICE static constexpr auto MakeALdsBlockDescriptor()
     {
@@ -421,133 +559,6 @@ struct UniversalGemmPipelineAgBgCrPolicy
 #endif
     }
 
-    template <typename Problem>
-    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSizeA()
-    {
-        constexpr index_t smem_size_a = sizeof(typename Problem::ADataType) *
-                                        MakeALdsBlockDescriptor<Problem>().get_element_space_size();
-        return smem_size_a;
-    }
-
-    template <typename Problem>
-    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSizeB()
-    {
-        constexpr index_t smem_size_b = sizeof(typename Problem::BDataType) *
-                                        MakeBLdsBlockDescriptor<Problem>().get_element_space_size();
-        return smem_size_b;
-    }
-
-    template <typename Problem>
-    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
-    {
-        constexpr index_t smem_size_a = GetSmemSizeA<Problem>();
-        constexpr index_t smem_size_b = GetSmemSizeB<Problem>();
-        index_t smem_size             = 0;
-        smem_size += smem_size_a + smem_size_b;
-
-        return smem_size;
-    }
-
-    template <typename Problem>
-    CK_TILE_HOST_DEVICE static constexpr auto MakeADramTileDistribution()
-    {
-        using ALayout = remove_cvref_t<typename Problem::ALayout>;
-
-        constexpr index_t BlockSize   = Problem::kBlockSize;
-        constexpr index_t MPerBlock   = Problem::BlockGemmShape::kM;
-        constexpr index_t KPerBlock   = Problem::BlockGemmShape::kK;
-        constexpr index_t VecLoadSize = GetVectorSizeA<Problem>();
-
-        // Tile: MPerBlock X KPerBlock
-        if constexpr(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::RowMajor>)
-        {
-            using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
-                                                                          MPerBlock,
-                                                                          KPerBlock,
-                                                                          VecLoadSize,
-                                                                          ATileAccessPattern>;
-            return TileEncodingPattern::Make2DStaticTileDistribution();
-        }
-        // Tile: KPerBlock X MPerBlock
-        else
-        {
-            using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
-                                                                          KPerBlock,
-                                                                          MPerBlock,
-                                                                          VecLoadSize,
-                                                                          ATileAccessPattern>;
-            return TileEncodingPattern::Make2DStaticTileDistribution();
-        }
-    }
-
-    template <typename Problem>
-    CK_TILE_HOST_DEVICE static constexpr auto MakeBDramTileDistribution()
-    {
-        using BLayout = remove_cvref_t<typename Problem::BLayout>;
-
-        constexpr index_t BlockSize   = Problem::kBlockSize;
-        constexpr index_t NPerBlock   = Problem::BlockGemmShape::kN;
-        constexpr index_t KPerBlock   = Problem::BlockGemmShape::kK;
-        constexpr index_t VecLoadSize = GetVectorSizeB<Problem>();
-
-        // Tile: KPerBlock X NPerBlock
-        if constexpr(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>)
-        {
-            using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
-                                                                          KPerBlock,
-                                                                          NPerBlock,
-                                                                          VecLoadSize,
-                                                                          BTileAccessPattern>;
-            return TileEncodingPattern::Make2DStaticTileDistribution();
-        }
-        // Tile: NPerBlock X KPerBlock
-        else
-        {
-            using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
-                                                                          NPerBlock,
-                                                                          KPerBlock,
-                                                                          VecLoadSize,
-                                                                          BTileAccessPattern>;
-            return TileEncodingPattern::Make2DStaticTileDistribution();
-        }
-    }
-
-    template <typename Problem>
-    CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledARegTileDistribution()
-    {
-        using ALayout = remove_cvref_t<typename Problem::ALayout>;
-        static_assert(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>);
-        constexpr index_t BlockSize   = Problem::kBlockSize;
-        constexpr index_t MPerBlock   = Problem::BlockGemmShape::kM;
-        constexpr index_t KPerBlock   = Problem::BlockGemmShape::kK;
-        constexpr index_t VecLoadSize = GetVectorSizeA<Problem>();
-
-        using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
-                                                                      KPerBlock,
-                                                                      MPerBlock,
-                                                                      VecLoadSize,
-                                                                      ATileAccessPattern>;
-        return TileEncodingPattern::MakeShuffled2DStaticTileDistribution();
-    }
-
-    template <typename Problem>
-    CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledBRegTileDistribution()
-    {
-        using BLayout = remove_cvref_t<typename Problem::BLayout>;
-        static_assert(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>);
-        constexpr index_t BlockSize   = Problem::kBlockSize;
-        constexpr index_t NPerBlock   = Problem::BlockGemmShape::kN;
-        constexpr index_t KPerBlock   = Problem::BlockGemmShape::kK;
-        constexpr index_t VecLoadSize = GetVectorSizeB<Problem>();
-
-        using TileEncodingPattern = TileDistributionEncodingPattern2D<BlockSize,
-                                                                      KPerBlock,
-                                                                      NPerBlock,
-                                                                      VecLoadSize,
-                                                                      BTileAccessPattern>;
-        return TileEncodingPattern::MakeShuffled2DStaticTileDistribution();
-    }
-
     template <typename Problem>
     CK_TILE_HOST_DEVICE static constexpr auto GetBlockGemm()
     {

include/ck_tile/ops/gemm/pipeline/tile_gemm_traits.hpp

@@ -32,6 +32,7 @@ struct TileGemmTraits
 template <bool kPadM_,
           bool kPadN_,
           bool kPadK_,
+          bool DoubleSmemBuffer_,
           typename ALayout_,
           typename BLayout_,
           typename CLayout_,
@@ -42,6 +43,8 @@ struct TileGemmUniversalTraits
     static constexpr bool kPadN = kPadN_;
     static constexpr bool kPadK = kPadK_;
 
+    static constexpr bool DoubleSmemBuffer = DoubleSmemBuffer_;
+
     using ALayout = ALayout_;
     using BLayout = BLayout_;
     using CLayout = CLayout_;

library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp

@@ -47,10 +47,9 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_generic_i
         //#########################################|          Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl| Scheduler|   Version|          |
         //#########################################|                 |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |          |          |          |
 #if defined(CK_USE_AMD_MFMA_GFX950)
-#else
-        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    16,    16,     32,   8,   16,   16,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              1,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              1,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1>
 #endif // defined(CK_USE_AMD_MFMA_GFX950)
-       // clang-format on
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    16,    16,     32,   8,   16,   16,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              1,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              1,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1>
+        // clang-format on
         >;
 
 template <ck::index_t NDimSpatial,
@@ -67,7 +66,7 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_instances
         //#########################################|          Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl| Scheduler|   Version|          |
         //#########################################|                 |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |          |          |          |
 #if defined(CK_USE_AMD_MFMA_GFX950)
-#else
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    16,    16,     32,   8,   16,   16,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              1,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              1,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1>,
 
         DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    32,    32,     32,   8,   32,   32,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              2,              2,      false,  S<4, 16,  1>, S<2, 0, 1>,  S<1, 0, 2>,                1,              2,              2,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 2>,
@@ -77,7 +76,6 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_instances
         DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    32,    32,     32,   8,   32,   32,    1,    1,  S<4, 16, 1>,  S<2, 0, 1>,  S<1, 0, 2>,                  1,              2,              2,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              2,              2,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 2>,
         DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    32,     32,   8,   32,   32,    2,    1,  S<4, 16, 1>,  S<2, 0, 1>,  S<1, 0, 2>,                  1,              4,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              4,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 4>,
         DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,   128,    32,     32,   8,   32,   32,    4,    1,  S<4, 16, 1>,  S<2, 0, 1>,  S<1, 0, 2>,                  1,              8,              8,      false,  S<4, 4,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              8,              8,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 8>
-#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 
@@ -118,10 +116,9 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_generic_
         //#########################################|          Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl| Scheduler|   Version|          |
         //#########################################|                 |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |          |          |          |
 #if defined(CK_USE_AMD_MFMA_GFX950)
-#else
-        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    16,    16,     32,   8,   16,   16,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              1,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              1,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1>
 #endif // defined(CK_USE_AMD_MFMA_GFX950)
-       // clang-format on
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    16,    16,     32,   8,   16,   16,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              1,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              1,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1>
+        // clang-format on
         >;
 
 template <ck::index_t NDimSpatial,
@@ -138,7 +135,7 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_instance
         //#########################################|          Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl| Scheduler|   Version|          |
         //#########################################|                 |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |          |          |          |
 #if defined(CK_USE_AMD_MFMA_GFX950)
-#else
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    16,    16,     32,   8,   16,   16,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              1,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              1,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1>,
 
         DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    32,    32,     32,   8,   32,   32,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              2,              2,      false,  S<4, 16,  1>, S<2, 0, 1>,  S<1, 0, 2>,                1,              2,              2,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 2>,
@@ -148,7 +145,6 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_instance
         DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    32,    32,     32,   8,   32,   32,    1,    1,  S<4, 16, 1>,  S<2, 0, 1>,  S<1, 0, 2>,                  1,              2,              2,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              2,              2,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 2>,
         DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    32,     32,   8,   32,   32,    2,    1,  S<4, 16, 1>,  S<2, 0, 1>,  S<1, 0, 2>,                  1,              4,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              4,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 4>,
         DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,   128,    32,     32,   8,   32,   32,    4,    1,  S<4, 16, 1>,  S<2, 0, 1>,  S<1, 0, 2>,                  1,              8,              8,      false,  S<4, 4,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              8,              8,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 8>
-#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 
@@ -189,10 +185,9 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_generic_i
         //#########################################|          Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl| Scheduler|   Version|          |
         //#########################################|                 |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |          |          |          |
 #if defined(CK_USE_AMD_MFMA_GFX950)
-#else
-        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    16,    16,     32,   8,   16,   16,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              1,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              1,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>
 #endif // defined(CK_USE_AMD_MFMA_GFX950)
-       // clang-format on
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    16,    16,     32,   8,   16,   16,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              1,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              1,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>
+        // clang-format on
         >;
 
 // NGCHW requires transpose, we use vector loads and stores params for them
@@ -210,7 +205,7 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_instances
         //#########################################|          Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl| Scheduler|   Version|          |
         //#########################################|                 |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |          |          |          |
 #if defined(CK_USE_AMD_MFMA_GFX950)
-#else
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    16,    16,     32,   8,   16,   16,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              1,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              1,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>,
     
         DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    32,    32,     32,   8,   32,   32,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              2,              2,      false,  S<4, 16,  1>, S<2, 0, 1>,  S<1, 0, 2>,                1,              2,              2,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>,
@@ -234,7 +229,6 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_instances
 
         DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    32,     32,   8,   32,   32,    2,    1,  S<4, 16, 1>,  S<2, 0, 1>,  S<1, 0, 2>,                  1,              4,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              4,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>,
         DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,   128,    32,     32,   8,   32,   32,    4,    1,  S<4, 16, 1>,  S<2, 0, 1>,  S<1, 0, 2>,                  1,              8,              8,      false,  S<4, 4,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              8,              8,      false,           1,           1,   S<1, 8, 1, 4>,                  1, Scheduler, PipelineVersion, 8, F16, F16, 8, 1>
-#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 
@@ -253,10 +247,9 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_generic_
         //#########################################|          Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl| Scheduler|   Version|          |
         //#########################################|                 |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |          |          |          |
 #if defined(CK_USE_AMD_MFMA_GFX950)
-#else
-        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,    F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    16,    16,     32,   8,   16,   16,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              1,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              1,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>
 #endif // defined(CK_USE_AMD_MFMA_GFX950)
-       // clang-format on
+        DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,    F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    16,    16,     32,   8,   16,   16,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              1,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              1,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>
+        // clang-format on
         >;
 
 template <ck::index_t NDimSpatial,
@@ -273,7 +266,7 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_instance
         //#########################################|          Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl| Scheduler|   Version|          |
         //#########################################|                 |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |          |          |          |
 #if defined(CK_USE_AMD_MFMA_GFX950)
-#else
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,    F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    16,    16,     32,   8,   16,   16,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              1,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              1,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>,
     
         DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,    F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    32,    32,     32,   8,   32,   32,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              2,              2,      false,  S<4, 16,  1>, S<2, 0, 1>,  S<1, 0, 2>,                1,              2,              2,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>,
@@ -297,7 +290,6 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_instance
 
         DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,    F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    32,     32,   8,   32,   32,    2,    1,  S<4, 16, 1>,  S<2, 0, 1>,  S<1, 0, 2>,                  1,              4,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              4,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>,
         DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,    F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,   128,    32,     32,   8,   32,   32,    4,    1,  S<4, 16, 1>,  S<2, 0, 1>,  S<1, 0, 2>,                  1,              8,              8,      false,  S<4, 4,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              8,              8,      false,           1,           1,   S<1, 8, 1, 4>,                  1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 1>
-#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 

library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_comp_instance.hpp

@@ -61,8 +61,9 @@ using device_grouped_conv_fwd_xdl_bf16_comp_instances = std::tuple<
         //########################################|    Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
         //########################################|           |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
         //########################################|           |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-#if defined(__gfx950__)
-#else
+#if defined(CK_USE_AMD_MFMA_GFX950)
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   128,    64,  16,  16,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          1,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          1,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
+#else // defined(CK_USE_AMD_MFMA_GFX950)
         // Compute friendly
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   256,   256,    32,   8,   8,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
@@ -81,7 +82,7 @@ using device_grouped_conv_fwd_xdl_bf16_comp_instances = std::tuple<
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,    64,    64,   8,   8,  32,   32,    2,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,    64,   128,    64,   8,   8,  32,   32,    1,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,    64,    64,    64,   8,   8,  32,   32,    1,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>
-#endif // defined(__gfx950__)
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 
@@ -97,8 +98,9 @@ using device_grouped_conv_fwd_xdl_f16_comp_instances = std::tuple<
         //########################################|    Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
         //########################################|           |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
         //########################################|           |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-#if defined(__gfx950__)
-#else
+#if defined(CK_USE_AMD_MFMA_GFX950)
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   128,    64,  16,  16,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          1,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          1,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
+#else // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   256,   256,    32,   8,   8,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   128,    32,   8,   8,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
@@ -113,7 +115,7 @@ using device_grouped_conv_fwd_xdl_f16_comp_instances = std::tuple<
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   256,    32,   8,   8,  32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   256,   128,    32,   8,   8,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
-#endif // defined(__gfx950__)
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 
@@ -148,8 +150,9 @@ using device_grouped_conv_fwd_xdl_int8_comp_instances = std::tuple<
         //########################################|    Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
         //########################################|           |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
         //########################################|           |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-#if defined(__gfx950__)
-#else
+#if defined(CK_USE_AMD_MFMA_GFX950)
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,DsLayout,ELayout,int8_t,int8_t,int32_t,   int8_t,    DsLayout,int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,             256,   128,   128,   128,  32,  32,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,              16,        1,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,         1,            1,           1,              S<1, 64, 1, 4>,              16,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
+#else // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   int8_t,   int8_t,     int32_t,      int8_t,    DsLayout,   int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   256,   256,    32,   8,   8,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   int8_t,   int8_t,     int32_t,      int8_t,    DsLayout,   int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   128,    32,   8,   8,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   int8_t,   int8_t,     int32_t,      int8_t,    DsLayout,   int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   256,   256,    32,   8,   8,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
@@ -160,7 +163,7 @@ using device_grouped_conv_fwd_xdl_int8_comp_instances = std::tuple<
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   int8_t,   int8_t,     int32_t,      int8_t,    DsLayout,   int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   256,    32,   8,   8,  32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   int8_t,   int8_t,     int32_t,      int8_t,    DsLayout,   int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   256,   128,    32,   8,   8,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   int8_t,   int8_t,     int32_t,      int8_t,    DsLayout,   int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
-#endif // defined(__gfx950__)
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 

library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp

@@ -45,13 +45,16 @@ using device_grouped_conv_fwd_xdl_merged_groups_bf16_instances = std::tuple<
         //########################################|    Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|     Type|     Type|  Pipeline|   ToMerge|
         //########################################|           |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|         |         | Scheduler|          |
         //########################################|           |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |         |         |          |          |
-#if defined(__gfx950__)
-#else
+#if defined(CK_USE_AMD_MFMA_GFX950)
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     32,   8, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, BF16, BF16, LoopScheduler::Default, 8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     32,   8, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, BF16, BF16, LoopScheduler::Default, 16>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     32,   8, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, BF16, BF16, LoopScheduler::Default, 32>
+#else // defined(CK_USE_AMD_MFMA_GFX950)
         // Instances with NumGroupsPerBatch > 1
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     16,   4, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, BF16, BF16, LoopScheduler::Default, 8>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     16,   4, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, BF16, BF16, LoopScheduler::Default, 16>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     16,   4, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, BF16, BF16, LoopScheduler::Default, 32>
-#endif // defined(__gfx950__)
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 
@@ -67,13 +70,17 @@ using device_grouped_conv_fwd_xdl_merged_groups_f16_instances = std::tuple<
         //########################################|    Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
         //########################################|           |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
         //########################################|           |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-#if defined(__gfx950__)
-#else
+#if defined(CK_USE_AMD_MFMA_GFX950)
+        // Instances with NumGroupsPerBatch > 1
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     32,   8, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, F16, F16, LoopScheduler::Default, 8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     32,   8, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, F16, F16, LoopScheduler::Default, 16>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     32,   8, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, F16, F16, LoopScheduler::Default, 32>
+#else // defined(CK_USE_AMD_MFMA_GFX950)
         // Instances with NumGroupsPerBatch > 1
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     16,   4, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, F16, F16, LoopScheduler::Default, 8>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     16,   4, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, F16, F16, LoopScheduler::Default, 16>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     16,   4, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, F16, F16, LoopScheduler::Default, 32>
-#endif // defined(__gfx950__)
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 

library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instance.cpp

@@ -34,8 +34,8 @@ using device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances = std::tuple<
         //##########|             |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |                       |                            |
         // pipeline v1, 1 wave
 #if defined(CK_USE_AMD_MFMA_GFX950)
-        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4, 16,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>
-#else
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4, 16,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
@@ -68,7 +68,6 @@ using device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances = std::tuple<
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>
 #endif
-#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 

library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instance.cpp

@@ -34,8 +34,8 @@ using device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances = std::tuple<
         //##########|             |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |                       |                            |
         // pipeline v1, 1 wave
 #if defined(CK_USE_AMD_MFMA_GFX950)
-        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4, 16,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>
-#else
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4, 16,   32,   32,    2,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
@@ -68,7 +68,6 @@ using device_batched_gemm_xdl_f16_f16_f16_gkm_gnk_gmn_instances = std::tuple<
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,   128,    64,     4,  8,   32,   32,    2,    1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Col,      Col,    Row, PassThrough, PassThrough, PassThrough,   256,    64,   128,     4,  8,   32,   32,    1,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              1,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>
 #endif
-#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 

library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instance.cpp

@@ -32,10 +32,9 @@ using device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_generic_instances = std::t
         //#################|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector| Stage   |                       |                            |
         //#################|      |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |                       |                            |
 #if defined(CK_USE_AMD_MFMA_GFX950)
-        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,    64,    16,    16,     4, 16,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>
-#else
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,    64,    16,    16,     4, 16,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>
-#endif
     // clang-format on
     >;
 
@@ -48,8 +47,8 @@ using device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances = std::tuple<
         //#################|      |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |                       |                            |
         // pipeline v1, 1 wave
 #if defined(CK_USE_AMD_MFMA_GFX950)
-        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4, 16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>
-#else
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4, 16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
@@ -109,7 +108,6 @@ using device_batched_gemm_xdl_f16_f16_f16_gmk_gkn_gmn_instances = std::tuple<
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,   128,    16,    32,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,      Row,    Row, PassThrough, PassThrough, PassThrough,    64,    16,    16,     4,  8,   16,   16,    1,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>
 #endif
-#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 

library/src/tensor_operation_instance/gpu/batched_gemm/device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instance.cpp

@@ -32,10 +32,9 @@ using device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_generic_instances = std::t
         //#################|      |      |      |        |        |        |        |   Operation|   Operation|   Operation|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |                |       PerVector| Stage   |                       |                            |
         //#################|      |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |                       |                            |
 #if defined(CK_USE_AMD_MFMA_GFX950)
-        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4, 16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>
-#else
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4, 16,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>
-#endif
     // clang-format on
     >;
 
@@ -48,8 +47,8 @@ using device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances = std::tuple<
         //#################|      |      |      |        |        |        |        |            |            |            |      |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |                |                |         |                       |                            |
         // pipeline v1, 1 wave
 #if defined(CK_USE_AMD_MFMA_GFX950)
-        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4, 16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>
-#else
+        DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   128,     4, 16,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   256,   128,     4,  8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   256,   128,   256,     4,  8,   32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,   128,   128,   128,     4,  8,   32,   32,    4,    2,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v1>,
@@ -97,7 +96,6 @@ using device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances = std::tuple<
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    64,    32,     4,  8,   32,   32,    2,    1,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>,
         DeviceBatchedGemmXdl<  F16,   F16,   F16,     F32,     Row,     Col,     Row, PassThrough, PassThrough, PassThrough,    64,    32,    64,     4,  8,   32,   32,    1,    2,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,      true,               7,               1,        1, LoopScheduler::Default,        PipelineVersion::v2>
 #endif
-#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 

library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm/device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp

@@ -26,10 +26,8 @@ using S = ck::Sequence<Is...>;
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using Scale       = ck::tensor_operation::element_wise::Scale;
 
-#if !defined(CK_USE_AMD_MFMA_GFX950)
 static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
 static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
-#endif
 
 // c[g, m, n] = a[g, m, k] * b[g, n, k]
 template <bool Masking>
@@ -41,7 +39,7 @@ using device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_
         //#######################################|        |         |         |        |      |       |       |      |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|   Triangle|
         //#######################################|        |         |         |        |      |       |       |      |        |         |            |            |            |            |            |               |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |           |
 #if defined(CK_USE_AMD_MFMA_GFX950)
-#else
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    256,   128,    32,    64,    32,   8,   8,    2,   32,   32,     2,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
         DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    256,   128,    32,   128,    32,   8,   8,    2,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
         DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault,        1,   256,    128,   256,    32,    64,    32,   8,   8,    2,   32,   32,     1,     8,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
@@ -57,8 +55,7 @@ using device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_
         // Padded fallback kernel
         DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
         DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>
-#endif // defined(CK_USE_AMD_MFMA_GFX950)
-       // clang-format on
+        // clang-format on
         >;
 
 template <bool Masking>
@@ -70,15 +67,14 @@ using device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_
         //#######################################|        |         |         |        |      |       |       |      |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|   Triangle|
         //#######################################|        |         |         |        |      |       |       |      |        |         |            |            |            |            |            |               |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |           |
 #if defined(CK_USE_AMD_MFMA_GFX950)
-#else
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    256,   128,    40,    64,    32,   4,   4,    2,   32,   32,     2,     4,     2,     S<2,128, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,     false,      S<2,128, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
         DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    256,   128,    40,   128,    32,   4,   4,    2,   32,   32,     2,     4,     4,     S<2,128, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,     false,      S<2,128, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
         DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   256,    40,    64,    32,   4,   4,    2,   32,   32,     1,     8,     2,     S<2,128, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,     false,      S<2,128, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
         DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   256,    40,   128,    32,   4,   4,    2,   32,   32,     1,     8,     4,     S<2,128, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,     false,      S<2,128, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
         DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   128,    40,    64,    32,   4,   4,    2,   32,   32,     1,     4,     2,     S<2,128, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,     false,      S<2,128, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>,
         DeviceBatchedGemmSoftmaxGemm_Xdl_CShuffle<     Row,      Col,      Row,     Row,   F16,    F16,    F16,   F16,     F32,      F16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded,        1,   256,    128,   128,    40,   128,    32,   4,   4,    2,   32,   32,     1,     4,     4,     S<2,128, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,     false,      S<2,128, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               4,               4,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8,   Masking>
-#endif // defined(CK_USE_AMD_MFMA_GFX950)
-       // clang-format on
+        // clang-format on
         >;
 
 void add_device_batched_gemm_softmax_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(

library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp

@@ -26,12 +26,10 @@ using S = ck::Sequence<Is...>;
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using ScaleAdd    = ck::tensor_operation::element_wise::ScaleAdd;
 
-#if !defined(CK_USE_AMD_MFMA_GFX950)
 static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
 static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
 
 static constexpr auto TensorDefault = ck::tensor_operation::device::TensorSpecialization::Default;
-#endif
 
 // c[g, m, n] = a[g, m, k] * b[g, n, k]
 template <index_t NumDimG,
@@ -48,7 +46,7 @@ using device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_
         // #############################################|         |        |        |        |        |       |        |        |       |                 |             |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |              |              |              |              |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|            | PerVector|
         // #############################################|         |        |        |        |        |       |        |        |       |                 |             |        |         |            |            |            |            |            |               |              |              |              |              |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |            |          |
 #if defined(CK_USE_AMD_MFMA_GFX950)
-#else
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec,        1>,
         DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,    64,    32,   8,   8,    2,   32,   32,     2,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
         DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,   128,    32,   8,   8,    2,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
@@ -66,8 +64,7 @@ using device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_
         DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec,        1>,  
         DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
         DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<BF16>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>
-#endif // defined(CK_USE_AMD_MFMA_GFX950)
-       // clang-format on
+        // clang-format on
         >;
 
 void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(

library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance.cpp

@@ -26,12 +26,10 @@ using S = ck::Sequence<Is...>;
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using ScaleAdd    = ck::tensor_operation::element_wise::ScaleAdd;
 
-#if !defined(CK_USE_AMD_MFMA_GFX950)
 static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
 static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
 
 static constexpr auto TensorDefault = ck::tensor_operation::device::TensorSpecialization::Default;
-#endif
 
 // c[g, m, n] = a[g, m, k] * b[g, n, k]
 template <index_t NumDimG,
@@ -48,7 +46,7 @@ using device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16
         // #############################################|         |        |        |        |        |      |       |       |      |                |             |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |              |              |              |              |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|            | PerVector|
         // #############################################|         |        |        |        |        |      |       |       |      |                |             |        |         |            |            |            |            |            |               |              |              |              |              |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |            |          |
 #if defined(CK_USE_AMD_MFMA_GFX950)
-#else
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec,       1>,
         DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,    64,    32,   8,   8,    2,   32,   32,     2,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
         DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,   128,    32,   8,   8,    2,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
@@ -68,8 +66,7 @@ using device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16
         DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec,       1>,
         DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
         DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   F16,    F16,    F16,   F16,  ck::Tuple<F16>,  ck::Tuple<>,     F32,      F16, PassThrough, PassThrough,    ScaleAdd, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>
-#endif // defined(CK_USE_AMD_MFMA_GFX950)
-       // clang-format on
+        // clang-format on
         >;
 
 void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(

library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp

@@ -26,12 +26,10 @@ using S = ck::Sequence<Is...>;
 using PassThrough = ck::tensor_operation::element_wise::PassThrough;
 using Scale       = ck::tensor_operation::element_wise::Scale;
 
-#if !defined(CK_USE_AMD_MFMA_GFX950)
 static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
 static constexpr auto GemmPadded  = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
 
 static constexpr auto TensorDefault = ck::tensor_operation::device::TensorSpecialization::Default;
-#endif
 
 // c[g, m, n] = a[g, m, k] * b[g, n, k]
 template <index_t NumDimG,
@@ -48,7 +46,7 @@ using device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_
         // #############################################|         |        |        |        |        |       |        |        |       |             |             |        |         |   Operation|   Operation|   Operation|   Operation|   Operation|               |              |              |              |              |    Stage|      |  Block| Block| Block| Block| Block|    |    |     |     |     |   Per|   Per|   Per| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  Lengths_K0_N_K1|    ArrangeOrder|                |                |       PerVector|    PerVector_K1|           |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|            |
         // #############################################|         |        |        |        |        |       |        |        |       |             |             |        |         |            |            |            |            |            |               |              |              |              |              |         |      |       |      |      |      |      |    |    |     |     |     |  Wave|  Wave|  Wave|                |               |               |               |               |               |          |                 |                |                |                |                |                |           |                 |                |                |                |                |                |           |            |            |                             |                |            |
 #if defined(CK_USE_AMD_MFMA_GFX950)
-#else
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,    64,    32,   128,    32,   8,   8,    2,   32,   32,     1,     2,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
         DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,    64,    32,   8,   8,    2,   32,   32,     2,     4,     2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
         DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    256,   128,    32,   128,    32,   8,   8,    2,   32,   32,     2,     4,     4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<4, 64, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>,
@@ -64,8 +62,7 @@ using device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_
         DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,    GemmDefault, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,     64,   256,    64,    64,    32,   8,   8,    2,   16,   16,     1,    16,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,      true,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,       true,     S<16, 16, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           4,               S<1, 32, 1, 8>,               8, MaskingSpec>,
         // Padded fallback kernel
         DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<  NumDimG, NumDimM, NumDimN, NumDimK, NumDimO,   BF16,    BF16,    BF16,   BF16,  ck::Tuple<>,  ck::Tuple<>,     F32,     BF16, PassThrough, PassThrough,       Scale, PassThrough, PassThrough,     GemmPadded, TensorDefault, TensorDefault, TensorDefault, TensorDefault,        1,   256,    128,   128,    64,   128,    32,   8,   8,    2,   32,   32,     1,     4,     4,     S<8, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,     false,      S<8, 32, 1>,      S<1, 0, 2>,      S<1, 0, 2>,               2,               8,               8,      false,     S< 8, 32, 1>,      S<0, 2, 1>,      S<0, 2, 1>,               1,               4,               2,      false,           1,           2,               S<1, 32, 1, 8>,               8, MaskingSpec>
-#endif // defined(CK_USE_AMD_MFMA_GFX950)
-       // clang-format on
+        // clang-format on
         >;
 
 void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(