compiled version of cross gpu connection

include/ck_tile/ops/gemm/kernel/batched_gemm_kernel.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/ops/gemm/kernel/gemm_kernel.hpp"
+
+namespace ck_tile {
+
+struct BatchedGemmHostArgs : public ck_tile::GemmHostArgs
+{
+    CK_TILE_HOST BatchedGemmHostArgs() = default;
+    CK_TILE_HOST BatchedGemmHostArgs(const void* a_ptr_,
+                                     const void* b_ptr_,
+                                     void* c_ptr_,
+                                     ck_tile::index_t k_batch_,
+                                     ck_tile::index_t M_,
+                                     ck_tile::index_t N_,
+                                     ck_tile::index_t K_,
+                                     ck_tile::index_t stride_A_,
+                                     ck_tile::index_t stride_B_,
+                                     ck_tile::index_t stride_C_,
+                                     ck_tile::index_t batch_stride_A_,
+                                     ck_tile::index_t batch_stride_B_,
+                                     ck_tile::index_t batch_stride_C_,
+                                     ck_tile::index_t batch_count_)
+        : GemmHostArgs(
+              a_ptr_, b_ptr_, c_ptr_, k_batch_, M_, N_, K_, stride_A_, stride_B_, stride_C_),
+          batch_stride_A(batch_stride_A_),
+          batch_stride_B(batch_stride_B_),
+          batch_stride_C(batch_stride_C_),
+          batch_count(batch_count_)
+    {
+    }
+
+    ck_tile::index_t batch_stride_A;
+    ck_tile::index_t batch_stride_B;
+    ck_tile::index_t batch_stride_C;
+    ck_tile::index_t batch_count;
+};
+
+template <typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>
+struct BatchedGemmKernel : public GemmKernel<TilePartitioner_, GemmPipeline_, EpiloguePipeline_>
+{
+    using Base = GemmKernel<TilePartitioner_, GemmPipeline_, EpiloguePipeline_>;
+
+    using GemmKernelArgs = typename Base::GemmKernelArgs;
+
+    using ADataType = typename Base::ADataType;
+    using BDataType = typename Base::BDataType;
+    using CDataType = typename Base::CDataType;
+
+    using TilePartitioner  = typename Base::TilePartitioner;
+    using GemmPipeline     = typename Base::GemmPipeline;
+    using EpiloguePipeline = typename Base::EpiloguePipeline;
+    using ALayout          = typename Base::ALayout;
+    using BLayout          = typename Base::BLayout;
+    using CLayout          = typename Base::CLayout;
+
+    struct BatchedGemmKernelArgs : GemmKernelArgs
+    {
+        index_t batch_stride_A;
+        index_t batch_stride_B;
+        index_t batch_stride_C;
+        index_t batch_count;
+    };
+
+    using KernelArgs = BatchedGemmKernelArgs;
+
+    __host__ static constexpr auto GridSize(index_t M, index_t N, index_t batch_count)
+    {
+        return TilePartitioner::GridSize(M, N, batch_count);
+    }
+
+    __host__ static constexpr auto BlockSize() { return dim3(Base::KernelBlockSize); }
+
+    CK_TILE_HOST static constexpr BatchedGemmKernelArgs
+    MakeKernelArgs(const BatchedGemmHostArgs& hostArgs)
+    {
+        return BatchedGemmKernelArgs{{hostArgs.a_ptr,
+                                      hostArgs.b_ptr,
+                                      hostArgs.c_ptr,
+                                      hostArgs.M,
+                                      hostArgs.N,
+                                      hostArgs.K,
+                                      hostArgs.stride_A,
+                                      hostArgs.stride_B,
+                                      hostArgs.stride_C},
+                                     hostArgs.batch_stride_A,
+                                     hostArgs.batch_stride_B,
+                                     hostArgs.batch_stride_C,
+                                     hostArgs.batch_count};
+    }
+
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
+    {
+        return max(GemmPipeline::GetSmemSize(), EpiloguePipeline::GetSmemSize());
+    }
+
+    CK_TILE_DEVICE void operator()(BatchedGemmKernelArgs kargs) const
+    {
+        const auto [i_m, i_n] = TilePartitioner{}();
+        const auto i_batch    = __builtin_amdgcn_readfirstlane(blockIdx.z);
+
+        //  options
+        const auto batch_stride_A = __builtin_amdgcn_readfirstlane(kargs.batch_stride_A);
+        const auto batch_offset_A = __builtin_amdgcn_readfirstlane(i_batch * batch_stride_A);
+        const ADataType* a_ptr    = static_cast<const ADataType*>(kargs.a_ptr) + batch_offset_A;
+
+        const auto batch_stride_B = __builtin_amdgcn_readfirstlane(kargs.batch_stride_B);
+        const auto batch_offset_B = __builtin_amdgcn_readfirstlane(i_batch * batch_stride_B);
+        const BDataType* b_ptr    = static_cast<const BDataType*>(kargs.b_ptr) + batch_offset_B;
+
+        const auto batch_stride_C = __builtin_amdgcn_readfirstlane(kargs.batch_stride_C);
+        const auto batch_offset_C = __builtin_amdgcn_readfirstlane(i_batch * batch_stride_C);
+        CDataType* c_ptr          = static_cast<CDataType*>(kargs.c_ptr) + batch_offset_C;
+
+        this->RunGemm(a_ptr, b_ptr, c_ptr, kargs, i_m, i_n);
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/kernel/gemm_kernel.hpp

@@ -12,6 +12,50 @@
 
 namespace ck_tile {
 
+struct GemmProblem
+{
+    CK_TILE_HOST GemmProblem() = default;
+    CK_TILE_HOST GemmProblem(
+        index_t M_, index_t N_, index_t K_, index_t stride_A_, index_t stride_B_, index_t stride_C_)
+        : M(M_), N(N_), K(K_), stride_A(stride_A_), stride_B(stride_B_), stride_C(stride_C_)
+    {
+    }
+
+    index_t M;
+    index_t N;
+    index_t K;
+    index_t stride_A;
+    index_t stride_B;
+    index_t stride_C;
+};
+
+struct GemmHostArgs : public GemmProblem
+{
+    CK_TILE_HOST GemmHostArgs() = default;
+    CK_TILE_HOST GemmHostArgs(const void* a_ptr_,
+                              const void* b_ptr_,
+                              void* c_ptr_,
+                              index_t k_batch_,
+                              index_t M_,
+                              index_t N_,
+                              index_t K_,
+                              index_t stride_A_,
+                              index_t stride_B_,
+                              index_t stride_C_)
+        : GemmProblem(M_, N_, K_, stride_A_, stride_B_, stride_C_),
+          a_ptr(a_ptr_),
+          b_ptr(b_ptr_),
+          c_ptr(c_ptr_),
+          k_batch(k_batch_)
+    {
+    }
+
+    const void* a_ptr;
+    const void* b_ptr;
+    void* c_ptr;
+    index_t k_batch;
+};
+
 template <typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>
 struct GemmKernel
 {
@@ -25,9 +69,12 @@ struct GemmKernel
 
     using ADataType = remove_cvref_t<typename GemmPipeline::ADataType>;
     using BDataType = remove_cvref_t<typename GemmPipeline::BDataType>;
-    // using CAccDataType = remove_cvref_t<typename GemmPipeline::CDataType>;
     using CDataType = remove_cvref_t<typename EpiloguePipeline::ODataType>;
 
+    static constexpr auto I0 = number<0>();
+    static constexpr auto I1 = number<1>();
+    static constexpr auto I2 = number<2>();
+
     __host__ static constexpr auto GridSize(index_t M, index_t N, index_t KBatch)
     {
         return TilePartitioner::GridSize(M, N, KBatch);
@@ -35,7 +82,7 @@ struct GemmKernel
 
     __host__ static constexpr auto BlockSize() { return dim3(KernelBlockSize); }
 
-    struct GemmCommonKargs
+    struct GemmKernelArgs
     {
         const void* a_ptr;
         const void* b_ptr;
@@ -48,36 +95,119 @@ struct GemmKernel
         index_t stride_C;
     };
 
-    CK_TILE_HOST static constexpr GemmCommonKargs MakeKargs(const void* a_ptr,
-                                                            const void* b_ptr,
-                                                            void* c_ptr,
-                                                            index_t M,
-                                                            index_t N,
-                                                            index_t K,
-                                                            index_t stride_A,
-                                                            index_t stride_B,
-                                                            index_t stride_C)
+    CK_TILE_HOST static constexpr GemmKernelArgs MakeKernelArgs(const GemmHostArgs& hostArgs)
     {
-        return GemmCommonKargs{a_ptr, b_ptr, c_ptr, M, N, K, stride_A, stride_B, stride_C};
+        return GemmKernelArgs{hostArgs.a_ptr,
+                              hostArgs.b_ptr,
+                              hostArgs.c_ptr,
+                              hostArgs.M,
+                              hostArgs.N,
+                              hostArgs.K,
+                              hostArgs.stride_A,
+                              hostArgs.stride_B,
+                              hostArgs.stride_C};
     }
+    // CK_TILE_HOST static constexpr GemmKernelArgs MakeKernelArgs(const void* a_ptr,
+    //                                                             const void* b_ptr,
+    //                                                             void* c_ptr,
+    //                                                             index_t M,
+    //                                                             index_t N,
+    //                                                             index_t K,
+    //                                                             index_t stride_A,
+    //                                                             index_t stride_B,
+    //                                                             index_t stride_C)
+    // {
+    //     return GemmKernelArgs{a_ptr, b_ptr, c_ptr, M, N, K, stride_A, stride_B, stride_C};
+    // }
 
     CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
     {
         return max(GemmPipeline::GetSmemSize(), EpiloguePipeline::GetSmemSize());
     }
 
-    CK_TILE_DEVICE void operator()(GemmCommonKargs kargs) const
+    CK_TILE_HOST static bool IsSupportedArgument(const GemmKernelArgs& kargs)
     {
-        const auto [i_m, i_n] = TilePartitioner{}();
-        // options
-        const ADataType* a_start = static_cast<const ADataType*>(kargs.a_ptr);
-        const BDataType* b_start = static_cast<const BDataType*>(kargs.b_ptr);
-        // Convert pointers to tensor views
-        auto a_tensor_view = [&]() {
+        if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
+        {
+            if(kargs.K % TilePartitioner::kK != 0 && GemmPipeline::kPadK == false)
+            {
+                return false;
+            }
+            if(kargs.K % GemmPipeline::VectorSizeA != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(kargs.M % TilePartitioner::kM != 0 && GemmPipeline::kPadM == false)
+            {
+                return false;
+            }
+            if(kargs.M % GemmPipeline::VectorSizeA != 0)
+            {
+                return false;
+            }
+        }
+
+        if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>)
+        {
+            if(kargs.N % TilePartitioner::kN != 0 && GemmPipeline::kPadN == false)
+            {
+                return false;
+            }
+            if(kargs.N % GemmPipeline::VectorSizeB != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(kargs.K % TilePartitioner::kK != 0 && GemmPipeline::kPadK == false)
+            {
+                return false;
+            }
+            if(kargs.K % GemmPipeline::VectorSizeB != 0)
+            {
+                return false;
+            }
+        }
+
+        if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
+        {
+            if(kargs.N % TilePartitioner::kN != 0 && GemmPipeline::kPadN == false)
+            {
+                return false;
+            }
+            if(kargs.N % GemmPipeline::VectorSizeC != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(kargs.M % TilePartitioner::kM != 0 && GemmPipeline::kPadM == false)
+            {
+                return false;
+            }
+            if(kargs.M % GemmPipeline::VectorSizeC != 0)
+            {
+                return false;
+            }
+        }
+        return true;
+    }
+
+    CK_TILE_DEVICE auto MakeGemmTensorViews(const ADataType* a_ptr,
+                                            const BDataType* b_ptr,
+                                            CDataType* c_ptr,
+                                            const GemmKernelArgs& kargs) const
+    {
+        const auto& a_tensor_view = [&]() {
             if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
             {
                 return make_naive_tensor_view<address_space_enum::global>(
-                    a_start,
+                    a_ptr,
                     make_tuple(kargs.M, kargs.K),
                     make_tuple(kargs.stride_A, 1),
                     number<GemmPipeline::VectorSizeA>{},
@@ -86,7 +216,7 @@ struct GemmKernel
             else
             {
                 return make_naive_tensor_view<address_space_enum::global>(
-                    a_start,
+                    a_ptr,
                     make_tuple(kargs.M, kargs.K),
                     make_tuple(1, kargs.stride_A),
                     number<1>{},
@@ -94,11 +224,11 @@ struct GemmKernel
             }
         }();
 
-        auto b_tensor_view = [&]() {
+        const auto& b_tensor_view = [&]() {
             if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>)
             {
                 return make_naive_tensor_view<address_space_enum::global>(
-                    b_start,
+                    b_ptr,
                     make_tuple(kargs.N, kargs.K),
                     make_tuple(1, kargs.stride_B),
                     number<1>{},
@@ -107,7 +237,7 @@ struct GemmKernel
             else
             {
                 return make_naive_tensor_view<address_space_enum::global>(
-                    b_start,
+                    b_ptr,
                     make_tuple(kargs.N, kargs.K),
                     make_tuple(kargs.stride_B, 1),
                     number<GemmPipeline::VectorSizeB>{},
@@ -115,46 +245,11 @@ struct GemmKernel
             }
         }();
 
-        auto a_pad_view = pad_tensor_view(
-            a_tensor_view,
-            make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kK>{}),
-            // somehow clang-format is splitting below line into multiple.
-            // clang-format off
-            sequence<false, GemmPipeline::kPadA>{});
-        // clang-format on
-
-        auto a_block_window = make_tile_window(
-            a_pad_view,
-            make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kK>{}),
-            {i_m, 0});
-
-        auto b_pad_view = pad_tensor_view(
-            b_tensor_view,
-            make_tuple(number<TilePartitioner::kN>{}, number<TilePartitioner::kK>{}),
-            // clang-format off
-            sequence<false, GemmPipeline::kPadB>{});
-        // clang-format on
-
-        auto b_block_window = make_tile_window(
-            b_pad_view,
-            make_tuple(number<TilePartitioner::kN>{}, number<TilePartitioner::kK>{}),
-            {i_n, 0});
-
-        // allocate LDS
-        __shared__ char smem_ptr[GetSmemSize()];
-
-        const index_t num_loop = TilePartitioner::GetLoopNum(kargs.K);
-
-        // Run GEMM cooperatively by whole wokrgroup.
-        auto c_block_tile =
-            GemmPipeline{}.template operator()(a_block_window, b_block_window, num_loop, smem_ptr);
-
-        CDataType* c_start = static_cast<CDataType*>(kargs.c_ptr);
-        auto c_tensor_view = [&]() {
+        const auto& c_tensor_view = [&]() {
             if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
             {
                 return make_naive_tensor_view<address_space_enum::global>(
-                    c_start,
+                    c_ptr,
                     make_tuple(kargs.M, kargs.N),
                     make_tuple(kargs.stride_C, 1),
                     number<GemmPipeline::VectorSizeC>{},
@@ -163,7 +258,7 @@ struct GemmKernel
             else
             {
                 return make_naive_tensor_view<address_space_enum::global>(
-                    c_start,
+                    c_ptr,
                     make_tuple(kargs.M, kargs.N),
                     make_tuple(1, kargs.stride_C),
                     number<1>{},
@@ -171,19 +266,142 @@ struct GemmKernel
             }
         }();
 
-        auto c_pad_view = pad_tensor_view(
-            c_tensor_view,
-            make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kN>{}),
-            // clang-format off
-            sequence<false, GemmPipeline::kPadC>{});
-        // clang-format on
-        auto c_block_window = make_tile_window(
+        return make_tuple(a_tensor_view, b_tensor_view, c_tensor_view);
+    }
+
+    template <typename TensorView>
+    CK_TILE_DEVICE auto MakeGemmPadViews(const TensorView& views) const
+    {
+        const auto& a_pad_view = [&]() {
+            const auto& a_tensor_view = views.at(I0);
+            if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
+            {
+                return pad_tensor_view(
+                    a_tensor_view,
+                    make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kK>{}),
+                    sequence<false, GemmPipeline::kPadK>{});
+            }
+            else
+            {
+                return pad_tensor_view(
+                    a_tensor_view,
+                    make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kK>{}),
+                    sequence<GemmPipeline::kPadM, false>{});
+            }
+        }();
+
+        const auto& b_pad_view = [&]() {
+            const auto& b_tensor_view = views.at(I1);
+            if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::ColumnMajor>)
+            {
+                return pad_tensor_view(
+                    b_tensor_view,
+                    make_tuple(number<TilePartitioner::kN>{}, number<TilePartitioner::kK>{}),
+                    sequence<false, GemmPipeline::kPadK>{});
+            }
+            else
+            {
+                return pad_tensor_view(
+                    b_tensor_view,
+                    make_tuple(number<TilePartitioner::kN>{}, number<TilePartitioner::kK>{}),
+                    sequence<GemmPipeline::kPadN, false>{});
+            }
+        }();
+
+        const auto& c_pad_view = [&]() {
+            const auto& c_tensor_view = views.at(I2);
+            if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
+            {
+                return pad_tensor_view(
+                    c_tensor_view,
+                    make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kN>{}),
+                    sequence<false, GemmPipeline::kPadN>{});
+            }
+            else
+            {
+                return pad_tensor_view(
+                    c_tensor_view,
+                    make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kN>{}),
+                    sequence<GemmPipeline::kPadM, false>{});
+            }
+        }();
+
+        return make_tuple(a_pad_view, b_pad_view, c_pad_view);
+    }
+
+    template <typename PadView>
+    CK_TILE_DEVICE auto
+    MakeGemmTileWindows(const PadView& views, const index_t i_m, const index_t i_n) const
+    {
+        const auto& a_pad_view     = views.at(I0);
+        const auto& a_block_window = make_tile_window(
+            a_pad_view,
+            make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kK>{}),
+            {i_m, 0});
+
+        const auto& b_pad_view     = views.at(I1);
+        const auto& b_block_window = make_tile_window(
+            b_pad_view,
+            make_tuple(number<TilePartitioner::kN>{}, number<TilePartitioner::kK>{}),
+            {i_n, 0});
+
+        const auto& c_pad_view = views.at(I2);
+        auto c_block_window    = make_tile_window(
             c_pad_view,
             make_tuple(number<TilePartitioner::kM>{}, number<TilePartitioner::kN>{}),
             {i_m, i_n});
 
+        return make_tuple(a_block_window, b_block_window, c_block_window);
+    }
+
+    /**
+     * @brief Runs single GEMM problem cooperatively by whole workgroup.
+     *
+     * @param a_ptr input A pointer
+     * @param b_ptr input B pointer
+     * @param c_ptr output C pointer
+     * @param kargs GEMM kernel arguments
+     * @param block_idx_m The GEMM's output M dimension tile index processed by this workgroup.
+     * @param block_idx_n The GEMM's output N dimension tile index processed by this workgroup.
+     */
+    CK_TILE_DEVICE void RunGemm(const ADataType* a_ptr,
+                                const BDataType* b_ptr,
+                                CDataType* c_ptr,
+                                const GemmKernelArgs& kargs,
+                                const index_t block_idx_m,
+                                const index_t block_idx_n) const
+    {
+        // Create Gemm tensor views, pad views and tile windows
+        const auto& gemm_tensor_views_tuple = MakeGemmTensorViews(a_ptr, b_ptr, c_ptr, kargs);
+        const auto& gemm_pad_views          = MakeGemmPadViews(gemm_tensor_views_tuple);
+        auto gemm_tile_windows = MakeGemmTileWindows(gemm_pad_views, block_idx_m, block_idx_n);
+
+        // allocate LDS
+        __shared__ char smem_ptr[GetSmemSize()];
+
+        const index_t num_loop = TilePartitioner::GetLoopNum(kargs.K);
+
+        // Run GEMM cooperatively by whole workgroup.
+        const auto& a_block_window = gemm_tile_windows.at(I0);
+        const auto& b_block_window = gemm_tile_windows.at(I1);
+        const auto& c_block_tile =
+            GemmPipeline{}.template operator()(a_block_window, b_block_window, num_loop, smem_ptr);
+
+        // Run Epilogue Pipeline
+        auto& c_block_window = gemm_tile_windows.at(I2);
         EpiloguePipeline{}(c_block_window, c_block_tile);
     }
+
+    CK_TILE_DEVICE void operator()(GemmKernelArgs kargs) const
+    {
+        const auto [i_m, i_n] = TilePartitioner{}();
+        // options
+        const ADataType* a_ptr = static_cast<const ADataType*>(kargs.a_ptr);
+        const BDataType* b_ptr = static_cast<const BDataType*>(kargs.b_ptr);
+        CDataType* c_ptr       = static_cast<CDataType*>(kargs.c_ptr);
+
+        RunGemm(a_ptr, b_ptr, c_ptr, kargs, i_m, i_n);
+    }
 };
 
 } // namespace ck_tile

include/ck_tile/ops/gemm/kernel/gemm_tile_partitioner.hpp

@@ -35,4 +35,40 @@ struct GemmTilePartitioner
         return make_tuple(iM, iN);
     }
 };
+
+template <typename BlockGemmShape_>
+struct GemmTile1DPartitioner
+{
+    using BlockGemmShape = remove_cvref_t<BlockGemmShape_>;
+
+    static constexpr index_t MPerBlock = BlockGemmShape::kM;
+    static constexpr index_t NPerBlock = BlockGemmShape::kN;
+    static constexpr index_t KPerBlock = BlockGemmShape::kK;
+
+    CK_TILE_HOST static constexpr auto GridSize(index_t M, index_t N)
+    {
+        index_t GridDimX = (M + MPerBlock - 1) / MPerBlock;
+        index_t GridDimY = (N + NPerBlock - 1) / NPerBlock;
+        return dim3(GridDimX * GridDimY, 1, 1);
+    }
+
+    CK_TILE_HOST_DEVICE static constexpr auto GetNBlock(index_t N)
+    {
+        return integer_divide_ceil(N, NPerBlock);
+    }
+
+    CK_TILE_HOST_DEVICE static constexpr auto GetLoopNum(index_t K)
+    {
+        return integer_divide_ceil(K, KPerBlock);
+    }
+
+    CK_TILE_DEVICE auto operator()(index_t blockOffset, index_t NBlockSize)
+    {
+        index_t iM = __builtin_amdgcn_readfirstlane((blockIdx.x - blockOffset) /
+                                                    GetNBlock(NBlockSize) * MPerBlock);
+        index_t iN = __builtin_amdgcn_readfirstlane((blockIdx.x - blockOffset) %
+                                                    GetNBlock(NBlockSize) * NPerBlock);
+        return make_tuple(iM, iN);
+    }
+};
 } // namespace ck_tile

include/ck_tile/ops/gemm/kernel/grouped_gemm_kernel.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <string>
+
+#include "ck_tile/core/numeric/math.hpp"
+#include "ck_tile/core/utility/literals.hpp"
+#include "ck_tile/core/utility/amd_address_space.hpp"
+#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/common.hpp"
+#include "ck_tile/host.hpp"
+
+namespace ck_tile {
+
+struct GroupedGemmHostArgs
+{
+    const void* a_ptr;
+    const void* b_ptr;
+    void* c_ptr;
+    index_t M;
+    index_t N;
+    index_t K;
+    index_t stride_A;
+    index_t stride_B;
+    index_t stride_C;
+};
+
+template <typename TilePartitioner_, typename GemmPipeline_, typename EpiloguePipeline_>
+struct GroupedGemmKernel
+{
+    using TilePartitioner                    = remove_cvref_t<TilePartitioner_>;
+    using GemmPipeline                       = remove_cvref_t<GemmPipeline_>;
+    using EpiloguePipeline                   = remove_cvref_t<EpiloguePipeline_>;
+    using ALayout                            = remove_cvref_t<typename GemmPipeline::ALayout>;
+    using BLayout                            = remove_cvref_t<typename GemmPipeline::BLayout>;
+    using CLayout                            = remove_cvref_t<typename GemmPipeline::CLayout>;
+    static constexpr index_t KernelBlockSize = GemmPipeline::BlockSize;
+
+    using ADataType = remove_cvref_t<typename GemmPipeline::ADataType>;
+    using BDataType = remove_cvref_t<typename GemmPipeline::BDataType>;
+    using CDataType = remove_cvref_t<typename EpiloguePipeline::ODataType>;
+
+    struct GemmTransKernelArg
+    {
+        GroupedGemmHostArgs group_karg;
+        ck_tile::index_t block_start;
+        ck_tile::index_t block_end;
+
+        GemmTransKernelArg() = default;
+        GemmTransKernelArg(GroupedGemmHostArgs&& karg, index_t bl_start, index_t bl_end)
+            : group_karg{karg}, block_start{bl_start}, block_end{bl_end}
+        {
+        }
+    };
+
+    __host__ static size_t GetWorkSpaceSize(const std::vector<GroupedGemmHostArgs>& gemm_descs)
+    {
+        return gemm_descs.size() * sizeof(GemmTransKernelArg);
+    }
+
+    __host__ static constexpr auto BlockSize() { return dim3(KernelBlockSize); }
+
+    using Hargs = GroupedGemmHostArgs;
+
+    __host__ static constexpr auto GridSize(const std::vector<Hargs>& gemm_descs)
+    {
+        index_t grid_size = 0;
+        for(const auto& it_desc : gemm_descs)
+        {
+            const auto dim3 = TilePartitioner::GridSize(it_desc.M, it_desc.N);
+            grid_size += dim3.x * dim3.y * 1;
+        }
+        return dim3(grid_size, 1, 1);
+    }
+
+    CK_TILE_HOST static auto MakeKargs(const std::vector<Hargs>& gemm_descs)
+    {
+        std::vector<GemmTransKernelArg> gemm_kernel_args_;
+        index_t group_count = ck_tile::type_convert<ck_tile::index_t>(gemm_descs.size());
+        index_t grid_size   = 0;
+        gemm_kernel_args_.reserve(group_count);
+
+        for(std::size_t i = 0; i < gemm_descs.size(); ++i)
+        {
+            const index_t M = gemm_descs[i].M;
+            const index_t N = gemm_descs[i].N;
+            const index_t K = gemm_descs[i].K;
+
+            if(M == 0 || N == 0 || K == 0)
+            {
+                continue;
+            }
+
+            const index_t stride_a = gemm_descs[i].stride_A;
+            const index_t stride_b = gemm_descs[i].stride_B;
+            const index_t stride_c = gemm_descs[i].stride_C;
+
+            const auto dim3             = TilePartitioner::GridSize(M, N);
+            const index_t grid_size_grp = dim3.x * 1 * 1;
+
+            const index_t block_start = grid_size;
+            const index_t block_end   = grid_size + grid_size_grp;
+
+            grid_size += grid_size_grp;
+
+            auto karg = GroupedGemmHostArgs{type_convert<const ADataType*>(gemm_descs[i].a_ptr),
+                                            type_convert<const BDataType*>(gemm_descs[i].b_ptr),
+                                            type_convert<CDataType*>(gemm_descs[i].c_ptr),
+                                            M,
+                                            N,
+                                            K,
+                                            stride_a,
+                                            stride_b,
+                                            stride_c};
+
+            gemm_kernel_args_.emplace_back(std::move(karg), block_start, block_end);
+        }
+
+        return gemm_kernel_args_;
+    }
+
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
+    {
+        return max(GemmPipeline::GetSmemSize(), EpiloguePipeline::GetSmemSize());
+    }
+
+    CK_TILE_DEVICE void Run(const Hargs& kargs, const index_t block_start) const
+    {
+        const auto [i_m, i_n] = TilePartitioner{}(block_start, kargs.N);
+        // options
+        const ADataType* a_start = static_cast<const ADataType*>(kargs.a_ptr);
+        const BDataType* b_start = static_cast<const BDataType*>(kargs.b_ptr);
+        // Convert pointers to tensor views
+        auto a_tensor_view = [&]() {
+            if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
+            {
+                return make_naive_tensor_view<address_space_enum::global>(
+                    a_start,
+                    make_tuple(kargs.M, kargs.K),
+                    make_tuple(kargs.stride_A, 1),
+                    number<GemmPipeline::VectorSizeA>{},
+                    number<1>{});
+            }
+            else
+            {
+                return make_naive_tensor_view<address_space_enum::global>(
+                    a_start,
+                    make_tuple(kargs.M, kargs.K),
+                    make_tuple(1, kargs.stride_A),
+                    number<1>{},
+                    number<1>{});
+            }
+        }();
+
+        auto b_tensor_view = [&]() {
+            if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>)
+            {
+                return make_naive_tensor_view<address_space_enum::global>(
+                    b_start,
+                    make_tuple(kargs.N, kargs.K),
+                    make_tuple(1, kargs.stride_B),
+                    number<1>{},
+                    number<1>{});
+            }
+            else
+            {
+                return make_naive_tensor_view<address_space_enum::global>(
+                    b_start,
+                    make_tuple(kargs.N, kargs.K),
+                    make_tuple(kargs.stride_B, 1),
+                    number<GemmPipeline::VectorSizeB>{},
+                    number<1>{});
+            }
+        }();
+
+        auto a_pad_view = [&]() {
+            if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
+            {
+                return pad_tensor_view(a_tensor_view,
+                                       make_tuple(number<TilePartitioner::MPerBlock>{},
+                                                  number<TilePartitioner::KPerBlock>{}),
+                                       sequence<false, GemmPipeline::kPadK>{});
+            }
+            else
+            {
+                return pad_tensor_view(a_tensor_view,
+                                       make_tuple(number<TilePartitioner::MPerBlock>{},
+                                                  number<TilePartitioner::KPerBlock>{}),
+                                       sequence<GemmPipeline::kPadM, false>{});
+            }
+        }();
+        // clang-format on
+
+        auto a_block_window = make_tile_window(
+            a_pad_view,
+            make_tuple(number<TilePartitioner::MPerBlock>{}, number<TilePartitioner::KPerBlock>{}),
+            {i_m, 0});
+
+        auto b_pad_view = [&]() {
+            if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::ColumnMajor>)
+            {
+                return pad_tensor_view(b_tensor_view,
+                                       make_tuple(number<TilePartitioner::NPerBlock>{},
+                                                  number<TilePartitioner::KPerBlock>{}),
+                                       sequence<false, GemmPipeline::kPadK>{});
+            }
+            else
+            {
+                return pad_tensor_view(b_tensor_view,
+                                       make_tuple(number<TilePartitioner::NPerBlock>{},
+                                                  number<TilePartitioner::KPerBlock>{}),
+                                       sequence<GemmPipeline::kPadN, false>{});
+            }
+        }();
+
+        auto b_block_window = make_tile_window(
+            b_pad_view,
+            make_tuple(number<TilePartitioner::NPerBlock>{}, number<TilePartitioner::KPerBlock>{}),
+            {i_n, 0});
+
+        // allocate LDS
+        __shared__ char smem_ptr[GetSmemSize()];
+
+        const index_t num_loop = TilePartitioner::GetLoopNum(kargs.K);
+
+        // Run GEMM cooperatively by whole wokrgroup.
+        auto c_block_tile =
+            GemmPipeline{}.template operator()(a_block_window, b_block_window, num_loop, smem_ptr);
+
+        CDataType* c_start = static_cast<CDataType*>(kargs.c_ptr);
+        auto c_tensor_view = [&]() {
+            if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
+            {
+                return make_naive_tensor_view<address_space_enum::global>(
+                    c_start,
+                    make_tuple(kargs.M, kargs.N),
+                    make_tuple(kargs.stride_C, 1),
+                    number<GemmPipeline::VectorSizeC>{},
+                    number<1>{});
+            }
+            else
+            {
+                return make_naive_tensor_view<address_space_enum::global>(
+                    c_start,
+                    make_tuple(kargs.M, kargs.N),
+                    make_tuple(1, kargs.stride_C),
+                    number<1>{},
+                    number<1>{});
+            }
+        }();
+
+        auto c_pad_view = [&]() {
+            if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
+            {
+                return pad_tensor_view(c_tensor_view,
+                                       make_tuple(number<TilePartitioner::MPerBlock>{},
+                                                  number<TilePartitioner::NPerBlock>{}),
+                                       sequence<false, GemmPipeline::kPadN>{});
+            }
+            else
+            {
+                return pad_tensor_view(c_tensor_view,
+                                       make_tuple(number<TilePartitioner::MPerBlock>{},
+                                                  number<TilePartitioner::NPerBlock>{}),
+                                       sequence<GemmPipeline::kPadM, false>{});
+            }
+        }();
+        auto CBlockWindow_pad = make_tile_window(
+            c_pad_view,
+            make_tuple(number<TilePartitioner::MPerBlock>{}, number<TilePartitioner::NPerBlock>{}),
+            {i_m, i_n});
+
+        EpiloguePipeline{}(CBlockWindow_pad, c_block_tile);
+    }
+
+    CK_TILE_DEVICE void operator()(const void CK_CONSTANT_ADDRESS_SPACE* gemm_descs_const,
+                                   int group_count) const
+    {
+        const index_t block_id   = ck_tile::get_block_1d_id();
+        const auto gemm_desc_ptr = reinterpret_cast<const GemmTransKernelArg*>(
+            cast_pointer_to_generic_address_space(gemm_descs_const));
+
+        index_t left     = 0;
+        index_t right    = group_count;
+        index_t group_id = index_t((left + right) / 2);
+
+        while((!(block_id >= gemm_desc_ptr[group_id].block_start &&
+                 block_id < gemm_desc_ptr[group_id].block_end)) &&
+              left <= right)
+        {
+            if(block_id < gemm_desc_ptr[group_id].block_start)
+            {
+                right = group_id;
+            }
+            else
+            {
+                left = group_id;
+            }
+            group_id = index_t((left + right) / 2);
+        }
+
+        Run(gemm_desc_ptr[group_id].group_karg, gemm_desc_ptr[group_id].block_start);
+    }
+};
+
+} // namespace ck_tile

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+
+namespace ck_tile {
+
+template <typename Problem, typename Policy>
+struct GemmPipelineAgBgCrImplBase
+{
+    using ADataType      = remove_cvref_t<typename Problem::ADataType>;
+    using BDataType      = remove_cvref_t<typename Problem::BDataType>;
+    using BlockGemmShape = remove_cvref_t<typename Problem::BlockGemmShape>;
+
+    static constexpr index_t MPerBlock = BlockGemmShape::kM;
+    static constexpr index_t NPerBlock = BlockGemmShape::kN;
+    static constexpr index_t KPerBlock = BlockGemmShape::kK;
+
+    template <typename DstBlockTile, typename SrcTileWindow>
+    CK_TILE_DEVICE void GlobalPrefetch(DstBlockTile& dst_block_tile,
+                                       SrcTileWindow& dram_tile_window) const
+    {
+        load_tile(dst_block_tile, dram_tile_window);
+        move_tile_window(dram_tile_window, {0, KPerBlock});
+    }
+
+    template <typename DstTileWindow, typename SrcBlockTile, typename ElementFunction>
+    CK_TILE_DEVICE void LocalPrefill(DstTileWindow& lds_tile_window,
+                                     const SrcBlockTile& src_block_tile,
+                                     const ElementFunction& element_func) const
+    {
+        const auto block_tile_tmp = tile_elementwise_in(element_func, src_block_tile);
+        store_tile(lds_tile_window, block_tile_tmp);
+    }
+
+    CK_TILE_DEVICE auto GetABLdsTensorViews(void* p_smem) const
+    {
+        // A tile in LDS
+        ADataType* p_a_lds              = static_cast<ADataType*>(p_smem);
+        constexpr auto a_lds_block_desc = Policy::template MakeALdsBlockDescriptor<Problem>();
+        auto a_lds_block = make_tensor_view<address_space_enum::lds>(p_a_lds, a_lds_block_desc);
+
+        // TODO: LDS alignment should come from Policy!
+        constexpr index_t a_lds_block_space_size_aligned =
+            integer_divide_ceil(sizeof(ADataType) * a_lds_block_desc.get_element_space_size(), 16) *
+            16;
+
+        // B tile in LDS
+        BDataType* p_b_lds = static_cast<BDataType*>(
+            static_cast<void*>(static_cast<char*>(p_smem) + a_lds_block_space_size_aligned));
+        constexpr auto b_lds_block_desc = Policy::template MakeBLdsBlockDescriptor<Problem>();
+        auto b_lds_block = make_tensor_view<address_space_enum::lds>(p_b_lds, b_lds_block_desc);
+
+        return make_tuple(std::move(a_lds_block), std::move(b_lds_block));
+    }
+
+    template <typename ADramBlockWindowTmp, typename ALdsTensorView>
+    CK_TILE_DEVICE auto GetAWindows(const ADramBlockWindowTmp& a_dram_block_window_tmp,
+                                    const ALdsTensorView& a_lds_block_view) const
+    {
+        // A DRAM tile window for load
+        auto a_copy_dram_window =
+            make_tile_window(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>());
+
+        // A LDS tile window for store
+        auto a_copy_lds_window =
+            make_tile_window(a_lds_block_view,
+                             make_tuple(number<MPerBlock>{}, number<KPerBlock>{}),
+                             {0, 0},
+                             a_copy_dram_window.get_tile_distribution());
+
+        auto a_lds_gemm_window = make_tile_window(
+            a_lds_block_view, make_tuple(number<MPerBlock>{}, number<KPerBlock>{}), {0, 0});
+
+        return make_tuple(std::move(a_copy_dram_window),
+                          std::move(a_copy_lds_window),
+                          std::move(a_lds_gemm_window));
+    }
+
+    template <typename BDramBlockWindowTmp, typename BLdsTensorView>
+    CK_TILE_DEVICE auto GetBWindows(const BDramBlockWindowTmp& b_dram_block_window_tmp,
+                                    const BLdsTensorView& b_lds_block_view) const
+    {
+        auto b_copy_dram_window =
+            make_tile_window(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>());
+
+        // B LDS tile window for store
+        auto b_copy_lds_window =
+            make_tile_window(b_lds_block_view,
+                             make_tuple(number<NPerBlock>{}, number<KPerBlock>{}),
+                             {0, 0},
+                             b_copy_dram_window.get_tile_distribution());
+
+        auto b_lds_gemm_window = make_tile_window(
+            b_lds_block_view, make_tuple(number<NPerBlock>{}, number<KPerBlock>{}), {0, 0});
+
+        return make_tuple(std::move(b_copy_dram_window),
+                          std::move(b_copy_lds_window),
+                          std::move(b_lds_gemm_window));
+    }
+};
+
+} // namespace ck_tile

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1_default_policy.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"
+
+namespace ck_tile {
+
+//  A Tile Window: global memory
+//  B Tile Window: global memory
+//  C Distributed tensor: register
+template <typename Problem>
+struct BaseGemmPipelineAgBgCrCompV3
+{
+    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)
+    {
+        ignore = num_loop;
+        return TailNumber::Full;
+    }
+};
+
+// Compute optimized pipeline
+// GlobalPrefetchStages: 2
+// LocalPreFillStages: 1
+// LocalPreFetchStages: 1
+// LocalSharedMemoryBuffer: 1
+template <typename Problem, typename Policy = GemmPipelineAGmemBGmemCRegV1DefaultPolicy>
+struct GemmPipelineAgBgCrCompV3 : public BaseGemmPipelineAgBgCrCompV3<Problem>
+{
+    using Base             = BaseGemmPipelineAgBgCrCompV3<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 VectorSizeA = Problem::VectorSizeA;
+    static constexpr index_t VectorSizeB = Problem::VectorSizeB;
+    static constexpr index_t VectorSizeC = Problem::VectorSizeC;
+
+    static constexpr bool kPadM = Problem::kPadM;
+    static constexpr bool kPadN = Problem::kPadN;
+    static constexpr bool kPadK = Problem::kPadK;
+
+    // Where is the right place for HasHotLoop and TailNum ???
+    static constexpr bool HasHotLoop = Problem::HasHotLoop;
+    static constexpr auto TailNum    = Problem::TailNum;
+    static constexpr auto Scheduler  = Problem::Scheduler;
+
+    using Base::PrefetchStages;
+
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
+    {
+        return Policy::template GetSmemSize<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 * VectorSizeA);
+            constexpr index_t B_Buffer_Load_Inst_Num =
+                NPerBlock * KPerBlock / (BlockSize * VectorSizeB);
+
+            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);
+
+            // A/B split schedule
+            // compiler is likely to use ds_read2 when instruction width smaller than 16bytes
+            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_write_inst_a = A_LDS_Write_Inst_Num;
+            constexpr auto num_ds_write_inst_b = B_LDS_Write_Inst_Num;
+
+            constexpr auto num_buffer_load_inst_a = A_Buffer_Load_Inst_Num;
+            constexpr auto num_buffer_load_inst_b = B_Buffer_Load_Inst_Num;
+
+            constexpr auto num_mfma_inst = C_MFMA_Inst_Num;
+
+            constexpr auto mfma_cycle = NPerXDL == 16 ? 16 : 32;
+            constexpr auto ds_read_a_issue_cycle =
+                A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
+            constexpr auto ds_read_b_issue_cycle =
+                B_LDS_Read_Width * sizeof(BDataType) == 16 ? 8 : 4;
+            constexpr auto ds_read_a_mfma_rate =
+                (mfma_cycle - 4 + 2 * ds_read_a_issue_cycle - 1) / (2 * ds_read_a_issue_cycle);
+            constexpr auto ds_read_b_mfma_rate =
+                (mfma_cycle - 4 + 2 * ds_read_b_issue_cycle - 1) / (2 * ds_read_b_issue_cycle);
+
+            constexpr auto num_dsread_a_mfma =
+                (num_ds_read_inst_a + ds_read_a_mfma_rate - 1) / ds_read_a_mfma_rate;
+            constexpr auto num_dsread_b_mfma =
+                (num_ds_read_inst_b + ds_read_b_mfma_rate - 1) / ds_read_b_mfma_rate;
+
+            // stage 1
+            // Separate this part?
+            // constexpr auto num_mfma_per_ds_read = sizeof(ComputeDataType) / sizeof(ADataType) >
+            //                                               sizeof(ComputeDataType) /
+            //                                               sizeof(BDataType)
+            //                                           ? sizeof(ComputeDataType) /
+            //                                           sizeof(ADataType) : sizeof(ComputeDataType)
+            //                                           / sizeof(BDataType);
+            constexpr auto num_mfma_stage1 =
+                num_mfma_inst - (num_dsread_a_mfma + num_dsread_b_mfma);
+            constexpr auto num_mfma_per_issue =
+                num_mfma_stage1 / (num_buffer_load_inst_a + num_buffer_load_inst_b);
+            constexpr auto num_dswrite_per_issue_a = num_ds_write_inst_a / num_buffer_load_inst_a;
+            constexpr auto num_dswrite_per_issue_b = num_ds_write_inst_b / num_buffer_load_inst_b;
+
+            static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i) {
+                ignore = i;
+                static_for<0, num_dswrite_per_issue_a, 1>{}([&](auto idswrite) {
+                    ignore = idswrite;
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                });
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(
+                    0x008, num_mfma_per_issue - num_dswrite_per_issue_a, 0); // MFMA
+            });
+            static_for<0, num_buffer_load_inst_b, 1>{}([&](auto i) {
+                ignore = i;
+                static_for<0, num_dswrite_per_issue_b, 1>{}([&](auto idswrite) {
+                    ignore = idswrite;
+                    __builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
+                    __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+                });
+                __builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
+                __builtin_amdgcn_sched_group_barrier(
+                    0x008, num_mfma_per_issue - num_dswrite_per_issue_b, 0); // MFMA
+            });
+
+            // stage 2
+            static_for<0, num_dsread_a_mfma, 1>{}([&](auto i) {
+                if constexpr((num_ds_read_inst_a - (i + 1) * ds_read_a_mfma_rate) >=
+                             ds_read_a_mfma_rate)
+                {
+                    __builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read
+                }
+                else
+                {
+                    __builtin_amdgcn_sched_group_barrier(
+                        0x100,
+                        num_ds_read_inst_a - (num_dsread_a_mfma - 1) * ds_read_a_mfma_rate,
+                        0); // DS read
+                }
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            });
+
+            static_for<0, num_dsread_b_mfma, 1>{}([&](auto i) {
+                if constexpr((num_ds_read_inst_b - (i + 1) * ds_read_b_mfma_rate) >=
+                             ds_read_b_mfma_rate)
+                {
+                    __builtin_amdgcn_sched_group_barrier(0x100, ds_read_b_mfma_rate, 0); // DS read
+                }
+                else
+                {
+                    __builtin_amdgcn_sched_group_barrier(
+                        0x100,
+                        num_ds_read_inst_b - (num_dsread_b_mfma - 1) * ds_read_b_mfma_rate,
+                        0); // DS read
+                }
+                __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
+            });
+        }
+
+        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* p_smem) 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>>,
+                "A/B Dram block window should have the same data type as appropriate "
+                "([A|B]DataType) defined in Problem definition!");
+
+            static_assert(MPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
+                              NPerBlock == BDramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
+                              KPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I1{}],
+                          "A/B block window appropriate sizes must be equal to MPerBlock/NPerblock"
+                          " or KPerBlock!");
+
+            // ------------------------------------------------------------------------------------
+            // Definitions of all needed tiles
+
+            // A/B tiles in LDS
+            auto&& [a_lds_block, b_lds_block] = Base::GetABLdsTensorViews(p_smem);
+
+            // A DRAM tile window for load
+            // A LDS tile window for store
+            // A LDS tile for block GEMM
+            auto&& [a_copy_dram_window, a_copy_lds_window, a_lds_gemm_window] =
+                Base::GetAWindows(a_dram_block_window_tmp, a_lds_block);
+
+            // B DRAM tile window for load
+            // B LDS tile window for store
+            // B LDS tile for block GEMM
+            auto&& [b_copy_dram_window, b_copy_lds_window, b_lds_gemm_window] =
+                Base::GetBWindows(b_dram_block_window_tmp, b_lds_block);
+
+            // Block GEMM
+            auto block_gemm   = BlockGemm();
+            auto c_block_tile = block_gemm.MakeCBlockTile();
+
+            using ABlockTileDistr = decltype(a_copy_dram_window.get_tile_distribution());
+            using BBlockTileDistr = decltype(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_block_tile;
+            BBlockTile b_block_tile;
+
+            // -----------------------------------------------------------------------------------------
+            // Gemm pipeline start
+
+            // prefetch
+            // global read 0
+            Base::GlobalPrefetch(a_block_tile, a_copy_dram_window);
+            Base::GlobalPrefetch(b_block_tile, b_copy_dram_window);
+
+            // initialize C
+            tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
+
+            // LDS write 0
+            Base::LocalPrefill(a_copy_lds_window, a_block_tile, a_element_func);
+            Base::LocalPrefill(b_copy_lds_window, b_block_tile, b_element_func);
+
+            Base::GlobalPrefetch(a_block_tile, a_copy_dram_window);
+            Base::GlobalPrefetch(b_block_tile, b_copy_dram_window);
+
+            block_sync_lds();
+            block_gemm.LocalPrefetch(a_lds_gemm_window, b_lds_gemm_window);
+
+            __builtin_amdgcn_sched_barrier(0);
+
+            // main body
+            if constexpr(HasHotLoop)
+            {
+                index_t i = 0;
+                do
+                {
+                    block_sync_lds();
+
+                    Base::LocalPrefill(a_copy_lds_window, a_block_tile, a_element_func);
+                    Base::LocalPrefill(b_copy_lds_window, b_block_tile, b_element_func);
+
+                    Base::GlobalPrefetch(a_block_tile, a_copy_dram_window);
+                    Base::GlobalPrefetch(b_block_tile, b_copy_dram_window);
+
+                    block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
+
+                    block_sync_lds();
+                    block_gemm.LocalPrefetch(a_lds_gemm_window, b_lds_gemm_window);
+                    HotLoopScheduler();
+                    __builtin_amdgcn_sched_barrier(0);
+
+                    i += 1;
+                } while(i < (num_loop - 1));
+            }
+            // tail
+            if constexpr(TailNum == TailNumber::Full)
+            {
+                block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
+            }
+            // Let's leak last MFMA block to epilogue region, cover the potential lds-shuffle
+            // latency
+            // __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) 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);
+    }
+
+    template <typename ADramBlockWindowTmp, typename BDramBlockWindowTmp>
+    CK_TILE_DEVICE auto operator()(const ADramBlockWindowTmp& a_dram_block_window_tmp,
+                                   const BDramBlockWindowTmp& b_dram_block_window_tmp,
+                                   index_t num_loop,
+                                   void* p_smem) 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);
+    }
+};
+
+} // namespace ck_tile

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

@@ -6,6 +6,7 @@
 #include "ck_tile/core.hpp"
 #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1_default_policy.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"
 
 namespace ck_tile {
 
@@ -90,7 +91,8 @@ struct BaseGemmPipelineAgBgCrMem
 template <typename Problem, typename Policy = GemmPipelineAGmemBGmemCRegV1DefaultPolicy>
 struct GemmPipelineAgBgCrMem : public BaseGemmPipelineAgBgCrMem<Problem>
 {
-    using Base = BaseGemmPipelineAgBgCrMem<Problem>;
+    using Base             = BaseGemmPipelineAgBgCrMem<Problem>;
+    using PipelineImplBase = GemmPipelineAgBgCrImplBase<Problem, Policy>;
 
     using ADataType      = remove_cvref_t<typename Problem::ADataType>;
     using BDataType      = remove_cvref_t<typename Problem::BDataType>;
@@ -103,8 +105,9 @@ struct GemmPipelineAgBgCrMem : public BaseGemmPipelineAgBgCrMem<Problem>
 
     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;
@@ -113,9 +116,9 @@ struct GemmPipelineAgBgCrMem : public BaseGemmPipelineAgBgCrMem<Problem>
     static constexpr index_t VectorSizeB = Problem::VectorSizeB;
     static constexpr index_t VectorSizeC = Problem::VectorSizeC;
 
-    static constexpr bool kPadA = Problem::kPadA;
-    static constexpr bool kPadB = Problem::kPadB;
-    static constexpr bool kPadC = Problem::kPadC;
+    static constexpr bool kPadM = Problem::kPadM;
+    static constexpr bool kPadN = Problem::kPadN;
+    static constexpr bool kPadK = Problem::kPadK;
 
     // Where is the right place for HasHotLoop and TailNum ???
     static constexpr bool HasHotLoop = Problem::HasHotLoop;
@@ -124,46 +127,20 @@ struct GemmPipelineAgBgCrMem : public BaseGemmPipelineAgBgCrMem<Problem>
 
     using Base::PrefetchStages;
 
-    CK_TILE_HOST_DEVICE constexpr index_t GetStaticLdsSize()
-    {
-        return integer_divide_ceil(
-                   sizeof(ADataType) *
-                       Policy::template MakeALdsBlockDescriptor<Problem>().get_element_space_size(),
-                   16) *
-                   16 +
-               sizeof(BDataType) *
-                   Policy::template MakeBLdsBlockDescriptor<Problem>().get_element_space_size();
-    }
-
     CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
     {
         return Policy::template GetSmemSize<Problem>();
     }
 
     template <GemmPipelineScheduler Scheduler>
-    struct PipelineImpl
+    struct PipelineImpl : public PipelineImplBase
     {
     };
 
     template <>
-    struct PipelineImpl<GemmPipelineScheduler::Intrawave>
+    struct PipelineImpl<GemmPipelineScheduler::Intrawave> : public PipelineImplBase
     {
-        template <typename DstBlockTile, typename SrcTileWindow>
-        CK_TILE_DEVICE void GlobalPrefetch(DstBlockTile& dst_block_tile,
-                                           SrcTileWindow& dram_tile_window) const
-        {
-            load_tile(dst_block_tile, dram_tile_window);
-            move_tile_window(dram_tile_window, {0, KPerBlock});
-        }
-
-        template <typename DstTileWindow, typename SrcBlockTile, typename ElementFunction>
-        CK_TILE_DEVICE void LocalPrefill(DstTileWindow& lds_tile_window,
-                                         const SrcBlockTile& src_block_tile,
-                                         const ElementFunction& element_func) const
-        {
-            const auto block_tile_tmp = tile_elementwise_in(element_func, src_block_tile);
-            store_tile(lds_tile_window, block_tile_tmp);
-        }
+        using Base = PipelineImplBase;
 
         template <bool HasHotLoop,
                   TailNumber TailNum,
@@ -185,70 +162,42 @@ struct GemmPipelineAgBgCrMem : public BaseGemmPipelineAgBgCrMem<Problem>
                 "A/B Dram block window should have the same data type as appropriate "
                 "([A|B]DataType) defined in Problem definition!");
 
-            static_assert(MPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[number<0>{}] &&
-                              NPerBlock ==
-                                  BDramBlockWindowTmp{}.get_window_lengths()[number<0>{}] &&
-                              KPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[number<1>{}],
+            static_assert(MPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
+                              NPerBlock == BDramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
+                              KPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I1{}],
                           "A/B block window appropriate sizes must be equal to MPerBlock/NPerblock"
                           " or KPerBlock!");
 
             // ------------------------------------------------------------------------------------
             // Definitions of all needed tiles
 
-            // A tile in LDS
-            ADataType* p_a_lds              = static_cast<ADataType*>(p_smem);
-            constexpr auto a_lds_block_desc = Policy::template MakeALdsBlockDescriptor<Problem>();
-            auto a_lds_block = make_tensor_view<address_space_enum::lds>(p_a_lds, a_lds_block_desc);
-
-            // TODO: LDS alignment should come from Policy!
-            constexpr index_t a_lds_block_space_size_aligned =
-                integer_divide_ceil(sizeof(ADataType) * a_lds_block_desc.get_element_space_size(),
-                                    16) *
-                16;
-
-            // B tile in LDS
-            BDataType* p_b_lds = static_cast<BDataType*>(
-                static_cast<void*>(static_cast<char*>(p_smem) + a_lds_block_space_size_aligned));
-            constexpr auto b_lds_block_desc = Policy::template MakeBLdsBlockDescriptor<Problem>();
-            auto b_lds_block = make_tensor_view<address_space_enum::lds>(p_b_lds, b_lds_block_desc);
+            // A/B tiles in LDS
+            // With c++20 could simplify to below line.
+            // Currently get error: captured structured bindings are a C++20 extension
+            // auto&& [a_lds_block, b_lds_block] = Base::GetABLdsTensorViews(p_smem);
+            auto ab_lds_blocks = Base::GetABLdsTensorViews(p_smem);
+            auto& a_lds_block  = ab_lds_blocks.at(I0{});
+            auto& b_lds_block  = ab_lds_blocks.at(I1{});
 
             // A DRAM tile window for load
-            auto a_copy_dram_window =
-                make_tile_window(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>());
-
             // A LDS tile window for store
-            auto a_copy_lds_window =
-                make_tile_window(a_lds_block,
-                                 make_tuple(number<MPerBlock>{}, number<KPerBlock>{}),
-                                 {0, 0},
-                                 a_copy_dram_window.get_tile_distribution());
-            // B DRAM tile window for load
-            auto b_copy_dram_window =
-                make_tile_window(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 LDS tile for block GEMM
+            auto a_windows           = Base::GetAWindows(a_dram_block_window_tmp, a_lds_block);
+            auto& a_copy_dram_window = a_windows.at(I0{});
+            auto& a_copy_lds_window  = a_windows.at(I1{});
+            auto& a_lds_gemm_window  = a_windows.at(I2{});
 
+            // B DRAM tile window for load
             // B LDS tile window for store
-            auto b_copy_lds_window =
-                make_tile_window(b_lds_block,
-                                 make_tuple(number<NPerBlock>{}, number<KPerBlock>{}),
-                                 {0, 0},
-                                 b_copy_dram_window.get_tile_distribution());
-
-            // A LDS tile for block GEMM
-            auto a_lds_gemm_window = make_tile_window(
-                a_lds_block, make_tuple(number<MPerBlock>{}, number<KPerBlock>{}), {0, 0});
             // B LDS tile for block GEMM
-            auto b_lds_gemm_window = make_tile_window(
-                b_lds_block, make_tuple(number<NPerBlock>{}, number<KPerBlock>{}), {0, 0});
+            auto b_windows           = Base::GetBWindows(b_dram_block_window_tmp, b_lds_block);
+            auto& b_copy_dram_window = b_windows.at(I0{});
+            auto& b_copy_lds_window  = b_windows.at(I1{});
+            auto& b_lds_gemm_window  = b_windows.at(I2{});
 
             // Block GEMM
-            constexpr auto block_gemm = BlockGemm();
-            auto c_block_tile         = block_gemm.MakeCBlockTile();
+            auto block_gemm   = BlockGemm();
+            auto c_block_tile = block_gemm.MakeCBlockTile();
 
             using ABlockTileDistr = decltype(a_copy_dram_window.get_tile_distribution());
             using BBlockTileDistr = decltype(b_copy_dram_window.get_tile_distribution());
@@ -266,20 +215,20 @@ struct GemmPipelineAgBgCrMem : public BaseGemmPipelineAgBgCrMem<Problem>
 
             // prefetch
             // global read 0
-            GlobalPrefetch(a_block_tiles.get(I0{}), a_copy_dram_window);
-            GlobalPrefetch(b_block_tiles.get(I0{}), b_copy_dram_window);
+            Base::GlobalPrefetch(a_block_tiles.get(I0{}), a_copy_dram_window);
+            Base::GlobalPrefetch(b_block_tiles.get(I0{}), b_copy_dram_window);
 
             // initialize C
             tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
 
             // LDS write 0
-            LocalPrefill(a_copy_lds_window, a_block_tiles.get(I0{}), a_element_func);
-            LocalPrefill(b_copy_lds_window, b_block_tiles.get(I0{}), b_element_func);
+            Base::LocalPrefill(a_copy_lds_window, a_block_tiles.get(I0{}), a_element_func);
+            Base::LocalPrefill(b_copy_lds_window, b_block_tiles.get(I0{}), b_element_func);
 
             // Global prefetch [1, PrefetchStages]
             static_for<1, PrefetchStages, 1>{}([&](auto prefetch_idx) {
-                GlobalPrefetch(a_block_tiles.get(number<prefetch_idx>{}), a_copy_dram_window);
-                GlobalPrefetch(b_block_tiles.get(number<prefetch_idx>{}), b_copy_dram_window);
+                Base::GlobalPrefetch(a_block_tiles.get(number<prefetch_idx>{}), a_copy_dram_window);
+                Base::GlobalPrefetch(b_block_tiles.get(number<prefetch_idx>{}), b_copy_dram_window);
             });
 
             // main body
@@ -290,24 +239,24 @@ struct GemmPipelineAgBgCrMem : public BaseGemmPipelineAgBgCrMem<Problem>
                 {
                     static_for<0, PrefetchStages, 1>{}([&](auto prefetch_idx) {
                         block_sync_lds();
-                        // block_gemm.LocalPrefetch();
+                        block_gemm.LocalPrefetch(a_lds_gemm_window, b_lds_gemm_window);
                         block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
 
                         block_sync_lds();
 
-                        LocalPrefill(
+                        Base::LocalPrefill(
                             a_copy_lds_window,
                             a_block_tiles.get(number<(prefetch_idx + 1) % PrefetchStages>{}),
                             a_element_func);
-                        LocalPrefill(
+                        Base::LocalPrefill(
                             b_copy_lds_window,
                             b_block_tiles.get(number<(prefetch_idx + 1) % PrefetchStages>{}),
                             b_element_func);
 
-                        GlobalPrefetch(a_block_tiles.get(number<prefetch_idx>{}),
-                                       a_copy_dram_window);
-                        GlobalPrefetch(b_block_tiles.get(number<prefetch_idx>{}),
-                                       b_copy_dram_window);
+                        Base::GlobalPrefetch(a_block_tiles.get(number<prefetch_idx>{}),
+                                             a_copy_dram_window);
+                        Base::GlobalPrefetch(b_block_tiles.get(number<prefetch_idx>{}),
+                                             b_copy_dram_window);
                     });
 
                     i += PrefetchStages;
@@ -318,27 +267,208 @@ struct GemmPipelineAgBgCrMem : public BaseGemmPipelineAgBgCrMem<Problem>
                 static_for<1, tail_num, 1>{}([&](auto prefetch_idx) {
                     block_sync_lds();
 
-                    // block_gemm.LocalPrefetch();
+                    block_gemm.LocalPrefetch(a_lds_gemm_window, b_lds_gemm_window);
                     block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
 
                     block_sync_lds();
-                    LocalPrefill(a_copy_lds_window,
-                                 a_block_tiles.get(number<prefetch_idx>{}),
-                                 a_element_func);
-                    LocalPrefill(b_copy_lds_window,
-                                 b_block_tiles.get(number<prefetch_idx>{}),
-                                 b_element_func);
+
+                    Base::LocalPrefill(a_copy_lds_window,
+                                       a_block_tiles.get(number<prefetch_idx>{}),
+                                       a_element_func);
+                    Base::LocalPrefill(b_copy_lds_window,
+                                       b_block_tiles.get(number<prefetch_idx>{}),
+                                       b_element_func);
+                });
+
+                block_sync_lds();
+                block_gemm.LocalPrefetch(a_lds_gemm_window, b_lds_gemm_window);
+                block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
+            };
+
+            if constexpr(TailNum == TailNumber::One)
+            {
+                block_sync_lds();
+                block_gemm.LocalPrefetch(a_lds_gemm_window, b_lds_gemm_window);
+                block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
+            }
+            else if constexpr(TailNum == TailNumber::Two)
+            {
+                HotLoopTail(number<2>{});
+            }
+            else if constexpr(TailNum == TailNumber::Three)
+            {
+                HotLoopTail(number<3>{});
+            }
+            else if constexpr(TailNum == TailNumber::Four)
+            {
+                HotLoopTail(number<4>{});
+            }
+            else if constexpr(TailNum == TailNumber::Five)
+            {
+                HotLoopTail(number<5>{});
+            }
+            else if constexpr(TailNum == TailNumber::Six)
+            {
+                HotLoopTail(number<6>{});
+            }
+            else if constexpr(TailNum == TailNumber::Seven)
+            {
+                HotLoopTail(number<7>{});
+            }
+            else if constexpr(TailNum == TailNumber::Full)
+            {
+                HotLoopTail(number<PrefetchStages>{});
+            }
+
+            return c_block_tile;
+        }
+    };
+
+    template <>
+    struct PipelineImpl<GemmPipelineScheduler::Interwave> : public PipelineImplBase
+    {
+        using Base = PipelineImplBase;
+
+        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* p_smem) 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>>,
+                "A/B Dram block window should have the same data type as appropriate "
+                "([A|B]DataType) defined in Problem definition!");
+
+            static_assert(MPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
+                              NPerBlock == BDramBlockWindowTmp{}.get_window_lengths()[I0{}] &&
+                              KPerBlock == ADramBlockWindowTmp{}.get_window_lengths()[I1{}],
+                          "A/B block window appropriate sizes must be equal to MPerBlock/NPerblock"
+                          " or KPerBlock!");
+
+            // ------------------------------------------------------------------------------------
+            // Definitions of all needed tiles
+
+            // A/B tiles in LDS
+            // With c++20 could simplify to below line.
+            // Currently get error: captured structured bindings are a C++20 extension
+            // auto&& [a_lds_block, b_lds_block] = Base::GetABLdsTensorViews(p_smem);
+            auto ab_lds_blocks = Base::GetABLdsTensorViews(p_smem);
+            auto& a_lds_block  = ab_lds_blocks.at(I0{});
+            auto& b_lds_block  = ab_lds_blocks.at(I1{});
+
+            // A DRAM tile window for load
+            // A LDS tile window for store
+            // A LDS tile for block GEMM
+            auto a_windows           = Base::GetAWindows(a_dram_block_window_tmp, a_lds_block);
+            auto& a_copy_dram_window = a_windows.at(I0{});
+            auto& a_copy_lds_window  = a_windows.at(I1{});
+            auto& a_lds_gemm_window  = a_windows.at(I2{});
+
+            // B DRAM tile window for load
+            // B LDS tile window for store
+            // B LDS tile for block GEMM
+            auto b_windows           = Base::GetBWindows(b_dram_block_window_tmp, b_lds_block);
+            auto& b_copy_dram_window = b_windows.at(I0{});
+            auto& b_copy_lds_window  = b_windows.at(I1{});
+            auto& b_lds_gemm_window  = b_windows.at(I2{});
+
+            // Block GEMM
+            auto block_gemm   = BlockGemm();
+            auto c_block_tile = block_gemm.MakeCBlockTile();
+
+            using ABlockTileDistr = decltype(a_copy_dram_window.get_tile_distribution());
+            using BBlockTileDistr = decltype(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{}));
+
+            tuple_array<ABlockTile, PrefetchStages> a_block_tiles;
+            tuple_array<BBlockTile, PrefetchStages> b_block_tiles;
+
+            // -----------------------------------------------------------------------------------------
+            // Gemm pipeline start
+
+            // prefetch
+            // global read 0
+            Base::GlobalPrefetch(a_block_tiles.get(I0{}), a_copy_dram_window);
+            Base::GlobalPrefetch(b_block_tiles.get(I0{}), b_copy_dram_window);
+
+            // initialize C
+            tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
+
+            // LDS write 0
+            Base::LocalPrefill(a_copy_lds_window, a_block_tiles.get(I0{}), a_element_func);
+            Base::LocalPrefill(b_copy_lds_window, b_block_tiles.get(I0{}), b_element_func);
+
+            // Global prefetch [1, PrefetchStages]
+            static_for<1, PrefetchStages, 1>{}([&](auto prefetch_idx) {
+                Base::GlobalPrefetch(a_block_tiles.get(number<prefetch_idx>{}), a_copy_dram_window);
+                Base::GlobalPrefetch(b_block_tiles.get(number<prefetch_idx>{}), b_copy_dram_window);
+            });
+
+            // main body
+            if constexpr(HasHotLoop)
+            {
+                index_t i = 0;
+                do
+                {
+                    static_for<0, PrefetchStages, 1>{}([&](auto prefetch_idx) {
+                        block_sync_lds();
+                        block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
+                        // no second block_sync_lds because it's interwave
+
+                        Base::LocalPrefill(
+                            a_copy_lds_window,
+                            a_block_tiles.get(number<(prefetch_idx + 1) % PrefetchStages>{}),
+                            a_element_func);
+                        Base::LocalPrefill(
+                            b_copy_lds_window,
+                            b_block_tiles.get(number<(prefetch_idx + 1) % PrefetchStages>{}),
+                            b_element_func);
+
+                        Base::GlobalPrefetch(a_block_tiles.get(number<prefetch_idx>{}),
+                                             a_copy_dram_window);
+                        Base::GlobalPrefetch(b_block_tiles.get(number<prefetch_idx>{}),
+                                             b_copy_dram_window);
+                    });
+
+                    i += PrefetchStages;
+                } while(i < (num_loop - PrefetchStages));
+            }
+
+            auto HotLoopTail = [&](auto tail_num) {
+                static_for<1, tail_num, 1>{}([&](auto prefetch_idx) {
+                    block_sync_lds();
+                    block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
+                    // no second block_sync_lds because it's interwave
+
+                    Base::LocalPrefill(a_copy_lds_window,
+                                       a_block_tiles.get(number<prefetch_idx>{}),
+                                       a_element_func);
+                    Base::LocalPrefill(b_copy_lds_window,
+                                       b_block_tiles.get(number<prefetch_idx>{}),
+                                       b_element_func);
                 });
 
                 block_sync_lds();
-                // block_gemm.LocalPrefetch();
                 block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
             };
 
             if constexpr(TailNum == TailNumber::One)
             {
                 block_sync_lds();
-                // block_gemm.LocalPrefetch();
                 block_gemm(c_block_tile, a_lds_gemm_window, b_lds_gemm_window);
             }
             else if constexpr(TailNum == TailNumber::Two)

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

@@ -11,6 +11,7 @@ namespace ck_tile {
 
 enum struct GemmPipelineScheduler
 {
+    Default,
     Intrawave,
     Interwave,
 };
@@ -43,6 +44,7 @@ inline std::ostream& operator<<(std::ostream& os, const ck_tile::GemmPipelineSch
 {
     switch(s)
     {
+    case ck_tile::GemmPipelineScheduler::Default: os << "Default"; break;
     case ck_tile::GemmPipelineScheduler::Intrawave: os << "Intrawave"; break;
     case ck_tile::GemmPipelineScheduler::Interwave: os << "Interwave"; break;
     default: os << "";

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

@@ -33,9 +33,9 @@ struct GemmPipelineAGmemBGmemCRegV1
     static constexpr index_t VectorSizeB = Problem::VectorSizeB;
     static constexpr index_t VectorSizeC = Problem::VectorSizeC;
 
-    static constexpr bool kPadA = Problem::kPadA;
-    static constexpr bool kPadB = Problem::kPadB;
-    static constexpr bool kPadC = Problem::kPadC;
+    static constexpr bool kPadM = Problem::kPadM;
+    static constexpr bool kPadN = Problem::kPadN;
+    static constexpr bool kPadK = Problem::kPadK;
 
     CK_TILE_HOST_DEVICE static constexpr index_t GetStaticLdsSize()
     {
@@ -101,11 +101,8 @@ struct GemmPipelineAGmemBGmemCRegV1
                              Policy::template MakeADramTileDistribution<Problem>());
 
         // A LDS tile window for store
-        auto a_copy_lds_window =
-            make_tile_window(a_lds_block,
-                             make_tuple(number<kMPerBlock>{}, number<kKPerBlock>{}),
-                             {0, 0},
-                             a_copy_dram_window.get_tile_distribution());
+        auto a_copy_lds_window = make_tile_window(
+            a_lds_block, make_tuple(number<kMPerBlock>{}, number<kKPerBlock>{}), {0, 0});
 
         // B DRAM tile window for load
         auto b_copy_dram_window =
@@ -115,11 +112,8 @@ struct GemmPipelineAGmemBGmemCRegV1
                              Policy::template MakeBDramTileDistribution<Problem>());
 
         // B LDS tile window for store
-        auto b_copy_lds_window =
-            make_tile_window(b_lds_block,
-                             make_tuple(number<kNPerBlock>{}, number<kKPerBlock>{}),
-                             {0, 0},
-                             b_copy_dram_window.get_tile_distribution());
+        auto b_copy_lds_window = make_tile_window(
+            b_lds_block, make_tuple(number<kNPerBlock>{}, number<kKPerBlock>{}), {0, 0});
 
         // A LDS tile for block GEMM
         auto a_lds_gemm_window = make_tile_window(
@@ -130,7 +124,7 @@ struct GemmPipelineAGmemBGmemCRegV1
             b_lds_block, make_tuple(number<kNPerBlock>{}, number<kKPerBlock>{}), {0, 0});
 
         // Block GEMM
-        constexpr auto block_gemm = Policy::template GetBlockGemm<Problem>();
+        auto block_gemm = Policy::template GetBlockGemm<Problem>();
 
         // Acc register tile
         auto c_block_tile = decltype(block_gemm(a_lds_gemm_window, b_lds_gemm_window)){};
@@ -149,12 +143,32 @@ struct GemmPipelineAGmemBGmemCRegV1
             tile_elementwise_inout([](auto& c) { c = 0; }, c_block_tile);
 
             // LDS write 0
-            const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_block_tile);
-            store_tile(a_copy_lds_window, a_block_tile_tmp);
+            if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::ColumnMajor>)
+            {
+                auto a_shuffle_tmp = make_static_distributed_tensor<ADataType>(
+                    Policy::template MakeShuffledARegBlockDescriptor<Problem>());
+                shuffle_tile(a_shuffle_tmp, a_block_tile);
+                const auto a_block_tile_tmp = tile_elementwise_in(a_element_func, a_shuffle_tmp);
+                store_tile(a_copy_lds_window, a_block_tile_tmp);
+            }
+            else
+            {
+                store_tile(a_copy_lds_window, tile_elementwise_in(a_element_func, a_block_tile));
+            }
 
             // LDS write 0
-            const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
-            store_tile(b_copy_lds_window, b_block_tile_tmp);
+            if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>)
+            {
+                auto b_shuffle_tmp = make_static_distributed_tensor<BDataType>(
+                    Policy::template MakeShuffledBRegBlockDescriptor<Problem>());
+                shuffle_tile(b_shuffle_tmp, b_block_tile);
+                const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_shuffle_tmp);
+                store_tile(b_copy_lds_window, b_block_tile_tmp);
+            }
+            else
+            {
+                store_tile(b_copy_lds_window, tile_elementwise_in(b_element_func, b_block_tile));
+            }
         }
 
         index_t iCounter = num_loop - 1;
@@ -180,8 +194,19 @@ struct GemmPipelineAGmemBGmemCRegV1
             store_tile(a_copy_lds_window, a_block_tile_tmp);
 
             // LDS write i + 1
-            const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
-            store_tile(b_copy_lds_window, b_block_tile_tmp);
+            if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>)
+            {
+                auto b_shuffle_tmp_loop = make_static_distributed_tensor<BDataType>(
+                    Policy::template MakeShuffledBRegBlockDescriptor<Problem>());
+                shuffle_tile(b_shuffle_tmp_loop, b_block_tile);
+                store_tile(b_copy_lds_window,
+                           tile_elementwise_in(b_element_func, b_shuffle_tmp_loop));
+            }
+            else
+            {
+                const auto b_block_tile_tmp = tile_elementwise_in(b_element_func, b_block_tile);
+                store_tile(b_copy_lds_window, b_block_tile_tmp);
+            }
 
             iCounter--;
         }

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

@@ -4,6 +4,7 @@
 #pragma once
 
 #include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp"
 
 namespace ck_tile {
 
@@ -11,6 +12,7 @@ namespace ck_tile {
 // Default policy class should not be templated, put template on member functions instead
 struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
 {
+
 #if 0
     // 2d
     template <typename Problem>
@@ -51,6 +53,7 @@ struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
         constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
         constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
 
+        // TODO: this 8 is AK1! should be a policy parameter!
         constexpr auto a_lds_block_desc_0 = make_naive_tensor_descriptor(
             make_tuple(number<kKPerBlock / 8>{}, number<kMPerBlock>{}, number<8>{}),
             make_tuple(number<(kMPerBlock + 1) * 8>{}, number<8>{}, number<1>{}),
@@ -116,6 +119,20 @@ struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
 
         return smem_size;
     }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetSmemPackA()
+    {
+        using ADataType = remove_cvref_t<typename Problem::ADataType>;
+        return Problem::VectorLoadSize / sizeof(ADataType);
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetSmemPackB()
+    {
+        using BDataType = remove_cvref_t<typename Problem::BDataType>;
+        return Problem::VectorLoadSize / sizeof(BDataType);
+    }
 #elif 1
     // fake XOR
     template <typename Problem>
@@ -192,88 +209,307 @@ struct GemmPipelineAGmemBGmemCRegV1DefaultPolicy
     CK_TILE_HOST_DEVICE static constexpr auto MakeADramTileDistribution()
     {
         using ADataType = remove_cvref_t<typename Problem::ADataType>;
-
-        constexpr index_t kBlockSize = Problem::kBlockSize;
-
-        constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
-        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
-
-        constexpr index_t K1 = 16 / sizeof(ADataType);
-        constexpr index_t K0 = kKPerBlock / K1;
-        constexpr index_t M2 = get_warp_size() / K0;
-#if 1 // coalesce reading for each blocks
-        constexpr index_t M1 = kBlockSize / get_warp_size();
-        static_assert(M2 != 0, "M2 is zero, which will lead to a division by zero error.");
-        static_assert(M1 != 0, "M1 is zero, which will lead to a division by zero error.");
-        constexpr index_t M0 = kMPerBlock / (M2 * M1);
-
-        return make_static_tile_distribution(
-            tile_distribution_encoding<sequence<1>,
-                                       tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
-                                       tuple<sequence<1>, sequence<1, 2>>,
-                                       tuple<sequence<1>, sequence<2, 0>>,
-                                       sequence<1, 2>,
-                                       sequence<0, 1>>{});
-#else // coalesce reading for each warps
-        constexpr index_t M0 = kBlockSize / get_warp_size();
-        constexpr index_t M1 = kMPerBlock / (M2 * M0);
-
-        return make_static_tile_distribution(
-            tile_distribution_encoding<sequence<1>,
-                                       tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
-                                       tuple<sequence<1>, sequence<1, 2>>,
-                                       tuple<sequence<0>, sequence<2, 0>>,
-                                       sequence<1, 2>,
-                                       sequence<1, 1>>{});
-#endif
+        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;
+
+        if constexpr(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>)
+        {
+            constexpr index_t M1           = Problem::VectorLoadSize / sizeof(ADataType);
+            constexpr index_t M0           = MPerBlock / M1;
+            constexpr index_t total_pixels = MPerBlock * KPerBlock / BlockSize;
+            static_assert(total_pixels % M1 == 0);
+            constexpr index_t K3    = total_pixels / M1;
+            constexpr index_t KPack = GetSmemPackA<Problem>();
+            static_assert(KPack % K3 == 0);
+            constexpr index_t K2 = KPack / K3;
+            if constexpr(get_warp_size() % (K2 * M0))
+            {
+                constexpr index_t K1 = get_warp_size() / (K2 * M0);
+                constexpr index_t K0 = BlockSize / get_warp_size();
+                static_assert(KPerBlock == K0 * K1 * K2 * K3);
+                return make_static_tile_distribution(
+                    tile_distribution_encoding<sequence<1>,
+                                               tuple<sequence<M0, M1>, sequence<K0, K1, K2, K3>>,
+                                               tuple<sequence<2>, sequence<2, 1, 2>>,
+                                               tuple<sequence<0>, sequence<1, 0, 2>>,
+                                               sequence<2, 1>,
+                                               sequence<3, 1>>{});
+            }
+            else
+            {
+                constexpr index_t K1   = (K2 * M0) / get_warp_size();
+                constexpr index_t K2_m = K2 / K1;
+                constexpr index_t K0   = BlockSize / get_warp_size() / K1;
+                static_assert(KPerBlock == K0 * K1 * K2_m * K3);
+                return make_static_tile_distribution(
+                    tile_distribution_encoding<sequence<1>,
+                                               tuple<sequence<M0, M1>, sequence<K0, K1, K2_m, K3>>,
+                                               tuple<sequence<2, 2>, sequence<1, 2>>,
+                                               tuple<sequence<0, 1>, sequence<0, 2>>,
+                                               sequence<2, 1>,
+                                               sequence<3, 1>>{});
+            }
+        }
+        else
+        {
+            constexpr index_t K1 = 16 / sizeof(ADataType);
+            constexpr index_t K0 = KPerBlock / K1;
+            constexpr index_t M2 = get_warp_size() / K0;
+            // coalesce reading for each blocks
+            if constexpr(get_warp_size() % (M2 * K0) == 0)
+            {
+                constexpr index_t M1 = BlockSize / get_warp_size();
+                static_assert(M2 != 0, "M2 is zero, which will lead to a division by zero error.");
+                static_assert(M1 != 0, "M1 is zero, which will lead to a division by zero error.");
+                constexpr index_t M0 = MPerBlock / (M2 * M1);
+                static_assert(M0 * M1 * M2 == MPerBlock,
+                              "Incorrect M0, M2, M1 configuration! "
+                              "M0, M1, M2 must cover whole MPerBlock!");
+
+                return make_static_tile_distribution(
+                    tile_distribution_encoding<sequence<1>,
+                                               tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
+                                               tuple<sequence<1>, sequence<1, 2>>,
+                                               tuple<sequence<1>, sequence<2, 0>>,
+                                               sequence<1, 2>,
+                                               sequence<0, 1>>{});
+            }
+            else
+            {
+                constexpr index_t M0 = BlockSize / get_warp_size();
+                constexpr index_t M1 = MPerBlock / (M2 * M0);
+                static_assert(M0 * M1 * M2 == MPerBlock,
+                              "Incorrect M0, M1, M2 configuration! "
+                              "M0, M1, M2 must cover whole MPerBlock!");
+                return make_static_tile_distribution(
+                    tile_distribution_encoding<sequence<1>,
+                                               tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
+                                               tuple<sequence<1>, sequence<1, 2>>,
+                                               tuple<sequence<0>, sequence<2, 0>>,
+                                               sequence<1, 2>,
+                                               sequence<1, 1>>{});
+            }
+        }
     }
 
     template <typename Problem>
     CK_TILE_HOST_DEVICE static constexpr auto MakeBDramTileDistribution()
     {
         using BDataType = remove_cvref_t<typename Problem::BDataType>;
+        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;
+
+        if constexpr(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>)
+        {
+            constexpr index_t N1           = Problem::VectorLoadSize / sizeof(BDataType);
+            constexpr index_t N0           = NPerBlock / N1;
+            constexpr index_t total_pixels = NPerBlock * KPerBlock / BlockSize;
+            static_assert(total_pixels % N1 == 0);
+            constexpr index_t K3    = total_pixels / N1;
+            constexpr index_t KPack = GetSmemPackB<Problem>();
+            static_assert(KPack % K3 == 0);
+            constexpr index_t K2 = KPack / K3;
+            if constexpr(get_warp_size() % (K2 * N0) == 0)
+            {
+                constexpr index_t K1 = get_warp_size() / (K2 * N0);
+                constexpr index_t K0 = BlockSize / get_warp_size();
+                static_assert(KPerBlock == K0 * K1 * K2 * K3);
+                return make_static_tile_distribution(
+                    tile_distribution_encoding<sequence<1>,
+                                               tuple<sequence<N0, N1>, sequence<K0, K1, K2, K3>>,
+                                               tuple<sequence<2>, sequence<2, 1, 2>>,
+                                               tuple<sequence<0>, sequence<1, 0, 2>>,
+                                               sequence<2, 1>,
+                                               sequence<3, 1>>{});
+            }
+            else
+            {
+                constexpr index_t K1   = (K2 * N0) / get_warp_size();
+                constexpr index_t K2_m = K2 / K1;
+                constexpr index_t K0   = BlockSize / get_warp_size() / K1;
+                static_assert(KPerBlock == K0 * K1 * K2_m * K3);
+                return make_static_tile_distribution(
+                    tile_distribution_encoding<sequence<1>,
+                                               tuple<sequence<N0, N1>, sequence<K0, K1, K2_m, K3>>,
+                                               tuple<sequence<2, 2>, sequence<1, 2>>,
+                                               tuple<sequence<0, 1>, sequence<0, 2>>,
+                                               sequence<2, 1>,
+                                               sequence<3, 1>>{});
+            }
+        }
+        else
+        {
+
+            constexpr index_t K1 = Problem::VectorLoadSize / sizeof(BDataType);
+            constexpr index_t K0 = KPerBlock / K1;
+            constexpr index_t N2 = get_warp_size() / K0;
+            // coalesce reading for each blocks
+            if constexpr(get_warp_size() % (N2 * K0) == 0)
+            {
+                constexpr index_t N1 = BlockSize / get_warp_size();
+                static_assert(N2 != 0, "N2 is zero, which will lead to a division by zero error.");
+                static_assert(N1 != 0, "N1 is zero, which will lead to a division by zero error.");
+                constexpr index_t N0 = NPerBlock / (N2 * N1);
+                static_assert(N0 * N1 * N2 == NPerBlock,
+                              "Incorrect N0, N1, N2 configuration! "
+                              "N0, N1, N2 must cover whole NPerBlock!");
+
+                return make_static_tile_distribution(
+                    tile_distribution_encoding<sequence<1>,
+                                               tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
+                                               tuple<sequence<1>, sequence<1, 2>>,
+                                               tuple<sequence<1>, sequence<2, 0>>,
+                                               sequence<1, 2>,
+                                               sequence<0, 1>>{});
+            }
+            // coalesce reading for each warps
+            else
+            {
+                constexpr index_t N0 = BlockSize / get_warp_size();
+                constexpr index_t N1 = NPerBlock / (N2 * N0);
+                static_assert(N0 * N1 * N2 == NPerBlock,
+                              "Incorrect N0, N1, N2 configuration! "
+                              "N0, N1, N2 must cover whole NPerBlock!");
+                return make_static_tile_distribution(
+                    tile_distribution_encoding<sequence<1>,
+                                               tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
+                                               tuple<sequence<1>, sequence<1, 2>>,
+                                               tuple<sequence<0>, sequence<2, 0>>,
+                                               sequence<1, 2>,
+                                               sequence<1, 1>>{});
+            }
+        }
+    }
 
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledBRegBlockDescriptor()
+    {
+        using BLayout   = remove_cvref_t<typename Problem::BLayout>;
+        using BDataType = remove_cvref_t<typename Problem::BDataType>;
+        static_assert(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>);
         constexpr index_t kBlockSize = Problem::kBlockSize;
-
         constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
         constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
 
-        constexpr index_t K1 = 16 / sizeof(BDataType);
-        constexpr index_t K0 = kKPerBlock / K1;
-        constexpr index_t N2 = get_warp_size() / K0;
-#if 1 // coalesce reading for each blocks
-        constexpr index_t N1 = kBlockSize / get_warp_size();
-        static_assert(N2 != 0, "M2 is zero, which will lead to a division by zero error.");
-        static_assert(N1 != 0, "M1 is zero, which will lead to a division by zero error.");
-        constexpr index_t N0 = kNPerBlock / (N2 * N1);
-
-        return make_static_tile_distribution(
-            tile_distribution_encoding<sequence<1>,
-                                       tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
-                                       tuple<sequence<1>, sequence<1, 2>>,
-                                       tuple<sequence<1>, sequence<2, 0>>,
-                                       sequence<1, 2>,
-                                       sequence<0, 1>>{});
-#else // coalesce reading for each warps
-        constexpr index_t N0 = kBlockSize / get_warp_size();
-        constexpr index_t N1 = kNPerBlock / (N2 * N0);
-
-        return make_static_tile_distribution(
-            tile_distribution_encoding<sequence<1>,
-                                       tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
-                                       tuple<sequence<1>, sequence<1, 2>>,
-                                       tuple<sequence<0>, sequence<2, 0>>,
-                                       sequence<1, 2>,
-                                       sequence<1, 1>>{});
-#endif
+        constexpr index_t N1           = Problem::VectorLoadSize / sizeof(BDataType);
+        constexpr index_t N0           = kNPerBlock / N1;
+        constexpr index_t total_pixels = kNPerBlock * kKPerBlock / kBlockSize;
+        static_assert(total_pixels % N1 == 0);
+        constexpr index_t K3     = total_pixels / N1;
+        constexpr index_t kKPack = GetSmemPackB<Problem>();
+        static_assert(kKPack % K3 == 0);
+        constexpr index_t K2 = kKPack / K3; // TODO: this dimention could be outside single wave
+        constexpr index_t warp_size = get_warp_size();
+        if constexpr(warp_size % (K2 * N0) == 0)
+        {
+            constexpr index_t K1 = warp_size / (K2 * N0);
+            constexpr index_t K0 = kBlockSize / warp_size;
+
+            return make_static_tile_distribution(
+                tile_distribution_encoding<sequence<1>,
+                                           tuple<sequence<N0, N1>, sequence<K0, K1, K2, K3>>,
+                                           tuple<sequence<2>, sequence<2, 1, 2>>,
+                                           tuple<sequence<0>, sequence<1, 0, 2>>,
+                                           sequence<1, 2>,
+                                           sequence<1, 3>>{});
+        }
+        else
+        {
+            constexpr index_t K1   = (K2 * N0) / get_warp_size();
+            constexpr index_t K2_m = K2 / K1;
+            constexpr index_t K0   = kBlockSize / get_warp_size() / K1;
+            static_assert(kKPerBlock == K0 * K1 * K2_m * K3);
+            return make_static_tile_distribution(
+                tile_distribution_encoding<sequence<1>,
+                                           tuple<sequence<N0, N1>, sequence<K0, K1, K2_m, K3>>,
+                                           tuple<sequence<2, 2>, sequence<1, 2>>,
+                                           tuple<sequence<0, 1>, sequence<0, 2>>,
+                                           sequence<1, 2>,
+                                           sequence<1, 3>>{});
+        }
     }
 
     template <typename Problem>
-    CK_TILE_HOST_DEVICE static constexpr auto GetBlockGemm()
+    CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledARegBlockDescriptor()
     {
-        using BlockGemmPolicy = BlockGemmASmemBSmemCRegV1DefaultPolicy;
+        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>);
+        constexpr index_t kBlockSize = Problem::kBlockSize;
+        constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr index_t M1           = Problem::VectorLoadSize / sizeof(ADataType);
+        constexpr index_t M0           = kMPerBlock / M1;
+        constexpr index_t total_pixels = kMPerBlock * kKPerBlock / kBlockSize;
+        static_assert(total_pixels % M1 == 0);
+        constexpr index_t K3     = total_pixels / M1;
+        constexpr index_t kKPack = GetSmemPackA<Problem>();
+        static_assert(kKPack % K3 == 0);
+        constexpr index_t K2 = kKPack / K3; // TODO: this dimention could be outside single wave
+        constexpr index_t warp_size = get_warp_size();
+        if constexpr(warp_size % (K2 * M0) == 0)
+        {
+            constexpr index_t K1 = warp_size / (K2 * M0);
+            constexpr index_t K0 = kBlockSize / warp_size;
+
+            return make_static_tile_distribution(
+                tile_distribution_encoding<sequence<1>,
+                                           tuple<sequence<M0, M1>, sequence<K0, K1, K2, K3>>,
+                                           tuple<sequence<2>, sequence<2, 1, 2>>,
+                                           tuple<sequence<0>, sequence<1, 0, 2>>,
+                                           sequence<1, 2>,
+                                           sequence<1, 3>>{});
+        }
+        else
+        {
+            constexpr index_t K1   = (K2 * M0) / get_warp_size();
+            constexpr index_t K2_m = K2 / K1;
+            constexpr index_t K0   = kBlockSize / get_warp_size() / K1;
+            static_assert(kKPerBlock == K0 * K1 * K2_m * K3);
+            return make_static_tile_distribution(
+                tile_distribution_encoding<sequence<1>,
+                                           tuple<sequence<M0, M1>, sequence<K0, K1, K2_m, K3>>,
+                                           tuple<sequence<2, 2>, sequence<1, 2>>,
+                                           tuple<sequence<0, 1>, sequence<0, 2>>,
+                                           sequence<1, 2>,
+                                           sequence<1, 3>>{});
+        }
+    }
 
-        return BlockGemmASmemBSmemCRegV1<Problem, BlockGemmPolicy>{};
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetBlockGemm()
+    {
+        constexpr bool TransposeC = false;
+        constexpr auto I0         = number<0>{};
+        constexpr auto I1         = number<1>{};
+        constexpr auto I2         = number<2>{};
+
+        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),
+                                                TransposeC>;
+        using BlockGemmPolicy = BlockGemmASmemBSmemCRegV1CustomPolicy<typename Problem::ADataType,
+                                                                      typename Problem::BDataType,
+                                                                      typename Problem::CDataType,
+                                                                      BlockWarps,
+                                                                      WarpGemm>;
+
+        return BlockUniversalGemmAsBsCr<Problem, BlockGemmPolicy>{};
     }
 };
 

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

@@ -3,40 +3,135 @@
 
 #pragma once
 
-#include "ck_tile/core.hpp"
 #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
 
 namespace ck_tile {
 
-static constexpr int _VectorSize = 16;
-
 template <typename ADataType_,
           typename BDataType_,
           typename CDataType_,
           typename BlockGemmShape_,
           typename TileGemmTraits_>
-struct GemmPipelineProblem
+struct GemmPipelineProblemBase
 {
-    using ADataType      = remove_cvref_t<ADataType_>;
-    using BDataType      = remove_cvref_t<BDataType_>;
-    using CDataType      = remove_cvref_t<CDataType_>;
+    using GemmTraits = remove_cvref_t<TileGemmTraits_>;
+
+    using ADataType = remove_cvref_t<ADataType_>;
+    using BDataType = remove_cvref_t<BDataType_>;
+    using CDataType = remove_cvref_t<CDataType_>;
+
     using BlockGemmShape = remove_cvref_t<BlockGemmShape_>;
-    using GemmTraits     = remove_cvref_t<TileGemmTraits_>;
 
     using ALayout = remove_cvref_t<typename GemmTraits::ALayout>;
     using BLayout = remove_cvref_t<typename GemmTraits::BLayout>;
     using CLayout = remove_cvref_t<typename GemmTraits::CLayout>;
 
-    static constexpr index_t kBlockSize = BlockGemmShape::NumWarps * get_warp_size();
-    static constexpr bool kPadA         = GemmTraits::kPadA;
-    static constexpr bool kPadB         = GemmTraits::kPadB;
-    static constexpr bool kPadC         = GemmTraits::kPadC;
+    static constexpr index_t VectorLoadSize = GemmTraits::_VectorSize;
+    static constexpr index_t kBlockSize     = BlockGemmShape::NumWarps * get_warp_size();
+
+    static constexpr bool kPadM = GemmTraits::kPadM;
+    static constexpr bool kPadN = GemmTraits::kPadN;
+    static constexpr bool kPadK = GemmTraits::kPadK;
+
+    static constexpr auto Scheduler = GemmPipelineScheduler::Default;
+
+    CK_TILE_HOST_DEVICE static constexpr auto GetAlignmentA()
+    {
+        if constexpr(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>)
+        {
+            constexpr index_t pixels_per_thread =
+                BlockGemmShape::kM * BlockGemmShape::kK / kBlockSize;
+            return pixels_per_thread < VectorLoadSize / sizeof(ADataType)
+                       ? pixels_per_thread
+                       : VectorLoadSize / sizeof(ADataType);
+        }
+        else
+        {
+            return VectorLoadSize / sizeof(ADataType);
+        }
+    }
+
+    CK_TILE_HOST_DEVICE static constexpr auto GetAlignmentB()
+    {
+        if constexpr(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>)
+        {
+            constexpr index_t pixels_per_thread =
+                BlockGemmShape::kN * BlockGemmShape::kK / kBlockSize;
+            return pixels_per_thread < VectorLoadSize / sizeof(BDataType)
+                       ? pixels_per_thread
+                       : VectorLoadSize / sizeof(BDataType);
+        }
+        else
+        {
+            return VectorLoadSize / sizeof(BDataType);
+        }
+    }
 
-    static constexpr index_t VectorSizeA = kPadA ? 1 : _VectorSize / sizeof(ADataType);
-    static constexpr index_t VectorSizeB = kPadB ? 1 : _VectorSize / sizeof(BDataType);
-    static constexpr index_t VectorSizeC = kPadC ? 1 : _VectorSize / sizeof(CDataType);
+    CK_TILE_HOST_DEVICE static constexpr auto GetAlignmentC()
+    {
+        if constexpr(std::is_same_v<CLayout, ck_tile::tensor_layout::gemm::ColumnMajor>)
+        {
+            constexpr index_t N1 = kBlockSize / get_warp_size();
+            constexpr index_t N2 = std::min(BlockGemmShape::kN / N1, get_warp_size());
+            constexpr index_t M0 = get_warp_size() / N2;
+            constexpr index_t M1 = BlockGemmShape::kM / M0;
+
+            return std::min(M1, static_cast<index_t>(VectorLoadSize / sizeof(CDataType)));
+        }
+        else
+        {
+            constexpr index_t M1 = kBlockSize / get_warp_size();
+            constexpr index_t M2 = std::min(BlockGemmShape::kM / M1, get_warp_size());
+            constexpr index_t N0 = get_warp_size() / M2;
+            constexpr index_t N1 = BlockGemmShape::kN / N0;
+
+            return std::min(N1, static_cast<index_t>(VectorLoadSize / sizeof(CDataType)));
+        }
+    }
+
+    static constexpr index_t VectorSizeA = []() {
+        if constexpr(std::is_same_v<ALayout, tensor_layout::gemm::RowMajor>)
+        {
+            return kPadK ? 1 : GetAlignmentA();
+        }
+        else
+        {
+            return kPadM ? 1 : GetAlignmentA();
+        }
+    }();
+
+    static constexpr index_t VectorSizeB = []() {
+        if constexpr(std::is_same_v<BLayout, tensor_layout::gemm::ColumnMajor>)
+        {
+            return kPadN ? 1 : GetAlignmentB();
+        }
+        else
+        {
+            return kPadK ? 1 : GetAlignmentB();
+        }
+    }();
+
+    static constexpr index_t VectorSizeC = []() {
+        if constexpr(std::is_same_v<CLayout, tensor_layout::gemm::RowMajor>)
+        {
+            return kPadN ? 1 : GetAlignmentC();
+        }
+        else
+        {
+            return kPadM ? 1 : GetAlignmentC();
+        }
+    }();
 };
 
+// Alias for GemmPipelineProblem
+template <typename ADataType_,
+          typename BDataType_,
+          typename CDataType_,
+          typename BlockGemmShape_,
+          typename TileGemmTraits_>
+using GemmPipelineProblem =
+    GemmPipelineProblemBase<ADataType_, BDataType_, CDataType_, BlockGemmShape_, TileGemmTraits_>;
+
 template <typename ADataType_,
           typename BDataType_,
           typename CDataType_,
@@ -45,30 +140,15 @@ template <typename ADataType_,
           GemmPipelineScheduler Scheduler_ = GemmPipelineScheduler::Intrawave,
           bool HasHotLoop_                 = true,
           TailNumber TailNum_              = TailNumber::Full>
-struct UniversalGemmPipelineProblem
+struct UniversalGemmPipelineProblem : public GemmPipelineProblemBase<ADataType_,
+                                                                     BDataType_,
+                                                                     CDataType_,
+                                                                     BlockGemmShape_,
+                                                                     TileGemmTraits_>
 {
-    using ADataType      = remove_cvref_t<ADataType_>;
-    using BDataType      = remove_cvref_t<BDataType_>;
-    using CDataType      = remove_cvref_t<CDataType_>;
-    using BlockGemmShape = remove_cvref_t<BlockGemmShape_>;
-    using GemmTraits     = remove_cvref_t<TileGemmTraits_>;
-
-    using ALayout = remove_cvref_t<typename GemmTraits::ALayout>;
-    using BLayout = remove_cvref_t<typename GemmTraits::BLayout>;
-    using CLayout = remove_cvref_t<typename GemmTraits::CLayout>;
-
-    static constexpr auto Scheduler     = Scheduler_;
-    static constexpr auto HasHotLoop    = HasHotLoop_;
-    static constexpr auto TailNum       = TailNum_;
-    static constexpr index_t kBlockSize = BlockGemmShape::NumWarps * get_warp_size();
-
-    static constexpr bool kPadA = GemmTraits::kPadA;
-    static constexpr bool kPadB = GemmTraits::kPadB;
-    static constexpr bool kPadC = GemmTraits::kPadC;
-
-    static constexpr index_t VectorSizeA = kPadA ? _VectorSize / sizeof(ADataType) : 1;
-    static constexpr index_t VectorSizeB = kPadB ? _VectorSize / sizeof(BDataType) : 1;
-    static constexpr index_t VectorSizeC = kPadC ? _VectorSize / sizeof(CDataType) : 1;
+    static constexpr auto Scheduler  = Scheduler_;
+    static constexpr auto HasHotLoop = HasHotLoop_;
+    static constexpr auto TailNum    = TailNum_;
 };
 
 } // namespace ck_tile

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

@@ -9,12 +9,8 @@
 namespace ck_tile {
 
 // UniversalGemm Policy
-template <typename LayoutA_, typename LayoutB_, typename LayoutC_>
 struct UniversalGemmPipelineAgBgCrPolicy
 {
-    using LayoutA = remove_cvref_t<LayoutA_>;
-    using LayoutB = remove_cvref_t<LayoutB_>;
-    using LayoutC = remove_cvref_t<LayoutC_>;
 
     static constexpr auto I0 = number<0>{};
     static constexpr auto I1 = number<1>{};
@@ -22,286 +18,136 @@ struct UniversalGemmPipelineAgBgCrPolicy
 
     static constexpr bool TransposeC = true;
 
+    template <typename Problem, typename DataType, index_t MNPerBlock>
+    CK_TILE_HOST_DEVICE static constexpr auto GetVectorLoadSize()
+    {
+        constexpr index_t BlockSize           = Problem::kBlockSize;
+        constexpr index_t KPerBlock           = Problem::BlockGemmShape::kK;
+        constexpr index_t elements_per_thread = MNPerBlock * KPerBlock / BlockSize;
+
+        if constexpr(elements_per_thread % (16 / sizeof(DataType)) == 0)
+        {
+            return (16 / sizeof(DataType));
+        }
+        else if constexpr(elements_per_thread % (8 / sizeof(DataType)) == 0)
+        {
+            return (8 / sizeof(DataType));
+        }
+        else if constexpr(elements_per_thread % (4 / sizeof(DataType)) == 0 &&
+                          sizeof(DataType) >= 4)
+        {
+            return (4 / sizeof(DataType));
+        }
+        else if constexpr(elements_per_thread % (2 / sizeof(DataType)) == 0 &&
+                          sizeof(DataType) >= 2)
+        {
+            return (2 / sizeof(DataType));
+        }
+        else
+        {
+            return 1;
+        }
+    }
+
     template <typename Problem>
     CK_TILE_HOST_DEVICE static constexpr auto MakeALdsBlockDescriptor()
     {
-        using WarpGemm = WarpGemmMfmaDispatcher<typename Problem::ADataType,
-                                                typename Problem::BDataType,
-                                                typename Problem::CDataType,
-                                                Problem::BlockGemmShape::WarpTile::at(I0),
-                                                Problem::BlockGemmShape::WarpTile::at(I1),
-                                                Problem::BlockGemmShape::WarpTile::at(I2),
-                                                TransposeC>;
 
         using ADataType = remove_cvref_t<typename Problem::ADataType>;
 
         constexpr index_t MPerBlock = Problem::BlockGemmShape::kM;
         constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
-        constexpr index_t K1        = WarpGemm::kK;
-        constexpr index_t K0        = KPerBlock / K1;
-
-        if constexpr(std::is_same<tensor_layout::gemm::RowMajor, LayoutA>::value)
-        {
-            constexpr auto MLdsLayer        = 32 * 4 / KPerBlock / sizeof(ADataType) < 1
-                                                  ? 1
-                                                  : 32 * 4 / KPerBlock / sizeof(ADataType);
-            constexpr auto a_lds_block_desc = make_naive_tensor_descriptor(
-                make_tuple(K0 * number<MLdsLayer>{}, number<MPerBlock / MLdsLayer>{}, K1),
-                make_tuple(K1, number<KPerBlock * MLdsLayer>{}, I1));
-
-            constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
-                a_lds_block_desc,
-                make_tuple(make_xor_transform(make_tuple(number<MPerBlock / MLdsLayer>{},
-                                                         number<K0 * MLdsLayer>{})),
-                           make_pass_through_transform(K1)),
-                make_tuple(sequence<1, 0>{}, sequence<2>{}),
-                make_tuple(sequence<1, 0>{}, sequence<2>{}));
-
-            constexpr auto a_lds_block_desc_ak0_kMLdsLayer_m_ak1 = transform_tensor_descriptor(
-                a_lds_block_desc_permuted,
-                make_tuple(make_unmerge_transform(make_tuple(K0, number<MLdsLayer>{})),
-                           make_pass_through_transform(number<MPerBlock / MLdsLayer>{}),
-                           make_pass_through_transform(K1)),
-                make_tuple(sequence<0>{}, sequence<1>{}, sequence<2>{}),
-                make_tuple(sequence<0, 2>{}, sequence<1>{}, sequence<3>{}));
-
-            constexpr auto a_lds_block_desc_m_k = transform_tensor_descriptor(
-                a_lds_block_desc_ak0_kMLdsLayer_m_ak1,
-                make_tuple(make_merge_transform_v3_division_mod(make_tuple(K0, K1)),
-                           make_merge_transform_v3_division_mod(
-                               make_tuple(number<MPerBlock / MLdsLayer>{}, number<MLdsLayer>{}))),
-                make_tuple(sequence<0, 3>{}, sequence<1, 2>{}),
-                make_tuple(sequence<1>{}, sequence<0>{}));
-
-            return a_lds_block_desc_m_k;
-        }
-        else // ColumnMajor A
-        {
-            // kfold and mpair dimension is not always required.
-            // more dimension in merge_transform increase the difficulty of generating immarg offset
-            // for compiler.
-            constexpr auto M0 = get_warp_size() * Problem::BlockGemmShape::BlockWarps::at(I0);
-            constexpr auto M1 = MPerBlock / M0;
-
-            constexpr auto KThreadWrite     = Problem::kBlockSize / M0;
-            constexpr auto K0PerThreadWrite = K0 / KThreadWrite;
-            constexpr auto KThreadRead      = 64 / WarpGemm::kM;
-            constexpr auto K0PerThreadRead  = K0 / KThreadRead;
-
-            constexpr auto kfold =
-                (K1 * M0 * sizeof(ADataType) > 128) ? 1 : 128 / (K1 * M0 * sizeof(ADataType));
-            constexpr auto KThreadReadPerm =
-                (kfold * K0PerThreadWrite / K0PerThreadRead) > 1
-                    ? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead)
-                    : KThreadRead;
-
-            // 1<=mpair<=kN0
-            constexpr auto mpair = (K1 * WarpGemm::kM * sizeof(ADataType) > 128)
-                                       ? 1
-                                       : ((128 / (K1 * WarpGemm::kM * sizeof(ADataType))) > M0
-                                              ? M0
-                                              : 128 / (K1 * WarpGemm::kM * sizeof(ADataType)));
-
-            constexpr auto a_lds_block_desc = make_naive_tensor_descriptor_packed(
-                make_tuple(number<KThreadWrite / kfold / KThreadReadPerm>{},
-                           number<K0PerThreadWrite>{},
-                           number<KThreadReadPerm * M1>{},
-                           number<kfold * M0 / mpair>{},
-                           number<mpair>{},
-                           K1));
-
-            constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
-                a_lds_block_desc,
-                make_tuple(
-                    make_pass_through_transform(number<KThreadWrite / kfold / KThreadReadPerm>{}),
-                    make_pass_through_transform(number<K0PerThreadWrite>{}),
-                    make_xor_transform(
-                        make_tuple(number<KThreadReadPerm * M1>{}, number<kfold * M0 / mpair>{})),
-                    make_pass_through_transform(number<mpair>{}),
-                    make_pass_through_transform(K1)),
-                make_tuple(
-                    sequence<0>{}, sequence<1>{}, sequence<2, 3>{}, sequence<4>{}, sequence<5>{}),
-                make_tuple(
-                    sequence<0>{}, sequence<1>{}, sequence<2, 3>{}, sequence<4>{}, sequence<5>{}));
-
-            constexpr auto a_lds_block_desc_unmerged = transform_tensor_descriptor(
-                a_lds_block_desc_permuted,
-                make_tuple(
-                    make_pass_through_transform(number<KThreadWrite / kfold / KThreadReadPerm>{}),
-                    make_pass_through_transform(number<K0PerThreadWrite>{}),
-                    make_unmerge_transform(make_tuple(number<KThreadReadPerm>{}, number<M1>{})),
-                    make_unmerge_transform(make_tuple(number<kfold>{}, number<M0 / mpair>{})),
-                    make_pass_through_transform(number<mpair>{}),
-                    make_pass_through_transform(K1)),
-                make_tuple(sequence<0>{},
-                           sequence<1>{},
-                           sequence<2>{},
-                           sequence<3>{},
-                           sequence<4>{},
-                           sequence<5>{}),
-                make_tuple(sequence<1>{},
-                           sequence<2>{},
-                           sequence<0, 3>{},
-                           sequence<4, 5>{},
-                           sequence<6>{},
-                           sequence<7>{}));
-
-            constexpr auto a_lds_block_desc_m_k = transform_tensor_descriptor(
-                a_lds_block_desc_unmerged,
-                make_tuple(make_merge_transform_v3_division_mod(
-                               make_tuple(number<KThreadReadPerm>{},
-                                          number<KThreadWrite / kfold / KThreadReadPerm>{},
-                                          number<kfold>{},
-                                          number<K0PerThreadWrite>{},
-                                          K1)),
-                           make_merge_transform_v3_division_mod(
-                               make_tuple(number<M0 / mpair>{}, number<mpair>{}, number<M1>{}))),
-                make_tuple(sequence<0, 1, 4, 2, 7>{}, sequence<5, 6, 3>{}),
-                make_tuple(sequence<1>{}, sequence<0>{}));
-
-            return a_lds_block_desc_m_k;
-        }
+        constexpr index_t KPack     = GetVectorLoadSize<Problem, ADataType, MPerBlock>();
+
+        constexpr auto DataTypeSize = sizeof(ADataType);
+        constexpr auto MLdsLayer =
+            (32 * 4 / KPerBlock / DataTypeSize) < 1 ? 1 : (32 * 4 / KPerBlock / DataTypeSize);
+
+        constexpr auto a_lds_block_desc_0 = make_naive_tensor_descriptor(
+            make_tuple(number<KPerBlock / KPack * MLdsLayer>{},
+                       number<MPerBlock / MLdsLayer>{},
+                       number<KPack>{}),
+            make_tuple(number<KPack>{}, number<KPerBlock * MLdsLayer>{}, number<1>{}),
+            number<KPack>{},
+            number<1>{});
+
+        constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
+            a_lds_block_desc_0,
+            make_tuple(make_xor_transform(make_tuple(number<MPerBlock / MLdsLayer>{},
+                                                     number<KPerBlock / KPack * MLdsLayer>{})),
+                       make_pass_through_transform(number<KPack>{})),
+            make_tuple(sequence<1, 0>{}, sequence<2>{}),
+            make_tuple(sequence<1, 0>{}, sequence<2>{}));
+
+        constexpr auto a_lds_block_desc_xk0_mnldslayer_mn_xk1 = transform_tensor_descriptor(
+            a_lds_block_desc_permuted,
+            make_tuple(make_unmerge_transform(
+                           make_tuple(number<KPerBlock / KPack>{}, number<MLdsLayer>{})),
+                       make_pass_through_transform(number<MPerBlock / MLdsLayer>{}),
+                       make_pass_through_transform(number<KPack>{})),
+            make_tuple(sequence<0>{}, sequence<1>{}, sequence<2>{}),
+            make_tuple(sequence<0, 2>{}, sequence<1>{}, sequence<3>{}));
+
+        constexpr auto a_lds_block_desc = transform_tensor_descriptor(
+            a_lds_block_desc_xk0_mnldslayer_mn_xk1,
+            make_tuple(make_merge_transform_v3_division_mod(
+                           make_tuple(number<MPerBlock / MLdsLayer>{}, number<MLdsLayer>{})),
+                       make_merge_transform_v3_division_mod(
+                           make_tuple(number<KPerBlock / KPack>{}, number<KPack>{}))),
+            make_tuple(sequence<1, 2>{}, sequence<0, 3>{}),
+            make_tuple(sequence<0>{}, sequence<1>{}));
+
+        return a_lds_block_desc;
     }
 
     template <typename Problem>
     CK_TILE_HOST_DEVICE static constexpr auto MakeBLdsBlockDescriptor()
     {
-        using WarpGemm = WarpGemmMfmaDispatcher<typename Problem::ADataType,
-                                                typename Problem::BDataType,
-                                                typename Problem::CDataType,
-                                                Problem::BlockGemmShape::WarpTile::at(I0),
-                                                Problem::BlockGemmShape::WarpTile::at(I1),
-                                                Problem::BlockGemmShape::WarpTile::at(I2),
-                                                TransposeC>;
 
         using BDataType = remove_cvref_t<typename Problem::BDataType>;
 
         constexpr index_t NPerBlock = Problem::BlockGemmShape::kN;
         constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
-
-        constexpr index_t K1 = WarpGemm::kK;
-        constexpr index_t K0 = KPerBlock / K1;
-
-        if constexpr(std::is_same<tensor_layout::gemm::ColumnMajor, LayoutB>::value)
-        {
-            // NLdsLayer * K0 as logical Bank
-            constexpr auto NLdsLayer = 32 * 4 / KPerBlock / sizeof(BDataType) < 1
-                                           ? 1
-                                           : 32 * 4 / KPerBlock / sizeof(BDataType);
-            ;
-            constexpr auto b_lds_block_desc = make_naive_tensor_descriptor(
-                make_tuple(K0 * number<NLdsLayer>{}, number<NPerBlock / NLdsLayer>{}, K1),
-                make_tuple(K1, number<KPerBlock * NLdsLayer>{}, I1));
-
-            constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor(
-                b_lds_block_desc,
-                make_tuple(make_xor_transform(make_tuple(number<NPerBlock / NLdsLayer>{},
-                                                         number<K0 * NLdsLayer>{})),
-                           make_pass_through_transform(K1)),
-                make_tuple(sequence<1, 0>{}, sequence<2>{}),
-                make_tuple(sequence<1, 0>{}, sequence<2>{}));
-
-            constexpr auto b_lds_block_desc_bk0_kNLdsLayer_n_bk1 = transform_tensor_descriptor(
-                b_lds_block_desc_permuted,
-                make_tuple(make_unmerge_transform(make_tuple(K0, number<NLdsLayer>{})),
-                           make_pass_through_transform(number<NPerBlock / NLdsLayer>{}),
-                           make_pass_through_transform(K1)),
-                make_tuple(sequence<0>{}, sequence<1>{}, sequence<2>{}),
-                make_tuple(sequence<0, 2>{}, sequence<1>{}, sequence<3>{}));
-
-            constexpr auto b_lds_block_desc_n_k = transform_tensor_descriptor(
-                b_lds_block_desc_bk0_kNLdsLayer_n_bk1,
-                make_tuple(make_merge_transform_v3_division_mod(make_tuple(K0, K1)),
-                           make_merge_transform_v3_division_mod(
-                               make_tuple(number<NPerBlock / NLdsLayer>{}, number<NLdsLayer>{}))),
-                make_tuple(sequence<0, 3>{}, sequence<1, 2>{}),
-                make_tuple(sequence<1>{}, sequence<0>{}));
-
-            return b_lds_block_desc_n_k;
-        }
-        else // RowMajor B
-        {
-            constexpr auto N0 = get_warp_size() * Problem::BlockGemmShape::BlockWarps::at(I1);
-            constexpr auto N1 = NPerBlock / N0;
-
-            constexpr auto KThreadWrite     = Problem::kBlockSize / N0;
-            constexpr auto K0PerThreadWrite = K0 / KThreadWrite;
-            constexpr auto KThreadRead      = 64 / WarpGemm::kN;
-            constexpr auto K0PerThreadRead  = K0 / KThreadRead;
-
-            constexpr auto kfold =
-                (K1 * N0 * sizeof(BDataType) > 128) ? 1 : 128 / (K1 * N0 * sizeof(BDataType));
-            constexpr auto KThreadReadPerm =
-                (kfold * K0PerThreadWrite / K0PerThreadRead) > 1
-                    ? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead)
-                    : KThreadRead;
-
-            // 1<=npair<=kN0
-            constexpr auto npair = (K1 * WarpGemm::kN * sizeof(BDataType) > 128)
-                                       ? 1
-                                       : ((128 / (K1 * WarpGemm::kN * sizeof(BDataType))) > N0
-                                              ? N0
-                                              : 128 / (K1 * WarpGemm::kN * sizeof(BDataType)));
-
-            constexpr auto b_lds_block_desc = make_naive_tensor_descriptor_packed(
-                make_tuple(number<KThreadWrite / kfold / KThreadReadPerm>{},
-                           number<K0PerThreadWrite>{},
-                           number<KThreadReadPerm * N1>{},
-                           number<kfold * N0 / npair>{},
-                           number<npair>{},
-                           K1));
-
-            constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor(
-                b_lds_block_desc,
-                make_tuple(
-                    make_pass_through_transform(number<KThreadWrite / kfold / KThreadReadPerm>{}),
-                    make_pass_through_transform(number<K0PerThreadWrite>{}),
-                    make_xor_transform(
-                        make_tuple(number<KThreadReadPerm * N1>{}, number<kfold * N0 / npair>{})),
-                    make_pass_through_transform(number<npair>{}),
-                    make_pass_through_transform(K1)),
-                make_tuple(
-                    sequence<0>{}, sequence<1>{}, sequence<2, 3>{}, sequence<4>{}, sequence<5>{}),
-                make_tuple(
-                    sequence<0>{}, sequence<1>{}, sequence<2, 3>{}, sequence<4>{}, sequence<5>{}));
-
-            constexpr auto b_lds_block_desc_unmerged = transform_tensor_descriptor(
-                b_lds_block_desc_permuted,
-                make_tuple(
-                    make_pass_through_transform(number<KThreadWrite / kfold / KThreadReadPerm>{}),
-                    make_pass_through_transform(number<K0PerThreadWrite>{}),
-                    make_unmerge_transform(make_tuple(number<KThreadReadPerm>{}, number<N1>{})),
-                    make_unmerge_transform(make_tuple(number<kfold>{}, number<N0 / npair>{})),
-                    make_pass_through_transform(number<npair>{}),
-                    make_pass_through_transform(K1)),
-                make_tuple(sequence<0>{},
-                           sequence<1>{},
-                           sequence<2>{},
-                           sequence<3>{},
-                           sequence<4>{},
-                           sequence<5>{}),
-                make_tuple(sequence<1>{},
-                           sequence<2>{},
-                           sequence<0, 3>{},
-                           sequence<4, 5>{},
-                           sequence<6>{},
-                           sequence<7>{}));
-
-            constexpr auto b_lds_block_desc_n_k = transform_tensor_descriptor(
-                b_lds_block_desc_unmerged,
-                make_tuple(make_merge_transform_v3_division_mod(
-                               make_tuple(number<KThreadReadPerm>{},
-                                          number<KThreadWrite / kfold / KThreadReadPerm>{},
-                                          number<kfold>{},
-                                          number<K0PerThreadWrite>{},
-                                          K1)),
-                           make_merge_transform_v3_division_mod(
-                               make_tuple(number<N0 / npair>{}, number<npair>{}, number<N1>{}))),
-                make_tuple(sequence<0, 1, 4, 2, 7>{}, sequence<5, 6, 3>{}),
-                make_tuple(sequence<1>{}, sequence<0>{}));
-
-            return b_lds_block_desc_n_k;
-        }
+        constexpr index_t KPack     = GetVectorLoadSize<Problem, BDataType, NPerBlock>();
+
+        constexpr auto DataTypeSize = sizeof(BDataType);
+        constexpr auto NLdsLayer =
+            (32 * 4 / KPerBlock / DataTypeSize) < 1 ? 1 : (32 * 4 / KPerBlock / DataTypeSize);
+
+        constexpr auto b_lds_block_desc_0 = make_naive_tensor_descriptor(
+            make_tuple(number<KPerBlock / KPack * NLdsLayer>{},
+                       number<NPerBlock / NLdsLayer>{},
+                       number<KPack>{}),
+            make_tuple(number<KPack>{}, number<KPerBlock * NLdsLayer>{}, number<1>{}),
+            number<KPack>{},
+            number<1>{});
+
+        constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor(
+            b_lds_block_desc_0,
+            make_tuple(make_xor_transform(make_tuple(number<NPerBlock / NLdsLayer>{},
+                                                     number<KPerBlock / KPack * NLdsLayer>{})),
+                       make_pass_through_transform(number<KPack>{})),
+            make_tuple(sequence<1, 0>{}, sequence<2>{}),
+            make_tuple(sequence<1, 0>{}, sequence<2>{}));
+
+        constexpr auto b_lds_block_desc_xk0_mnldslayer_mn_xk1 = transform_tensor_descriptor(
+            b_lds_block_desc_permuted,
+            make_tuple(make_unmerge_transform(
+                           make_tuple(number<KPerBlock / KPack>{}, number<NLdsLayer>{})),
+                       make_pass_through_transform(number<NPerBlock / NLdsLayer>{}),
+                       make_pass_through_transform(number<KPack>{})),
+            make_tuple(sequence<0>{}, sequence<1>{}, sequence<2>{}),
+            make_tuple(sequence<0, 2>{}, sequence<1>{}, sequence<3>{}));
+
+        constexpr auto b_lds_block_desc = transform_tensor_descriptor(
+            b_lds_block_desc_xk0_mnldslayer_mn_xk1,
+            make_tuple(make_merge_transform_v3_division_mod(
+                           make_tuple(number<NPerBlock / NLdsLayer>{}, number<NLdsLayer>{})),
+                       make_merge_transform_v3_division_mod(
+                           make_tuple(number<KPerBlock / KPack>{}, number<KPack>{}))),
+            make_tuple(sequence<1, 2>{}, sequence<0, 3>{}),
+            make_tuple(sequence<0>{}, sequence<1>{}));
+        return b_lds_block_desc;
     }
 
     template <typename Problem>
@@ -334,69 +180,268 @@ struct UniversalGemmPipelineAgBgCrPolicy
     template <typename Problem>
     CK_TILE_HOST_DEVICE static constexpr auto MakeADramTileDistribution()
     {
-        using WarpGemm = WarpGemmMfmaDispatcher<typename Problem::ADataType,
-                                                typename Problem::BDataType,
-                                                typename Problem::CDataType,
-                                                Problem::BlockGemmShape::WarpTile::at(I0),
-                                                Problem::BlockGemmShape::WarpTile::at(I1),
-                                                Problem::BlockGemmShape::WarpTile::at(I2),
-                                                TransposeC>;
+        using ADataType = remove_cvref_t<typename Problem::ADataType>;
+        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 K1 = WarpGemm::kK;
-        constexpr index_t K0 = KPerBlock / K1;
-        constexpr index_t M2 = get_warp_size() / K0;
-
-        constexpr index_t M1 = BlockSize / get_warp_size();
-        static_assert(M2 != 0, "M2 is zero, which will lead to a division by zero error.");
-        static_assert(M1 != 0, "M1 is zero, which will lead to a division by zero error.");
-        constexpr index_t M0 = MPerBlock / (M2 * M1);
-
-        return make_static_tile_distribution(
-            tile_distribution_encoding<sequence<1>,
-                                       tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
-                                       tuple<sequence<1>, sequence<1, 2>>,
-                                       tuple<sequence<1>, sequence<2, 0>>,
-                                       sequence<1, 2>,
-                                       sequence<0, 1>>{});
+        if constexpr(std::is_same_v<ALayout, ck_tile::tensor_layout::gemm::ColumnMajor>)
+        {
+            constexpr index_t M1           = Problem::VectorLoadSize / sizeof(ADataType);
+            constexpr index_t M0           = MPerBlock / M1;
+            constexpr index_t total_pixels = MPerBlock * KPerBlock / BlockSize;
+            static_assert(total_pixels % M1 == 0);
+            constexpr index_t K3    = total_pixels / M1;
+            constexpr index_t KPack = GetVectorLoadSize<Problem, ADataType, MPerBlock>();
+            static_assert(KPack % K3 == 0);
+            constexpr index_t K2 = KPack / K3;
+            if constexpr(get_warp_size() % (K2 * M0) == 0)
+            {
+                constexpr index_t K1 = get_warp_size() / (K2 * M0);
+                constexpr index_t K0 = BlockSize / get_warp_size();
+                static_assert(KPerBlock == K0 * K1 * K2 * K3);
+                return make_static_tile_distribution(
+                    tile_distribution_encoding<sequence<1>,
+                                               tuple<sequence<M0, M1>, sequence<K0, K1, K2, K3>>,
+                                               tuple<sequence<2>, sequence<2, 1, 2>>,
+                                               tuple<sequence<0>, sequence<1, 0, 2>>,
+                                               sequence<2, 1>,
+                                               sequence<3, 1>>{});
+            }
+            else
+            {
+                constexpr index_t K1   = (K2 * M0) / get_warp_size();
+                constexpr index_t K2_m = K2 / K1;
+                constexpr index_t K0   = BlockSize / get_warp_size() / K1;
+                static_assert(KPerBlock == K0 * K1 * K2_m * K3);
+                return make_static_tile_distribution(
+                    tile_distribution_encoding<sequence<1>,
+                                               tuple<sequence<M0, M1>, sequence<K0, K1, K2_m, K3>>,
+                                               tuple<sequence<2, 2>, sequence<1, 2>>,
+                                               tuple<sequence<0, 1>, sequence<0, 2>>,
+                                               sequence<2, 1>,
+                                               sequence<3, 1>>{});
+            }
+        }
+        else
+        {
+            constexpr index_t K1 = Problem::VectorLoadSize / sizeof(ADataType);
+            constexpr index_t K0 = KPerBlock / K1;
+            constexpr index_t M2 = get_warp_size() / K0;
+            if constexpr(get_warp_size() % (M2 * K0) == 0)
+            {
+                constexpr index_t M1 = BlockSize / get_warp_size();
+                static_assert(M2 != 0, "M2 is zero, which will lead to a division by zero error.");
+                static_assert(M1 != 0, "M1 is zero, which will lead to a division by zero error.");
+                constexpr index_t M0 = MPerBlock / (M2 * M1);
+                return make_static_tile_distribution(
+                    tile_distribution_encoding<sequence<1>,
+                                               tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
+                                               tuple<sequence<1>, sequence<1, 2>>,
+                                               tuple<sequence<1>, sequence<2, 0>>,
+                                               sequence<1, 2>,
+                                               sequence<0, 1>>{});
+            }
+            else
+            {
+                constexpr index_t M0 = BlockSize / get_warp_size();
+                constexpr index_t M1 = MPerBlock / (M2 * M0);
+                return make_static_tile_distribution(
+                    tile_distribution_encoding<sequence<1>,
+                                               tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
+                                               tuple<sequence<1>, sequence<1, 2>>,
+                                               tuple<sequence<0>, sequence<2, 0>>,
+                                               sequence<1, 2>,
+                                               sequence<1, 1>>{});
+            }
+        }
     }
 
     template <typename Problem>
     CK_TILE_HOST_DEVICE static constexpr auto MakeBDramTileDistribution()
     {
-        using WarpGemm = WarpGemmMfmaDispatcher<typename Problem::ADataType,
-                                                typename Problem::BDataType,
-                                                typename Problem::CDataType,
-                                                Problem::BlockGemmShape::WarpTile::at(I0),
-                                                Problem::BlockGemmShape::WarpTile::at(I1),
-                                                Problem::BlockGemmShape::WarpTile::at(I2),
-                                                TransposeC>;
+        using BDataType = remove_cvref_t<typename Problem::BDataType>;
+        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;
 
+        if constexpr(std::is_same_v<BLayout, ck_tile::tensor_layout::gemm::RowMajor>)
+        {
+            constexpr index_t N1           = Problem::VectorLoadSize / sizeof(BDataType);
+            constexpr index_t N0           = NPerBlock / N1;
+            constexpr index_t total_pixels = NPerBlock * KPerBlock / BlockSize;
+            static_assert(total_pixels % N1 == 0);
+            constexpr index_t K3    = total_pixels / N1;
+            constexpr index_t KPack = GetVectorLoadSize<Problem, BDataType, NPerBlock>();
+            static_assert(KPack % K3 == 0);
+            constexpr index_t K2 = KPack / K3;
+            if constexpr(get_warp_size() % (K2 * N0) == 0)
+            {
+                constexpr index_t K1 = get_warp_size() / (K2 * N0);
+                constexpr index_t K0 = BlockSize / get_warp_size();
+                static_assert(KPerBlock == K0 * K1 * K2 * K3);
+                return make_static_tile_distribution(
+                    tile_distribution_encoding<sequence<1>,
+                                               tuple<sequence<N0, N1>, sequence<K0, K1, K2, K3>>,
+                                               tuple<sequence<2>, sequence<2, 1, 2>>,
+                                               tuple<sequence<0>, sequence<1, 0, 2>>,
+                                               sequence<2, 1>,
+                                               sequence<3, 1>>{});
+            }
+            else
+            {
+                constexpr index_t K1   = (K2 * N0) / get_warp_size();
+                constexpr index_t K2_m = K2 / K1;
+                constexpr index_t K0   = BlockSize / get_warp_size() / K1;
+                static_assert(KPerBlock == K0 * K1 * K2_m * K3);
+                return make_static_tile_distribution(
+                    tile_distribution_encoding<sequence<1>,
+                                               tuple<sequence<N0, N1>, sequence<K0, K1, K2_m, K3>>,
+                                               tuple<sequence<2, 2>, sequence<1, 2>>,
+                                               tuple<sequence<0, 1>, sequence<0, 2>>,
+                                               sequence<2, 1>,
+                                               sequence<3, 1>>{});
+            }
+        }
+        else
+        {
+
+            constexpr index_t K1 = Problem::VectorLoadSize / sizeof(BDataType);
+            constexpr index_t K0 = KPerBlock / K1;
+            constexpr index_t N2 = get_warp_size() / K0;
+            // coalesce reading for each blocks
+            if constexpr(get_warp_size() % (N2 * K0) == 0)
+            {
+                constexpr index_t N1 = BlockSize / get_warp_size();
+                static_assert(N2 != 0, "N2 is zero, which will lead to a division by zero error.");
+                static_assert(N1 != 0, "N1 is zero, which will lead to a division by zero error.");
+                constexpr index_t N0 = NPerBlock / (N2 * N1);
+
+                return make_static_tile_distribution(
+                    tile_distribution_encoding<sequence<1>,
+                                               tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
+                                               tuple<sequence<1>, sequence<1, 2>>,
+                                               tuple<sequence<1>, sequence<2, 0>>,
+                                               sequence<1, 2>,
+                                               sequence<0, 1>>{});
+            }
+            // coalesce reading for each warps
+            else
+            {
+                constexpr index_t N0 = BlockSize / get_warp_size();
+                constexpr index_t N1 = NPerBlock / (N2 * N0);
+
+                return make_static_tile_distribution(
+                    tile_distribution_encoding<sequence<1>,
+                                               tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
+                                               tuple<sequence<1>, sequence<1, 2>>,
+                                               tuple<sequence<0>, sequence<2, 0>>,
+                                               sequence<1, 2>,
+                                               sequence<1, 1>>{});
+            }
+        }
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledARegBlockDescriptor()
+    {
+        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::ColumnMajor>);
         constexpr index_t BlockSize = Problem::kBlockSize;
+        constexpr index_t MPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr index_t M1           = Problem::VectorLoadSize / sizeof(ADataType);
+        constexpr index_t M0           = MPerBlock / M1;
+        constexpr index_t total_pixels = MPerBlock * KPerBlock / BlockSize;
+        static_assert(total_pixels % M1 == 0);
+        constexpr index_t K3     = total_pixels / M1;
+        constexpr index_t kKPack = GetVectorLoadSize<Problem, ADataType, MPerBlock>();
+        static_assert(kKPack % K3 == 0);
+        constexpr index_t K2 = kKPack / K3; // TODO: this dimention could be outside single wave
+        constexpr index_t warp_size = get_warp_size();
+        if constexpr(warp_size % (K2 * M0) == 0)
+        {
+            constexpr index_t K1 = warp_size / (K2 * M0);
+            constexpr index_t K0 = BlockSize / warp_size;
+
+            return make_static_tile_distribution(
+                tile_distribution_encoding<sequence<1>,
+                                           tuple<sequence<M0, M1>, sequence<K0, K1, K2, K3>>,
+                                           tuple<sequence<2>, sequence<2, 1, 2>>,
+                                           tuple<sequence<0>, sequence<1, 0, 2>>,
+                                           sequence<1, 2>,
+                                           sequence<1, 3>>{});
+        }
+        else
+        {
+            constexpr index_t K1   = (K2 * M0) / get_warp_size();
+            constexpr index_t K2_m = K2 / K1;
+            constexpr index_t K0   = BlockSize / get_warp_size() / K1;
+            static_assert(KPerBlock == K0 * K1 * K2_m * K3);
+            return make_static_tile_distribution(
+                tile_distribution_encoding<sequence<1>,
+                                           tuple<sequence<M0, M1>, sequence<K0, K1, K2_m, K3>>,
+                                           tuple<sequence<2, 2>, sequence<1, 2>>,
+                                           tuple<sequence<0, 1>, sequence<0, 2>>,
+                                           sequence<1, 2>,
+                                           sequence<1, 3>>{});
+        }
+    }
 
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeShuffledBRegBlockDescriptor()
+    {
+        using BLayout   = remove_cvref_t<typename Problem::BLayout>;
+        using BDataType = remove_cvref_t<typename Problem::BDataType>;
+        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 K1 = WarpGemm::kK;
-        constexpr index_t K0 = KPerBlock / K1;
-        constexpr index_t N2 = get_warp_size() / K0;
-
-        constexpr index_t N1 = BlockSize / get_warp_size();
-        static_assert(N2 != 0, "M2 is zero, which will lead to a division by zero error.");
-        static_assert(N1 != 0, "M1 is zero, which will lead to a division by zero error.");
-        constexpr index_t N0 = NPerBlock / (N2 * N1);
-
-        return make_static_tile_distribution(
-            tile_distribution_encoding<sequence<1>,
-                                       tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
-                                       tuple<sequence<1>, sequence<1, 2>>,
-                                       tuple<sequence<1>, sequence<2, 0>>,
-                                       sequence<1, 2>,
-                                       sequence<0, 1>>{});
+        constexpr index_t N1           = Problem::VectorLoadSize / sizeof(BDataType);
+        constexpr index_t N0           = NPerBlock / N1;
+        constexpr index_t total_pixels = NPerBlock * KPerBlock / BlockSize;
+        static_assert(total_pixels % N1 == 0);
+        constexpr index_t K3     = total_pixels / N1;
+        constexpr index_t kKPack = GetVectorLoadSize<Problem, BDataType, NPerBlock>();
+        static_assert(kKPack % K3 == 0);
+        constexpr index_t K2 = kKPack / K3; // TODO: this dimention could be outside single wave
+        constexpr index_t warp_size = get_warp_size();
+        if constexpr(warp_size % (K2 * N0) == 0)
+        {
+            constexpr index_t K1 = warp_size / (K2 * N0);
+            constexpr index_t K0 = BlockSize / warp_size;
+
+            return make_static_tile_distribution(
+                tile_distribution_encoding<sequence<1>,
+                                           tuple<sequence<N0, N1>, sequence<K0, K1, K2, K3>>,
+                                           tuple<sequence<2>, sequence<2, 1, 2>>,
+                                           tuple<sequence<0>, sequence<1, 0, 2>>,
+                                           sequence<1, 2>,
+                                           sequence<1, 3>>{});
+        }
+        else
+        {
+            constexpr index_t K1   = (K2 * N0) / get_warp_size();
+            constexpr index_t K2_m = K2 / K1;
+            constexpr index_t K0   = BlockSize / get_warp_size() / K1;
+            static_assert(KPerBlock == K0 * K1 * K2_m * K3);
+            return make_static_tile_distribution(
+                tile_distribution_encoding<sequence<1>,
+                                           tuple<sequence<N0, N1>, sequence<K0, K1, K2_m, K3>>,
+                                           tuple<sequence<2, 2>, sequence<1, 2>>,
+                                           tuple<sequence<0, 1>, sequence<0, 2>>,
+                                           sequence<1, 2>,
+                                           sequence<1, 3>>{});
+        }
     }
 
     template <typename Problem>

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

@@ -3,19 +3,23 @@
 
 #pragma once
 
+#include "ck_tile/core.hpp"
+
 namespace ck_tile {
 
-template <bool kPadA_,
-          bool kPadB_,
-          bool kPadC_,
+template <bool kPadM_,
+          bool kPadN_,
+          bool kPadK_,
           typename ALayout_,
           typename BLayout_,
           typename CLayout_>
 struct TileGemmTraits
 {
-    static constexpr bool kPadA = kPadA_;
-    static constexpr bool kPadB = kPadB_;
-    static constexpr bool kPadC = kPadC_;
+    static constexpr bool kPadM = kPadM_;
+    static constexpr bool kPadN = kPadN_;
+    static constexpr bool kPadK = kPadK_;
+
+    static constexpr int _VectorSize = 16;
 
     using ALayout = ALayout_;
     using BLayout = BLayout_;

include/ck_tile/ops/gemm/warp/warp_gemm.hpp

@@ -10,114 +10,150 @@
 namespace ck_tile {
 
 // fp16
-using WarpGemmMfmaF16F16F32M32N32K8 =
-    WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplF16F16F32M32N32K8>>;
 
-using WarpGemmMfmaF16F16F32M16N16K16 =
-    WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplF16F16F32M16N16K16>>;
+using WarpGemmMfmaF16F16F32M32N32K8 = WarpGemmImpl<
+    WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
 
-using WarpGemmMfmaF16F16F32M32N32K16 =
-    WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<WarpGemmAttributeMfmaImplF16F16F32M32N32K8, 2>>;
+using WarpGemmMfmaF16F16F32M16N16K16 = WarpGemmImpl<
+    WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
 
-using WarpGemmMfmaF16F16F32M16N16K32 =
-    WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<WarpGemmAttributeMfmaImplF16F16F32M16N16K16, 2>>;
+using WarpGemmMfmaF16F16F32M32N32K16 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
+    WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
+    2>>;
 
-using WarpGemmMfmaF16F16F32M32N32K8SwizzleA = WarpGemmImpl<
-    WarpGemmAtrributeMfmaIterateK_SwizzleA<WarpGemmAttributeMfmaImplF16F16F32M32N32K8, 1>>;
+using WarpGemmMfmaF16F16F32M16N16K32 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
+    WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>,
+    2>>;
 
-using WarpGemmMfmaF16F16F32M32N32K16SwizzleA = WarpGemmImpl<
-    WarpGemmAtrributeMfmaIterateK_SwizzleA<WarpGemmAttributeMfmaImplF16F16F32M32N32K8, 2>>;
+using WarpGemmMfmaF16F16F32M32N32K8SwizzleA = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK_SwizzleA<
+    WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
+    1>>;
 
-using WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution = WarpGemmImpl<
-    WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImplF16F16F32M32N32K8>>;
+using WarpGemmMfmaF16F16F32M32N32K16SwizzleA = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK_SwizzleA<
+    WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
+    2>>;
 
-using WarpGemmMfmaF16F16F32M16N16K16TransposedCDistribution = WarpGemmImpl<
-    WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImplF16F16F32M16N16K16>>;
+using WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution =
+    WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
+        WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
+
+using WarpGemmMfmaF16F16F32M16N16K16TransposedCDistribution =
+    WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
+        WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
 
 using WarpGemmMfmaF16F16F32M32N32K16TransposedCDistribution =
     WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution<
-        WarpGemmAttributeMfmaImplF16F16F32M32N32K8,
+        WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
         2>>;
 
 using WarpGemmMfmaF16F16F32M16N16K32TransposedCDistribution =
     WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution<
-        WarpGemmAttributeMfmaImplF16F16F32M16N16K16,
+        WarpGemmAttributeMfmaImplF16F16F32M16N16K16<WGAttrCtlEnum::Default_>,
         2>>;
 
 using WarpGemmMfmaF16F16F32M32N32K16SwizzleBTransposedCDistribution =
     WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB<
-        WarpGemmAttributeMfmaImplF16F16F32M32N32K8,
+        WarpGemmAttributeMfmaImplF16F16F32M32N32K8<WGAttrCtlEnum::Default_>,
         2>>;
 
+using WarpGemmMfmaF16F16F32M4N64K16 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
+    WarpGemmAttributeMfmaImplF16F16F32M4N64K4<WGAttrCtlEnum::Default_>,
+    4>>;
+
+using WarpGemmMfmaF16F16F32M64N4K16 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
+    WarpGemmAttributeMfmaImplF16F16F32M64N4K4<WGAttrCtlEnum::Default_>,
+    4>>;
+
 // bf16
-using WarpGemmMfmaBf16Bf16F32M32N32K8 =
-    WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8>>;
 
-using WarpGemmMfmaBf16Bf16F32M16N16K16 =
-    WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16>>;
+using WarpGemmMfmaBf16Bf16F32M32N32K8 = WarpGemmImpl<
+    WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
+
+using WarpGemmMfmaBf16Bf16F32M16N16K16 = WarpGemmImpl<
+    WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
 
-using WarpGemmMfmaBf16Bf16F32M32N32K16 =
-    WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8, 2>>;
+using WarpGemmMfmaBf16Bf16F32M32N32K16 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
+    WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
+    2>>;
 
-using WarpGemmMfmaBf16Bf16F32M16N16K32 =
-    WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16, 2>>;
+using WarpGemmMfmaBf16Bf16F32M16N16K32 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
+    WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>,
+    2>>;
 
-using WarpGemmMfmaBf16Bf16F32M32N32K8SwizzleA = WarpGemmImpl<
-    WarpGemmAtrributeMfmaIterateK_SwizzleA<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8, 1>>;
+using WarpGemmMfmaBf16Bf16F32M32N32K8SwizzleA = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK_SwizzleA<
+    WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
+    1>>;
 
-using WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleA = WarpGemmImpl<
-    WarpGemmAtrributeMfmaIterateK_SwizzleA<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8, 2>>;
+using WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleA =
+    WarpGemmImpl<WarpGemmAtrributeMfmaIterateK_SwizzleA<
+        WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
+        2>>;
 
-using WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution = WarpGemmImpl<
-    WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8>>;
+using WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution =
+    WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
+        WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>>>;
 
-using WarpGemmMfmaBf16Bf16F32M16N16K16TransposedCDistribution = WarpGemmImpl<
-    WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16>>;
+using WarpGemmMfmaBf16Bf16F32M16N16K16TransposedCDistribution =
+    WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
+        WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>>>;
 
 using WarpGemmMfmaBf16Bf16F32M32N32K16TransposedCDistribution =
     WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution<
-        WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8,
+        WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
         2>>;
 
 using WarpGemmMfmaBf16Bf16F32M16N16K32TransposedCDistribution =
     WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution<
-        WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16,
+        WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16<WGAttrCtlEnum::Default_>,
         2>>;
 
 using WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleBTransposedCDistribution =
     WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB<
-        WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8,
+        WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8<WGAttrCtlEnum::Default_>,
         2>>;
 
+using WarpGemmMfmaBf16Bf16F32M4N64K16 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
+    WarpGemmAttributeMfmaImplBf16Bf16F32M4N64K4<WGAttrCtlEnum::Default_>,
+    4>>;
+
+using WarpGemmMfmaBf16Bf16F32M64N4K16 = WarpGemmImpl<WarpGemmAtrributeMfmaIterateK<
+    WarpGemmAttributeMfmaImplBf16Bf16F32M64N4K4<WGAttrCtlEnum::Default_>,
+    4>>;
+
 // fp8
-using WarpGemmMfma_f32_32x32x16_fp8_fp8 =
-    WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8>>;
 
-using WarpGemmMfma_f32_32x32x16_fp8_bf8 =
-    WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8>>;
+using WarpGemmMfma_f32_32x32x16_fp8_fp8 = WarpGemmImpl<
+    WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8<WGAttrCtlEnum::Default_>>>;
+
+using WarpGemmMfma_f32_32x32x16_fp8_bf8 = WarpGemmImpl<
+    WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8<WGAttrCtlEnum::Default_>>>;
 
-using WarpGemmMfma_f32_32x32x16_bf8_fp8 =
-    WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_fp8>>;
+using WarpGemmMfma_f32_32x32x16_bf8_fp8 = WarpGemmImpl<
+    WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_fp8<WGAttrCtlEnum::Default_>>>;
 
-using WarpGemmMfma_f32_32x32x16_bf8_bf8 =
-    WarpGemmImpl<WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8>>;
+using WarpGemmMfma_f32_32x32x16_bf8_bf8 = WarpGemmImpl<
+    WarpGemmAtrributeMfma<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8<WGAttrCtlEnum::Default_>>>;
 
-using WarpGemmMfma_f32_32x32x16_fp8_fp8_CTransposed = WarpGemmImpl<
-    WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8>>;
+using WarpGemmMfma_f32_32x32x16_fp8_fp8_CTransposed =
+    WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
+        WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8<WGAttrCtlEnum::Default_>>>;
 
-using WarpGemmMfma_f32_32x32x16_fp8_bf8_CTransposed = WarpGemmImpl<
-    WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8>>;
+using WarpGemmMfma_f32_32x32x16_fp8_bf8_CTransposed =
+    WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
+        WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8<WGAttrCtlEnum::Default_>>>;
 
-using WarpGemmMfma_f32_32x32x16_bf8_fp8_CTransposed = WarpGemmImpl<
-    WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_fp8>>;
+using WarpGemmMfma_f32_32x32x16_bf8_fp8_CTransposed =
+    WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
+        WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_fp8<WGAttrCtlEnum::Default_>>>;
 
-using WarpGemmMfma_f32_32x32x16_bf8_bf8_CTransposed = WarpGemmImpl<
-    WarpGemmAtrributeMfmaTransposedCDistribution<WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8>>;
+using WarpGemmMfma_f32_32x32x16_bf8_bf8_CTransposed =
+    WarpGemmImpl<WarpGemmAtrributeMfmaTransposedCDistribution<
+        WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8<WGAttrCtlEnum::Default_>>>;
 
 template <index_t swizzle_factor = 2>
 using WarpGemmMfmaFp8Fp8F32M32N32K16SwizzleBTransposedCDistribution =
     WarpGemmImpl<WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB<
-        WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<fp8_t, fp8_t>,
+        WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<fp8_t, fp8_t, WGAttrCtlEnum::Default_>,
         2,
         swizzle_factor>>;
 

include/ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -21,9 +21,15 @@ struct WarpGemmAtrributeMfma
     using BVecType = typename Impl::BVecType;
     using CVecType = typename Impl::CVecType;
 
-    static constexpr index_t kM = Impl::kM;
-    static constexpr index_t kN = Impl::kN;
-    static constexpr index_t kK = Impl::kK;
+    static constexpr index_t kM          = Impl::kM;
+    static constexpr index_t kN          = Impl::kN;
+    static constexpr index_t kK          = Impl::kK;
+    static constexpr index_t kKPerThread = Impl::kABKPerLane;
+
+    CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return 1; }
+
+    static_assert(Impl::kAMBlock == 1 && Impl::kBNBlock == 1,
+                  "Multi-block WarpGemmAttributeMfmaImpl is not supported");
 
     using AWarpDstrEncoding = tile_distribution_encoding<
         sequence<>,
@@ -51,10 +57,13 @@ struct WarpGemmAtrributeMfma
         sequence<0, 2>>;
 
     // c_vec += a_vec * b_vec
-    CK_TILE_DEVICE void
-    operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
+    template <bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   bool_constant<post_nop_> = {}) const
     {
-        Impl{}(c_vec, a_vec, b_vec);
+        Impl{}(c_vec, a_vec, b_vec, bool_constant<post_nop_>{});
     }
 
     // c_vec = a_vec * b_vec
@@ -81,38 +90,144 @@ struct WarpGemmAtrributeMfmaIterateK
         ext_vector_t<BDataType, vector_traits<typename Impl::BVecType>::vector_size * kKIter>;
     using CVecType = typename Impl::CVecType;
 
-    static constexpr index_t kM = Impl::kM;
-    static constexpr index_t kN = Impl::kN;
-    static constexpr index_t kK = Impl::kK * kKIter;
+    static constexpr index_t kM          = Impl::kM;
+    static constexpr index_t kN          = Impl::kN;
+    static constexpr index_t kK          = Impl::kK * kKIter;
+    static constexpr index_t kKPerThread = Impl::kABKPerLane * kKIter;
 
-    using AWarpDstrEncoding = tile_distribution_encoding<
-        sequence<>,
-        tuple<sequence<Impl::kAMLane>, sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
-        tuple<sequence<2, 1>>,
-        tuple<sequence<0, 0>>,
-        sequence<2>,
-        sequence<1>>;
+    CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return kKIter; }
 
-    using BWarpDstrEncoding = tile_distribution_encoding<
-        sequence<>,
-        tuple<sequence<Impl::kBNLane>, sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
-        tuple<sequence<2, 1>>,
-        tuple<sequence<0, 0>>,
-        sequence<2>,
-        sequence<1>>;
+    static_assert(Impl::kAMBlock == 1 || Impl::kBNBlock == 1,
+                  "Multi-block on both M & N directions is not supported");
 
-    using CWarpDstrEncoding = tile_distribution_encoding<
-        sequence<>,
-        tuple<sequence<Impl::kCM0PerLane, Impl::kCMLane, Impl::kCM1PerLane>,
-              sequence<Impl::kCNLane>>,
-        tuple<sequence<1, 2>>,
-        tuple<sequence<1, 0>>,
-        sequence<1, 1>,
-        sequence<0, 2>>;
+    CK_TILE_DEVICE static constexpr auto get_awarp_dstr_encoding()
+    {
+        if constexpr(Impl::kAMBlock == 1 && Impl::kBNBlock == 1)
+        {
+            return tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<Impl::kAMLane>,
+                      sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
+                tuple<sequence<2, 1>>,
+                tuple<sequence<0, 0>>,
+                sequence<2>,
+                sequence<1>>{};
+        }
+        else if constexpr(Impl::kAMBlock == 1 && 1 < Impl::kBNBlock)
+        {
+            // each M blocks share the same data
+            return tile_distribution_encoding<
+                sequence<Impl::kBNBlock>,
+                tuple<sequence<Impl::kAMLane>,
+                      sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
+                tuple<sequence<0, 2, 1>>,
+                tuple<sequence<0, 0, 0>>,
+                sequence<2>,
+                sequence<1>>{};
+        }
+        else if constexpr(1 < Impl::kAMBlock && Impl::kBNBlock == 1)
+        {
+            // single block to multi-block thread mapping
+            return tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<Impl::kAMBlock, Impl::kAMLane>,
+                      sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
+                tuple<sequence<1, 2, 1>>,
+                tuple<sequence<0, 0, 1>>,
+                sequence<2>,
+                sequence<1>>{};
+        }
+    }
+
+    CK_TILE_DEVICE static constexpr auto get_bwarp_dstr_encoding()
+    {
+        if constexpr(Impl::kAMBlock == 1 && Impl::kBNBlock == 1)
+        {
+            return tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<Impl::kBNLane>,
+                      sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
+                tuple<sequence<2, 1>>,
+                tuple<sequence<0, 0>>,
+                sequence<2>,
+                sequence<1>>{};
+        }
+        else if constexpr(Impl::kAMBlock == 1 && 1 < Impl::kBNBlock)
+        {
+            // single block to multi-block thread mapping
+            return tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<Impl::kBNBlock, Impl::kBNLane>,
+                      sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
+                tuple<sequence<1, 2, 1>>,
+                tuple<sequence<0, 0, 1>>,
+                sequence<2>,
+                sequence<1>>{};
+        }
+        else if constexpr(1 < Impl::kAMBlock && Impl::kBNBlock == 1)
+        {
+            // each N blocks share the same data
+            return tile_distribution_encoding<
+                sequence<Impl::kAMBlock>,
+                tuple<sequence<Impl::kBNLane>,
+                      sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
+                tuple<sequence<0, 2, 1>>,
+                tuple<sequence<0, 0, 0>>,
+                sequence<2>,
+                sequence<1>>{};
+        }
+    }
+
+    CK_TILE_DEVICE static constexpr auto get_cwarp_dstr_encoding()
+    {
+        if constexpr(Impl::kAMBlock == 1 && Impl::kBNBlock == 1)
+        {
+            return tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<Impl::kCM0PerLane, Impl::kCMLane, Impl::kCM1PerLane>,
+                      sequence<Impl::kCNLane>>,
+                tuple<sequence<1, 2>>,
+                tuple<sequence<1, 0>>,
+                sequence<1, 1>,
+                sequence<0, 2>>{};
+        }
+        else if constexpr(Impl::kAMBlock == 1 && 1 < Impl::kBNBlock)
+        {
+            return tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<Impl::kCM0PerLane, Impl::kCMLane, Impl::kCM1PerLane>,
+                      sequence<Impl::kBNBlock * Impl::kCNLane>>,
+                tuple<sequence<1, 2>>,
+                tuple<sequence<1, 0>>,
+                sequence<1, 1>,
+                sequence<0, 2>>{};
+        }
+        else if constexpr(1 < Impl::kAMBlock && Impl::kBNBlock == 1)
+        {
+            return tile_distribution_encoding<
+                sequence<>,
+                tuple<
+                    sequence<Impl::kCM0PerLane, Impl::kAMBlock * Impl::kCMLane, Impl::kCM1PerLane>,
+                    sequence<Impl::kCNLane>>,
+                tuple<sequence<1, 2>>,
+                tuple<sequence<1, 0>>,
+                sequence<1, 1>,
+                sequence<0, 2>>{};
+        }
+    }
+
+    using AWarpDstrEncoding = decltype(get_awarp_dstr_encoding());
+
+    using BWarpDstrEncoding = decltype(get_bwarp_dstr_encoding());
+
+    using CWarpDstrEncoding = decltype(get_cwarp_dstr_encoding());
 
     // c_vec += a_vec * b_vec
-    CK_TILE_DEVICE void
-    operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
+    template <bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   bool_constant<post_nop_> = {}) const
     {
         using buf_a = thread_buffer<typename Impl::AVecType, kKIter>;
         using buf_b = thread_buffer<typename Impl::BVecType, kKIter>;
@@ -122,10 +237,33 @@ struct WarpGemmAtrributeMfmaIterateK
                    reinterpret_cast<const buf_a&>(a_vec)
                        .template get_as<typename Impl::AVecType>()[iKIter],
                    reinterpret_cast<const buf_b&>(b_vec)
-                       .template get_as<typename Impl::BVecType>()[iKIter]);
+                       .template get_as<typename Impl::BVecType>()[iKIter],
+                   bool_constant<post_nop_>{});
         });
     }
 
+    template <index_t iKIter, bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   number<iKIter>,
+                                   bool_constant<post_nop_> = {}) const
+    {
+        using buf_a = thread_buffer<typename Impl::AVecType, kKIter>;
+        using buf_b = thread_buffer<typename Impl::BVecType, kKIter>;
+
+        static_assert(iKIter < kKIter);
+
+        // static_for<0, kKIter, 1>{}([&](auto iKIter) {
+        Impl{}(c_vec,
+               reinterpret_cast<const buf_a&>(a_vec)
+                   .template get_as<typename Impl::AVecType>()[iKIter],
+               reinterpret_cast<const buf_b&>(b_vec)
+                   .template get_as<typename Impl::BVecType>()[iKIter],
+               bool_constant<post_nop_>{});
+        //});
+    }
+
     // c_vec = a_vec * b_vec
     CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
     {
@@ -164,9 +302,15 @@ struct WarpGemmAtrributeMfmaTransposedCDistribution
     using BVecType = typename Impl::AVecType;
     using CVecType = typename Impl::CVecType;
 
-    static constexpr index_t kM = Impl::kN;
-    static constexpr index_t kN = Impl::kM;
-    static constexpr index_t kK = Impl::kK;
+    static constexpr index_t kM          = Impl::kN;
+    static constexpr index_t kN          = Impl::kM;
+    static constexpr index_t kK          = Impl::kK;
+    static constexpr index_t kKPerThread = Impl::kABKPerLane;
+
+    CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return 1; }
+
+    static_assert(Impl::kAMBlock == 1 && Impl::kBNBlock == 1,
+                  "Multi-block WarpGemmAttributeMfmaImpl is not supported");
 
     using AWarpDstrEncoding = tile_distribution_encoding<
         sequence<>,
@@ -194,11 +338,14 @@ struct WarpGemmAtrributeMfmaTransposedCDistribution
         sequence<0, 2>>;
 
     // c_vec += a_vec * b_vec
-    CK_TILE_DEVICE void
-    operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
+    template <bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   bool_constant<post_nop_> = {}) const
     {
         // swap A and B
-        Impl{}(c_vec, b_vec, a_vec);
+        Impl{}(c_vec, b_vec, a_vec, bool_constant<post_nop_>{});
     }
 
     // c_vec = a_vec * b_vec
@@ -222,9 +369,15 @@ struct WarpGemmAtrributeMfmaTransposedCDistribution_SwizzleB
     using BVecType = typename Impl::AVecType;
     using CVecType = typename Impl::CVecType;
 
-    static constexpr index_t kM = Impl::kN;
-    static constexpr index_t kN = Impl::kM;
-    static constexpr index_t kK = Impl::kK;
+    static constexpr index_t kM          = Impl::kN;
+    static constexpr index_t kN          = Impl::kM;
+    static constexpr index_t kK          = Impl::kK;
+    static constexpr index_t kKPerThread = Impl::kABKPerLane;
+
+    CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return 1; }
+
+    static_assert(Impl::kAMBlock == 1 && Impl::kBNBlock == 1,
+                  "Multi-block WarpGemmAttributeMfmaImpl is not supported");
 
     using AWarpDstrEncoding = tile_distribution_encoding<
         sequence<>,
@@ -255,12 +408,15 @@ struct WarpGemmAtrributeMfmaTransposedCDistribution_SwizzleB
         sequence<2, 2>,
         sequence<0, 2>>;
 
+    template <bool post_nop_ = false>
     // c_vec += a_vec * b_vec
-    CK_TILE_DEVICE void
-    operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   bool_constant<post_nop_> = {}) const
     {
         // swap A and B
-        Impl{}(c_vec, b_vec, a_vec);
+        Impl{}(c_vec, b_vec, a_vec, bool_constant<post_nop_>{});
     }
 
     // c_vec = a_vec * b_vec
@@ -287,38 +443,144 @@ struct WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution
         ext_vector_t<BDataType, vector_traits<typename Impl::BVecType>::vector_size * kKIter>;
     using CVecType = typename Impl::CVecType;
 
-    static constexpr index_t kM = Impl::kN;
-    static constexpr index_t kN = Impl::kM;
-    static constexpr index_t kK = Impl::kK * kKIter;
+    static constexpr index_t kM          = Impl::kN;
+    static constexpr index_t kN          = Impl::kM;
+    static constexpr index_t kK          = Impl::kK * kKIter;
+    static constexpr index_t kKPerThread = Impl::kABKPerLane * kKIter;
 
-    using AWarpDstrEncoding = tile_distribution_encoding<
-        sequence<>,
-        tuple<sequence<Impl::kBNLane>, sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
-        tuple<sequence<2, 1>>,
-        tuple<sequence<0, 0>>,
-        sequence<2>,
-        sequence<1>>;
+    CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return kKIter; }
 
-    using BWarpDstrEncoding = tile_distribution_encoding<
-        sequence<>,
-        tuple<sequence<Impl::kAMLane>, sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
-        tuple<sequence<2, 1>>,
-        tuple<sequence<0, 0>>,
-        sequence<2>,
-        sequence<1>>;
+    static_assert(Impl::kAMBlock == 1 || Impl::kBNBlock == 1,
+                  "Multi-block on both M & N directions is not supported");
 
-    using CWarpDstrEncoding = tile_distribution_encoding<
-        sequence<>,
-        tuple<sequence<Impl::kCNLane>,
-              sequence<Impl::kCM0PerLane, Impl::kCMLane, Impl::kCM1PerLane>>,
-        tuple<sequence<2, 1>>,
-        tuple<sequence<1, 0>>,
-        sequence<2, 2>,
-        sequence<0, 2>>;
+    CK_TILE_DEVICE static constexpr auto get_awarp_dstr_encoding()
+    {
+        if constexpr(Impl::kAMBlock == 1 && Impl::kBNBlock == 1)
+        {
+            return tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<Impl::kBNLane>,
+                      sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
+                tuple<sequence<2, 1>>,
+                tuple<sequence<0, 0>>,
+                sequence<2>,
+                sequence<1>>{};
+        }
+        else if constexpr(Impl::kAMBlock == 1 && 1 < Impl::kBNBlock)
+        {
+            // single block to multi-block thread mapping
+            return tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<Impl::kBNBlock, Impl::kBNLane>,
+                      sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
+                tuple<sequence<1, 2, 1>>,
+                tuple<sequence<0, 0, 1>>,
+                sequence<2>,
+                sequence<1>>{};
+        }
+        else if constexpr(1 < Impl::kAMBlock && Impl::kBNBlock == 1)
+        {
+            // each N blocks share the same data
+            return tile_distribution_encoding<
+                sequence<Impl::kAMBlock>,
+                tuple<sequence<Impl::kBNLane>,
+                      sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
+                tuple<sequence<0, 2, 1>>,
+                tuple<sequence<0, 0, 0>>,
+                sequence<2>,
+                sequence<1>>{};
+        }
+    }
 
+    CK_TILE_DEVICE static constexpr auto get_bwarp_dstr_encoding()
+    {
+        if constexpr(Impl::kAMBlock == 1 && Impl::kBNBlock == 1)
+        {
+            return tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<Impl::kAMLane>,
+                      sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
+                tuple<sequence<2, 1>>,
+                tuple<sequence<0, 0>>,
+                sequence<2>,
+                sequence<1>>{};
+        }
+        else if constexpr(Impl::kAMBlock == 1 && 1 < Impl::kBNBlock)
+        {
+            // each M blocks share the same data
+            return tile_distribution_encoding<
+                sequence<Impl::kBNBlock>,
+                tuple<sequence<Impl::kAMLane>,
+                      sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
+                tuple<sequence<0, 2, 1>>,
+                tuple<sequence<0, 0, 0>>,
+                sequence<2>,
+                sequence<1>>{};
+        }
+        else if constexpr(1 < Impl::kAMBlock && Impl::kBNBlock == 1)
+        {
+            // single block to multi-block thread mapping
+            return tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<Impl::kAMBlock, Impl::kAMLane>,
+                      sequence<Impl::kABKLane, Impl::kABKPerLane * kKIter>>,
+                tuple<sequence<1, 2, 1>>,
+                tuple<sequence<0, 0, 1>>,
+                sequence<2>,
+                sequence<1>>{};
+        }
+    }
+
+    CK_TILE_DEVICE static constexpr auto get_cwarp_dstr_encoding()
+    {
+        if constexpr(Impl::kAMBlock == 1 && Impl::kBNBlock == 1)
+        {
+            return tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<Impl::kCNLane>,
+                      sequence<Impl::kCM0PerLane, Impl::kCMLane, Impl::kCM1PerLane>>,
+                tuple<sequence<2, 1>>,
+                tuple<sequence<1, 0>>,
+                sequence<2, 2>,
+                sequence<0, 2>>{};
+        }
+        else if constexpr(Impl::kAMBlock == 1 && 1 < Impl::kBNBlock)
+        {
+            return tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<Impl::kBNBlock * Impl::kCNLane>,
+                      sequence<Impl::kCM0PerLane, Impl::kCMLane, Impl::kCM1PerLane>>,
+                tuple<sequence<2, 1>>,
+                tuple<sequence<1, 0>>,
+                sequence<2, 2>,
+                sequence<0, 2>>{};
+        }
+        else if constexpr(1 < Impl::kAMBlock && Impl::kBNBlock == 1)
+        {
+            return tile_distribution_encoding<
+                sequence<>,
+                tuple<
+                    sequence<Impl::kCNLane>,
+                    sequence<Impl::kCM0PerLane, Impl::kAMBlock * Impl::kCMLane, Impl::kCM1PerLane>>,
+                tuple<sequence<2, 1>>,
+                tuple<sequence<1, 0>>,
+                sequence<2, 2>,
+                sequence<0, 2>>{};
+        }
+    }
+
+    using AWarpDstrEncoding = decltype(get_awarp_dstr_encoding());
+
+    using BWarpDstrEncoding = decltype(get_bwarp_dstr_encoding());
+
+    using CWarpDstrEncoding = decltype(get_cwarp_dstr_encoding());
+
+    template <bool post_nop_ = false>
     // c_vec += a_vec * b_vec
-    CK_TILE_DEVICE void
-    operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   bool_constant<post_nop_> = {}) const
     {
         using buf_a = thread_buffer<typename Impl::AVecType, kKIter>;
         using buf_b = thread_buffer<typename Impl::BVecType, kKIter>;
@@ -328,10 +590,34 @@ struct WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution
                    reinterpret_cast<const buf_b&>(b_vec)
                        .template get_as<typename Impl::BVecType>()[iKIter],
                    reinterpret_cast<const buf_a&>(a_vec)
-                       .template get_as<typename Impl::AVecType>()[iKIter]);
+                       .template get_as<typename Impl::AVecType>()[iKIter],
+                   bool_constant<post_nop_>{});
         });
     }
 
+    template <index_t iKIter, bool post_nop_ = false>
+    // c_vec += a_vec * b_vec
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   number<iKIter>,
+                                   bool_constant<post_nop_> = {}) const
+    {
+        using buf_a = thread_buffer<typename Impl::AVecType, kKIter>;
+        using buf_b = thread_buffer<typename Impl::BVecType, kKIter>;
+
+        static_assert(iKIter < kKIter);
+        // swap A and B, value and type
+        // static_for<0, kKIter, 1>{}([&](auto iKIter) {
+        Impl{}(c_vec,
+               reinterpret_cast<const buf_b&>(b_vec)
+                   .template get_as<typename Impl::BVecType>()[iKIter],
+               reinterpret_cast<const buf_a&>(a_vec)
+                   .template get_as<typename Impl::AVecType>()[iKIter],
+               bool_constant<post_nop_>{});
+        //});
+    }
+
     // c_vec = a_vec * b_vec
     CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
     {
@@ -372,10 +658,16 @@ struct WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB
         ext_vector_t<BDataType, vector_traits<typename Impl::BVecType>::vector_size * kKIter>;
     using CVecType = typename Impl::CVecType;
 
-    static constexpr index_t kM      = Impl::kN;
-    static constexpr index_t kN      = Impl::kM;
-    static constexpr index_t kK      = Impl::kK * kKIter;
-    static constexpr index_t SFactor = SFactor_; // group how many CM1 together
+    static constexpr index_t kM          = Impl::kN;
+    static constexpr index_t kN          = Impl::kM;
+    static constexpr index_t kK          = Impl::kK * kKIter;
+    static constexpr index_t kKPerThread = Impl::kABKPerLane * kKIter;
+    static constexpr index_t SFactor     = SFactor_; // group how many CM1 together
+
+    CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return kKIter; }
+
+    static_assert(Impl::kAMBlock == 1 && Impl::kBNBlock == 1,
+                  "Multi-block WarpGemmAttributeMfmaImpl is not supported");
 
     using AWarpDstrEncoding = tile_distribution_encoding<
         sequence<>,
@@ -429,8 +721,11 @@ struct WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB
         sequence<0, 2>>;
 #endif
     // c_vec += a_vec * b_vec
-    CK_TILE_DEVICE void
-    operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
+    template <bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   bool_constant<post_nop_> = {}) const
     {
         using buf_a = thread_buffer<typename Impl::AVecType, kKIter>;
         using buf_b = thread_buffer<typename Impl::BVecType, kKIter>;
@@ -440,10 +735,33 @@ struct WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB
                    reinterpret_cast<const buf_b&>(b_vec)
                        .template get_as<typename Impl::BVecType>()[iKIter],
                    reinterpret_cast<const buf_a&>(a_vec)
-                       .template get_as<typename Impl::AVecType>()[iKIter]);
+                       .template get_as<typename Impl::AVecType>()[iKIter],
+                   bool_constant<post_nop_>{});
         });
     }
 
+    template <index_t iKIter, bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   number<iKIter>,
+                                   bool_constant<post_nop_> = {}) const
+    {
+        using buf_a = thread_buffer<typename Impl::AVecType, kKIter>;
+        using buf_b = thread_buffer<typename Impl::BVecType, kKIter>;
+
+        static_assert(iKIter < kKIter);
+        // swap A and B, value and type
+        // static_for<0, kKIter, 1>{}([&](auto iKIter) {
+        Impl{}(c_vec,
+               reinterpret_cast<const buf_b&>(b_vec)
+                   .template get_as<typename Impl::BVecType>()[iKIter],
+               reinterpret_cast<const buf_a&>(a_vec)
+                   .template get_as<typename Impl::AVecType>()[iKIter],
+               bool_constant<post_nop_>{});
+        //});
+    }
+
     // c_vec = a_vec * b_vec
     CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
     {
@@ -483,10 +801,16 @@ struct WarpGemmAtrributeMfmaIterateK_SwizzleA
         ext_vector_t<BDataType, vector_traits<typename Impl::BVecType>::vector_size * kKIter>;
     using CVecType = typename Impl::CVecType;
 
-    static constexpr index_t kM      = Impl::kM;
-    static constexpr index_t kN      = Impl::kN;
-    static constexpr index_t kK      = Impl::kK * kKIter;
-    static constexpr index_t SFactor = SFactor_; // group how many CM1 together
+    static constexpr index_t kM          = Impl::kM;
+    static constexpr index_t kN          = Impl::kN;
+    static constexpr index_t kK          = Impl::kK * kKIter;
+    static constexpr index_t kKPerThread = Impl::kABKPerLane * kKIter;
+    static constexpr index_t SFactor     = SFactor_; // group how many CM1 together
+
+    CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return kKIter; }
+
+    static_assert(Impl::kAMBlock == 1 && Impl::kBNBlock == 1,
+                  "Multi-block WarpGemmAttributeMfmaImpl is not supported");
 
     using AWarpDstrEncoding = tile_distribution_encoding<
         sequence<>,
@@ -518,8 +842,11 @@ struct WarpGemmAtrributeMfmaIterateK_SwizzleA
         sequence<0, 2>>;
 
     // c_vec += a_vec * b_vec
-    CK_TILE_DEVICE void
-    operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
+    template <bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   bool_constant<post_nop_> = {}) const
     {
         using buf_a = thread_buffer<typename Impl::AVecType, kKIter>;
         using buf_b = thread_buffer<typename Impl::BVecType, kKIter>;
@@ -529,10 +856,33 @@ struct WarpGemmAtrributeMfmaIterateK_SwizzleA
                    reinterpret_cast<const buf_a&>(a_vec)
                        .template get_as<typename Impl::AVecType>()[iKIter],
                    reinterpret_cast<const buf_b&>(b_vec)
-                       .template get_as<typename Impl::BVecType>()[iKIter]);
+                       .template get_as<typename Impl::BVecType>()[iKIter],
+                   bool_constant<post_nop_>{});
         });
     }
 
+    template <index_t iKIter, bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   number<iKIter>,
+                                   bool_constant<post_nop_> = {}) const
+    {
+        using buf_a = thread_buffer<typename Impl::AVecType, kKIter>;
+        using buf_b = thread_buffer<typename Impl::BVecType, kKIter>;
+
+        static_assert(iKIter < kKIter);
+
+        // static_for<0, kKIter, 1>{}([&](auto iKIter) {
+        Impl{}(c_vec,
+               reinterpret_cast<const buf_a&>(a_vec)
+                   .template get_as<typename Impl::AVecType>()[iKIter],
+               reinterpret_cast<const buf_b&>(b_vec)
+                   .template get_as<typename Impl::BVecType>()[iKIter],
+               bool_constant<post_nop_>{});
+        //});
+    }
+
     // c_vec = a_vec * b_vec
     CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
     {

include/ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma_impl.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -7,12 +7,68 @@
 
 namespace ck_tile {
 
+// TODO: refactor warp-gemm
+// currently there is a discrepency for vav/vva if we need transpose C/D
+// e.g. if we want A:agpr, B:vgpr, we have to use vva in WGAttrEnum
+// because we swap the A/B pointer in _impl code (but not known this info here)
+enum class WGAttrCtlEnum
+{
+    Default_ = 0,
+    Raw_vvv  = 1, // c-vgpr, a-vgpr, b-vgpr
+    Raw_vaa  = 2, // c-vgpr, a-agpr, b-agpr
+    Raw_vav  = 3, // c-vgpr, a-agpr, b-vgpr
+    Raw_vva  = 4, // c-vgpr, a-vgpr, b-agpr
+    Raw_avv  = 5, // c-agpr, a-vgpr, b-vgpr
+    // raw_a_a_a = 3,  // c-agpr, a-agpr, b-agpr
+};
+
+#define DISPATCH_MFMA_(mfma_, dmod_, amod_, bmod_, cmod_)       \
+    if constexpr(post_nop_)                                     \
+    {                                                           \
+        asm volatile(mfma_ " %0, %1, %2, %3 ; yyy\n"            \
+                           "s_nop 3"                            \
+                     : dmod_(c_vec)                             \
+                     : amod_(a_vec), bmod_(b_vec), cmod_(c_vec) \
+                     :);                                        \
+    }                                                           \
+    else                                                        \
+    {                                                           \
+        asm volatile(mfma_ " %0, %1, %2, %3\n"                  \
+                     : dmod_(c_vec)                             \
+                     : amod_(a_vec), bmod_(b_vec), cmod_(c_vec) \
+                     :);                                        \
+    }
+
+#define DISPATCH_MFMA_CTRL_(mfma_, ctrl_)              \
+    if constexpr(ctrl_ == WGAttrCtlEnum::Raw_vvv)      \
+    {                                                  \
+        DISPATCH_MFMA_(mfma_, "+v", "v", "v", "v")     \
+    }                                                  \
+    else if constexpr(ctrl_ == WGAttrCtlEnum::Raw_vaa) \
+    {                                                  \
+        DISPATCH_MFMA_(mfma_, "+v", "a", "a", "v")     \
+    }                                                  \
+    else if constexpr(ctrl_ == WGAttrCtlEnum::Raw_vav) \
+    {                                                  \
+        DISPATCH_MFMA_(mfma_, "+v", "a", "v", "v")     \
+    }                                                  \
+    else if constexpr(ctrl_ == WGAttrCtlEnum::Raw_vva) \
+    {                                                  \
+        DISPATCH_MFMA_(mfma_, "+v", "v", "a", "v")     \
+    }                                                  \
+    else if constexpr(ctrl_ == WGAttrCtlEnum::Raw_avv) \
+    {                                                  \
+        DISPATCH_MFMA_(mfma_, "+a", "v", "v", "a")     \
+    }
+
 // FP16
+template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
 struct WarpGemmAttributeMfmaImplF16F16F32M32N32K8
 {
-    using ADataType = fp16_t;
-    using BDataType = fp16_t;
-    using CDataType = float;
+    static constexpr WGAttrCtlEnum Ctrl = Ctrl_;
+    using ADataType                     = fp16_t;
+    using BDataType                     = fp16_t;
+    using CDataType                     = float;
 
     using AVecType = ext_vector_t<fp16_t, 4>;
     using BVecType = ext_vector_t<fp16_t, 4>;
@@ -22,6 +78,9 @@ struct WarpGemmAttributeMfmaImplF16F16F32M32N32K8
     static constexpr index_t kN = 32;
     static constexpr index_t kK = 8;
 
+    static constexpr index_t kAMBlock = 1;
+    static constexpr index_t kBNBlock = 1;
+
     static constexpr index_t kAMLane     = 32;
     static constexpr index_t kBNLane     = 32;
     static constexpr index_t kABKLane    = 2;
@@ -33,16 +92,23 @@ struct WarpGemmAttributeMfmaImplF16F16F32M32N32K8
     static constexpr index_t kCM1PerLane = 4;
 
     // c_vec += a_vec * b_vec
-    CK_TILE_DEVICE void
-    operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
+    template <bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   bool_constant<post_nop_> = {}) const
     {
+        DISPATCH_MFMA_CTRL_("v_mfma_f32_32x32x8f16", Ctrl)
+        else
+        {
 #if defined(__gfx9__)
-        c_vec = __builtin_amdgcn_mfma_f32_32x32x8f16(a_vec, b_vec, c_vec, 0, 0, 0);
+            c_vec = __builtin_amdgcn_mfma_f32_32x32x8f16(a_vec, b_vec, c_vec, 0, 0, 0);
 #else
-        ignore = c_vec;
-        ignore = a_vec;
-        ignore = b_vec;
+            ck_tile::ignore = c_vec;
+            ck_tile::ignore = a_vec;
+            ck_tile::ignore = b_vec;
 #endif
+        }
     }
 
     // c_vec = a_vec * b_vec
@@ -52,18 +118,20 @@ struct WarpGemmAttributeMfmaImplF16F16F32M32N32K8
         return bit_cast<CVecType>(
             __builtin_amdgcn_mfma_f32_32x32x8f16(a_vec, b_vec, fp32x16_t{0.f}, 0, 0, 0));
 #else
-        ignore = a_vec;
-        ignore = b_vec;
+        ck_tile::ignore = a_vec;
+        ck_tile::ignore = b_vec;
         return CVecType{0.f};
 #endif
     }
 };
 
+template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
 struct WarpGemmAttributeMfmaImplF16F16F32M16N16K16
 {
-    using ADataType = fp16_t;
-    using BDataType = fp16_t;
-    using CDataType = float;
+    static constexpr WGAttrCtlEnum Ctrl = Ctrl_;
+    using ADataType                     = fp16_t;
+    using BDataType                     = fp16_t;
+    using CDataType                     = float;
 
     using AVecType = ext_vector_t<fp16_t, 4>;
     using BVecType = ext_vector_t<fp16_t, 4>;
@@ -73,6 +141,9 @@ struct WarpGemmAttributeMfmaImplF16F16F32M16N16K16
     static constexpr index_t kN = 16;
     static constexpr index_t kK = 16;
 
+    static constexpr index_t kAMBlock = 1;
+    static constexpr index_t kBNBlock = 1;
+
     static constexpr index_t kAMLane     = 16;
     static constexpr index_t kBNLane     = 16;
     static constexpr index_t kABKLane    = 4;
@@ -84,16 +155,151 @@ struct WarpGemmAttributeMfmaImplF16F16F32M16N16K16
     static constexpr index_t kCM1PerLane = 4;
 
     // c_vec += a_vec * b_vec
-    CK_TILE_DEVICE void
-    operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
+    template <bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   bool_constant<post_nop_> = {}) const
+    {
+        DISPATCH_MFMA_CTRL_("v_mfma_f32_16x16x16f16", Ctrl)
+        else
+        {
+#if defined(__gfx9__)
+            c_vec = __builtin_amdgcn_mfma_f32_16x16x16f16(a_vec, b_vec, c_vec, 0, 0, 0);
+#else
+            ck_tile::ignore = c_vec;
+            ck_tile::ignore = a_vec;
+            ck_tile::ignore = b_vec;
+#endif
+        }
+    }
+
+    // c_vec = a_vec * b_vec
+    CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
     {
 #if defined(__gfx9__)
-        c_vec = __builtin_amdgcn_mfma_f32_16x16x16f16(a_vec, b_vec, c_vec, 0, 0, 0);
+        return bit_cast<CVecType>(
+            __builtin_amdgcn_mfma_f32_16x16x16f16(a_vec, b_vec, fp32x4_t{0.f}, 0, 0, 0));
+#else
+        ck_tile::ignore = a_vec;
+        ck_tile::ignore = b_vec;
+        return CVecType{0.f};
+#endif
+    }
+};
+
+template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
+struct WarpGemmAttributeMfmaImplF16F16F32M4N64K4
+{
+    static constexpr WGAttrCtlEnum Ctrl = Ctrl_;
+    using ADataType                     = fp16_t;
+    using BDataType                     = fp16_t;
+    using CDataType                     = float;
+
+    using AVecType = ext_vector_t<fp16_t, 4>;
+    using BVecType = ext_vector_t<fp16_t, 4>;
+    using CVecType = ext_vector_t<float, 4>;
+
+    static constexpr index_t kM = 4;
+    static constexpr index_t kN = 64;
+    static constexpr index_t kK = 4;
+
+    static constexpr index_t kAMBlock = 1;
+    static constexpr index_t kBNBlock = 16;
+
+    // we only write down single block (4 threads) thread mapping here
+    static constexpr index_t kAMLane     = 4;
+    static constexpr index_t kBNLane     = 4;
+    static constexpr index_t kABKLane    = 1;
+    static constexpr index_t kABKPerLane = 4;
+
+    static constexpr index_t kCMLane     = 1;
+    static constexpr index_t kCNLane     = 4;
+    static constexpr index_t kCM0PerLane = 1;
+    static constexpr index_t kCM1PerLane = 4;
+
+    // c_vec += a_vec * b_vec
+    template <bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   bool_constant<post_nop_> = {}) const
+    {
+        DISPATCH_MFMA_CTRL_("v_mfma_f32_4x4x4f16", Ctrl)
+        else
+        {
+#if defined(__gfx9__)
+            c_vec = __builtin_amdgcn_mfma_f32_4x4x4f16(a_vec, b_vec, c_vec, 0, 0, 0);
+#else
+            ignore = c_vec;
+            ignore = a_vec;
+            ignore = b_vec;
+#endif
+        }
+    }
+
+    // c_vec = a_vec * b_vec
+    CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
+    {
+#if defined(__gfx9__)
+        return bit_cast<CVecType>(
+            __builtin_amdgcn_mfma_f32_4x4x4f16(a_vec, b_vec, fp32x4_t{0.f}, 0, 0, 0));
 #else
-        ignore = c_vec;
         ignore = a_vec;
         ignore = b_vec;
+        return CVecType{0.f};
+#endif
+    }
+};
+
+template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
+struct WarpGemmAttributeMfmaImplF16F16F32M64N4K4
+{
+    static constexpr WGAttrCtlEnum Ctrl = Ctrl_;
+    using ADataType                     = fp16_t;
+    using BDataType                     = fp16_t;
+    using CDataType                     = float;
+
+    using AVecType = ext_vector_t<fp16_t, 4>;
+    using BVecType = ext_vector_t<fp16_t, 4>;
+    using CVecType = ext_vector_t<float, 4>;
+
+    static constexpr index_t kM = 64;
+    static constexpr index_t kN = 4;
+    static constexpr index_t kK = 4;
+
+    static constexpr index_t kAMBlock = 16;
+    static constexpr index_t kBNBlock = 1;
+
+    // we only write down single block (4 threads) thread mapping here
+    static constexpr index_t kAMLane     = 4;
+    static constexpr index_t kBNLane     = 4;
+    static constexpr index_t kABKLane    = 1;
+    static constexpr index_t kABKPerLane = 4;
+
+    static constexpr index_t kCMLane     = 1;
+    static constexpr index_t kCNLane     = 4;
+    static constexpr index_t kCM0PerLane = 1;
+    static constexpr index_t kCM1PerLane = 4;
+
+    // c_vec += a_vec * b_vec
+    template <bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   bool_constant<post_nop_> = {}) const
+    {
+        DISPATCH_MFMA_CTRL_("v_mfma_f32_4x4x4f16", Ctrl)
+        else
+        {
+#if defined(__gfx9__)
+            c_vec = __builtin_amdgcn_mfma_f32_4x4x4f16(a_vec, b_vec, c_vec, 0, 0, 0);
+#else
+            ignore = c_vec;
+            ignore = a_vec;
+            ignore = b_vec;
 #endif
+        }
     }
 
     // c_vec = a_vec * b_vec
@@ -101,7 +307,7 @@ struct WarpGemmAttributeMfmaImplF16F16F32M16N16K16
     {
 #if defined(__gfx9__)
         return bit_cast<CVecType>(
-            __builtin_amdgcn_mfma_f32_16x16x16f16(a_vec, b_vec, fp32x4_t{0.f}, 0, 0, 0));
+            __builtin_amdgcn_mfma_f32_4x4x4f16(a_vec, b_vec, fp32x4_t{0.f}, 0, 0, 0));
 #else
         ignore = a_vec;
         ignore = b_vec;
@@ -111,11 +317,13 @@ struct WarpGemmAttributeMfmaImplF16F16F32M16N16K16
 };
 
 // Bf16
+template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
 struct WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8
 {
-    using ADataType = bf16_t;
-    using BDataType = bf16_t;
-    using CDataType = float;
+    static constexpr WGAttrCtlEnum Ctrl = Ctrl_;
+    using ADataType                     = bf16_t;
+    using BDataType                     = bf16_t;
+    using CDataType                     = float;
 
     using AVecType = ext_vector_t<bf16_t, 4>;
     using BVecType = ext_vector_t<bf16_t, 4>;
@@ -125,6 +333,9 @@ struct WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8
     static constexpr index_t kN = 32;
     static constexpr index_t kK = 8;
 
+    static constexpr index_t kAMBlock = 1;
+    static constexpr index_t kBNBlock = 1;
+
     static constexpr index_t kAMLane     = 32;
     static constexpr index_t kBNLane     = 32;
     static constexpr index_t kABKLane    = 2;
@@ -136,28 +347,35 @@ struct WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8
     static constexpr index_t kCM1PerLane = 4;
 
     // c_vec += a_vec * b_vec
-    CK_TILE_DEVICE void
-    operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
+    template <bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   bool_constant<post_nop_> = {}) const
     {
+        DISPATCH_MFMA_CTRL_("v_mfma_f32_32x32x8bf16_1k", Ctrl)
+        else
+        {
 #if defined(__gfx90a__) || defined(__gfx94__)
-        c_vec = __builtin_amdgcn_mfma_f32_32x32x8bf16_1k(a_vec, b_vec, c_vec, 0, 0, 0);
+            c_vec = __builtin_amdgcn_mfma_f32_32x32x8bf16_1k(a_vec, b_vec, c_vec, 0, 0, 0);
 #elif defined(__gfx908__)
-        static_for<0, 2, 1>{}([&](auto k) {
-            c_vec = __builtin_amdgcn_mfma_f32_32x32x4bf16(
-                reinterpret_cast<const thread_buffer<ADataType, 4>&>(a_vec)
-                    .template get_as<ext_vector_t<bf16_t, 2>>()[number<k>{}],
-                reinterpret_cast<const thread_buffer<BDataType, 4>&>(b_vec)
-                    .template get_as<ext_vector_t<bf16_t, 2>>()[number<k>{}],
-                c_vec,
-                0,
-                0,
-                0);
-        });
+            static_for<0, 2, 1>{}([&](auto k) {
+                c_vec = __builtin_amdgcn_mfma_f32_32x32x4bf16(
+                    reinterpret_cast<const thread_buffer<ADataType, 4>&>(a_vec)
+                        .template get_as<ext_vector_t<bf16_t, 2>>()[number<k>{}],
+                    reinterpret_cast<const thread_buffer<BDataType, 4>&>(b_vec)
+                        .template get_as<ext_vector_t<bf16_t, 2>>()[number<k>{}],
+                    c_vec,
+                    0,
+                    0,
+                    0);
+            });
 #else
-        ignore = c_vec;
-        ignore = a_vec;
-        ignore = b_vec;
+            ck_tile::ignore = c_vec;
+            ck_tile::ignore = a_vec;
+            ck_tile::ignore = b_vec;
 #endif
+        }
     }
 
     // c_vec = a_vec * b_vec
@@ -181,18 +399,20 @@ struct WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8
         });
         return c_vec;
 #else
-        ignore = a_vec;
-        ignore = b_vec;
+        ck_tile::ignore = a_vec;
+        ck_tile::ignore = b_vec;
         return CVecType{0.f};
 #endif
     }
 };
 
+template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
 struct WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16
 {
-    using ADataType = bf16_t;
-    using BDataType = bf16_t;
-    using CDataType = float;
+    static constexpr WGAttrCtlEnum Ctrl = Ctrl_;
+    using ADataType                     = bf16_t;
+    using BDataType                     = bf16_t;
+    using CDataType                     = float;
 
     using AVecType = ext_vector_t<bf16_t, 4>;
     using BVecType = ext_vector_t<bf16_t, 4>;
@@ -202,6 +422,9 @@ struct WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16
     static constexpr index_t kN = 16;
     static constexpr index_t kK = 16;
 
+    static constexpr index_t kAMBlock = 1;
+    static constexpr index_t kBNBlock = 1;
+
     static constexpr index_t kAMLane     = 16;
     static constexpr index_t kBNLane     = 16;
     static constexpr index_t kABKLane    = 4;
@@ -213,28 +436,34 @@ struct WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16
     static constexpr index_t kCM1PerLane = 4;
 
     // c_vec += a_vec * b_vec
-    CK_TILE_DEVICE void
-    operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
+    template <bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   bool_constant<post_nop_> = {}) const
     {
+        DISPATCH_MFMA_CTRL_("v_mfma_f32_16x16x16bf16_1k", Ctrl)
+        {
 #if defined(__gfx90a__) || defined(__gfx94__)
-        c_vec = __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(a_vec, b_vec, c_vec, 0, 0, 0);
+            c_vec = __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(a_vec, b_vec, c_vec, 0, 0, 0);
 #elif defined(__gfx908__)
-        static_for<0, 2, 1>{}([&](auto k) {
-            c_vec = __builtin_amdgcn_mfma_f32_16x16x8bf16(
-                reinterpret_cast<const thread_buffer<ADataType, 4>&>(a_vec)
-                    .template get_as<ext_vector_t<bf16_t, 2>>()[number<k>{}],
-                reinterpret_cast<const thread_buffer<BDataType, 4>&>(b_vec)
-                    .template get_as<ext_vector_t<bf16_t, 2>>()[number<k>{}],
-                c_vec,
-                0,
-                0,
-                0);
-        });
+            static_for<0, 2, 1>{}([&](auto k) {
+                c_vec = __builtin_amdgcn_mfma_f32_16x16x8bf16(
+                    reinterpret_cast<const thread_buffer<ADataType, 4>&>(a_vec)
+                        .template get_as<ext_vector_t<bf16_t, 2>>()[number<k>{}],
+                    reinterpret_cast<const thread_buffer<BDataType, 4>&>(b_vec)
+                        .template get_as<ext_vector_t<bf16_t, 2>>()[number<k>{}],
+                    c_vec,
+                    0,
+                    0,
+                    0);
+            });
 #else
-        ignore = c_vec;
-        ignore = a_vec;
-        ignore = b_vec;
+            ck_tile::ignore = c_vec;
+            ck_tile::ignore = a_vec;
+            ck_tile::ignore = b_vec;
 #endif
+        }
     }
 
     // c_vec = a_vec * b_vec
@@ -257,6 +486,134 @@ struct WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16
                 0);
         });
         return c_vec;
+#else
+        ck_tile::ignore = a_vec;
+        ck_tile::ignore = b_vec;
+        return CVecType{0.f};
+#endif
+    }
+};
+
+template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
+struct WarpGemmAttributeMfmaImplBf16Bf16F32M4N64K4
+{
+    static constexpr WGAttrCtlEnum Ctrl = Ctrl_;
+    using ADataType                     = bf16_t;
+    using BDataType                     = bf16_t;
+    using CDataType                     = float;
+
+    using AVecType = ext_vector_t<bf16_t, 4>;
+    using BVecType = ext_vector_t<bf16_t, 4>;
+    using CVecType = ext_vector_t<float, 4>;
+
+    static constexpr index_t kM = 4;
+    static constexpr index_t kN = 64;
+    static constexpr index_t kK = 4;
+
+    static constexpr index_t kAMBlock = 1;
+    static constexpr index_t kBNBlock = 16;
+
+    // we only write down single block (4 threads) thread mapping here
+    static constexpr index_t kAMLane     = 4;
+    static constexpr index_t kBNLane     = 4;
+    static constexpr index_t kABKLane    = 1;
+    static constexpr index_t kABKPerLane = 4;
+
+    static constexpr index_t kCMLane     = 1;
+    static constexpr index_t kCNLane     = 4;
+    static constexpr index_t kCM0PerLane = 1;
+    static constexpr index_t kCM1PerLane = 4;
+
+    // c_vec += a_vec * b_vec
+    template <bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   bool_constant<post_nop_> = {}) const
+    {
+        DISPATCH_MFMA_CTRL_("v_mfma_f32_4x4x4bf16_1k", Ctrl)
+        else
+        {
+#if defined(__gfx9__)
+            c_vec = __builtin_amdgcn_mfma_f32_4x4x4bf16_1k(a_vec, b_vec, c_vec, 0, 0, 0);
+#else
+            ignore = c_vec;
+            ignore = a_vec;
+            ignore = b_vec;
+#endif
+        }
+    }
+
+    // c_vec = a_vec * b_vec
+    CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
+    {
+#if defined(__gfx9__)
+        return bit_cast<CVecType>(
+            __builtin_amdgcn_mfma_f32_4x4x4bf16_1k(a_vec, b_vec, fp32x4_t{0.f}, 0, 0, 0));
+#else
+        ignore = a_vec;
+        ignore = b_vec;
+        return CVecType{0.f};
+#endif
+    }
+};
+
+template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
+struct WarpGemmAttributeMfmaImplBf16Bf16F32M64N4K4
+{
+    static constexpr WGAttrCtlEnum Ctrl = Ctrl_;
+    using ADataType                     = bf16_t;
+    using BDataType                     = bf16_t;
+    using CDataType                     = float;
+
+    using AVecType = ext_vector_t<bf16_t, 4>;
+    using BVecType = ext_vector_t<bf16_t, 4>;
+    using CVecType = ext_vector_t<float, 4>;
+
+    static constexpr index_t kM = 64;
+    static constexpr index_t kN = 4;
+    static constexpr index_t kK = 4;
+
+    static constexpr index_t kAMBlock = 16;
+    static constexpr index_t kBNBlock = 1;
+
+    // we only write down single block (4 threads) thread mapping here
+    static constexpr index_t kAMLane     = 4;
+    static constexpr index_t kBNLane     = 4;
+    static constexpr index_t kABKLane    = 1;
+    static constexpr index_t kABKPerLane = 4;
+
+    static constexpr index_t kCMLane     = 1;
+    static constexpr index_t kCNLane     = 4;
+    static constexpr index_t kCM0PerLane = 1;
+    static constexpr index_t kCM1PerLane = 4;
+
+    // c_vec += a_vec * b_vec
+    template <bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   bool_constant<post_nop_> = {}) const
+    {
+        DISPATCH_MFMA_CTRL_("v_mfma_f32_4x4x4bf16_1k", Ctrl)
+        else
+        {
+#if defined(__gfx9__)
+            c_vec = __builtin_amdgcn_mfma_f32_4x4x4bf16_1k(a_vec, b_vec, c_vec, 0, 0, 0);
+#else
+            ignore = c_vec;
+            ignore = a_vec;
+            ignore = b_vec;
+#endif
+        }
+    }
+
+    // c_vec = a_vec * b_vec
+    CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
+    {
+#if defined(__gfx9__)
+        return bit_cast<CVecType>(
+            __builtin_amdgcn_mfma_f32_4x4x4bf16_1k(a_vec, b_vec, fp32x4_t{0.f}, 0, 0, 0));
 #else
         ignore = a_vec;
         ignore = b_vec;
@@ -266,12 +623,13 @@ struct WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16
 };
 
 // FP8
-template <typename AType_, typename BType_>
+template <typename AType_, typename BType_, WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
 struct WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base
 {
-    using ADataType = AType_;
-    using BDataType = BType_;
-    using CDataType = float;
+    static constexpr WGAttrCtlEnum Ctrl = Ctrl_;
+    using ADataType                     = AType_;
+    using BDataType                     = BType_;
+    using CDataType                     = float;
 
     using AVecType = ext_vector_t<ADataType, 8>;
     using BVecType = ext_vector_t<BDataType, 8>;
@@ -281,6 +639,9 @@ struct WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base
     static constexpr index_t kN = 32;
     static constexpr index_t kK = 16;
 
+    static constexpr index_t kAMBlock = 1;
+    static constexpr index_t kBNBlock = 1;
+
     static constexpr index_t kAMLane     = 32;
     static constexpr index_t kBNLane     = 32;
     static constexpr index_t kABKLane    = 2;
@@ -292,38 +653,120 @@ struct WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base
     static constexpr index_t kCM1PerLane = 4;
 
     // c_vec += a_vec * b_vec
-    CK_TILE_DEVICE void
-    operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const
+    template <bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   bool_constant<post_nop_> = {}) const
     {
+        if constexpr(Ctrl == WGAttrCtlEnum::Raw_vvv)
+        {
+            if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, fp8_t>)
+            {
+                DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_fp8", "+v", "v", "v", "v")
+            }
+            else if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, bf8_t>)
+            {
+                DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_bf8", "+v", "v", "v", "v")
+            }
+            else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, fp8_t>)
+            {
+                DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_fp8", "+v", "v", "v", "v")
+            }
+            else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, bf8_t>)
+            {
+                DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_bf8", "+v", "v", "v", "v")
+            }
+        }
+        else if constexpr(Ctrl == WGAttrCtlEnum::Raw_vaa)
+        {
+            if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, fp8_t>)
+            {
+                DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_fp8", "+v", "a", "a", "v")
+            }
+            else if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, bf8_t>)
+            {
+                DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_bf8", "+v", "a", "a", "v")
+            }
+            else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, fp8_t>)
+            {
+                DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_fp8", "+v", "a", "a", "v")
+            }
+            else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, bf8_t>)
+            {
+                DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_bf8", "+v", "a", "a", "v")
+            }
+        }
+        else if constexpr(Ctrl == WGAttrCtlEnum::Raw_vav)
+        {
+            if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, fp8_t>)
+            {
+                DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_fp8", "+v", "a", "v", "v")
+            }
+            else if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, bf8_t>)
+            {
+                DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_bf8", "+v", "a", "v", "v")
+            }
+            else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, fp8_t>)
+            {
+                DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_fp8", "+v", "a", "v", "v")
+            }
+            else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, bf8_t>)
+            {
+                DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_bf8", "+v", "a", "v", "v")
+            }
+        }
+        else if constexpr(Ctrl == WGAttrCtlEnum::Raw_vva)
+        {
+            if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, fp8_t>)
+            {
+                DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_fp8", "+v", "v", "a", "v")
+            }
+            else if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, bf8_t>)
+            {
+                DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_bf8", "+v", "v", "a", "v")
+            }
+            else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, fp8_t>)
+            {
+                DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_fp8", "+v", "v", "a", "v")
+            }
+            else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, bf8_t>)
+            {
+                DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_bf8", "+v", "v", "a", "v")
+            }
+        }
+        else
+        {
 #if defined(__gfx94__)
-        if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, fp8_t>)
-            c_vec = __builtin_amdgcn_mfma_f32_32x32x16_fp8_fp8(
-                bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
-        else if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, bf8_t>)
-            c_vec = __builtin_amdgcn_mfma_f32_32x32x16_fp8_bf8(
-                bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
-        else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, fp8_t>)
-            c_vec = __builtin_amdgcn_mfma_f32_32x32x16_bf8_fp8(
-                bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
-        else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, bf8_t>)
-            c_vec = __builtin_amdgcn_mfma_f32_32x32x16_bf8_bf8(
-                bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
+            if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, fp8_t>)
+                c_vec = __builtin_amdgcn_mfma_f32_32x32x16_fp8_fp8(
+                    bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
+            else if constexpr(std::is_same_v<ADataType, fp8_t> && std::is_same_v<BDataType, bf8_t>)
+                c_vec = __builtin_amdgcn_mfma_f32_32x32x16_fp8_bf8(
+                    bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
+            else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, fp8_t>)
+                c_vec = __builtin_amdgcn_mfma_f32_32x32x16_bf8_fp8(
+                    bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
+            else if constexpr(std::is_same_v<ADataType, bf8_t> && std::is_same_v<BDataType, bf8_t>)
+                c_vec = __builtin_amdgcn_mfma_f32_32x32x16_bf8_bf8(
+                    bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
 #elif defined(__gfx908__) || defined(__gfx90a__)
-        static_for<0, 8, 1>{}([&](auto k) {
-            float a_f32 =
-                type_convert<float>(reinterpret_cast<const thread_buffer<ADataType, 8>&>(a_vec)
-                                        .template get_as<ADataType>()[number<k>{}]);
-            float b_f32 =
-                type_convert<float>(reinterpret_cast<const thread_buffer<BDataType, 8>&>(b_vec)
-                                        .template get_as<BDataType>()[number<k>{}]);
-
-            c_vec = __builtin_amdgcn_mfma_f32_32x32x2f32(a_f32, b_f32, c_vec, 0, 0, 0);
-        });
+            static_for<0, 8, 1>{}([&](auto k) {
+                float a_f32 =
+                    type_convert<float>(reinterpret_cast<const thread_buffer<ADataType, 8>&>(a_vec)
+                                            .template get_as<ADataType>()[number<k>{}]);
+                float b_f32 =
+                    type_convert<float>(reinterpret_cast<const thread_buffer<BDataType, 8>&>(b_vec)
+                                            .template get_as<BDataType>()[number<k>{}]);
+
+                c_vec = __builtin_amdgcn_mfma_f32_32x32x2f32(a_f32, b_f32, c_vec, 0, 0, 0);
+            });
 #else
-        ignore = c_vec;
-        ignore = a_vec;
-        ignore = b_vec;
+            ck_tile::ignore = c_vec;
+            ck_tile::ignore = a_vec;
+            ck_tile::ignore = b_vec;
 #endif
+        }
     }
 
     // c_vec = a_vec * b_vec
@@ -356,20 +799,100 @@ struct WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base
         });
         return c_vec;
 #else
-        ignore = a_vec;
-        ignore = b_vec;
+        ck_tile::ignore = a_vec;
+        ck_tile::ignore = b_vec;
         return CVecType{0.f};
 #endif
     }
 };
 
+template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
 using WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8 =
-    WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<fp8_t, fp8_t>;
+    WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<fp8_t, fp8_t, Ctrl_>;
+
+template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
 using WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8 =
-    WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<fp8_t, bf8_t>;
+    WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<fp8_t, bf8_t, Ctrl_>;
+
+template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
 using WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_fp8 =
-    WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<bf8_t, fp8_t>;
+    WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<bf8_t, fp8_t, Ctrl_>;
+
+template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
 using WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8 =
-    WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<bf8_t, bf8_t>;
+    WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base<bf8_t, bf8_t, Ctrl_>;
+
+// int8
+template <WGAttrCtlEnum Ctrl_ = WGAttrCtlEnum::Default_>
+struct WarpGemmAttributeMfmaImpl_i32_32x32x16_i8
+{
+    static constexpr WGAttrCtlEnum Ctrl = Ctrl_;
+    using ADataType                     = int8_t;
+    using BDataType                     = int8_t;
+    using CDataType                     = int32_t;
+
+    using AVecType = ext_vector_t<ADataType, 8>;
+    using BVecType = ext_vector_t<BDataType, 8>;
+    using CVecType = ext_vector_t<CDataType, 16>;
+
+    static constexpr index_t kM = 32;
+    static constexpr index_t kN = 32;
+    static constexpr index_t kK = 16;
+
+    static constexpr index_t kAMBlock = 1;
+    static constexpr index_t kBNBlock = 1;
+
+    static constexpr index_t kAMLane     = 32;
+    static constexpr index_t kBNLane     = 32;
+    static constexpr index_t kABKLane    = 2;
+    static constexpr index_t kABKPerLane = 8;
+
+    static constexpr index_t kCMLane     = 2;
+    static constexpr index_t kCNLane     = 32;
+    static constexpr index_t kCM0PerLane = 4;
+    static constexpr index_t kCM1PerLane = 4;
+
+    // c_vec += a_vec * b_vec
+    template <bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CVecType& c_vec,
+                                   const AVecType& a_vec,
+                                   const BVecType& b_vec,
+                                   bool_constant<post_nop_> = {}) const
+    {
+        DISPATCH_MFMA_CTRL_("v_mfma_i32_32x32x16_i8", Ctrl)
+        else
+        {
+#if defined(__gfx94__)
+            c_vec = __builtin_amdgcn_mfma_i32_32x32x8i8(
+                bit_cast<long>(a_vec), bit_cast<long>(b_vec), c_vec, 0, 0, 0);
+#elif defined(__gfx908__) || defined(__gfx90a__)
+            static_for<0, 8, 1>{}([&](auto k) {
+                float a_f32 =
+                    type_convert<float>(reinterpret_cast<const thread_buffer<ADataType, 8>&>(a_vec)
+                                            .template get_as<ADataType>()[number<k>{}]);
+                float b_f32 =
+                    type_convert<float>(reinterpret_cast<const thread_buffer<BDataType, 8>&>(b_vec)
+                                            .template get_as<BDataType>()[number<k>{}]);
+
+                c_vec = __builtin_amdgcn_mfma_f32_32x32x2f32(a_f32, b_f32, c_vec, 0, 0, 0);
+            });
+#else
+            ck_tile::ignore = c_vec;
+            ck_tile::ignore = a_vec;
+            ck_tile::ignore = b_vec;
+#endif
+        }
+    }
+
+    // c_vec = a_vec * b_vec
+    CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const
+    {
+        CVecType c_vec{0};
+        operator()(c_vec, a_vec, b_vec);
+        return c_vec;
+    }
+};
+
+#undef DISPATCH_MFMA_
 
 } // namespace ck_tile

include/ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -21,40 +21,44 @@ struct WarpGemmMfmaDispatcher;
 
 // clang-format off
 // fp16
-template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 32, 32,  8, false> { using Type = WarpGemmMfmaF16F16F32M32N32K8; };
-template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 32, 32,  8, true> { using Type = WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution; };
-template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 32, 32, 16, false> { using Type = WarpGemmMfmaF16F16F32M32N32K16; };
-template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 32, 32, 16, true> { using Type = WarpGemmMfmaF16F16F32M32N32K16TransposedCDistribution; };
-template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 16, 16, 16, false> { using Type = WarpGemmMfmaF16F16F32M16N16K16; };
-template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 16, 16, 16, true> { using Type = WarpGemmMfmaF16F16F32M16N16K16TransposedCDistribution; };
-template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 16, 16, 32, false> { using Type = WarpGemmMfmaF16F16F32M16N16K32; };
-template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 16, 16, 32, true> { using Type = WarpGemmMfmaF16F16F32M16N16K32TransposedCDistribution; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 32, 32,  8, false> { using Type = WarpGemmMfmaF16F16F32M32N32K8; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 32, 32,  8, true> { using Type = WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 32, 32, 16, false> { using Type = WarpGemmMfmaF16F16F32M32N32K16; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 32, 32, 16, true> { using Type = WarpGemmMfmaF16F16F32M32N32K16TransposedCDistribution; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 16, 16, 16, false> { using Type = WarpGemmMfmaF16F16F32M16N16K16; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 16, 16, 16, true> { using Type = WarpGemmMfmaF16F16F32M16N16K16TransposedCDistribution; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 16, 16, 32, false> { using Type = WarpGemmMfmaF16F16F32M16N16K32; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 16, 16, 32, true> { using Type = WarpGemmMfmaF16F16F32M16N16K32TransposedCDistribution; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 4, 64, 16, false> { using Type = WarpGemmMfmaF16F16F32M4N64K16; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 64, 4, 16, false> { using Type = WarpGemmMfmaF16F16F32M64N4K16; };
 
-template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 32, 32,  8, false, true> { using Type = WarpGemmMfmaF16F16F32M32N32K8SwizzleA; };
-template<> struct WarpGemmMfmaDispatcher<half_t, half_t, float, 32, 32, 16, false, true> { using Type = WarpGemmMfmaF16F16F32M32N32K16SwizzleA; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 32, 32,  8, false, true> { using Type = WarpGemmMfmaF16F16F32M32N32K8SwizzleA; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::half_t, ck_tile::half_t, float, 32, 32, 16, false, true> { using Type = WarpGemmMfmaF16F16F32M32N32K16SwizzleA; };
 
 // bf16
-template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 32, 32,  8, false> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K8; };
-template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 32, 32,  8, true> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution; };
-template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 32, 32, 16, false> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16; };
-template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 32, 32, 16, true> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16TransposedCDistribution; };
-template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 16, 16, 16, false> { using Type = WarpGemmMfmaBf16Bf16F32M16N16K16; };
-template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 16, 16, 16, true> { using Type = WarpGemmMfmaBf16Bf16F32M16N16K16TransposedCDistribution; };
-template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 16, 16, 32, false> { using Type = WarpGemmMfmaBf16Bf16F32M16N16K32; };
-template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 16, 16, 32, true> { using Type = WarpGemmMfmaBf16Bf16F32M16N16K32TransposedCDistribution; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 32, 32,  8, false> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K8; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 32, 32,  8, true> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 32, 32, 16, false> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 32, 32, 16, true> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16TransposedCDistribution; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 16, 16, 16, false> { using Type = WarpGemmMfmaBf16Bf16F32M16N16K16; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 16, 16, 16, true> { using Type = WarpGemmMfmaBf16Bf16F32M16N16K16TransposedCDistribution; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 16, 16, 32, false> { using Type = WarpGemmMfmaBf16Bf16F32M16N16K32; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 16, 16, 32, true> { using Type = WarpGemmMfmaBf16Bf16F32M16N16K32TransposedCDistribution; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 4, 64, 16, false> { using Type = WarpGemmMfmaBf16Bf16F32M4N64K16; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 64, 4, 16, false> { using Type = WarpGemmMfmaBf16Bf16F32M64N4K16; };
 
-template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 32, 32,  8, false, true> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K8SwizzleA; };
-template<> struct WarpGemmMfmaDispatcher<bf16_t, bf16_t, float, 32, 32, 16, false, true> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleA; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 32, 32,  8, false, true> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K8SwizzleA; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::bf16_t, ck_tile::bf16_t, float, 32, 32, 16, false, true> { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleA; };
 
 // fp8
-template<> struct WarpGemmMfmaDispatcher<fp8_t, fp8_t, float, 32, 32,  16, false> { using Type = WarpGemmMfma_f32_32x32x16_fp8_fp8; };
-template<> struct WarpGemmMfmaDispatcher<fp8_t, fp8_t, float, 32, 32,  16, true> { using Type = WarpGemmMfma_f32_32x32x16_fp8_fp8_CTransposed; };
-template<> struct WarpGemmMfmaDispatcher<fp8_t, bf8_t, float, 32, 32,  16, false> { using Type = WarpGemmMfma_f32_32x32x16_fp8_bf8; };
-template<> struct WarpGemmMfmaDispatcher<fp8_t, bf8_t, float, 32, 32,  16, true> { using Type = WarpGemmMfma_f32_32x32x16_fp8_bf8_CTransposed; };
-template<> struct WarpGemmMfmaDispatcher<bf8_t, fp8_t, float, 32, 32,  16, false> { using Type = WarpGemmMfma_f32_32x32x16_bf8_fp8; };
-template<> struct WarpGemmMfmaDispatcher<bf8_t, fp8_t, float, 32, 32,  16, true> { using Type = WarpGemmMfma_f32_32x32x16_bf8_fp8_CTransposed; };
-template<> struct WarpGemmMfmaDispatcher<bf8_t, bf8_t, float, 32, 32,  16, false> { using Type = WarpGemmMfma_f32_32x32x16_bf8_bf8; };
-template<> struct WarpGemmMfmaDispatcher<bf8_t, bf8_t, float, 32, 32,  16, true> { using Type = WarpGemmMfma_f32_32x32x16_bf8_bf8_CTransposed; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::fp8_t, ck_tile::fp8_t, float, 32, 32,  16, false> { using Type = WarpGemmMfma_f32_32x32x16_fp8_fp8; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::fp8_t, ck_tile::fp8_t, float, 32, 32,  16, true> { using Type = WarpGemmMfma_f32_32x32x16_fp8_fp8_CTransposed; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::fp8_t, ck_tile::bf8_t, float, 32, 32,  16, false> { using Type = WarpGemmMfma_f32_32x32x16_fp8_bf8; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::fp8_t, ck_tile::bf8_t, float, 32, 32,  16, true> { using Type = WarpGemmMfma_f32_32x32x16_fp8_bf8_CTransposed; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::bf8_t, ck_tile::fp8_t, float, 32, 32,  16, false> { using Type = WarpGemmMfma_f32_32x32x16_bf8_fp8; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::bf8_t, ck_tile::fp8_t, float, 32, 32,  16, true> { using Type = WarpGemmMfma_f32_32x32x16_bf8_fp8_CTransposed; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::bf8_t, ck_tile::bf8_t, float, 32, 32,  16, false> { using Type = WarpGemmMfma_f32_32x32x16_bf8_bf8; };
+template<> struct WarpGemmMfmaDispatcher<ck_tile::bf8_t, ck_tile::bf8_t, float, 32, 32,  16, true> { using Type = WarpGemmMfma_f32_32x32x16_bf8_bf8_CTransposed; };
 
 // clang-format on
 } // namespace impl

include/ck_tile/ops/gemm/warp/warp_gemm_impl.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -14,6 +14,11 @@ struct WarpGemmImpl
     static constexpr index_t kM = WarpGemmAttribute::kM;
     static constexpr index_t kN = WarpGemmAttribute::kN;
     static constexpr index_t kK = WarpGemmAttribute::kK;
+    /// @brief The number of elements in K dimension processed by single thread in wavefront.
+    ///
+    /// @note  Note that WarpGemm may run MFMA instruction multiple times (on different K).
+    ///        In such situation this value reflects this fact.
+    static constexpr index_t kKPerThread = WarpGemmAttribute::kKPerThread;
 
     using ADataType = typename WarpGemmAttribute::ADataType;
     using BDataType = typename WarpGemmAttribute::BDataType;
@@ -31,11 +36,21 @@ struct WarpGemmImpl
     using BWarpTensor = static_distributed_tensor<BDataType, BWarpDstr>;
     using CWarpTensor = static_distributed_tensor<CDataType, CWarpDstr>;
 
-    CK_TILE_DEVICE void operator()(CWarpTensor& c, const AWarpTensor& a, const BWarpTensor& b) const
+    CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access()
     {
-        using AVec = ext_vector_t<ADataType, AWarpTensor::get_thread_buffer_size()>;
-        using BVec = ext_vector_t<BDataType, BWarpTensor::get_thread_buffer_size()>;
-        using CVec = ext_vector_t<CDataType, CWarpTensor::get_thread_buffer_size()>;
+        return WarpGemmAttribute_::get_num_of_access();
+    }
+
+    template <typename CTensor, typename ATensor, typename BTensor, bool post_nop_ = false>
+    CK_TILE_DEVICE void
+    operator()(CTensor& c, const ATensor& a, const BTensor& b, bool_constant<post_nop_> = {}) const
+    {
+        static_assert(detail::is_similiar_distributed_tensor_v<CTensor, CWarpTensor> &&
+                      detail::is_similiar_distributed_tensor_v<ATensor, AWarpTensor> &&
+                      detail::is_similiar_distributed_tensor_v<BTensor, BWarpTensor>);
+        using AVec = ext_vector_t<ADataType, ATensor::get_thread_buffer_size()>;
+        using BVec = ext_vector_t<BDataType, BTensor::get_thread_buffer_size()>;
+        using CVec = ext_vector_t<CDataType, CTensor::get_thread_buffer_size()>;
 
         constexpr auto I0 = number<0>{};
 
@@ -44,18 +59,49 @@ struct WarpGemmImpl
         auto c_vec       = c.get_thread_buffer().template get_as<CVec>()[I0];
 
         // c_vec += a_vec * b_vec
-        WarpGemmAttribute{}(c_vec, a_vec, b_vec);
+        WarpGemmAttribute{}(c_vec, a_vec, b_vec, bool_constant<post_nop_>{});
 
         c.get_thread_buffer().template set_as<CVec>(I0, c_vec);
     }
 
-    CK_TILE_DEVICE auto operator()(const AWarpTensor& a, const BWarpTensor& b) const
+    template <typename CTensor,
+              typename ATensor,
+              typename BTensor,
+              index_t i_subk,
+              bool post_nop_ = false>
+    CK_TILE_DEVICE void operator()(CTensor& c,
+                                   const ATensor& a,
+                                   const BTensor& b,
+                                   number<i_subk>,
+                                   bool_constant<post_nop_> = {}) const
     {
-        CWarpTensor c;
+        using AVec = ext_vector_t<ADataType, ATensor::get_thread_buffer_size()>;
+        using BVec = ext_vector_t<BDataType, BTensor::get_thread_buffer_size()>;
+        using CVec = ext_vector_t<CDataType, CTensor::get_thread_buffer_size()>;
+
+        constexpr auto I0 = number<0>{};
 
-        using AVec = ext_vector_t<ADataType, AWarpTensor::get_thread_buffer_size()>;
-        using BVec = ext_vector_t<BDataType, BWarpTensor::get_thread_buffer_size()>;
-        using CVec = ext_vector_t<CDataType, CWarpTensor::get_thread_buffer_size()>;
+        const auto a_vec = a.get_thread_buffer().template get_as<AVec>()[I0];
+        const auto b_vec = b.get_thread_buffer().template get_as<BVec>()[I0];
+        auto c_vec       = c.get_thread_buffer().template get_as<CVec>()[I0];
+
+        // c_vec += a_vec * b_vec
+        WarpGemmAttribute{}(c_vec, a_vec, b_vec, number<i_subk>{}, bool_constant<post_nop_>{});
+
+        c.get_thread_buffer().template set_as<CVec>(I0, c_vec);
+    }
+
+    template <typename ATensor, typename BTensor>
+    CK_TILE_DEVICE auto operator()(const ATensor& a, const BTensor& b) const
+    {
+        using CTensor = CWarpTensor;
+        static_assert(detail::is_similiar_distributed_tensor_v<ATensor, AWarpTensor> &&
+                      detail::is_similiar_distributed_tensor_v<BTensor, BWarpTensor>);
+        CTensor c;
+
+        using AVec = ext_vector_t<ADataType, ATensor::get_thread_buffer_size()>;
+        using BVec = ext_vector_t<BDataType, BTensor::get_thread_buffer_size()>;
+        using CVec = ext_vector_t<CDataType, CTensor::get_thread_buffer_size()>;
 
         constexpr auto I0 = number<0>{};
 

include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_one_pass.hpp

@@ -121,7 +121,7 @@ struct Layernorm2dFwdPipelineOnePass
         auto [mean, var] = block_welford(acc, cur_count, max_count);
         block_welford_sync(mean, var, cur_count);
         block_welford_cross_warp_sync(mean, var, cur_count, smem);
-        block_tile_welford_post_scale_var(var, cur_count);
+        block_tile_welford_post_scale_var(var, cur_count, constant<kFastFDiv>{});
 
         // compute inv-std
         auto inv_std = tile_elementwise_in(

include/ck_tile/ops/layernorm2d/pipeline/layernorm2d_fwd_pipeline_two_pass.hpp

@@ -35,6 +35,7 @@ struct Layernorm2dFwdPipelineTwoPass
     static constexpr bool kNeedCrossWarpSync = Problem::kNeedCrossWarpSync;
     static constexpr bool kPadM              = false; // TODO - BlockLayernorm2dFwdProblem::kPadM
     static constexpr bool kPadN              = Problem::Traits::kPadN;
+    static constexpr bool kFastFDiv          = Problem::Traits::kFastFDiv;
     static constexpr auto kFusedAdd          = Problem::Traits::kFusedAdd;
     static constexpr auto kFusedQuant        = Problem::Traits::kFusedQuant;
 
@@ -137,15 +138,22 @@ struct Layernorm2dFwdPipelineTwoPass
 
         block_welford_sync(mean, var, cur_count);
         block_welford_cross_warp_sync(mean, var, cur_count, smem);
-        block_tile_welford_post_scale_var(var, cur_count);
+        block_tile_welford_post_scale_var(var, cur_count, constant<kFastFDiv>{});
 
         // compute inv-std
         auto inv_std = tile_elementwise_in(
             [&](const auto& v_) {
-                return type_convert<ComputeDataType>(1.0f) / (sqrt(v_ + epsilon));
+                if(kFastFDiv && std::is_same_v<ComputeDataType, float>)
+                {
+                    return type_convert<ComputeDataType>(1.0f) *
+                           __builtin_amdgcn_rcpf(sqrt(v_ + epsilon));
+                }
+                else
+                {
+                    return type_convert<ComputeDataType>(1.0f) / sqrt(v_ + epsilon);
+                }
             },
             var);
-
         if constexpr(kSaveMean)
             store_tile(mean_window, cast_tile<MeanDataType>(mean));
         if constexpr(kSaveInvStd)