Export ROCm/rocm-libraries@2d4a3223cb

include/ck_tile/ops/gemm/block/block_gemm_areg_bgmem_creg_v1.hpp

@@ -0,0 +1,136 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/block/block_gemm_areg_bgmem_creg_v1_default_policy.hpp"
+#include "ck_tile/ops/gemm/block/block_gemm_problem.hpp"
+
+namespace ck_tile {
+
+// A is block distributed tensor
+// B is block window on global memory
+// C is block distributed tensor
+// This will:
+//   1. load B from global memory into shared memory and then
+//   2. Call BlockGemmARegSGmemCRegV1
+template <typename Problem_, typename Policy_ = BlockGemmARegBGmemCRegV1DefaultPolicy>
+struct BlockGemmARegBGmemCRegV1
+{
+    using Problem        = remove_cvref_t<Problem_>;
+    using Policy         = remove_cvref_t<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>;
+
+    static constexpr index_t kBlockSize = Problem::kBlockSize;
+
+    // use BlockGemmARegBSmemCRegV1 as the underlying block-GEMM implementation
+    using BlockGemmARegBGmemCRegImpl = BlockGemmARegBGmemCRegV1<
+        BlockGemmProblem<ADataType, BDataType, CDataType, kBlockSize, BlockGemmShape>,
+        BlockGemmARegBGmemCRegV1DefaultPolicy>;
+
+    CK_TILE_HOST_DEVICE static constexpr index_t GetStaticLdsSize()
+    {
+        return sizeof(BDataType) *
+               Policy::template MakeBSmemBlockDescriptor<Problem>().get_element_space_size();
+    }
+
+    // C += A * B
+    template <typename CBlockTensor, typename ABlockTensor, typename BBlockGmemWindowTmp>
+    CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
+                                   const ABlockTensor& a_block_tensor,
+                                   const BBlockGmemWindowTmp& b_block_gmem_window_tmp,
+                                   void* smem_ptr) const
+    {
+        static_assert(
+            std::is_same_v<ADataType, remove_cv_t<typename ABlockTensor::DataType>> &&
+                std::is_same_v<BDataType, remove_cv_t<typename BBlockGmemWindowTmp::DataType>> &&
+                std::is_same_v<CDataType, remove_cv_t<typename CBlockTensor::DataType>>,
+            "wrong!");
+
+        constexpr index_t MPerBlock = ABlockTensor{}.get_lengths()[number<0>{}];
+        constexpr index_t NPerBlock = BBlockGmemWindowTmp{}.get_window_lengths()[number<0>{}];
+        constexpr index_t KPerBlock = ABlockTensor{}.get_lengths()[number<1>{}];
+
+        static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
+                          KPerBlock == BlockGemmShape::kK,
+                      "wrong!");
+
+        const auto b_block_gmem_window =
+            make_tile_window(b_block_gmem_window_tmp.get_bottom_tensor_view(),
+                             make_tuple(number<NPerBlock>{}, number<KPerBlock>{}),
+                             b_block_gmem_window_tmp.get_window_origin(),
+                             Policy::template MakeBGmemTileDistribution<Problem>());
+
+        // B LDS and LDS window
+        auto b_block_smem = make_tensor_view<address_space_enum::lds>(
+            reinterpret_cast<BDataType*>(smem_ptr),
+            Policy::template MakeBSmemBlockDescriptor<Problem>());
+
+        auto b_block_smem_window = make_tile_window(
+            b_block_smem, make_tuple(number<MPerBlock>{}, number<KPerBlock>{}), {0, 0});
+
+        // load B tile from global mem
+        const auto b_block_tile = load_tile(b_block_gmem_window);
+
+        // store B tile into shared mem
+        store_tile(b_block_smem_window, b_block_tile);
+
+        // wait for store_tile to finish
+        block_sync_lds();
+
+        // block GEMM
+        BlockGemmARegBGmemCRegImpl{}(c_block_tensor, a_block_tensor, b_block_smem_window);
+    }
+
+    // C = A * B
+    template <typename ABlockTensor, typename BBlockGmemWindowTmp>
+    CK_TILE_DEVICE auto operator()(const ABlockTensor& a_block_tensor,
+                                   const BBlockGmemWindowTmp& b_block_gmem_window_tmp,
+                                   void* smem_ptr) const
+    {
+        static_assert(
+            std::is_same_v<ADataType, remove_cv_t<typename ABlockTensor::DataType>> &&
+                std::is_same_v<BDataType, remove_cv_t<typename BBlockGmemWindowTmp::DataType>>,
+            "wrong!");
+
+        constexpr index_t MPerBlock = ABlockTensor{}.get_lengths()[number<0>{}];
+        constexpr index_t NPerBlock = BBlockGmemWindowTmp{}.get_window_lengths()[number<0>{}];
+        constexpr index_t KPerBlock = ABlockTensor{}.get_lengths()[number<1>{}];
+
+        static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
+                          KPerBlock == BlockGemmShape::kK,
+                      "wrong!");
+
+        const auto b_block_gmem_window =
+            make_tile_window(b_block_gmem_window_tmp.get_bottom_tensor_view(),
+                             make_tuple(number<NPerBlock>{}, number<KPerBlock>{}),
+                             b_block_gmem_window_tmp.get_window_origin(),
+                             Policy::template MakeBGmemTileDistribution<Problem>());
+
+        // B LDS and LDS window
+        auto b_block_smem = make_tensor_view<address_space_enum::lds>(
+            reinterpret_cast<BDataType*>(smem_ptr),
+            Policy::template MakeBSmemBlockDescriptor<Problem>());
+
+        auto b_block_smem_window = make_tile_window(
+            b_block_smem, make_tuple(number<MPerBlock>{}, number<KPerBlock>{}), {0, 0});
+
+        // load B tile from global mem
+        const auto b_block_tile = load_tile(b_block_gmem_window);
+
+        // store B tile into shared mem
+        store_tile(b_block_smem_window, b_block_tile);
+
+        // wait for store_tile to finish
+        block_sync_lds();
+
+        // block GEMM
+        return BlockGemmARegBGmemCRegImpl{}(a_block_tensor, b_block_smem_window);
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_areg_bgmem_creg_v1_default_policy.hpp

@@ -0,0 +1,110 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+
+namespace ck_tile {
+
+// Default policy for BlockGemmARegBGmemCRegV1
+// Default policy class should not be templated, put template on member functions instead
+struct BlockGemmARegBGmemCRegV1DefaultPolicy
+{
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeBGmemTileDistribution()
+    {
+        using BDataType = remove_cvref_t<typename Problem::BDataType>;
+
+        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;
+        constexpr index_t N1 = kBlockSize / get_warp_size();
+        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>>{});
+    }
+
+#if 0
+    // 2d
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeBLdsBlockDescriptor()
+    {
+        constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr auto b_lds_block_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(kNPerBlock, kKPerBlock), number<32>{});
+
+        return b_lds_block_desc;
+    }
+#elif 0
+    // 3d + padding
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeBSmemBlockDescriptor()
+    {
+        constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr auto b_lds_block_desc_0 = make_naive_tensor_descriptor(
+            make_tuple(number<kKPerBlock / 8>{}, number<kNPerBlock>{}, number<8>{}),
+            make_tuple(number<(kNPerBlock + 1) * 8>{}, number<8>{}, number<1>{}),
+            number<8>{},
+            number<1>{});
+
+        constexpr auto b_lds_block_desc = transform_tensor_descriptor(
+            b_lds_block_desc_0,
+            make_tuple(make_pass_through_transform(kNPerBlock),
+                       make_merge_transform(make_tuple(kKPerBlock / 8, 8))),
+            make_tuple(sequence<1>{}, sequence<0, 2>{}),
+            make_tuple(sequence<0>{}, sequence<1>{}));
+
+        return b_lds_block_desc;
+    }
+#elif 1
+    // fake XOR
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeBSmemBlockDescriptor()
+    {
+        using BDataType = remove_cvref_t<typename Problem::BDataType>;
+
+        constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr auto b_lds_block_desc_d1_d2_d3 = make_naive_tensor_descriptor_packed(
+            make_tuple(number<kNPerBlock / 2>{}, number<2>{}, number<kKPerBlock>{}),
+            number<kKPerBlock>{});
+
+        constexpr index_t kK1 = 16 / sizeof(BDataType);
+
+        constexpr auto b_lds_block_desc_d4_d5_d6 = transform_tensor_descriptor(
+            b_lds_block_desc_d1_d2_d3,
+            make_tuple(
+                make_xor_transform(make_tuple(number<kNPerBlock / 2>{}, number<kKPerBlock>{}), kK1),
+                make_pass_through_transform(2)),
+            make_tuple(sequence<0, 2>{}, sequence<1>{}),
+            make_tuple(sequence<0, 2>{}, sequence<1>{}));
+
+        constexpr auto b_lds_block_desc_n_k = transform_tensor_descriptor(
+            b_lds_block_desc_d4_d5_d6,
+            make_tuple(make_merge_transform(make_tuple(number<kNPerBlock / 2>{}, number<2>{})),
+                       make_pass_through_transform(kKPerBlock)),
+            make_tuple(sequence<0, 1>{}, sequence<2>{}),
+            make_tuple(sequence<0>{}, sequence<1>{}));
+
+        return b_lds_block_desc_n_k;
+    }
+#endif
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_areg_breg_creg_v1.hpp

@@ -0,0 +1,407 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/block/block_gemm_areg_breg_creg_v1_default_policy.hpp"
+
+namespace ck_tile {
+
+// A is block distributed tensor
+// B is block distributed tensor
+// C is block distributed tensor
+template <typename Problem_,
+          typename Policy_ = BlockGemmARegBRegCRegV1DefaultPolicy,
+          bool TransposeC_ = false>
+struct BlockGemmARegBRegCRegV1
+{
+    private:
+    template <typename PipelineProblem_, typename GemmPolicy_>
+    struct GemmTraits_
+    {
+        using Problem        = remove_cvref_t<PipelineProblem_>;
+        using Policy         = remove_cvref_t<GemmPolicy_>;
+        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>;
+
+        static constexpr index_t kBlockSize = 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 auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+        using WarpGemm               = remove_cvref_t<decltype(config.template at<0>())>;
+
+        static constexpr index_t MWarp        = config.template at<1>();
+        static constexpr index_t NWarp        = config.template at<2>();
+        static constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WarpGemm::kM);
+        static constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WarpGemm::kN);
+        static constexpr index_t KIterPerWarp = KPerBlock / WarpGemm::kK;
+
+        static constexpr index_t KPack = WarpGemm::kKPerThread;
+    };
+
+    public:
+    using Problem                    = remove_cvref_t<Problem_>;
+    using Policy                     = remove_cvref_t<Policy_>;
+    static constexpr bool TransposeC = TransposeC_;
+
+    using Traits = GemmTraits_<Problem, Policy>;
+
+    using WarpGemm       = typename Traits::WarpGemm;
+    using BlockGemmShape = typename Traits::BlockGemmShape;
+
+    using ADataType = remove_cvref_t<typename Traits::ADataType>;
+    using BDataType = remove_cvref_t<typename Traits::BDataType>;
+    using CDataType = remove_cvref_t<typename Traits::CDataType>;
+
+    static constexpr index_t KIterPerWarp = Traits::KIterPerWarp;
+    static constexpr index_t MIterPerWarp = Traits::MIterPerWarp;
+    static constexpr index_t NIterPerWarp = Traits::NIterPerWarp;
+
+    static constexpr index_t MWarp            = Traits::MWarp;
+    static constexpr index_t NWarp            = Traits::NWarp;
+    static constexpr bool UseDefaultScheduler = (Problem::NumWaveGroups != 1);
+
+    CK_TILE_DEVICE static constexpr auto MakeABlockDistributionEncode()
+    {
+        if constexpr(UseDefaultScheduler)
+        {
+            constexpr auto a_block_outer_dstr_encoding =
+                tile_distribution_encoding<sequence<NWarp>,
+                                           tuple<sequence<MIterPerWarp>, sequence<KIterPerWarp>>,
+                                           tuple<>,
+                                           tuple<>,
+                                           sequence<1, 2>,
+                                           sequence<0, 0>>{};
+
+            constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+                a_block_outer_dstr_encoding, typename WarpGemm::AWarpDstrEncoding{});
+
+            return a_block_dstr_encode;
+        }
+        else
+        {
+            constexpr auto a_block_outer_dstr_encoding = tile_distribution_encoding<
+                sequence<NWarp>,
+                tuple<sequence<MIterPerWarp, MWarp>, sequence<KIterPerWarp>>,
+                tuple<sequence<1, 0>>,
+                tuple<sequence<1, 0>>,
+                sequence<1, 2>,
+                sequence<0, 0>>{};
+            constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+                a_block_outer_dstr_encoding, typename WarpGemm::AWarpDstrEncoding{});
+
+            return a_block_dstr_encode;
+        }
+    }
+
+    CK_TILE_DEVICE static constexpr auto MakeBBlockDistributionEncode()
+    {
+        if constexpr(UseDefaultScheduler)
+        {
+            constexpr auto b_block_outer_dstr_encoding =
+                tile_distribution_encoding<sequence<MWarp>,
+                                           tuple<sequence<NIterPerWarp>, sequence<KIterPerWarp>>,
+                                           tuple<>,
+                                           tuple<>,
+                                           sequence<1, 2>,
+                                           sequence<0, 0>>{};
+            constexpr auto b_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+                b_block_outer_dstr_encoding, typename WarpGemm::BWarpDstrEncoding{});
+
+            return b_block_dstr_encode;
+        }
+        else
+        {
+            constexpr auto b_block_outer_dstr_encoding = tile_distribution_encoding<
+                sequence<MWarp>,
+                tuple<sequence<NIterPerWarp, NWarp>, sequence<KIterPerWarp>>,
+                tuple<sequence<0, 1>>,
+                tuple<sequence<0, 1>>,
+                sequence<1, 2>,
+                sequence<0, 0>>{};
+            constexpr auto b_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+                b_block_outer_dstr_encoding, typename WarpGemm::BWarpDstrEncoding{});
+
+            return b_block_dstr_encode;
+        }
+    }
+
+    CK_TILE_DEVICE static constexpr auto MakeCBlockDistributionEncode()
+    {
+        using c_distr_ys_major = std::conditional_t<TransposeC, sequence<2, 1>, sequence<1, 2>>;
+        if constexpr(UseDefaultScheduler)
+        {
+            constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+                sequence<MWarp>,
+                tuple<sequence<MIterPerWarp>, sequence<NIterPerWarp, NWarp>>,
+                tuple<>,
+                tuple<>,
+                c_distr_ys_major,
+                sequence<0, 0>>{};
+            constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+                c_block_outer_dstr_encoding, typename WarpGemm::CWarpDstrEncoding{});
+
+            return c_block_dstr_encode;
+        }
+        else
+        {
+            constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
+                tuple<sequence<1, 2>>,
+                tuple<sequence<1, 1>>,
+                c_distr_ys_major,
+                sequence<0, 0>>{};
+            constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+                c_block_outer_dstr_encoding, typename WarpGemm::CWarpDstrEncoding{});
+
+            return c_block_dstr_encode;
+        }
+    }
+
+    // C += A * B
+    template <typename CBlockTensor, typename ABlockTensor, typename BBlockTensor>
+    CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
+                                   const ABlockTensor& a_block_tensor,
+                                   const BBlockTensor& b_block_tensor) const
+    {
+        static_assert(std::is_same_v<ADataType, remove_cv_t<typename ABlockTensor::DataType>> &&
+                          std::is_same_v<BDataType, remove_cv_t<typename BBlockTensor::DataType>> &&
+                          std::is_same_v<CDataType, remove_cv_t<typename CBlockTensor::DataType>>,
+                      "wrong!");
+
+        // check ABC-block-distribution
+        static_assert(
+            std::is_same_v<remove_cvref_t<decltype(MakeABlockDistributionEncode())>,
+                           remove_cvref_t<decltype(ABlockTensor::get_tile_distribution()
+                                                       .get_static_tile_distribution_encoding())>>,
+            "A distribution is wrong!");
+        static_assert(
+            std::is_same_v<remove_cvref_t<decltype(MakeBBlockDistributionEncode())>,
+                           remove_cvref_t<decltype(BBlockTensor::get_tile_distribution()
+                                                       .get_static_tile_distribution_encoding())>>,
+            "B distribution is wrong!");
+        static_assert(
+            std::is_same_v<remove_cvref_t<decltype(MakeCBlockDistributionEncode())>,
+                           remove_cvref_t<decltype(CBlockTensor::get_tile_distribution()
+                                                       .get_static_tile_distribution_encoding())>>,
+            "C distribution is wrong!");
+
+        using AWarpDstr = typename WarpGemm::AWarpDstr;
+        using BWarpDstr = typename WarpGemm::BWarpDstr;
+        using CWarpDstr = typename WarpGemm::CWarpDstr;
+
+        using AWarpTensor = typename WarpGemm::AWarpTensor;
+        using BWarpTensor = typename WarpGemm::BWarpTensor;
+        using CWarpTensor = typename WarpGemm::CWarpTensor;
+
+        constexpr auto a_warp_y_lengths =
+            to_sequence(AWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+        constexpr auto b_warp_y_lengths =
+            to_sequence(BWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+        constexpr auto c_warp_y_lengths =
+            to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+
+        constexpr auto a_warp_y_index_zeros = uniform_sequence_gen_t<AWarpDstr::NDimY, 0>{};
+        constexpr auto b_warp_y_index_zeros = uniform_sequence_gen_t<BWarpDstr::NDimY, 0>{};
+        constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
+
+        // hot loop:
+        static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+            static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                // read A warp tensor from A Block window
+                AWarpTensor a_warp_tensor;
+                a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
+                    merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
+                    merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
+
+                static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                    // read B warp tensor from B block tensor
+                    BWarpTensor b_warp_tensor;
+                    b_warp_tensor.get_thread_buffer() = b_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(sequence<nIter, kIter>{}, b_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
+
+                    // read C warp tensor from C block tensor
+                    using c_iter_idx = std::
+                        conditional_t<TransposeC, sequence<nIter, mIter>, sequence<mIter, nIter>>;
+                    CWarpTensor c_warp_tensor;
+                    c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(c_iter_idx{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
+
+                    // warp GEMM
+                    WarpGemm{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
+
+                    // write C warp tensor into C block tensor
+                    c_block_tensor.set_y_sliced_thread_data(
+                        merge_sequences(c_iter_idx{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
+                        c_warp_tensor.get_thread_buffer());
+                });
+            });
+        });
+    }
+
+    // C += A * B with MX scaling
+    // ScaleATensor: [MIterPerWarp, KIterPerWarp] -> int32_t
+    // ScaleBTensor: [NIterPerWarp, KIterPerWarp] -> int32_t
+    template <typename CBlockTensor,
+              typename ABlockTensor,
+              typename BBlockTensor,
+              typename ScaleATensor,
+              typename ScaleBTensor>
+    CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
+                                   const ABlockTensor& a_block_tensor,
+                                   const BBlockTensor& b_block_tensor,
+                                   const ScaleATensor& scale_a_tensor,
+                                   const ScaleBTensor& scale_b_tensor) const
+    {
+        static_assert(std::is_same_v<ADataType, remove_cv_t<typename ABlockTensor::DataType>> &&
+                          std::is_same_v<BDataType, remove_cv_t<typename BBlockTensor::DataType>> &&
+                          std::is_same_v<CDataType, remove_cv_t<typename CBlockTensor::DataType>>,
+                      "wrong!");
+
+        // check ABC-block-distribution
+        static_assert(
+            std::is_same_v<remove_cvref_t<decltype(MakeABlockDistributionEncode())>,
+                           remove_cvref_t<decltype(ABlockTensor::get_tile_distribution()
+                                                       .get_static_tile_distribution_encoding())>>,
+            "A distribution is wrong!");
+        static_assert(
+            std::is_same_v<remove_cvref_t<decltype(MakeBBlockDistributionEncode())>,
+                           remove_cvref_t<decltype(BBlockTensor::get_tile_distribution()
+                                                       .get_static_tile_distribution_encoding())>>,
+            "B distribution is wrong!");
+        static_assert(
+            std::is_same_v<remove_cvref_t<decltype(MakeCBlockDistributionEncode())>,
+                           remove_cvref_t<decltype(CBlockTensor::get_tile_distribution()
+                                                       .get_static_tile_distribution_encoding())>>,
+            "C distribution is wrong!");
+
+        using AWarpDstr = typename WarpGemm::AWarpDstr;
+        using BWarpDstr = typename WarpGemm::BWarpDstr;
+        using CWarpDstr = typename WarpGemm::CWarpDstr;
+
+        using AWarpTensor = typename WarpGemm::AWarpTensor;
+        using BWarpTensor = typename WarpGemm::BWarpTensor;
+        using CWarpTensor = typename WarpGemm::CWarpTensor;
+
+        constexpr auto a_warp_y_lengths =
+            to_sequence(AWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+        constexpr auto b_warp_y_lengths =
+            to_sequence(BWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+        constexpr auto c_warp_y_lengths =
+            to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+
+        constexpr auto a_warp_y_index_zeros = uniform_sequence_gen_t<AWarpDstr::NDimY, 0>{};
+        constexpr auto b_warp_y_index_zeros = uniform_sequence_gen_t<BWarpDstr::NDimY, 0>{};
+        constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
+
+        // hot loop with MX scaling:
+        static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+            static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                // read A warp tensor from A Block window
+                AWarpTensor a_warp_tensor;
+                a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
+                    merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
+                    merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
+
+                // get A scale for this M-K tile using get_y_sliced_thread_data
+                auto scale_a_slice = scale_a_tensor.get_y_sliced_thread_data(
+                    sequence<kIter, mIter, 0>{}, sequence<1, 1, 1>{});
+                const auto a_scale_e8m0 = scale_a_slice[number<0>{}];
+                const int32_t a_scale   = static_cast<int32_t>(a_scale_e8m0.get());
+
+                static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                    // read B warp tensor from B block tensor
+                    BWarpTensor b_warp_tensor;
+                    b_warp_tensor.get_thread_buffer() = b_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(sequence<nIter, kIter>{}, b_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
+
+                    // get B scale for this N-K tile using get_y_sliced_thread_data
+                    auto scale_b_slice = scale_b_tensor.get_y_sliced_thread_data(
+                        sequence<kIter, nIter, 0>{}, sequence<1, 1, 1>{});
+                    const auto b_scale_e8m0 = scale_b_slice[number<0>{}];
+                    const int32_t b_scale   = static_cast<int32_t>(b_scale_e8m0.get());
+
+                    // read C warp tensor from C block tensor
+                    using c_iter_idx = std::
+                        conditional_t<TransposeC, sequence<nIter, mIter>, sequence<mIter, nIter>>;
+                    CWarpTensor c_warp_tensor;
+                    c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(c_iter_idx{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
+
+                    // warp GEMM with MX scaling
+                    // Cast e8m0_t to int32_t, use OpSel=0 (least significant byte)
+                    constexpr index_t kOpSel = 0; // Always use OpSel=0
+                    WarpGemm{}.template operator()<kOpSel, kOpSel>(
+                        c_warp_tensor, a_warp_tensor, b_warp_tensor, a_scale, b_scale);
+
+                    // write C warp tensor into C block tensor
+                    c_block_tensor.set_y_sliced_thread_data(
+                        merge_sequences(c_iter_idx{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
+                        c_warp_tensor.get_thread_buffer());
+                });
+            });
+        });
+    }
+
+    CK_TILE_DEVICE static constexpr auto MakeCBlockTile()
+    {
+        using c_distr_ys_major = std::conditional_t<TransposeC, sequence<2, 1>, sequence<1, 2>>;
+        if constexpr(UseDefaultScheduler)
+        {
+            constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+                sequence<MWarp>,
+                tuple<sequence<MIterPerWarp>, sequence<NIterPerWarp, NWarp>>,
+                tuple<>,
+                tuple<>,
+                c_distr_ys_major,
+                sequence<0, 0>>{};
+
+            constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+                c_block_outer_dstr_encoding, typename WarpGemm::CWarpDstrEncoding{});
+            constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
+            auto c_block_tensor         = make_static_distributed_tensor<CDataType>(c_block_dstr);
+            return c_block_tensor;
+        }
+        else
+        {
+            constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
+                tuple<sequence<1, 2>>,
+                tuple<sequence<1, 1>>,
+                c_distr_ys_major,
+                sequence<0, 0>>{};
+
+            constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+                c_block_outer_dstr_encoding, typename WarpGemm::CWarpDstrEncoding{});
+            constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
+            auto c_block_tensor         = make_static_distributed_tensor<CDataType>(c_block_dstr);
+            return c_block_tensor;
+        }
+    }
+
+    // C = A * B
+    template <typename ABlockTensor, typename BBlockTensor>
+    CK_TILE_DEVICE auto operator()(const ABlockTensor& a_block_tensor,
+                                   const BBlockTensor& b_block_tensor) const
+    {
+        auto c_block_tensor = MakeCBlockTile();
+        operator()(c_block_tensor, a_block_tensor, b_block_tensor);
+        return c_block_tensor;
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_areg_breg_creg_v1_custom_policy.hpp

@@ -0,0 +1,36 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+
+namespace ck_tile {
+
+template <typename AType_,
+          typename BType_,
+          typename CType_,
+          typename BlockWarps_,
+          typename WarpGemm_>
+struct BlockGemmARegBRegCRegV1CustomPolicy
+{
+    using AType = remove_cvref_t<AType_>;
+    using BType = remove_cvref_t<BType_>;
+    using CType = remove_cvref_t<CType_>;
+
+    using BlockWarps = remove_cvref_t<BlockWarps_>;
+
+    static constexpr index_t kMWarps = BlockWarps::at(number<0>{});
+    static constexpr index_t kNWarps = BlockWarps::at(number<1>{});
+    static constexpr index_t kKWarps = BlockWarps::at(number<2>{});
+
+    using WarpGemm = remove_cvref_t<WarpGemm_>;
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
+    {
+        return make_tuple(WarpGemm{}, kMWarps, kNWarps);
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_areg_breg_creg_v1_default_policy.hpp

@@ -0,0 +1,33 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/warp/warp_gemm.hpp"
+
+namespace ck_tile {
+
+// Default policy for BlockGemmARegBRegCRegV1
+// Default policy class should not be templated, put template on member functions instead
+struct BlockGemmARegBRegCRegV1DefaultPolicy
+{
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
+    {
+        if constexpr(std::is_same_v<typename Problem::ADataType, half_t> &&
+                     std::is_same_v<typename Problem::BDataType, half_t> &&
+                     std::is_same_v<typename Problem::CDataType, float>)
+        {
+            return make_tuple(WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution{}, 4, 1);
+        }
+        else if constexpr(std::is_same_v<typename Problem::ADataType, bf16_t> &&
+                          std::is_same_v<typename Problem::BDataType, bf16_t> &&
+                          std::is_same_v<typename Problem::CDataType, float>)
+        {
+            return make_tuple(WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution{}, 4, 1);
+        }
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_areg_breg_creg_v2.hpp

@@ -0,0 +1,372 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/block/block_gemm_areg_breg_creg_v2_custom_policy.hpp"
+
+namespace ck_tile {
+
+// This BlockGemm enhanced the control over inst issue order
+// A is block distributed tensor
+// B is block distributed tensor
+// C is block distributed tensor
+template <typename Problem_, typename Policy_>
+struct BlockGemmARegBRegCRegV2
+{
+    private:
+    template <typename PipelineProblem_, typename GemmPolicy_>
+    struct GemmTraits_
+    {
+        using Problem        = remove_cvref_t<PipelineProblem_>;
+        using Policy         = remove_cvref_t<GemmPolicy_>;
+        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>;
+
+        static constexpr index_t kBlockSize = 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 auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+        using WarpGemm               = remove_cvref_t<decltype(config.template at<0>())>;
+
+        static constexpr index_t MWarp        = config.template at<1>();
+        static constexpr index_t NWarp        = config.template at<2>();
+        static constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WarpGemm::kM);
+        static constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WarpGemm::kN);
+        static constexpr index_t KIterPerWarp = KPerBlock / WarpGemm::kK;
+
+        static constexpr auto BlockGemmLoopOrder = Policy::BlockGemmLoopOrder;
+
+        static constexpr index_t KPack = WarpGemm::kKPerThread;
+    };
+
+    public:
+    using Problem = remove_cvref_t<Problem_>;
+    using Policy  = remove_cvref_t<Policy_>;
+
+    using Traits = GemmTraits_<Problem, Policy>;
+
+    using WarpGemm                           = typename Traits::WarpGemm;
+    using BlockGemmShape                     = typename Traits::BlockGemmShape;
+    static constexpr auto BlockGemmLoopOrder = Traits::BlockGemmLoopOrder;
+
+    using ADataType = remove_cvref_t<typename Traits::ADataType>;
+    using BDataType = remove_cvref_t<typename Traits::BDataType>;
+    using CDataType = remove_cvref_t<typename Traits::CDataType>;
+
+    static constexpr index_t KIterPerWarp = Traits::KIterPerWarp;
+    static constexpr index_t MIterPerWarp = Traits::MIterPerWarp;
+    static constexpr index_t NIterPerWarp = Traits::NIterPerWarp;
+
+    static constexpr index_t MWarp            = Traits::MWarp;
+    static constexpr index_t NWarp            = Traits::NWarp;
+    static constexpr bool UseDefaultScheduler = (Problem::NumWaveGroups != 1);
+
+    CK_TILE_DEVICE static constexpr auto MakeABlockDistributionEncode()
+    {
+        if constexpr(UseDefaultScheduler)
+        {
+            constexpr auto a_block_outer_dstr_encoding =
+                tile_distribution_encoding<sequence<NWarp>,
+                                           tuple<sequence<MIterPerWarp>, sequence<KIterPerWarp>>,
+                                           tuple<>,
+                                           tuple<>,
+                                           sequence<1, 2>,
+                                           sequence<0, 0>>{};
+
+            constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+                a_block_outer_dstr_encoding, typename WarpGemm::AWarpDstrEncoding{});
+
+            return a_block_dstr_encode;
+        }
+        else
+        {
+            if constexpr(BlockGemmLoopOrder == GemmLoopOrder::KMN)
+            {
+                constexpr auto a_block_outer_dstr_encoding = tile_distribution_encoding<
+                    sequence<NWarp>,
+                    tuple<sequence<MIterPerWarp, MWarp>, sequence<KIterPerWarp>>,
+                    tuple<sequence<1, 0>>,
+                    tuple<sequence<1, 0>>,
+                    sequence<2, 1>,
+                    sequence<0, 0>>{};
+
+                constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+                    a_block_outer_dstr_encoding, typename WarpGemm::AWarpDstrEncoding{});
+
+                return a_block_dstr_encode;
+            }
+            else if constexpr(BlockGemmLoopOrder == GemmLoopOrder::MNK)
+            {
+                constexpr auto a_block_outer_dstr_encoding = tile_distribution_encoding<
+                    sequence<NWarp>,
+                    tuple<sequence<MIterPerWarp, MWarp>, sequence<KIterPerWarp>>,
+                    tuple<sequence<1, 0>>,
+                    tuple<sequence<1, 0>>,
+                    sequence<1, 2>,
+                    sequence<0, 0>>{};
+
+                constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+                    a_block_outer_dstr_encoding, typename WarpGemm::AWarpDstrEncoding{});
+
+                return a_block_dstr_encode;
+            }
+        }
+    }
+
+    CK_TILE_DEVICE static constexpr auto MakeBBlockDistributionEncode()
+    {
+        if constexpr(UseDefaultScheduler)
+        {
+            constexpr auto b_block_outer_dstr_encoding =
+                tile_distribution_encoding<sequence<MWarp>,
+                                           tuple<sequence<NIterPerWarp>, sequence<KIterPerWarp>>,
+                                           tuple<>,
+                                           tuple<>,
+                                           sequence<1, 2>,
+                                           sequence<0, 0>>{};
+            constexpr auto b_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+                b_block_outer_dstr_encoding, typename WarpGemm::BWarpDstrEncoding{});
+
+            return b_block_dstr_encode;
+        }
+        else
+        {
+            if constexpr(BlockGemmLoopOrder == GemmLoopOrder::KMN)
+            {
+                constexpr auto b_block_outer_dstr_encoding = tile_distribution_encoding<
+                    sequence<MWarp>,
+                    tuple<sequence<NIterPerWarp, NWarp>, sequence<KIterPerWarp>>,
+                    tuple<sequence<0, 1>>,
+                    tuple<sequence<0, 1>>,
+                    sequence<2, 1>,
+                    sequence<0, 0>>{};
+                constexpr auto b_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+                    b_block_outer_dstr_encoding, typename WarpGemm::BWarpDstrEncoding{});
+
+                return b_block_dstr_encode;
+            }
+            else if constexpr(BlockGemmLoopOrder == GemmLoopOrder::MNK)
+            {
+                constexpr auto b_block_outer_dstr_encoding = tile_distribution_encoding<
+                    sequence<MWarp>,
+                    tuple<sequence<NIterPerWarp, NWarp>, sequence<KIterPerWarp>>,
+                    tuple<sequence<0, 1>>,
+                    tuple<sequence<0, 1>>,
+                    sequence<1, 2>,
+                    sequence<0, 0>>{};
+                constexpr auto b_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+                    b_block_outer_dstr_encoding, typename WarpGemm::BWarpDstrEncoding{});
+                return b_block_dstr_encode;
+            }
+        }
+    }
+
+    CK_TILE_DEVICE static constexpr auto MakeCBlockDistributionEncode()
+    {
+        if constexpr(UseDefaultScheduler)
+        {
+            constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+                sequence<MWarp>,
+                tuple<sequence<MIterPerWarp>, sequence<NIterPerWarp, NWarp>>,
+                tuple<>,
+                tuple<>,
+                sequence<1, 2>,
+                sequence<0, 0>>{};
+            constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+                c_block_outer_dstr_encoding, typename WarpGemm::CWarpDstrEncoding{});
+
+            return c_block_dstr_encode;
+        }
+        else
+        {
+            constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
+                tuple<sequence<1, 2>>,
+                tuple<sequence<1, 1>>,
+                sequence<1, 2>,
+                sequence<0, 0>>{};
+            constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+                c_block_outer_dstr_encoding, typename WarpGemm::CWarpDstrEncoding{});
+
+            return c_block_dstr_encode;
+        }
+    }
+
+    // C += A * B
+    template <typename CBlockTensor, typename ABlockTensor, typename BBlockTensor>
+    CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
+                                   const ABlockTensor& a_block_tensor,
+                                   const BBlockTensor& b_block_tensor) const
+    {
+        static_assert(std::is_same_v<ADataType, remove_cv_t<typename ABlockTensor::DataType>> &&
+                          std::is_same_v<BDataType, remove_cv_t<typename BBlockTensor::DataType>> &&
+                          std::is_same_v<CDataType, remove_cv_t<typename CBlockTensor::DataType>>,
+                      "wrong!");
+
+        // check ABC-block-distribution
+        static_assert(
+            std::is_same_v<remove_cvref_t<decltype(MakeABlockDistributionEncode())>,
+                           remove_cvref_t<decltype(ABlockTensor::get_tile_distribution()
+                                                       .get_static_tile_distribution_encoding())>>,
+            "A distribution is wrong!");
+        static_assert(
+            std::is_same_v<remove_cvref_t<decltype(MakeBBlockDistributionEncode())>,
+                           remove_cvref_t<decltype(BBlockTensor::get_tile_distribution()
+                                                       .get_static_tile_distribution_encoding())>>,
+            "B distribution is wrong!");
+        static_assert(
+            std::is_same_v<remove_cvref_t<decltype(MakeCBlockDistributionEncode())>,
+                           remove_cvref_t<decltype(CBlockTensor::get_tile_distribution()
+                                                       .get_static_tile_distribution_encoding())>>,
+            "C distribution is wrong!");
+
+        using AWarpDstr = typename WarpGemm::AWarpDstr;
+        using BWarpDstr = typename WarpGemm::BWarpDstr;
+        using CWarpDstr = typename WarpGemm::CWarpDstr;
+
+        using AWarpTensor = typename WarpGemm::AWarpTensor;
+        using BWarpTensor = typename WarpGemm::BWarpTensor;
+        using CWarpTensor = typename WarpGemm::CWarpTensor;
+
+        constexpr auto a_warp_y_lengths =
+            to_sequence(AWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+        constexpr auto b_warp_y_lengths =
+            to_sequence(BWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+        constexpr auto c_warp_y_lengths =
+            to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+
+        constexpr auto a_warp_y_index_zeros = uniform_sequence_gen_t<AWarpDstr::NDimY, 0>{};
+        constexpr auto b_warp_y_index_zeros = uniform_sequence_gen_t<BWarpDstr::NDimY, 0>{};
+        constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
+
+        // hot loop:
+        if constexpr(BlockGemmLoopOrder == GemmLoopOrder::KMN)
+        {
+            static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+                static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                    // read A warp tensor from A Block window
+                    AWarpTensor a_warp_tensor;
+                    a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(sequence<kIter, mIter>{}, a_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
+
+                    static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                        // read B warp tensor from B block tensor
+                        BWarpTensor b_warp_tensor;
+                        b_warp_tensor.get_thread_buffer() = b_block_tensor.get_y_sliced_thread_data(
+                            merge_sequences(sequence<kIter, nIter>{}, b_warp_y_index_zeros),
+                            merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
+
+                        CWarpTensor c_warp_tensor;
+                        c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
+                            merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                            merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
+
+                        // warp GEMM
+                        WarpGemm{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
+
+                        // write C warp tensor into C block tensor
+                        c_block_tensor.set_y_sliced_thread_data(
+                            merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                            merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
+                            c_warp_tensor.get_thread_buffer());
+                    });
+                });
+            });
+        }
+        else if constexpr(BlockGemmLoopOrder == GemmLoopOrder::MNK)
+        {
+            static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                    static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+                        // read A warp tensor from A Block window
+                        AWarpTensor a_warp_tensor;
+
+                        a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
+                            merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
+                            merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
+
+                        // read B warp tensor from B block tensor
+                        BWarpTensor b_warp_tensor;
+
+                        b_warp_tensor.get_thread_buffer() = b_block_tensor.get_y_sliced_thread_data(
+                            merge_sequences(sequence<nIter, kIter>{}, b_warp_y_index_zeros),
+                            merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
+
+                        // read C warp tensor from C block tensor
+                        CWarpTensor c_warp_tensor;
+
+                        c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
+                            merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                            merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
+
+                        // warp GEMM
+                        WarpGemm{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
+
+                        // write C warp tensor into C block tensor
+                        c_block_tensor.set_y_sliced_thread_data(
+                            merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                            merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
+                            c_warp_tensor.get_thread_buffer());
+                    });
+                });
+            });
+        }
+    }
+
+    CK_TILE_DEVICE static constexpr auto MakeCBlockTile()
+    {
+        if constexpr(UseDefaultScheduler)
+        {
+            constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+                sequence<MWarp>,
+                tuple<sequence<MIterPerWarp>, sequence<NIterPerWarp, NWarp>>,
+                tuple<>,
+                tuple<>,
+                sequence<1, 2>,
+                sequence<0, 0>>{};
+
+            constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+                c_block_outer_dstr_encoding, typename WarpGemm::CWarpDstrEncoding{});
+            constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
+            auto c_block_tensor         = make_static_distributed_tensor<CDataType>(c_block_dstr);
+            return c_block_tensor;
+        }
+        else
+        {
+            constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+                sequence<>,
+                tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
+                tuple<sequence<1, 2>>,
+                tuple<sequence<1, 1>>,
+                sequence<1, 2>,
+                sequence<0, 0>>{};
+
+            constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+                c_block_outer_dstr_encoding, typename WarpGemm::CWarpDstrEncoding{});
+            constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
+            auto c_block_tensor         = make_static_distributed_tensor<CDataType>(c_block_dstr);
+            return c_block_tensor;
+        }
+    }
+
+    // C = A * B
+    template <typename ABlockTensor, typename BBlockTensor>
+    CK_TILE_DEVICE auto operator()(const ABlockTensor& a_block_tensor,
+                                   const BBlockTensor& b_block_tensor) const
+    {
+        auto c_block_tensor = MakeCBlockTile();
+        operator()(c_block_tensor, a_block_tensor, b_block_tensor);
+        return c_block_tensor;
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_areg_breg_creg_v2_custom_policy.hpp

@@ -0,0 +1,45 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+
+namespace ck_tile {
+
+enum struct GemmLoopOrder
+{
+    KMN,
+    MNK,
+};
+
+template <typename AType_,
+          typename BType_,
+          typename CType_,
+          typename BlockWarps_,
+          typename WarpGemm_,
+          GemmLoopOrder BlockGemmLoopOrder_ = GemmLoopOrder::KMN>
+struct BlockGemmARegBRegCRegV2CustomPolicy
+{
+    using AType = remove_cvref_t<AType_>;
+    using BType = remove_cvref_t<BType_>;
+    using CType = remove_cvref_t<CType_>;
+
+    using BlockWarps = remove_cvref_t<BlockWarps_>;
+
+    static constexpr index_t kMWarps = BlockWarps::at(number<0>{});
+    static constexpr index_t kNWarps = BlockWarps::at(number<1>{});
+    static constexpr index_t kKWarps = BlockWarps::at(number<2>{});
+
+    using WarpGemm = remove_cvref_t<WarpGemm_>;
+
+    static constexpr auto BlockGemmLoopOrder = BlockGemmLoopOrder_;
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
+    {
+        return make_tuple(WarpGemm{}, kMWarps, kNWarps);
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_one_warp_v1.hpp

@@ -0,0 +1,251 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v1_default_policy.hpp"
+
+namespace ck_tile {
+
+// A is block distributed tensor
+// B is block window on shared memory
+// C is block distributed tensor
+template <typename Problem_, typename Policy_ = BlockGemmARegBSmemCRegV1DefaultPolicy>
+struct BlockGemmARegBSmemCRegOneWarpV1
+{
+    using Problem        = remove_cvref_t<Problem_>;
+    using Policy         = remove_cvref_t<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>;
+
+    static constexpr index_t kBlockSize = Problem::kBlockSize;
+    static_assert(kBlockSize == get_warp_size(), "Check failed!");
+
+    // C += A * B
+    template <typename CBlockTensor, typename ABlockTensorTmp, typename BBlockWindowTmp>
+    CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
+                                   const ABlockTensorTmp& a_block_tensor_tmp,
+                                   const BBlockWindowTmp& b_block_window_tmp) const
+    {
+        static_assert(
+            std::is_same_v<ADataType, remove_cv_t<typename ABlockTensorTmp::DataType>> &&
+                std::is_same_v<BDataType, remove_cv_t<typename BBlockWindowTmp::DataType>> &&
+                std::is_same_v<CDataType, remove_cv_t<typename CBlockTensor::DataType>>,
+            "wrong!");
+
+        // constexpr index_t MPerBlock = ABlockTensorTmp{}.get_lengths()[number<0>{}];
+        // constexpr index_t NPerBlock = BBlockWindowTmp{}.get_window_lengths()[number<0>{}];
+        // constexpr index_t KPerBlock = ABlockTensorTmp{}.get_lengths()[number<1>{}];
+        constexpr index_t MPerBlock = BlockGemmShape::kM;
+        constexpr index_t NPerBlock = BlockGemmShape::kN;
+        constexpr index_t KPerBlock = BlockGemmShape::kK;
+
+        // static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
+        //                   KPerBlock == BlockGemmShape::kK,
+        //               "wrong!");
+
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template at<0>())>;
+
+        constexpr index_t MWarp = config.template at<1>();
+        constexpr index_t NWarp = config.template at<2>();
+
+        static_assert(MWarp == 1 && NWarp == 1, "Check failed!");
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
+        constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        constexpr index_t NPerBlockPerIter = NPerBlock / NIterPerWarp;
+        constexpr index_t KPerBlockPerIter = KPerBlock / KIterPerWarp;
+
+        const index_t iNWarp = 0;
+
+        constexpr auto c_block_outer_dstr_encoding =
+            tile_distribution_encoding<sequence<>,
+                                       tuple<sequence<MIterPerWarp>, sequence<NIterPerWarp>>,
+                                       tuple<>,
+                                       tuple<>,
+                                       sequence<1, 2>,
+                                       sequence<0, 0>>{};
+
+        constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
+
+        // constrcut from A-block-tensor from A-Block-tensor-tmp
+        // FIXME: need method to check a_block_tensor and a_block_tensor_tmp have equivalent
+        // distribution
+        auto a_block_tensor = make_static_distributed_tensor<typename ABlockTensorTmp::DataType>(
+            MakeABlockTileDistribution());
+
+        a_block_tensor.get_thread_buffer() = a_block_tensor_tmp.get_thread_buffer();
+
+        // construct B-warp-window
+        auto b_warp_window_tmp = make_tile_window(
+            b_block_window_tmp.get_bottom_tensor_view(),
+            make_tuple(number<WG::kN>{}, number<WG::kK>{}),
+            b_block_window_tmp.get_window_origin() + multi_index<2>{iNWarp * WG::kN, 0},
+            make_static_tile_distribution(typename WG::BWarpDstrEncoding{}));
+
+#if 0 // FIXME: using array will cause register spill
+        array<array<decltype(b_warp_window_tmp), KIterPerWarp>, NIterPerWarp> b_warp_windows{
+            {b_warp_window_tmp}};
+
+        for(index_t nIter = 0; nIter < NIterPerWarp; nIter++)
+        {
+            for(index_t kIter = 0; kIter < KIterPerWarp; kIter++)
+            {
+                move_tile_window(b_warp_windows(nIter)(kIter),
+                                 {nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
+            }
+        }
+#else
+        statically_indexed_array<
+            statically_indexed_array<decltype(b_warp_window_tmp), KIterPerWarp>,
+            NIterPerWarp>
+            b_warp_windows;
+
+        static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+            static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+                b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
+
+                move_tile_window(b_warp_windows(nIter)(kIter),
+                                 {nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
+            });
+        });
+#endif
+
+        // check C-block-distribution
+        static_assert(
+            std::is_same_v<remove_cvref_t<decltype(c_block_dstr_encode)>,
+                           remove_cvref_t<decltype(CBlockTensor::get_tile_distribution()
+                                                       .get_static_tile_distribution_encoding())>>,
+            "wrong!");
+
+        using AWarpDstr = typename WG::AWarpDstr;
+        using CWarpDstr = typename WG::CWarpDstr;
+
+        using AWarpTensor = typename WG::AWarpTensor;
+        using CWarpTensor = typename WG::CWarpTensor;
+
+        constexpr auto a_warp_y_lengths =
+            to_sequence(AWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+        constexpr auto c_warp_y_lengths =
+            to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+
+        constexpr auto a_warp_y_index_zeros = uniform_sequence_gen_t<AWarpDstr::NDimY, 0>{};
+        constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
+
+        // hot loop:
+        static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+            static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                // read A warp tensor from A block tensor
+                AWarpTensor a_warp_tensor;
+
+                a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
+                    merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
+                    merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
+
+                static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                    // read B warp tensor from B Block window
+                    const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
+
+                    // read C warp tensor from C block tensor
+                    CWarpTensor c_warp_tensor;
+
+                    c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
+
+                    // warp GEMM
+                    WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
+
+                    // write C warp tensor into C block tensor
+                    c_block_tensor.set_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
+                        c_warp_tensor.get_thread_buffer());
+                });
+            });
+        });
+    }
+
+    template <index_t MPerBlock = BlockGemmShape::kM, index_t KPerBlock = BlockGemmShape::kK>
+    CK_TILE_DEVICE static constexpr auto MakeABlockTileDistribution()
+    {
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template at<0>())>;
+
+        constexpr index_t MWarp = config.template at<1>();
+        constexpr index_t NWarp = config.template at<2>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        constexpr auto a_block_outer_dstr_encoding =
+            tile_distribution_encoding<sequence<NWarp>,
+                                       tuple<sequence<MIterPerWarp, MWarp>, sequence<KIterPerWarp>>,
+                                       tuple<sequence<1, 0>>,
+                                       tuple<sequence<1, 0>>,
+                                       sequence<1, 2>,
+                                       sequence<0, 0>>{};
+
+        constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            a_block_outer_dstr_encoding, typename WG::AWarpDstrEncoding{});
+
+        return make_static_tile_distribution(a_block_dstr_encode);
+    }
+
+    CK_TILE_DEVICE static constexpr auto MakeCBlockTile()
+    {
+        constexpr index_t MPerBlock = BlockGemmShape::kM;
+        constexpr index_t NPerBlock = BlockGemmShape::kN;
+
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template at<0>())>;
+
+        constexpr index_t MWarp = config.template at<1>();
+        constexpr index_t NWarp = config.template at<2>();
+
+        static_assert(MWarp == 1 && NWarp == 1, "Check failed!");
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
+        // constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        constexpr auto c_block_outer_dstr_encoding =
+            tile_distribution_encoding<sequence<>,
+                                       tuple<sequence<MIterPerWarp>, sequence<NIterPerWarp>>,
+                                       tuple<>,
+                                       tuple<>,
+                                       sequence<1, 2>,
+                                       sequence<0, 0>>{};
+
+        constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
+
+        static_assert(decltype(c_block_dstr_encode)::NDimP == 1, "Check failed!");
+
+        constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
+        auto c_block_tensor         = make_static_distributed_tensor<CDataType>(c_block_dstr);
+        return c_block_tensor;
+    }
+
+    // C = A * B
+    template <typename ABlockTensorTmp, typename BBlockWindowTmp>
+    CK_TILE_DEVICE auto operator()(const ABlockTensorTmp& a_block_tensor_tmp,
+                                   const BBlockWindowTmp& b_block_window_tmp) const
+    {
+        auto c_block_tensor = MakeCBlockTile();
+        operator()(c_block_tensor, a_block_tensor_tmp, b_block_window_tmp);
+        return c_block_tensor;
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v1.hpp

@@ -0,0 +1,226 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v1_default_policy.hpp"
+
+namespace ck_tile {
+
+// A is block distributed tensor
+// B is block window on shared memory
+// C is block distributed tensor
+template <typename Problem_, typename Policy_ = BlockGemmARegBSmemCRegV1DefaultPolicy>
+struct BlockGemmARegBSmemCRegV1
+{
+    using Problem        = remove_cvref_t<Problem_>;
+    using Policy         = remove_cvref_t<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>;
+
+    static constexpr index_t kBlockSize = Problem::kBlockSize;
+
+    // C += A * B
+    template <typename CBlockTensor, typename ABlockTensorTmp, typename BBlockWindowTmp>
+    CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
+                                   const ABlockTensorTmp& a_block_tensor_tmp,
+                                   const BBlockWindowTmp& b_block_window_tmp) const
+    {
+        static_assert(
+            std::is_same_v<ADataType, remove_cv_t<typename ABlockTensorTmp::DataType>> &&
+                std::is_same_v<BDataType, remove_cv_t<typename BBlockWindowTmp::DataType>> &&
+                std::is_same_v<CDataType, remove_cv_t<typename CBlockTensor::DataType>>,
+            "wrong!");
+
+        constexpr index_t MPerBlock = ABlockTensorTmp{}.get_lengths()[number<0>{}];
+        constexpr index_t NPerBlock = BBlockWindowTmp{}.get_window_lengths()[number<0>{}];
+        constexpr index_t KPerBlock = ABlockTensorTmp{}.get_lengths()[number<1>{}];
+
+        static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
+                          KPerBlock == BlockGemmShape::kK,
+                      "wrong!");
+
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template at<0>())>;
+
+        constexpr index_t MWarp = config.template at<1>();
+        constexpr index_t NWarp = config.template at<2>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
+        constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        constexpr index_t NPerBlockPerIter = NPerBlock / NIterPerWarp;
+        constexpr index_t KPerBlockPerIter = KPerBlock / KIterPerWarp;
+
+        const index_t iNWarp = get_warp_id() % NWarp;
+
+        constexpr auto a_block_outer_dstr_encoding =
+            tile_distribution_encoding<sequence<NWarp>,
+                                       tuple<sequence<MIterPerWarp, MWarp>, sequence<KIterPerWarp>>,
+                                       tuple<sequence<1, 0>>,
+                                       tuple<sequence<1, 0>>,
+                                       sequence<1, 2>,
+                                       sequence<0, 0>>{};
+
+        constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+            sequence<>,
+            tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
+            tuple<sequence<1, 2>>,
+            tuple<sequence<1, 1>>,
+            sequence<1, 2>,
+            sequence<0, 0>>{};
+
+        constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            a_block_outer_dstr_encoding, typename WG::AWarpDstrEncoding{});
+
+        constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
+
+        constexpr auto a_block_dstr = make_static_tile_distribution(a_block_dstr_encode);
+
+        // constrcut from A-block-tensor from A-Block-tensor-tmp
+        // FIXME: need method to check a_block_tensor and a_block_tensor_tmp have equivalent
+        // distribution
+        auto a_block_tensor =
+            make_static_distributed_tensor<typename ABlockTensorTmp::DataType>(a_block_dstr);
+
+        a_block_tensor.get_thread_buffer() = a_block_tensor_tmp.get_thread_buffer();
+
+        // construct B-warp-window
+        auto b_warp_window_tmp = make_tile_window(
+            b_block_window_tmp.get_bottom_tensor_view(),
+            make_tuple(number<WG::kN>{}, number<WG::kK>{}),
+            b_block_window_tmp.get_window_origin() + multi_index<2>{iNWarp * WG::kN, 0},
+            make_static_tile_distribution(typename WG::BWarpDstrEncoding{}));
+
+#if 0 // FIXME: using array will cause register spill
+        array<array<decltype(b_warp_window_tmp), KIterPerWarp>, NIterPerWarp> b_warp_windows{
+            {b_warp_window_tmp}};
+
+        for(index_t nIter = 0; nIter < NIterPerWarp; nIter++)
+        {
+            for(index_t kIter = 0; kIter < KIterPerWarp; kIter++)
+            {
+                move_tile_window(b_warp_windows(nIter)(kIter),
+                                 {nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
+            }
+        }
+#else
+        statically_indexed_array<
+            statically_indexed_array<decltype(b_warp_window_tmp), KIterPerWarp>,
+            NIterPerWarp>
+            b_warp_windows;
+
+        static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+            static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+                b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
+
+                move_tile_window(b_warp_windows(nIter)(kIter),
+                                 {nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
+            });
+        });
+#endif
+
+        // check C-block-distribution
+        static_assert(
+            std::is_same_v<remove_cvref_t<decltype(c_block_dstr_encode)>,
+                           remove_cvref_t<decltype(CBlockTensor::get_tile_distribution()
+                                                       .get_static_tile_distribution_encoding())>>,
+            "wrong!");
+
+        using AWarpDstr = typename WG::AWarpDstr;
+        using CWarpDstr = typename WG::CWarpDstr;
+
+        using AWarpTensor = typename WG::AWarpTensor;
+        using CWarpTensor = typename WG::CWarpTensor;
+
+        constexpr auto a_warp_y_lengths =
+            to_sequence(AWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+        constexpr auto c_warp_y_lengths =
+            to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+
+        constexpr auto a_warp_y_index_zeros = uniform_sequence_gen_t<AWarpDstr::NDimY, 0>{};
+        constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
+
+        // hot loop:
+        static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+            static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                // read A warp tensor from A block tensor
+                AWarpTensor a_warp_tensor;
+
+                a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
+                    merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
+                    merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
+
+                static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                    // read B warp tensor from B Block window
+                    const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
+
+                    // read C warp tensor from C block tensor
+                    CWarpTensor c_warp_tensor;
+
+                    c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
+
+                    // warp GEMM
+                    WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
+
+                    // write C warp tensor into C block tensor
+                    c_block_tensor.set_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
+                        c_warp_tensor.get_thread_buffer());
+                });
+            });
+        });
+    }
+
+    CK_TILE_DEVICE static constexpr auto MakeCBlockTile()
+    {
+        constexpr index_t MPerBlock = BlockGemmShape::kM;
+        constexpr index_t NPerBlock = BlockGemmShape::kN;
+
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template at<0>())>;
+
+        constexpr index_t MWarp = config.template at<1>();
+        constexpr index_t NWarp = config.template at<2>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
+        // constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+            sequence<>,
+            tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
+            tuple<sequence<1, 2>>,
+            tuple<sequence<1, 1>>,
+            sequence<1, 2>,
+            sequence<0, 0>>{};
+
+        constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
+        constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
+        auto c_block_tensor         = make_static_distributed_tensor<CDataType>(c_block_dstr);
+        return c_block_tensor;
+    }
+
+    // C = A * B
+    template <typename ABlockTensorTmp, typename BBlockWindowTmp>
+    CK_TILE_DEVICE auto operator()(const ABlockTensorTmp& a_block_tensor_tmp,
+                                   const BBlockWindowTmp& b_block_window_tmp) const
+    {
+        auto c_block_tensor = MakeCBlockTile();
+        operator()(c_block_tensor, a_block_tensor_tmp, b_block_window_tmp);
+        return c_block_tensor;
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v1_custom_policy.hpp

@@ -0,0 +1,36 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+
+namespace ck_tile {
+
+template <typename AType_,
+          typename BType_,
+          typename CType_,
+          typename BlockWarps_,
+          typename WarpGemm_>
+struct BlockGemmARegBSmemCRegV1CustomPolicy
+{
+    using AType = remove_cvref_t<AType_>;
+    using BType = remove_cvref_t<BType_>;
+    using CType = remove_cvref_t<CType_>;
+
+    using BlockWarps = remove_cvref_t<BlockWarps_>;
+
+    static constexpr index_t kMWarps = BlockWarps::at(number<0>{});
+    static constexpr index_t kNWarps = BlockWarps::at(number<1>{});
+    static constexpr index_t kKWarps = BlockWarps::at(number<2>{});
+
+    using WarpGemm = remove_cvref_t<WarpGemm_>;
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
+    {
+        return make_tuple(WarpGemm{}, kMWarps, kNWarps);
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v1_default_policy.hpp

@@ -0,0 +1,56 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/warp/warp_gemm.hpp"
+
+namespace ck_tile {
+
+// Default policy for BlockGemmARegBSmemCRegV1
+// Default policy class should not be templated, put template on member functions instead
+struct BlockGemmARegBSmemCRegV1DefaultPolicy
+{
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
+    {
+        if constexpr(std::is_same_v<typename Problem::ADataType, half_t> &&
+                     std::is_same_v<typename Problem::BDataType, half_t> &&
+                     std::is_same_v<typename Problem::CDataType, float>)
+        {
+#if 0
+            constexpr index_t kBlockSize = Problem::kBlockSize;
+
+            constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
+            constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+            constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+            static_assert(kBlockSize % get_warp_size() == 0, "wrong!");
+
+            constexpr index_t NumWarp = kBlockSize / get_warp_size();
+
+            // FIXME
+            if constexpr(NumWarp == 4 && kMPerBlock % 128 == 0 &&
+                         kNPerBlock % 128 == 0 % kKPerBlock % 16 == 0)
+            {
+                return make_tuple(WarpGemmMfmaF16F16F32M32N32K8{}, 4, 1);
+            }
+            else
+            {
+                return make_tuple(WarpGemmMfmaF16F16F32M32N32K8{}, 4, 1);
+            }
+#else
+            return make_tuple(WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution{}, 4, 1);
+#endif
+        }
+        else if constexpr(std::is_same_v<typename Problem::ADataType, bf16_t> &&
+                          std::is_same_v<typename Problem::BDataType, bf16_t> &&
+                          std::is_same_v<typename Problem::CDataType, float>)
+        {
+            return make_tuple(WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution{}, 4, 1);
+        }
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v2.hpp

@@ -0,0 +1,241 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v2_default_policy.hpp"
+
+namespace ck_tile {
+
+// A is block distributed tensor
+// B is block window on shared memory
+// C is block distributed tensor
+template <typename Problem_, typename Policy_ = BlockGemmARegBSmemCRegV2DefaultPolicy>
+struct BlockGemmARegBSmemCRegV2
+{
+    using Problem        = remove_cvref_t<Problem_>;
+    using Policy         = remove_cvref_t<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>;
+
+    static constexpr index_t kBlockSize = Problem::kBlockSize;
+
+    // C += A * B
+    template <typename CBlockTensor, typename ABlockTensorTmp, typename BBlockWindowTmp>
+    CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
+                                   const ABlockTensorTmp& a_block_tensor_tmp,
+                                   const BBlockWindowTmp& b_block_window_tmp) const
+    {
+        static_assert(
+            std::is_same_v<ADataType, remove_cv_t<typename ABlockTensorTmp::DataType>> &&
+                std::is_same_v<BDataType, remove_cv_t<typename BBlockWindowTmp::DataType>> &&
+                std::is_same_v<CDataType, remove_cv_t<typename CBlockTensor::DataType>>,
+            "wrong!");
+
+        constexpr index_t MPerBlock = ABlockTensorTmp{}.get_lengths()[number<0>{}];
+        constexpr index_t NPerBlock = BBlockWindowTmp{}.get_window_lengths()[number<0>{}];
+        constexpr index_t KPerBlock = ABlockTensorTmp{}.get_lengths()[number<1>{}];
+
+        static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
+                          KPerBlock == BlockGemmShape::kK,
+                      "wrong!");
+
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template at<0>())>;
+
+        constexpr index_t MWarp = config.template at<1>();
+        constexpr index_t NWarp = config.template at<2>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
+        constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        constexpr index_t NPerBlockPerIter = NPerBlock / NIterPerWarp;
+        constexpr index_t KPerBlockPerIter = KPerBlock / KIterPerWarp;
+
+        const index_t iNWarp = get_warp_id() % NWarp;
+
+        constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+            sequence<>,
+            tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
+            tuple<sequence<1, 2>>,
+            tuple<sequence<1, 1>>,
+            sequence<1, 2>,
+            sequence<0, 0>>{};
+
+        constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
+
+        // constrcut from A-block-tensor from A-Block-tensor-tmp
+        // FIXME: need method to check a_block_tensor and a_block_tensor_tmp have equivalent
+        // distribution
+        auto a_block_tensor = make_static_distributed_tensor<typename ABlockTensorTmp::DataType>(
+            MakeABlockTileDistribution());
+
+        a_block_tensor.get_thread_buffer() = a_block_tensor_tmp.get_thread_buffer();
+
+        // construct B-warp-window
+        auto b_warp_window_tmp = make_tile_window(
+            b_block_window_tmp.get_bottom_tensor_view(),
+            make_tuple(number<WG::kN>{}, number<WG::kK>{}),
+            b_block_window_tmp.get_window_origin() + multi_index<2>{iNWarp * WG::kN, 0},
+            make_static_tile_distribution(typename WG::BWarpDstrEncoding{}));
+
+#if 0 // FIXME: using array will cause register spill
+        array<array<decltype(b_warp_window_tmp), KIterPerWarp>, NIterPerWarp> b_warp_windows{
+            {b_warp_window_tmp}};
+
+        for(index_t nIter = 0; nIter < NIterPerWarp; nIter++)
+        {
+            for(index_t kIter = 0; kIter < KIterPerWarp; kIter++)
+            {
+                move_tile_window(b_warp_windows(nIter)(kIter),
+                                 {nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
+            }
+        }
+#else
+        statically_indexed_array<
+            statically_indexed_array<decltype(b_warp_window_tmp), KIterPerWarp>,
+            NIterPerWarp>
+            b_warp_windows;
+
+        static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+            static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+                b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
+
+                move_tile_window(b_warp_windows(nIter)(kIter),
+                                 {nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
+            });
+        });
+#endif
+
+        // check C-block-distribution
+        static_assert(
+            std::is_same_v<remove_cvref_t<decltype(c_block_dstr_encode)>,
+                           remove_cvref_t<decltype(CBlockTensor::get_tile_distribution()
+                                                       .get_static_tile_distribution_encoding())>>,
+            "wrong!");
+
+        using AWarpDstr = typename WG::AWarpDstr;
+        using CWarpDstr = typename WG::CWarpDstr;
+
+        using AWarpTensor = typename WG::AWarpTensor;
+        using CWarpTensor = typename WG::CWarpTensor;
+
+        constexpr auto a_warp_y_lengths =
+            to_sequence(AWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+        constexpr auto c_warp_y_lengths =
+            to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+
+        constexpr auto a_warp_y_index_zeros = uniform_sequence_gen_t<AWarpDstr::NDimY, 0>{};
+        constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
+
+        // hot loop:
+        static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+            static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                // read B warp tensor from B Block window
+                const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
+
+                static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                    // read A warp tensor from A block tensor
+                    AWarpTensor a_warp_tensor;
+
+                    a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
+
+                    // read C warp tensor from C block tensor
+                    CWarpTensor c_warp_tensor;
+
+                    c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
+
+                    // warp GEMM
+                    WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
+                    // WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor_array[nIter]);
+
+                    // write C warp tensor into C block tensor
+                    c_block_tensor.set_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
+                        c_warp_tensor.get_thread_buffer());
+                });
+            });
+        });
+    }
+
+    template <index_t MPerBlock = BlockGemmShape::kM, index_t KPerBlock = BlockGemmShape::kK>
+    CK_TILE_DEVICE static constexpr auto MakeABlockTileDistribution()
+    {
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template at<0>())>;
+
+        constexpr index_t MWarp = config.template at<1>();
+        constexpr index_t NWarp = config.template at<2>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        constexpr auto a_block_outer_dstr_encoding =
+            tile_distribution_encoding<sequence<NWarp>,
+                                       tuple<sequence<MIterPerWarp, MWarp>, sequence<KIterPerWarp>>,
+                                       tuple<sequence<1, 0>>,
+                                       tuple<sequence<1, 0>>,
+                                       sequence<1, 2>,
+                                       sequence<0, 0>>{};
+
+        constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            a_block_outer_dstr_encoding, typename WG::AWarpDstrEncoding{});
+
+        return make_static_tile_distribution(a_block_dstr_encode);
+    }
+
+    CK_TILE_DEVICE static constexpr auto MakeCBlockTile()
+    {
+        constexpr index_t MPerBlock = BlockGemmShape::kM;
+        constexpr index_t NPerBlock = BlockGemmShape::kN;
+
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template at<0>())>;
+
+        constexpr index_t MWarp = config.template at<1>();
+        constexpr index_t NWarp = config.template at<2>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
+        // constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+            sequence<>,
+            tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
+            tuple<sequence<1, 2>>,
+            tuple<sequence<1, 1>>,
+            sequence<1, 2>,
+            sequence<0, 0>>{};
+
+        constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
+        constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
+        auto c_block_tensor         = make_static_distributed_tensor<CDataType>(c_block_dstr);
+        return c_block_tensor;
+    }
+
+    // C = A * B
+    template <typename ABlockTensorTmp, typename BBlockWindowTmp>
+    CK_TILE_DEVICE auto operator()(const ABlockTensorTmp& a_block_tensor_tmp,
+                                   const BBlockWindowTmp& b_block_window_tmp) const
+    {
+        auto c_block_tensor = MakeCBlockTile();
+        operator()(c_block_tensor, a_block_tensor_tmp, b_block_window_tmp);
+        return c_block_tensor;
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v2_custom_policy.hpp

@@ -0,0 +1,36 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+
+namespace ck_tile {
+
+template <typename AType_,
+          typename BType_,
+          typename CType_,
+          typename BlockWarps_,
+          typename WarpGemm_>
+struct BlockGemmARegBSmemCRegV2CustomPolicy
+{
+    using AType = remove_cvref_t<AType_>;
+    using BType = remove_cvref_t<BType_>;
+    using CType = remove_cvref_t<CType_>;
+
+    using BlockWarps = remove_cvref_t<BlockWarps_>;
+
+    static constexpr index_t kMWarps = BlockWarps::at(number<0>{});
+    static constexpr index_t kNWarps = BlockWarps::at(number<1>{});
+    static constexpr index_t kKWarps = BlockWarps::at(number<2>{});
+
+    using WarpGemm = remove_cvref_t<WarpGemm_>;
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
+    {
+        return make_tuple(WarpGemm{}, kMWarps, kNWarps);
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v2_default_policy.hpp

@@ -0,0 +1,46 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/warp/warp_gemm.hpp"
+
+namespace ck_tile {
+
+// Default policy for BlockGemmARegBSmemCRegV2
+// Default policy class should not be templated, put template on member functions instead
+struct BlockGemmARegBSmemCRegV2DefaultPolicy
+{
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
+    {
+
+#if 0
+        constexpr index_t kBlockSize = Problem::kBlockSize;
+
+        constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
+        constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+        static_assert(kBlockSize % get_warp_size() == 0, "wrong!");
+
+        constexpr index_t NumWarp = kBlockSize / get_warp_size();
+
+        // FIXME
+        if constexpr(NumWarp == 4 && kMPerBlock % 128 == 0 &&
+                     kNPerBlock % 128 == 0 % kKPerBlock % 16 == 0)
+        {
+            return make_tuple(WarpGemmMfmaF16F16F32M32N32K8{}, 4, 1);
+        }
+        else
+        {
+            return make_tuple(WarpGemmMfmaF16F16F32M32N32K8{}, 4, 1);
+        }
+#else
+        return make_tuple(WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution{}, 4, 1);
+#endif
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v2r1.hpp

@@ -0,0 +1,247 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/block/block_gemm_areg_bsmem_creg_v2_default_policy.hpp"
+
+namespace ck_tile {
+
+// A is block distributed tensor
+// B is block window on shared memory
+// C is block distributed tensor
+template <typename Problem_, typename Policy_ = BlockGemmARegBSmemCRegV2DefaultPolicy>
+struct BlockGemmARegBSmemCRegV2R1
+{
+    using Problem        = remove_cvref_t<Problem_>;
+    using Policy         = remove_cvref_t<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>;
+
+    static constexpr index_t kBlockSize = Problem::kBlockSize;
+
+    // C += A * B
+    template <typename CBlockTensor, typename ABlockTensorTmp, typename BBlockWindowTmp>
+    CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
+                                   const ABlockTensorTmp& a_block_tensor_tmp,
+                                   const BBlockWindowTmp& b_block_window_tmp) const
+    {
+        static_assert(
+            std::is_same_v<ADataType, remove_cv_t<typename ABlockTensorTmp::DataType>> &&
+                std::is_same_v<BDataType, remove_cv_t<typename BBlockWindowTmp::DataType>> &&
+                std::is_same_v<CDataType, remove_cv_t<typename CBlockTensor::DataType>>,
+            "wrong!");
+
+        constexpr index_t MPerBlock = ABlockTensorTmp{}.get_lengths()[number<0>{}];
+        constexpr index_t NPerBlock = BBlockWindowTmp{}.get_window_lengths()[number<0>{}];
+        constexpr index_t KPerBlock = ABlockTensorTmp{}.get_lengths()[number<1>{}];
+
+        static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
+                          KPerBlock == BlockGemmShape::kK,
+                      "wrong!");
+
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template at<0>())>;
+
+        constexpr index_t MWarp = config.template at<1>();
+        constexpr index_t NWarp = config.template at<2>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
+        constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        constexpr index_t NPerBlockPerIter = NPerBlock / NIterPerWarp;
+        constexpr index_t KPerBlockPerIter = KPerBlock / KIterPerWarp;
+
+        const index_t iNWarp = get_warp_id() % NWarp;
+
+        constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+            sequence<>,
+            tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
+            tuple<sequence<1, 2>>,
+            tuple<sequence<1, 1>>,
+            sequence<1, 2>,
+            sequence<0, 0>>{};
+
+        constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
+
+        // constrcut from A-block-tensor from A-Block-tensor-tmp
+        // FIXME: need method to check a_block_tensor and a_block_tensor_tmp have equivalent
+        // distribution
+        auto a_block_tensor = make_static_distributed_tensor<typename ABlockTensorTmp::DataType>(
+            MakeABlockTileDistribution());
+
+        a_block_tensor.get_thread_buffer() = a_block_tensor_tmp.get_thread_buffer();
+
+        // construct B-warp-window
+        auto b_warp_window_tmp = make_tile_window(
+            b_block_window_tmp.get_bottom_tensor_view(),
+            make_tuple(number<WG::kN>{}, number<WG::kK>{}),
+            b_block_window_tmp.get_window_origin() + multi_index<2>{iNWarp * WG::kN, 0},
+            make_static_tile_distribution(typename WG::BWarpDstrEncoding{}));
+
+        statically_indexed_array<
+            statically_indexed_array<decltype(b_warp_window_tmp), KIterPerWarp>,
+            NIterPerWarp>
+            b_warp_windows;
+
+        static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+            static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+                b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
+
+                move_tile_window(b_warp_windows(nIter)(kIter),
+                                 {nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
+            });
+        });
+
+        // check C-block-distribution
+        static_assert(
+            std::is_same_v<remove_cvref_t<decltype(c_block_dstr_encode)>,
+                           remove_cvref_t<decltype(CBlockTensor::get_tile_distribution()
+                                                       .get_static_tile_distribution_encoding())>>,
+            "wrong!");
+
+        using AWarpDstr = typename WG::AWarpDstr;
+        using CWarpDstr = typename WG::CWarpDstr;
+
+        using AWarpTensor = typename WG::AWarpTensor;
+        using CWarpTensor = typename WG::CWarpTensor;
+
+        constexpr auto a_warp_y_lengths =
+            to_sequence(AWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+        constexpr auto c_warp_y_lengths =
+            to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+
+        constexpr auto a_warp_y_index_zeros = uniform_sequence_gen_t<AWarpDstr::NDimY, 0>{};
+        constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
+
+        statically_indexed_array<
+            statically_indexed_array<decltype(load_tile(decltype(b_warp_window_tmp){})),
+                                     KIterPerWarp>,
+            NIterPerWarp>
+            b_warp_tensors;
+
+        static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+            static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                b_warp_tensors(nIter)(kIter) = load_tile(b_warp_windows(nIter)(kIter));
+            });
+        });
+
+        // hot loop:
+        static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+            static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                // read B warp tensor from B Block window
+                const auto b_warp_tensor = b_warp_tensors(nIter)(kIter);
+
+                static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                    // read A warp tensor from A block tensor
+                    AWarpTensor a_warp_tensor;
+
+                    a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
+
+                    // read C warp tensor from C block tensor
+                    CWarpTensor c_warp_tensor;
+
+                    c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
+
+                    // warp GEMM
+                    WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
+                    // WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor_array[nIter]);
+
+                    // write C warp tensor into C block tensor
+                    c_block_tensor.set_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
+                        c_warp_tensor.get_thread_buffer());
+                });
+            });
+        });
+
+        __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+        static_for<0, KIterPerWarp, 1>{}([&](auto) {
+            static_for<0, NIterPerWarp, 1>{}([&](auto) {
+                __builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
+                __builtin_amdgcn_sched_group_barrier(0x008, 2, 0); // MFMA
+            });
+        });
+    }
+
+    template <index_t MPerBlock = BlockGemmShape::kM, index_t KPerBlock = BlockGemmShape::kK>
+    CK_TILE_DEVICE static constexpr auto MakeABlockTileDistribution()
+    {
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template at<0>())>;
+
+        constexpr index_t MWarp = config.template at<1>();
+        constexpr index_t NWarp = config.template at<2>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        constexpr auto a_block_outer_dstr_encoding =
+            tile_distribution_encoding<sequence<NWarp>,
+                                       tuple<sequence<MIterPerWarp, MWarp>, sequence<KIterPerWarp>>,
+                                       tuple<sequence<1, 0>>,
+                                       tuple<sequence<1, 0>>,
+                                       sequence<1, 2>,
+                                       sequence<0, 0>>{};
+
+        constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            a_block_outer_dstr_encoding, typename WG::AWarpDstrEncoding{});
+
+        return make_static_tile_distribution(a_block_dstr_encode);
+    }
+
+    CK_TILE_DEVICE static constexpr auto MakeCBlockTile()
+    {
+        constexpr index_t MPerBlock = BlockGemmShape::kM;
+        constexpr index_t NPerBlock = BlockGemmShape::kN;
+
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template at<0>())>;
+
+        constexpr index_t MWarp = config.template at<1>();
+        constexpr index_t NWarp = config.template at<2>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
+        // constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+            sequence<>,
+            tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
+            tuple<sequence<1, 2>>,
+            tuple<sequence<1, 1>>,
+            sequence<1, 2>,
+            sequence<0, 0>>{};
+
+        constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
+        constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
+        auto c_block_tensor         = make_static_distributed_tensor<CDataType>(c_block_dstr);
+        return c_block_tensor;
+    }
+
+    // C = A * B
+    template <typename ABlockTensorTmp, typename BBlockWindowTmp>
+    CK_TILE_DEVICE auto operator()(const ABlockTensorTmp& a_block_tensor_tmp,
+                                   const BBlockWindowTmp& b_block_window_tmp) const
+    {
+        auto c_block_tensor = MakeCBlockTile();
+        operator()(c_block_tensor, a_block_tensor_tmp, b_block_window_tmp);
+        return c_block_tensor;
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_asmem_breg_creg_v1.hpp

@@ -0,0 +1,225 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/block/block_gemm_asmem_breg_creg_v1_default_policy.hpp"
+
+namespace ck_tile {
+
+// A is block window on shared memory
+// B is block distributed tensor
+// C is block distributed tensor
+template <typename Problem_, typename Policy_ = BlockGemmASmemBRegCRegV1DefaultPolicy>
+struct BlockGemmASmemBRegCRegV1
+{
+    using Problem        = remove_cvref_t<Problem_>;
+    using Policy         = remove_cvref_t<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>;
+
+    static constexpr index_t kBlockSize = Problem::kBlockSize;
+
+    // C += A * B
+    template <typename CBlockTensor, typename ABlockWindowTmp, typename BBlockTensorTmp>
+    CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
+                                   const ABlockWindowTmp& a_block_window_tmp,
+                                   const BBlockTensorTmp& b_block_tensor_tmp) const
+    {
+        static_assert(
+            std::is_same_v<ADataType, remove_cv_t<typename ABlockWindowTmp::DataType>> &&
+                std::is_same_v<BDataType, remove_cv_t<typename BBlockTensorTmp::DataType>> &&
+                std::is_same_v<CDataType, remove_cv_t<typename CBlockTensor::DataType>>,
+            "wrong!");
+
+        constexpr index_t MPerBlock = ABlockWindowTmp{}.get_window_lengths()[number<0>{}];
+        constexpr index_t NPerBlock = BBlockTensorTmp{}.get_lengths()[number<0>{}];
+        constexpr index_t KPerBlock = ABlockWindowTmp{}.get_window_lengths()[number<1>{}];
+
+        static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
+                          KPerBlock == BlockGemmShape::kK,
+                      "wrong!");
+
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template at<0>())>;
+
+        constexpr index_t MWarp = config.template at<1>();
+        constexpr index_t NWarp = config.template at<2>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
+        constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        constexpr index_t MPerBlockPerIter = MPerBlock / MIterPerWarp;
+        constexpr index_t KPerBlockPerIter = KPerBlock / KIterPerWarp;
+
+        const index_t iMWarp = get_warp_id() / NWarp;
+
+        constexpr auto b_block_outer_dstr_encoding =
+            tile_distribution_encoding<sequence<MWarp>,
+                                       tuple<sequence<NIterPerWarp, NWarp>, sequence<KIterPerWarp>>,
+                                       tuple<sequence<0, 1>>,
+                                       tuple<sequence<0, 1>>,
+                                       sequence<1, 2>,
+                                       sequence<0, 0>>{};
+
+        constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+            sequence<>,
+            tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
+            tuple<sequence<1, 2>>,
+            tuple<sequence<1, 1>>,
+            sequence<1, 2>,
+            sequence<0, 0>>{};
+
+        constexpr auto b_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            b_block_outer_dstr_encoding, typename WG::BWarpDstrEncoding{});
+
+        constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
+
+        constexpr auto b_block_dstr = make_static_tile_distribution(b_block_dstr_encode);
+
+        // constrcut from B-block-tensor from B-Block-tensor-tmp
+        // FIXME: need method to check b_block_tensor and b_block_tensor_tmp have equivalent
+        // distribution
+        auto b_block_tensor =
+            make_static_distributed_tensor<typename BBlockTensorTmp::DataType>(b_block_dstr);
+
+        b_block_tensor.get_thread_buffer() = b_block_tensor_tmp.get_thread_buffer();
+
+        // construct A-warp-window
+        auto a_warp_window_tmp = make_tile_window(
+            a_block_window_tmp.get_bottom_tensor_view(),
+            make_tuple(number<WG::kM>{}, number<WG::kK>{}),
+            a_block_window_tmp.get_window_origin() + multi_index<2>{iMWarp * WG::kM, 0},
+            make_static_tile_distribution(typename WG::AWarpDstrEncoding{}));
+
+#if 0 // FIXME: using array will cause register spill
+        array<array<decltype(b_warp_window_tmp), KIterPerWarp>, NIterPerWarp> b_warp_windows{
+            {b_warp_window_tmp}};
+
+        for(index_t nIter = 0; nIter < NIterPerWarp; nIter++)
+        {
+            for(index_t kIter = 0; kIter < KIterPerWarp; kIter++)
+            {
+                move_tile_window(b_warp_windows(nIter)(kIter),
+                                 {nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
+            }
+        }
+#else
+        statically_indexed_array<
+            statically_indexed_array<decltype(a_warp_window_tmp), KIterPerWarp>,
+            MIterPerWarp>
+            a_warp_windows;
+
+        static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+            static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+                a_warp_windows(mIter)(kIter) = a_warp_window_tmp;
+
+                move_tile_window(a_warp_windows(mIter)(kIter),
+                                 {mIter * MPerBlockPerIter, kIter * KPerBlockPerIter});
+            });
+        });
+#endif
+
+        // check C-block-distribution
+        static_assert(
+            std::is_same_v<remove_cvref_t<decltype(c_block_dstr_encode)>,
+                           remove_cvref_t<decltype(CBlockTensor::get_tile_distribution()
+                                                       .get_static_tile_distribution_encoding())>>,
+            "wrong!");
+
+        using BWarpDstr = typename WG::BWarpDstr;
+        using CWarpDstr = typename WG::CWarpDstr;
+
+        using BWarpTensor = typename WG::BWarpTensor;
+        using CWarpTensor = typename WG::CWarpTensor;
+
+        constexpr auto b_warp_y_lengths =
+            to_sequence(BWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+        constexpr auto c_warp_y_lengths =
+            to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+
+        constexpr auto b_warp_y_index_zeros = uniform_sequence_gen_t<BWarpDstr::NDimY, 0>{};
+        constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
+
+        // hot loop:
+        static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+            static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                // read A warp tensor from A Block window
+                const auto a_warp_tensor = load_tile(a_warp_windows(mIter)(kIter));
+                static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                    // read B warp tensor from B block tensor
+                    BWarpTensor b_warp_tensor;
+
+                    b_warp_tensor.get_thread_buffer() = b_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(sequence<nIter, kIter>{}, b_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
+
+                    // read C warp tensor from C block tensor
+                    CWarpTensor c_warp_tensor;
+
+                    c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
+
+                    // warp GEMM
+                    WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
+
+                    // write C warp tensor into C block tensor
+                    c_block_tensor.set_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
+                        c_warp_tensor.get_thread_buffer());
+                });
+            });
+        });
+    }
+
+    CK_TILE_DEVICE static constexpr auto MakeCBlockTile()
+    {
+        constexpr index_t MPerBlock = BlockGemmShape::kM;
+        constexpr index_t NPerBlock = BlockGemmShape::kN;
+
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template at<0>())>;
+
+        constexpr index_t MWarp = config.template at<1>();
+        constexpr index_t NWarp = config.template at<2>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
+        // constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+            sequence<>,
+            tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
+            tuple<sequence<1, 2>>,
+            tuple<sequence<1, 1>>,
+            sequence<1, 2>,
+            sequence<0, 0>>{};
+
+        constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
+        constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
+        auto c_block_tensor         = make_static_distributed_tensor<CDataType>(c_block_dstr);
+        return c_block_tensor;
+    }
+
+    // C = A * B
+    template <typename ABlockWindowTmp, typename BBlockTensorTmp>
+    CK_TILE_DEVICE auto operator()(const ABlockWindowTmp& a_block_window_tmp,
+                                   const BBlockTensorTmp& b_block_tensor_tmp) const
+    {
+        auto c_block_tensor = MakeCBlockTile();
+        operator()(c_block_tensor, a_block_window_tmp, b_block_tensor_tmp);
+        return c_block_tensor;
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_asmem_breg_creg_v1_custom_policy.hpp

@@ -0,0 +1,36 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+
+namespace ck_tile {
+
+template <typename AType_,
+          typename BType_,
+          typename CType_,
+          typename BlockWarps_,
+          typename WarpGemm_>
+struct BlockGemmASmemBRegCRegV1CustomPolicy
+{
+    using AType = remove_cvref_t<AType_>;
+    using BType = remove_cvref_t<BType_>;
+    using CType = remove_cvref_t<CType_>;
+
+    using BlockWarps = remove_cvref_t<BlockWarps_>;
+
+    static constexpr index_t kMWarps = BlockWarps::at(number<0>{});
+    static constexpr index_t kNWarps = BlockWarps::at(number<1>{});
+    static constexpr index_t kKWarps = BlockWarps::at(number<2>{});
+
+    using WarpGemm = remove_cvref_t<WarpGemm_>;
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
+    {
+        return make_tuple(WarpGemm{}, kMWarps, kNWarps);
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_asmem_breg_creg_v1_default_policy.hpp

@@ -0,0 +1,56 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/warp/warp_gemm.hpp"
+
+namespace ck_tile {
+
+// Default policy for BlockGemmASmemBRegCRegV1
+// Default policy class should not be templated, put template on member functions instead
+struct BlockGemmASmemBRegCRegV1DefaultPolicy
+{
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
+    {
+        if constexpr(std::is_same_v<typename Problem::ADataType, half_t> &&
+                     std::is_same_v<typename Problem::BDataType, half_t> &&
+                     std::is_same_v<typename Problem::CDataType, float>)
+        {
+#if 0
+            constexpr index_t kBlockSize = Problem::kBlockSize;
+
+            constexpr index_t kMPerBlock = Problem::BlockGemmShape::kM;
+            constexpr index_t kNPerBlock = Problem::BlockGemmShape::kN;
+            constexpr index_t kKPerBlock = Problem::BlockGemmShape::kK;
+
+            static_assert(kBlockSize % get_warp_size() == 0, "wrong!");
+
+            constexpr index_t NumWarp = kBlockSize / get_warp_size();
+
+            // FIXME
+            if constexpr(NumWarp == 4 && kMPerBlock % 128 == 0 &&
+                         kNPerBlock % 128 == 0 % kKPerBlock % 16 == 0)
+            {
+                return make_tuple(WarpGemmMfmaF16F16F32M32N32K8{}, 4, 1);
+            }
+            else
+            {
+                return make_tuple(WarpGemmMfmaF16F16F32M32N32K8{}, 4, 1);
+            }
+#else
+            return make_tuple(WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution{}, 4, 1);
+#endif
+        }
+        else if constexpr(std::is_same_v<typename Problem::ADataType, bf16_t> &&
+                          std::is_same_v<typename Problem::BDataType, bf16_t> &&
+                          std::is_same_v<typename Problem::CDataType, float>)
+        {
+            return make_tuple(WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution{}, 4, 1);
+        }
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_asmem_bsmem_creg_v1.hpp

@@ -0,0 +1,213 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/block/block_gemm_asmem_bsmem_creg_v1_default_policy.hpp"
+
+namespace ck_tile {
+
+// A is block window on shared memory
+// B is block window on shared memory
+// C is block distributed tensor
+template <typename Problem_, typename Policy_ = BlockGemmASmemBSmemCRegV1DefaultPolicy>
+struct BlockGemmASmemBSmemCRegV1
+{
+    using Problem        = remove_cvref_t<Problem_>;
+    using Policy         = remove_cvref_t<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>;
+
+    static constexpr index_t kBlockSize = Problem::kBlockSize;
+
+    // C += A * B
+    template <typename CBlockTensor, typename ABlockWindow, typename BBlockWindow>
+    CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
+                                   const ABlockWindow& a_block_window,
+                                   const BBlockWindow& b_block_window) const
+    {
+        static_assert(std::is_same_v<ADataType, typename ABlockWindow::DataType> &&
+                          std::is_same_v<BDataType, typename BBlockWindow::DataType> &&
+                          std::is_same_v<CDataType, typename CBlockTensor::DataType>,
+                      "wrong!");
+
+        constexpr index_t MPerBlock = ABlockWindow{}.get_window_lengths()[number<0>{}];
+        constexpr index_t NPerBlock = BBlockWindow{}.get_window_lengths()[number<0>{}];
+        constexpr index_t KPerBlock = ABlockWindow{}.get_window_lengths()[number<1>{}];
+
+        static_assert(MPerBlock == BlockGemmShape::kM && NPerBlock == BlockGemmShape::kN &&
+                          KPerBlock == BlockGemmShape::kK,
+                      "wrong!");
+
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template at<0>())>;
+
+        constexpr index_t MWarp = config.template at<1>();
+        constexpr index_t NWarp = config.template at<2>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
+        constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        constexpr index_t MPerBlockPerIter = MPerBlock / MIterPerWarp;
+        constexpr index_t NPerBlockPerIter = NPerBlock / NIterPerWarp;
+        constexpr index_t KPerBlockPerIter = KPerBlock / KIterPerWarp;
+
+        const index_t iMWarp = get_warp_id() / NWarp;
+        const index_t iNWarp = get_warp_id() % NWarp;
+
+        // construct A-warp-window
+        auto a_warp_window_tmp = make_tile_window(
+            a_block_window.get_bottom_tensor_view(),
+            make_tuple(number<WG::kM>{}, number<WG::kK>{}),
+            a_block_window.get_window_origin() + multi_index<2>{iMWarp * WG::kM, 0},
+            make_static_tile_distribution(typename WG::AWarpDstrEncoding{}));
+
+#if 0 // FIXME: using array will cause register spill
+        array<array<decltype(a_warp_window_tmp), KIterPerWarp>, MIterPerWarp> a_warp_windows{
+            {a_warp_window_tmp}};
+
+        for(index_t mIter = 0; mIter < MIterPerWarp; mIter++)
+        {
+            for(index_t kIter = 0; kIter < KIterPerWarp; kIter++)
+            {
+                move_tile_window(a_warp_windows(mIter)(kIter),
+                                 {mIter * MPerBlockPerIter, kIter * KPerBlockPerIter});
+            }
+        }
+#else
+        statically_indexed_array<
+            statically_indexed_array<decltype(a_warp_window_tmp), KIterPerWarp>,
+            MIterPerWarp>
+            a_warp_windows;
+
+        static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+            static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+                a_warp_windows(mIter)(kIter) = a_warp_window_tmp;
+
+                move_tile_window(a_warp_windows(mIter)(kIter),
+                                 {mIter * MPerBlockPerIter, kIter * KPerBlockPerIter});
+            });
+        });
+#endif
+
+        // construct B-warp-window
+        auto b_warp_window_tmp = make_tile_window(
+            b_block_window.get_bottom_tensor_view(),
+            make_tuple(number<WG::kN>{}, number<WG::kK>{}),
+            b_block_window.get_window_origin() + multi_index<2>{iNWarp * WG::kN, 0},
+            make_static_tile_distribution(typename WG::BWarpDstrEncoding{}));
+
+#if 0 // FIXME: using array will cause register spill
+        array<array<decltype(b_warp_window_tmp), KIterPerWarp>, NIterPerWarp> b_warp_windows{
+            {b_warp_window_tmp}};
+
+        for(index_t nIter = 0; nIter < NIterPerWarp; nIter++)
+        {
+            for(index_t kIter = 0; kIter < KIterPerWarp; kIter++)
+            {
+                move_tile_window(b_warp_windows(nIter)(kIter),
+                                 {nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
+            }
+        }
+#else
+        statically_indexed_array<
+            statically_indexed_array<decltype(b_warp_window_tmp), KIterPerWarp>,
+            NIterPerWarp>
+            b_warp_windows;
+
+        static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+            static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+                b_warp_windows(nIter)(kIter) = b_warp_window_tmp;
+
+                move_tile_window(b_warp_windows(nIter)(kIter),
+                                 {nIter * NPerBlockPerIter, kIter * KPerBlockPerIter});
+            });
+        });
+#endif
+
+        using CWarpDstr   = typename WG::CWarpDstr;
+        using CWarpTensor = typename WG::CWarpTensor;
+
+        constexpr auto c_warp_y_lengths =
+            to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+        constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
+
+        // hot loop:
+        static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+            static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                // read A warp tensor from A block window
+                const auto a_warp_tensor = load_tile(a_warp_windows(mIter)(kIter));
+
+                static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                    // read B warp tensor from B Block window
+                    const auto b_warp_tensor = load_tile(b_warp_windows(nIter)(kIter));
+
+                    // read C warp tensor from C block tensor
+                    CWarpTensor c_warp_tensor;
+
+                    c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
+
+                    // warp GEMM
+                    WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
+
+                    // write C warp tensor into C block tensor
+                    c_block_tensor.set_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
+                        c_warp_tensor.get_thread_buffer());
+                });
+            });
+        });
+    }
+
+    CK_TILE_DEVICE static constexpr auto MakeCBlockTile()
+    {
+        constexpr index_t MPerBlock = BlockGemmShape::kM;
+        constexpr index_t NPerBlock = BlockGemmShape::kN;
+
+        constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WG = remove_cvref_t<decltype(config.template at<0>())>;
+
+        constexpr index_t MWarp = config.template at<1>();
+        constexpr index_t NWarp = config.template at<2>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
+
+        constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+            sequence<>,
+            tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
+            tuple<sequence<1, 2>>,
+            tuple<sequence<1, 1>>,
+            sequence<1, 2>,
+            sequence<0, 0>>{};
+
+        constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
+
+        constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
+
+        auto c_block_tensor = make_static_distributed_tensor<CDataType>(c_block_dstr);
+        return c_block_tensor;
+    }
+
+    // C = A * B
+    template <typename ABlockTensorTmp, typename BBlockWindow>
+    CK_TILE_DEVICE auto operator()(const ABlockTensorTmp& a_block_tensor_tmp,
+                                   const BBlockWindow& b_block_window) const
+    {
+        auto c_block_tensor = MakeCBlockTile();
+        operator()(c_block_tensor, a_block_tensor_tmp, b_block_window);
+        return c_block_tensor;
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_asmem_bsmem_creg_v1_custom_policy.hpp

@@ -0,0 +1,38 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+
+namespace ck_tile {
+
+// Default policy for BlockGemmASmemBSmemCRegV1
+// Default policy class should not be templated, put template on member functions instead
+template <typename AType_,
+          typename BType_,
+          typename CType_,
+          typename BlockWarps_,
+          typename WarpGemm_>
+struct BlockGemmASmemBSmemCRegV1CustomPolicy
+{
+    using AType = remove_cvref_t<AType_>;
+    using BType = remove_cvref_t<BType_>;
+    using CType = remove_cvref_t<CType_>;
+
+    using BlockWarps = remove_cvref_t<BlockWarps_>;
+
+    static constexpr index_t kMWarps = BlockWarps::at(number<0>{});
+    static constexpr index_t kNWarps = BlockWarps::at(number<1>{});
+    static constexpr index_t kKWarps = BlockWarps::at(number<2>{});
+
+    using WarpGemm = remove_cvref_t<WarpGemm_>;
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
+    {
+        return make_tuple(WarpGemm{}, kMWarps, kNWarps);
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_asmem_bsmem_creg_v1_default_policy.hpp

@@ -0,0 +1,76 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/common/tensor_layout.hpp"
+#include "ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp"
+
+namespace ck_tile {
+
+// Default policy for BlockGemmASmemBSmemCRegV1
+// Default policy class should not be templated, put template on member functions instead
+struct BlockGemmASmemBSmemCRegV1DefaultPolicy
+{
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
+    {
+#if defined(__gfx950__)
+        constexpr bool is_a_load_tr = std::is_same_v<remove_cvref_t<typename Problem::ALayout>,
+                                                     tensor_layout::gemm::ColumnMajor> &&
+                                      !std::is_same_v<typename Problem::ADataType, float>;
+        constexpr bool is_b_load_tr = std::is_same_v<remove_cvref_t<typename Problem::BLayout>,
+                                                     tensor_layout::gemm::RowMajor> &&
+                                      !std::is_same_v<typename Problem::BDataType, float>;
+#else
+        constexpr bool is_a_load_tr = false;
+        constexpr bool is_b_load_tr = false;
+#endif
+        constexpr auto wg_attr_num_access = (is_a_load_tr || is_b_load_tr)
+                                                ? WGAttrNumAccessEnum::Double
+                                                : WGAttrNumAccessEnum::Single;
+
+        if constexpr(((std::is_same_v<typename Problem::ADataType, half_t> &&
+                       std::is_same_v<typename Problem::BDataType, half_t>) ||
+                      (std::is_same_v<typename Problem::ADataType, bf16_t> &&
+                       std::is_same_v<typename Problem::BDataType, bf16_t>)) &&
+                     std::is_same_v<typename Problem::CDataType, float>)
+        {
+            if constexpr(get_warp_size() == 64)
+            {
+                using WG = WarpGemmDispatcher<typename Problem::ADataType,
+                                              typename Problem::BDataType,
+                                              typename Problem::CDataType,
+                                              32,
+                                              32,
+                                              16,
+                                              true,
+                                              false,
+                                              false,
+                                              wg_attr_num_access>;
+                return make_tuple(WG{}, 4, 1);
+            }
+            else
+            {
+                using WG = WarpGemmDispatcher<typename Problem::ADataType,
+                                              typename Problem::BDataType,
+                                              typename Problem::CDataType,
+                                              16,
+                                              16,
+                                              16,
+                                              true,
+                                              false,
+                                              false,
+                                              wg_attr_num_access>;
+                return make_tuple(WG{}, 4, 1);
+            }
+        }
+        else
+        {
+            static_assert(false, "Unsupported data type configuration for GEMM warp execution.");
+        }
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_mx_areg_bsmem_creg_v1.hpp

@@ -0,0 +1,374 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+
+namespace ck_tile {
+
+// A is block distributed tensor
+// A scale is block distributed tensor
+// B is block window on shared memory
+// B scale is block distributed tensor
+// C is block distributed tensor
+// It supports only warp gemms with transposed C.
+// TargetCMPerLane_ controls how many consecutive elements of matrix C are calculated by each lane.
+template <typename Problem_, typename Policy_, index_t TargetCMPerLane_ = -1>
+struct BlockGemmMxARegBSmemCRegV1
+{
+    using Problem        = remove_cvref_t<Problem_>;
+    using Policy         = remove_cvref_t<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>;
+
+    static constexpr index_t kBlockSize = 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 auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+    using WarpGemm = remove_cvref_t<decltype(config.template at<0>())>;
+
+    static constexpr index_t MWarp = config.template at<1>();
+    static constexpr index_t NWarp = config.template at<2>();
+
+    static constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WarpGemm::kM);
+    static constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WarpGemm::kN);
+    static constexpr index_t KIterPerWarp = KPerBlock / WarpGemm::kK;
+
+    static constexpr index_t CMPerLane = WarpGemm::WarpGemmAttribute::Impl::kCM0PerLane *
+                                         WarpGemm::WarpGemmAttribute::Impl::kCM1PerLane;
+    static constexpr index_t TargetCMPerLane = max(CMPerLane, TargetCMPerLane_);
+
+    static_assert(TargetCMPerLane % CMPerLane == 0);
+    static constexpr index_t NIterPack = TargetCMPerLane / CMPerLane;
+
+    // C += A * B
+    template <typename CBlockTensor,
+              typename ABlockTensorTmp,
+              typename AScaleBlockTensorTmp,
+              typename BBlockWindowTmp,
+              typename BScaleBlockTensorTmp>
+    CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
+                                   const ABlockTensorTmp& a_block_tensor_tmp,
+                                   const AScaleBlockTensorTmp& a_scale_block_tensor_tmp,
+                                   const BBlockWindowTmp& b_block_window_tmp,
+                                   const BScaleBlockTensorTmp& b_scale_block_tensor_tmp) const
+    {
+        static_assert(std::is_same_v<ADataType, remove_cv_t<typename ABlockTensorTmp::DataType>> &&
+                      std::is_same_v<BDataType, remove_cv_t<typename BBlockWindowTmp::DataType>> &&
+                      std::is_same_v<CDataType, remove_cv_t<typename CBlockTensor::DataType>>);
+
+        static_assert(MPerBlock == ABlockTensorTmp{}.get_lengths()[number<0>{}] &&
+                      NPerBlock == BBlockWindowTmp{}.get_window_lengths()[number<0>{}] &&
+                      KPerBlock == ABlockTensorTmp{}.get_lengths()[number<1>{}]);
+
+        const index_t iNWarp = get_warp_id() % NWarp;
+
+        // construct A-block-tensor from A-Block-tensor-tmp
+        auto a_block_tensor = make_static_distributed_tensor<typename ABlockTensorTmp::DataType>(
+            MakeABlockTileDistribution());
+        a_block_tensor.get_thread_buffer() = a_block_tensor_tmp.get_thread_buffer();
+
+        auto a_scale_block_tensor =
+            make_static_distributed_tensor<remove_cv_t<typename AScaleBlockTensorTmp::DataType>>(
+                MakeAScaleBlockTileDistribution());
+        a_scale_block_tensor.get_thread_buffer() = a_scale_block_tensor_tmp.get_thread_buffer();
+
+        auto b_scale_block_tensor =
+            make_static_distributed_tensor<remove_cv_t<typename BScaleBlockTensorTmp::DataType>>(
+                MakeBScaleBlockTileDistribution());
+        b_scale_block_tensor.get_thread_buffer() = b_scale_block_tensor_tmp.get_thread_buffer();
+
+        // Construct B-warp-window
+        // Matrix B is shuffled in such a way that each lane calculates TargetCMPerLane consecutive
+        // elements of matrix C. See MakeBScaleBlockTileDistribution and MakeCBlockTile that shuffle
+        // B scale and C in the same way.
+        auto b_warp_window_tmp = [&] {
+            using Impl = typename WarpGemm::WarpGemmAttribute::Impl;
+
+            constexpr index_t N3 = Impl::kCM1PerLane;
+            constexpr index_t N2 = TargetCMPerLane / N3;
+            constexpr index_t N1 = Impl::kCMLane;
+            constexpr index_t N0 = NPerBlock / (N1 * N2 * N3);
+
+            const auto b_lds_unmerged = transform_tensor_view(
+                b_block_window_tmp.get_bottom_tensor_view(),
+                make_tuple(make_unmerge_transform(
+                               make_tuple(number<N0>{}, number<N1>{}, number<N2>{}, number<N3>{})),
+                           make_pass_through_transform(number<KPerBlock>{})),
+                make_tuple(sequence<0>{}, sequence<1>{}),
+                make_tuple(sequence<0, 2, 1, 3>{}, sequence<4>{}));
+
+            const auto b_lds_merged = transform_tensor_view(
+                b_lds_unmerged,
+                make_tuple(make_merge_transform(
+                               make_tuple(number<N0>{}, number<N2>{}, number<N1>{}, number<N3>{})),
+                           make_pass_through_transform(number<KPerBlock>{})),
+                make_tuple(sequence<0, 1, 2, 3>{}, sequence<4>{}),
+                make_tuple(sequence<0>{}, sequence<1>{}));
+
+            return make_tile_window(
+                b_lds_merged,
+                make_tuple(number<WarpGemm::kN>{}, number<WarpGemm::kK>{}),
+                b_block_window_tmp.get_window_origin() + multi_index<2>{iNWarp * WarpGemm::kN, 0},
+                make_static_tile_distribution(typename WarpGemm::BWarpDstrEncoding{}));
+        }();
+
+        // check C-block-distribution
+        static_assert(
+            std::is_same_v<remove_cvref_t<decltype(MakeCBlockTile()
+                                                       .get_tile_distribution()
+                                                       .get_static_tile_distribution_encoding())>,
+                           remove_cvref_t<decltype(CBlockTensor::get_tile_distribution()
+                                                       .get_static_tile_distribution_encoding())>>);
+
+        using AWarpDstr = typename WarpGemm::AWarpDstr;
+        using CWarpDstr = typename WarpGemm::CWarpDstr;
+
+        using AWarpTensor = typename WarpGemm::AWarpTensor;
+        using CWarpTensor = typename WarpGemm::CWarpTensor;
+
+        using AScaleWarpDstr =
+            remove_cvref_t<decltype(make_static_tile_distribution(MakeAScaleWarpDstrEncoding()))>;
+        using AScaleWarpTensor =
+            static_distributed_tensor<remove_cv_t<typename AScaleBlockTensorTmp::DataType>,
+                                      AScaleWarpDstr>;
+
+        using BScaleWarpDstr =
+            remove_cvref_t<decltype(make_static_tile_distribution(MakeBScaleWarpDstrEncoding()))>;
+        using BScaleWarpTensor =
+            static_distributed_tensor<remove_cv_t<typename BScaleBlockTensorTmp::DataType>,
+                                      BScaleWarpDstr>;
+
+        constexpr auto a_warp_y_lengths =
+            to_sequence(AWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+        constexpr auto c_warp_y_lengths =
+            to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+
+        constexpr auto a_warp_y_index_zeros = uniform_sequence_gen_t<AWarpDstr::NDimY, 0>{};
+        constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
+
+        constexpr auto a_scale_warp_y_lengths =
+            to_sequence(AScaleWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+        constexpr auto b_scale_warp_y_lengths =
+            to_sequence(BScaleWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+
+        constexpr auto a_scale_warp_y_index_zeros =
+            uniform_sequence_gen_t<AScaleWarpDstr::NDimY, 0>{};
+        constexpr auto b_scale_warp_y_index_zeros =
+            uniform_sequence_gen_t<BScaleWarpDstr::NDimY, 0>{};
+
+        // hot loop:
+        static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+            static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                auto b_warp_window = b_warp_window_tmp;
+                move_tile_window(
+                    b_warp_window,
+                    {nIter * (NPerBlock / NIterPerWarp), kIter * (KPerBlock / KIterPerWarp)});
+                // read B warp tensor from B Block window
+                const auto b_warp_tensor = load_tile(b_warp_window);
+
+                BScaleWarpTensor b_scale_warp_tensor;
+
+                b_scale_warp_tensor.get_thread_buffer() =
+                    b_scale_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(sequence<nIter / NIterPack, nIter % NIterPack, kIter>{},
+                                        b_scale_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1, 1>{}, b_scale_warp_y_lengths));
+
+                static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                    // read A warp tensor from A block tensor
+                    AWarpTensor a_warp_tensor;
+
+                    a_warp_tensor.get_thread_buffer() = a_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
+
+                    AScaleWarpTensor a_scale_warp_tensor;
+
+                    a_scale_warp_tensor.get_thread_buffer() =
+                        a_scale_block_tensor.get_y_sliced_thread_data(
+                            merge_sequences(sequence<mIter, kIter>{}, a_scale_warp_y_index_zeros),
+                            merge_sequences(sequence<1, 1>{}, a_scale_warp_y_lengths));
+
+                    // read C warp tensor from C block tensor
+                    CWarpTensor c_warp_tensor;
+
+                    c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, nIter / NIterPack, nIter % NIterPack>{},
+                                        c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1, 1>{}, c_warp_y_lengths));
+
+                    // warp GEMM
+                    WarpGemm{}.template operator()<0, 0>(
+                        c_warp_tensor,
+                        a_warp_tensor,
+                        b_warp_tensor,
+                        int32_t(a_scale_warp_tensor.get_thread_buffer()[0]),
+                        int32_t(b_scale_warp_tensor.get_thread_buffer()[0]));
+
+                    // write C warp tensor into C block tensor
+                    c_block_tensor.set_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, nIter / NIterPack, nIter % NIterPack>{},
+                                        c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1, 1>{}, c_warp_y_lengths),
+                        c_warp_tensor.get_thread_buffer());
+                });
+            });
+        });
+    }
+
+    template <index_t MPerBlock_ = MPerBlock, index_t KPerBlock_ = KPerBlock>
+    CK_TILE_DEVICE static constexpr auto MakeABlockTileDistribution()
+    {
+        constexpr index_t MIterPerWarp_ = MPerBlock_ / (MWarp * WarpGemm::kM);
+        constexpr index_t KIterPerWarp_ = KPerBlock_ / WarpGemm::kK;
+
+        constexpr auto a_block_outer_dstr_encoding = tile_distribution_encoding<
+            sequence<NWarp>,
+            tuple<sequence<MIterPerWarp_, MWarp>, sequence<KIterPerWarp_>>,
+            tuple<sequence<1, 0>>,
+            tuple<sequence<1, 0>>,
+            sequence<1, 2>,
+            sequence<0, 0>>{};
+
+        constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            a_block_outer_dstr_encoding, typename WarpGemm::AWarpDstrEncoding{});
+
+        return make_static_tile_distribution(a_block_dstr_encode);
+    }
+
+    CK_TILE_DEVICE static constexpr auto MakeAScaleWarpDstrEncoding()
+    {
+        using Impl = typename WarpGemm::WarpGemmAttribute::Impl;
+
+        constexpr index_t AScaleMLane     = Impl::kAMLane;
+        constexpr index_t ABScaleKLane    = Impl::kABKLane;
+        constexpr index_t ABScaleKPerLane = Impl::kABKPerLane / Impl::kScaleGranularity;
+
+        return ck_tile::tile_distribution_encoding<
+            ck_tile::sequence<>,
+            ck_tile::tuple<ck_tile::sequence<AScaleMLane>,
+                           ck_tile::sequence<ABScaleKLane, ABScaleKPerLane>>,
+            ck_tile::tuple<ck_tile::sequence<2, 1>>,
+            ck_tile::tuple<ck_tile::sequence<0, 0>>,
+            ck_tile::sequence<2>,
+            ck_tile::sequence<1>>{};
+    }
+
+    CK_TILE_DEVICE static constexpr auto MakeBScaleWarpDstrEncoding()
+    {
+        using Impl = typename WarpGemm::WarpGemmAttribute::Impl;
+
+        constexpr index_t BScaleNLane     = Impl::kBNLane;
+        constexpr index_t ABScaleKLane    = Impl::kABKLane;
+        constexpr index_t ABScaleKPerLane = Impl::kABKPerLane / Impl::kScaleGranularity;
+
+        return ck_tile::tile_distribution_encoding<
+            ck_tile::sequence<>,
+            ck_tile::tuple<ck_tile::sequence<BScaleNLane>,
+                           ck_tile::sequence<ABScaleKLane, ABScaleKPerLane>>,
+            ck_tile::tuple<ck_tile::sequence<2, 1>>,
+            ck_tile::tuple<ck_tile::sequence<0, 0>>,
+            ck_tile::sequence<2>,
+            ck_tile::sequence<1>>{};
+    }
+
+    template <index_t MPerBlock_ = MPerBlock, index_t KPerBlock_ = KPerBlock>
+    CK_TILE_DEVICE static constexpr auto MakeAScaleBlockTileDistribution()
+    {
+        constexpr index_t MIterPerWarp_ = MPerBlock_ / (MWarp * WarpGemm::kM);
+        constexpr index_t KIterPerWarp_ = KPerBlock_ / WarpGemm::kK;
+
+        constexpr auto a_scale_block_outer_dstr_encoding = tile_distribution_encoding<
+            sequence<NWarp>,
+            tuple<sequence<MIterPerWarp_, MWarp>, sequence<KIterPerWarp_>>,
+            tuple<sequence<1, 0>>,
+            tuple<sequence<1, 0>>,
+            sequence<1, 2>,
+            sequence<0, 0>>{};
+
+        constexpr auto a_scale_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            a_scale_block_outer_dstr_encoding, MakeAScaleWarpDstrEncoding());
+
+        return make_static_tile_distribution(a_scale_block_dstr_encode);
+    }
+
+    template <index_t NPerBlock_ = NPerBlock, index_t KPerBlock_ = KPerBlock>
+    CK_TILE_DEVICE static constexpr auto MakeBScaleBlockTileDistribution()
+    {
+        constexpr index_t NIterPerWarp_ = NPerBlock_ / (NWarp * WarpGemm::kN);
+        constexpr index_t KIterPerWarp_ = KPerBlock_ / WarpGemm::kK;
+
+        using Impl = typename WarpGemm::WarpGemmAttribute::Impl;
+
+        constexpr index_t ABScaleKLane    = Impl::kABKLane;
+        constexpr index_t ABScaleKPerLane = Impl::kABKPerLane / Impl::kScaleGranularity;
+
+        constexpr auto b_scale_block_dstr_encode = ck_tile::tile_distribution_encoding<
+            ck_tile::sequence<MWarp>,
+            ck_tile::tuple<ck_tile::sequence<NIterPerWarp_ / NIterPack,
+                                             NWarp,
+                                             Impl::kCMLane,
+                                             NIterPack,
+                                             Impl::kCM0PerLane,
+                                             Impl::kCM1PerLane>,
+                           ck_tile::sequence<KIterPerWarp_, ABScaleKLane, ABScaleKPerLane>>,
+            ck_tile::tuple<ck_tile::sequence<0, 1>, ck_tile::sequence<2, 1, 1, 1>>,
+            ck_tile::tuple<ck_tile::sequence<0, 1>, ck_tile::sequence<1, 4, 2, 5>>,
+            ck_tile::sequence<1, 1, 2, 2>,
+            ck_tile::sequence<0, 3, 0, 2>>{};
+
+        return make_static_tile_distribution(b_scale_block_dstr_encode);
+    }
+
+    CK_TILE_DEVICE static constexpr auto MakeCBlockTile()
+    {
+        using Impl = typename WarpGemm::WarpGemmAttribute::Impl;
+
+        constexpr auto c_block_dstr_encode = ck_tile::tile_distribution_encoding<
+            ck_tile::sequence<>,
+            ck_tile::tuple<ck_tile::sequence<MIterPerWarp, MWarp, Impl::kCNLane>,
+                           ck_tile::sequence<NIterPerWarp / NIterPack,
+                                             NWarp,
+                                             Impl::kCMLane,
+                                             NIterPack,
+                                             Impl::kCM0PerLane,
+                                             Impl::kCM1PerLane>>,
+            ck_tile::tuple<ck_tile::sequence<1, 2>, ck_tile::sequence<2, 1>>,
+            ck_tile::tuple<ck_tile::sequence<1, 1>, ck_tile::sequence<2, 2>>,
+            ck_tile::sequence<1, 2, 2, 2, 2>,
+            ck_tile::sequence<0, 0, 3, 4, 5>>{};
+
+        constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
+        auto c_block_tensor         = make_static_distributed_tensor<CDataType>(c_block_dstr);
+        return c_block_tensor;
+    }
+
+    // C = A * B
+    template <typename ABlockTensorTmp,
+              typename AScaleBlockTensorTmp,
+              typename BBlockWindowTmp,
+              typename BScaleBlockTensorTmp>
+    CK_TILE_DEVICE auto operator()(const ABlockTensorTmp& a_block_tensor_tmp,
+                                   const AScaleBlockTensorTmp& a_scale_block_tensor_tmp,
+                                   const BBlockWindowTmp& b_block_window_tmp,
+                                   const BScaleBlockTensorTmp& b_scale_block_tensor_tmp) const
+    {
+        auto c_block_tensor = MakeCBlockTile();
+        operator()(c_block_tensor,
+                   a_block_tensor_tmp,
+                   a_scale_block_tensor_tmp,
+                   b_block_window_tmp,
+                   b_scale_block_tensor_tmp);
+        return c_block_tensor;
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_mx_areg_bsmem_creg_v1_custom_policy.hpp

@@ -0,0 +1,36 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+
+namespace ck_tile {
+
+template <typename AType_,
+          typename BType_,
+          typename CType_,
+          typename BlockWarps_,
+          typename WarpGemm_>
+struct BlockGemmMxARegBSmemCRegV1CustomPolicy
+{
+    using AType = remove_cvref_t<AType_>;
+    using BType = remove_cvref_t<BType_>;
+    using CType = remove_cvref_t<CType_>;
+
+    using BlockWarps = remove_cvref_t<BlockWarps_>;
+
+    static constexpr index_t kMWarps = BlockWarps::at(number<0>{});
+    static constexpr index_t kNWarps = BlockWarps::at(number<1>{});
+    static constexpr index_t kKWarps = BlockWarps::at(number<2>{});
+
+    using WarpGemm = remove_cvref_t<WarpGemm_>;
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
+    {
+        return make_tuple(WarpGemm{}, kMWarps, kNWarps);
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_gemm_problem.hpp

@@ -0,0 +1,28 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+
+namespace ck_tile {
+
+// Problem Description for BlockGemm
+template <typename ADataType_,
+          typename BDataType_,
+          typename CDataType_,
+          index_t kBlockSize_,
+          typename BlockGemmShape_,
+          index_t NumWaveGroups_ = 1>
+struct BlockGemmProblem
+{
+    using ADataType      = remove_cvref_t<ADataType_>;
+    using BDataType      = remove_cvref_t<BDataType_>;
+    using CDataType      = remove_cvref_t<CDataType_>;
+    using BlockGemmShape = remove_cvref_t<BlockGemmShape_>;
+
+    static constexpr index_t kBlockSize    = kBlockSize_;
+    static constexpr index_t NumWaveGroups = NumWaveGroups_;
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_universal_gemm_as_bs_cr.hpp

@@ -0,0 +1,527 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/common/load_and_convert_tile.hpp"
+#include "ck_tile/ops/gemm/block/block_gemm_asmem_bsmem_creg_v1_default_policy.hpp"
+#include "ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_scheduler.hpp"
+#include "ck_tile/ops/elementwise.hpp"
+
+namespace ck_tile {
+
+// A is block window on shared memory
+// B is block window on shared memory
+// C is block distributed tensor
+template <typename Problem_,
+          typename Policy_     = BlockGemmASmemBSmemCRegV1DefaultPolicy,
+          index_t UnaryOpSize_ = 8>
+struct BlockUniversalGemmAsBsCr
+{
+    private:
+    // TODO: This should be in Policy - UniversalGemmPolicyBase ?
+    template <typename PipelineProblem_, typename GemmPolicy_>
+    struct GemmTraits_
+    {
+        using Problem         = remove_cvref_t<PipelineProblem_>;
+        using Policy          = remove_cvref_t<GemmPolicy_>;
+        using ADataType       = remove_cvref_t<typename Problem::ADataType>;
+        using BDataType       = remove_cvref_t<typename Problem::BDataType>;
+        using ComputeDataType = remove_cvref_t<typename Problem::ComputeDataType>;
+        using CDataType       = remove_cvref_t<typename Problem::CDataType>;
+        using BlockGemmShape  = remove_cvref_t<typename Problem::BlockGemmShape>;
+
+        static constexpr index_t kBlockSize = Problem::kBlockSize;
+        static constexpr auto Scheduler     = Problem::Scheduler;
+
+        static constexpr index_t MPerBlock = BlockGemmShape::kM;
+        static constexpr index_t NPerBlock = BlockGemmShape::kN;
+        static constexpr index_t KPerBlock = BlockGemmShape::kK;
+
+        static constexpr auto config = Policy::template GetWarpGemmMWarpNWarp<Problem>();
+
+        using WarpGemm = remove_cvref_t<decltype(config.template at<0>())>;
+
+        static constexpr index_t MWarp = config.template at<1>();
+        static constexpr index_t NWarp = config.template at<2>();
+
+        using I0 = number<0>;
+        using I1 = number<1>;
+
+        static_assert(MWarp == BlockGemmShape::BlockWarps::at(I0{}),
+                      "Error! WarpGemm's MWarp is not consisten with BlockGemmShape!");
+        static_assert(NWarp == BlockGemmShape::BlockWarps::at(I1{}),
+                      "Error! WarpGemm's NWarp is not consisten with BlockGemmShape!");
+        static_assert(WarpGemm::kM == BlockGemmShape::WarpTile::at(I0{}),
+                      "Error! WarpGemm's M is not consisten with BlockGemmShape!");
+        static_assert(WarpGemm::kN == BlockGemmShape::WarpTile::at(I1{}),
+                      "Error! WarpGemm's N is not consisten with BlockGemmShape!");
+
+        static constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WarpGemm::kM);
+        static constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WarpGemm::kN);
+        static constexpr index_t KIterPerWarp = KPerBlock / WarpGemm::kK;
+
+        static_assert(MIterPerWarp * MWarp * WarpGemm::kM == MPerBlock,
+                      "Error! Warps should cover all Block tile!");
+        static_assert(NIterPerWarp * NWarp * WarpGemm::kN == NPerBlock,
+                      "Error! Warps should cover all Block tile!");
+
+        static constexpr index_t MPerBlockPerIter = MWarp * WarpGemm::kM;
+        static constexpr index_t NPerBlockPerIter = NWarp * WarpGemm::kN;
+        static constexpr index_t KPerBlockPerIter = WarpGemm::kK;
+
+        // Controls how many MAC clusters (MFMA blocks) we have per wave
+        // Ie if
+        // InterWaveSchedulingMacClusters = 1;
+        // KPerBlock == 32
+        // WarpGemm::kK = 8
+        // Then we would group all 4 WarpGemms into single MAC cluster.
+        // But if we would set InterWaveSchedulingMacClusters = 2, then we would
+        // split those 4 warp gemms into two groups.
+        static constexpr index_t InterWaveSchedulingMacClusters = 1;
+
+        // should be at least equal to: WarpGemm::Impl::kABKPerLane
+        static constexpr index_t KPack      = WarpGemm::kKPerThread;
+        static constexpr index_t KPerThread = KIterPerWarp * WarpGemm::kKPerThread;
+    };
+
+    public:
+    using Traits = GemmTraits_<Problem_, Policy_>;
+
+    using ADataType       = remove_cvref_t<typename Traits::ADataType>;
+    using BDataType       = remove_cvref_t<typename Traits::BDataType>;
+    using ComputeDataType = remove_cvref_t<typename Traits::ComputeDataType>;
+    using CDataType       = remove_cvref_t<typename Traits::CDataType>;
+
+    using ATypeToUse =
+        std::conditional_t<std::is_same_v<ADataType, pk_int4_t>, BDataType, ADataType>;
+    using BTypeToUse = std::conditional_t<std::is_same_v<BDataType, pk_int4_t> ||
+                                              std::is_same_v<BDataType, pk_fp4_t> ||
+                                              sizeof(BDataType) < sizeof(ADataType),
+                                          ADataType,
+                                          BDataType>;
+
+    using WarpGemm = remove_cvref_t<typename Traits::WarpGemm>;
+
+    static constexpr index_t KIterPerWarp = Traits::KIterPerWarp;
+    static constexpr index_t MIterPerWarp = Traits::MIterPerWarp;
+    static constexpr index_t NIterPerWarp = Traits::NIterPerWarp;
+
+    static constexpr index_t MWarp = Traits::MWarp;
+    static constexpr index_t NWarp = Traits::NWarp;
+
+    static constexpr auto Scheduler = Traits::Scheduler;
+
+    using AWarpDstr = typename WarpGemm::AWarpDstr;
+    using BWarpDstr = typename WarpGemm::BWarpDstr;
+    using CWarpDstr = typename WarpGemm::CWarpDstr;
+
+    using AWarpTensor = typename WarpGemm::AWarpTensor;
+    using BWarpTensor = typename WarpGemm::BWarpTensor;
+    using CWarpTensor = typename WarpGemm::CWarpTensor;
+
+    static constexpr auto a_warp_y_lengths =
+        to_sequence(AWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+    static constexpr auto b_warp_y_lengths =
+        to_sequence(BWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+    static constexpr auto c_warp_y_lengths =
+        to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+
+    static constexpr auto a_warp_y_index_zeros = uniform_sequence_gen_t<AWarpDstr::NDimY, 0>{};
+    static constexpr auto b_warp_y_index_zeros = uniform_sequence_gen_t<BWarpDstr::NDimY, 0>{};
+    static constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
+
+    static constexpr index_t APackedSize =
+        ck_tile::numeric_traits<remove_cvref_t<ADataType>>::PackedSize;
+    static constexpr index_t BPackedSize =
+        ck_tile::numeric_traits<remove_cvref_t<BDataType>>::PackedSize;
+
+    using I0 = number<0>;
+    using I1 = number<1>;
+
+    CK_TILE_DEVICE static constexpr auto MakeABlockDistributionEncode()
+    {
+        constexpr index_t KPerThread     = Traits::KPerThread;
+        constexpr index_t NumMacClusters = Traits::InterWaveSchedulingMacClusters;
+        constexpr index_t KPerInnerLoop =
+            ck_tile::max(KPerThread / NumMacClusters, WarpGemm::kKPerThread);
+        constexpr index_t KIterInterwave = KPerInnerLoop / WarpGemm::kKPerThread;
+
+        using KIterSeq = std::conditional_t<Scheduler == GemmPipelineScheduler::Interwave,
+                                            sequence<KIterInterwave>,
+                                            sequence<KIterPerWarp>>;
+
+        constexpr auto a_block_outer_dstr_encoding =
+            tile_distribution_encoding<sequence<NWarp>,
+                                       tuple<sequence<MIterPerWarp, MWarp>, KIterSeq>,
+                                       tuple<sequence<1, 0>>,
+                                       tuple<sequence<1, 0>>,
+                                       sequence<1, 2>,
+                                       sequence<0, 0>>{};
+        constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            a_block_outer_dstr_encoding, typename WarpGemm::AWarpDstrEncoding{});
+
+        return a_block_dstr_encode;
+    }
+
+    CK_TILE_DEVICE static constexpr auto MakeBBlockDistributionEncode()
+    {
+        constexpr index_t KPerThread     = Traits::KPerThread;
+        constexpr index_t NumMacClusters = Traits::InterWaveSchedulingMacClusters;
+        constexpr index_t KPerInnerLoop =
+            ck_tile::max(KPerThread / NumMacClusters, WarpGemm::kKPerThread);
+        constexpr index_t KIterInterwave = KPerInnerLoop / WarpGemm::kKPerThread;
+
+        using KIterSeq = std::conditional_t<Scheduler == GemmPipelineScheduler::Interwave,
+                                            sequence<KIterInterwave>,
+                                            sequence<KIterPerWarp>>;
+
+        constexpr auto b_block_outer_dstr_encoding =
+            tile_distribution_encoding<sequence<MWarp>,
+                                       tuple<sequence<NIterPerWarp, NWarp>, KIterSeq>,
+                                       tuple<sequence<0, 1>>,
+                                       tuple<sequence<0, 1>>,
+                                       sequence<1, 2>,
+                                       sequence<0, 0>>{};
+        constexpr auto b_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            b_block_outer_dstr_encoding, typename WarpGemm::BWarpDstrEncoding{});
+
+        return b_block_dstr_encode;
+    }
+
+    template <GemmPipelineScheduler Scheduler, typename GemmTraits>
+    struct BlockGemmImpl
+    {
+    };
+
+    template <typename GemmTraits>
+    struct BlockGemmImpl<GemmPipelineScheduler::Intrawave, GemmTraits>
+    {
+        static constexpr auto ALdsTileDistr =
+            decltype(make_static_tile_distribution(MakeABlockDistributionEncode())){};
+        static constexpr auto BLdsTileDistr =
+            decltype(make_static_tile_distribution(MakeBBlockDistributionEncode())){};
+
+        using ALdsTile = decltype(make_static_distributed_tensor<ATypeToUse>(ALdsTileDistr));
+        using BLdsTile = decltype(make_static_distributed_tensor<BTypeToUse>(BLdsTileDistr));
+
+        ALdsTile a_warp_tile_;
+        BLdsTile b_warp_tile_;
+
+        template <typename ASmemBlockWindow,
+                  typename BSmemBlockWindow,
+                  bool ALoadTranspose = false,
+                  bool BLoadTranspose = false>
+        CK_TILE_DEVICE void LocalPrefetch(const ASmemBlockWindow& a_block_window,
+                                          const BSmemBlockWindow& b_block_window,
+                                          bool_constant<ALoadTranspose> = {},
+                                          bool_constant<BLoadTranspose> = {})
+        {
+            load_and_convert_tile<UnaryOpSize_, ALoadTranspose>(a_warp_tile_, a_block_window);
+            load_and_convert_tile<UnaryOpSize_, BLoadTranspose>(b_warp_tile_, b_block_window);
+        }
+
+        // C += A * B
+        template <typename CBlockTensor,
+                  typename ASmemBlockWindow,
+                  typename BSmemBlockWindow,
+                  bool ALoadTranspose = false,
+                  bool BLoadTranspose = false>
+        CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
+                                       const ASmemBlockWindow&,
+                                       const BSmemBlockWindow&,
+                                       bool_constant<ALoadTranspose> = {},
+                                       bool_constant<BLoadTranspose> = {})
+        {
+            static_assert(std::is_same_v<CDataType, typename CBlockTensor::DataType>,
+                          "The CDataType as defined in traits should be the same as correspoinding "
+                          "C block tensor data type!");
+
+            // hot loop:
+            static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+                static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                    // read A warp tensor from A block tensor
+                    AWarpTensor a_warp_tensor;
+
+                    a_warp_tensor.get_thread_buffer() = a_warp_tile_.get_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, kIter>{}, a_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
+
+                    static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                        // read B warp tensor from B block tensor
+                        BWarpTensor b_warp_tensor;
+
+                        b_warp_tensor.get_thread_buffer() = b_warp_tile_.get_y_sliced_thread_data(
+                            merge_sequences(sequence<nIter, kIter>{}, b_warp_y_index_zeros),
+                            merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
+
+                        // read C warp tensor from C block tensor
+                        CWarpTensor c_warp_tensor;
+
+                        c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
+                            merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                            merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
+
+                        // warp GEMM
+                        WarpGemm{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
+
+                        // write C warp tensor into C block tensor
+                        c_block_tensor.set_y_sliced_thread_data(
+                            merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                            merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
+                            c_warp_tensor.get_thread_buffer());
+                    });
+                });
+            });
+        }
+    };
+
+    template <typename GemmTraits>
+    struct BlockGemmImpl<GemmPipelineScheduler::Interwave, GemmTraits>
+    {
+        static constexpr index_t KPerThread     = GemmTraits::KPerThread;
+        static constexpr index_t NumMacClusters = GemmTraits::InterWaveSchedulingMacClusters;
+        static constexpr index_t KPerInnerLoop =
+            ck_tile::max(KPerThread / NumMacClusters, WarpGemm::kKPerThread);
+        static constexpr index_t KRepeat        = KPerThread / KPerInnerLoop;
+        static constexpr index_t KInnerLoopIter = KPerInnerLoop / WarpGemm::kKPerThread;
+
+        static constexpr auto ALdsTileDistr =
+            decltype(make_static_tile_distribution(MakeABlockDistributionEncode())){};
+        static constexpr auto BLdsTileDistr =
+            decltype(make_static_tile_distribution(MakeBBlockDistributionEncode())){};
+
+        using ALdsTile = decltype(make_static_distributed_tensor<ATypeToUse>(ALdsTileDistr));
+        using BLdsTile = decltype(make_static_distributed_tensor<BTypeToUse>(BLdsTileDistr));
+
+        ALdsTile a_warp_tile_;
+        BLdsTile b_warp_tile_;
+
+        template <index_t KIdx,
+                  typename ASmemBlockWindow,
+                  typename BSmemBlockWindow,
+                  bool ALoadTranspose = false,
+                  bool BLoadTranspose = false>
+        CK_TILE_DEVICE void LocalPrefetch(const ASmemBlockWindow& a_block_window,
+                                          const BSmemBlockWindow& b_block_window,
+                                          bool_constant<ALoadTranspose> = {},
+                                          bool_constant<BLoadTranspose> = {})
+        {
+            constexpr auto a_lds_load_distr = [&]() {
+                if constexpr(ALoadTranspose)
+                    return make_static_tile_distribution(typename InputTileDistributionTraits<
+                                                         decltype(MakeABlockDistributionEncode()),
+                                                         ADataType>::TransposedDstrEncode{});
+                else
+                    return make_static_tile_distribution(MakeABlockDistributionEncode());
+            }();
+            constexpr auto b_lds_load_distr = [&]() {
+                if constexpr(BLoadTranspose)
+                    return make_static_tile_distribution(typename InputTileDistributionTraits<
+                                                         decltype(MakeBBlockDistributionEncode()),
+                                                         BDataType>::TransposedDstrEncode{});
+                else
+                    return make_static_tile_distribution(MakeBBlockDistributionEncode());
+            }();
+            constexpr auto a_lds_shape = []() {
+                if constexpr(ALoadTranspose)
+                    return make_tuple(number<KPerInnerLoop>{}, number<GemmTraits::MPerBlock>{});
+                else
+                    return make_tuple(number<GemmTraits::MPerBlock>{}, number<KPerInnerLoop>{});
+            }();
+            constexpr auto b_lds_shape = []() {
+                if constexpr(BLoadTranspose)
+                    return make_tuple(number<KPerInnerLoop>{}, number<GemmTraits::NPerBlock>{});
+                else
+                    return make_tuple(number<GemmTraits::NPerBlock>{}, number<KPerInnerLoop>{});
+            }();
+            constexpr auto k_idx_offset = KIdx * KPerInnerLoop;
+            constexpr auto a_offset =
+                ALoadTranspose ? multi_index<2>{k_idx_offset, 0} : multi_index<2>{0, k_idx_offset};
+            constexpr auto b_offset =
+                BLoadTranspose ? multi_index<2>{k_idx_offset, 0} : multi_index<2>{0, k_idx_offset};
+
+            auto a_lds_gemm_window = make_tile_window(
+                a_block_window.get_bottom_tensor_view(), a_lds_shape, a_offset, a_lds_load_distr);
+            auto b_lds_gemm_window = make_tile_window(
+                b_block_window.get_bottom_tensor_view(), b_lds_shape, b_offset, b_lds_load_distr);
+
+            load_and_convert_tile<UnaryOpSize_, ALoadTranspose>(a_warp_tile_, a_lds_gemm_window);
+            load_and_convert_tile<UnaryOpSize_, BLoadTranspose>(b_warp_tile_, b_lds_gemm_window);
+        }
+
+        // C += A * B
+        template <typename CBlockTensor,
+                  typename ASmemBlockWindow,
+                  typename BSmemBlockWindow,
+                  bool ALoadTranspose = false,
+                  bool BLoadTranspose = false>
+        CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
+                                       const ASmemBlockWindow& a_block_window,
+                                       const BSmemBlockWindow& b_block_window,
+                                       bool_constant<ALoadTranspose> a_load_tr = {},
+                                       bool_constant<BLoadTranspose> b_load_tr = {})
+        {
+            static_assert(std::is_same_v<CDataType, typename CBlockTensor::DataType>,
+                          "The CDataType as defined in traits should be the same as correspoinding "
+                          "C block tensor data type!");
+
+            // hot loop:
+            static_for<0, KRepeat, 1>{}([&](auto kIter) {
+                LocalPrefetch<kIter.value>(a_block_window, b_block_window, a_load_tr, b_load_tr);
+                __builtin_amdgcn_sched_barrier(
+                    0); // Complete scheduling all pending instruction groups before this point
+
+                // NOTE: Synchronize threads in a workgroup at the start of each MAC
+                // cluster, but except the first, as we can shorten non-MAC cluster a bit
+                // and there's no observable negative impact. The desired effect is waves in
+                // a workgroup executing MAC in sync. This avoids some out-of-sync waves
+                // hijacking MAC resource from other workgroups and reducing the chance of
+                // latency hiding by waiting for the rest of the workgroup at the eventual
+                // sync point.
+                if constexpr(kIter.value != 0 || KRepeat == 1)
+                {
+                    // This pattern ensures:
+                    // At runtime: All waves synchronize (hardware barrier)
+                    // At compile-time: Instructions after the barrier don't get moved before it
+                    // (scheduling barrier)
+                    __builtin_amdgcn_s_barrier(); // Blocks execution until all waves (threads) in
+                                                  // the workgroup reach this point
+                    __builtin_amdgcn_sched_barrier(
+                        0); // Prevents instruction reordering across this boundary
+                }
+
+                static_for<0, KInnerLoopIter, 1>{}([&](auto kInnerIter) {
+                    static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                        // read A warp tensor from A block tensor
+                        AWarpTensor a_warp_tensor;
+
+                        a_warp_tensor.get_thread_buffer() = a_warp_tile_.get_y_sliced_thread_data(
+                            merge_sequences(sequence<mIter, kInnerIter>{}, a_warp_y_index_zeros),
+                            merge_sequences(sequence<1, 1>{}, a_warp_y_lengths));
+                        static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                            // read B warp tensor from B block tensor
+                            BWarpTensor b_warp_tensor;
+
+                            b_warp_tensor.get_thread_buffer() =
+                                b_warp_tile_.get_y_sliced_thread_data(
+                                    merge_sequences(sequence<nIter, kInnerIter>{},
+                                                    b_warp_y_index_zeros),
+                                    merge_sequences(sequence<1, 1>{}, b_warp_y_lengths));
+                            // read C warp tensor from C block tensor-
+                            CWarpTensor c_warp_tensor;
+
+                            c_warp_tensor.get_thread_buffer() =
+                                c_block_tensor.get_y_sliced_thread_data(
+                                    merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                                    merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
+
+                            // The block_sync_lds() here performs double duty:
+                            // A) safeguard against data hazard because barrier from
+                            // blockwise_gemm is moved here B) reduce VMEM FIFO congestion
+                            // by applying small delays to different wavefronts It is
+                            // performed near the end of MAC cluster to minimize lgkmcnt
+                            // penalty
+                            if constexpr(kIter.value == KRepeat - 1 &&
+                                         kInnerIter.value == KInnerLoopIter - 1 &&
+                                         mIter.value == MIterPerWarp - 1 &&
+                                         nIter.value == NIterPerWarp - 1)
+                            {
+                                __builtin_amdgcn_sched_barrier(0);
+                                block_sync_lds();
+                                __builtin_amdgcn_sched_barrier(0);
+                            }
+                            // warp GEMM
+                            WarpGemm{}(c_warp_tensor, a_warp_tensor, b_warp_tensor);
+
+                            // write C warp tensor into C block tensor
+                            c_block_tensor.set_y_sliced_thread_data(
+                                merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                                merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
+                                c_warp_tensor.get_thread_buffer());
+
+                            if constexpr(kInnerIter.value == 0 && mIter.value == 0 &&
+                                         nIter.value == 0)
+                            {
+                                __builtin_amdgcn_sched_barrier(0);
+                                __builtin_amdgcn_s_setprio(1);
+                                __builtin_amdgcn_sched_barrier(0);
+                            }
+                        });
+                    });
+                });
+
+                __builtin_amdgcn_sched_barrier(0);
+                __builtin_amdgcn_s_setprio(0);
+                __builtin_amdgcn_sched_barrier(0);
+            });
+        }
+    };
+
+    public:
+    CK_TILE_DEVICE static constexpr auto MakeCBlockTile()
+    {
+        constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+            sequence<>,
+            tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
+            tuple<sequence<1, 2>>,
+            tuple<sequence<1, 1>>,
+            sequence<1, 2>,
+            sequence<0, 0>>{};
+
+        constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            c_block_outer_dstr_encoding, typename WarpGemm::CWarpDstrEncoding{});
+        constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
+        auto c_block_tensor         = make_static_distributed_tensor<CDataType>(c_block_dstr);
+
+        return c_block_tensor;
+    }
+
+    template <typename ASmemBlockWindow,
+              typename BSmemBlockWindow,
+              bool ALoadTranspose = false,
+              bool BLoadTranspose = false>
+    CK_TILE_DEVICE void LocalPrefetch(const ASmemBlockWindow& a_block_window,
+                                      const BSmemBlockWindow& b_block_window,
+                                      bool_constant<ALoadTranspose> a_load_tr = {},
+                                      bool_constant<BLoadTranspose> b_load_tr = {})
+    {
+        block_gemm_impl_.LocalPrefetch(a_block_window, b_block_window, a_load_tr, b_load_tr);
+    }
+
+    // C += A * B
+    template <typename CBlockTensor,
+              typename ASmemBlockWindow,
+              typename BSmemBlockWindow,
+              bool ALoadTranspose = false,
+              bool BLoadTranspose = false>
+    CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
+                                   const ASmemBlockWindow& a_block_window,
+                                   const BSmemBlockWindow& b_block_window,
+                                   bool_constant<ALoadTranspose> a_load_tr = {},
+                                   bool_constant<BLoadTranspose> b_load_tr = {})
+    {
+        block_gemm_impl_(c_block_tensor, a_block_window, b_block_window, a_load_tr, b_load_tr);
+    }
+
+    // C = A * B
+    template <typename ASmemBlockWindow,
+              typename BSmemBlockWindow,
+              bool ALoadTranspose = false,
+              bool BLoadTranspose = false>
+    CK_TILE_DEVICE auto operator()(const ASmemBlockWindow& a_block_window,
+                                   const BSmemBlockWindow& b_block_window,
+                                   bool_constant<ALoadTranspose> a_load_tr = {},
+                                   bool_constant<BLoadTranspose> b_load_tr = {})
+    {
+        auto c_block_tensor = MakeCBlockTile();
+        block_gemm_impl_(c_block_tensor, a_block_window, b_block_window, a_load_tr, b_load_tr);
+        return c_block_tensor;
+    }
+
+    private:
+    BlockGemmImpl<Scheduler, Traits> block_gemm_impl_{};
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_wp_asmem_breg_creg.hpp

@@ -0,0 +1,212 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/block/block_wp_asmem_bsmem_creg_v1_custom_policy.hpp"
+
+namespace ck_tile {
+
+// A is block window on shared memory
+// B is block window on register
+// C is block distributed tensor
+template <typename Problem_, typename BlockPolicy_>
+struct BlockWeightPreshuffleASmemBRegCReg
+{
+    using Problem        = remove_cvref_t<Problem_>;
+    using BlockPolicy    = remove_cvref_t<BlockPolicy_>;
+    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>;
+
+    static constexpr auto I0   = number<0>();
+    static constexpr auto I1   = number<1>();
+    static constexpr auto I2   = number<2>();
+    static constexpr auto idxM = I0;
+    static constexpr auto idxN = I1;
+    static constexpr auto idxK = I2;
+    using BlockTile            = remove_cvref_t<typename BlockGemmShape::BlockTile>;
+    using BlockWarps           = remove_cvref_t<typename BlockGemmShape::BlockWarps>;
+    using WarpTile             = remove_cvref_t<typename BlockGemmShape::WarpTile>;
+
+    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 kBlockSize = Problem::kBlockSize;
+
+    static constexpr auto config = BlockPolicy::template GetWarpGemmMWarpNWarp<Problem>();
+    using WarpGemm               = remove_cvref_t<decltype(config.template at<0>())>;
+
+    static constexpr index_t MWarp = config.template at<1>();
+    static constexpr index_t NWarp = config.template at<2>();
+
+    static constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WarpGemm::kM);
+    static constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WarpGemm::kN);
+    static constexpr index_t KIterPerWarp = KPerBlock / WarpGemm::kK;
+
+    static constexpr index_t MPerBlockPerIter = MWarp * WarpGemm::kM;
+    static constexpr index_t KPerBlockPerIter = WarpGemm::kK;
+
+    static constexpr index_t DsReadPreload = 2; // default 2, preload 2 ds read
+
+    static constexpr index_t m_preload = (MIterPerWarp * KIterPerWarp >= DsReadPreload)
+                                             ? DsReadPreload
+                                             : MIterPerWarp * KIterPerWarp;
+
+    using AWarpTensor = typename WarpGemm::AWarpTensor;
+    statically_indexed_array<AWarpTensor, m_preload> preloaded_a_warp_tensor;
+
+    CK_TILE_DEVICE static constexpr auto MakeABlockDistributionEncode()
+    {
+        constexpr auto a_block_outer_dstr_encoding =
+            tile_distribution_encoding<sequence<NWarp>,
+                                       tuple<sequence<1, MWarp>, sequence<1>>,
+                                       tuple<sequence<1, 0>>,
+                                       tuple<sequence<1, 0>>,
+                                       sequence<1, 2>,
+                                       sequence<0, 0>>{};
+        constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            a_block_outer_dstr_encoding, typename WarpGemm::AWarpDstrEncoding{});
+
+        return a_block_dstr_encode;
+    }
+
+    template <typename SmemBlockWindow>
+    CK_TILE_DEVICE auto MakeALoadWindows(SmemBlockWindow& a_block_window) const
+    {
+        constexpr auto a_load_dstr = make_static_tile_distribution(MakeABlockDistributionEncode());
+
+        // create MIterPerWarp × KIterPerWarp window
+        return generate_tuple(
+            [&](auto kIter) {
+                return generate_tuple(
+                    [&](auto mIter) {
+                        return make_tile_window(
+                            get_slice_tile(
+                                a_block_window,
+                                sequence<mIter * MPerBlockPerIter, kIter * KPerBlockPerIter>{},
+                                sequence<(mIter + 1) * MPerBlockPerIter,
+                                         (kIter + 1) * KPerBlockPerIter>{}),
+                            a_load_dstr);
+                    },
+                    number<MIterPerWarp>{});
+            },
+            number<KIterPerWarp>{});
+    }
+
+    template <typename ALoadWindows>
+    CK_TILE_DEVICE void LocalPrefetch(const ALoadWindows& a_load_windows)
+    {
+
+        static_for<0, m_preload, 1>{}([&](auto loadIter) {
+            constexpr auto mIter = loadIter % MIterPerWarp;
+            constexpr auto kIter = loadIter / MIterPerWarp;
+
+            load_tile(preloaded_a_warp_tensor(loadIter),
+                      a_load_windows[number<kIter>{}][number<mIter>{}]);
+        });
+    }
+
+    CK_TILE_DEVICE static constexpr auto MakeCBlockTile()
+    {
+        constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+            sequence<>,
+            tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
+            tuple<sequence<1, 2>>,
+            tuple<sequence<1, 1>>,
+            sequence<1, 2>,
+            sequence<0, 0>>{};
+
+        constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            c_block_outer_dstr_encoding, typename WarpGemm::CWarpDstrEncoding{});
+
+        constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
+
+        auto c_block_tensor = make_static_distributed_tensor<CDataType>(c_block_dstr);
+        return c_block_tensor;
+    }
+
+    // C += A * B
+    template <typename CBlockTensor,
+              typename ALoadWindows,
+              typename BFlatBlockTensor,
+              typename BFlatDistribution>
+    CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
+                                   const ALoadWindows& a_load_windows,
+                                   BFlatBlockTensor& b_block_tensor,
+                                   const BFlatDistribution&)
+    {
+        constexpr auto MIter_2nd_last = (MIterPerWarp >= 2) ? MIterPerWarp - 2 : MIterPerWarp - 1;
+
+        using CWarpDstr   = typename WarpGemm::CWarpDstr;
+        using CWarpTensor = typename WarpGemm::CWarpTensor;
+
+        using BWarpTensor = typename WarpGemm::BWarpTensor;
+
+        constexpr auto b_block_y_lengths =
+            to_sequence(BFlatDistribution{}.get_ys_to_d_descriptor().get_lengths());
+
+        constexpr auto c_warp_y_lengths =
+            to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+
+        constexpr auto b_block_y_index_zeros =
+            uniform_sequence_gen_t<BFlatDistribution::NDimY, 0>{};
+        constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
+
+        static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+            static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                constexpr auto AwarpIter = (kIter * MIterPerWarp + mIter) % m_preload;
+                static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                    // read C warp tensor from C block tensor
+                    BWarpTensor b_warp_tensor;
+                    CWarpTensor c_warp_tensor;
+
+                    b_warp_tensor.get_thread_buffer() = b_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(sequence<nIter, kIter>{},
+                                        typename sequence_split<decltype(b_block_y_index_zeros),
+                                                                2>::right_type{}),
+                        merge_sequences(
+                            sequence<1, 1>{},
+                            typename sequence_split<decltype(b_block_y_lengths), 2>::right_type{}));
+
+                    c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
+
+                    // warp GEMM
+                    WarpGemm{}(
+                        c_warp_tensor, preloaded_a_warp_tensor(number<AwarpIter>{}), b_warp_tensor);
+
+                    // write C warp tensor into C block tensor
+                    c_block_tensor.set_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
+                        c_warp_tensor.get_thread_buffer());
+
+                    __builtin_amdgcn_sched_barrier(0x7F6);
+                });
+                // preload next A from lds
+                if constexpr((kIter * MIterPerWarp + mIter) <
+                             (KIterPerWarp * MIterPerWarp - m_preload))
+                {
+                    constexpr auto AmIter = (mIter + m_preload) % MIterPerWarp;
+                    constexpr auto AkIter = (kIter + (mIter + m_preload) / MIterPerWarp);
+
+                    load_tile(preloaded_a_warp_tensor(number<AwarpIter>{}),
+                              a_load_windows[number<AkIter>{}][number<AmIter>{}]);
+                }
+
+                // barrier
+                if constexpr((kIter == KIterPerWarp - 1) && (mIter == MIter_2nd_last))
+                {
+                    block_sync_lds();
+                }
+            });
+        });
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_wp_asmem_bsmem_creg_v1.hpp

@@ -0,0 +1,118 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/block/block_wp_asmem_bsmem_creg_v1_custom_policy.hpp"
+
+namespace ck_tile {
+
+// A is block window on shared memory
+// B is block window on shared memory
+// C is block distributed tensor
+template <typename Problem_, typename BlockPolicy_>
+struct BlockWeightPreshuffleASmemBSmemCRegV1
+{
+    using Problem        = remove_cvref_t<Problem_>;
+    using BlockPolicy    = remove_cvref_t<BlockPolicy_>;
+    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>;
+
+    static constexpr auto I0   = number<0>();
+    static constexpr auto I1   = number<1>();
+    static constexpr auto I2   = number<2>();
+    static constexpr auto idxM = I0;
+    static constexpr auto idxN = I1;
+    static constexpr auto idxK = I2;
+    using BlockTile            = remove_cvref_t<typename BlockGemmShape::BlockTile>;
+    using BlockWarps           = remove_cvref_t<typename BlockGemmShape::BlockWarps>;
+    using WarpTile             = remove_cvref_t<typename BlockGemmShape::WarpTile>;
+
+    static constexpr index_t kBlockSize = Problem::kBlockSize;
+
+    static constexpr auto config = BlockPolicy::template GetWarpGemmMWarpNWarp<Problem>();
+    using WG                     = remove_cvref_t<decltype(config.template at<0>())>;
+
+    CK_TILE_DEVICE static constexpr auto MakeCBlockTile()
+    {
+        constexpr index_t MPerBlock = BlockGemmShape::kM;
+        constexpr index_t NPerBlock = BlockGemmShape::kN;
+
+        constexpr index_t MWarp = config.template at<1>();
+        constexpr index_t NWarp = config.template at<2>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t NIterPerWarp = NPerBlock / (NWarp * WG::kN);
+
+        constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding<
+            sequence<>,
+            tuple<sequence<MIterPerWarp, MWarp>, sequence<NIterPerWarp, NWarp>>,
+            tuple<sequence<1, 2>>,
+            tuple<sequence<1, 1>>,
+            sequence<1, 2>,
+            sequence<0, 0>>{};
+
+        constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding(
+            c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{});
+
+        constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode);
+
+        auto c_block_tensor = make_static_distributed_tensor<CDataType>(c_block_dstr);
+        return c_block_tensor;
+    }
+
+    // C += A * B
+    template <typename CBlockTensor, typename ABlockWindow, typename BFlatBlockTensor>
+    CK_TILE_DEVICE void operator()(CBlockTensor& c_block_tensor,
+                                   ABlockWindow& a_warp_windows,
+                                   BFlatBlockTensor& b_warp_tensor) const
+    {
+        constexpr index_t MPerBlock = BlockGemmShape::kM;
+        constexpr index_t KPerBlock = BlockGemmShape::kK;
+
+        constexpr index_t MWarp = config.template at<1>();
+
+        constexpr index_t MIterPerWarp = MPerBlock / (MWarp * WG::kM);
+        constexpr index_t NIterPerWarp =
+            BlockTile::at(idxN) / (WarpTile::at(idxN) * BlockWarps::at(idxN));
+        constexpr index_t KIterPerWarp = KPerBlock / WG::kK;
+
+        using CWarpDstr   = typename WG::CWarpDstr;
+        using CWarpTensor = typename WG::CWarpTensor;
+
+        constexpr auto c_warp_y_lengths =
+            to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths());
+        constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t<CWarpDstr::NDimY, 0>{};
+
+        // hot loop:
+        static_for<0, KIterPerWarp, 1>{}([&](auto kIter) {
+            static_for<0, MIterPerWarp, 1>{}([&](auto mIter) {
+                // read A warp tensor from A block window
+                const auto a_warp_tensor = load_tile(a_warp_windows(mIter)(kIter));
+
+                static_for<0, NIterPerWarp, 1>{}([&](auto nIter) {
+                    // read C warp tensor from C block tensor
+                    CWarpTensor c_warp_tensor;
+
+                    c_warp_tensor.get_thread_buffer() = c_block_tensor.get_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths));
+
+                    // warp GEMM
+                    WG{}(c_warp_tensor, a_warp_tensor, b_warp_tensor(nIter)(kIter));
+
+                    // write C warp tensor into C block tensor
+                    c_block_tensor.set_y_sliced_thread_data(
+                        merge_sequences(sequence<mIter, nIter>{}, c_warp_y_index_zeros),
+                        merge_sequences(sequence<1, 1>{}, c_warp_y_lengths),
+                        c_warp_tensor.get_thread_buffer());
+                });
+            });
+        });
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/gemm/block/block_wp_asmem_bsmem_creg_v1_custom_policy.hpp

@@ -0,0 +1,38 @@
+// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+
+namespace ck_tile {
+
+// Default policy for BlockGemmASmemBSmemCRegV1
+// Default policy class should not be templated, put template on member functions instead
+template <typename AType_,
+          typename BType_,
+          typename CType_,
+          typename BlockWarps_,
+          typename WarpGemm_>
+struct BlockWeightPreshuffleASmemBSmemCRegV1CustomPolicy
+{
+    using AType = remove_cvref_t<AType_>;
+    using BType = remove_cvref_t<BType_>;
+    using CType = remove_cvref_t<CType_>;
+
+    using BlockWarps = remove_cvref_t<BlockWarps_>;
+
+    static constexpr index_t kMWarps = BlockWarps::at(number<0>{});
+    static constexpr index_t kNWarps = BlockWarps::at(number<1>{});
+    static constexpr index_t kKWarps = BlockWarps::at(number<2>{});
+
+    using WarpGemm = remove_cvref_t<WarpGemm_>;
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemmMWarpNWarp()
+    {
+        return make_tuple(WarpGemm{}, kMWarps, kNWarps);
+    }
+};
+
+} // namespace ck_tile