Merge branch 'develop' into moe_cross_reduce

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_b_scale.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
+
+namespace ck {
+
+// Naive pipeline with lowest resource request per WGP
+// GlobalPrefetchStages: 1
+// LocalPreFillStages: 1
+// LocalPreFetchStages: 0
+// LocalSharedMemoryBuffer: 1
+
+template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
+          index_t BlockSize,
+          typename ADataType,
+          typename BDataType,
+          typename ComputeDataType,
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPacks>
+struct BlockwiseGemmXdlops_pipeline_v1_b_scale
+{
+};
+
+template <index_t BlockSize,
+          typename ADataType,
+          typename BDataType,
+          typename ComputeDataType,
+          typename AccDataType,
+          typename ATileDesc,
+          typename BTileDesc,
+          typename AMmaTileDesc,
+          typename BMmaTileDesc,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MRepeat,
+          index_t NRepeat,
+          index_t KPack
+          // ,bool TransposeC //disable transposec right now...
+          >
+struct BlockwiseGemmXdlops_pipeline_v1_b_scale<BlockGemmPipelineScheduler::Intrawave,
+                                               BlockSize,
+                                               ADataType,
+                                               BDataType,
+                                               ComputeDataType,
+                                               AccDataType,
+                                               ATileDesc,
+                                               BTileDesc,
+                                               AMmaTileDesc,
+                                               BMmaTileDesc,
+                                               ABlockTransferSrcScalarPerVector,
+                                               BBlockTransferSrcScalarPerVector,
+                                               MPerBlock,
+                                               NPerBlock,
+                                               KPerBlock,
+                                               MPerXDL,
+                                               NPerXDL,
+                                               MRepeat,
+                                               NRepeat,
+                                               KPack>
+    : BlockwiseGemmXdlops_pipeline_base<BlockSize,
+                                        ADataType,
+                                        BDataType,
+                                        ComputeDataType,
+                                        AccDataType,
+                                        ATileDesc,
+                                        BTileDesc,
+                                        AMmaTileDesc,
+                                        BMmaTileDesc,
+                                        ABlockTransferSrcScalarPerVector,
+                                        BBlockTransferSrcScalarPerVector,
+                                        MPerBlock,
+                                        NPerBlock,
+                                        KPerBlock,
+                                        MPerXDL,
+                                        NPerXDL,
+                                        MRepeat,
+                                        NRepeat,
+                                        KPack>
+
+{
+    using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
+                                                   ADataType,
+                                                   BDataType,
+                                                   ComputeDataType,
+                                                   AccDataType,
+                                                   ATileDesc,
+                                                   BTileDesc,
+                                                   AMmaTileDesc,
+                                                   BMmaTileDesc,
+                                                   ABlockTransferSrcScalarPerVector,
+                                                   BBlockTransferSrcScalarPerVector,
+                                                   MPerBlock,
+                                                   NPerBlock,
+                                                   KPerBlock,
+                                                   MPerXDL,
+                                                   NPerXDL,
+                                                   MRepeat,
+                                                   NRepeat,
+                                                   KPack>;
+    using Base::I0;
+    using Base::KRepeat;
+    using Base::xdlops_gemm;
+
+    using Base::CalculateCThreadOriginDataIndex;
+    using Base::CalculateCThreadOriginDataIndex8D;
+    using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::GetCThreadBuffer;
+    using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
+    using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
+    using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
+
+    using Base::a_block_desc_m0_m1_m2_k;
+    using Base::b_block_desc_n0_n1_n2_k;
+
+    using Base::AMmaKStride;
+    using Base::BMmaKStride;
+
+    static constexpr index_t PrefetchStages  = 1;
+    static constexpr index_t PrefillStages   = 1;
+    static constexpr index_t GlobalBufferNum = 1;
+
+    __host__ static constexpr bool BlockHasHotloop(index_t num_loop)
+    {
+        return num_loop > PrefetchStages;
+    }
+
+    __host__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
+    {
+        ignore = num_loop;
+        return TailNumber::Full;
+    }
+
+    template <bool HasMainLoop,
+              TailNumber TailNum,
+              typename AGridDesc,
+              typename ABlockDesc,
+              typename ABlockTransfer,
+              typename AGridBuffer,
+              typename ABlockBuffer,
+              typename ABlockTransferStep,
+              typename BGridDesc,
+              typename BBlockDesc,
+              typename BBlockTransfer,
+              typename BGridBuffer,
+              typename BBlockBuffer,
+              typename BBlockTransferStep,
+              typename CThreadBuffer,
+              // BScale Thread Copy
+              typename BScaleGridBuffer,
+              typename BScaleGridDesc,
+              typename BScaleThreadDesc,
+              typename BScaleThreadTransfer,
+              typename BScaleThreadTransferStep>
+    __device__ void Run(
+        // ABlockCopy
+        const AGridDesc& a_grid_desc,
+        const ABlockDesc& a_block_desc,
+        ABlockTransfer& a_blockwise_copy,
+        const AGridBuffer& a_grid_buf,
+        ABlockBuffer& a_block_buf,
+        const ABlockTransferStep& a_block_copy_step,
+        // BBlockCopy
+        const BGridDesc& b_grid_desc,
+        const BBlockDesc& b_block_desc,
+        BBlockTransfer& b_blockwise_copy,
+        const BGridBuffer& b_grid_buf,
+        BBlockBuffer& b_block_buf,
+        const BBlockTransferStep& b_block_copy_step,
+        // CThread
+        CThreadBuffer& c_thread_buf,
+        // BScaleThreadCopy
+        const BScaleGridDesc& b_scale_grid_desc,
+        const BScaleThreadDesc& b_scale_thread_desc,
+        BScaleThreadTransfer& b_scale_thread_copy,
+        const BScaleGridBuffer& b_scale_grid_buf,
+        const BScaleThreadTransferStep& b_scale_thread_copy_step,
+        // num_loop
+        index_t num_loop,
+        index_t num_loop_per_scale) const
+    {
+        // assume kperblock = scaleblockk
+        ignore            = num_loop_per_scale;
+        auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            a_thread_desc_.GetElementSpaceSize());
+        auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            b_thread_desc_.GetElementSpaceSize());
+
+        auto b_scale_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
+            b_scale_thread_desc.GetElementSpaceSize());
+
+        // Global prefetch 1
+        a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+        b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+        a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+        b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+        static_for<0, NRepeat, 1>{}([&](auto n0) {
+            b_scale_thread_copy.Run(b_scale_grid_desc,
+                                    b_scale_grid_buf,
+                                    b_scale_thread_desc,
+                                    make_tuple(n0, I0),
+                                    b_scale_thread_buf);
+
+            b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
+                                                   b_scale_thread_copy_step.At(Number<0>{}));
+        });
+        b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
+                                               b_scale_thread_copy_step.At(Number<1>{}));
+
+        // Local prefill 1
+        a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+        b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+        // Initialize C
+        c_thread_buf.Clear();
+
+        auto c_thread_buf_per_scale = remove_cvref_t<decltype(c_thread_buf)>();
+
+        // main body
+        if constexpr(HasMainLoop)
+        {
+            index_t i = 0;
+            do
+            {
+                // -------------------------------------------------------------------------------------------
+                a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
+                b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
+
+                a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
+                b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
+
+                block_sync_lds();
+                static_for<0, KRepeat, 1>{}([&](auto k) {
+                    static_for<0, MRepeat, 1>{}([&](auto m0) {
+                        a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                           make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                           a_block_buf,
+                                           a_thread_desc_,
+                                           make_tuple(m0, I0, k, I0),
+                                           a_thread_buf);
+                    });
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                           make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                           b_block_buf,
+                                           b_thread_desc_,
+                                           make_tuple(n0, I0, k, I0),
+                                           b_thread_buf);
+                    });
+                });
+
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    static_for<0, NRepeat, 1>{}([&](auto n0) {
+                        c_thread_buf_per_scale.Clear();
+                        static_for<0, KRepeat, 1>{}([&](auto k0) {
+                            vector_type<ComputeDataType, KPack> a_thread_vec;
+                            vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                            static_for<0, KPack, 1>{}([&](auto ik) {
+                                a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                    a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                        make_tuple(m0, I0, k0, ik))>{}];
+                                b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                    b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                        make_tuple(n0, I0, k0, ik))>{}];
+                            });
+
+                            using mfma_input_type =
+                                typename vector_type<ComputeDataType,
+                                                     xdlops_gemm.K1PerXdlops>::type;
+
+                            xdlops_gemm.template Run<>(
+                                a_thread_vec.template AsType<mfma_input_type>(),
+                                b_thread_vec.template AsType<mfma_input_type>(),
+                                c_thread_buf_per_scale.GetVectorTypeReference(I0));
+                        });
+                        static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
+                            constexpr index_t c_offset =
+                                c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
+                            c_thread_buf(Number<c_offset>{}) +=
+                                c_thread_buf_per_scale[Number<t>{}] *
+                                type_convert<AccDataType>(b_scale_thread_buf[n0]);
+                        });
+                    });
+                });
+
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    b_scale_thread_copy.Run(b_scale_grid_desc,
+                                            b_scale_grid_buf,
+                                            b_scale_thread_desc,
+                                            make_tuple(n0, I0),
+                                            b_scale_thread_buf);
+
+                    b_scale_thread_copy.MoveSrcSliceWindow(
+                        b_scale_grid_desc, b_scale_thread_copy_step.At(Number<0>{}));
+                });
+
+                b_scale_thread_copy.MoveSrcSliceWindow(b_scale_grid_desc,
+                                                       b_scale_thread_copy_step.At(Number<1>{}));
+
+                block_sync_lds();
+                a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
+                b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
+
+                i += 1;
+
+            } while(i < (num_loop - 1));
+        }
+
+        // tail
+        if constexpr(TailNum == TailNumber::Full)
+        {
+            block_sync_lds();
+            static_for<0, KRepeat, 1>{}([&](auto k) {
+                static_for<0, MRepeat, 1>{}([&](auto m0) {
+                    a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
+                                       make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
+                                       a_block_buf,
+                                       a_thread_desc_,
+                                       make_tuple(m0, I0, k, I0),
+                                       a_thread_buf);
+                });
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
+                                       make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
+                                       b_block_buf,
+                                       b_thread_desc_,
+                                       make_tuple(n0, I0, k, I0),
+                                       b_thread_buf);
+                });
+            });
+
+            static_for<0, MRepeat, 1>{}([&](auto m0) {
+                static_for<0, NRepeat, 1>{}([&](auto n0) {
+                    c_thread_buf_per_scale.Clear();
+                    static_for<0, KRepeat, 1>{}([&](auto k0) {
+                        vector_type<ComputeDataType, KPack> a_thread_vec;
+                        vector_type<ComputeDataType, KPack> b_thread_vec;
+
+                        static_for<0, KPack, 1>{}([&](auto ik) {
+                            a_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                a_thread_buf[Number<a_thread_desc_.CalculateOffset(
+                                    make_tuple(m0, I0, k0, ik))>{}];
+                            b_thread_vec.template AsType<ComputeDataType>()(ik) =
+                                b_thread_buf[Number<b_thread_desc_.CalculateOffset(
+                                    make_tuple(n0, I0, k0, ik))>{}];
+                        });
+
+                        using mfma_input_type =
+                            typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
+
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf_per_scale.GetVectorTypeReference(I0));
+                    });
+                    static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
+                        constexpr index_t c_offset =
+                            c_thread_desc_.CalculateOffset(make_tuple(m0, n0, t));
+                        c_thread_buf(Number<c_offset>{}) +=
+                            c_thread_buf_per_scale[Number<t>{}] *
+                            type_convert<AccDataType>(b_scale_thread_buf[n0]);
+                    });
+                });
+            });
+        }
+    }
+
+    protected:
+    using Base::a_thread_copy_;
+    using Base::a_thread_desc_;
+    using Base::b_thread_copy_;
+    using Base::b_thread_desc_;
+    using Base::c_thread_desc_;
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/device/device_gemm_v2.hpp

@@ -36,6 +36,10 @@ struct DeviceGemmV2 : public BaseOperator
                         CElementwiseOperation c_element_op) = 0;
 
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+
+    virtual bool GetPermuteA()         = 0;
+    virtual bool GetPermuteB()         = 0;
+    virtual ck::index_t GetKPerBlock() = 0;
 };
 
 template <typename ALayout,
@@ -73,6 +77,43 @@ struct DeviceGemmV2R1 : public BaseOperator
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };
 
+template <typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename ADataType,
+          typename BDataType,
+          typename BScaleType,
+          typename CDataType,
+          index_t ScaleBlockN,
+          index_t ScaleBlockK,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+struct DeviceGemmV2BScale : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_a,
+                        const void* p_b,
+                        void* p_c,
+                        ck::index_t M,
+                        ck::index_t N,
+                        ck::index_t K,
+                        ck::index_t StrideA,
+                        ck::index_t StrideB,
+                        ck::index_t StrideC,
+                        ck::index_t StrideScaleB,
+                        const void* p_b_scale,
+                        ck::index_t KSplit,
+                        AElementwiseOperation a_element_op,
+                        BElementwiseOperation b_element_op,
+                        CElementwiseOperation c_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+
+    virtual bool GetPermuteB()         = 0;
+    virtual ck::index_t GetKPerBlock() = 0;
+};
+
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_streamk_v3.hpp

@@ -469,7 +469,11 @@ struct DeviceGemm_Xdl_CShuffle_Streamk_V3 : public DeviceGemm_Streamk_V2<ALayout
         {
             return false;
         }
-
+        if(!is_bf16_atomic_supported() && std::is_same_v<CDataType, ck::bhalf_t> &&
+           arg.Streamk_sel > 0)
+        {
+            return false;
+        }
         if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
                                                        GemmSpec == GemmSpecialization::NKPadding ||
                                                        GemmSpec == GemmSpecialization::MNKPadding ||

include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp

@@ -64,7 +64,9 @@ template <typename ALayout,
           BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
           BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
           typename ComputeTypeA                       = CDataType,
-          typename ComputeTypeB                       = ComputeTypeA>
+          typename ComputeTypeB                       = ComputeTypeA,
+          bool PermuteA                               = false,
+          bool PermuteB                               = false>
 struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
                                                        BLayout,
                                                        CLayout,
@@ -122,7 +124,9 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
         BlkGemmPipeSched,
         BlkGemmPipelineVer,
         ComputeTypeA,
-        ComputeTypeB>;
+        ComputeTypeB,
+        PermuteA,
+        PermuteB>;
 
     using Argument = typename GridwiseGemm::Argument;
 
@@ -633,6 +637,11 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
         return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
     }
 
+    index_t GetKPerBlock() override { return KPerBlock; }
+
+    bool GetPermuteA() override { return PermuteA; }
+    bool GetPermuteB() override { return PermuteB; }
+
     static auto MakeArgument(const ADataType* p_a,
                              const BDataType* p_b,
                              CDataType* p_c,

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3.hpp

@@ -127,7 +127,9 @@ template <typename ALayout,
           BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
           BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v4,
           typename ComputeTypeA                       = CDataType,
-          typename ComputeTypeB                       = ComputeTypeA>
+          typename ComputeTypeB                       = ComputeTypeA,
+          bool PermuteA                               = false,
+          bool PermuteB                               = false>
 struct GridwiseGemm_xdl_cshuffle_v3
 {
     static constexpr auto I0 = Number<0>{};
@@ -151,6 +153,20 @@ struct GridwiseGemm_xdl_cshuffle_v3
 
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
 
+    static constexpr index_t APackedSize = []() {
+        if constexpr(is_same_v<remove_cvref_t<ADataType>, pk_i4_t>)
+            return 2;
+        else
+            return 1;
+    }();
+
+    static constexpr index_t BPackedSize = []() {
+        if constexpr(is_same_v<remove_cvref_t<BDataType>, pk_i4_t>)
+            return 2;
+        else
+            return 1;
+    }();
+
     __host__ static auto CalculateGridSize(index_t M, index_t N, index_t KBatch)
     {
         return std::make_tuple(Block2CTileMap::CalculateGridSize(M, N), 1, KBatch);
@@ -319,6 +335,10 @@ struct GridwiseGemm_xdl_cshuffle_v3
 
         using GemmSpecialization = tensor_operation::device::GemmSpecialization;
 
+        static_assert(!(is_same_v<remove_cvref_t<ADataType>, pk_i4_t> &&
+                        GemmSpec != GemmSpecialization::Default),
+                      "pk_i4_t does not support padding");
+
         if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
                      GemmSpec == GemmSpecialization::MNKPadding)
         {
@@ -372,6 +392,8 @@ struct GridwiseGemm_xdl_cshuffle_v3
             return b_grid_desc_bk0_n_bk1;
         }
         else
+        {
+            if constexpr(!PermuteB)
             {
                 // not pad N or K
                 const auto b_grid_desc_bk0_n_bk1 = transform_tensor_descriptor(
@@ -383,6 +405,28 @@ struct GridwiseGemm_xdl_cshuffle_v3
 
                 return b_grid_desc_bk0_n_bk1;
             }
+            else
+            {
+                // Pre-shuffled Weight
+                // BGlobal[K / KPerBlock, N, KPerBlock / K1, K1] -> BTile[K / K1, N, K1]
+                constexpr index_t BK01 = KPerBlock / BK1Value;
+                const index_t BK0_     = StrideB / BK1Value;
+                const index_t BK00     = BK0_ / BK01;
+
+                const auto b_grid_desc_bk00_n_bk01_bk1_permute =
+                    make_naive_tensor_descriptor_packed(make_tuple(BK00, N, BK01, BK1Value));
+
+                const auto b_grid_desc_bk0_n_bk1_permute = transform_tensor_descriptor(
+                    b_grid_desc_bk00_n_bk01_bk1_permute,
+                    make_tuple(make_merge_transform(make_tuple(BK00, BK01)),
+                               make_pass_through_transform(make_tuple(N)),
+                               make_pass_through_transform(BK1Value)),
+                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+                return b_grid_desc_bk0_n_bk1_permute;
+            }
+        }
     }
 
     template <typename ABlockDesc_AK0_M_AK1>
@@ -572,7 +616,7 @@ struct GridwiseGemm_xdl_cshuffle_v3
         {
             if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
             {
-                a_k_split_offset = blockIdx.z * karg.KRead;
+                a_k_split_offset = blockIdx.z * karg.KRead / APackedSize;
             }
             else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
             {
@@ -585,7 +629,15 @@ struct GridwiseGemm_xdl_cshuffle_v3
             }
             else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, BLayout>)
             {
-                b_k_split_offset = blockIdx.z * karg.KRead;
+                if constexpr(!PermuteB)
+                {
+                    b_k_split_offset = blockIdx.z * karg.KRead / BPackedSize;
+                }
+                else
+                {
+                    const int k0_offset = karg.KRead * karg.N;
+                    b_k_split_offset    = blockIdx.z * k0_offset / BPackedSize;
+                }
             }
 
             if(blockIdx.z < static_cast<uint32_t>(karg.KBatch - 1))
@@ -625,9 +677,8 @@ struct GridwiseGemm_xdl_cshuffle_v3
         // in some cases.
         else if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
         {
-            constexpr auto MLdsLayer        = 32 * 4 / KPerBlock / sizeof(ADataType) < 1
-                                                  ? 1
-                                                  : 32 * 4 / KPerBlock / sizeof(ADataType);
+            constexpr index_t LdsSize       = 32 * 4 / KPerBlock / sizeof(ADataType) / APackedSize;
+            constexpr auto MLdsLayer        = LdsSize < 1 ? 1 : LdsSize;
             constexpr auto a_lds_block_desc = make_naive_tensor_descriptor(
                 make_tuple(
                     AK0Number * Number<MLdsLayer>{}, Number<MPerBlock / MLdsLayer>{}, AK1Number),
@@ -761,10 +812,8 @@ struct GridwiseGemm_xdl_cshuffle_v3
         else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
         {
             // NLdsLayer * K0 as logical Bank
-            constexpr auto NLdsLayer = 32 * 4 / KPerBlock / sizeof(BDataType) < 1
-                                           ? 1
-                                           : 32 * 4 / KPerBlock / sizeof(BDataType);
-            ;
+            constexpr index_t LdsSize       = 32 * 4 / KPerBlock / sizeof(BDataType) / BPackedSize;
+            constexpr index_t NLdsLayer     = LdsSize < 1 ? 1 : LdsSize;
             constexpr auto b_lds_block_desc = make_naive_tensor_descriptor(
                 make_tuple(
                     BK0Number * Number<NLdsLayer>{}, Number<NPerBlock / NLdsLayer>{}, BK1Number),
@@ -946,8 +995,8 @@ struct GridwiseGemm_xdl_cshuffle_v3
         constexpr auto c_block_size =
             c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
 
-        return math::max((a_block_space_size_aligned * sizeof(ADataType) +
-                          b_block_space_size_aligned * sizeof(BDataType)),
+        return math::max((a_block_space_size_aligned * sizeof(ADataType) / APackedSize +
+                          b_block_space_size_aligned * sizeof(BDataType) / BPackedSize),
                          c_block_size * sizeof(CShuffleDataType));
     }
 
@@ -1312,8 +1361,9 @@ struct GridwiseGemm_xdl_cshuffle_v3
             static_cast<ADataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
 
         auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<BDataType*>(p_shared) +
-                a_block_space_size_aligned * sizeof(ADataType) / sizeof(BDataType),
+            reinterpret_cast<BDataType*>(static_cast<char*>(p_shared) + a_block_space_size_aligned *
+                                                                            sizeof(ADataType) /
+                                                                            APackedSize),
             b_block_desc_bk0_n_bk1.GetElementSpaceSize());
 
         constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0);
@@ -1706,16 +1756,16 @@ struct GridwiseGemm_xdl_cshuffle_v3
             static_cast<ADataType*>(p_shared_0), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
 
         auto b_block_buf_ping = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<BDataType*>(p_shared_0) +
-                a_block_space_size_aligned * sizeof(ADataType) / sizeof(BDataType),
+            bit_cast<BDataType*>(static_cast<char*>(p_shared_0) +
+                                 a_block_space_size_aligned * sizeof(ADataType)),
             b_block_desc_bk0_n_bk1.GetElementSpaceSize());
 
         auto a_block_buf_pong = make_dynamic_buffer<AddressSpaceEnum::Lds>(
             static_cast<ADataType*>(p_shared_1), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
 
         auto b_block_buf_pong = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<BDataType*>(p_shared_1) +
-                a_block_space_size_aligned * sizeof(ADataType) / sizeof(BDataType),
+            bit_cast<BDataType*>(bit_cast<char*>(p_shared_1) +
+                                 a_block_space_size_aligned * sizeof(ADataType)),
             b_block_desc_bk0_n_bk1.GetElementSpaceSize());
 
         auto a_block_bufs = make_tuple(a_block_buf_ping, a_block_buf_pong);

include/ck/tensor_operation/gpu/warp/wmma_gemm.hpp

@@ -307,7 +307,7 @@ struct wmma_type<WmmaInstr::wmma_f32_16x16x16_f16_gfx12,
 
     // Wave mode dependent propety
     static constexpr index_t wave_size = Number<WaveSize>{};
-    // * Fixed in Navi3x, Will be wave mode dependent on Navi4x
+    // * Fixed for gfx11, Will be wave mode dependent on gfx12
     // static constexpr index_t num_src_a_vgprs_per_wave = k_per_wmma / 2 * src_a_data_size / 4;
     // static constexpr index_t num_src_b_vgprs_per_wave = k_per_wmma / 2 * src_b_data_size / 4;
     // * num_acc_vgprs_per_wave alone M direction

include/ck/utility/amd_buffer_addressing.hpp

@@ -429,7 +429,8 @@ __device__ typename vector_type<T, N>::type amd_buffer_load_impl(int32x4_t src_w
             (is_same<T, f8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
             (is_same<T, bf8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
             (is_same<T, int8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
-            (is_same<T, uint8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)),
+            (is_same<T, uint8_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)) ||
+            (is_same<T, pk_i4_t>::value && (N == 1 || N == 2 || N == 4 || N == 8 || N == 16)),
         "wrong! not implemented");
 
     using r_t     = typename vector_type<T, N>::type;

include/ck/utility/amd_inline_asm.hpp

@@ -4,13 +4,34 @@
 #ifndef CK_AMD_INLINE_ASM_HPP
 #define CK_AMD_INLINE_ASM_HPP
 
-#include "data_type.hpp"
 #include "c_style_pointer_cast.hpp"
+#include "data_type.hpp"
 
 // TODO: deprecate all amd_assembly_outer_product_xxx
 
 namespace ck {
 
+inline __device__ int amd_assembly_and_or_b32(int a, int b, int d)
+{
+    int c;
+    asm volatile("v_and_or_b32 %0, %1, %2, %3" : "=v"(c) : "v"(a), "v"(b), "v"(d));
+    return c;
+}
+
+inline __device__ half2_t amd_assembly_pk_fma_f16(half2_t a, half2_t b, half2_t c)
+{
+    half2_t d;
+    asm volatile("v_pk_fma_f16 %0, %1, %2, %3" : "=v"(d) : "v"(a), "v"(b), "v"(c));
+    return d;
+}
+
+inline __device__ half2_t amd_assembly_pk_add_f16(half2_t a, half2_t b)
+{
+    half2_t c;
+    asm volatile("v_pk_add_f16 %0, %1, %2" : "=v"(c) : "v"(a), "v"(b));
+    return c;
+}
+
 // c0 += inner_product(a, b0)
 // c1 += inner_product(a, b1)
 __device__ void amd_assembly_outer_product_1x2(float a, float b0, float b1, float& c0, float& c1)