Merge the int4 kernel and profiling in one commit for RTP.

Add int4+scale based on Zhang, Jing pk_i4. Compile pass, function pass.
Modify the kernel to 128x128x128, and use mfma_32x32x4
Move the weight permute from host to device

Modified the scale init method.

Modified the init method, the function is failed, need to debug.

Added init method

Support group=128 for Llam2-7B-int4

Move the weight permute from host to device

Add ckProfiler for GEMM b scale (int4)

Add reference function.

Add pipeline v4 (2 LDS pingpong)

Add more int4-Gemm kernel profiling instances.

Modify the in4-Gemm kernel instances

Move the pk_i4 permute in kernel

example/01_gemm/CMakeLists.txt

@@ -30,6 +30,7 @@ add_example_executable(example_gemm_xdl_fp8_v3 gemm_xdl_fp8_v3.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8_v3)
 add_example_executable(example_gemm_xdl_fp16_fp8_v3 gemm_xdl_fp16_fp8_v3.cpp)
 add_example_executable(example_gemm_xdl_fp16_pk_i4_v3 gemm_xdl_fp16_pk_i4_v3.cpp)
+add_example_executable(example_gemm_xdl_fp16_pk_i4_v3_b_scale gemm_xdl_fp16_pk_i4_v3_b_scale.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_fp8_v3)
 add_example_executable(example_gemm_xdl_bf16_v3 gemm_xdl_bf16_v3.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_bf16_v3)

example/01_gemm/gemm_xdl_fp16_pk_i4_v3.cpp

@@ -23,7 +23,7 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
 
 static constexpr bool PermuteB = true;
 
-static constexpr ck::index_t KPerBlock = 128;
+static constexpr ck::index_t KPerBlock = 64;
 
 // clang-format off
 using DeviceGemmV2Instance = 
@@ -43,22 +43,22 @@ using DeviceGemmV2Instance =
         2, 32, 32, 0,
         1, 1, S<1, 16, 1, 8>, 4,
 #else
-        128,
-        16, 64,
+        256,
+        128, 128,
         KPerBlock, 8, 32,
-        16,   16,
-        1,    2,
-        S<16, 8, 1>,  S<1, 0, 2>,  S<1, 0, 2>,
+        32,   32,
+        2,    2,
+        S<8, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>,
         2, 8, 8, 0,
-        S<4, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>,
+        S<2, 128, 1>,  S<1, 0, 2>,  S<1, 0, 2>,
         2, 32, 32, 0,
         1, 1, S<1, 16, 1, 8>, 4,
 #endif
-        ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v2, CDataType, CDataType, false, PermuteB>;
+        ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, CDataType, CDataType, false, PermuteB>;
 
 // clang-format on
 
-    using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
                                                                         BDataType,
                                                                         CDataType,
                                                                         AccDataType,

example/01_gemm/gemm_xdl_fp16_pk_i4_v3_b_scale.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_b_scale.hpp"
+
+using ADataType        = ck::half_t;
+using BDataType        = ck::pk_i4_t;
+using BScaleDataType   = ck::half_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::half_t;
+using CDataType        = ck::half_t;
+
+using ALayout = Row;
+using BLayout = Col;
+using CLayout = Row;
+
+using AElementOp = PassThrough;
+using BElementOp = PassThrough;
+using CElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+static constexpr bool PermuteB = false;
+
+static constexpr ck::index_t Scale_Block_N = 1;
+static constexpr ck::index_t Scale_Block_K = 128;
+
+static constexpr ck::index_t KPerBlock = 128;
+
+// clang-format off
+using DeviceGemmV2Instance = 
+    ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3<
+        ALayout,   BLayout,  CLayout,   
+        ADataType, BDataType, BScaleDataType, CDataType, AccDataType, CShuffleDataType, 
+        AElementOp, BElementOp, CElementOp, GemmDefault, 
+#if 0
+        128, Scale_Block_N, Scale_Block_K,
+        16, 128,
+        KPerBlock, 8, 32,
+        16,   16,
+        1,    4,
+        S<16, 8, 1>,  S<1, 0, 2>,  S<1, 0, 2>,
+        2, 8, 8, 0,
+        S<4, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>,
+        2, 32, 32, 0,
+        1, 1, S<1, 16, 1, 8>, 4,
+#else
+        256, Scale_Block_N, Scale_Block_K,
+        128, 128,
+        KPerBlock, 8, 32,
+        32,   32,
+        2,    2,
+        S<16, 16, 1>,  S<1, 0, 2>,  S<1, 0, 2>,
+        2, 8, 8, 0,
+        S<4, 64, 1>,  S<1, 0, 2>,  S<1, 0, 2>,
+        2, 32, 32, 0,
+        1, 1, S<1, 16, 1, 8>, 4,
+#endif
+        ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, CDataType, CDataType, false, PermuteB>;
+
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        AccDataType,
+                                                                        CDataType,
+                                                                        AccDataType,
+                                                                        PassThrough,
+                                                                        PassThrough,
+                                                                        PassThrough>;
+template <typename ProblemType>
+bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
+{
+    using namespace ck::literals;
+
+    auto M       = problem_size.M;
+    auto N       = problem_size.N;
+    auto K       = problem_size.K;
+    auto StrideA = problem_size.StrideA;
+    auto StrideB = problem_size.StrideB;
+    auto StrideC = problem_size.StrideC;
+    auto KBatch  = problem_size.KBatch;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    auto f_get_default_stride =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            if(stride == 0)
+            {
+                // give a chance if stride is zero, return a default packed stride
+                if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+                {
+                    return col;
+                }
+                else
+                {
+                    return row;
+                }
+            }
+            else
+                return stride;
+        };
+
+    ck::index_t Scale_Stride_BN = (K + Scale_Block_K - 1) / Scale_Block_K;
+
+    StrideA = f_get_default_stride(M, K, StrideA, ALayout{});
+    StrideB = f_get_default_stride(K, N, StrideB, BLayout{});
+    StrideC = f_get_default_stride(M, N, StrideC, CLayout{});
+
+    Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<BDataType> b_k_n_permute(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<BScaleDataType> b1_k_n(f_host_tensor_descriptor((K + Scale_Block_K - 1) / Scale_Block_K,
+                                                           (N + Scale_Block_N - 1) / Scale_Block_N,
+                                                           Scale_Stride_BN,
+                                                           BLayout{}));
+
+    switch(config.init_method)
+    {
+    case 0:
+        a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_1<BScaleDataType>{1});
+        break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
+        break;
+    case 2:
+        a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_1<BScaleDataType>{1});
+        break;
+    case 3:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_1<BScaleDataType>{1});
+        break;
+    case 4:
+        a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
+        break;
+    case 5:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_1<BScaleDataType>{1});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.5, 0.5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_3<BScaleDataType>{0, 1.0});
+    }
+
+    Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_host_result.mDesc << std::endl;
+
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n_permute.mDesc.GetElementSpaceSize());
+    DeviceMem b1_scale_device_buf(sizeof(BScaleDataType) * b1_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    printf("b_k_n element space size: %zu, b_k_n device size: %lu, BDataType size: %lu\n",
+           b_k_n_permute.mDesc.GetElementSpaceSize(),
+           sizeof(BDataType) * b_k_n_permute.mDesc.GetElementSpaceSize(),
+           sizeof(BDataType));
+
+    // weight permute
+    if constexpr(PermuteB)
+    {
+        int K1 = KPerBlock;
+        int K0 = K / KPerBlock;
+
+        // int K0, N, K1
+        for(int j = 0; j < K0; j++)
+        {
+            for(int i = 0; i < N; i++)
+            {
+                for(int jj = 0; jj < K1; jj++)
+                {
+                    b_k_n_permute(j * N * K1 + i * K1 + jj) = b_k_n(i * K + (j * K1 + jj));
+                }
+            }
+        }
+    }
+    else
+    {
+        for(int i = 0; i < N; i++)
+        {
+            for(int j = 0; j < K; j++)
+            {
+                b_k_n_permute(i * K + j) = b_k_n(i * K + j);
+            }
+        }
+    }
+
+    // vector pk_i4x4 permute
+#if 0
+    for(int i = 0; i < N; i++)
+    {
+        for(int j = 0; j < K; j += 8)
+        {
+            int input[8];
+
+            for(int k = 0; k < 4; k++)
+            {
+                int i4x2         = b_k_n_permute(j + k * 2, i);
+                input[k * 2 + 0] = (i4x2 >> 4) & 0xf;
+                input[k * 2 + 1] = (i4x2 >> 0) & 0xf;
+            }
+
+            // permute 01234567->20643175
+            {
+                int hi   = input[2];
+                int lo   = input[0];
+                int i4x2 = (hi << 4) | lo;
+
+                b_k_n_permute(j + 0, i) = i4x2;
+            }
+
+            {
+                int hi   = input[6];
+                int lo   = input[4];
+                int i4x2 = (hi << 4) | lo;
+
+                b_k_n_permute(j + 2, i) = i4x2;
+            }
+
+            {
+                int hi   = input[3];
+                int lo   = input[1];
+                int i4x2 = (hi << 4) | lo;
+
+                b_k_n_permute(j + 4, i) = i4x2;
+            }
+
+            {
+                int hi   = input[7];
+                int lo   = input[5];
+                int i4x2 = (hi << 4) | lo;
+
+                b_k_n_permute(j + 6, i) = i4x2;
+            }
+        }
+    }
+#endif
+
+    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n_permute.mData.data());
+    b1_scale_device_buf.ToDevice(b1_k_n.mData.data());
+    DeviceMem workspace;
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CElementOp{};
+
+    // do GEMM
+    auto gemm      = DeviceGemmV2Instance{};
+    auto invoker   = gemm.MakeInvoker();
+    float ave_time = 0;
+
+    auto argument =
+        gemm.MakeArgument(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+                          static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+                          static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+                          M,
+                          N,
+                          K,
+                          StrideA,
+                          StrideB,
+                          StrideC,
+                          static_cast<BScaleDataType*>(b1_scale_device_buf.GetDeviceBuffer()),
+                          KBatch,
+                          a_element_op,
+                          b_element_op,
+                          c_element_op);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        std::cerr << gemm.GetTypeString() << " does not support this problem" << std::endl;
+
+        return true;
+    }
+
+    std::size_t workspace_size = gemm.GetWorkSpaceSize(&argument);
+    printf("workspace_size: %zu\n", workspace_size);
+
+    bool pass = true;
+    if(config.do_verification)
+    {
+        Tensor<float> b_k_n_dequant({K, N});
+
+        float v_b = 0;
+        for(int n = 0; n < N; n++)
+        {
+            for(int k = 0; k < K; k++)
+            {
+                ck::pk_i4_t i4x2 = b_k_n(k, n);
+                int8_t i4        = 0;
+                if(k % 2 == 1)
+                    i4 = (i4x2 >> 0) & 0xf;
+                else
+                    i4 = (i4x2 >> 4) & 0xf;
+                i4  = i4 - 8;
+                v_b = ck::type_convert<float>(i4);
+
+                b_k_n_dequant(k, n) =
+                    ck::type_convert<float>(v_b) *
+                    ck::type_convert<float>(b1_k_n(k / Scale_Block_K, n / Scale_Block_N));
+            }
+        }
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n_dequant, c_m_n_host_result, PassThrough{}, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        ave_time = invoker.Run(argument, StreamConfig{nullptr, false, 0});
+        c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+        pass &= ck::utils::check_err(c_m_n_device_result,
+                                     c_m_n_host_result,
+                                     "Error: Incorrect results!",
+                                     get_rtol<CDataType>(),
+                                     get_atol<CDataType>());
+
+#if 0
+        std::cout << "a_m_k: " << std::endl;
+        for(int i = 0; i < M; i++)
+        {
+            for(int j = 0; j < K; j++)
+            {
+                std::cout << ck::type_convert<float>(a_m_k(i, j)) << ",";
+            }
+            std::cout << std::endl;
+        }
+
+        std::cout << "b_k_n: " << std::endl;
+        for(int i = 0; i < N; i++)
+        {
+            for(int j = 0; j < K; j++)
+            {
+                ck::pk_i4_t i4x2 = b_k_n(j, i);
+                int8_t i4 = 0;
+                if( j % 2 == 1)
+                    i4 = (i4x2 >> 0) & 0xf;
+                else
+                    i4 = (i4x2 >> 4) & 0xf;
+                i4 = i4 - 8;
+                std::cout << ck::type_convert<float>(i4) << ",";
+            }
+            std::cout << std::endl;
+        }
+
+        std::cout<<"scale_b1_k_n: "<<std::endl;
+        for(int i = 0; i < N; i++)
+        {
+            for(int j = 0; j < K; j++)
+            {
+                std::cout << ck::type_convert<float>(b1_k_n(i,j)) << ",";
+            }
+            std::cout << std::endl;
+        }
+
+        std::cout << "c_m_n_device_result: " << std::endl;
+        for(int i = 0; i < M; i++)
+        {
+            for(int j = 0; j < N; j++)
+            {
+                std::cout << ck::type_convert<float>(c_m_n_device_result(i, j)) << ",";
+            }
+            std::cout << std::endl;
+        }
+
+        std::cout << "c_m_n_host_result: " << std::endl;
+        for(int i = 0; i < M; i++)
+        {
+            for(int j = 0; j < N; j++)
+            {
+                std::cout << ck::type_convert<float>(c_m_n_host_result(i, j)) << ",";
+            }
+            std::cout << std::endl;
+        }
+#endif
+    }
+
+    if(config.time_kernel)
+    {
+        ave_time =
+            invoker.Run(argument, StreamConfig{nullptr, config.time_kernel, 0, 20, 50, true, 50});
+
+        std::size_t flop = 2_uz * M * N * K;
+        std::size_t num_btype =
+            sizeof(ADataType) * M * K +
+            sizeof(BDataType) * K * N /
+                (ck::is_same_v<ck::remove_cvref_t<BDataType>, ck::pk_i4_t> ? 2 : 1) +
+            sizeof(CDataType) * M * N;
+
+        float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s, " << gemm.GetTypeString() << std::endl;
+    }
+    return pass;
+}
+
+bool run_gemm_splitk_example(int argc, char* argv[])
+{
+    ProblemSizeSplitK problem_size;
+    ExecutionConfig config;
+
+    return !parse_cmd_args(argc, argv, problem_size, config) || run_gemm(problem_size, config);
+}
+
+int main(int argc, char* argv[]) { return !run_gemm_splitk_example(argc, argv); }

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_b_scale.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_b_scale.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_b_scale.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v4_b_scale.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v5.hpp"
+
+namespace ck {
+
+enum struct BlockGemmPipelineVersion
+{
+    v1, // Naive
+    v2, // Mem
+    v3, // Comp
+    v4, // Comp, double lds buffer
+    v5, // Comp, double global prefetch register buffer
+};
+
+template <BlockGemmPipelineVersion BlkGemmPipelineVer,
+          BlockGemmPipelineScheduler BlkGemmPipeSche,
+          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>
+constexpr auto BlockGemmPipeline_Selector()
+{
+    if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
+    {
+        return BlockwiseGemmXdlops_pipeline_v1_b_scale<BlkGemmPipeSche,
+                                                       BlockSize,
+                                                       ADataType,
+                                                       BDataType,
+                                                       ComputeDataType,
+                                                       AccDataType,
+                                                       ATileDesc,
+                                                       BTileDesc,
+                                                       AMmaTileDesc,
+                                                       BMmaTileDesc,
+                                                       ABlockTransferSrcScalarPerVector,
+                                                       BBlockTransferSrcScalarPerVector,
+                                                       MPerBlock,
+                                                       NPerBlock,
+                                                       KPerBlock,
+                                                       MPerXDL,
+                                                       NPerXDL,
+                                                       MRepeat,
+                                                       NRepeat,
+                                                       KPack>{};
+    }
+    else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
+    {
+        return BlockwiseGemmXdlops_pipeline_v2_b_scale<BlkGemmPipeSche,
+                                                       BlockSize,
+                                                       ADataType,
+                                                       BDataType,
+                                                       ComputeDataType,
+                                                       AccDataType,
+                                                       ATileDesc,
+                                                       BTileDesc,
+                                                       AMmaTileDesc,
+                                                       BMmaTileDesc,
+                                                       ABlockTransferSrcScalarPerVector,
+                                                       BBlockTransferSrcScalarPerVector,
+                                                       MPerBlock,
+                                                       NPerBlock,
+                                                       KPerBlock,
+                                                       MPerXDL,
+                                                       NPerXDL,
+                                                       MRepeat,
+                                                       NRepeat,
+                                                       KPack>{};
+    }
+    else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
+    {
+        return BlockwiseGemmXdlops_pipeline_v3_b_scale<BlkGemmPipeSche,
+                                                        BlockSize,
+                                                        ADataType,
+                                                        BDataType,
+                                                        ComputeDataType,
+                                                        AccDataType,
+                                                        ATileDesc,
+                                                        BTileDesc,
+                                                        AMmaTileDesc,
+                                                        BMmaTileDesc,
+                                                        ABlockTransferSrcScalarPerVector,
+                                                        BBlockTransferSrcScalarPerVector,
+                                                        MPerBlock,
+                                                        NPerBlock,
+                                                        KPerBlock,
+                                                        MPerXDL,
+                                                        NPerXDL,
+                                                        MRepeat,
+                                                        NRepeat,
+                                                        KPack>{};
+    }
+    else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v4)
+    {
+        return BlockwiseGemmXdlops_pipeline_v4_b_scale<BlkGemmPipeSche,
+                                                       BlockSize,
+                                                       ADataType,
+                                                       BDataType,
+                                                       ComputeDataType,
+                                                       AccDataType,
+                                                       ATileDesc,
+                                                       BTileDesc,
+                                                       AMmaTileDesc,
+                                                       BMmaTileDesc,
+                                                       ABlockTransferSrcScalarPerVector,
+                                                       BBlockTransferSrcScalarPerVector,
+                                                       MPerBlock,
+                                                       NPerBlock,
+                                                       KPerBlock,
+                                                       MPerXDL,
+                                                       NPerXDL,
+                                                       MRepeat,
+                                                       NRepeat,
+                                                       KPack>{};
+    }
+    else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v5)
+    {
+        return BlockwiseGemmXdlops_pipeline_v5<BlkGemmPipeSche,
+                                                       BlockSize,
+                                                       ADataType,
+                                                       BDataType,
+                                                       ComputeDataType,
+                                                       AccDataType,
+                                                       ATileDesc,
+                                                       BTileDesc,
+                                                       AMmaTileDesc,
+                                                       BMmaTileDesc,
+                                                       ABlockTransferSrcScalarPerVector,
+                                                       BBlockTransferSrcScalarPerVector,
+                                                       MPerBlock,
+                                                       NPerBlock,
+                                                       KPerBlock,
+                                                       MPerXDL,
+                                                       NPerXDL,
+                                                       MRepeat,
+                                                       NRepeat,
+                                                       KPack>{};
+    }
+    else
+    {
+        std::cerr << "BlockGemmPipeline configuration is not available" << std::endl;
+    }
+}
+
+} // namespace ck

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

@@ -76,6 +76,42 @@ 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,                        
+                        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/element/unary_element_wise_operation.hpp

@@ -79,6 +79,33 @@ struct PassThroughPack8
     __host__ __device__ constexpr void operator()(ck::half8_t& y, const ck::pk_i4x4_t& x) const
     {
 #if 1
+        int x_permute = 0;
+        int bits4_0 = (bit_cast<int>(x) >> 0) & 0xF; 
+        int bits4_1 = (bit_cast<int>(x) >> 4) & 0xF;
+        int bits4_2 = (bit_cast<int>(x) >> 8) & 0xF;
+        int bits4_3 = (bit_cast<int>(x) >> 12) & 0xF;
+        int bits4_4 = (bit_cast<int>(x) >> 16) & 0xF;
+        int bits4_5 = (bit_cast<int>(x) >> 20) & 0xF;
+        int bits4_6 = (bit_cast<int>(x) >> 24) & 0xF;
+        int bits4_7 = (bit_cast<int>(x) >> 28) & 0xF;
+
+        x_permute |= (bits4_1 << 0);
+        x_permute |= (bits4_3 << 4);
+        x_permute |= (bits4_5 << 8);
+        x_permute |= (bits4_7 << 12);
+        x_permute |= (bits4_0 << 16);
+        x_permute |= (bits4_2 << 20);
+        x_permute |= (bits4_4 << 24);
+        x_permute |= (bits4_6 << 28);
+
+        vector_type<half_t, 8> result;
+
+        result.template AsType<half4_t>()(Number<0>{}) = pki4_to_half4(x_permute);
+        result.template AsType<half4_t>()(Number<1>{}) = pki4_to_half4(x_permute >> 8);
+
+        y = result.template AsType<half8_t>()[Number<0>{}];
+
+#elif 1
         vector_type<half_t, 8> result;
 
         result.template AsType<half4_t>()(Number<0>{}) = pki4_to_half4(bit_cast<int>(x));

library/include/ck/library/tensor_operation_instance/gpu/gemm_b_scale.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3_b_scale.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+#if(defined(CK_ENABLE_FP16) || defined(CK_ENABLE_FP8))
+void add_device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn_mem_v2_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmV2BScale<Row,
+                                                   Col,
+                                                   Row,
+                                                   F16,
+                                                   I4,
+                                                   F16,
+                                                   F16,
+                                                   1,
+                                                   128,
+                                                   PassThrough,
+                                                   PassThrough,
+                                                   PassThrough>>>& instances);
+#endif
+
+template <typename ADataType,
+          typename BDataType,
+          typename BScaleDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          index_t  ScaleBlockK>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceGemmV2BScale<ALayout,
+                                                     BLayout,
+                                                     CLayout,
+                                                     ADataType,
+                                                     BDataType,
+                                                     BScaleDataType,
+                                                     CDataType,
+                                                     1,
+                                                     ScaleBlockK,
+                                                     ck::tensor_operation::element_wise::PassThrough,
+                                                     ck::tensor_operation::element_wise::PassThrough,
+                                                     ck::tensor_operation::element_wise::PassThrough>>
+{
+    using DeviceOp = DeviceGemmV2BScale<ALayout,
+                                        BLayout,
+                                        CLayout,
+                                        ADataType,
+                                        BDataType,
+                                        BScaleDataType,
+                                        CDataType,
+                                        1,
+                                        ScaleBlockK,
+                                        ck::tensor_operation::element_wise::PassThrough,
+                                        ck::tensor_operation::element_wise::PassThrough,
+                                        ck::tensor_operation::element_wise::PassThrough>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+        
+        if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, pk_i4_t> &&
+                     is_same_v<CDataType, half_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn_mem_v2_default_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_b_scale/CMakeLists.txt

+# ONLY XDL_KERNELS
+set(GEMM_B_SCALE_INSTANCES)
+
+list(APPEND GEMM_B_SCALE_INSTANCES 
+        device_gemm_b_scale_xdl_f16_i4_f16/device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn_mem_v2_default_instance.cpp
+        )
+
+set_source_files_properties(device_gemm_b_scale_xdl_f16_i4_f16/device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn_mem_v2_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
+
+add_instance_library(device_gemm_b_scale_instance ${GEMM_B_SCALE_INSTANCES})
\ No newline at end of file

library/src/tensor_operation_instance/gpu/gemm_b_scale/device_gemm_b_scale_xdl_f16_i4_f16/device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn_mem_v2_default_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+void add_device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn_mem_v2_default_instances(
+    std::vector<std::unique_ptr<
+        DeviceGemmV2BScale<Row, Col, Row, F16, I4, F16, F16, 1, 128, PassThrough, PassThrough, PassThrough>>>&
+        instances)
+{
+    add_device_operation_instances(
+        instances,
+        //device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn_mem_instances<Interwave, GemmDefault>{});
+        device_gemm_b_scale_xdl_f16_i4_f16_mk_nk_mn_mem_instances<Intrawave, GemmDefault>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
\ No newline at end of file

profiler/src/CMakeLists.txt

@@ -59,6 +59,7 @@ set(PROFILER_SOURCES
 #  list(APPEND PROFILER_SOURCES profile_gemm_bias_add_reduce.cpp)
 #  list(APPEND PROFILER_SOURCES profile_gemm_splitk.cpp)
    list(APPEND PROFILER_SOURCES profile_gemm_universal.cpp)
+   list(APPEND PROFILER_SOURCES profile_gemm_b_scale.cpp)
 #  list(APPEND PROFILER_SOURCES profile_gemm_universal_reduce.cpp)
 #  list(APPEND PROFILER_SOURCES profile_gemm_universal_streamk.cpp)
 #  list(APPEND PROFILER_SOURCES profile_conv_fwd_bias_relu.cpp)
@@ -141,6 +142,7 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE utility getopt::getopt)
 #  endif()
 #  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
    target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_instance)
+   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_b_scale_instance)
 #  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_reduce_instance)
 #  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_universal_streamk_instance)
 #  target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_add_multiply_instance)

profiler/src/profile_gemm_b_scale.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "profiler/profile_gemm_b_scale_impl.hpp"
+#include "profiler_operation_registry.hpp"
+
+enum struct GemmMatrixLayout
+{
+    MK_KN_MN, // 0
+    MK_NK_MN, // 1
+    KM_KN_MN, // 2
+    KM_NK_MN, // 3
+};
+
+enum struct GemmDataType
+{
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
+    F8_F16_F16,     // 4
+    F16_F8_F16,     // 5
+    F16_F16_F16_F8, // 6
+    F8_F8_BF16,     // 7
+    F16_I4_F16,     // 8
+};
+
+enum struct BScaleBlockTile
+{
+    K_64, // 0
+    K_128,  // 1
+};
+
+#define OP_NAME "gemm_b_scale"
+#define OP_DESC "Int4-dequant GEMM"
+
+int profile_gemm_b_scale(int argc, char* argv[])
+{
+    if(argc != 16 && argc != 19)
+    {
+        printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: "
+               "f16->f8; 7: f8->bf16, "
+               "comp f8; 8: f16@i4)\n");
+        printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
+        printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
+        printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
+        printf("                     3: A[k, m] * B[n, k] = C[m, n])\n");
+        printf("arg4: B scale block tile (0: 64, 1: 128):\n");
+        printf("arg5: verification (0: no; 1: yes)\n");
+        printf("arg6: initialization (0: no init; 1: integer value; 2: decimal value)\n");
+        printf("arg7: print tensor value (0: no; 1: yes)\n");
+        printf("arg8: time kernel (0=no, 1=yes)\n");
+        printf("arg9 to 14: M, N, K, StrideA, StrideB, StrideC\n");
+        printf("arg15: split k into  mulitiple batch\n");
+        printf("optional:\n");
+        printf("arg16: number of warm-up cycles (default 1)\n");
+        printf("arg17: number of iterations (default 10)\n");
+        printf("arg18: memory for rotating buffer (default 0, size in MB)\n");
+        exit(1);
+    }
+
+    printf("Start profiling\n");
+    const auto data_type       = static_cast<GemmDataType>(std::stoi(argv[2]));
+    const auto layout          = static_cast<GemmMatrixLayout>(std::stoi(argv[3]));
+    const auto B_scale_block   = static_cast<BScaleBlockTile>(std::stoi(argv[4]));
+    const bool do_verification = std::stoi(argv[5]);
+    const int init_method      = std::stoi(argv[6]);
+    const bool do_log          = std::stoi(argv[7]);
+    const bool time_kernel     = std::stoi(argv[8]);
+
+    const int M = std::stoi(argv[9]);
+    const int N = std::stoi(argv[10]);
+    const int K = std::stoi(argv[11]);
+
+    const int StrideA = std::stoi(argv[12]);
+    const int StrideB = std::stoi(argv[13]);
+    const int StrideC = std::stoi(argv[14]);
+    const int KBatch  = std::stoi(argv[15]);
+    printf("M:%d, N:%d, K:%d, StrideA:%d, StrideB:%d, StrideC:%d, KBatch:%d\n", M, N, K, StrideA, StrideB, StrideC, KBatch);
+
+    int n_warmup      = 1;
+    int n_iter        = 10;
+    uint64_t rotating = 0;
+    if(argc == 18)
+    {
+        n_warmup = std::stoi(argv[16]);
+        n_iter   = std::stoi(argv[17]);
+        rotating = std::stoull(argv[18]) * 1024 * 1024;
+    }
+
+    using F32  = float;
+    using F16  = ck::half_t;
+    using BF16 = ck::bhalf_t;
+    using F8   = ck::f8_t;
+    using I4   = ck::pk_i4_t;
+
+    using Row = ck::tensor_layout::gemm::RowMajor;
+    using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+    auto profile = [&](auto a_type,
+                       auto b_type,
+                       auto b_scale_type,
+                       auto comp_type,
+                       auto acc_type,
+                       auto c_type,
+                       auto scale_block_k,
+                       auto a_layout,
+                       auto b_layout,
+                       auto c_layout) {
+        using ADataType       = decltype(a_type);
+        using BDataType       = decltype(b_type);
+        using BScaleDataType  = decltype(b_scale_type);
+        using ComputeDataType = decltype(comp_type);
+        using AccDataType     = decltype(acc_type);
+        using CDataType       = decltype(c_type);
+
+        using ALayout = decltype(a_layout);
+        using BLayout = decltype(b_layout);
+        using CLayout = decltype(c_layout);
+
+        const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
+        const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
+        const int DefaultStrideC = ck::is_same_v<CLayout, Row> ? N : M;
+
+        bool pass = ck::profiler::profile_gemm_b_scale_impl<ADataType,
+                                                            BDataType,
+                                                            BScaleDataType,
+                                                            ComputeDataType,
+                                                            AccDataType,
+                                                            CDataType,
+                                                            scale_block_k,
+                                                            ALayout,
+                                                            BLayout,
+                                                            CLayout>(
+            do_verification,
+            init_method,
+            do_log,
+            time_kernel,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? DefaultStrideA : StrideA,
+            (StrideB < 0) ? DefaultStrideB : StrideB,
+            (StrideC < 0) ? DefaultStrideC : StrideC,
+            KBatch,
+            n_warmup,
+            n_iter,
+            rotating);
+
+        return pass ? 0 : 1;
+    };
+
+    // if(data_type == GemmDataType::F16_I4_F16 && layout == GemmMatrixLayout::MK_NK_MN && B_scale_block == BScaleBlockTile::K_64)
+    // {
+    //     return profile(F16{}, I4{}, F16{}, F16{}, F32{}, F16{}, ck::Number<64>{}, Row{}, Col{}, Row{});
+    // }
+    if(data_type == GemmDataType::F16_I4_F16 && layout == GemmMatrixLayout::MK_NK_MN && B_scale_block == BScaleBlockTile::K_128)
+    {
+        printf("F16_I4_F16 MK_NK_MN K_128\n");
+        return profile(F16{}, I4{}, F16{}, F16{}, F32{}, F16{}, ck::Number<128>{}, Row{}, Col{}, Row{});
+    }
+    else
+    {
+        std::cout << "this data_type & layout is not implemented" << std::endl;
+
+        return 1;
+    }
+}
+
+REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_gemm_b_scale);