Merge branch 'develop' into codegen_hiprtc

.github/CODEOWNERS

-* @junliume @illsilin @carlushuang @qianfengz @aosewski @poyenc @geyyer @bartekxk @andriy-ca
+* @junliume @illsilin @carlushuang @qianfengz @aosewski @poyenc @geyyer @bartekxk @andriy-ca @afagaj
 # Documentation files
-docs/ @ROCm/rocm-documentation @junliume @illsilin @carlushuang @qianfengz @aosewski @poyenc @geyyer @bartekxk @andriy-ca
-*.md @ROCm/rocm-documentation @junliume @illsilin @carlushuang @qianfengz @aosewski @poyenc @geyyer @bartekxk @andriy-ca
-*.rst @ROCm/rocm-documentation @junliume @illsilin @carlushuang @qianfengz @aosewski @poyenc @geyyer @bartekxk @andriy-ca
-.readthedocs.yaml @ROCm/rocm-documentation @junliume @illsilin @carlushuang @qianfengz @aosewski @poyenc @geyyer @bartekxk @andriy-ca
+docs/ @ROCm/rocm-documentation @junliume @illsilin @carlushuang @qianfengz @aosewski @poyenc @geyyer @bartekxk @andriy-ca @afagaj
+*.md @ROCm/rocm-documentation @junliume @illsilin @carlushuang @qianfengz @aosewski @poyenc @geyyer @bartekxk @andriy-ca @afagaj
+*.rst @ROCm/rocm-documentation @junliume @illsilin @carlushuang @qianfengz @aosewski @poyenc @geyyer @bartekxk @andriy-ca @afagaj
+.readthedocs.yaml @ROCm/rocm-documentation @junliume @illsilin @carlushuang @qianfengz @aosewski @poyenc @geyyer @bartekxk @andriy-ca @afagaj
 # Header directory for Doxygen documentation
-library/include/ @ROCm/rocm-documentation @junliume @illsilin @carlushuang @qianfengz @aosewski @poyenc @geyyer @bartekxk @andriy-ca
+library/include/ @ROCm/rocm-documentation @junliume @illsilin @carlushuang @qianfengz @aosewski @poyenc @geyyer @bartekxk @andriy-ca @afagaj

.github/CONTRIBUTING.md

+We'd love for you to contribute to our source code!
+
+Some helpful links:
+
+- [Code of Conduct guidelines](https://www.contributor-covenant.org/version/2/1/code_of_conduct/code_of_conduct.txt)
+- [New issue guidelines](https://github.com/rocm/composable_kernel/blob/develop/.github/ISSUE_TEMPLATE.md)
+- [Submitting a pull request guidelines](https://github.com/rocm/composable_kernel/blob/develop/.github/PULL_REQUEST_TEMPLATE.md)
+- [Maintainers](https://github.com/rocm/composable_kernel/blob/develop/CONTRIBUTORS.md)
+- [General information](https://github.com/rocm/composable_kernel/blob/develop/README.md)
+- [ROCm documentation](https://rocm.docs.amd.com/en/latest/how-to/llm-fine-tuning-optimization/optimizing-with-composable-kernel.html)
\ No newline at end of file

.github/ISSUE_TEMPLATE.md

+When creating an issue, please check if a similar issue already exists.
+
+### When reporting a bug, please include:
+- [ ] A descriptive title
+- [ ] An isolated way to reproduce the behavior (preferably a docker container with a repro)
+- [ ] ROCm version, clang version, Composable Kernel commit pin
+- [ ] Environment variables
+- [ ] The behavior you expect to see, and the behavior you actually see
+
+### When requesting a feature, please include:
+- [ ] A descriptive title
+- [ ] A detailed description of the problem you are trying to solve
+- [ ] An overview of the suggested solution
+- [ ] Explanation why the solution is an improvement
\ No newline at end of file

.github/PULL_REQUEST_TEMPLATE.md

+## Proposed changes
+
+Please describe the motivation behind the pull request, whether it enables a new feature or fixes a bug. If there are associated pull requests or issues, please link them to the pull request.
+
+## Checklist
+
+Please put an `x` into the boxes that apply. You can also fill these out after creating the PR. If you're not sure, please don't hesitate to ask.
+
+- [ ] I have added tests relevant to the introduced functionality, and the unit tests are passing locally
+- [ ] I have added inline documentation which enables the maintainers with understanding the motivation
+- [ ] I have removed the stale documentation which is no longer relevant after this pull request
+- [ ] (If this change is user-facing) I have added release notes which provide the end users with a brief summary of the improvement from this pull request
+- [ ] I have run `clang-format` on all changed files
+- [ ] Any dependent changes have been merged
+
+## Discussion
+
+If this is a relatively large or complex change, feel free to start a discussion by explaining why you chose the solution you did and what alternatives you considered
+

CMakeLists.txt

@@ -183,14 +183,17 @@ message("Building CK for the following targets: ${SUPPORTED_GPU_TARGETS}")
 if (SUPPORTED_GPU_TARGETS MATCHES "gfx9")
     message("Enabling XDL instances")
     add_definitions(-DCK_USE_XDL)
+    set(CK_USE_XDL "ON")
 endif()
 if (SUPPORTED_GPU_TARGETS MATCHES "gfx94")
     message("Enabling FP8 gemms on native architectures")
     add_definitions(-DCK_USE_GFX94)
+    set(CK_USE_GFX94 "ON")
 endif()
 if (SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12")
     message("Enabling WMMA instances")
     add_definitions(-DCK_USE_WMMA)
+    set(CK_USE_WMMA "ON")
 endif()
 if (SUPPORTED_GPU_TARGETS MATCHES "gfx12")
     add_definitions(-DCK_USE_OCP_FP8)
@@ -204,6 +207,7 @@ endif()
 option(CK_USE_FP8_ON_UNSUPPORTED_ARCH "Enable FP8 GEMM instances on older architectures" OFF)
 if(CK_USE_FP8_ON_UNSUPPORTED_ARCH AND (SUPPORTED_GPU_TARGETS MATCHES "gfx90a" OR SUPPORTED_GPU_TARGETS MATCHES "gfx908"))
     add_definitions(-DCK_USE_FP8_ON_UNSUPPORTED_ARCH)
+    set(CK_USE_FP8_ON_UNSUPPORTED_ARCH "ON")
 endif()
 
 # CK config file to record supported datatypes, etc.
@@ -581,7 +585,7 @@ if(NOT GPU_ARCHS AND USER_GPU_TARGETS)
    )
    add_subdirectory(example)
    if(BUILD_TESTING)
-	   add_subdirectory(test)
+       add_subdirectory(test)
    endif()
 endif()
 

Dockerfile

-FROM ubuntu:20.04
+FROM ubuntu:22.04
 ARG DEBIAN_FRONTEND=noninteractive
-ARG ROCMVERSION=6.2
+ARG ROCMVERSION=6.3
 ARG compiler_version=""
 ARG compiler_commit=""
 ARG CK_SCCACHE=""
@@ -13,17 +13,12 @@ RUN set -xe && \
     apt-get update && apt-get install -y --allow-unauthenticated apt-utils wget gnupg2 curl && \
     curl -fsSL https://repo.radeon.com/rocm/rocm.gpg.key | gpg --dearmor -o /etc/apt/trusted.gpg.d/rocm-keyring.gpg
 
-RUN if [ "$ROCMVERSION" != "6.3" ]; then \
-        sh -c "wget https://repo.radeon.com/amdgpu-install/$ROCMVERSION/ubuntu/focal/amdgpu-install_6.2.60200-1_all.deb  --no-check-certificate" && \
-        apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated ./amdgpu-install_6.2.60200-1_all.deb && \
+RUN if [ "$ROCMVERSION" != "6.4" ]; then \
+        sh -c "wget https://repo.radeon.com/amdgpu-install/$ROCMVERSION/ubuntu/focal/amdgpu-install_6.3.60300-1_all.deb  --no-check-certificate" && \
+        apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-unauthenticated ./amdgpu-install_6.3.60300-1_all.deb && \
         wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add - && \
         sh -c "echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] $DEB_ROCM_REPO focal main > /etc/apt/sources.list.d/rocm.list" && \
         sh -c 'echo deb [arch=amd64 signed-by=/etc/apt/trusted.gpg.d/rocm-keyring.gpg] https://repo.radeon.com/amdgpu/$ROCMVERSION/ubuntu focal main > /etc/apt/sources.list.d/amdgpu.list'; \
-    elif [ "$ROCMVERSION" = "6.3" ] && [ "$compiler_version" = "rc1" ]; then \
-        sh -c "wget http://artifactory-cdn.amd.com/artifactory/list/amdgpu-deb/amdgpu-install-internal_6.3-20.04-1_all.deb --no-check-certificate" && \
-        apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install dialog libpopt0 rsync && DEBIAN_FRONTEND=noninteractive apt-get install ./amdgpu-install-internal_6.3-20.04-1_all.deb && \
-        sh -c 'echo deb [arch=amd64 trusted=yes] http://compute-artifactory.amd.com/artifactory/list/rocm-release-archive-20.04-deb/ 6.3 rel-20 > /etc/apt/sources.list.d/rocm-build.list' && \
-        amdgpu-repo --amdgpu-build=2074281; \
     fi
 
 RUN sh -c "echo deb http://mirrors.kernel.org/ubuntu focal main universe | tee -a /etc/apt/sources.list" && \
@@ -53,6 +48,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
     libnuma-dev \
     libpthread-stubs0-dev \
     llvm-amdgpu \
+    mpich \
     net-tools \
     pkg-config \
     python \
@@ -68,6 +64,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
     nano \
     zlib1g-dev \
     zip \
+    libzstd-dev \
     openssh-server \
     clang-format-12 \
     kmod && \
@@ -75,7 +72,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
     rm -rf /var/lib/apt/lists/* && \
     rm -rf amdgpu-install* && \
 # Remove unnecessary rocm components that take a lot of space
-    apt-get remove -y rocblas rocfft rocsparse composablekernel-dev
+    apt-get remove -y rocblas rocfft rocsparse composablekernel-dev hipblaslt
 
 # Update the cmake to version 3.27.5
 RUN pip install --upgrade cmake==3.27.5 && \
@@ -97,7 +94,7 @@ RUN pip install --upgrade cmake==3.27.5 && \
     dpkg -i dumb-init_*.deb && rm dumb-init_*.deb && \
 # Install packages for processing the performance results
     pip3 install --upgrade pip && \
-    pip3 install sqlalchemy==1.4.46 pymysql pandas==2.0.3 setuptools-rust sshtunnel==0.4.0 && \
+    pip3 install sqlalchemy==2.0.36 pymysql pandas==2.2.3 setuptools-rust sshtunnel==0.4.0 && \
 # Add render group
     groupadd -f render && \
 # Install the new rocm-cmake version

Dockerfile.compiler

-ARG BASE_DOCKER="rocm/composable_kernel:ck_ub20.04_rocm6.2"
+ARG BASE_DOCKER="rocm/composable_kernel:ck_ub22.04_rocm6.3"
 FROM $BASE_DOCKER
 ARG compiler_version=""
 ARG compiler_commit=""

cmake/EnableCompilerWarnings.cmake

@@ -66,7 +66,7 @@ else()
             -Wunreachable-code
             -Wunused
             -Wno-reserved-identifier
-	    -Werror
+            -Werror
             -Wno-option-ignored
             -Wsign-compare
             -Wno-extra-semi-stmt

codegen/test/rtc/include/rtc/hip.hpp

@@ -5,6 +5,7 @@
 #include <memory>
 #include <stdexcept>
 #include <string>
+#include <stdexcept>
 
 namespace rtc {
 

docs/sphinx/requirements.in

-rocm-docs-core==1.10.0
+rocm-docs-core==1.12.1
 sphinxcontrib-bibtex==2.6.3

docs/sphinx/requirements.txt

@@ -103,7 +103,7 @@ requests==2.32.3
     # via
     #   pygithub
     #   sphinx
-rocm-docs-core==1.10.0
+rocm-docs-core==1.12.1
     # via -r requirements.in
 six==1.16.0
     # via pybtex

example/01_gemm/CMakeLists.txt

@@ -29,10 +29,16 @@ add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_v3)
 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_executable(example_gemm_xdl_bf16_pk_i4_v3 gemm_xdl_bf16_pk_i4_v3.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)
 
+add_example_executable(example_gemm_xdl_bf16_streamk_v3 gemm_xdl_bf16_streamk_v3.cpp)
+add_example_dependencies(example_gemm_xdl example_gemm_xdl_bf16_streamk_v3)
+
 add_example_executable(example_gemm_xdl_wavelet_fp16 gemm_xdl_wavelet_fp16.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_wavelet_fp16)
 

example/01_gemm/common.hpp

@@ -287,3 +287,85 @@ bool parse_cmd_args<ProblemSizeSplitK>(int argc,
 
     return true;
 }
+
+template <typename DataType>
+inline __host__ __device__ constexpr double get_rtol()
+{
+    if constexpr(std::is_same_v<DataType, float>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, double>)
+    {
+        return 1e-6;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::half_t>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
+    {
+        return 5e-2;
+    }
+    else if constexpr(std::is_same_v<DataType, int32_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, int8_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
+    {
+        return 1e-1; // 240 and 224 are acceptable
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
+    {
+        return 1.5e-1; // 57344 and 49152 are acceptable
+    }
+    else
+    {
+        return 1e-3;
+    }
+}
+
+template <typename DataType>
+inline __host__ __device__ constexpr double get_atol()
+{
+    if constexpr(std::is_same_v<DataType, float>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, double>)
+    {
+        return 1e-6;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::half_t>)
+    {
+        return 1e-3;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bhalf_t>)
+    {
+        return 5e-2;
+    }
+    else if constexpr(std::is_same_v<DataType, int32_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, int8_t>)
+    {
+        return 1e-1;
+    }
+    else if constexpr(std::is_same_v<DataType, ck::f8_t>)
+    {
+        return 16.1; // 240 and 224 are acceptable
+    }
+    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
+    {
+        return 8192.1; // 57344 and 49152 are acceptable
+    }
+    else
+    {
+        return 1e-3;
+    }
+}

example/01_gemm/gemm_xdl_bf16_pk_i4_v3.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp"
+
+using ADataType        = ck::bhalf_t;
+using BDataType        = ck::pk_i4_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::bhalf_t;
+using CDataType        = ck::bhalf_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 PermuteA         = false;
+static constexpr bool PermuteB         = true;
+static constexpr ck::index_t KPerBlock = 128;
+
+// clang-format off
+using DeviceGemmV2Instance = 
+    ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3<
+        ALayout,   BLayout,  CLayout,   
+        ADataType, BDataType, CDataType, AccDataType, CShuffleDataType, 
+        AElementOp, BElementOp, CElementOp, GemmDefault, 
+        128,
+        16, 64,
+        KPerBlock, 8, 32,
+        16,   16,
+        1,    2,
+        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,
+        ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v2, ADataType, ADataType, PermuteA, PermuteB>;
+
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        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, ck::index_t stride, auto layout) {
+            if(stride == -1)
+            {
+                // give a chance if stride is -1, return a default packed stride
+                if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+                {
+                    return static_cast<std::size_t>(col);
+                }
+                else
+                {
+                    return static_cast<std::size_t>(row);
+                }
+            }
+            else
+                return static_cast<std::size_t>(stride);
+        };
+
+    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{}));
+
+    switch(config.init_method)
+    {
+    case 0:
+        a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        break;
+    case 2:
+        a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        break;
+    case 3:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+    }
+
+    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 << "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 c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    // 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);
+            }
+        }
+    }
+
+    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n_permute.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,
+                                      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;
+    }
+
+    bool pass = true;
+    if(config.do_verification)
+    {
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, 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(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); }

example/01_gemm/gemm_xdl_bf16_streamk_v3.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_streamk_v3.hpp"
+
+using ADataType        = ck::bhalf_t;
+using BDataType        = ck::bhalf_t;
+using CDataType        = ck::bhalf_t;
+using AccDataType      = float;
+using CShuffleDataType = ck::bhalf_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;
+
+// clang-format off
+using DeviceGemmV2_Streamk_Instance = 
+    ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_Streamk_V3<
+        ALayout,   BLayout,  CLayout,   
+        ADataType,   BDataType,  CDataType,  AccDataType,  CShuffleDataType, 
+        PassThrough, PassThrough, PassThrough, GemmDefault, 
+        256,
+        128, 128, 
+        64, 8, 8,
+        16,   16,
+        4,    4,
+        S<8, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
+        2, 8, 8, 0,
+        S<8, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
+        2, 8, 8, 0,
+        1, 2, S<1, 32, 1, 8>, 8,
+        ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3>;
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp>;
+
+#include "run_gemm_example_streamk_v2.inc"
+
+int main(int argc, char* argv[]) { return !run_gemm_universal_streamk_example(argc, argv); }

example/01_gemm/gemm_xdl_fp16_fp8_v3.cpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 
 #include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp"
 
-using ADataType        = ck::f8_t;
-using BDataType        = ck::half_t;
+using ADataType        = ck::half_t;
+using BDataType        = ck::f8_t;
 using AccDataType      = float;
 using CShuffleDataType = ck::half_t;
 using CDataType        = ck::half_t;
@@ -29,15 +29,15 @@ using DeviceGemmV2Instance =
         AElementOp, BElementOp, CElementOp, GemmDefault, 
         64,
         16, 16, 
-        64, 16, 8,
+        256, 8, 16,
         16,   16,
         1,    1, 
-        S<4, 16, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
-        2, 16, 16, 0,
-        S<8, 8, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
+        S<32, 2, 1>,  S<1, 0, 2>,  S<1, 0, 2>,
         2, 8, 8, 0,
+        S<16, 4, 1>,  S<1, 0, 2>,  S<1, 0, 2>,
+        2, 16, 16, 0,
         1, 1, S<1, 16, 1, 4>, 4,
-        ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v1>;
+        ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1>;
 // clang-format on
 
 using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,

example/01_gemm/gemm_xdl_fp16_pk_i4_v3.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp"
+
+using ADataType        = ck::half_t;
+using BDataType        = ck::pk_i4_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 PermuteA         = false;
+static constexpr bool PermuteB         = true;
+static constexpr ck::index_t KPerBlock = 128;
+
+// clang-format off
+using DeviceGemmV2Instance = 
+    ck::tensor_operation::device::DeviceGemm_Xdl_CShuffleV3<
+        ALayout,   BLayout,  CLayout,   
+        ADataType, BDataType, CDataType, AccDataType, CShuffleDataType, 
+        AElementOp, BElementOp, CElementOp, GemmDefault, 
+        128,
+        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,
+        ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v2, ADataType, ADataType, PermuteA, PermuteB>;
+
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                        BDataType,
+                                                                        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, ck::index_t stride, auto layout) {
+            if(stride == -1)
+            {
+                // give a chance if stride is -1, return a default packed stride
+                if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+                {
+                    return static_cast<std::size_t>(col);
+                }
+                else
+                {
+                    return static_cast<std::size_t>(row);
+                }
+            }
+            else
+                return static_cast<std::size_t>(stride);
+        };
+
+    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{}));
+
+    switch(config.init_method)
+    {
+    case 0:
+        a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        break;
+    case 2:
+        a_m_k.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        break;
+    case 3:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n.GenerateTensorValue(GeneratorTensor_1<BDataType>{1});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+    }
+
+    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 << "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 c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    // 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
+    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).data;
+                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;
+            }
+        }
+    }
+
+    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n_permute.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,
+                                      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;
+    }
+
+    bool pass = true;
+    if(config.do_verification)
+    {
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_m_k, b_k_n, 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(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); }

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 PermuteA = false;
+static constexpr bool PermuteB = true;
+
+static constexpr ck::index_t Scale_Block_N = 1;
+static constexpr ck::index_t Scale_Block_K = 128;
+
+static constexpr ck::index_t KPerBlock = 64;
+
+// 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, 
+        256, Scale_Block_N, Scale_Block_K,
+        128, 128,
+        KPerBlock, 8, 32,
+        32,   32,
+        4,    1,
+        S<8, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>,
+        2, 8, 8, 0,
+        S<2, 128, 1>,  S<1, 0, 2>,  S<1, 0, 2>,
+        2, 32, 32, 0,
+        1, 1, S<1, 32, 1, 8>, 8,
+        ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, CDataType, CDataType, PermuteA, 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, ck::index_t stride, auto layout) {
+            if(stride == -1)
+            {
+                // give a chance if stride is -1, return a default packed stride
+                if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
+                {
+                    return static_cast<std::size_t>(col);
+                }
+                else
+                {
+                    return static_cast<std::size_t>(row);
+                }
+            }
+            else
+                return static_cast<std::size_t>(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());
+
+    // 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
+    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).data;
+                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;
+            }
+        }
+    }
+
+    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,
+                          Scale_Stride_BN,
+                          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;
+    }
+
+    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).data;
+                int8_t i4        = 0;
+                if(k % 2 == 1)
+                    i4 = (i4x2.data >> 0) & 0xf;
+                else
+                    i4 = (i4x2.data >> 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(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); }