Merge branch 'develop' into moe_cross_reduce

.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

CMakeLists.txt

@@ -585,7 +585,7 @@ if(NOT GPU_ARCHS AND USER_GPU_TARGETS)
    )
    add_subdirectory(example)
    if(BUILD_TESTING)
-	   add_subdirectory(test)
+       add_subdirectory(test)
    endif()
 endif()
 

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

docs/sphinx/requirements.in

-rocm-docs-core==1.12.0
+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==1.12.0
+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); }

example/01_gemm/gemm_xdl_fp16_v3.cpp

@@ -12,7 +12,7 @@ using CShuffleDataType = ck::half_t;
 using CDataType        = ck::half_t;
 
 using ALayout = Row;
-using BLayout = Row;
+using BLayout = Col;
 using CLayout = Row;
 
 using AElementOp = PassThrough;
@@ -27,17 +27,17 @@ using DeviceGemmV2Instance =
         ALayout,   BLayout,  CLayout,   
         ADataType,   BDataType,  CDataType,  AccDataType,  CShuffleDataType, 
         PassThrough, PassThrough, PassThrough, GemmDefault, 
-        256,
-        224, 256, 
-        64, 8, 2,
+        64,
+        16, 16, 
+        256, 8, 8,
         16,   16,
-        7,    8,
-        S<8, 32, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
+        1,    1,
+        S<32, 2, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
         2, 8, 8, 0,
-        S<8, 32, 1>,  S<0, 2, 1>,  S<0, 2, 1>, 
-        1, 8, 2, 0,
-        1, 2, S<1, 32, 1, 8>, 8,
-        ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3>;
+        S<32, 2, 1>,  S<1, 0, 2>,  S<1, 0, 2>, 
+        2, 8, 8, 0,
+        1, 1, S<1, 16, 1, 4>, 4,
+        ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v2>;
 // clang-format on
 
 using ReferenceGemmInstance = ck::tensor_operation::host::

example/01_gemm/gemm_xdl_streamk.cpp

@@ -15,7 +15,6 @@ using F16 = ck::half_t;
 
 using ALayout = Row;
 using BLayout = Row;
-// using BLayout = Col;
 using CLayout = Row;
 
 using AElementOp = PassThrough;

example/01_gemm/run_gemm_example.inc

@@ -5,88 +5,6 @@
 
 #include "ck/tensor_operation/gpu/device/device_gemm_streamk.hpp"
 
-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 2e-1;
-    }
-    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
-    {
-        return 2e-1;
-    }
-    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 2e-1;
-    }
-    else if constexpr(std::is_same_v<DataType, ck::bf8_t>)
-    {
-        return 2e-1;
-    }
-    else
-    {
-        return 1e-3;
-    }
-}
-
 template <typename ProblemType>
 bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
 {

example/01_gemm/run_gemm_example_streamk_v2.inc

@@ -3,88 +3,6 @@
 
 #pragma once
 
-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;
-    }
-}
-
 template <typename ProblemType>
 bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
 {

example/01_gemm/run_gemm_example_v2.inc

@@ -3,88 +3,6 @@
 
 #pragma once
 
-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;
-    }
-}
-
 template <typename ProblemType>
 bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
 {

example/ck_tile/01_fmha/README.md

@@ -15,8 +15,7 @@ This will result in an executable `build/bin/tile_example_fmha_fwd`
 ## kernel
 The kernel template is `fmha_fwd_kernel.hpp`, this is the grid-wise op in old ck_tile's terminology. We put it here purposely, to demonstrate one can construct a kernel by using various internal component from ck_tile. We may still have an implementation under ck_tile's include path (in the future) for the kernel template.
 
-There are 3 template parameters for this kernel template.
-* `TilePartitioner` is used to map the workgroup to corresponding tile, `fmha_fwd_tile_partitioner.hpp` in this folder served as this purpose.
+There are 2 template parameters for this kernel template.
 * `FmhaPipeline` is one of the block_tile_pipeline(under `include/ck_tile/tile_program/block_tile_pipeline`) which is a performance critical component. Indeed, we did a lot of optimization and trials to optimize the pipeline and may still workout more performance pipeline and update into that folder. People only need to replace this pipeline type and would be able to enjoy the benefit of different performant implementations (stay tuned for updated pipeline(s)).
 * `EpiloguePipeline` will modify and store out the result in the last phase. People usually will do lot of post-fusion at this stage, so we also abstract this concept. Currently we didn't do much thing at the epilogue stage but leave the room for future possible support.
 

example/ck_tile/01_fmha/codegen/ops/fmha_fwd.py

@@ -29,11 +29,6 @@ K0_MAX_SUBMAX_MAP = {
     256: 256
 }
 
-TILE_PARTITIONER_MAP = {
-    "shb" : "ck_tile::FmhaFwdTilePartitioner_SHB",
-    "hbs" : "ck_tile::FmhaFwdTilePartitioner_HBS",
-}
-
 FMHA_FWD_KERNEL_HEADER = """// SPDX-License-Identifier: MIT
 // Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.\n
 // auto generated by generate.py
@@ -90,9 +85,7 @@ using fmha_epilogue_{F_idx} =
                                            {F_spad}, {F_dvpad}>>;
 
 using fmha_kernel_{F_idx} =
-    ck_tile::FmhaFwdKernel<{F_tile_partitioner}<fmha_shape_{F_idx}>,
-                  fmha_pipeline_{F_idx},
-                  fmha_epilogue_{F_idx}>;
+    ck_tile::FmhaFwdKernel<fmha_pipeline_{F_idx}, fmha_epilogue_{F_idx}>;
 
 using trait_{F_idx} = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode},{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout},
                         {F_pipeline_enum}, fmha_mask_{F_idx}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
@@ -329,12 +322,6 @@ class FmhaFwdKernel:
     F_pipeline      : FmhaFwdPipeline
     mask_impl       : str
 
-    def get_tp(self) -> str:
-        if self.F_mode == 'group':
-            return 'hbs'
-        else:
-            return 'shb'
-
     @property
     def template(self) -> str:
         kernel_body = str()
@@ -374,13 +361,12 @@ class FmhaFwdKernel:
                 F_pipeline_enum = PIPELINE_ENUM_MAP[self.F_pipeline.tag],
                 F_mask          = get_mask_map(self.mask_impl)[self.F_pipeline.F_mask],
                 F_mode          = MODE_MAP[self.F_mode],
-                F_pipeline      = PIPELINE_MAP[self.F_pipeline.tag],
-                F_tile_partitioner = TILE_PARTITIONER_MAP[self.get_tp()])
+                F_pipeline      = PIPELINE_MAP[self.F_pipeline.tag])
 
     @property
     def name(self) -> str:
         # TODO: we don't encode idx here
-        return f"fmha_fwd_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_{self.get_tp()}_" + \
+        return f"fmha_fwd_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_" + \
                 self.F_tile.name + '_' + self.F_pipeline.name
 
     @property