Merge branch 'develop' into add_contraction_example_fp64

CHANGELOG.md

+# Change Log for Composable Kernel
+
+Full documentation for Composable Kernel is not yet available.
+
+## CK 0.1.1 for ROCm 5.5.0
+
+### Fixed
+- Fixed a bug in 6-dimensional kernels (#555).
+- Fixed grouped ConvBwdWeight test case failure (#524).
+
+### Optimizations
+- Optimized ...
+
+### Added
+- Added user tutorial (#563).
+- Added more instances for irregular GEMM sizes (#560).
+- Added inter-wave consumer-producer programming model for GEMM kernels (#310).
+- Added multi-D GEMM client APIs (#534).
+- Added multi-embeddings support (#542).
+- Added Navi3x blockwise GEMM and real GEMM support (#541).
+
+### Changed
+- Changed ...

Dockerfile

@@ -60,7 +60,7 @@ RUN dpkg -i dumb-init_*.deb && rm dumb-init_*.deb
 ARG PREFIX=/opt/rocm
 # Install packages for processing the performance results
 RUN pip3 install --upgrade pip
-RUN pip3 install sqlalchemy
+RUN pip3 install sqlalchemy==1.4.46
 RUN pip3 install pymysql
 RUN pip3 install pandas
 RUN pip3 install setuptools-rust

Jenkinsfile

@@ -19,7 +19,14 @@ def runShell(String command){
 }
 
 def getDockerImageName(){
-    def img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}"
+    def img
+    if (params.COMPILER_COMMIT == ""){
+        img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}"
+    }
+    else{
+        def commit = "${params.COMPILER_COMMIT}"[0..6]
+        img = "${env.CK_DOCKERHUB}:ck_ub20.04_rocm${params.ROCMVERSION}_${params.COMPILER_VERSION}_${commit}"
+    }
     return img
 }
 
@@ -550,8 +557,9 @@ def process_results(Map conf=[:]){
 }
 
 //launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version
-CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;COMPILER_VERSION=release
-                                              0 19 * * * % BUILD_DOCKER=true;COMPILER_VERSION=amd-stg-open''' : ""
+CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true
+                                              0 21 * * * % RUN_FULL_QA=false;COMPILER_VERSION=release;COMPILER_COMMIT=
+                                              0 19 * * * % BUILD_DOCKER=true;COMPILER_VERSION=amd-stg-open;COMPILER_COMMIT=''' : ""
 
 pipeline {
     agent none
@@ -568,16 +576,16 @@ pipeline {
             description: "Force building docker image (default: false), set to true if docker image needs to be updated.")
         string(
             name: 'ROCMVERSION', 
-            defaultValue: '5.3', 
-            description: 'Specify which ROCM version to use: 5.2.3, or 5.3 (default), etc.')
+            defaultValue: '5.4.3', 
+            description: 'Specify which ROCM version to use: 5.4.3 (default).')
         string(
             name: 'COMPILER_VERSION', 
-            defaultValue: 'release', 
-            description: 'Specify which version of compiler to use: ck-9110, release (default), or amd-stg-open.')
+            defaultValue: 'amd-stg-open', 
+            description: 'Specify which version of compiler to use: ck-9110, release, or amd-stg-open (default).')
         string(
             name: 'COMPILER_COMMIT', 
-            defaultValue: '', 
-            description: 'Specify which commit of compiler branch to use: leave empty to use the latest commit (default), or use 8a82e4eb7ba28521ba9a9424a0315a8a16590424 commit of amd-stg-open branch.')
+            defaultValue: '5541927df00eabd6a110180170eca7785d436ee3', 
+            description: 'Specify which commit of compiler branch to use: leave empty to use the latest commit, or use 5541927df00eabd6a110180170eca7785d436ee3 (default) commit of amd-stg-open branch.')
         string(
             name: 'BUILD_COMPILER', 
             defaultValue: 'hipcc', 

client_example/01_gemm/gemm.cpp

@@ -83,7 +83,7 @@ int main(int argc, char* argv[])
         [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
             using Layout = decltype(layout);
 
-            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
             {
                 return (nRow - 1) * stride + nCol;
             }

client_example/02_gemm_add_add_fastgelu/gemm_add_add_fastgelu.cpp

@@ -92,7 +92,7 @@ int main(int argc, char* argv[])
         [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
             using Layout = decltype(layout);
 
-            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
             {
                 return (nRow - 1) * stride + nCol;
             }

client_example/02_gemm_add_add_fastgelu/gemm_add_fastgelu.cpp

@@ -88,7 +88,7 @@ int main(int argc, char* argv[])
         [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
             using Layout = decltype(layout);
 
-            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
             {
                 return (nRow - 1) * stride + nCol;
             }

client_example/02_gemm_add_add_fastgelu/gemm_fastgelu.cpp

@@ -84,7 +84,7 @@ int main(int argc, char* argv[])
         [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
             using Layout = decltype(layout);
 
-            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
             {
                 return (nRow - 1) * stride + nCol;
             }

client_example/03_gemm_layernorm/CMakeLists.txt

-add_executable(client_gemm_add_add_reduce_normalize gemm_add_add_layernorm.cpp)
-target_link_libraries(client_gemm_add_add_reduce_normalize PRIVATE composable_kernel::device_operations)
+add_executable(client_gemm_add_add_layernorm_naive gemm_add_add_layernorm_naive.cpp)
+target_link_libraries(client_gemm_add_add_layernorm_naive PRIVATE composable_kernel::device_operations)
+
+add_executable(client_gemm_add_relu_add_layernorm_welford gemm_add_relu_add_layernorm_welford.cpp)
+target_link_libraries(client_gemm_add_relu_add_layernorm_welford PRIVATE composable_kernel::device_operations)

client_example/03_gemm_layernorm/gemm_add_add_layernorm.cpp → client_example/03_gemm_layernorm/gemm_add_add_layernorm_naive.cpp

@@ -190,7 +190,7 @@ int main()
         [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
             using Layout = decltype(layout);
 
-            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
             {
                 return (nRow - 1) * stride + nCol;
             }

client_example/03_gemm_layernorm/gemm_add_relu_add_layernorm_welford.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/gemm_add_relu_add_layernorm.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_gemm_multiple_d_layernorm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using F16 = ck::half_t;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddReluAdd  = ck::tensor_operation::element_wise::AddReluAdd;
+
+// DataType
+using ADataType     = F16;
+using BDataType     = F16;
+using D0DataType    = F16;
+using D1DataType    = F16;
+using GammaDataType = F16;
+using BetaDataType  = F16;
+using HDataType     = F16;
+
+// Layout
+using ALayout  = Row;
+using BLayout  = Col;
+using D0Layout = Row;
+using D1Layout = Row;
+using HLayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = AddReluAdd;
+using HElementOp   = PassThrough;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}, mMemSize_(mem_size)
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    void SetZero() const { (void)hipMemset(p_mem_, 0, mMemSize_); }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+    std::size_t mMemSize_;
+};
+
+int main(int argc, char* argv[])
+{
+    // GEMM shape
+    ck::index_t M = 1024;
+    ck::index_t N = 1024;
+    ck::index_t K = 1024;
+
+    ck::index_t StrideA  = K;
+    ck::index_t StrideB  = K;
+    ck::index_t StrideD0 = 0;
+    ck::index_t StrideD1 = N;
+    ck::index_t StrideH  = N;
+
+    float epsilon = 1e-5;
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * f_matrix_space_size(M, K, StrideA, ALayout{}));
+    SimpleDeviceMem b_device_buf(sizeof(BDataType) * f_matrix_space_size(K, N, StrideB, BLayout{}));
+    SimpleDeviceMem d0_device_buf(sizeof(D0DataType) *
+                                  f_matrix_space_size(M, N, StrideD0, D0Layout{}));
+    SimpleDeviceMem d1_device_buf(sizeof(D1DataType) *
+                                  f_matrix_space_size(M, N, StrideD1, D1Layout{}));
+    SimpleDeviceMem gamma_device_buf(sizeof(GammaDataType) * N);
+    SimpleDeviceMem beta_device_buf(sizeof(BetaDataType) * N);
+    SimpleDeviceMem h_device_buf(sizeof(HDataType) * f_matrix_space_size(M, N, StrideH, HLayout{}));
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGemmMultipleDLayernorm<
+        ALayout,
+        BLayout,
+        ck::Tuple<D0Layout, D1Layout>,
+        HLayout,
+        ADataType,
+        BDataType,
+        ck::Tuple<D0DataType, D1DataType>,
+        GammaDataType,
+        BetaDataType,
+        HDataType,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::PassThrough,
+        ck::tensor_operation::element_wise::AddReluAdd,
+        ck::tensor_operation::element_wise::PassThrough>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+    const auto h_element_op   = HElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            {d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
+            gamma_device_buf.GetDeviceBuffer(),
+            beta_device_buf.GetDeviceBuffer(),
+            h_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            {StrideD0, StrideD1},
+            StrideH,
+            epsilon,
+            a_element_op,
+            b_element_op,
+            cde_element_op,
+            h_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace_dev(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+            h_device_buf.SetZero();
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_byte =
+                sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                (sizeof(D0DataType) + sizeof(D1DataType) + sizeof(HDataType)) * M * N +
+                (sizeof(GammaDataType) + sizeof(BetaDataType)) * N;
+
+            float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(ave_time < best_ave_time)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+              << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b_device_buf.GetDeviceBuffer(),
+            {d0_device_buf.GetDeviceBuffer(), d1_device_buf.GetDeviceBuffer()},
+            gamma_device_buf.GetDeviceBuffer(),
+            beta_device_buf.GetDeviceBuffer(),
+            h_device_buf.GetDeviceBuffer(),
+            M,
+            N,
+            K,
+            StrideA,
+            StrideB,
+            {StrideD0, StrideD1},
+            StrideH,
+            epsilon,
+            a_element_op,
+            b_element_op,
+            cde_element_op,
+            h_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            size_t workspace_sz = op_ptr->GetWorkSpaceSize(argument_ptr.get());
+            SimpleDeviceMem workspace_dev(workspace_sz);
+            op_ptr->SetWorkSpacePointer(argument_ptr.get(), workspace_dev.GetDeviceBuffer());
+            h_device_buf.SetZero();
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}
\ No newline at end of file

client_example/08_fused_attention/CMakeLists.txt

 add_executable(client_fused_attention fused_attention.cpp)
 target_link_libraries(client_fused_attention PRIVATE composable_kernel::device_operations)
+
+add_executable(client_fused_attention_bias fused_attention_bias.cpp)
+target_link_libraries(client_fused_attention_bias PRIVATE composable_kernel::device_operations)

client_example/08_fused_attention/fused_attention_bias.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/batched_gemm_bias_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+using AElementOp    = ck::tensor_operation::element_wise::PassThrough;
+using B0ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using Acc0ElementOp = ck::tensor_operation::element_wise::ScaleAdd;
+using B1ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CElementOp    = ck::tensor_operation::element_wise::PassThrough;
+
+constexpr static auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+
+using ADataType   = ck::half_t;
+using B0DataType  = ck::half_t;
+using B1DataType  = ck::half_t;
+using CDataType   = ck::half_t;
+using D0DataType  = ck::half_t;
+using AccDataType = float;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main(int argc, char* argv[])
+{
+    int G0 = 48;
+    int G1 = 16;
+    int M  = 1024;
+    int N  = 1024;
+    int K  = 64;
+    int O  = 64;
+
+    // A layout [G0, M, G1, K]
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{M * G1 * K, K, G1 * K, 1};
+
+    // B0 layout [G0, N, G1, K]
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides{N * G1 * K, K, G1 * K, 1};
+
+    // B1 layout [G0, N, G1, O]
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides{N * G1 * O, O, 1, G1 * O};
+
+    // C layout [G0, M, G1, O]
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides{M * G1 * O, O, G1 * O, 1};
+
+    // D layout [G0, M, G1, N]
+    std::vector<ck::index_t> d0_gs_ms_ns_lengths{G0, G1, M, N};
+    std::vector<ck::index_t> d0_gs_ms_ns_strides{M * G1 * N, N, G1 * N, 1};
+
+    SimpleDeviceMem a_device_buf(sizeof(ADataType) * G0 * G1 * M * K);
+    SimpleDeviceMem b0_device_buf(sizeof(B0DataType) * G0 * G1 * N * K);
+    SimpleDeviceMem d0_device_buf(sizeof(D0DataType) * G0 * G1 * M * N);
+    SimpleDeviceMem b1_device_buf(sizeof(B1DataType) * G0 * G1 * O * N);
+    SimpleDeviceMem c_device_buf(sizeof(CDataType) * G0 * G1 * M * O);
+
+    using DeviceOp =
+        ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute<2,
+                                                                          1,
+                                                                          1,
+                                                                          1,
+                                                                          1,
+                                                                          ADataType,
+                                                                          B0DataType,
+                                                                          B1DataType,
+                                                                          CDataType,
+                                                                          ck::Tuple<D0DataType>,
+                                                                          ck::Tuple<>,
+                                                                          AElementOp,
+                                                                          B0ElementOp,
+                                                                          Acc0ElementOp,
+                                                                          B1ElementOp,
+                                                                          CElementOp,
+                                                                          MaskingSpec>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device op instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr      = op_ptrs[i];
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b0_device_buf.GetDeviceBuffer(),
+            b1_device_buf.GetDeviceBuffer(),
+            c_device_buf.GetDeviceBuffer(),
+            std::array<void*, 1>{d0_device_buf.GetDeviceBuffer()}, // p_acc0_biases
+            {},                                                    // p_acc1_biases
+            a_gs_ms_ks_lengths,
+            a_gs_ms_ks_strides,
+            b0_gs_ns_ks_lengths,
+            b0_gs_ns_ks_strides,
+            b1_gs_os_ns_lengths,
+            b1_gs_os_ns_strides,
+            c_gs_ms_os_lengths,
+            c_gs_ms_os_strides,
+            std::array<std::vector<ck::index_t>, 1>{
+                d0_gs_ms_ns_lengths}, // acc0_biases_gs_ms_ns_lengths
+            std::array<std::vector<ck::index_t>, 1>{
+                d0_gs_ms_ns_strides}, // acc0_biases_gs_ms_ns_strides
+            {},                       // acc1_biases_gs_ms_os_lengths
+            {},                       // acc1_biases_gs_ms_os_strides
+            AElementOp{},
+            B0ElementOp{},
+            Acc0ElementOp{1 / sqrtf(K)},
+            B1ElementOp{},
+            CElementOp{});
+
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop      = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * G0 * G1;
+            std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
+                                     sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O +
+                                     sizeof(D0DataType) * M * N) *
+                                    G0 * G1;
+
+            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, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    // run the best instance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            a_device_buf.GetDeviceBuffer(),
+            b0_device_buf.GetDeviceBuffer(),
+            b1_device_buf.GetDeviceBuffer(),
+            c_device_buf.GetDeviceBuffer(),
+            std::array<void*, 1>{d0_device_buf.GetDeviceBuffer()}, // p_acc0_biases
+            {},                                                    // p_acc1_biases
+            a_gs_ms_ks_lengths,
+            a_gs_ms_ks_strides,
+            b0_gs_ns_ks_lengths,
+            b0_gs_ns_ks_strides,
+            b1_gs_os_ns_lengths,
+            b1_gs_os_ns_strides,
+            c_gs_ms_os_lengths,
+            c_gs_ms_os_strides,
+            std::array<std::vector<ck::index_t>, 1>{
+                d0_gs_ms_ns_lengths}, // acc0_biases_gs_ms_ns_lengths
+            std::array<std::vector<ck::index_t>, 1>{
+                d0_gs_ms_ns_strides}, // acc0_biases_gs_ms_ns_strides
+            {},                       // acc1_biases_gs_ms_os_lengths
+            {},                       // acc1_biases_gs_ms_os_strides
+            AElementOp{},
+            B0ElementOp{},
+            Acc0ElementOp{1 / sqrtf(K)},
+            B1ElementOp{},
+            CElementOp{});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+
+    return 0;
+}

client_example/15_gemm_add_multiply/gemm_add_multiply.cpp

@@ -92,7 +92,7 @@ int main(int argc, char* argv[])
         [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
             using Layout = decltype(layout);
 
-            if(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
             {
                 return (nRow - 1) * stride + nCol;
             }

doc/markdown/tutorial_hello_world.md

+## CK Hello world
+
+## Motivation
+
+This tutorial is aimed at engineers dealing with artificial intelligence and machine learning who would like to optimize their pipelines and squeeze every performance drop by adding Composable Kernel (CK) library to their projects. We would like to make the CK library approachable so the tutorial is not based on the latest release and doesn't have all the bleeding edge features, but it will be reproducible now and forever.
+
+During this tutorial we will have an introduction to the CK library, we will build it and run some examples and tests, so to say we will run a "Hello world" example. In future tutorials we will go in depth and breadth and get familiar with other tools and ways to integrate CK into your project.
+
+## Description
+
+Modern AI technology solves more and more problems in all imaginable fields, but crafting fast and efficient workflows is still challenging. CK is one of the tools to make AI heavy lifting as fast and efficient as possible. CK is a collection of optimized AI operator kernels and tools to create new ones. The library has components required for majority of modern neural networks architectures including matrix multiplication, convolution, contraction, reduction, attention modules, variety of activation functions, fused operators and many more.
+
+So how do we (almost) reach the speed of light? CK acceleration abilities are based on:
+
+* Layered structure.
+* Tile-based computation model.
+* Tensor coordinate transformation.
+* Hardware acceleration use.
+* Support of low precision data types including fp16, bf16, int8 and int4.
+
+If you are excited and need more technical details and benchmarking results - read this awesome blog [post](https://community.amd.com/t5/instinct-accelerators/amd-composable-kernel-library-efficient-fused-kernels-for-ai/ba-p/553224). 
+
+For more details visit our [github repo](https://github.com/ROCmSoftwarePlatform/composable_kernel).
+
+## Hardware targets
+
+CK library fully supports "gfx908" and "gfx90a" GPU architectures and only some operators are supported for "gfx1030". Let's check the hardware you have at hand and decide on the target GPU architecture 
+
+GPU Target	AMD GPU
+gfx908 	Radeon Instinct MI100
+gfx90a 	Radeon Instinct MI210, MI250, MI250X
+gfx1030 	Radeon PRO V620, W6800, W6800X, W6800X Duo, W6900X, RX 6800, RX 6800 XT, RX 6900 XT, RX 6900 XTX, RX 6950 XT
+
+There are also [cloud options](https://aws.amazon.com/ec2/instance-types/g4/) you can find if you don't have an AMD GPU at hand.
+
+## Build the library
+
+First let's clone the library and rebase to the tested version:
+
+```
+git clone https://github.com/ROCmSoftwarePlatform/composable_kernel.git
+cd composable_kernel/
+git checkout tutorial_hello_world
+```
+
+To make our lives easier we prepared [docker images](https://hub.docker.com/r/rocm/composable_kernel) with all the necessary dependencies. Pick the right image and create a container. In this tutorial we use "rocm/composable_kernel:ck_ub20.04_rocm5.3_release" image, it is based on Ubuntu 20.04, ROCm v5.3, compiler release version.
+
+If your current folder is ${HOME}, start the docker container with
+
+```
+docker run  \
+-it  \
+--privileged  \
+--group-add sudo  \
+-w /root/workspace  \
+-v ${HOME}:/root/workspace  \
+rocm/composable_kernel:ck_ub20.04_rocm5.3_release  \
+/bin/bash
+```
+
+If your current folder is different from ${HOME}, adjust the line `-v ${HOME}:/root/workspace` to fit your folder structure.
+
+Inside the docker container current folder is "~/workspace", library path is "~/workspace/composable_kernel", navigate to the library
+
+```
+cd composable_kernel/
+```
+
+Create and go to the "build" directory
+
+```
+mkdir build && cd build
+```
+
+In the previous section we talked about target GPU architecture. Once you decide which one is right for you, run cmake using the right GPU_TARGETS flag
+
+```
+cmake  \
+-D CMAKE_PREFIX_PATH=/opt/rocm  \
+-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc  \
+-D CMAKE_CXX_FLAGS="-O3"  \
+-D CMAKE_BUILD_TYPE=Release  \
+-D BUILD_DEV=OFF  \
+-D GPU_TARGETS="gfx908;gfx90a;gfx1030" ..
+```
+
+If everything went well the cmake run will end up with:
+
+```
+-- Configuring done
+-- Generating done
+-- Build files have been written to: "/root/workspace/composable_kernel/build"
+```
+
+Finally, we can build examples and tests
+
+```
+make -j examples tests
+```
+
+If everything is smooth, you'll see
+
+```
+Scanning dependencies of target tests
+[100%] Built target tests
+```
+
+## Run examples and tests
+
+Examples are listed as test cases as well, so we can run all examples and tests with
+
+```
+ctest
+```
+
+You can check the list of all tests by running
+
+```
+ctest -N
+```
+
+We can also run them separately, here is a separate example execution. 
+
+```
+./bin/example_gemm_xdl_fp16 1 1 1
+```
+
+The arguments "1 1 1" mean that we want to run this example in the mode: verify results with CPU, initialize matrices with integers and benchmark the kernel execution. You can play around with these parameters and see how output and execution results change.
+
+If everything goes well and you have a device based on gfx908 or gfx90a architecture you should see something like
+
+```
+a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
+c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+launch_and_time_kernel: grid_dim {480, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+Perf: 1.10017 ms, 117.117 TFlops, 87.6854 GB/s, DeviceGemmXdl<256, 256, 128, 4, 8, 32, 32, 4, 2> NumPrefetch: 1, LoopScheduler: Default, PipelineVersion: v1
+```
+
+Meanwhile, running it on a gfx1030 device should result in
+
+```
+a_m_k: dim 2, lengths {3840, 4096}, strides {4096, 1}
+b_k_n: dim 2, lengths {4096, 4096}, strides {1, 4096}
+c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+DeviceGemmXdl<256, 256, 128, 4, 8, 32, 32, 4, 2> NumPrefetch: 1, LoopScheduler: Default, PipelineVersion: v1 does not support this problem
+```
+
+But don't panic, some of the operators are supported on gfx1030 architecture, so you can run a separate example like
+
+```
+./bin/example_gemm_dl_fp16 1 1 1
+```
+
+and it should result in something nice similar to
+
+```
+a_m_k: dim 2, lengths {3840, 4096}, strides {1, 4096}
+b_k_n: dim 2, lengths {4096, 4096}, strides {4096, 1}
+c_m_n: dim 2, lengths {3840, 4096}, strides {4096, 1}
+arg.a_grid_desc_k0_m0_m1_k1_{2048, 3840, 2}
+arg.b_grid_desc_k0_n0_n1_k1_{2048, 4096, 2}
+arg.c_grid_desc_m_n_{ 3840, 4096}
+launch_and_time_kernel: grid_dim {960, 1, 1}, block_dim {256, 1, 1}
+Warm up 1 time
+Start running 10 times...
+Perf: 3.65695 ms, 35.234 TFlops, 26.3797 GB/s, DeviceGemmDl<256, 128, 128, 16, 2, 4, 4, 1>
+```
+
+Or we can run a separate test
+
+```
+ctest -R test_gemm_fp16
+```
+
+If everything goes well you should see something like
+
+```
+Start 121: test_gemm_fp16
+1/1 Test #121: test_gemm_fp16 ...................   Passed   51.81 sec
+
+100% tests passed, 0 tests failed out of 1
+```
+
+## Summary
+
+In this tutorial we took the first look at the Composable Kernel library, built it on your system and ran some examples and tests. Stay tuned, in the next tutorial we will run kernels with different configs to find out the best one for your hardware and task.
+
+P.S.: Don't forget to switch out the cloud instance if you have launched one, you can find better ways to spend your money for sure!

example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp

@@ -4,7 +4,6 @@
 #include <iostream>
 #include <numeric>
 #include <initializer_list>
-#include <cstdlib>
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
@@ -116,7 +115,7 @@ auto f_host_tensor_descriptor2d =
     [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
         using namespace ck::literals;
 
-        if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        if constexpr(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
         {
             return HostTensorDescriptor({row, col}, {stride, 1_uz});
         }

example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp

@@ -4,7 +4,6 @@
 #include <iostream>
 #include <numeric>
 #include <initializer_list>
-#include <cstdlib>
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
@@ -15,6 +14,7 @@
 #include "ck/library/utility/device_memory.hpp"
 #include "ck/library/utility/host_tensor.hpp"
 #include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
 #include "ck/library/utility/check_err.hpp"
@@ -69,21 +69,20 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleDLayern
 // clang-format on
 
 auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
-    return HostTensorDescriptor(std::vector<std::size_t>({len}),
-                                std::vector<std::size_t>({stride}));
+    return HostTensorDescriptor({len}, {stride});
 };
 
 auto f_host_tensor_descriptor2d =
     [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
-        if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        using namespace ck::literals;
+
+        if constexpr(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
         {
-            return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                        std::vector<std::size_t>({stride, 1}));
+            return HostTensorDescriptor({row, col}, {stride, 1_uz});
         }
         else
         {
-            return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                        std::vector<std::size_t>({1, stride}));
+            return HostTensorDescriptor({row, col}, {1_uz, stride});
         }
     };
 
@@ -97,6 +96,7 @@ void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
                          AElementOp a_element_op,
                          BElementOp b_element_op,
                          CDEElementOp cde_element_op,
+                         HElementOp h_element_op,
                          int M,
                          int N,
                          AccDataType epsilon = 1e-5)
@@ -145,7 +145,7 @@ void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
     auto ref_layernorm_invoker = ref_layernorm.MakeInvoker();
 
     auto ref_layernorm_argument = ref_layernorm.MakeArgument(
-        e_m_n, gamma_n, beta_n, h_m_n, HElementOp{}, {M, N}, {1}, epsilon);
+        e_m_n, gamma_n, beta_n, h_m_n, h_element_op, {M, N}, {1}, epsilon);
     ref_layernorm_invoker.Run(ref_layernorm_argument);
 }
 
@@ -249,6 +249,7 @@ int main()
                             a_element_op,
                             b_element_op,
                             cde_element_op,
+                            h_element_op,
                             M,
                             N,
                             epsilon);

example/21_gemm_layernorm/gemm_layernorm_xdl_naive_fp16.cpp

@@ -4,7 +4,6 @@
 #include <iostream>
 #include <numeric>
 #include <initializer_list>
-#include <cstdlib>
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
@@ -115,7 +114,7 @@ auto f_host_tensor_descriptor2d =
     [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
         using namespace ck::literals;
 
-        if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+        if constexpr(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
         {
             return HostTensorDescriptor({row, col}, {stride, 1_uz});
         }

example/21_gemm_layernorm/gemm_xdl_layernorm_naive_single_kernel_fp16.cpp

@@ -135,7 +135,7 @@ int main(int argc, char* argv[])
         [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
             using namespace ck::literals;
 
-            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            if constexpr(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
             {
                 return HostTensorDescriptor({row, col}, {stride, 1_uz});
             }

example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt

+add_example_executable(example_gemm_bias_softmax_gemm_permute gemm_bias_softmax_gemm_permute.cpp)

example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp    = ck::tensor_operation::element_wise::PassThrough;
+using B0ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using C0DEElementOp = ck::tensor_operation::element_wise::ScaleAdd;
+using Acc0ElementOp = ck::tensor_operation::element_wise::PassThrough;
+using B1ElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CElementOp    = ck::tensor_operation::element_wise::PassThrough;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
+constexpr static auto MaskingSpec =
+    ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
+static constexpr auto TensorSpecA  = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
+static constexpr auto TensorSpecC  = ck::tensor_operation::device::TensorSpecialization::Default;
+
+using F16              = ck::half_t;
+using F32              = float;
+using ADataType        = F16;
+using B0DataType       = F16;
+using B1DataType       = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using CDataType        = F16;
+using D0DataType       = F16;
+using Acc0BiasDataType = ck::Tuple<D0DataType>;
+using Acc1BiasDataType = ck::Tuple<>;
+
+static constexpr ck::index_t NumDimG = 2;
+static constexpr ck::index_t NumDimM = 1;
+static constexpr ck::index_t NumDimN = 1;
+static constexpr ck::index_t NumDimK = 1;
+static constexpr ck::index_t NumDimO = 1;
+
+using DeviceOpInstance =
+    ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle<
+        NumDimG,
+        NumDimM,
+        NumDimN,
+        NumDimK,
+        NumDimO,
+        ADataType,
+        B0DataType,
+        B1DataType,
+        CDataType,
+        Acc0BiasDataType,
+        Acc1BiasDataType,
+        AccDataType,
+        CShuffleDataType,
+        AElementOp,
+        B0ElementOp,
+        C0DEElementOp,
+        B1ElementOp,
+        CElementOp,
+        GemmSpec,
+        TensorSpecA,
+        TensorSpecB0,
+        TensorSpecB1,
+        TensorSpecC,
+        1,
+        256,
+        128,         // MPerBlock
+        128,         // NPerBlock
+        32,          // KPerBlock
+        64,          // Gemm1NPerBlock
+        32,          // Gemm1KPerBlock
+        8,           // AK1
+        8,           // BK1
+        2,           // B1K1
+        32,          // MPerXDL
+        32,          // NPerXDL
+        1,           // MXdlPerWave
+        4,           // NXdlPerWave
+        2,           // Gemm1NXdlPerWave
+        S<4, 64, 1>, // ABlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<4, 64, 1>, // BBlockTransfer
+        S<1, 0, 2>,
+        S<1, 0, 2>,
+        2,
+        8,
+        8,
+        true,
+        S<16, 16, 1>, // B1BlockTransfer
+        S<0, 2, 1>,
+        S<0, 2, 1>,
+        1,
+        4,
+        2,
+        false,
+        1,              // CShuffleMXdlPerWavePerShuffle
+        2,              // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        8,              // CShuffleBlockTransferScalarPerVector_NPerBlock
+        MaskingSpec>;   // MaskingSpecialization
+
+// Ref Gemm0: fp16 in, fp32 out
+using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B0DataType,
+                                                                                AccDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B0ElementOp,
+                                                                                Acc0ElementOp>;
+
+// Ref Softmax: fp32 in, fp16 out
+using ReferenceSoftmaxInstance =
+    ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
+
+// Ref Gemm1: fp16 in, fp16 out
+using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
+                                                                                B1DataType,
+                                                                                CDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                B1ElementOp,
+                                                                                CElementOp>;
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    int G0      = 3;
+    int G1      = 2;
+    int M       = 1024;
+    int N       = 1024;
+    int K       = 64;
+    int O       = 64;
+    float alpha = 1;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 11)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M  = std::stoi(argv[4]);
+        N  = std::stoi(argv[5]);
+        K  = std::stoi(argv[6]);
+        O  = std::stoi(argv[7]);
+        G0 = std::stoi(argv[8]);
+        G1 = std::stoi(argv[9]);
+
+        alpha = std::stof(argv[10]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 11: M, N, K, O, G0, G1\n");
+        printf("arg10: scale (alpha)\n");
+        exit(0);
+    }
+
+    std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
+    std::vector<ck::index_t> a_gs_ms_ks_strides{
+        M * G1 * K, K, G1 * K, 1}; // A layout [G0, M, G1, K]
+
+    std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
+    std::vector<ck::index_t> b0_gs_ns_ks_strides{
+        N * G1 * K, K, G1 * K, 1}; // B0 layout [G0, N, G1, K]
+
+    std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
+    std::vector<ck::index_t> b1_gs_os_ns_strides{
+        N * G1 * O, O, 1, G1 * O}; // B1 layout [G0, N, G1, O]
+
+    std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
+    std::vector<ck::index_t> c_gs_ms_os_strides{
+        M * G1 * O, O, G1 * O, 1}; // C layout [G0, M, G1, O]
+
+    // D layout [G0, M, G1, N]
+    std::vector<ck::index_t> d0_gs_ms_ns_lengths{G0, G1, M, N};
+    std::vector<ck::index_t> d0_gs_ms_ns_strides{M * G1 * N, N, G1 * N, 1};
+
+    Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
+    Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
+    Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
+    Tensor<D0DataType> d0_gs_ms_ns(d0_gs_ms_ns_lengths, d0_gs_ms_ns_strides);
+    Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+    Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
+
+    std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
+    std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
+    std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
+    std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-2, 2});
+        break;
+    case 2:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-1, 1});
+        break;
+    case 3:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
+        break;
+    default:
+        a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
+        b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
+        b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
+        d0_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<D0DataType>{1});
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * G0 * G1 * M * K);
+    DeviceMem b0_device_buf(sizeof(B0DataType) * G0 * G1 * N * K);
+    DeviceMem d0_device_buf(sizeof(D0DataType) * G0 * G1 * M * N);
+    DeviceMem b1_device_buf(sizeof(B1DataType) * G0 * G1 * O * N);
+    DeviceMem c_device_buf(sizeof(CDataType) * G0 * G1 * M * O);
+
+    a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
+    b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
+    b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
+    d0_device_buf.ToDevice(d0_gs_ms_ns.mData.data());
+
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+
+    auto a_element_op    = AElementOp{};
+    auto b0_element_op   = B0ElementOp{};
+    auto c0de_element_op = C0DEElementOp{alpha};
+    auto acc0_element_op = Acc0ElementOp{};
+    auto b1_element_op   = B1ElementOp{};
+    auto c_element_op    = CElementOp{};
+
+    auto argument = device_op.MakeArgument(
+        static_cast<const ADataType*>(a_device_buf.GetDeviceBuffer()),
+        static_cast<const B0DataType*>(b0_device_buf.GetDeviceBuffer()),
+        static_cast<const B1DataType*>(b1_device_buf.GetDeviceBuffer()),
+        static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
+        std::array<void*, 1>{d0_device_buf.GetDeviceBuffer()}, // p_acc0_biases
+        {},                                                    // p_acc1_biases
+        a_gs_ms_ks_lengths,
+        a_gs_ms_ks_strides,
+        b0_gs_ns_ks_lengths,
+        b0_gs_ns_ks_strides,
+        b1_gs_os_ns_lengths,
+        b1_gs_os_ns_strides,
+        c_gs_ms_os_lengths,
+        c_gs_ms_os_strides,
+        std::array<std::vector<ck::index_t>, 1>{
+            d0_gs_ms_ns_lengths}, // acc0_biases_gs_ms_ns_lengths
+        std::array<std::vector<ck::index_t>, 1>{
+            d0_gs_ms_ns_strides}, // acc0_biases_gs_ms_ns_strides
+        {},                       // acc1_biases_gs_ms_os_lengths
+        {},                       // acc1_biases_gs_ms_os_strides
+        a_element_op,
+        b0_element_op,
+        c0de_element_op,
+        b1_element_op,
+        c_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error("wrong! this device_op instance does not support this problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    ck::index_t BatchCount = G0 * G1;
+    std::size_t flop       = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
+    std::size_t num_btype =
+        (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N + sizeof(B1DataType) * N * O +
+         sizeof(CDataType) * M * O + sizeof(D0DataType) * M * N) *
+        BatchCount;
+
+    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, "
+              << std::endl;
+
+    if(do_verification)
+    {
+        c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
+
+        Tensor<ADataType> a_g_m_k({BatchCount, M, K});
+        Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
+        Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
+        Tensor<AccDataType> acc0_g_m_n({BatchCount, M, N});        // scratch object after gemm0
+        Tensor<ADataType> a1_g_m_n({BatchCount, M, N});            // scratch object after softmax
+        Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
+        Tensor<D0DataType> d0_g_m_n({BatchCount, M, N});
+
+        // permute
+        a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
+            a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+        b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
+            b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
+            b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
+        });
+        d0_gs_ms_ns.ForEach([&](auto& self, auto idx) {
+            d0_g_m_n(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
+        });
+
+        // gemm 0
+        auto ref_gemm0          = ReferenceGemm0Instance{};
+        auto ref_gemm0_invoker  = ref_gemm0.MakeInvoker();
+        auto ref_gemm0_argument = ref_gemm0.MakeArgument(
+            a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
+
+        ref_gemm0_invoker.Run(ref_gemm0_argument);
+
+        acc0_g_m_n.ForEach([&](auto&, auto idx) {
+            c0de_element_op(acc0_g_m_n(idx), acc0_g_m_n(idx), d0_g_m_n(idx));
+        });
+        // masking
+        const auto mask = DeviceOpInstance::C0MatrixMask(N);
+        acc0_g_m_n.ForEach([&](auto& self, auto idx) {
+            if(mask.IsMaskedElement(idx[1], idx[2]))
+                self(idx) = -ck::NumericLimits<float>::Infinity();
+        });
+
+        // softmax
+        auto ref_softmax          = ReferenceSoftmaxInstance{};
+        auto ref_softmax_invoker  = ref_softmax.MakeInvoker();
+        auto ref_softmax_argument = ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2});
+
+        ref_softmax_invoker.Run(ref_softmax_argument);
+
+        // gemm1
+        auto ref_gemm1          = ReferenceGemm1Instance{};
+        auto ref_gemm1_invoker  = ref_gemm1.MakeInvoker();
+        auto ref_gemm1_argument = ref_gemm1.MakeArgument(
+            a1_g_m_n, b1_g_n_o, c_g_m_o_host_result, PassThrough{}, b1_element_op, c_element_op);
+
+        ref_gemm1_invoker.Run(ref_gemm1_argument);
+
+        // permute
+        c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
+            const size_t& g0 = idx[0];
+            const size_t& g1 = idx[1];
+
+            const size_t g = g0 * G1 + g1;
+
+            self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
+        });
+
+        // default absolute error and relative error is 0.001
+        double rtol = 1e-3;
+        double atol = 1e-3;
+
+        return ck::utils::check_err(c_gs_ms_os_device_result.mData,
+                                    c_gs_ms_os_host_result.mData,
+                                    "Error: Incorrect results!",
+                                    rtol,
+                                    atol)
+                   ? 0
+                   : 1;
+    }
+
+    return 0;
+}