CGEMM examples bf16, fp32, int8 (#332)

* Add int8 specialization for elementwise Add and Subtract.

* CGEMM examples bf16, fp32, int8

* Add convert reference output to CDataType.

* Skip BF16 data type during testing.

* Lower K value to get rid of accumulation error.

* Fix merge artifact.

* Fix changed function name: GetElementSpaceSize()

* Fix merge artifact.
Co-authored-by: Adam Osewski <aosewski@amd.com>

example/22_cgemm/CMakeLists.txt

+add_custom_target(example_cgemm_xdl)
+
+add_example_executable(example_cgemm_xdl_bf16 cgemm_xdl_bf16.cpp)
 add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)
+add_example_executable(example_cgemm_xdl_fp32 cgemm_xdl_fp32.cpp)
+add_example_executable(example_cgemm_xdl_int8 cgemm_xdl_int8.cpp)
+
+add_dependencies(example_cgemm_xdl example_cgemm_xdl_bf16)
+add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp16)
+add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp32)
+add_dependencies(example_cgemm_xdl example_cgemm_xdl_int8)

example/22_cgemm/cgemm_xdl_bf16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "cgemm_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/device_cgemm_4gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+using ADataType   = BF16;
+using BDataType   = BF16;
+using CDataType   = BF16;
+using AccDataType = F32;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using ReferenceCGemmInstance = ck::tensor_operation::host::
+    ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+
+// clang-format off
+using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
+    <ALayout,                    // typename ALayout
+     BLayout,                    // typename BLayout
+     CLayout,                    // typename CLayout
+     ADataType,                  // typename ADataType
+     BDataType,                  // typename BDataType
+     CDataType,                  // typename CDataType
+     AccDataType,                // typename GemmAccDataType
+     CDataType,                  // typename CShuffleDataType
+     PassThrough,                // typename AElementwiseOperation
+     PassThrough,                // typename BElementwiseOperation
+     PassThrough,                // typename CElementwiseOperation
+     GemmDefault,                // GemmSpecialization GemmSpec
+     1,                          // index_t NumGemmKPrefetchStage
+     256,                        // index_t BlockSize
+     256,                        // index_t MPerBlock
+     128,                        // index_t NPerBlock
+     32,                         // index_t KPerBlock
+     8,                          // index_t AK1
+     8,                          // index_t BK1
+     32,                         // index_t MPerXDL
+     32,                         // index_t NPerXDL
+     4,                          // index_t MXdlPerWave
+     2,                          // index_t NXdlPerWave
+     S<4, 64, 1>,                // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
+     S<1, 0, 2>,                 // typename ABlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename ABlockTransferSrcAccessOrder
+     2,                          // index_t ABlockTransferSrcVectorDim
+     8,                          // index_t ABlockTransferSrcScalarPerVector
+     8,                          // index_t ABlockTransferDstScalarPerVector_AK1
+     1,                          // index_t ABlockLdsExtraM
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder
+     2,                          // index_t BBlockTransferSrcVectorDim
+     8,                          // index_t BBlockTransferSrcScalarPerVector
+     8,                          // index_t BBlockTransferDstScalarPerVector_BK1
+     1,                          // index_t BBlockLdsExtraN
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle
+     S<1, 32, 1, 8>,             // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+     8>;                         // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // CGEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 416;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        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]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << "arg3: run kernel # of times (>1)\n"
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
+                  << std::endl;
+        exit(0);
+    }
+
+    return run_cgemm_xdl<ADataType,
+                         BDataType,
+                         CDataType,
+                         ALayout,
+                         BLayout,
+                         CLayout,
+                         PassThrough,
+                         PassThrough,
+                         PassThrough,
+                         DeviceCGemmInstance,
+                         ReferenceCGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
+}

example/22_cgemm/cgemm_xdl_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/stream_config.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/tensor_operation/gpu/device/tensor_layout.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16   = ck::half_t;
+using F32   = float;
+using BF16  = ck::bhalf_t;
+using INT8  = std::int8_t;
+using INT32 = std::int32_t;
+
+template <typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          typename DeviceCGemmInstance,
+          typename ReferenceCGemmInstance>
+int run_cgemm_xdl(ck::index_t M,
+                  ck::index_t N,
+                  ck::index_t K,
+                  ck::index_t StrideA,
+                  ck::index_t StrideB,
+                  ck::index_t StrideC,
+                  bool do_verification,
+                  int init_method,
+                  bool time_kernel)
+{
+    auto f_host_tensor_descriptor =
+        [](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)
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({stride, 1}));
+            }
+            else
+            {
+                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
+                                            std::vector<std::size_t>({1, stride}));
+            }
+        };
+
+    Tensor<ADataType> a_m_k_real(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<ADataType> a_m_k_imag(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n_real(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<BDataType> b_k_n_imag(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_real_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_imag_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k_real: " << a_m_k_real.mDesc << std::endl;
+    std::cout << "a_m_k_imag: " << a_m_k_imag.mDesc << std::endl;
+    std::cout << "b_k_n_real: " << b_k_n_real.mDesc << std::endl;
+    std::cout << "b_k_n_imag: " << b_k_n_imag.mDesc << std::endl;
+    std::cout << "c_m_n_real: " << c_m_n_real_device_result.mDesc << std::endl;
+    std::cout << "c_m_n_imag: " << c_m_n_imag_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k_real.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        a_m_k_imag.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
+        b_k_n_real.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        b_k_n_imag.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
+        break;
+    default:
+        a_m_k_real.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
+        a_m_k_imag.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
+        b_k_n_real.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        b_k_n_imag.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    auto cgemm = DeviceCGemmInstance{};
+
+    DeviceMem a_m_k_real_device_buf(sizeof(ADataType) * a_m_k_real.mDesc.GetElementSpaceSize());
+    DeviceMem a_m_k_imag_device_buf(sizeof(ADataType) * a_m_k_imag.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_real_device_buf(sizeof(BDataType) * b_k_n_real.mDesc.GetElementSpaceSize());
+    DeviceMem b_k_n_imag_device_buf(sizeof(BDataType) * b_k_n_imag.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_real_device_buf(sizeof(CDataType) *
+                                    c_m_n_real_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_imag_device_buf(sizeof(CDataType) *
+                                    c_m_n_imag_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem workspace_device_buf(cgemm.GetWorkspaceSize(M, N, K, StrideA, StrideB, StrideC));
+
+    a_m_k_real_device_buf.ToDevice(a_m_k_real.mData.data());
+    a_m_k_imag_device_buf.ToDevice(a_m_k_imag.mData.data());
+    b_k_n_real_device_buf.ToDevice(b_k_n_real.mData.data());
+    b_k_n_imag_device_buf.ToDevice(b_k_n_imag.mData.data());
+
+    auto a_element_op = AElementwiseOperation{};
+    auto b_element_op = BElementwiseOperation{};
+    auto c_element_op = CElementwiseOperation{};
+
+    // do GEMM
+    auto invoker = cgemm.MakeInvoker();
+    auto argument =
+        cgemm.MakeArgument(static_cast<ADataType*>(a_m_k_real_device_buf.GetDeviceBuffer()),
+                           static_cast<ADataType*>(a_m_k_imag_device_buf.GetDeviceBuffer()),
+                           static_cast<BDataType*>(b_k_n_real_device_buf.GetDeviceBuffer()),
+                           static_cast<BDataType*>(b_k_n_imag_device_buf.GetDeviceBuffer()),
+                           static_cast<CDataType*>(c_m_n_real_device_buf.GetDeviceBuffer()),
+                           static_cast<CDataType*>(c_m_n_imag_device_buf.GetDeviceBuffer()),
+                           static_cast<CDataType*>(workspace_device_buf.GetDeviceBuffer()),
+                           M,
+                           N,
+                           K,
+                           StrideA,
+                           StrideB,
+                           StrideC,
+                           a_element_op,
+                           b_element_op,
+                           c_element_op);
+
+    if(!cgemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_cgemm with the specified compilation parameters does "
+            "not support this CGEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(8) * M * N * K;
+    std::size_t num_btype =
+        std::size_t(2) *
+        (sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + 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, "
+              << cgemm.GetTypeString() << std::endl;
+
+    c_m_n_real_device_buf.FromDevice(c_m_n_real_device_result.mData.data());
+    c_m_n_imag_device_buf.FromDevice(c_m_n_imag_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CDataType> c_m_n_real_host_result(
+            f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+        Tensor<CDataType> c_m_n_imag_host_result(
+            f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+        auto ref_cgemm   = ReferenceCGemmInstance{};
+        auto ref_invoker = ref_cgemm.MakeInvoker();
+
+        auto ref_argument = ref_cgemm.MakeArgument(a_m_k_real,
+                                                   a_m_k_imag,
+                                                   b_k_n_real,
+                                                   b_k_n_imag,
+                                                   c_m_n_real_host_result,
+                                                   c_m_n_imag_host_result,
+                                                   a_element_op,
+                                                   b_element_op,
+                                                   c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        bool result = true;
+        result      = ck::utils::check_err(c_m_n_real_device_result.mData,
+                                      c_m_n_real_host_result.mData,
+                                      "Verification error: incorrect results in real part!",
+                                      1e-2f,
+                                      1e-1f);
+        result      = result &&
+                 ck::utils::check_err(c_m_n_imag_device_result.mData,
+                                      c_m_n_imag_host_result.mData,
+                                      "Verification error: incorrect results in imaginary part!",
+                                      1e-2f,
+                                      1e-1f);
+        return result ? 0 : 1;
+    }
+    return 0;
+}

example/22_cgemm/cgemm_xdl_fp16.cpp

@@ -2,43 +2,30 @@
 // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
 
 #include <iostream>
-#include <numeric>
-#include <initializer_list>
-#include <cstdlib>
 
-#include "ck/ck.hpp"
+#include "cgemm_xdl_common.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
-#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
 #include "ck/tensor_operation/gpu/device/device_cgemm_4gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.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/reference_tensor_operation/cpu/reference_cgemm.hpp"
-
-template <ck::index_t... Is>
-using S = ck::Sequence<Is...>;
-
-using F16 = ck::half_t;
-using F32 = float;
-
-using Row = ck::tensor_layout::gemm::RowMajor;
-using Col = ck::tensor_layout::gemm::ColumnMajor;
-
-using PassThrough = ck::tensor_operation::element_wise::PassThrough;
-
-using ADataType   = F16;
-using BDataType   = F16;
-using CDataType   = F16;
-using AccDataType = F32;
+using ADataType        = F16;
+using BDataType        = F16;
+using CDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
 
 using ALayout = ck::tensor_layout::gemm::RowMajor;
 using BLayout = ck::tensor_layout::gemm::ColumnMajor;
 using CLayout = ck::tensor_layout::gemm::RowMajor;
 
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
 static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
 
+using ReferenceCGemmInstance = ck::tensor_operation::host::
+    ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+
 // clang-format off
 using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
     <ALayout,                    // typename ALayout
@@ -48,7 +35,7 @@ using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_
      BDataType,                  // typename BDataType
      CDataType,                  // typename CDataType
      AccDataType,                // typename GemmAccDataType
-     CDataType,                  // typename CShuffleDataType
+     CShuffleDataType,           // typename CShuffleDataType
      PassThrough,                // typename AElementwiseOperation
      PassThrough,                // typename BElementwiseOperation
      PassThrough,                // typename CElementwiseOperation
@@ -84,9 +71,6 @@ using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_
      8>;                         // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
 // clang-format on
 
-using ReferenceCGemmInstance = ck::tensor_operation::host::
-    ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
-
 int main(int argc, char* argv[])
 {
     bool do_verification = true;
@@ -124,155 +108,24 @@ int main(int argc, char* argv[])
     }
     else
     {
-        printf("arg1: verification (0=no, 1=yes)\n");
-        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
-        printf("arg3: run kernel # of times (>1)\n");
-        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << "arg3: run kernel # of times (>1)\n"
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
+                  << std::endl;
         exit(0);
     }
 
-    auto f_host_tensor_descriptor =
-        [](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)
-            {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({stride, 1}));
-            }
-            else
-            {
-                return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
-                                            std::vector<std::size_t>({1, stride}));
-            }
-        };
-
-    Tensor<ADataType> a_m_k_real(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
-    Tensor<ADataType> a_m_k_imag(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
-    Tensor<BDataType> b_k_n_real(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
-    Tensor<BDataType> b_k_n_imag(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
-    Tensor<CDataType> c_m_n_real_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
-    Tensor<CDataType> c_m_n_imag_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
-
-    std::cout << "a_m_k_real: " << a_m_k_real.mDesc << std::endl;
-    std::cout << "a_m_k_imag: " << a_m_k_imag.mDesc << std::endl;
-    std::cout << "b_k_n_real: " << b_k_n_real.mDesc << std::endl;
-    std::cout << "b_k_n_imag: " << b_k_n_imag.mDesc << std::endl;
-    std::cout << "c_m_n_real: " << c_m_n_real_device_result.mDesc << std::endl;
-    std::cout << "c_m_n_imag: " << c_m_n_imag_device_result.mDesc << std::endl;
-
-    switch(init_method)
-    {
-    case 0: break;
-    case 1:
-        a_m_k_real.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
-        a_m_k_imag.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
-        b_k_n_real.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
-        b_k_n_imag.GenerateTensorValue(GeneratorTensor_2<BDataType>{-2, 2});
-        break;
-    default:
-        a_m_k_real.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
-        a_m_k_imag.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
-        b_k_n_real.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
-        b_k_n_imag.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
-    }
-
-    auto cgemm = DeviceCGemmInstance{};
-
-    DeviceMem a_m_k_real_device_buf(sizeof(ADataType) * a_m_k_real.mDesc.GetElementSpaceSize());
-    DeviceMem a_m_k_imag_device_buf(sizeof(ADataType) * a_m_k_imag.mDesc.GetElementSpaceSize());
-    DeviceMem b_k_n_real_device_buf(sizeof(BDataType) * b_k_n_real.mDesc.GetElementSpaceSize());
-    DeviceMem b_k_n_imag_device_buf(sizeof(BDataType) * b_k_n_imag.mDesc.GetElementSpaceSize());
-    DeviceMem c_m_n_real_device_buf(sizeof(CDataType) *
-                                    c_m_n_real_device_result.mDesc.GetElementSpaceSize());
-    DeviceMem c_m_n_imag_device_buf(sizeof(CDataType) *
-                                    c_m_n_imag_device_result.mDesc.GetElementSpaceSize());
-    DeviceMem workspace_device_buf(cgemm.GetWorkspaceSize(M, N, K, StrideA, StrideB, StrideC));
-
-    a_m_k_real_device_buf.ToDevice(a_m_k_real.mData.data());
-    a_m_k_imag_device_buf.ToDevice(a_m_k_imag.mData.data());
-    b_k_n_real_device_buf.ToDevice(b_k_n_real.mData.data());
-    b_k_n_imag_device_buf.ToDevice(b_k_n_imag.mData.data());
-
-    auto a_element_op = PassThrough{};
-    auto b_element_op = PassThrough{};
-    auto c_element_op = PassThrough{};
-
-    // do GEMM
-    auto invoker = cgemm.MakeInvoker();
-    auto argument =
-        cgemm.MakeArgument(static_cast<ADataType*>(a_m_k_real_device_buf.GetDeviceBuffer()),
-                           static_cast<ADataType*>(a_m_k_imag_device_buf.GetDeviceBuffer()),
-                           static_cast<BDataType*>(b_k_n_real_device_buf.GetDeviceBuffer()),
-                           static_cast<BDataType*>(b_k_n_imag_device_buf.GetDeviceBuffer()),
-                           static_cast<CDataType*>(c_m_n_real_device_buf.GetDeviceBuffer()),
-                           static_cast<CDataType*>(c_m_n_imag_device_buf.GetDeviceBuffer()),
-                           static_cast<CDataType*>(workspace_device_buf.GetDeviceBuffer()),
-                           M,
-                           N,
-                           K,
-                           StrideA,
-                           StrideB,
-                           StrideC,
-                           a_element_op,
-                           b_element_op,
-                           c_element_op);
-
-    if(!cgemm.IsSupportedArgument(argument))
-    {
-        throw std::runtime_error(
-            "wrong! device_cgemm with the specified compilation parameters does "
-            "not support this CGEMM problem");
-    }
-
-    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
-
-    std::size_t flop = std::size_t(8) * M * N * K;
-    std::size_t num_btype =
-        std::size_t(2) *
-        (sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + 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, "
-              << cgemm.GetTypeString() << std::endl;
-
-    c_m_n_real_device_buf.FromDevice(c_m_n_real_device_result.mData.data());
-    c_m_n_imag_device_buf.FromDevice(c_m_n_imag_device_result.mData.data());
-
-    if(do_verification)
-    {
-        Tensor<CDataType> c_m_n_real_host_result(
-            f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
-        Tensor<CDataType> c_m_n_imag_host_result(
-            f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
-
-        auto ref_cgemm   = ReferenceCGemmInstance{};
-        auto ref_invoker = ref_cgemm.MakeInvoker();
-
-        auto ref_argument = ref_cgemm.MakeArgument(a_m_k_real,
-                                                   a_m_k_imag,
-                                                   b_k_n_real,
-                                                   b_k_n_imag,
-                                                   c_m_n_real_host_result,
-                                                   c_m_n_imag_host_result,
-                                                   a_element_op,
-                                                   b_element_op,
-                                                   c_element_op);
-
-        ref_invoker.Run(ref_argument);
-
-        ck::utils::check_err(c_m_n_real_device_result.mData,
-                             c_m_n_real_host_result.mData,
-                             "Verification error: incorrect results in real part!",
-                             1e-2f,
-                             1e-1f);
-        ck::utils::check_err(c_m_n_imag_device_result.mData,
-                             c_m_n_imag_host_result.mData,
-                             "Verification error: incorrect results in imaginary part!",
-                             1e-2f,
-                             1e-1f);
-    }
-
-    return 0;
+    return run_cgemm_xdl<ADataType,
+                         BDataType,
+                         CDataType,
+                         ALayout,
+                         BLayout,
+                         CLayout,
+                         PassThrough,
+                         PassThrough,
+                         PassThrough,
+                         DeviceCGemmInstance,
+                         ReferenceCGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
 }

example/22_cgemm/cgemm_xdl_fp32.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "cgemm_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/device_cgemm_4gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+using ADataType   = F32;
+using BDataType   = F32;
+using CDataType   = F32;
+using AccDataType = F32;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using ReferenceCGemmInstance = ck::tensor_operation::host::
+    ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+
+// clang-format off
+using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
+    <ALayout,                    // typename ALayout
+     BLayout,                    // typename BLayout
+     CLayout,                    // typename CLayout
+     ADataType,                  // typename ADataType
+     BDataType,                  // typename BDataType
+     CDataType,                  // typename CDataType
+     AccDataType,                // typename GemmAccDataType
+     CDataType,                  // typename CShuffleDataType
+     PassThrough,                // typename AElementwiseOperation
+     PassThrough,                // typename BElementwiseOperation
+     PassThrough,                // typename CElementwiseOperation
+     GemmDefault,                // GemmSpecialization GemmSpec
+     1,                          // index_t NumGemmKPrefetchStage
+     256,                        // index_t BlockSize
+     256,                        // index_t MPerBlock
+     128,                        // index_t NPerBlock
+     16,                         // index_t KPerBlock
+     4,                          // index_t AK1
+     4,                          // index_t BK1
+     32,                         // index_t MPerXDL
+     32,                         // index_t NPerXDL
+     4,                          // index_t MXdlPerWave
+     2,                          // index_t NXdlPerWave
+     S<4, 64, 1>,                // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
+     S<1, 0, 2>,                 // typename ABlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename ABlockTransferSrcAccessOrder
+     2,                          // index_t ABlockTransferSrcVectorDim
+     4,                          // index_t ABlockTransferSrcScalarPerVector
+     4,                          // index_t ABlockTransferDstScalarPerVector_AK1
+     1,                          // index_t ABlockLdsExtraM
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder
+     2,                          // index_t BBlockTransferSrcVectorDim
+     4,                          // index_t BBlockTransferSrcScalarPerVector
+     4,                          // index_t BBlockTransferDstScalarPerVector_BK1
+     1,                          // index_t BBlockLdsExtraN
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle
+     S<1, 16, 1, 16>,            // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+     4>;                         // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // CGEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        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]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << "arg3: run kernel # of times (>1)\n"
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
+                  << std::endl;
+        exit(0);
+    }
+
+    return run_cgemm_xdl<ADataType,
+                         BDataType,
+                         CDataType,
+                         ALayout,
+                         BLayout,
+                         CLayout,
+                         PassThrough,
+                         PassThrough,
+                         PassThrough,
+                         DeviceCGemmInstance,
+                         ReferenceCGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
+}

example/22_cgemm/cgemm_xdl_int8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "cgemm_xdl_common.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/device_cgemm_4gemm_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
+using ADataType   = INT8;
+using BDataType   = INT8;
+using CDataType   = INT8;
+using AccDataType = INT32;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using ReferenceCGemmInstance = ck::tensor_operation::host::
+    ReferenceCGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+
+// clang-format off
+using DeviceCGemmInstance = ck::tensor_operation::device::DeviceCGemm_4Gemm_Xdl_CShuffle
+    <ALayout,                    // typename ALayout
+     BLayout,                    // typename BLayout
+     CLayout,                    // typename CLayout
+     ADataType,                  // typename ADataType
+     BDataType,                  // typename BDataType
+     CDataType,                  // typename CDataType
+     AccDataType,                // typename GemmAccDataType
+     CDataType,                  // typename CShuffleDataType
+     PassThrough,                // typename AElementwiseOperation
+     PassThrough,                // typename BElementwiseOperation
+     PassThrough,                // typename CElementwiseOperation
+     GemmDefault,                // GemmSpecialization GemmSpec
+     1,                          // index_t NumGemmKPrefetchStage
+     256,                        // index_t BlockSize
+     256,                        // index_t MPerBlock
+     128,                        // index_t NPerBlock
+     64,                         // index_t KPerBlock
+     16,                         // index_t AK1
+     16,                         // index_t BK1
+     32,                         // index_t MPerXDL
+     32,                         // index_t NPerXDL
+     4,                          // index_t MXdlPerWave
+     2,                          // index_t NXdlPerWave
+     S<4, 64, 1>,                // typename ABlockTransferThreadClusterLengths_AK0_M_AK1
+     S<1, 0, 2>,                 // typename ABlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename ABlockTransferSrcAccessOrder
+     2,                          // index_t ABlockTransferSrcVectorDim
+     16,                         // index_t ABlockTransferSrcScalarPerVector
+     16,                         // index_t ABlockTransferDstScalarPerVector_AK1
+     1,                          // index_t ABlockLdsExtraM
+     S<4, 64, 1>,                // typename BBlockTransferThreadClusterLengths_BK0_N_BK1
+     S<1, 0, 2>,                 // typename BBlockTransferThreadClusterArrangeOrder
+     S<1, 0, 2>,                 // typename BBlockTransferSrcAccessOrder
+     2,                          // index_t BBlockTransferSrcVectorDim
+     8,                          // index_t BBlockTransferSrcScalarPerVector
+     8,                          // index_t BBlockTransferDstScalarPerVector_BK1
+     1,                          // index_t BBlockLdsExtraN
+     1,                          // index_t CShuffleMXdlPerWavePerShuffle
+     1,                          // index_t CShuffleNXdlPerWavePerShuffle
+     S<1, 64, 1, 4>,             // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+     16>;                        // index_t CShuffleBlockTransferScalarPerVector_NPerBlock
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // CGEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideC = 4096;
+
+    if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        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]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideC = std::stoi(argv[9]);
+    }
+    else
+    {
+        std::cout << "arg1: verification (0=no, 1=yes)\n"
+                  << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+                  << "arg3: run kernel # of times (>1)\n"
+                  << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n"
+                  << std::endl;
+        exit(0);
+    }
+
+    return run_cgemm_xdl<ADataType,
+                         BDataType,
+                         CDataType,
+                         ALayout,
+                         BLayout,
+                         CLayout,
+                         PassThrough,
+                         PassThrough,
+                         PassThrough,
+                         DeviceCGemmInstance,
+                         ReferenceCGemmInstance>(
+        M, N, K, StrideA, StrideB, StrideC, do_verification, init_method, time_kernel);
+}

include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp

@@ -51,6 +51,13 @@ struct Add
         const float y_tmp  = x1_tmp + x2_tmp;
         y                  = ck::type_convert<bhalf_t>(y_tmp);
     }
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<int8_t>(int8_t& y, const int8_t& x0, const int8_t& x1) const
+    {
+        y = x0 + x1;
+    };
 };
 
 struct Subtract
@@ -88,6 +95,13 @@ struct Subtract
         const float y_tmp  = x1_tmp - x2_tmp;
         y                  = ck::type_convert<bhalf_t>(y_tmp);
     }
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<int8_t>(int8_t& y, const int8_t& x0, const int8_t& x1) const
+    {
+        y = x0 - x1;
+    };
 };
 
 struct Bilinear

include/ck/utility/dynamic_buffer.hpp

@@ -4,6 +4,7 @@
 #pragma once
 
 #include "ck/ck.hpp"
+#include "ck/utility/data_type.hpp"
 #include "enable_if.hpp"
 #include "c_style_pointer_cast.hpp"
 #include "amd_buffer_addressing.hpp"

library/include/ck/library/reference_tensor_operation/cpu/reference_cgemm.hpp

@@ -6,8 +6,9 @@
 #include <iostream>
 #include <sstream>
 
-#include "ck/tensor_operation/gpu/device/device_base.hpp"
 #include "ck/library/utility/host_tensor.hpp"
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 namespace ck {
 namespace tensor_operation {
@@ -91,7 +92,7 @@ struct ReferenceCGemm : public device::BaseOperator
                     v_c_real += v_a_real * v_b_real - v_a_imag * v_b_imag;
                 }
 
-                arg.c_m_n_real_(m, n) = v_c_real;
+                arg.c_m_n_real_(m, n) = ck::type_convert<CDataType>(v_c_real);
             };
 
             auto f_mk_kn_mn_imag = [&](auto m, auto n) {
@@ -107,7 +108,7 @@ struct ReferenceCGemm : public device::BaseOperator
                     v_c_imag += v_a_real * v_b_imag + v_a_imag * v_b_real;
                 }
 
-                arg.c_m_n_imag_(m, n) = v_c_imag;
+                arg.c_m_n_imag_(m, n) = ck::type_convert<CDataType>(v_c_imag);
             };
 
             make_ParallelTensorFunctor(f_mk_kn_mn_real,