[Bf16 & int8] [example & ckprofiler] (#100)

* Add int8 of mk_nk_mn to the ckProfiler

* Add example of int8 gemm

* Fix typo, use ushort instead of half_t for bfloat16

* replace ushortXXX_t to bhalfXXX_t

* rename ushort to bhalf_t

* Add bf16 example

* Add bf16 gemm to ckProfiler

* Fix alignment

* Fix typo

* Add unit test for gemm_xdl int8

* Add gemm_xdl fp32 unit test

* Add gemm_xdl bf16 unit test

* fix build

* fix build issue due to merge conflict

* Fix build

* Fix build error
Co-authored-by: rocking <chunylai@amd.com>
Co-authored-by: Chao Liu <chao.liu2@amd.com>

profiler/src/profile_gemm.cpp

@@ -20,8 +20,10 @@ enum GemmMatrixLayout
 
 enum GemmDataType
 {
-    F32_F32_F32, // 0
-    F16_F16_F16, // 1
+    F32_F32_F32,    // 0
+    F16_F16_F16,    // 1
+    BF16_BF16_BF16, // 2
+    INT8_INT8_INT8, // 3
 };
 
 int profile_gemm(int argc, char* argv[])
@@ -29,7 +31,7 @@ int profile_gemm(int argc, char* argv[])
     if(!(argc == 14 || argc == 15))
     {
         printf("arg1: tensor operation (gemm: GEMM)\n");
-        printf("arg2: data type (0: fp32; 1: fp16)\n");
+        printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
         printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
         printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");
         printf("                     2: A[k, m] * B[k, n] = C[m, n];\n");
@@ -221,6 +223,46 @@ int profile_gemm(int argc, char* argv[])
             (StrideC < 0) ? N : StrideC,
             KBatch);
     }
+    else if(data_type == GemmDataType::INT8_INT8_INT8 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        ck::profiler::profile_gemm_impl<int8_t,
+                                        int8_t,
+                                        int8_t,
+                                        ck::tensor_layout::gemm::RowMajor,
+                                        ck::tensor_layout::gemm::ColumnMajor,
+                                        ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            nrepeat,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? K : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            KBatch);
+    }
+    else if(data_type == GemmDataType::BF16_BF16_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        ck::profiler::profile_gemm_impl<ck::bhalf_t,
+                                        ck::bhalf_t,
+                                        ck::bhalf_t,
+                                        ck::tensor_layout::gemm::RowMajor,
+                                        ck::tensor_layout::gemm::ColumnMajor,
+                                        ck::tensor_layout::gemm::RowMajor>(
+            do_verification,
+            init_method,
+            do_log,
+            nrepeat,
+            M,
+            N,
+            K,
+            (StrideA < 0) ? M : StrideA,
+            (StrideB < 0) ? K : StrideB,
+            (StrideC < 0) ? N : StrideC,
+            KBatch);
+    }
     else
     {
         throw std::runtime_error("wrong! this GEMM data_type & layout is not implemented");

profiler/src/profile_gemm_bias_2d.cpp

@@ -28,7 +28,7 @@ int profile_gemm_bias_2d(int argc, char* argv[])
 {
     if(!(argc == 16 || argc == 17))
     {
-        printf("arg1: tensor operation (gemm: GEMM+Bias)\n");
+        printf("arg1: tensor operation (gemm: GEMM+Bias_2d)\n");
         printf("arg2: data type (0: fp32; 1: fp16)\n");
         printf("arg3: matrix layout (0: A[m, k] * B[k, n] = C[m, n];\n");
         printf("                     1: A[m, k] * B[n, k] = C[m, n];\n");

test/CMakeLists.txt

@@ -34,3 +34,21 @@ foreach(TEST ${TESTS})
     message("adding test ${BASE_NAME}")
     add_test_executeable(test_${BASE_NAME} ${TEST})
 endforeach(TEST ${TESTS})
+
+# test_gemm_xdl_fp32
+set(GEMM_XDL_FP32_SOURCE gemm_xdl/test_gemm_fp32.cpp)
+add_executable(test_gemm_xdl_fp32 ${GEMM_XDL_FP32_SOURCE})
+target_link_libraries(test_gemm_xdl_fp32 PRIVATE host_tensor)
+target_link_libraries(test_gemm_xdl_fp32 PRIVATE device_gemm_instance)
+
+# test_gemm_xdl_bf16
+set(GEMM_XDL_BF16_SOURCE gemm_xdl/test_gemm_bf16.cpp)
+add_executable(test_gemm_xdl_bf16 ${GEMM_XDL_BF16_SOURCE})
+target_link_libraries(test_gemm_xdl_bf16 PRIVATE host_tensor)
+target_link_libraries(test_gemm_xdl_bf16 PRIVATE device_gemm_instance)
+
+# test_gemm_xdl_int8
+set(GEMM_XDL_INT8_SOURCE gemm_xdl/test_gemm_int8.cpp)
+add_executable(test_gemm_xdl_int8 ${GEMM_XDL_INT8_SOURCE})
+target_link_libraries(test_gemm_xdl_int8 PRIVATE host_tensor)
+target_link_libraries(test_gemm_xdl_int8 PRIVATE device_gemm_instance)

test/conv2d_fwd.cpp

@@ -202,9 +202,9 @@ int main(int argc, char* argv[])
             ck::tensor_operation::device::device_conv2d_fwd_instance::
                 add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
         }
-        else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ushort> &&
-                          ck::is_same_v<ck::remove_cv_t<WeiDataType>, ushort> &&
-                          ck::is_same_v<ck::remove_cv_t<OutDataType>, ushort>)
+        else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::bhalf_t> &&
+                          ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::bhalf_t> &&
+                          ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::bhalf_t>)
         {
             ck::tensor_operation::device::device_conv2d_fwd_instance::
                 add_device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
@@ -298,7 +298,7 @@ int main(int argc, char* argv[])
     }
     else if(data_type == 2)
     {
-        res = Run(ushort(), ushort(), ushort());
+        Run(ck::bhalf_t(), ck::bhalf_t(), ck::bhalf_t());
     }
     else if(data_type == 3)
     {

test/gemm_xdl/gemm_util.hpp

+#ifndef GEMM_UTILS_HPP
+#define GEMM_UTILS_HPP
+
+#include "config.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+
+namespace ck {
+namespace gemm_util {
+
+struct GemmParams
+{
+    GemmParams()
+        : M(1024), N(1024), K(1024), StrideA(1024), StrideB(1024), StrideC(1024), alpha(1), beta(0)
+    {
+    }
+
+    ck::index_t M;
+    ck::index_t N;
+    ck::index_t K;
+
+    ck::index_t StrideA;
+    ck::index_t StrideB;
+    ck::index_t StrideC;
+
+    float alpha;
+    float beta;
+};
+
+template <typename GemmInstance,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+void RunHostGEMM(const Tensor<ADataType>& A,
+                 const Tensor<BDataType>& B,
+                 Tensor<CDataType>& C,
+                 AElementwiseOperation a_element_op,
+                 BElementwiseOperation b_element_op,
+                 CElementwiseOperation c_element_op)
+{
+    auto ref_gemm    = GemmInstance{};
+    auto ref_invoker = ref_gemm.MakeInvoker();
+
+    auto ref_argument = ref_gemm.MakeArgument(A, B, C, a_element_op, b_element_op, c_element_op);
+
+    ref_invoker.Run(ref_argument);
+}
+
+template <typename DeviceGemmPtr_,
+          typename ADataType,
+          typename BDataType,
+          typename CDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation>
+void RunDeviceGEMM(DeviceGemmPtr_& gemmPtr,
+                   const ck::gemm_util::GemmParams& params,
+                   const Tensor<ADataType>& A,
+                   const Tensor<BDataType>& B,
+                   Tensor<CDataType>& C,
+                   AElementwiseOperation a_element_op,
+                   BElementwiseOperation b_element_op,
+                   CElementwiseOperation c_element_op)
+{
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * A.mDesc.GetElementSpace());
+    DeviceMem b_k_n_device_buf(sizeof(BDataType) * B.mDesc.GetElementSpace());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * C.mDesc.GetElementSpace());
+
+    a_m_k_device_buf.ToDevice(A.mData.data());
+    b_k_n_device_buf.ToDevice(B.mData.data());
+
+    auto invoker_ptr = gemmPtr->MakeInvokerPointer();
+    auto argument_ptr =
+        gemmPtr->MakeArgumentPointer(static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+                                     static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+                                     static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()),
+                                     params.M,
+                                     params.N,
+                                     params.K,
+                                     params.StrideA,
+                                     params.StrideB,
+                                     params.StrideC,
+                                     a_element_op,
+                                     b_element_op,
+                                     c_element_op);
+
+    if(!gemmPtr->IsSupportedArgument(argument_ptr.get()))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    invoker_ptr->Run(argument_ptr.get());
+    c_m_n_device_buf.FromDevice(C.mData.data());
+}
+
+} // namespace gemm_util
+} // namespace ck
+#endif

test/gemm_xdl/test_gemm_bf16.cpp

+#include <algorithm>
+#include <cstdlib>
+#include <half.hpp>
+#include <iostream>
+#include <numeric>
+#include <tuple>
+#include <vector>
+
+#include "gemm_util.hpp"
+#include "config.hpp"
+#include "print.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "host_gemm.hpp"
+#include "device_tensor.hpp"
+#include "device_gemm_xdl.hpp"
+#include "device_gemm_xdl_c_shuffle.hpp"
+#include "element_wise_operation.hpp"
+#include "reference_gemm.hpp"
+#include "gemm_specialization.hpp"
+#include "test_util.hpp"
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using DeviceGemmPtr_ =
+    ck::tensor_operation::device::DeviceGemmPtr<ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::PassThrough>;
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace device_gemm_instance {
+void add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(std::vector<DeviceGemmPtr_>&);
+}
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace {
+
+using BF16 = ck::bhalf_t;
+
+using ADataType   = BF16;
+using BDataType   = BF16;
+using CDataType   = BF16;
+using AccDataType = float;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+auto PrepareGemmTensor(const ck::gemm_util::GemmParams& params)
+{
+    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}));
+            }
+        };
+
+    // use fp32 host kernel to verify bf16 device kernel
+    Tensor<ADataType> a_m_k_bf16(
+        f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n_bf16(
+        f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_device_bf16(
+        f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
+
+    Tensor<float> a_m_k_fp32(
+        f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
+    Tensor<float> b_k_n_fp32(
+        f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
+    Tensor<float> c_m_n_host_fp32(
+        f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
+    Tensor<float> c_m_n_device_fp32(
+        f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
+
+    a_m_k_bf16.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
+    b_k_n_bf16.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+
+    bf16_to_f32_(a_m_k_bf16, a_m_k_fp32);
+    bf16_to_f32_(b_k_n_bf16, b_k_n_fp32);
+
+    return std::make_tuple(a_m_k_bf16,
+                           b_k_n_bf16,
+                           c_m_n_device_bf16,
+                           a_m_k_fp32,
+                           b_k_n_fp32,
+                           c_m_n_host_fp32,
+                           c_m_n_device_fp32);
+}
+
+bool TestGemm(DeviceGemmPtr_& gemmPtr)
+{
+    // Arrange
+    ck::gemm_util::GemmParams params;
+    params.M       = 1024;
+    params.N       = 1024;
+    params.K       = 1024;
+    params.StrideA = 1024;
+    params.StrideB = 1024;
+    params.StrideC = 1024;
+
+    auto host_tensors                = PrepareGemmTensor(params);
+    const Tensor<ADataType>& a_bf16  = std::get<0>(host_tensors);
+    const Tensor<BDataType>& b_bf16  = std::get<1>(host_tensors);
+    Tensor<CDataType>& c_device_bf16 = std::get<2>(host_tensors);
+    Tensor<float>& a_fp32            = std::get<3>(host_tensors);
+    Tensor<float>& b_fp32            = std::get<4>(host_tensors);
+    Tensor<float>& c_host_fp32       = std::get<5>(host_tensors);
+    Tensor<float>& c_device_fp32     = std::get<6>(host_tensors);
+
+    auto a_element_op = PassThrough{};
+    auto b_element_op = PassThrough{};
+    auto c_element_op = PassThrough{};
+
+    // use fp32 host kernel to verify bf16 device kernel
+    using ReferenceGemmInstance = ck::tensor_operation::host::
+        ReferenceGemm<float, float, float, PassThrough, PassThrough, PassThrough>;
+    ck::gemm_util::RunHostGEMM<ReferenceGemmInstance>(
+        a_fp32, b_fp32, c_host_fp32, a_element_op, b_element_op, c_element_op);
+
+    // Act
+    ck::gemm_util::RunDeviceGEMM(
+        gemmPtr, params, a_bf16, b_bf16, c_device_bf16, a_element_op, b_element_op, c_element_op);
+
+    bf16_to_f32_(c_device_bf16, c_device_fp32);
+
+    // Assert
+    bool res = test_util::check_err(
+        c_device_fp32.mData, c_host_fp32.mData, "Error: incorrect results!", 1e-2f, 1e-3f);
+
+    std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+
+    return res;
+}
+
+} // anonymous namespace
+
+int main()
+{
+    std::vector<DeviceGemmPtr_> gemmPtrs;
+    ck::tensor_operation::device::device_gemm_instance::
+        add_device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances(gemmPtrs);
+
+    bool res = true;
+
+    for(auto& gemmPtr : gemmPtrs)
+    {
+        res &= TestGemm(gemmPtr);
+    }
+
+    std::cout << "TestGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
+}

test/gemm_xdl/test_gemm_fp32.cpp

+#include <algorithm>
+#include <cstdlib>
+#include <half.hpp>
+#include <iostream>
+#include <numeric>
+#include <tuple>
+#include <vector>
+
+#include "gemm_util.hpp"
+#include "config.hpp"
+#include "print.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "host_gemm.hpp"
+#include "device_tensor.hpp"
+#include "device_gemm_xdl.hpp"
+#include "device_gemm_xdl_c_shuffle.hpp"
+#include "element_wise_operation.hpp"
+#include "reference_gemm.hpp"
+#include "gemm_specialization.hpp"
+#include "test_util.hpp"
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using DeviceGemmPtr_ =
+    ck::tensor_operation::device::DeviceGemmPtr<ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::PassThrough>;
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace device_gemm_instance {
+void add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(std::vector<DeviceGemmPtr_>&);
+}
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace {
+
+using ADataType   = float;
+using BDataType   = float;
+using CDataType   = float;
+using AccDataType = float;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+auto PrepareGemmTensor(const ck::gemm_util::GemmParams& params)
+{
+    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(
+        f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(
+        f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_host_result(
+        f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(
+        f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
+
+    a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{-0.5, 0.5});
+    b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+
+    return std::make_tuple(a_m_k, b_k_n, c_m_n_host_result, c_m_n_device_result);
+}
+
+bool TestGemm(DeviceGemmPtr_& gemmPtr)
+{
+    // Arrange
+    ck::gemm_util::GemmParams params;
+    params.M       = 1024;
+    params.N       = 1024;
+    params.K       = 1024;
+    params.StrideA = 1024;
+    params.StrideB = 1024;
+    params.StrideC = 1024;
+
+    auto host_tensors           = PrepareGemmTensor(params);
+    const Tensor<ADataType>& a  = std::get<0>(host_tensors);
+    const Tensor<BDataType>& b  = std::get<1>(host_tensors);
+    Tensor<CDataType>& c_host   = std::get<2>(host_tensors);
+    Tensor<CDataType>& c_device = std::get<3>(host_tensors);
+
+    auto a_element_op = PassThrough{};
+    auto b_element_op = PassThrough{};
+    auto c_element_op = PassThrough{};
+
+    using ReferenceGemmInstance = ck::tensor_operation::host::
+        ReferenceGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+    ck::gemm_util::RunHostGEMM<ReferenceGemmInstance>(
+        a, b, c_host, a_element_op, b_element_op, c_element_op);
+
+    // Act
+    ck::gemm_util::RunDeviceGEMM(
+        gemmPtr, params, a, b, c_device, a_element_op, b_element_op, c_element_op);
+
+    // Assert
+    bool res = test_util::check_err(
+        c_device.mData, c_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
+
+    std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+
+    return res;
+}
+
+} // anonymous namespace
+
+int main()
+{
+    std::vector<DeviceGemmPtr_> gemmPtrs;
+    ck::tensor_operation::device::device_gemm_instance::
+        add_device_gemm_xdl_f32_f32_f32_mk_nk_mn_instances(gemmPtrs);
+
+    bool res = true;
+
+    for(auto& gemmPtr : gemmPtrs)
+    {
+        res &= TestGemm(gemmPtr);
+    }
+
+    std::cout << "TestGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
+}

test/gemm_xdl/test_gemm_int8.cpp

+#include <algorithm>
+#include <cstdlib>
+#include <half.hpp>
+#include <iostream>
+#include <numeric>
+#include <tuple>
+#include <vector>
+
+#include "gemm_util.hpp"
+#include "config.hpp"
+#include "print.hpp"
+#include "device.hpp"
+#include "host_tensor.hpp"
+#include "host_tensor_generator.hpp"
+#include "host_gemm.hpp"
+#include "device_tensor.hpp"
+#include "device_gemm_xdl.hpp"
+#include "device_gemm_xdl_c_shuffle.hpp"
+#include "element_wise_operation.hpp"
+#include "reference_gemm.hpp"
+#include "gemm_specialization.hpp"
+#include "test_util.hpp"
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using DeviceGemmPtr_ =
+    ck::tensor_operation::device::DeviceGemmPtr<ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                ck::tensor_operation::element_wise::PassThrough>;
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace device_gemm_instance {
+void add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instances(std::vector<DeviceGemmPtr_>&);
+}
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+
+namespace {
+
+using ADataType   = int8_t;
+using BDataType   = int8_t;
+using CDataType   = int8_t;
+using AccDataType = int32_t;
+
+using ALayout = ck::tensor_layout::gemm::RowMajor;
+using BLayout = ck::tensor_layout::gemm::ColumnMajor;
+using CLayout = ck::tensor_layout::gemm::RowMajor;
+
+auto PrepareGemmTensor(const ck::gemm_util::GemmParams& params)
+{
+    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(
+        f_host_tensor_descriptor(params.M, params.K, params.StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(
+        f_host_tensor_descriptor(params.K, params.N, params.StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_host_result(
+        f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_result(
+        f_host_tensor_descriptor(params.M, params.N, params.StrideC, CLayout{}));
+
+    a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+    b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+
+    return std::make_tuple(a_m_k, b_k_n, c_m_n_host_result, c_m_n_device_result);
+}
+
+bool TestGemm(DeviceGemmPtr_& gemmPtr)
+{
+    // Arrange
+    ck::gemm_util::GemmParams params;
+    params.M       = 1024;
+    params.N       = 1024;
+    params.K       = 1024;
+    params.StrideA = 1024;
+    params.StrideB = 1024;
+    params.StrideC = 1024;
+
+    auto host_tensors           = PrepareGemmTensor(params);
+    const Tensor<ADataType>& a  = std::get<0>(host_tensors);
+    const Tensor<BDataType>& b  = std::get<1>(host_tensors);
+    Tensor<CDataType>& c_host   = std::get<2>(host_tensors);
+    Tensor<CDataType>& c_device = std::get<3>(host_tensors);
+
+    auto a_element_op = PassThrough{};
+    auto b_element_op = PassThrough{};
+    auto c_element_op = PassThrough{};
+
+    using ReferenceGemmInstance = ck::tensor_operation::host::
+        ReferenceGemm<ADataType, BDataType, CDataType, PassThrough, PassThrough, PassThrough>;
+    ck::gemm_util::RunHostGEMM<ReferenceGemmInstance>(
+        a, b, c_host, a_element_op, b_element_op, c_element_op);
+
+    // Act
+    ck::gemm_util::RunDeviceGEMM(
+        gemmPtr, params, a, b, c_device, a_element_op, b_element_op, c_element_op);
+
+    // Assert
+    bool res = test_util::check_err(c_device.mData, c_host.mData, "Error: incorrect results!");
+
+    std::cout << (res ? "SUCCESS" : "FAILURE") << std::endl;
+
+    return res;
+}
+
+} // anonymous namespace
+
+int main()
+{
+    std::vector<DeviceGemmPtr_> gemmPtrs;
+    ck::tensor_operation::device::device_gemm_instance::
+        add_device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instances(gemmPtrs);
+
+    bool res = true;
+
+    for(auto& gemmPtr : gemmPtrs)
+    {
+        res &= TestGemm(gemmPtr);
+    }
+
+    std::cout << "TestGemm ..... " << (res ? "SUCCESS" : "FAILURE") << std::endl;
+}