add int8 gemm multiply multiply a8w8 (#1591)

* add int8 gemm multiply multiply a8w8

* uncomment

* clang-format-12

* Add example_gemm_multiply_multiply_xdl_int8

* Remove shell scripts

* update preprocess number for mi308; bring back printout in ckprofiler

* format

---------
Co-authored-by: chenjun <junchen2@amd.com>
Co-authored-by: Haocong WANG <haocwang@amd.com>
Co-authored-by: carlushuang <carlus.huang@amd.com>

example/65_gemm_multiply_multiply/CMakeLists.txt

 add_example_executable(example_gemm_multiply_multiply_xdl_fp8 gemm_multiply_multiply_xdl_fp8.cpp)
 add_example_executable(example_gemm_multiply_multiply_xdl_fp8_ab_scale gemm_multiply_multiply_xdl_fp8_ab_scale.cpp)
 add_example_executable(example_gemm_add_add_xdl_fp16 gemm_add_add_xdl_fp16.cpp)
+add_example_executable(example_gemm_multiply_multiply_xdl_int8 gemm_multiply_multiply_xdl_int8.cpp)
\ No newline at end of file

example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_int8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.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_gemm.hpp"
+#include "ck/library/utility/check_err.hpp"
+
+#include "ck/utility/blkgemmpipe_scheduler.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using I8  = int8_t;
+using I32 = int;
+using F16 = ck::half_t;
+using FP8 = ck::f8_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = I8;
+using B0DataType       = I8;
+using AccDataType      = I32;
+using CShuffleDataType = I32;
+using D0DataType       = F32;
+using D1DataType       = F32;
+using DsDataType       = ck::Tuple<D0DataType, D1DataType>;
+using EDataType        = F16;
+
+using A0Layout = Row;
+using B0Layout = Col;
+using D0Layout = Row;
+using D1Layout = Col;
+using DsLayout = ck::Tuple<D0Layout, D1Layout>;
+using ELayout  = Row;
+
+struct MultiplyMultiply
+{
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ constexpr void
+    operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::half_t, float, float, float>(
+        ck::half_t& e, const float& c, const float& d0, const float& d1) const
+    {
+        const float x0_f = c * d0 * d1;
+
+        e = ck::type_convert<ck::half_t>(x0_f);
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::half_t, int, float, float>(
+        ck::half_t& e, const int& c, const float& d0, const float& d1) const
+    {
+        const float x0_f =
+            ck::type_convert<float>(c) * ck::type_convert<float>(d0) * ck::type_convert<float>(d1);
+
+        e = ck::type_convert<ck::half_t>(x0_f);
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::bhalf_t, int, float, float>(
+        ck::bhalf_t& e, const int& c, const float& d0, const float& d1) const
+    {
+        const float x0_f =
+            ck::type_convert<float>(c) * ck::type_convert<float>(d0) * ck::type_convert<float>(d1);
+
+        e = ck::type_convert<ck::bhalf_t>(x0_f);
+    }
+};
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = MultiplyMultiply;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShuffle_V3
+    // clang-format off
+///######|  ALayout|  BLayout| DsLayout| ELayout|      AData|      BData|     DsData|     EData|     AccData|         CShuffle|           A|           B|          CDE|           GEMM| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+///######|         |         |         |        |       Type|       Type|       Type|      Type|        Type|         DataType| Elementwise| Elementwise|  Elementwise| Spacialization|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+///######|         |         |         |        |           |           |           |          |            |                 |   Operation|   Operation|    Operation|               |      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+///######|         |         |         |        |           |           |           |          |            |                 |            |            |             |               |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |    S<C, D0, D1>|
+///###### RRR
+      ///<      Row,      Row, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,   256,   256,   128,    64,  16,   4,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,               8,              4,          0,          1,           1,               S<1, 32, 1, 8>,      S<8, 8, 1>,  ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, I8>;
+///###### RCR
+         <      Row,      Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,   256,   256,   128,    64,  16,  16,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,          0,     S<4, 64, 1>,    S<1, 0, 2>,     S<1, 0, 2>,             2,              16,             16,          0,          1,           1,               S<1, 32, 1, 8>,      S<8, 8, 1>,  ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, I8>;
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = K;
+    ck::index_t StrideD = 0;
+    ck::index_t StrideE = N;
+
+    ck::index_t KBatch = 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 == 12)
+    {
+        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]);
+        StrideD = std::stoi(argv[9]);
+        StrideE = std::stoi(argv[10]);
+
+        KBatch = std::stoi(argv[11]);
+    }
+    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 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE, KBatch\n");
+        exit(0);
+    }
+    do_verification = false;
+    auto f_host_tensor_descriptor =
+        [](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)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
+    Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+    Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
+    Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD, D1Layout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
+    std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
+    std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
+    std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{0, 2});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{0, 2});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{0, 2});
+        break;
+    default:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+        d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+        d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a0_device_buf.ToDevice(a0_m_k.mData.data());
+    b0_device_buf.ToDevice(b0_k_n.mData.data());
+    d0_device_buf.ToDevice(d0_m_n.mData.data());
+    d1_device_buf.ToDevice(d1_m_n.mData.data());
+    e_device_buf.ToDevice(e_m_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumDTensor = DsDataType::Size();
+
+    constexpr auto I0 = ck::Number<0>{};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(a0_device_buf.GetDeviceBuffer(),
+                               b0_device_buf.GetDeviceBuffer(),
+                               std::array<const void*, NumDTensor>{d0_device_buf.GetDeviceBuffer(),
+                                                                   d1_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               StrideA,
+                               StrideB,
+                               std::array<ck::index_t, NumDTensor>{I0, I0},
+                               StrideE,
+                               KBatch,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 20, 50});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * 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"
+              << std::endl;
+
+    if(do_verification)
+    {
+        invoker.Run(argument, StreamConfig{nullptr, false});
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        Tensor<CShuffleDataType> c_m_n({M, N});
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
+                                                                                B0DataType,
+                                                                                CShuffleDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                PassThrough,
+                                                                                PassThrough>;
+        auto ref_gemm               = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a0_m_k, b0_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

include/ck/host_utility/flush_cache.hpp

@@ -237,7 +237,7 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
                                              Args... args)
 {
 #if CK_TIME_KERNEL
-#define MEDIAN 1
+#define MEDIAN 0
     if(stream_config.time_kernel_)
     {
         if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
@@ -275,6 +275,14 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
 #else
         float total_time = 0;
 #endif
+        hipEvent_t start, stop;
+
+        hip_check_error(hipEventCreate(&start));
+        hip_check_error(hipEventCreate(&stop));
+
+        hip_check_error(hipDeviceSynchronize());
+        hip_check_error(hipEventRecord(start, stream_config.stream_id_));
+
         for(int i = 0; i < nrepeat; ++i)
         {
             if constexpr(!TimePreprocess)
@@ -282,13 +290,13 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
                 preprocess();
             }
 
-            hipEvent_t start, stop;
+            // hipEvent_t start, stop;
 
-            hip_check_error(hipEventCreate(&start));
-            hip_check_error(hipEventCreate(&stop));
+            // hip_check_error(hipEventCreate(&start));
+            // hip_check_error(hipEventCreate(&stop));
 
-            hip_check_error(hipDeviceSynchronize());
-            hip_check_error(hipEventRecord(start, stream_config.stream_id_));
+            // hip_check_error(hipDeviceSynchronize());
+            // hip_check_error(hipEventRecord(start, stream_config.stream_id_));
             // calculate preprocess time
             if constexpr(TimePreprocess)
             {
@@ -299,25 +307,34 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
             hip_check_error(hipGetLastError());
             // end real kernel
 
-            hip_check_error(hipEventRecord(stop, stream_config.stream_id_));
-            hip_check_error(hipEventSynchronize(stop));
-            float cur_time = 0;
-            hip_check_error(hipEventElapsedTime(&cur_time, start, stop));
-#if MEDIAN
-            times.insert(cur_time);
-#else
-            total_time += cur_time;
-#endif
+            //             hip_check_error(hipEventRecord(stop, stream_config.stream_id_));
+            //             hip_check_error(hipEventSynchronize(stop));
+            //             float cur_time = 0;
+            //             hip_check_error(hipEventElapsedTime(&cur_time, start, stop));
+            // #if MEDIAN
+            //             times.insert(cur_time);
+            // #else
+            //             total_time += cur_time;
+            // #endif
 
             if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
             {
-                std::cout << "i: " << i << " cur_time: " << cur_time << std::endl;
+                // std::cout << "i: " << i << " cur_time: " << cur_time << std::endl;
 
                 printf("gemm_args.p_a_grid: %p, gemm_args.p_b_grid:%p\n",
                        static_cast<const void*>(gemm_args.p_a_grid),
                        static_cast<const void*>(gemm_args.p_b_grid));
             }
         }
+        hip_check_error(hipEventRecord(stop, stream_config.stream_id_));
+        hip_check_error(hipEventSynchronize(stop));
+        float cur_time = 0;
+        hip_check_error(hipEventElapsedTime(&cur_time, start, stop));
+#if MEDIAN
+        times.insert(cur_time);
+#else
+        total_time += cur_time;
+#endif
 
 #if MEDIAN
         auto mid = times.begin();
@@ -333,7 +350,11 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
             return (*mid + *mid_next) / 2;
         }
 #else
-        return total_time / nrepeat;
+        // return total_time / nrepeat;
+        hipDeviceProp_t deviceProps;
+        hip_check_error(hipGetDeviceProperties(&deviceProps, 0));
+        float preprocess_offset = deviceProps.multiProcessorCount == 80 ? 0.005 : 0.01;
+        return (total_time - preprocess_offset * nrepeat) / nrepeat;
 #endif
     }
     else

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

@@ -272,6 +272,26 @@ struct MultiplyMultiply
 
         e = ck::type_convert<ck::bhalf_t>(x0_f);
     }
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::half_t, int, ck::half_t, ck::half_t>(
+        ck::half_t& e, const int& c, const ck::half_t& d0, const ck::half_t& d1) const
+    {
+        const float x0_f =
+            ck::type_convert<float>(c) * ck::type_convert<float>(d0) * ck::type_convert<float>(d1);
+
+        e = ck::type_convert<ck::half_t>(x0_f);
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::bhalf_t, int, float, float>(
+        ck::bhalf_t& e, const int& c, const float& d0, const float& d1) const
+    {
+        const float x0_f =
+            ck::type_convert<float>(c) * ck::type_convert<float>(d0) * ck::type_convert<float>(d1);
+
+        e = ck::type_convert<ck::bhalf_t>(x0_f);
+    }
 };
 
 struct MultiplyAddFastGelu

include/ck/utility/amd_xdlops.hpp

@@ -327,12 +327,12 @@ struct intrin_mfma_i32_16x16x32i8<16, 16>
     __device__ static void Run(const int8x8_t& reg_a, const int8x8_t& reg_b, FloatC& reg_c)
     {
         reg_c.template AsType<int32x4_t>()(Number<0>{}) =
-            __builtin_amdgcn_mfma_i32_16x16x32i8(bit_cast<int64_t>(reg_a),
-                                                 bit_cast<int64_t>(reg_b),
-                                                 reg_c.template AsType<int32x4_t>()[Number<0>{}],
-                                                 0,
-                                                 0,
-                                                 0);
+            __builtin_amdgcn_mfma_i32_16x16x32_i8(bit_cast<int64_t>(reg_a),
+                                                  bit_cast<int64_t>(reg_b),
+                                                  reg_c.template AsType<int32x4_t>()[Number<0>{}],
+                                                  0,
+                                                  0,
+                                                  0);
     }
 };
 

library/include/ck/library/tensor_operation_instance/gpu/gemm_multiply_multiply.hpp

@@ -96,6 +96,87 @@ void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_kpadding_i
                                                           MultiplyMultiply>>>& instances);
 #endif
 
+#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_INT8))
+void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
+                                                          Col,
+                                                          Tuple<Row, Col>,
+                                                          Row,
+                                                          I8,
+                                                          I8,
+                                                          Tuple<F32, F32>,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          MultiplyMultiply>>>& instances);
+
+void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
+                                                          Col,
+                                                          Tuple<Row, Col>,
+                                                          Row,
+                                                          I8,
+                                                          I8,
+                                                          Tuple<F32, F32>,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          MultiplyMultiply>>>& instances);
+
+void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
+                                                          Col,
+                                                          Tuple<Row, Col>,
+                                                          Row,
+                                                          I8,
+                                                          I8,
+                                                          Tuple<F32, F32>,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          MultiplyMultiply>>>& instances);
+
+void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
+                                                          Col,
+                                                          Tuple<Row, Col>,
+                                                          Row,
+                                                          I8,
+                                                          I8,
+                                                          Tuple<F32, F32>,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          MultiplyMultiply>>>& instances);
+
+void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
+                                                          Col,
+                                                          Tuple<Row, Col>,
+                                                          Row,
+                                                          I8,
+                                                          I8,
+                                                          Tuple<F32, F32>,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          MultiplyMultiply>>>& instances);
+
+void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
+                                                          Col,
+                                                          Tuple<Row, Col>,
+                                                          Row,
+                                                          I8,
+                                                          I8,
+                                                          Tuple<F32, F32>,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          MultiplyMultiply>>>& instances);
+
+#endif
+
 template <typename ADataType,
           typename BDataType,
           typename CDataType,
@@ -155,6 +236,30 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGemmMu
                     op_ptrs);
             }
         }
+#endif
+#if(defined(CK_ENABLE_BF16) || defined(CK_ENABLE_INT8))
+        if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
+                     is_same_v<CDataType, bhalf_t>)
+        {
+            if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
+                         is_same_v<CLayout, Row>)
+            {
+                add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_default_instances(
+                    op_ptrs);
+                add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_kpadding_instances(
+                    op_ptrs);
+
+                add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_default_instances(
+                    op_ptrs);
+                add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_kpadding_instances(
+                    op_ptrs);
+
+                add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_default_instances(
+                    op_ptrs);
+                add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_kpadding_instances(
+                    op_ptrs);
+            }
+        }
 #endif
         return op_ptrs;
     }

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/CMakeLists.txt

@@ -8,9 +8,19 @@ list(APPEND GEMM_MULTIPLY_MULTIPLY_INSTANCES
         device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v1_kpadding_instance.cpp
         device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_default_instance.cpp
         device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_mem_v2_kpadding_instance.cpp
+
+        device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_default_instance.cpp
+        device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_kpadding_instance.cpp
+        device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_default_instance.cpp
+        device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_kpadding_instance.cpp
+        device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_default_instance.cpp
+        device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_kpadding_instance.cpp
         )
 
 set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
 set_source_files_properties(device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
 
+set_source_files_properties(device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_default_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
+set_source_files_properties(device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_kpadding_instance.cpp PROPERTIES COMPILE_OPTIONS ";-mllvm;-greedy-reverse-local-assignment=1")
+
 add_instance_library(device_gemm_multiply_multiply_instance ${GEMM_MULTIPLY_MULTIPLY_INSTANCES})

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_default_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
+                                                          Col,
+                                                          Tuple<Row, Col>,
+                                                          Row,
+                                                          I8,
+                                                          I8,
+                                                          Tuple<F32, F32>,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          MultiplyMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_instances<GemmDefault>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_kpadding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
+                                                          Col,
+                                                          Tuple<Row, Col>,
+                                                          Row,
+                                                          I8,
+                                                          I8,
+                                                          Tuple<F32, F32>,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          MultiplyMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_comp_instances<GemmKPadding>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_default_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
+                                                          Col,
+                                                          Tuple<Row, Col>,
+                                                          Row,
+                                                          I8,
+                                                          I8,
+                                                          Tuple<F32, F32>,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          MultiplyMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_instances<Intrawave,
+                                                                            GemmDefault>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_kpadding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v1_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
+                                                          Col,
+                                                          Tuple<Row, Col>,
+                                                          Row,
+                                                          I8,
+                                                          I8,
+                                                          Tuple<F32, F32>,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          MultiplyMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_instances<Intrawave,
+                                                                            GemmKPadding>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_default_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_default_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
+                                                          Col,
+                                                          Tuple<Row, Col>,
+                                                          Row,
+                                                          I8,
+                                                          I8,
+                                                          Tuple<F32, F32>,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          MultiplyMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_instances<Interwave,
+                                                                            GemmDefault>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_i8_i8_bf16/device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_kpadding_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_v2_kpadding_instances(
+    std::vector<std::unique_ptr<DeviceGemmMultipleDSplitK<Row,
+                                                          Col,
+                                                          Tuple<Row, Col>,
+                                                          Row,
+                                                          I8,
+                                                          I8,
+                                                          Tuple<F32, F32>,
+                                                          BF16,
+                                                          PassThrough,
+                                                          PassThrough,
+                                                          MultiplyMultiply>>>& instances)
+{
+    add_device_operation_instances(
+        instances,
+        device_gemm_multiply_multiply_xdl_i8_i8_bf16_mk_nk_mn_mem_instances<Interwave,
+                                                                            GemmKPadding>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

profiler/include/profiler/profile_gemm_multiply_multiply_impl.hpp

@@ -271,10 +271,12 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
                           << " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
                           << kbatch_curr << std::endl;
 
-#if defined CK_ENABLE_FP8
+#if defined CK_ENABLE_FP8 || defined CK_ENABLE_INT8
                 // set softer tolerances for fp8
-                if constexpr(is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
-                             is_same_v<EDataType, f8_t>)
+                if constexpr((is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
+                              is_same_v<EDataType, f8_t>) ||
+                             (is_same_v<ADataType, int8_t> || is_same_v<BDataType, int8_t> ||
+                              is_same_v<EDataType, int8_t>))
                 {
                     std::string msg = "Error: Incorrect results!";
                     double rtol     = 1e-1;
@@ -286,7 +288,7 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
                 {
 #endif
                     pass = pass & ck::utils::check_err(e_m_n_device_result, e_m_n_host_result);
-#if defined CK_ENABLE_FP8
+#if defined CK_ENABLE_FP8 || defined CK_ENABLE_INT8
                 }
 #endif
 

profiler/src/profile_gemm_multiply_multiply.cpp

@@ -27,6 +27,7 @@ enum struct GemmDataType
     F16_F8_F16,     // 5
     F16_F16_F16_F8, // 6
     F8_F8_BF16,     // 7
+    INT8_INT8_BF16, // 8
 };
 
 #define OP_NAME "gemm_multiply_multiply"
@@ -39,7 +40,7 @@ int profile_gemm_multiply_multiply(int argc, char* argv[])
         printf("arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n");
         printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8; 4: f8@f16; 5: f16@f8; 6: "
                "f16->f8; 7: f8->bf16, "
-               "comp f8)\n");
+               "comp f8; 8: int8->bf16)\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");
@@ -89,6 +90,8 @@ int profile_gemm_multiply_multiply(int argc, char* argv[])
     using F32  = float;
     using BF16 = ck::bhalf_t;
     using F8   = ck::f8_t;
+    using I8   = int8_t;
+    using I32  = int;
 
     using Row = ck::tensor_layout::gemm::RowMajor;
     using Col = ck::tensor_layout::gemm::ColumnMajor;
@@ -162,6 +165,11 @@ int profile_gemm_multiply_multiply(int argc, char* argv[])
         return profile(
             F8{}, F8{}, F8{}, F32{}, F32{}, F32{}, BF16{}, Row{}, Col{}, Row{}, Col{}, Row{});
     }
+    else if(data_type == GemmDataType::INT8_INT8_BF16 && layout == GemmMatrixLayout::MK_NK_MN)
+    {
+        return profile(
+            I8{}, I8{}, I8{}, I32{}, F32{}, F32{}, BF16{}, Row{}, Col{}, Row{}, Col{}, Row{});
+    }
     else
     {
         std::cout << "this data_type & layout is not implemented" << std::endl;