Add unit test for gemm_xdl int8

test/CMakeLists.txt

@@ -46,6 +46,12 @@ set(CONVND_FWD_XDL_SOURCE convnd_fwd_xdl/main.cpp)
 add_executable(test_convnd_fwd_xdl ${CONVND_FWD_XDL_SOURCE})
 target_link_libraries(test_convnd_fwd_xdl PRIVATE host_tensor)
 
+# 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 space_filling_curve_
 set(SPACE_FILLING_CURVE_SOURCE space_filling_curve/space_filling_curve.cpp)
 add_executable(space_filling_curve ${SPACE_FILLING_CURVE_SOURCE})

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_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;
+}