Add bf16 example

example/1_gemm_xdl/gemm_xdl_bf16.cpp

+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+#include <stdlib.h>
+#include <half.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"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using BF16 = ck::bhalf_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   = 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;
+
+// clang-format off
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle<
+    ADataType,              // ADataType
+    BDataType,              // BDataType
+    CDataType,              // CDataType
+    AccDataType,            // AccDataType
+    ALayout,                // ALayout
+    BLayout,                // BLayout
+    CLayout,                // CLayout
+    PassThrough,            // AElementwiseOperation
+    PassThrough,            // BElementwiseOperation
+    PassThrough,            // CElementwiseOperation
+    256,                    // BlockSize
+    256,                    // MPerBlock
+    128,                    // NPerBlock
+    4,                      // K0PerBlock
+    8,                      // K1
+    32,                     // MPerXDL
+    32,                     // NPerXDL
+    4,                      // MXdlPerWave
+    2,                      // NXdlPerWave
+    S<4, 64, 1>,            // ABlockTransferThreadClusterLengths_K0_M_K1
+    S<1, 0, 2>,             // ABlockTransferThreadClusterArrangeOrder
+    S<1, 0, 2>,             // ABlockTransferSrcAccessOrder
+    2,                      // ABlockTransferSrcVectorDim
+    8,                      // ABlockTransferSrcScalarPerVector
+    8,                      // ABlockTransferDstScalarPerVector_K1
+    true,                   // ABlockLdsAddExtraM
+    S<4, 64, 1>,            // BBlockTransferThreadClusterLengths_K0_N_K1
+    S<1, 0, 2>,             // BBlockTransferThreadClusterArrangeOrder
+    S<1, 0, 2>,             // BBlockTransferSrcAccessOrder
+    2,                      // BBlockTransferSrcVectorDim
+    8,                      // BBlockTransferSrcScalarPerVector
+    8,                      // BBlockTransferDstScalarPerVector_K1
+    true,                   // BBlockLdsAddExtraN
+    1,                      // CShuffleMXdlPerWavePerShuffle
+    1,                      // CShuffleNXdlPerWavePerShuffle
+    S<1, 1, 32, 1, 1, 8>,   // CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
+    8>;                     // CBlockTransferScalarPerVector_NWaveNPerXdl
+// clang-format on
+
+using ReferenceGemmInstance = ck::tensor_operation::host::
+    ReferenceGemm<float, float, float, PassThrough, PassThrough, PassThrough>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = 0;
+    int init_method      = 0;
+    int nrepeat          = 5;
+
+    // GEMM 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]);
+        nrepeat         = std::stoi(argv[3]);
+    }
+    else if(argc == 10)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        nrepeat         = 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
+    {
+        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");
+        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(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+    std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
+    DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
+    DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpace());
+
+    a_m_k_device_buf.ToDevice(a_m_k.mData.data());
+    b_k_n_device_buf.ToDevice(b_k_n.mData.data());
+
+    auto a_element_op = PassThrough{};
+    auto b_element_op = PassThrough{};
+    auto c_element_op = PassThrough{};
+
+    // do GEMM
+    auto gemm     = DeviceGemmInstance{};
+    auto invoker  = gemm.MakeInvoker();
+    auto argument = gemm.MakeArgument(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()),
+                                      M,
+                                      N,
+                                      K,
+                                      StrideA,
+                                      StrideB,
+                                      StrideC,
+                                      a_element_op,
+                                      b_element_op,
+                                      c_element_op);
+
+    if(!gemm.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, nrepeat);
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        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, "
+              << gemm.GetTypeString() << std::endl;
+
+    c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<float> a_f32_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+        Tensor<float> b_f32_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+        Tensor<float> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+        Tensor<float> c_m_n_device_f32_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+
+        bf16_to_f32_(a_m_k, a_f32_m_k);
+        bf16_to_f32_(b_k_n, b_f32_k_n);
+        bf16_to_f32_(c_m_n_device_result, c_m_n_device_f32_result);
+
+        auto ref_gemm    = ReferenceGemmInstance{};
+        auto ref_invoker = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a_f32_m_k, b_f32_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        check_error(c_m_n_host_result, c_m_n_device_f32_result);
+    }
+
+    return 0;
+}

example/CMakeLists.txt

@@ -15,6 +15,7 @@ include_directories(BEFORE
 
 set(GEMM_XDL_SOURCE 1_gemm_xdl/gemm_xdl.cpp)
 set(GEMM_XDL_INT8_SOURCE 1_gemm_xdl/gemm_xdl_int8.cpp)
+set(GEMM_XDL_BF16_SOURCE 1_gemm_xdl/gemm_xdl_bf16.cpp)
 set(GEMM_XDL_BIAS_RELU_SOURCE 2_gemm_xdl_bias_relu/gemm_xdl_bias_relu.cpp)
 set(GEMM_XDL_BIAS_RELU_ADD_SOURCE 3_gemm_xdl_bias_relu_add/gemm_xdl_bias_relu_add.cpp)
 set(CONV2D_FWD_XDL_SOURCE 4_conv2d_fwd_xdl/conv2d_fwd_xdl.cpp)
@@ -28,6 +29,7 @@ set(CONVND_FWD_XDL_SOURCE 11_convnd_fwd_xdl/convnd_fwd_xdl.cpp)
 
 add_executable(gemm_xdl ${GEMM_XDL_SOURCE})
 add_executable(gemm_xdl_int8 ${GEMM_XDL_INT8_SOURCE})
+add_executable(gemm_xdl_bf16 ${GEMM_XDL_BF16_SOURCE})
 add_executable(gemm_xdl_bias_relu ${GEMM_XDL_BIAS_RELU_SOURCE})
 add_executable(gemm_xdl_bias_relu_add ${GEMM_XDL_BIAS_RELU_ADD_SOURCE})
 add_executable(conv2d_fwd_xdl ${CONV2D_FWD_XDL_SOURCE})
@@ -41,6 +43,7 @@ add_executable(convnd_fwd_xdl ${CONVND_FWD_XDL_SOURCE})
 
 target_link_libraries(gemm_xdl PRIVATE host_tensor)
 target_link_libraries(gemm_xdl_int8 PRIVATE host_tensor)
+target_link_libraries(gemm_xdl_bf16 PRIVATE host_tensor)
 target_link_libraries(gemm_xdl_bias_relu PRIVATE host_tensor)
 target_link_libraries(gemm_xdl_bias_relu_add PRIVATE host_tensor)
 target_link_libraries(conv2d_fwd_xdl PRIVATE host_tensor)

host/host_tensor/include/host_tensor.hpp

@@ -315,6 +315,8 @@ void ostream_HostTensorDescriptor(const HostTensorDescriptor& desc, std::ostream
 
 float bf16_to_f32_(ck::bhalf_t src_val);
 
+void bf16_to_f32_(const Tensor<ck::bhalf_t>& src, Tensor<float>& dst);
+
 template <typename T>
 void check_error(const Tensor<T>& ref, const Tensor<T>& result)
 {

host/host_tensor/src/host_tensor.cpp

@@ -73,3 +73,9 @@ float bf16_to_f32_(ck::bhalf_t src_val)
     } u = {uint32_t(src_val) << 16};
     return u.fp32;
 }
+
+void bf16_to_f32_(const Tensor<ck::bhalf_t>& src, Tensor<float>& dst)
+{
+    for(int i = 0; i < src.mData.size(); ++i)
+        dst.mData[i] = bf16_to_f32_(src.mData[i]);
+}