gemm_xdl_bf16_rtn.cpp

// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.

#include "common.hpp"

#include "ck/utility/type_convert.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle.hpp"

using ADataType        = ck::bhalf_t;
using BDataType        = ck::bhalf_t;
using CDataType        = ck::bhalf_t;
using AccDataType      = float;
using CShuffleDataType = float;

using ALayout = Row;
using BLayout = Col;
using CLayout = Row;

struct bfp16_to_fp16
{
    __host__ __device__ void operator()(ck::half_t& y, const ck::bhalf_t& x) const
    {
        y = ck::type_convert<ck::half_t>(x);
    }
};

struct fp32_to_bfp16
{
    __host__ __device__ void operator()(ck::bhalf_t& y, const float& x) const
    {
        y = ck::bf16_convert_rtn<ck::bhalf_t>(x);
    }
};

using AElementOp = bfp16_to_fp16;
using BElementOp = bfp16_to_fp16;
using CElementOp = fp32_to_bfp16;

static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;

// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
// ######| ALayout| BLayout| CLayout|     AData|     BData|     CData|     AccData|         CShuffle|           A|           B|           C|           GEMM| NumGemmK| 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|              LoopScheduler|                Pipeline| ComputeType|
// ######|        |        |        |      Type|      Type|      Type|        Type|         DataType| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch|  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|                           |                 Version|            |
// ######|        |        |        |          |          |          |            |                 |   Operation|   Operation|   Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|                           |                        |            |
// ######|        |        |        |          |          |          |            |                 |            |            |            |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                           |                        |            |
         < ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,  AElementOp,  BElementOp,  CElementOp,    GemmDefault,        1,   256,   256,   128,    32,   8,   8,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8, ck::LoopScheduler::Default, ck::PipelineVersion::v1, ck::half_t>;
// clang-format on

using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
                                                                        BDataType,
                                                                        CDataType,
                                                                        AccDataType,
                                                                        AElementOp,
                                                                        BElementOp,
                                                                        CElementOp,
                                                                        ck::half_t>;

#include "run_gemm_example.inc"

int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }