MNN/source/backend/opencl/execution/buffer/MatmulBufExecution.cpp

//
//  MatmulBufExecution.cpp
//  MNN
//
//  Created by MNN on 2019/02/28.
//  Copyright © 2018, Alibaba Group Holding Limited
//

#ifndef MNN_OPENCL_BUFFER_CLOSED

#include "backend/opencl/execution/buffer/MatmulBufExecution.hpp"

namespace MNN {
namespace OpenCL {

MatMulBufExecution::MatMulBufExecution(const std::vector<Tensor *> &inputs, const MNN::Op *op, Backend *backend,
                                 bool transposeA, bool transposeB) : CommonExecution(backend, op)
                                 , mTransposeA(transposeA), mTransposeB(transposeB){
    mOpenCLBackend = static_cast<OpenCLBackend *>(backend);
}
ErrorCode MatMulBufExecution::onEncode(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs) {
    mUnits.resize(1);
    auto &unit = mUnits[0];
    auto runtime = mOpenCLBackend->getOpenCLRuntime();

    Tensor *input0 = inputs[0];
    Tensor *input1 = inputs[1];
    Tensor *output = outputs[0];

    std::vector<int> input0Shape = tensorShapeFormat(input0);
    std::vector<int> input1Shape = tensorShapeFormat(input1);
    std::vector<int> outputShape = tensorShapeFormat(output);

    std::set<std::string> buildOptions;
    int M = input0Shape[0];
    int K = input0Shape[3];
    if(mTransposeA) {
        M = input0Shape[3];
        K = input0Shape[0];
    }
    int N = mTransposeB ? input1Shape[0]: input1Shape[3];
    
    const int K_4 = UP_DIV(K, 4);
    const int N_4 = UP_DIV(N, 4);
    const int M_4 = UP_DIV(M, 4);
    
    // set large tile
    unsigned int tileM = 128;
    unsigned int tileN = 128;
    unsigned int tileK = 32;
    unsigned int localM = 32;
    unsigned int localN = 8;
    
    if(inputs.size() > 2) {
        buildOptions.emplace("-DBIAS");
    }
    
    bool canUseTile = (M % tileM == 0) && \
        (N % tileN == 0) && \
        (K % tileK == 0);
    bool canUseLargeTile = canUseTile && mTransposeA && !mTransposeB && inputs.size() == 2;
    if (!canUseLargeTile) {
        // set small tile
        tileM = 64;
        tileN = 128;
        tileK = 8;
        localM = 16;
        localN = 16;
        
        canUseTile = (M % tileM == 0) && (N % tileN == 0) && (K % tileK == 0);
    }
    
    if(canUseLargeTile) {
        // Match with Large tileM->MWG tileN->NWG tileK->KWG localM->MDIMA localN->NDIMC
        buildOptions.emplace(" -DGEMMK=0 -DKREG=1 -DKWG=32 -DKWI=2 -DMDIMA=32 -DMDIMC=32 -DMWG=128 -DNDIMB=8 -DNDIMC=8 -DNWG=128 -DSA=0 -DSB=0 -DSTRM=0 -DSTRN=1 -DVWM=2 -DVWN=8 -DOUTPUTMN");
        if(mOpenCLBackend->getOpenCLRuntime()->getGpuType() == GpuType::ADRENO) {
            buildOptions.emplace("-DUSE_CL_MAD=1");
            buildOptions.emplace("-DRELAX_WORKGROUP_SIZE=1");
        }
        if(runtime->isSupportedFP16()){
            buildOptions.emplace(" -DPRECISION=16");
        } else {
            buildOptions.emplace(" -DPRECISION=32");
        }
        unit.kernel       = runtime->buildKernel("matmul_params_buf", "Xgemm", buildOptions);

     } else if(canUseTile) {
        if(mTransposeA) {
            buildOptions.emplace(" -DTRANSPOSE_A");
        }
        if(mTransposeB) {
            buildOptions.emplace(" -DTRANSPOSE_B");
        }
        // Match with Small tileM->OPWM tileN->OPWN tileK->CPWK localM->OPWM/OPTM localN->OPWN/OPTN
        buildOptions.emplace(" -DOPWM=64 -DOPWN=128 -DCPWK=8 -DOPTM=4 -DOPTN=8");
        
        unit.kernel       = runtime->buildKernel("matmul_local_buf", "matmul_local_buf", buildOptions);
    } else {
        if(mTransposeA) {
            mKernelName = mTransposeB ? "matmul_transA_transB_buf":"matmul_transA_buf";
        } else {
            mKernelName = mTransposeB ? "matmul_transB_buf":"matmul_buf";
        }
        unit.kernel       = runtime->buildKernel("matmul_buf", mKernelName, buildOptions);
    }
    
    mMaxWorkGroupSize = static_cast<uint32_t>(runtime->getMaxWorkGroupSize(unit.kernel));

    cl_int ret = CL_SUCCESS;
    if(canUseLargeTile) {
        
        int out_per_thread_m = tileM / localM;
        int out_per_thread_n = tileN / localN;
        
        mGlobalWorkSize = {static_cast<uint32_t>(M/out_per_thread_m), static_cast<uint32_t>(N/out_per_thread_n)};
        mLocalWorkSize = {localM, localN};
        
        float alpha = 1.0;
        float beta = 0.0f;
        int offset = 0;
        int idx            = 0;
        ret |= unit.kernel->get().setArg(idx++, static_cast<int>(M));
        ret |= unit.kernel->get().setArg(idx++, static_cast<int>(N));
        ret |= unit.kernel->get().setArg(idx++, static_cast<int>(K));
        ret |= unit.kernel->get().setArg(idx++, alpha);
        ret |= unit.kernel->get().setArg(idx++, beta);
        ret |= unit.kernel->get().setArg(idx++, openCLBuffer(input0));
        ret |= unit.kernel->get().setArg(idx++, openCLBuffer(input1));
        if (inputs.size() > 2) {
            ret |= unit.kernel->get().setArg(idx++, openCLBuffer(inputs[2]));
        }
        ret |= unit.kernel->get().setArg(idx++, openCLBuffer(output));
        ret |= unit.kernel->get().setArg(idx++, offset);
        ret |= unit.kernel->get().setArg(idx++, offset);
        ret |= unit.kernel->get().setArg(idx++, offset);
        
        MNN_CHECK_CL_SUCCESS(ret, "setArg MatMulBufExecution use large tile opt");
        
    } else if(canUseTile) {
        int out_per_thread_m = tileM / localM;
        int out_per_thread_n = tileN / localN;
        
        mGlobalWorkSize = {static_cast<uint32_t>(M/out_per_thread_m), static_cast<uint32_t>(N/out_per_thread_n)};
        mLocalWorkSize = {localM, localN};

        int idx            = 0;
        ret |= unit.kernel->get().setArg(idx++, static_cast<int>(M));
        ret |= unit.kernel->get().setArg(idx++, static_cast<int>(N));
        ret |= unit.kernel->get().setArg(idx++, static_cast<int>(K));
        ret |= unit.kernel->get().setArg(idx++, openCLBuffer(input0));
        ret |= unit.kernel->get().setArg(idx++, openCLBuffer(input1));
        if(inputs.size() > 2) {
            ret |= unit.kernel->get().setArg(idx++, openCLBuffer(inputs[2]));
        }
        ret |= unit.kernel->get().setArg(idx++, openCLBuffer(output));

        MNN_CHECK_CL_SUCCESS(ret, "setArg MatMulBufExecution use tile opt");

    } else {
        if(mTransposeA) {
            mGlobalWorkSize = {static_cast<uint32_t>(N_4), static_cast<uint32_t>(M_4)};
            int idx            = 0;
            ret |= unit.kernel->get().setArg(idx++, mGlobalWorkSize[0]);
            ret |= unit.kernel->get().setArg(idx++, mGlobalWorkSize[1]);
            ret |= unit.kernel->get().setArg(idx++, openCLBuffer(input0));
            ret |= unit.kernel->get().setArg(idx++, openCLBuffer(input1));
            if(inputs.size() > 2) {
                ret |= unit.kernel->get().setArg(idx++, openCLBuffer(inputs[2]));
            }
            ret |= unit.kernel->get().setArg(idx++, openCLBuffer(output));
            ret |= unit.kernel->get().setArg(idx++, static_cast<int>(K));
            ret |= unit.kernel->get().setArg(idx++, static_cast<int>(K_4));
            ret |= unit.kernel->get().setArg(idx++, static_cast<int>(M));
            ret |= unit.kernel->get().setArg(idx++, static_cast<int>(M_4));
            ret |= unit.kernel->get().setArg(idx++, static_cast<int>(N_4));
            ret |= unit.kernel->get().setArg(idx++, static_cast<int>(N));
            MNN_CHECK_CL_SUCCESS(ret, "setArg MatMulBufExecution mTransposeA");
            
            mLocalWorkSize = localWS2DDefault(mGlobalWorkSize, mMaxWorkGroupSize, mOpenCLBackend->getOpenCLRuntime(), mKernelName, unit.kernel).first;
        }
        else {
            
            mGlobalWorkSize = {static_cast<uint32_t>(N_4), static_cast<uint32_t>(M)};
            int idx            = 0;
            ret |= unit.kernel->get().setArg(idx++, mGlobalWorkSize[0]);
            ret |= unit.kernel->get().setArg(idx++, mGlobalWorkSize[1]);
            ret |= unit.kernel->get().setArg(idx++, openCLBuffer(input0));
            ret |= unit.kernel->get().setArg(idx++, openCLBuffer(input1));
            if(inputs.size() > 2) {
                ret |= unit.kernel->get().setArg(idx++, openCLBuffer(inputs[2]));
            }
            ret |= unit.kernel->get().setArg(idx++, openCLBuffer(output));
            ret |= unit.kernel->get().setArg(idx++, static_cast<int>(K));
            ret |= unit.kernel->get().setArg(idx++, static_cast<int>(K_4));
            ret |= unit.kernel->get().setArg(idx++, static_cast<int>(N_4));
            ret |= unit.kernel->get().setArg(idx++, static_cast<int>(N));
            MNN_CHECK_CL_SUCCESS(ret, "setArg MatMulBufExecution");
            mLocalWorkSize = localWS2DDefault(mGlobalWorkSize, mMaxWorkGroupSize, mOpenCLBackend->getOpenCLRuntime(), mKernelName, unit.kernel).first;
        }
    }
    mOpenCLBackend->recordKernel2d(unit.kernel, mGlobalWorkSize, mLocalWorkSize);
    unit.globalWorkSize = {mGlobalWorkSize[0], mGlobalWorkSize[1]};
    unit.localWorkSize = {mLocalWorkSize[0], mLocalWorkSize[1]};
    return NO_ERROR;
}

class MatMulBufCreator : public OpenCLBackend::Creator {
public:
    virtual Execution *onCreate(const std::vector<Tensor *> &inputs, const std::vector<Tensor *> &outputs,
                                const MNN::Op *op, Backend *backend) const override {
        for (int i = 0; i < inputs.size(); ++i) {
            TensorUtils::setTensorSupportPack(inputs[i], false);
        }
        for (int i = 0; i < outputs.size(); ++i) {
            TensorUtils::setTensorSupportPack(outputs[i], false);
        }
        auto param = op->main_as_MatMul();
        return new MatMulBufExecution(inputs, op, backend, param->transposeA(), param->transposeB());
    }
};

REGISTER_OPENCL_OP_CREATOR(MatMulBufCreator, OpType_MatMul, BUFFER);

} // namespace OpenCL
} // namespace MNN
#endif /* MNN_OPENCL_BUFFER_CLOSED */