SwiGLU

#include "kernel_operator.h"

template <typename srcType>
class KernelSwiGLU {
public:
    __aicore__ inline KernelSwiGLU(){}
    __aicore__ inline void Init(GM_ADDR src0Gm, GM_ADDR src1Gm, GM_ADDR dstGm, uint32_t srcSize, float beta)
    {
        betaValue = beta;
        dataSize = srcSize;
        src0Global.SetGlobalBuffer(reinterpret_cast<__gm__ srcType *>(src0Gm), dataSize);
        src1Global.SetGlobalBuffer(reinterpret_cast<__gm__ srcType *>(src1Gm), dataSize);
        dst_global.SetGlobalBuffer(reinterpret_cast<__gm__ srcType *>(dstGm), dataSize);
        pipe.InitBuffer(inQueueX, 1, dataSize * sizeof(srcType));
        pipe.InitBuffer(inQueueY, 1, dataSize * sizeof(srcType));
        pipe.InitBuffer(outQueue, 1, dataSize * sizeof(srcType));
        if (sizeof(srcType) != sizeof(float)) {
            pipe.InitBuffer(calcBufs, dataSize * (sizeof(float) / sizeof(uint8_t)) * 3);
        }

    }
    __aicore__ inline void Process()
    {
        CopyIn();
        Compute();
        CopyOut();
    }

private:
    __aicore__ inline void CopyIn()
    {
        AscendC::LocalTensor<srcType> src0Local = inQueueX.AllocTensor<srcType>();
        AscendC::LocalTensor<srcType> src1Local = inQueueY.AllocTensor<srcType>();
        AscendC::DataCopy(src0Local, src0Global, dataSize);
        AscendC::DataCopy(src1Local, src1Global, dataSize);
        inQueueX.EnQue(src0Local);
        inQueueY.EnQue(src1Local);
    }
    __aicore__ inline void Compute()
    {
        AscendC::LocalTensor<srcType> dstLocal = outQueue.AllocTensor<srcType>();
        AscendC::LocalTensor<srcType> src0Local = inQueueX.DeQue<srcType>();
        AscendC::LocalTensor<srcType> src1Local = inQueueY.DeQue<srcType>();
        AscendC::LocalTensor<uint8_t> tmpLocal;
        if (sizeof(srcType) != sizeof(float)) {
            tmpLocal = calcBufs.Get<uint8_t>();
            AscendC::SwiGLU<srcType, false>(dstLocal, src0Local, src1Local, betaValue, tmpLocal, dataSize);
        } else {
            AscendC::SwiGLU<srcType, false>(dstLocal, src0Local, src1Local, betaValue, dataSize);
        }

        outQueue.EnQue<srcType>(dstLocal);
        inQueueX.FreeTensor(src0Local);
        inQueueY.FreeTensor(src1Local);
    }
    __aicore__ inline void CopyOut()
    {
        AscendC::LocalTensor<srcType> dstLocal = outQueue.DeQue<srcType>();
        AscendC::DataCopy(dst_global, dstLocal, dataSize);
        outQueue.FreeTensor(dstLocal);
    }

private:
    AscendC::GlobalTensor<srcType> src0Global;
    AscendC::GlobalTensor<srcType> src1Global;
    AscendC::GlobalTensor<srcType> dst_global;

    AscendC::TPipe pipe;
    AscendC::TQue<AscendC::QuePosition::VECIN, 1> inQueueX;
    AscendC::TQue<AscendC::QuePosition::VECIN, 1> inQueueY;
    AscendC::TQue<AscendC::QuePosition::VECOUT, 1> outQueue;
    AscendC::TBuf<AscendC::TPosition::VECCALC> calcBufs;
    uint32_t dataSize = 0;
    float betaValue = 0;
};

template <typename dataType>
__aicore__ void kernel_swiglu_operator(GM_ADDR src0Gm, GM_ADDR src1Gm, GM_ADDR dstGm, uint32_t srcSize)
{
    KernelSwiGLU<dataType> op;
    float scalarValue = 1;
    op.Init(src0Gm, src1Gm, dstGm, srcSize, scalarValue);
    op.Process();
}