static void inline _smpy2_hi_lo (int src1, int src2, int *restrict out_hi, int *restrict out_lo)
{
    long long out = _smpy2ll(src1, src2);
    *out_hi = _hill(out);
    *out_lo = _loll(out);
}
void ym_mult_q15(short *restrict pSrcA,
                 short *restrict pSrcB,
                 short *restrict pDst,
                 unsigned int blockSize)
{
    unsigned int blkCnt;                               /* loop counters */
    int inA1, inA2, inB1, inB2;                  /* temporary input variables */
    long long inA, inB;
    long long *restrict p_srcA = (long long*)pSrcA;
    long long *restrict p_srcB = (long long*)pSrcB;
    long long *restrict p_dst = (long long*)pDst;

    _nassert(((int)pSrcA & 0x07) == 0);
    _nassert(((int)pSrcB & 0x07) == 0);
    _nassert(((int)pDst & 0x07) == 0);
    _nassert((int)blockSize > 0);
    _nassert((int)(blockSize & 7) == 0);
    /* loop Unrolling */
    blkCnt = blockSize >> 3U;
    while (blkCnt > 0U)
    {
        int out_lo1, out_hi1, out_hi2, out_lo2;
        /* read two samples at a time from sourceA */
        inA = _amem8(p_srcA);
        p_srcA++;
        inA1 = _loll(inA);
        inA2 = _hill(inA);
        /* read two samples at a time from sourceB */
        inB = _amem8(p_srcB);
        p_srcB++;
        inB1 = _loll(inB);
        inB2 = _hill(inB);

        /* multiply mul = sourceA * sourceB */
        _smpy2_hi_lo(inA1, inB1, &out_hi1, &out_lo1);
        _smpy2_hi_lo(inA2, inB2, &out_hi2, &out_lo2);
        /* store the result */
        _amem8(p_dst) = _itoll(_packh2(out_hi2, out_lo2), _packh2(out_hi1, out_lo1));
        p_dst++;

        inA = _amem8(p_srcA);
        p_srcA++;
        inA1 = _loll(inA);
        inA2 = _hill(inA);
        /* read two samples at a time from sourceB */
        inB = _amem8(p_srcB);
        p_srcB++;
        inB1 = _loll(inB);
        inB2 = _hill(inB);

        /* multiply mul = sourceA * sourceB */
        _smpy2_hi_lo(inA1, inB1, &out_hi1, &out_lo1);
        _smpy2_hi_lo(inA2, inB2, &out_hi2, &out_lo2);
        /* store the result */
        _amem8(p_dst) = _itoll(_packh2(out_hi2, out_lo2), _packh2(out_hi1, out_lo1));
        p_dst++;
        /* Decrement the blockSize loop counter */
        blkCnt--;
    }
}