Optimization of simple function using linear assembly

Hello

I have tried to write a simple function that adds with saturation two vectors and saves the result in the 3rd vector (because the current version in C wasn't fast enough). So far the attached code performs at ~0.5 cycles/element. I have a feeling it could be faster. Could you suggest something? I have compiled it with -mh24 and -O3 and have no further ideas. The DSP is c6400.

    .sect ".text:_func"
    .global    _func
        
   
_func:    .cproc    A4, B4, A6, B6
        .reg    vec_a, vec_b, vec_out, count, ah:al, bh:bl, ch:cl
        .no_mdep
       
            mv    A4, vec_a
            mv    B4, vec_b
            mv    A6, count
            mv    B6, vec_out
       
       
            shr    count, 3, count
           
loop:        .trip    2, 0x20000000, 2
            lddw    *vec_a++, ah:al
            lddw    *vec_b++, bh:bl
            saddu4    ah, bh, ch
            saddu4    al, bl, cl
            stdw    ch:cl, *vec_out++
    [count]    sub        count, 1, count
    [count]    b        loop
   
            .return
            .endproc

Here is the C++ equivalent. But it runs only ~4cycles/element, which is really slow. Something's definitely wrong here. Could you suggest any modifications here as well?

    inline void IntrinsicSADD
    (
        const uint8    * restrict inVector1,
        const uint8    * restrict inVector2,
        const int      inLength,
        uint8        * restrict outVector
    )
    {
        for(int i = 0; i < inLength; ++i)
        {
            uint32 out_hi = _saddu4( _hi( _amemd8_const( &inVector1[i] ) ), _hi( _amemd8_const( &inVector2[i] ) ) );
            uint32 out_lo = _saddu4( _lo( _amemd8_const( &inVector1[i] ) ), _lo( _amemd8_const( &inVector2[i] ) ) );
            _amemd8( &outVector[i] ) = _itod( out_hi, out_lo ); \
        }
    }