breakpoint();
        for (x in range(loops))
        {
            var a = aload16b(src+x*16+0*stride);
            var b = aload16b(src+x*16+1*stride);
            var c = aload16b(src+x*16+2*stride);
            var d = aload16b(src+x*16+3*stride);
            var e = aload16b(src+x*16+4*stride);
            var f = aload16b(src+x*16+5*stride);
            var g = aload16b(src+x*16+6*stride);
            var zero = Vec[Byte, 16](0u8);
            
            var ag = avg_b(a,g);
            var bf = avg_b(b,f);
            var ce = avg_b(c,e);
            var d2 = avg_b(d,zero);
            
            var agbf = avg_b(ag,bf);
            var ced = avg_b(ce,d2);
            
            var t = avg_b(agbf, ced);
            
            astore16b(dst+x*16, t);
        }