Boxcar filter too slow

Hi all,

I've wrote a simple boxcar monochrome (8 bit/pixel) image  scale method, and it is too slow. The following algorithm's running time is 9.3 ms for a 1280x960 => 640x480 conversion (500 MHz, so ~4.6 million cycle for 1.3M pixel, eg 3.8 clock/pixel). is there any official lib to do this faster? (For example I cannot find any similar algorithm in IMGLIB).

void scale(UInt32 width, UInt32 height, char *restrict out, const char *restrict in)
{
    int y;
    // Two rows step
    const UInt32 IN_INC = 2*width;
    const UInt32 OUT_INC = width>>1;
    const UInt32 CYCLE = height >> 1;
    const char *restrict x0 = in;           // even row pointer
    const char *restrict x1 = in+width;     // odd row pointer
    #pragma MUST_ITERATE(90)
    for (y = 0; y < CYCLE; ++y)
    {
        scale_inner(width, out, x0, x1);
        x0 += IN_INC;
        x1 += IN_INC;
        out += OUT_INC;
    }
}

static void scale_inner(UInt32 width, char *restrict xo, const char *restrict x0, const char *restrict x1)
{
    int i;
    const UInt32 CYCLE = width >> 3; // 8 pixel unrolling
    #pragma MUST_ITERATE(20)
    for (i = 0; i < CYCLE; i++)
    {
        xo[4*i+0] = ((Uint16)x0[8*i+0] + (Uint16)x1[8*i+0] + (Uint16)x0[8*i+1] + (Uint16)x1[8*i+1])>>2;
        xo[4*i+1] = ((Uint16)x0[8*i+2] + (Uint16)x1[8*i+2] + (Uint16)x0[8*i+3] + (Uint16)x1[8*i+3])>>2;
        xo[4*i+2] = ((Uint16)x0[8*i+4] + (Uint16)x1[8*i+4] + (Uint16)x0[8*i+5] + (Uint16)x1[8*i+5])>>2;
        xo[4*i+3] = ((Uint16)x0[8*i+6] + (Uint16)x1[8*i+6] + (Uint16)x0[8*i+7] + (Uint16)x1[8*i+7])>>2;
    }
}