Const time box blur (no scale version)

box_blur_noscale.c

void box_blur_noscale(const uint8_t* input, int32_t* out, const int w, const int h, int hwin)
{
    const int win = 2*hwin+1;
    int *_buf = (int*)alloca((w*win+w) * sizeof(int));
    int* sums = _buf + w*win;
    int* next_row = _buf;
    int* oldest_row = _buf;
    int i, j;

    memset(sums, 0, w*sizeof(int));

    int* output = out + hwin*w - hwin;

    for (i=0; i<h; i++) 
    {
        int hsum = 0;
        const uint8_t* back = input;

        for (j = 0; j < win-1; j++) hsum += input[j];
        for (; j < w; j++) 
        {
            hsum += input[j];
            next_row[j] = hsum;
            sums[j] += hsum;
            hsum -= *(back++);
        }
        if (i >= win-1) 
        {
            for(j = win-1; j < w; j++)
            {
                hsum = sums[j];
                output[j] = hsum; // mult with 1/(win*win) if required
                sums[j] = hsum - oldest_row[j];
            }
            output += w;
            oldest_row += w;
            if (oldest_row == _buf + w*win) oldest_row = _buf;
        }
        input += w;
        next_row += w;
        if (next_row == _buf + w*win) next_row = _buf;
    }
}