Question

I have an image represented by a 2D array of chars, I need to perform some operations on this image and store the result in another 2D array, these operations vary from calculating average value of neighbour cells to reordering rows.
what are possible optimisations I can do to get better performance?
any possible techniques are welcome (e.g locality of references, inline assembly, …)

I use c on a linux x86_64 machine

PS: I have raw color images with each pixel represented by a group of RGB values.

Answer 1

• First, allocate 4 bytes per rgb -triplet.
• Have a pointer for the beginning of each row, then shuffling can be done with pointer exchange

• plan your API as void process_row(int *out_row, int *in_row, int *row_above, int *row_below);

• some RGB calculations can be calculated in parallel with normal integer arithmetic.

  • e.g. 00rrrrrrrr 00bbbbbbbb 00gggggggg fits a single 32-bit integer.
    One can add 4 of those without overflow, then simply shift the result right by 2 bits
    and mask with 0011111111 0011111111 0011111111 b
    (this of course requires pre- and postprocessing the data)
  • on 64-bit processor packing of 00 rr 00 gg 00 bb 00 aa is also feasible approach
• possibly even easier approach is to use __sse intrinsics, from which there is just a small step to…
• inline assembler
• there is also a package ORC, which implements portable SIMD.