TODO

• [ ] One transformation matrix at the end of the 25 operations

• [ ] Affine transformation matrices.

• [ ] Sparsity: This year, we added another dimension of sparsity in addition to the space sparsity.

• [ ] Tiling

• [ ] Test with asymmetric image

• [ ] When initializing the orientation array, see if we can do a bit of loop unrolling

• [ ] Get rid of all helper functions, put everything into implementation driver at the end

• [ ] Premake 8 buffers for each of the types of rotations

  • Make each buffer 3x larger by padding 256 to the front and back
  • Then we can remove the memcpy, memset, memove
  • Take initial frame_buffer, generate every possible combination of rotation and flip state for a total of 8 different images
  • Now when we rotate clockwise, or flip, we just update a variable orientation to point to the correct pre-made image in the sequence

• [ ] Padding for translation

  Image translation (moving the image) can also be slow if it involves creating new image buffers. This code uses another trick: padding.

  1. Each of the 8 pre-made images is placed onto a much larger canvas by padding it with a white border. The border size (n) is equal to the size of the original image.
  2. Instead of moving the image, the code moves a "viewing window" over this large canvas.
  3. The translation point (x, y) is converted into a single offset.
  4. The final image buf is simply a slice of the correct padded buffer, starting from a position determined by this offset.

  This means translation is achieved just by changing the starting pointer for reading the image data, which is extremely fast.

Other tips

• Don’t use dynamic memory
• No matrix transformations
  • Instead, record flip, orientation, x offset and y offset