post-processing: negate: optimise for non-vectorising builds

[davidplowman]

The loop can be better optimised when the vectorising compiler cannot
be used.

Signed-off-by: David Plowman [email protected]

[davidplowman]

@naushir
For my amusement, I timed the loop for debug, release and vectorising builds, using either uint8_t, uint32_t or uint64_t operations. The results (in ms per 3MP frame, running single-threaded) were:

             uint8_t  uint32_t  uint64_t
Debug        900      230       120
Release      230      120       120
Vectorising  60       60        60

I've assumed we can rely on the buffers being 4-byte aligned. Do you think 8-byte alignment could be assumed, as that would run faster in a debug buidl?

[naushir]

For YUV420 formats, they will be 8-byte aligned. However, other formats will currently be 4-byte aligned. There's nothing to stop us from changing all to 8-byte alignment.... but it sort of does not seem worth it for only improving Debug builds.

[davidplowman]

Sounds godd as it is, then. Do you want to push the merge button?