ofnuts wrote:
Gimp v3 with GEGL use floating point instead of short integer arithmetic. This is going to be slower, but then we will lose less time trying to fight the side effects of the short precision of the current version.
I'm not so certain that FP will be slower on most modern hardware. With dedicated FP hardware (such as SSE), addition and multiplication occur in a single cycle (same as integer), and while FP division and higher order functions take longer, they are typically faster than with integers. Of course, FP will require a lot more memory and this will have a negative impact upon the processor's memory caching; but then a lot of current libgimp arithmetic involves operations on individual bytes and this has a negative impact upon the processor's instruction lookahead parallelization (the lookahead optimization usually gets blocked when non-word-sized operands are encountered, whereas for word-sized operations the next few instructions can potentially be executed in parallel to the current one).
I'm not sure how it will all wash out (and if you care to instruct me, I'd be glad to be enlightened) but I get the feeling that the days of integer ops being faster than floating are over. ... and then there is the potential for
OpenCL (which we should've had a decade ago had the graphics card vendors been more open with their hardware).