2021.05.02 00:36 "[Tiff] SIMD optimizations", by Larry Bank

2021.05.03 21:09 "Re: [Tiff] SIMD optimizations", by Even Rouault


There is a cost associated with any target-specific optimizations and that cost pertains to the continual maintenance and testing required to support them. The optimizations should have proven value with real-world test cases (e.g. the whole time and not just the time for the part sped-up in isolation) before a project agrees to assimilate them because there is also a cost.

The few tests I reported about yesterday were on whole image decoding, not just the optimized part.

If Larry Bank (the expert) loses interest after the code has been submitted then there might not be another expert to take his place if a problem comes up.

My quick look at the SSE optimizations is that it is not too complicated SIMD usage, at least I have written in the past similar or more involved SSE code. At worse, that should be something easily removable if that becomes a pain in the future, as the generic C code paths would remain.

If access to Apple's M1 processor is needed in order to verify that SIMD code is working, but access to it is not available, then that would be a problem for the libtiff maintainers.

Hopefully that should be only generic Neon stuff, or at least ARM64 one, that could be run as well on other ARM64 platforms such as Graviton2. I presume this is testable on x86 with a gcc ARM cross compiler + qemu user mode emulation.

My software is free, but my time generally not.