Merge performance improvements from zfex

[hacklschorsch]

zfex (also on PyPI) is a performance-focused fork of zfec exploiting SIMD for both, Intel and ARM, with impressive results:

Legacy zfec had both results slightly above 50 MB/sec. zfex in all cases ran faster, achieving best performance with -DZFEX_UNROLL_ADDMUL_SIMD=4 unrolling, giving almost 6-fold speed-up.

Having an automatic and seamless way to pick faster versions of the algorithm if the hardware capabilities are available would be really neat.

[sajith]

It is fantastic that zfex exists and zfex contributors have been able to achieve all those performance gains. It however comes at a cost of some additional complexity and a bigger maintenance overhead. Take a look at https://github.com/WojciechMigda/zfex/blob/main/zfex/zfex.c, for example, and decide for yourself if you really want to add those changes to zfec.

The zfex fork is a result of our refusal to merge hand-rolled assembly to zfec: see #71.

I think there is room for both zfec and zfex in the world: zfex can courageously make all the performance improvements that they can, and zfec can remain simpler and more conservative. There is value in both approaches.

[sajith]

The zfex fork is a result of our refusal to merge hand-rolled assembly to zfec: see #71.

Well, to be really accurate, it was our inability (mine really, and other folks' unavailability) to review the code, not outright refusal, that caused the fork. :-)

[hacklschorsch]

Thanks for the clarification @sajith !

Would still be great if we had some SIMD story - Maybe the "write the code in such a way that the compiler puts in the right opcodes" route is viable here? (Not my specialty...)