Hi folks, I have noticed that in the intel/2023 toolchain (and subchains), optimization has very wisely been set to a moderate level (-O2 -ftz -fp-speculation=safe -fp-model precise). By setting the toolchain option 'opt' to True, -O2 is changed to -O3 for a very moderate gain in performance.
For this specific package, I need to get rid of -fp-speculation=safe -fp-model precise, as these options are not necessary for correctness and are detrimental for performance (in this case the cost is a factor 2.5 - that is most certainly a special case and not typical of other applications). I can do that by setting the toolchain option 'extra_cflags': '-fp-speculation=fast -fp-model fast', (those are actually the default settings for the compiler). Is that the correct way to do something like that? Should setting 'opt': True for the Intel toolchain change these compiler options for the OneAPI compilers? I agree that for a generic framework compiling numerical codes, the default should be the more conservative settings, the Intel defaults may be too agressive. Best regards Jakob PS. This is for the ASAP3 package, I will submit the EasyConfigs once the fate of SciPy-bundle-2023.07-iimkl-2023a.eb (#18875) has been decided.
