Don't set -mno-sse / -mno-avx2

[haampie]

On PPC and other architecutres neither -msse nor -mno-sse are supported.

Yet

numcodecs/setup.py

Lines 25 to 32 in 45a8ef3

	if disable_sse2:
	base_compile_args.append('-mno-sse2')
	elif have_sse2:
	base_compile_args.append('-msse2')
	if disable_avx2:
	base_compile_args.append('-mno-avx2')
	elif have_avx2:
	base_compile_args.append('-mavx2')

numcodecs will set -mno-sse / -mno-avx2 on those platforms.

Either make it conditional on have_sse/avx2, or don't set any -mno-* flag at all, why bother?

[joshmoore]

Thanks for the issue, @haampie. Do you have the feeling that another PR like #483 would achieve what you're looking for?

[haampie]

The script would have to be updated either like:

if have_sse2:
    base_compile_args.append('-mno-sse2' if disable_sse2 else '-msse2')

but I'd be happier if it was just:

if have_sse2 and not disable_sse2:
    base_compiler_flags.append('-msse2')

I don't think I've ever seen a build system that explicitly disabled a certain micro-architecture. Most of the build systems do something like "if ENABLE_SSE2=1 then add -msse2", if not specified, don't add any flag. (And some projects that use cpuid to dynamically dispatch to micro arch optimized routines just test what the compiler can target, compile for each target, and use a generic arch for the entrypoint)

[sahiljhawar]

Is this issue fixed? I am having problem while building and pushing image to docker hub on (linux/arm64/v8). And the issue stems from -msse2 and -mavx2

[haampie]

it's not

[dstansby]

The change described in #486 (comment) seems reasonable. @haampie, do you think you'd be able to open a pull request with that change?