Switch to newer x86-64 microarchitecture by default

[itamarst]

x86-64-v2 should be supported on basically every extant computer (it's hardware from 2011 or later?).

• Write an article about why this is fine and good
• PyWavelets @rgommers
• NumPy next
• CPython
• Setuptools
• Rust
• Meson
• scikit-buildcore

Conda-forge might have data on per-micro-architecture downloads.

[itamarst]

Some useful references:

• https://fuchsia.dev/fuchsia-src/contribute/governance/rfcs/0073_x86_64_platform_requirement
• https://developers.redhat.com/blog/2021/01/05/building-red-hat-enterprise-linux-9-for-the-x86-64-v2-microarchitecture-level
• Some experiments with rustc suggesting 0.4% improvement with x86-64-v2 (i.e. rustc compiled with x86-64-v2, this is speed of compilation). x86-64-v3 has more noticeable performance improvements: [DO NOT MERGE] default x64 linux target to x64-v2 for perf test rust-lang/rust#95302

[itamarst]

Some experiments:

• https://github.com/pythonspeed/mandelbrot-simd single-threaded scalar backend is 1% faster when switching to x86-64-v2.
• Python projects are often difficult to benchmark reliably, so the ones I tried the noise overwhelms the improvement.
• https://github.com/tahoe-lafs/zfec/ had a 100% speedup but I assume this is a massive outlier.

[itamarst]

I have also found a bunch of cases where doing this makes the code slower. And others where it's impossible to tell if there's a speedup or not due to how much noise Python adds to everything.

For example, for PyWavelets for a couple of benchmarks (I can't get asv to give me anything useful in its output, so I'm just manually hacking in pytest-benchmark), AFAICT it just made things slower, as did x86-64-v3 for that matter. In contrast, "CFLAGS=-flto" did speed things up. And my experience so far has been that LTO either always helps, or at worst does no harm.

Given all that... this may not be worth pursuing.

[rgommers]

For the record, I started on migrating to rattler-build (reduces the build time by ~half) in preparation for adding x86-64-v3 builds (will ~double build time) for PyWavelets, but ran into several issues; it's a bit early days. Filed some issues, trying again in a couple of weeks.

[rgommers]

Agreed that -flto is higher prio. Unfortunate that -v3 would result in slowdowns, I wouldn't expect that.

[itamarst]

I released a package with x86-64-v2 (zfec) that had an actual significant speedup (double the speed!) and... people complained about lack of compatibility on actual hardware they have. So possibly changing the default isn't a good approach in general 😢

[rgommers]

Really? v2 should be pretty safe I'd have thought. Double the speed is nice either way!

[itamarst]

tahoe-lafs/zfec#114 (comment)

The doubled speed seems like a massive outlier, I assume it's some very specific thing about this particular code.