Support AVX-512 builds on Windows

[cmb69]

"Since limited support for /arch:AVX512 was added in Visual Studio 2017, and expanded in Visual Studio 2019"[1], we can safely offer this option, since PHP 8.4 is supposed to build with Visual Studio 2022, and it is unlikely that someone tries to build PHP 8.4 with Visual Studio, requesting AVX-512 support.

[1] https://learn.microsoft.com/en-us/cpp/build/reference/arch-x64?view=msvc-170

---

Note that this merely enables people to build for AVX-512; it doesn't change the default (that is still used for official Windows builds), nor does it update CI (which uses AVX2; I don't think AVX-512 is supported on hosted GH runners).

Also note that a full PHP build for AVX-512 succeeded, but apparently not even php -i can be run. Coreinfo reports here:

AVX-512-F - Supports AVX-512 Foundation instructions
AVX-512-DQ - Supports AVX-512 double and quadword instructions
AVX-512-IFAMA - Supports AVX-512 integer Fused multiply-add instructions
AVX-512-PF - Supports AVX-512 prefetch instructions
AVX-512-ER - Supports AVX-512 exponential and reciprocal instructions
AVX-512-CD - Supports AVX-512 conflict detection instructions
AVX-512-BW - Supports AVX-512 byte and word instructions
AVX-512-VL - Supports AVX-512 vector length instructions

I'll look into this.

[cmb69]

The problem is that vcvtusi2sd is not supported on my machine ("illegal instruction"), so apparently Coreinfo is wrong about my AVX-512F support. If somebody could run some tests, fine; otherwise I suggest to merge this PR.

[TimWolla]

so apparently Coreinfo is wrong about my AVX-512F support

I can't help with Windows, but what CPU are you using?

[cmb69]

[…], but what CPU are you using?

It's an AMD Ryzen 5 2400G. Actually, I was surprised that it supported AVX-512, but, hey, would have been cool. :)

And to clarify, the issue had nothing to do with PHP's AVX-512 support, but rather is due to compile time optimization of

php-src/Zend/zend_hrtime.c

Line 62 in 9083ce4

zend_hrtime_timer_scale= (double)ZEND_NANO_IN_SEC / (zend_hrtime_t)tf.QuadPart;

[Disservin]

which uses AVX2; I don't think AVX-512 is supported on hosted GH runners

I think some runners have avx512 but this isn't stable. Rather what one should do is use https://github.com/petarpetrovt/setup-sde and use the sde to emulate it. It's been great for testing avx512 builds.

[cmb69]

Rather what one should do is use https://github.com/petarpetrovt/setup-sde and use the sde to emulate it. It's been great for testing avx512 builds.

Interesting, thanks! I assume that the tests runs would be rather slow, so likely something for nightly builds only. And it may make more sense to use that for a Linux build, given that the official Windows binaries are still restricted to SSE2. @iluuu1994, what do you think?

[Disservin]

Interesting, thanks! I assume that the tests runs would be rather slow, so likely something for nightly builds only.

in my CI it seems to be twice as slow as the avx2 build

Are you aware that this currently fails because of missing gh command? (not sure what caddy is used for)
https://github.com/php/php-src/actions/runs/10150594501/job/28068047388#step:8:20

[iluuu1994]

I have no experience with Intel SDE, so I can't provide any helpful comments. A nightly build for this seems reasonable.

[cmb69]

Are you aware that this currently fails because of missing gh command?

Thanks for mentioning that! I have now forwarded to #13296 (comment).

[cmb69]

I have filed #15540 as separate ticket, and going to merge this soon; doesn't make sense to not enable explicit support for AVX-512 while we already support dynamic AVX-512 detection (and to my knowledge there is no code which only supports compile time AVX-512 in php-src).