when compiling for alderlake, 512 instructions are used

[Pleune]

Sorry for being breif, but I didn't really have time to dig deep into this one, as I'm really just trying to update my home immich install which happens to pull this in.

   last 25 log lines:
   >   cargo:warning=         ^
   >   cargo:warning=./src/f16.c:29:10: error: always_inline function '_mm512_fmadd_ph' requires target feature 'widekl', but would be inlined into function 'v_f16_cosine_avx512fp16' that is compiled without support for 'widekl'
   >   cargo:warning=    xx = _mm512_fmadd_ph(x, x, xx);
   >   cargo:warning=         ^
   >   cargo:warning=./src/f16.c:30:10: error: always_inline function '_mm512_fmadd_ph' requires target feature 'widekl', but would be inlined into function 'v_f16_cosine_avx512fp16' that is compiled without support for 'widekl'
   >   cargo:warning=    yy = _mm512_fmadd_ph(y, y, yy);
   >   cargo:warning=         ^
   >   cargo:warning=./src/f16.c:33:17: error: always_inline function '_mm512_reduce_add_ph' requires target feature 'widekl', but would be inlined into function 'v_f16_cosine_avx512fp16' that is compiled without support for 'widekl'
   >   cargo:warning=    float rxy = _mm512_reduce_add_ph(xy);
   >   cargo:warning=                ^
   >   cargo:warning=./src/f16.c:34:17: error: always_inline function '_mm512_reduce_add_ph' requires target feature 'widekl', but would be inlined into function 'v_f16_cosine_avx512fp16' that is compiled without support for 'widekl'
   >   cargo:warning=    float rxx = _mm512_reduce_add_ph(xx);
   >   cargo:warning=                ^
   >   cargo:warning=./src/f16.c:35:17: error: always_inline function '_mm512_reduce_add_ph' requires target feature 'widekl', but would be inlined into function 'v_f16_cosine_avx512fp16' that is compiled without support for 'widekl'
   >   cargo:warning=    float ryy = _mm512_reduce_add_ph(yy);
   >   cargo:warning=                ^
   >   cargo:warning=19 errors generated.
   >

Here I'm compiling with mtune and march set to alderlake. I must set march to x86-64-v3, or just not set it at all, and then it builds fine. It should not try to use avx512 and etc, as these are not available on my alderlake 12600k.

[usamoi]

I must set march to x86-64-v3, or just not set it at all, and then it builds fine. It should not try to use avx512 and etc, as these are not available on my alderlake 12600k.

You don't need to set them, and avx512 is enabled at runtime. If avx512 is not available on your machine, the code related to AVX512 will not be executed.

[Pleune]

Sorry, I mean that I believe cargo is for some reason using avx512 instructions, but then when clang is linking to it it's throwing a fit that they should not be there. Or something like that, I'm not really sure.

Either way I am not able to build at all with march=alderlake

[usamoi]

Could you provide your operating system, clang version, and the version of pgvecto.rs you are trying to compile?

[usamoi]

I believe this is a bug in clang. To avoid this bug, please do not set -march=alderlake.

[Pleune]

Thank you for spending the time to have dug into this a little. I greatly appreciate it.

I was building the v0.3.0 tag with clang 16.0.6 on NixOS. These are current on today's nixos-unstable branch