Added changes to support AVX 10.1 support #269
Conversation
detect AVX10.1 ISA
|
Hi @rrwinterton! Thank you for your pull request and welcome to our community. Action RequiredIn order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you. ProcessIn order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA. Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with If you have received this in error or have any questions, please contact us at cla@meta.com. Thanks! |
src/x86/isa.c
Outdated
| /* | ||
| * AVX 10.1 instructions: | ||
| */ | ||
| isa.avx10_1 = (structured_feature_info1.edx & UINT32_C(1 << 19)); |
There was a problem hiding this comment.
refer a UINT32_C constant like the others
Replace UINT32_C(1 << 19)
with UINT32_C(0x00080000)
To confirm the bits, I look at libyuv AVX10 support
It is in the same register as avxvnniint8
((cpu_einfo7[3] & 0x00000010) ? kCpuHasAVXVNNIINT8 : 0);
((cpu_einfo7[3] & 0x00080000) ? kCpuHasAVX10 : 0)
It looks right except in libyuv it checks for if ((GetXCR0() & 0xe0) == 0xe0) which says the OS supports saving ymm16 to ymm31.
So I think you should also check avx512_regs &&
There was a problem hiding this comment.
The reason I didn't check for AVX512 is because in the future there may not be 512 on some systems that support AVX10.2 in 256 bit form but that is still TBD. I think it is a good idea for now since we can change this in the future if this is the case.
|
Thank you for signing our Contributor License Agreement. We can now accept your code for this (and any) Meta Open Source project. Thanks! |
fbarchard
left a comment
fbarchard
left a comment
There was a problem hiding this comment.
Comment about other avx10 is not needed and subject to change.
The 'avx512 registers' is a misnomer. The OSS indicates the OS supports saving 32 vector registers.
src/x86/isa.c
Outdated
| /* | ||
| * AVX 10.1 instructions: | ||
| */ | ||
| isa.avx10_1 = avx512_regs && !!(structured_feature_info1.edx & UINT32_C(0x00080000)); |
There was a problem hiding this comment.
indent is wrong. Make it the same as lines above
There was a problem hiding this comment.
to be clear, you should use 1 tab, not 4 spaces
tools/isa-info.c
Outdated
| printf("\tAVX512_4VNNIW: %s\n", cpuinfo_has_x86_avx512_4vnniw() ? "yes" : "no"); | ||
| printf("\tAVX512_4FMAPS: %s\n", cpuinfo_has_x86_avx512_4fmaps() ? "yes" : "no"); | ||
|
|
||
| //avx10.1 (in future will add other avx10 version) |
There was a problem hiding this comment.
this comment is not needed and inconsistent with other printf's
tools/isa-info.c
Outdated
|
|
||
| //avx10.1 (in future will add other avx10 version) | ||
| printf("\tAVX10_1: %s\n", cpuinfo_has_x86_avx10_1() ? "yes" : "no"); | ||
|
|
There was a problem hiding this comment.
remove blank lines. avx10_1 is an ISA and should be grouped with the rest
There was a problem hiding this comment.
CI have linter automation that should detect this
rrwinterton
changed the title
fbarchard
left a comment
fbarchard
left a comment
There was a problem hiding this comment.
Change title for this PR to Added changes to support AVX 10.1
Have a follow up PR for 10.2
|
@rrwinterton can you please put a reference to some public doc on what AVX10 is and list 1-2 CPUs currently support it? |
rrwinterton
changed the title
fbarchard
left a comment
fbarchard
left a comment
There was a problem hiding this comment.
Looks good to me. If you can, test on real hardware that does not have avx512, but does have avx10.1
SDE doesnt really test the OSSAVE portion correctly, and the only publically released hardware that supports avx10.1 is granite rapids, but that also has avx512, so it doesnt test the OSSAVE.
This test is historically so new cpus can be used with old OS's, where the linux kernel does not support the new registers (32 ymm registers), so the cpuid needs to be turned off.
But the way you have it is safe... if it fails, it will just disable avx10 on a cpu/OS that it should have had it enabled.
6 participants
detect AVX10.1 ISA