On Tue, Oct 29, 2024 at 04:18:50PM +0100, Paolo Bonzini wrote:
> Most of the patches here are from Tao Su's v1.  The main issue in his
> version were two:
> 
> - overlooking kvm_cpu_xsave_init(), which currently looks at ExtSaveArea.
>   This would get a bit ugly for extended save states that are enabled
>   by both AVX512 and AVX10.  Patches 1-2 change kvm_cpu_xsave_init()
>   to look at ExtSaveArea's size field instead of testing features.
> 
> - downgrading silently to KVM reported value if the avx10_version property
>   is >= kvm reported value.  Xiaoyao Li suggested basing this on
>   cpu->check_cpuid and cpu->enforce_cpuid.  Also, the check must accept
>   any accelerator and not just KVM.  I moved the check to
>   x86_cpu_filter_features in patch 4.
> 
> I don't have a Granite Rapids machine, so please test! :)

I test it on Granite Rapids and all meet expection with my minor changes on
patch4 :)

> 
> Paolo
> 
> Paolo Bonzini (3):
>   target/i386: cpu: set correct supported XCR0 features for TCG
>   target/i386: do not rely on ExtSaveArea for accelerator-supported XCR0 bits
>   target/i386: return bool from x86_cpu_filter_features
> 
> Tao Su (5):
>   target/i386: add AVX10 feature and AVX10 version property
>   target/i386: add CPUID.24 features for AVX10
>   target/i386: Add feature dependencies for AVX10
>   target/i386: Add AVX512 state when AVX10 is supported
>   target/i386: Introduce GraniteRapids-v2 model
> 
>  target/i386/cpu.h         |  16 ++++
>  target/i386/cpu.c         | 175 ++++++++++++++++++++++++++++++++++----
>  target/i386/kvm/kvm-cpu.c |   8 --
>  target/i386/kvm/kvm.c     |   3 +-
>  4 files changed, 175 insertions(+), 27 deletions(-)
> 
> -- 
> 2.47.0
> 

Reply via email to