On Wed, Aug 02, 2023 at 04:36:10PM -0700, Sean Christopherson wrote: > On Wed, Aug 02, 2023, Tao Su wrote: > > Latest Intel platform GraniteRapids-D introduces AMX-COMPLEX, which adds > > two instructions to perform matrix multiplication of two tiles containing > > complex elements and accumulate the results into a packed single precision > > tile. > > > > AMX-COMPLEX is enumerated via CPUID.(EAX=7,ECX=1):EDX[bit 8] > > > > Since there are no new VMX controls or additional host enabling required > > for guests to use this feature, advertise the CPUID to userspace. > > Nit, I would rather justify this (last paragraph) with something like: > > Advertise AMX_COMPLEX if it's supported in hardware. There are no VMX > controls for the feature, i.e. the instructions can't be interecepted, and > KVM advertises base AMX in CPUID if AMX is supported in hardware, even if > KVM doesn't advertise AMX as being supported in XCR0, e.g. because the > process didn't opt-in to allocating tile data. > > If the above is accurate and there are no objections, I'll fixup the changelog > when applying. Totally agree. > > Side topic, this does make me wonder if advertising AMX when XTILE_DATA isn't > permitted is a bad idea. But no one has complained, and chasing down all the > dependent AMX features would get annoying, so I'm inclined to keep the status quo. >From the description of AMX exception, there is no CPUID checking and #UD will be produced if XCR0[18:17] != 0b11. Since user applications should check both the XCR0 and CPUIDs before using related AMX instructions, I don't think there should be bad effects in keeping the status quo. Thanks, Tao
