Re: AVX-512 instruction set

[Paolo Bonzini]

On Wed, Apr 12, 2023 at 3:41 PM Walid Ghandour
<waleedjghandour@gmail.com> wrote:
> I will try to work on this.

Even if the individual tasks I listed are not enough to implement
AVX512, I suggest that you post individual pieces of work so that you
can somewhat pipeline the work.

Also, please review the design of the AVX decoder and the
implementation of the translation and the tests. You can find the
patches at https://patchew.org/QEMU/20221013214651.672114-1-pbonzini@redhat.com/
(patches 3-4-6-8 especially).

Paolo

> Regards,
> Walid
>
> Le mer. 12 avr. 2023 à 15:30, Paolo Bonzini <pbonzini@redhat.com> a écrit :
>>
>> On Wed, Apr 12, 2023 at 2:17 PM Alex Bennée <alex.bennee@linaro.org> wrote:
>> > I don't think there is currently any active effort to add AVX512
>> > support. There have been various GSoC projects to improve the x86 SIMD
>> > emulation but I don't think they got merged.
>>
>> No, there isn't. However, the recent implementation of AVX in QEMU 7.2
>> is designed to make AVX512 at least doable.
>>
>> Adding support for AVX512 would be a very large work (at least 1
>> months full time plus time to get it merged), but not impossible. The
>> tasks could be something like this:
>>
>> 1. EVEX prefix decoding for AVX and AVX2 instructions
>> 2. operand broadcast
>> 3. VEX encoded mask instruction: kmov, kadd, kxnor, etc
>> 4. other instructions and permutations with mask operands and permutations
>> 5. opmask support including merging and zeroing
>> 6. Disp8*N addressing mode
>> 7. rounding control and exception suppression
>>
>> It's important that, at any given step, the new functionality is
>> tested comprehensively.
>>
>> > > 62 f2 7d 48 18 0d fa 0c 00 00 vbroadcastss 0xcfa(%rip),%zmm1
>> > >
>> > > qemu: uncaught target signal 4 (Illegal instruction) - core dumped
>> > >
>> > > I like to add support for broadcast and fmadd avx 512 instructions such 
>> > > as the following one:
>> > >
>> > > 62 e2 7d 48 b8 c9     vfmadd231ps %zmm1,%zmm0,%zmm17
>>
>> Both of these are using a small subset of AVX512 (step 1 above). Both
>> vbroadcastss and vfmadd231ps are already implemented in QEMU, but not
>> using ZMM registers.
>>
>> Unfortunately the base AVX512 functionality is large, and therefore
>> the above tasks are all needed to claim support for AVX512. On the
>> other hand, for historical reasons AVX512BW and AVX512VL extensions
>> are separate but in practice they are easier to just implement at the
>> same time as basic AVX512; my expectation is that they would come up
>> almost for free with the rest of the work.
>>
>> Paolo
>>