qemu-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [PATCH] target-i386: add Intel AVX-512 support


From: Paolo Bonzini
Subject: Re: [Qemu-devel] [PATCH] target-i386: add Intel AVX-512 support
Date: Mon, 27 Oct 2014 16:53:01 +0100
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:31.0) Gecko/20100101 Thunderbird/31.2.0


On 10/27/2014 04:48 PM, Eduardo Habkost wrote:
> On Fri, Oct 24, 2014 at 07:55:10AM +0200, Paolo Bonzini wrote:
> [...]
>> I think we can keep the macros.  The actual cleanup would be to have a
>> single member for the 32 512-bit ZMM registers, instead of splitting
>> xmm/ymmh/zmmh/zmm_hi16.  This will get rid of the YMM_* and ZMM_*
>> registers.  However, we could not use simple memcpy()s to marshal in and
>> out of the XSAVE data.  We can do it in 2.2.
> 
> Now, about the actual 2.3 plans: I believe the xmm/ymmh/zmmh/zmm_hi16
> separation actually makes the code simpler, because the only code
> touching those registers inside QEMU is xsave/vmstate state
> loading/saving, where the bits of the ZMM registers are split into
> different sections/fields exactly in the same way.
> 
> We can still eliminate the YMM_* and ZMM_* macros while keeping
> xmm/ymmh/zmmh/zmm_hi16 as separate fields, and having a single zmm_regs
> member would just add unnecessary complexity to the
> XSAVE<->CPUX86State<->VMState translation code.
> 
> (But if one day TCG starts implementing AVX or AVX512, then it may be
> worthwhile to move the ZMM registers into a single field.)

Yes, exactly---or even sooner, if we start dumping the 256 or 512 byte
registers.  I think the current separation of xmm/ymmh/zmmh/zmm_hi16 is
a KVM-ism.

Let's see what the patches look like, once they are ready.

Paolo



reply via email to

[Prev in Thread] Current Thread [Next in Thread]