lkml.org 
[lkml]   [2016]   [Oct]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [tip:x86/urgent] x86/cpufeature: Add AVX512_4VNNIW and AVX512_4FMAPS features

* Luc, Piotr <Piotr.Luc@intel.com> wrote:

> On Mon, 2016-10-17 at 09:55 +0200, Thomas Gleixner wrote:
> > On Sun, 16 Oct 2016, Borislav Petkov wrote:
> > >
> > > >
> > > > The spec can be found in Intel Software Developer Manual or in
> > > > Instruction Set Extensions Programming Reference. See
> > > > https://software.intel.com/sites/default/files/managed/69/78/3194
> > > > 33-025.pdf.
> > >
> > > >
> > > > +/* Intel-defined CPU features, CPUID level 0x00000007:0 (edx),
> > > > word 18 */
> > > > +#define X86_FEATURE_AVX512_4VNNIW  (18*32+2) /* AVX-512 Neural
> > > > Network Instructions */
> > > > +#define X86_FEATURE_AVX512_4FMAPS  (18*32+3) /* AVX-512 Multiply
> > > > Accumulation Single precision */
> > >
> > > This is getting ridiculous: we keep adding new leafs to
> > > ->x86_capability, thus bloating cpuinfo_x86 but then it is not even
> > > worth it - this patch defines only two bits.
> >
> > What's worse is that the Instruction Set Extensions Programming
> > Reference
> > manual says:
> >
> > CPUID.(EAX=07H, ECX=0):EDX[bit 02] AVX512_4FMAPS
> > CPUID.(EAX=07H, ECX=0):EBX[bit 03] AVX512_4VNNIW
> >
> > So AVX512_4VNNIW is in EBX not EDX. What's correct here? The manual
> > or the patch?
> >
> > I'm going to zap it.
> >
> The manual contains the typo in  table 2.1 on page 2.2.
> Please compare it to the detailed description of CPUID in table 4.8 on
> page 2-16.
> There manual groups both new bits under EDX:
>
> EDX    Bits 01 - 00: Reserved
>        Bit 02: AVX512_4VNNIW (Vector instructions for deep learning
> enhanced word variable precision.)
>        Bit 03: AVX512_4FMAPS (Vector instructions for deep learning
> floating-point single precision.)
>        Bits 31-04: Reserved
>
> The typo was acknowledged and is going to be fixed in next version of
> the document. 

All of this should be pointed out in the changelog.

I've zapped the commit for the time being - let's iterate this once more, ok?

Thanks,

Ingo

\
 
 \ /
  Last update: 2016-10-17 10:48    [W:0.054 / U:0.164 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site