[lkml]   [2019]   [Jun]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH v2 2/2] x86/cpufeatures: Enumerate new AVX512 BFLOAT16 instructions
On Mon, Jun 17, 2019 at 11:00:16AM -0700, Fenghua Yu wrote:
> AVX512 Vector Neural Network Instructions (VNNI) in Intel Deep Learning
> Boost support BFLOAT16 format (BF16).

That sentence is a mouthful and I have no clue what it means. Marketing
junk? If so, either rewrite it for mere mortals or kill it.

> BF16 is a short version of FP32


Please write out.

> and has several advantages over FP16.


> BF16 offers more than enough range for
> deep learning training tasks and doesn't need to handle hardware exception
> as this is a performance optimization. FP32 accumulation after the
> multiply is essential to achieve sufficient numerical behavior on an
> application level.
> AVX512 BFLOAT16 instructions can be enumerated by:
> CPUID.7.1:EAX[bit 5] AVX512_BF16
> Use word 12, which is empty now, to hold features in CPUID.7.1:EAX
> including AVX512_BF16.

... because that leaf is features only, right?

> Leaf CPUID_DUMMY is renamed as CPUID_7_1_EAX.

That's obvious from the patch, ain't it?

> Detailed information of the CPUID bit and AVX512 BFLOAT16 instructions
> can be found in the latest Intel Architecture Instruction Set Extensions
> and Future Features Programming Reference.


Good mailing practices for 400: avoid top-posting and trim the reply.

 \ /
  Last update: 2019-06-19 19:32    [W:0.066 / U:15.948 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site