[PATCH 1/1] x86: Optimizing memcpy for AMD Zen architecture.

[Florian Weimer]

* sajan karumanchi:

> diff --git a/sysdeps/x86/cacheinfo.h b/sysdeps/x86/cacheinfo.h
> index 7f342fdc23..d6d6877702 100644
> --- a/sysdeps/x86/cacheinfo.h
> +++ b/sysdeps/x86/cacheinfo.h
> @@ -303,6 +303,8 @@ init_cacheinfo (void)
>        data   = handle_amd (_SC_LEVEL1_DCACHE_SIZE);
>        long int core = handle_amd (_SC_LEVEL2_CACHE_SIZE);
>        shared = handle_amd (_SC_LEVEL3_CACHE_SIZE);
> +      unsigned int eax;
> +      unsigned int threads_per_ccx = 0;
>  
>        /* Get maximum extended function. */
>        __cpuid (0x80000000, max_cpuid_ex, ebx, ecx, edx);
> @@ -320,7 +322,7 @@ init_cacheinfo (void)
>  	      threads = 1 << ((ecx >> 12) & 0x0f);
>  	    }
>  
> -	  if (threads == 0)
> +	  if (threads == 0 || cpu_features->basic.family >= 0x17)
>  	    {
>  	      /* If APIC ID width is not available, use logical
>  		 processor count.  */
> @@ -335,13 +337,27 @@ init_cacheinfo (void)
>  	  if (threads > 0)
>  	    shared /= threads;
>  
> -	  /* Account for exclusive L2 and L3 caches.  */
> -	  shared += core;
> -	}
> +	  /* Get shared cache per ccx for Zen architectures */
> +	  if (cpu_features->basic.family >= 0x17)
> +	    {
> +	      /* Get number of threads share the L3 cache in CCX */
> +	      __cpuid_count(0x8000001D, 0x3, eax, ebx, ecx, edx);
> +	      threads_per_ccx = ((eax >> 14) & 0xfff) + 1;
> +	      shared = shared * threads_per_ccx;
> +	    }
> +	  else
> +	    {
> +	      /* Account for exclusive L2 and L3 caches.  */
> +	      shared += core;
> +            }
> +      }
>      }

Although not visible in the patch, these changes a properly guarded by
an arch_kind_amd check, as expected.

Beyond that, I can't comment on the substance of the patch, but I'd like
to request the follow style changes:

* Move the definitions of eax and threads_per_ccx closer
  to their usage site.  Initialize threads_per_ccx directly with its
  final variable.  (The separate variable is nice for documentation
  purposes.)
* Add a space after __cpuid_count (to follow GNU style).
* Add ".  " (period and two spaces) at the end of all new comments.
* Remove Signed-off-by.  glibc does not use DCO
  <https://developercertificate.org/>.  I assume this patch is covered
  by AMD's copyright assignment instead.

I can make these changes for you and push this, or you can post a new
patch.

Thanks,
Florian
-- 
Red Hat GmbH, https://de.redhat.com/ , Registered seat: Grasbrunn,
Commercial register: Amtsgericht Muenchen, HRB 153243,
Managing Directors: Charles Cachera, Brian Klemm, Laurie Krebs, Michael O'Neill