Re: [PATCH] cksum: Use pclmul hardware instruction for CRC32 calculation

[Pádraig Brady]

On 13/03/2021 20:37, Jim Meyering wrote:
> On Sat, Mar 13, 2021 at 10:19 AM Pádraig Brady <P@draigbrady.com> wrote:
> > On 13/03/2021 16:13, Pádraig Brady wrote:
> >
> > FYI testing on an older i3-2310M system
> > shows the bottleneck is not near I/O (cat is much faster).
> > A 500MiB file improves from 1.40s to 0.67s on the i3-2310M.
> >
> > $ time src/cksum file.in
> >     3404199294 524288000 file.in
> >     real 0m0.672s
> >     user 0m0.584s
> >     sys  0m0.084s
> >
> > I'm also considering applying the attached
> > to add a --debug option (present on a few other coreutils),
> > which will diagnose the implementation used
> > (since it's build time and run time variable).
>
> I like the new option, and the patch looks fine.
> I assume you'll mention the addition in NEWS.

Thanks for the reminder. I've adjusted like:


diff --git a/NEWS b/NEWS
index aad05df6d..5368a3eed 100644
--- a/NEWS
+++ b/NEWS
@@ -70,7 +70,9 @@ GNU coreutils NEWS                                    -*- 
outline -*-
   cat --show-ends will now show \r\n as ^M$.  Previously the \r was taken
   literally, thus overwriting the first character in the line with '$'.

-  cksum is now up to 4 times faster by using a slice by 8 algorithm.
+  cksum is now up to 4 times faster by using a slice by 8 algorithm,
+  and up to 8 times faster where pclmul instructions are supported.
+  A new --debug option will indicate if pclmul is being used.

   df now recognizes these file systems as remote:
   acfs, coda, fhgfs, gpfs, ibrix, ocfs2, and vxfs.