Re: [PATCH 2/2] Aarch64: Add branch diluter pass

[Segher Boessenkool <segher@kernel.crashing.org>]

Hi!

On Tue, Jul 28, 2020 at 07:55:58PM +0100, Andrea Corallo wrote:
> > [ Btw, the mailing list archive will not show your attachments (just lets
> > me download them); naming the files *.txt probably works, but you can
> > also use a sane mime type (like, text/plain) ].
> 
> [ Sure can do it np, I'm just less sure if text/x-diff is such and
> insane mime type for a mailing list software :) ]

Well, first and foremost, it does not *work* with our current mailing
list setup.  Also, anything x- only makes sense if you know the software
the receiver runs (and of course you don't, this is public email, with
hundreds or thousands of readers).

> > Can you share a geomean improvement as well?  Also something like 0.4%
> > is sounds like, or is it more?
> 
> After my first measure I was suggestted by a colleague a less noisy
> system to benchmark on and a more reproducable methodology.  I repeated
> the tests on N1 with the following results:
> 
> | Benchmark      | Est. Peak Rate ration |
> |                |    diluted / baseline |
> |----------------+-----------------------|
> | 400.perlbench  |                1.018x |
> | 401.bzip2      |                1.004x |
> | 403.gcc        |                0.987x |
> | 429.mcf        |                1.000x |
> | 445.gobmk      |                0.998x |
> | 456.hmmer      |                1.000x |
> | 458.sjeng      |                1.008x |
> | 462.libquantum |                1.014x |
> | 464.h264ref    |                1.004x |
> | 471.omnetpp    |                1.017x |
> | 473.astar      |                1.007x |
> | 483.xalancbmk  |                0.998x |

Cool.  Well, gcc has a pretty big drop, what's up with that?  Everything
else looks just great :-)

> I was explained xalanc tend to be very noisy being memory bound so this
> explains the difference, not sure why sjeng looks less good.

Sjeng is very jumpy code, so of course your patch will influence it a
lot.  No idea why it is less positive this run.

> The
> overall ratio comparing spec rates is +~0.44%.

Yup, a nice healthy improvement :-)


Segher