[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Rapp-dev] SSSE3 alignment and rapp_validate_alignment

From: Hans-Peter Nilsson
Subject: Re: [Rapp-dev] SSSE3 alignment and rapp_validate_alignment
Date: Fri, 26 Sep 2014 16:00:47 +0200

> From: Willie Betschart <address@hidden>
> Date: Thu, 17 Apr 2014 17:52:32 +0200

> Sure I will deliver in separate messages if I would have more than one upda=
> tes at the time.
> I have updated the code after the review comments.

Well, better thanks, but like I said, rapp_validate_buffer
won't be added until we add non-helper functionality to the
library to make a shared-library version bump worthwhile.
The tests were not separated from rapp_validate_buffer; I
didn't add any of that except a rapp_alignment == rapp_align(1)
test to the test mentioned below.

> Regarding SSSE3, we don't use that backend since SSE2 is
> faster, except for those operations with alignment. I found
> the bug when reviewing something else.

Looking closer, there's no bug per se, as __SSE2__ is always
defined whenever __SSSE3__ is, but I've added a "|| defined
__SSSE3__" to avoid future confusion.  I've also added a basic
framework test to catch this type of alignment bug earlier; I
saw it fail for the bitblt functions with an unfriendly failure
mode when I injected this type of bug.  A SEGV with a comment
in the backtrace path seems friendlier.  Using a signal handler
yet friendlier, but doing it portably doesn't seem worthwhile.

> I got about the same results as before, SSE2 was faster than
> SSSE3 when I retried to perform tuning on both backends.

Right.  FWIW, that oddness has been observed before IIRC both by
Almbladh and me.  Not a priority, though.

brgds, H-P

reply via email to

[Prev in Thread] Current Thread [Next in Thread]