This is the mail archive of the
glibc-bugs@sourceware.org
mailing list for the glibc project.
[Bug math/20033] [x86_64] vectorized math function don't call the __finite versions
- From: "jtaylor.debian at googlemail dot com" <sourceware-bugzilla at sourceware dot org>
- To: glibc-bugs at sourceware dot org
- Date: Tue, 26 Jul 2016 16:40:43 +0000
- Subject: [Bug math/20033] [x86_64] vectorized math function don't call the __finite versions
- Auto-submitted: auto-generated
- References: <bug-20033-131@http.sourceware.org/bugzilla/>
https://sourceware.org/bugzilla/show_bug.cgi?id=20033
--- Comment #3 from Julian Taylor <jtaylor.debian at googlemail dot com> ---
The question is what is the intention in the original code.
Why is there a scalar fallback at all? to catch the non finite input or some
other reason.
In this case the movmskps in svml_d_pow2_core_sse4.S:298 which is also present
in the avx2 variant.
I assume this code originates from intel svml? then it would be interesting
what is their documented behaviour for non finite input. Having the same result
would be beneficial for reproducibility and also make libmvec and svml
interchangeable.
Besides this, we currently have a property which is shared between vector and
scalar math and which is also tested so can be assumed to be intentional
behaviour to be relied upon. What is gained by removing it? Is there a relevant
speedup by doing so that makes the divergence worthwhile?
--
You are receiving this mail because:
You are on the CC list for the bug.