This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [patch, ARM] Make 128 bits the default vector size for NEON
- From: Ira Rosen <ira dot rosen at linaro dot org>
- To: Hans-Peter Nilsson <hans-peter dot nilsson at axis dot com>
- Cc: gcc-patches at gcc dot gnu dot org, patches at linaro dot org
- Date: Thu, 7 Apr 2011 08:34:19 +0200
- Subject: Re: [patch, ARM] Make 128 bits the default vector size for NEON
- References: <AANLkTinazE6UQu_WE_iwCFscjasV-CHNqNHZ5bO2zhdT@mail.gmail.com> <201104061407.p36E7bDe019262@ignucius.se.axis.com>
On 6 April 2011 16:07, Hans-Peter Nilsson <hans-peter.nilsson@axis.com> wrote:
>> Date: Thu, 31 Mar 2011 13:39:05 +0200
>> From: Ira Rosen <ira.rosen@linaro.org>
>
>> This patch changes NEON's default vector size from 64 to 128 bits.
>
> I'm wondering, are there NEON-specific measurements to support
> this change?
>
> A colleague of mine implemented support for 64- and 128-bit NEON
> for RAPP <http://savannah.nongnu.org/projects/rapp/>, but found
> that the 128-bit version was slower.
The best vector size depends on the particular target (a version of
ARM) and the benchmark. The vectorizer has a cost model that allows it
to fall back to 64-bit vectors if 128-bit version is not profitable
(-fvect-cost-model). We plan to enhance the cost model to estimate
both versions and vectorize with the most profitable vector size. For
now, you can use preferred-vector-size param if needed. Having 64 as a
default doesn't allow us to use different NEON instructions that use
both types of vectors.
Ira
>
> brgds, H-P
>