This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH][Aarch64] Fix vec_perm cost for thunderx2t99
- From: Anton Youdkevitch <anton dot youdkevitch at bell-sw dot com>
- To: Andrew Pinski <pinskia at gmail dot com>
- Cc: GCC Patches <gcc-patches at gcc dot gnu dot org>
- Date: Tue, 5 Nov 2019 14:48:20 +0300
- Subject: Re: [PATCH][Aarch64] Fix vec_perm cost for thunderx2t99
- References: <20191101140242.GA7200@bell-sw.com> <CA+=Sn1=bxU_+b+U+00rX7NK83Wd2rUxON3XFL5+ZUdY-8Y_85Q@mail.gmail.com>
Andrew,
On 02.11.2019 2:22, Andrew Pinski wrote:
On Fri, Nov 1, 2019 at 7:03 AM Anton Youdkevitch
<anton.youdkevitch@bell-sw.com> wrote:
Hello,
Here is the one-liner that fixes the incorrect
vec_perm cost for thunderx2t99 chip.
With the patch applied 526.blender of CPU2017
gets ~5% improvement with no measurable changes
for other benchmarks.
Bootstrapped OK on aarch64-linux-gnu.
OK for trunk?
Maybe the big question is vec_perm used for both 1 input and 2 input
cases? If so maybe splitting the two cases would be important too.
It is as there is no per-number-of-operands distinction while
computing the vector permutation cost.
However, since 1-operand permutes are rare this would be a good
approximation (statistically).
Otherwise this is ok from my point of view but I can't approve it.
Thanks,
Andrew Pinski
2019-11-01 Anton Youdkevitch <anton.youdkevitch@bell-sw.com>
* gcc/config/aarch64/aarch64.c (thunderx2t99_vector_cost):
change vec_perm field
--
Thanks,
Anton