Hi,
On Thu, Jan 28 2021, Richard Biener wrote:
> On Thu, Jan 28, 2021 at 7:32 AM Hongtao Liu via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
>>
>> Hi:
>> GCC11 will be the system GCC 2 years from now, and for the
>> processors then, they shouldn't even need to split a 256-bit vector
>> into 2 128-bits vectors.
>> .i.e. Test SPEC2017 with the below 2 options on Zen3/ICL show
>> option B is better than Option A.
>> Option A:
>> -march=x86-64 -mtune=generic -mavx2 -mfma -Ofast
>>
>> Option B:
>> Option A +
>> -mtune-ctrl="256_unaligned_load_optimal,256_unaligned_store_optimal"
>>
>> Bootstrapped and regtested on x86-64_iinux-gnu{-m32,}.
>
> Given the explicit list for unaligned loads it's a no-brainer to change that
> for X86_TUNE_AVX256_UNALIGNED_LOAD_OPTIMAL. Given both
> BDVER and ZNVER1 are listed for X86_TUNE_AVX256_UNALIGNED_STORE_OPTIMAL
> we should try to benchmark the effect on ZNVER1 - Martin, do we still
> have a znver1 machine around?
Sorry, I kept forgetting about this and when I did not, the machine was
busy. I did just one SPEC CPUrate (single threaded) reference run
without and with the patch (on top of 6b1633378b7), both times with
-Ofast -mavx2 -mtune=generic and with LTO, and the results were actually
rather good (smaller is better):
SPEC 2017 FPrate (time):
| Benchmark | Before | After | % |
|-----------------+--------+-------+--------|
| 503.bwaves_r | 217 | 209 | -3.69 |
| 507.cactuBSSN_r | 236 | 235 | -0.42 |
| 508.namd_r | 252 | 242 | -3.97 |
| 510.parest_r | 384 | 383 | -0.26 |
| 511.povray_r | 486 | 495 | +1.85 |
| 519.lbm_r | 172 | 173 | +0.58 |
| 521.wrf_r | 292 | 277 | -5.14 |
| 526.blender_r | 300 | 303 | +1.00 |
| 527.cam4_r | 255 | 248 | -2.75 |
| 538.imagick_r | 400 | 400 | +0.00 |
| 544.nab_r | 316 | 316 | +0.00 |
| 549.fotonik3d_r | 366 | 351 | -4.10 |
| 554.roms_r | 283 | 248 | -12.37 |
#+TBLFM: $4=100*$3/$2-100;%+.2f
SPEC 2017 INTrate (time):
| Benchmark | Before | After | % |
|-----------------+--------+-------+-------|
| 500.perlbench_r | 446 | 443 | -0.67 |
| 502.gcc_r | 267 | 267 | +0.00 |
| 505.mcf_r | 285 | 285 | +0.00 |
| 520.omnetpp_r | 437 | 436 | -0.23 |
| 523.xalancbmk_r | 302 | 308 | +1.99 |
| 525.x264_r | 217 | 219 | +0.92 |
| 531.deepsjeng_r | 316 | 311 | -1.58 |
| 541.leela_r | 500 | 499 | -0.20 |
| 548.exchange2_r | 314 | 315 | +0.32 |
| 557.xz_r | 391 | 392 | +0.26 |
#+TBLFM: $4=100*$3/$2-100;%+.2f
If we regard any regressions smaller than 2% as noise then there were
none. And 554.roms_r really liked the change, even on znver1.
Martin
