Benchmarking AMD's New AOCC Compiler For Ryzen

[Niarbeht]

https://community.amd.com/servlet/Ji..._13_33_Pro.jpg Help!
I need help! I can't use any distro of Ubuntu or similar. Use linux mint 18.1 with kernel 4.4. I cannot measure temperature or anything or nothing in linux mint. I tried to use ubuntu "any distribution" without success. Ryzen 1700, gigabyte ax370 gaming 5.

I.... think you posted in the wrong thread?

[wizard69]

Interesting. Looks like this is more of a beta release with development ongoing. It will be interesting to see if AMD can get real improvements out of code generation and thus better performance in a few months.

[chimpy]

https://community.amd.com/servlet/Ji..._13_33_Pro.jpg Help!
I need help! I can't use any distro of Ubuntu or similar. Use linux mint 18.1 with kernel 4.4. I cannot measure temperature or anything or nothing in linux mint. I tried to use ubuntu "any distribution" without success. Ryzen 1700, gigabyte ax370 gaming 5.

I thought you needed a newer kernel for Ryzen:

https://www.phoronix.com/scan.php?page=news_item&px=AMD-Ryzen-Newer-Kernel

[Brutalix]

https://community.amd.com/servlet/Ji..._13_33_Pro.jpg Help!
I need help! I can't use any distro of Ubuntu or similar. Use linux mint 18.1 with kernel 4.4. I cannot measure temperature or anything or nothing in linux mint. I tried to use ubuntu "any distribution" without success. Ryzen 1700, gigabyte ax370 gaming 5.

Well ubuntu has enabled some amd gpio flags in their kernels, which give the Irq trap on the gigabyte motherboards. Either switch to debian or some other non ubuntu distro, or compile your own. This is true for all the am4 gigabyte motherboards.

Kind regards
Brut.

[reavertm]

Would be interesting to see ICC in here. (and funny to see and quite possibly outperforming AOCC even at AMD hardware).

[Marc.2377]

Depends on the mtune setting, really. Gcc's znver1 model is largely copied from bdver4 (Carrizo/Excavator), and that sucks on Zen. -march=znver1 -mtune=haswell gives much better results. Noone has submitted a more accurate zen scheduling/cost model yet.

I remember reading your post here and it mentions "-mprefer-avx128". Does that flag still help in this newer case?

Would be interesting to see ICC in here. (and funny to see and quite possibly outperforming AOCC even at AMD hardware).

A bit sad to admit, but indeed I find this very likely.

[nevion]

Michael did you receive any suggested settings from AMD devs? You know, like -frelease-the-kraken?

[Michael]

Michael did you receive any suggested settings from AMD devs? You know, like -frelease-the-kraken?

No information at all.

[nevion]

Michael that's unfortunate... here's a few flags I think would be good to try after skimming the documentation It's not really clear what settings were used on the compiler from the result set. I'm guessing -O3 and not much else. No idea if any of those will help on anything but something should be advanced here and maybe one of these options stumbles onto it. The strided-vectorization feature is AOCC specific so it's worth a try.

-march=znver1 -O3  -mllvm -enable-strided-vectorization

[mlau]

I remember reading your post here and it mentions "-mprefer-avx128". Does that flag still help in this newer case?

I read Agner Fog's analysis of Zen, and he concluded that feeding Zen with 256bit AVX is worthwhile. So I dropped this flag.