I FIXED AMD's Preferred Core Linux patch to FINALLY work for HIGHER BOOST and PERFORMANCE !

trclst@lemmy.ml to Linux@lemmy.ml – 264 points – I FIXED AMD's Preferred Core Linux patch to FINALLY work for HIGHER BOOST and PERFORMANCE !
youtube.com
20

I FIXED my CAPSLOCK KEY to FINALLY enable CRUISE CONTROL for COOL

I can not tell you how many videos and channels I avoid because of that shit.

also ones that have dumb ass thumbnails that do nothing to hint about the content.

Agreed, have you tried Dearrow? It’s from the person who created SponsorBlock and is a crowdsourced way to get rid of the clickbait

I love this addon. It's made my YouTube experience significantly better

If the Daily Mail's editors do it, it must be a good idea!

"This is on version 4 and approved by many people and it doesn't even work. How was this tested?" types furiously in comic sans

This video is killing me. Very entertaining to share their frustration.

Edit: "When we reach 25k subscribers I will debug the Linux scheduler"

AMD's software has always been in a very sorry state. Only after non-inhouse developers created an alternative GPU driver from scratch for Linux it became feasible to use.

I am not the creator of the video but amazed what it all makes it into the stable branch.

TL;DW?

AMD patches for preferred core (prefer those cores which can clock higher) are a mess and ended up not working because of a wrong if condition. Showing that no one at AMD even tested it before submitting. The programmer in the video complains about AMDs developers being incompetent and shows how it's fixed.

Apparently there's a bug in an AMD's driver. It was supposed to assign processes based on each core's self reported performance, but because of the bug it was random.

This "self reported performance" is based on evaluation done to the cores in the fab process, by AMD. Meaning, due to imperfections some cores are a bit better than others.

It's not just random, it simply does not even work. Because they set this:

+/*Preferred Core featue is supported*/
+static bool prefcore = true;

And later in the code they do the if condition wrong:

+	if (prefcore)
+		WRITE_ONCE(cpudata->highest_perf, AMD_PSTATE_PREFCORE_THRESHOLD);
+	else
+		WRITE_ONCE(cpudata->highest_perf, AMD_CPPC_HIGHEST_PERF(cap1));

if should look like this:

+	if (prefcore)
+		WRITE_ONCE(cpudata->highest_perf, AMD_CPPC_HIGHEST_PERF(cap1));
+	else
+		WRITE_ONCE(cpudata->highest_perf, AMD_PSTATE_PREFCORE_THRESHOLD);

There is probably even more wrong, looking at the code quality, but this at least makes the preferred core work.

I don't know if we're discussing semantics. A performance score is attributed, and before the fix their scores were all 166. It doesn't work, as you said. So the consequence is the preferred core being "random", isn't it?

I don't know if it's random, the CPU scheduler still decides what thread to use. It will have its own semantics, but I don't know on what those are based.