Radeon 5700 XT : support in beta testing (FahCore_22)

[bruce]

Yes.

Several of you who are affected (and anybody who might buy one for FAH are included) should start a discussion on their support website saying OpenCL doesn't work reliably. Somebody can mention FAHBench, too.

[foldy]

I created one forum thread at AMD and one email customer complaint for driver 19.7.2. I mentioned FAHbench to reproduce too. Lets see if someone answers.

[bruce]

Even if they don't answer, AMD Developers will see it and start planning for a fix.

"Me too" complaints might raise the priority -- if only because FAH Donors will stop buying the latest AMD GPUs. FWIW

[foldy]

@Sven: Can you make a screenshot showing the FAHbench error? Then I can send it to AMD support.

[Sven]

Here the Errormessage:

Force RMSE error of 2.42883e+07 with threshold of 5 / status failed / single precision

Have to figure out a way to post Images.

[foldy]

You can use https://imgur.com/upload and then past the URL here.

[Sven]

Here the screenshot, forgot that I could use google foto sharing, as I didn't want to register at imgur.

https://photos.app.goo.gl/mGpfghzEqanmgNw79

[bruce]

BTW, there are virtually no new projects who do not require double precision, so validating that FAHBench runs in single isn't as useful as it used to be. That also means that OpenCL 1.1 installations need to be upgraded to 1.2

[foldy]

So FAHbench double precision mode matches FahClient mixed mode with lots single precision and some double precision? And FahBench single precision matches FAhClient single precision only?

@Sven: I sent your screenshot of FAHbench error to AMD support.
AMD driver 19.7.3 was released so you could test with FAHbench again.
If issue is reproducible then give AMD feedback https://www.feedback.amd.com/se/5A1E27D207AEC5E5

[bruce]

The GPU cores are being compiled with what they're calling MIXED PRECISION. This means that a major portion of the calculations are performed in single-precision, but a small fraction of them will be done in double precision when that's important to the success of the WU. In traditional FAH calculations (and by that, I mean CPU calculations) this was easy to do because CPUs have always supported both. In the early days, FAHCore for CPUs was complied for Single and then an experimental version was compiled for Double. It was terribly slow (and FAHBench may do that now) but the added precision doesn't produce any better science.

To produce Mixed Precision code, individual calculations had to be examined as to whether the extra time required by more precise calculations was useful or not and manually coded when appropriate.

[Sven]

@foldy: Nope, no luck with 19.7.3, same error in FahBench. Tested with DoublePrecision

Edit: I filled out the AMD Form with details description. I will report here if/when I get an answer.

[Sven]

No progress with 19.7.4, still the same FAHBench error.

[bruce]

AMD has a history of releasing a whole host of patches all at once. Do you suppose that will happen again?

https://techreport.com/news/31627/amd-p ... -on-linux/
https://phoronix.com/scan.php?page=news ... 10-Support

[Sven]

Hard to say, but I guess AMD has other software priorities at the moment.

They are having troubles with the AGESA-Code für the Ryen 3000 CPU's. Some Linux distro's couldn't start because of problems with the Ryzen microcode. They release a new one for the board makers, but the one contained another bug, so it was pulled again. So a lot of construction work at AMD software/driver at the moment.

I guess that's top software priority at the moment. We may have to wait longer for a Radeon driver that works correctly with the 5700 xx.

[foldy]

CPU and GPU have different AMD teams. There may be not enough support requests for the FAH issue currently to get higher priority. It is also possible that AMD OpenCL driver works as specified but FAH has a bug which only showed with the Radeon 5700 XT.