Part 2 of the 10^21 search will be over some time in March this year, so it's time to start getting ready for part 3! I found a few bugs in the current GPU applications that needed to be fixed (CPU applications are fine).

To make sure everything works as intended, I will launch a small beta-test of the part 3 search by the end of this week. The new work units will require ~2.5-2.7 GB RAM on CPU, and the same amount of RAM on GPU (for GPU applications), so the minimum requirement for GPUs will be 3 GB RAM.

The purpose of the test is to confirm that all OS/CPU/GPU combinations work properly and confirm that 3 GB GPUs can process the new work units.

The beta test is running now. I've added 2000 work units to find different ways of how new GPU applications can fail. You can expect many failed tasks there, so if you don't want it, you can disable "Run test applications" in https://sech.me/boinc/Amicable/prefs.php?subset=project

I have run several of the Beta Wus. They all failed like this one https://sech.me/boinc/Amicable/result.php?resultid=72223798
Its a Nvidia RTX 2080TI on a Linux GNOME 44 (Flatpak runtime) [6.5.0-14-generic|libc 2.35] System
Any hints?

"Self-test failed" means that the compiled OpenCL code couldn't process a known amicable number properly and returned false negative for at least one of them. The new GPU application has a built-in list of ~2000 numbers to check all the essential code paths. If you get this error, the only way to fix it is to try different driver versions. The driver/OS combination you use can't compile the OpenCL code correctly.

What did you change? The older Version of the WUs run smoothly. I will not change anything on my system only for your program. The other BOINC Programs run perfect with the combination of the Nvidia driver 535.99 and Linux Mint 21.3

The older version doesn't have this self-test, so even if it ran smoothly it could skip some numbers. Self-test makes sure that the OpenCL code is compiled correctly and works as it was supposed to work.

I've updated the test application to a new version, it has a few bugfixes and it also prints more detailed information if self-test fails. You can test how your computers work with the new version.

This test is to find out how many computers actually fail the self-test. If it prints that your GPU couldn't find some known amicable numbers, then it's useless for the project (with the current application). If it turns out that all Linux hosts fail the test, then I will need to figure out how to fix it, even if it's not a bug in my code (Windows hosts work fine).

Hmm, I'm actually getting useful data now, for example in https://sech.me/boinc/Amicable/result.php?resultid=72251208 there is a list of numbers which are supposed to be amicable, but they're not. It means there is some bug in the self test (in combination with GCC compiler), I'll look into it.

Got this error on a GPU that works perfectly with the regular application:
https://sech.me/boinc/Amicable/result.php?resultid=72251514

Amicable/Amicable_OpenCL.cpp, line 297: Kernel size for AMD GPU has been set to 21
Amicable/Amicable_OpenCL.cpp, line 1339: clEnqueueNDRangeKernel returned error -63
Amicable/Amicable_OpenCL.cpp, line 583: Self-test failed
2024-01-24 04:45:32 (3841781): called boinc_finish(-1)

The self-test should be fixed in v3.06 of the test application, please check it.

I hope this error is fixed in v3.06 as well.

I hope this error is fixed in v3.06 as well.

Yes, fixed!

I hope this error is fixed in v3.06 as well.

Yes, I can say it´s fixed. Thx

One of my computers (hostid=216029 with the GTX 750Ti) is not receiving any new tasks stating: "A minimum of 2700 MB (preferably 2700 MB) of video RAM is needed to process tasks using your computer's NVIDIA GPU"
Although my GPU has 2047 MB it had crunched and validated the new Beta Test Part 3 tasks earlier this evening. So is the memory limit temporary or a permanent change?

It requires ~2550 MB with kernel size 21 and ~2300 MB with kernel size 18, so it's definitely more than 2047 MB. GPUs with <= 2GB RAM can't run it, they will most likely fail with an error.

P.S. Maybe your GPU drivers used system memory?

I am also not being sent any of the original Part 2 tasks that all ran just fine on my GPU. (I run my tasks with kernel size 23.)

edit:
this page https://sech.me/boinc/Amicable/forum_thread.php?id=79 should probably be updated

I will reduce GPU memory requirements back to 1 GB as soon as this test is over, so you can keep crunching until the end of part 2.

Hi, apparently there were 0 macOS GPU tasks distributed, Glops for "Beta test of part 3 of the search (beta test)" macOS = 0

Yes, no one with macOS had test applications enabled.

I just did before I posted but I got no task.

If you enable it I'll be happy to test and let you know :)