log in |
1)
Message boards :
Number crunching :
Question/Problem with Intel GPU work units
(Message 9663)
Posted 6 Feb 2024 by AndyPC It looks like that integer is GPU_type.CE_per_multiprocessor, but where is it consumed? I only see that name referenced twice in the whole repo, both in mfakto.cpp. One is setting it and the other is printing it. |
2)
Message boards :
Number crunching :
Question/Problem with Intel GPU work units
(Message 9658)
Posted 5 Feb 2024 by AndyPC Trying to work on debugging this a bit. I recompiled mfakto from current Github sources on Pop!OS (based on Ubuntu) 22.04 LTS with an Intel Arc A770 16 GB gpu. CPU is AMD Ryzen 7 7700, OS kernel is 6.6.10. Turned on as much debug output as I could find including #define DETAILED_INFO. Self-test is failing using both OpenCL 2.2.0 from ROCm and OpenCL 3.0 from Intel Compute Runtime. Here's the output from the latter: https://gist.github.com/APCBoston/580385ff898cc013b34e2e7570b8b468 I'm not entirely satisfied with the explanation that Intel's drivers are bad, as Phoronix is succesfully running vector compute benchmarks on this GPU... either there's a very specific and obscure bug in the driver that is only showing up on this application, or, I suspect more plausible, a bug in mfakto. Will try working through this in gdb when I have a bit more time. |