I'm having a problem with a recently acquired HD7970 in a linux machine: quite often, usually after a few (non-opencl) applications have run, but other times right off the bat. The problem is that no OpenCL programmes will run against the GPU: they all fail when calling clCreateCommandQueue with an error CL_OUT_OF_HOST_MEMORY.
Simple samples like 'Reduction' also fail in the same way. The machine has a 64 bit intel cpu & 12gb ram and all the software i'm using is 64-bit: it is not short on free ram. The system has been crashing a lot so it's usually a fairly recent reboot with not much running. I'm not using any GL-compositing desktop, just plain X.
Rebooting, or even just restarting the X server seems to fix it (usually).
I only recently upgraded to 12.4, previously with a HD 6950 I didn't see this problem unless a few CL applications were active.
driver version: 8.961
I didn't see the problem when I swapped the card with a HD6950 with this driver, at least while I had it there a day or so.
I just noticed another problem with 12.6 - somecode i'm fairly confident is ok and i had working on 12.4 and earlier now doesn't work on 12.6 beta ... although if i add a bunch of clFinish() calls it works ok. It's all using a single queue.
I seem to have narrowed it down to a bit of code that uses a global atomic int32 in multiple calls of the same kernel. If I call clFinish() after invoking this kernel every time then it works, otherwise it crashes and brings down the system. I don't have a small test-case, but it is in public (java) code.
Should I be using the problem reporting form?http://www.amdsurveys.com/se.ashx?s=5A1E27D20B2F3EAC
I had a look at it, but it doesn't look very developer-oriented.
Yes, please use the AMD issue reporting form that you linked to above.
I'll double check with Support just to be 100% sure that feedback submitted via this form will get to the Catalyst developers.