Re: numerous kernel panics when using OpenCL inside FCPX
Re: numerous kernel panics when using OpenCL inside FCPX
- Subject: Re: numerous kernel panics when using OpenCL inside FCPX
- From: Paul Miller <email@hidden>
- Date: Fri, 26 Apr 2013 14:13:37 -0500
- Organization: FXTEC
On 4/26/2013 1:47 PM, Garrick Meeker wrote:
Some drivers do have bugs. The AMD driver in 10.6, in particular, may not be worth supporting. It doesn't support OpenCL images (requiring buffers only) and while we did work around that, we found the OpenCL/OpenGL interop so buggy that we disabled FxPlug hardware rendering for these cards/OS (instead transferring from host memory to the card). AMD 6xxx cards in 10.7 had issues (something I found with local memory) but an update seemed to fix that.
I should have mentioned this is on an NV 330M. I'm getting a *crash*
(instead of the full-on panic) in the same spot on an NV 8600M. I'm
falling back to software if images aren't supported.
However, you might be writing out of bounds to images or buffers, and very bad things can happen. Perhaps the first way to check this is to use the CPU device, call clEnqueueMapBuffer/clEnqueueMapImage on all buffers to print out the address, and see if you get a segfault (and compare that against your known buffer addresses).
I don't think this is happening, unless FCPX is lying about the bounds
of the images. In this particular case, I'm getting a 656x487 FLOAT
buffer from FCPX, and I'm creating an OCL image buffer of the same size
and writing the image and then reading back to the output image of the
same size.
I am running a swizzle kernel, but that's it. The crash happens when I
read back.
BTW, I'm about to open source a OpenCL logging utility that might interest you. It won't detect buffer overwrites (and kernel panics won't leave any log), but it's useful to track down errors on the host side like event handling.
Yeah, I could use something like that that. I've developed a bunch of my
own logging/config stuff as well. The big trick is trying to make it all
"just work", so we don't have to have a bunch of user-adjustable options.
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Pro-apps-dev mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
This email sent to email@hidden