PDA

View Full Version : OpenCL on Apple imac with ATI Radeon HD 4850 is badly broken



FuriousFingers
04-15-2010, 08:42 AM
I have tested some demos on Snow Leopard (10.6.3). Some work fine but some many examples seem to crash AMD's compiler. Even Apple's own OpenCL_Procedural_Grass demo is crashing. Most of the demos from here: http://davibu.interfree.it/index.html do not work either. (They had to be modified a bit to run on MacOS but most give compiler crashes).

Can't believe how Apple could push such bad drivers when they are the ones that spearheaded the OpenCL effort :(

Anyone knows if this crashing bug is confirmed? Any idea when AMD will fix this? And how soon Apple will push the updates.

YOYOSENIOR
04-16-2010, 06:56 AM
ATI drivers on OS X doesn't implement context sharing between OpenGL and OpenCL, you must check if the Apple gl sharing extension is available.

And the second issue is because ATI cards doesn't support image 2D and Image 3D, if the kernel
use this kind of cl memory the kernel cannot be executed.

It's exactly the same issue on Windows, ATI need to work on their driver but they focused their effort on 5xxx series which is actually not available on Mac....

FuriousFingers
04-17-2010, 09:29 AM
What? no context sharing either :s. But anyways, that's lack of features. I'm not complaining about that (currently). This is about stability of ATI's compiler. All those crashes occur during kernel compilation. Not during kernel execution. Their compiler is in very bad shape and from what it looks like they are actually using LLVM so, its not like they are developing it from scratch. First time I'm using an ATI GPU and this is my experience *sigh*.

dphil
04-17-2011, 07:09 PM
I have been getting the same problems with my 4850. I get strange compilation errors when I try to compile certain kernels. I say strange, because after much testing and fiddling with the kernel, there doesn't seem to be any rhyme or reason to it. It compiles ok if I comment out one of several bits of code, all of which are harmless (trust me, I know; their calls/calculations are repeated in different forms all over my kernel, and I've triple-checked the var names, types, etc).

Meanwhile, the same kernel compiles fine on another computer with a GeForce GT 120. Is there some way of resolving this?

david.garcia
04-18-2011, 09:02 AM
I would notify Apple's developer support giving them a short kernel that reproduces the issue.