As you may know, depending on the NVIDIA drivers, the Julia4D OpenCL demo of GPU Caps Viewer works fine (developer drivers) or crashes (public drivers).
Here is the explanation about the crash by NVIDIA:
This has been fixed in an upcoming driver. The problem with the OpenCL test in GPU caps viewer is that the code generation for the vector store is trying to optimize the case where the RHS completely overwrites the LHS. When checking for this, it compares the sizes of the destination and source vectors and goes down the optimized path if the sizes match. However, it uses the actual size of the vector3 (=4), rather than the accessible size (=3) and compares it to the size of the destination vector (=4).
Here, LHS stands for left-hand-side and RHS for right-hand-side and both terms reference the both side of an equation.
Anyway, the good news is that we’ll have the fix in the next R26.1.xx WHQL…
so?…this means that gpu caps viewer code is not optimised? or is it that the driver wont work on un-optomised paths?
Julia4D demo uses an OpenCL kernel from Apple’s SDK, without modification (the kernel source code is in the media folder of GPU Caps). This kernel works and worked fine on Radeon and GeForce, but recently some optimizations in NV drivers have introduced a bug that is visible with the Julia4D demo of GPU Caps Viewer.
ahh..i forgot the Julia4D code was from the AppleCL SDK…
i knew it couldnt be your code causing probs! 😉
thanks for the clarification
Ok.. now please explain why FluidMark crashes with more than 7 emitters 😀
@Zibri: http://www.geeks3d.com/20100524/physx-fluidmark-gpu-physx-is-fast-but-has-limitations-you-must-know/
Oh thanks I think I missed that one 🙂