broken OpenGL interop why release CUDA 2.2 ?

Hi there,
I’m trying to understand the reason why NVIDIA did no beta testing of CUDA 2.2 for Mac but sudently released it… with completelly broken OpenGL interop.

I think CUDA 2.0 also had broken OpenGL interop, CUDA 2.1 fixed that, now 2.2 broke it again. So this probably meany we will have to wait several months for CUDA 2.3 for this to be eventually fixed.

Come on - who approved this for release?

Very disppointing.

Mark

what’s wrong with it? the SDK samples work fine

The release notes say OpenGL interop is not working.

The SDK apps I tested that use OpenGL interop show abismal performance - giving the impression that reads/writes to OpenGL buffers are going through the host.

Or did I read the release notes wrong or have another problem?

Mark

MacPro 2008 with GT8800, OSX 10.5.7, CUDA 2.2

“OpenGL interop will always use a software path leading to reduced performance when compared to interop on other platforms.”

That doesn’t mean it’s broken, that means it’s not as fast on OSX (because of its driver layout) compared to other platforms.

I do not have access to my MacPro for the next few days, but I remember testing Nbody and it runns at about 5 FPS in CUDA 2.2.

In CUDA 2.1 (where OpenGL interop was working fine) it was running at 30 FPS IIRC.

So I am talking about something that worked well in CUDA 2.1 and now is not in CUDA 2.2 - thus the use of the word “broken”.

This really severelly impacts ANY application visualizing CUDA computations.

Could you explain why this stopped working? Or if there are any plans to release a fix, or if we have to wait for CUDA 2.3 to hopefully fix this?

Mark

I’m getting 104 FPS on my MacBook Pro (8600 GT with 256MB), so I bet what happened is that when you upgraded CUDA.Kext did not get installed correctly.

I did some tests comparing CUDA 2.1 and 2.2 on OSX vs CUDA 2.2 on Windows XP.

All samples computing stuff (not visualizing it) seem to perform almost the same on these setups.

The samples using OpenGL have very bad performance with CUDA 2.2 on OSX:

smokeParticles - with 65536 particles

OSX CUDA 2.2 = 4.4 fps

OSX CUDA 2.1 = 17.1 fps

XP CUDA 2.2 = 64.0 fps

volumeRender

OSX CUDA 2.2 = 35.0 fps

OSX CUDA 2.1 = 108.0 fps

XP CUDA 2.2 = 64.0 fps

particles - with 32768 particles

OSX CUDA 2.2 = 63.0 fps

OSX CUDA 2.1 = 170 fps

XP CUDA 2.2 = 280.0 fps

This is on a MacPro1,1 with a 8800 GT (PCIe 1.x version).

Is anybody else experiencing this?