So in the interest of better Mac dev communication, here’s a brief update of what’s going on with CUDA and SL right now:
Drivers: 2.3.0 definitely doesn’t work with Snow Leopard, 2.3.1 should? Honestly, I’m not sure there because I haven’t switched on my laptop yet. Regardless, if there’s not now, there will be a working driver version in the near future.
Compiler: the new gcc version has some incompatibilities with nvcc. We’ve fixed these internally already.
64-bit: Right now, you have to boot SL into 32-bit mode (which is apparently not the default for recent Mac Pros) to use 2.3. This will be fixed in the next release (or in a point update soon after–the CUDA team is not the blocking factor here).
We’ll be releasing an update to 2.3 to fix #1 and #2 late next week. It will fix the installer and permissions issues as well. This will probably be a semi-silent update–filenames of the libraries will be the same, some version numbers will change (nvcc will show a different build date, for example). I’ll try to make sure that there’s a different name for the package because I hate this “download the same thing again, see if it’s actually the same thing” thing as much as you guys do. Probably more, because everybody complains about it to me…
Does anyone know how long it’ll be? My 10.6 OSX disc came today and CUDA support is the only thing stopping me from upgrading. I’d do all the gcc symlink tricks but I’d like to keep the defaults the way they are and just use it with gcc 4.2
no, but changing symlinks isn’t a trick so much as it is a feature. Apple includes both compilers and uses the symlinks so that the end user can choose which to run by simply changing that symlink.
With the new packages, I managed to get my machine into the bad installer state that a lot of you have hit. Hooray! However, that means that something is broken, and unless I can concretely determine that it’s my machine (and how to fix it) I don’t think packages are going to be out today.
SORRY GUYS. I am more annoyed about this than you are, I promise.
I must confess though that what I really want to see is a full migration to OpenCL under Snow Leopard of all the nice SDK examples. Apple have not exactly released many examples, and while it is good to see the OCL working under Windows I have not had much luck migrating them to SL. I realize you will want to fix the CUDA issues first. Are Snowie versions of all these programs hiding somewhere? I still cannot find them despite the recent announcement of “comprehensive support”.
I adventured with the NVIDIA OpenCL examples on SL too. I din’t get far yet, but just looking at the bandwidthTest I found many bugs. Device->Device was not working at all due to an OpenCL buffer being wrongly initialized. The timings are all wrong since the timer used for profiling is not accurate enough - so at first OpenCL seems to perform worse than it actually does. Using clEvent’s shows the truth.
I’m currently planing to write a performance test app from scratch to compare both:
memory speeds (mostly done, but not well coded)
OpenGL interop speed
kernel execution speed (just some kernels tricky ones that can be greatly optimized like transpose, histogram, …?)
2D texturing from linear memory vs reading/writing to image objects
If you happen to be doing/planing similar things maybe we can work together on this?