Hello, this is my first post on this forum, so I hope I don’t make too much of an embarrassment out of myself right from the beginning :)
Anyway, my issue is that I have been trying to find information on CUDA 2.1’s new features, particularly the JIT functionality (to replace the modules I’m using at the moment). Now, the 2.1 toolkit, driver and SDK can all be found on the CUDA site, but there is no documentation of any kind visible yet (other than for the profiler).
So, where can I learn more about these features, and when can we expect the full documentation for 2.1? :)
Not really finding what I’m looking for in the doc so far… perhaps I’m not looking at it correctly. The JIT info wasn’t added to the modules section was it? It should be in the Runtime section at least, I think…
Pretty sure that the JIT stuff is only accessible from the driver API (how would it make sense to have the runtime API load modules and such? The runtime API doesn’t have any user-accessible concept of contexts or modules).
It’s there, look in the CUDA Reference Manual. It’s just a new function that loads PTX files instead of cubins. This is important, for future compatibility. Runtime API doesn’t need it because it works with “fatbins” which are more advanced than cubins and incoporate both PTX and pre-compiled machine code.
P.S. Runtime API can also work with external modules via the Device Code Repository. This is an automated system and accepts both PTX and machine code files.