CUDA toolkit 3.2 and runtime version mismatch


We are using C2050 with CUDA Toolkit 3.2 on Windows7/64bit machine.
The C2050 driver version is V263.06.

We are having problem with free() call in cuda, and according to the NVIDIA, malloc and free supports were from this CUDA Version 3.2.

Strange thing is that we do not have any issue with malloc calls. Only with free() calls.

When we use GPU SDK’s Device Inquiry, it shows CUDA Toolkit = 3.2, CUDA Runtime Version = 3.1.

We are wondering whethere our prolem with free() has anything to do with the older CUDA runtime version.

Specifically speaking, we are running CUDA from Matlab via MEX. When it reaches free(), the Matlab actually crashes. We verified that the same code can be properly running on Linux 64bit machine (Ubuntu). Therefore, this must be something to do with driver for Windows7.

I’ll appreciate if you have any feedback.

Uninstalling GPU SDK and reinstalling GPU SDK (v3.2) eliminated the mismatch.
Now CUDA Toolkit (compiler) and runtime version are both 3.2.