problem of cuda when used in shared library

I tried simple cuda operation in a shared library:

CUDA_SAFE_CALL( cudaMallocArray( &cu_Index, &index.channelDesc, _xIndex, _yIndex )); 

Then the host program gives the following error:

Non-handled exception in Test.exe at 0x03b638d1: 0xC0000005: access confliction when reading 0x00000000

Before this is executed, content of cu_Index is 0x00000000, but I think this is fine since we passed in its address.

After I use the same code to a static linked library, it works fine.

So, is there anything I need pay attention when use CUDA in a shared library ( dll )?

I was using WinXP Pro, VC8, 8800GTS, cuda0.8


This is a known bug (see ).

A fix will be available in the near future.

Thank you for your information.

May I know when this will be released?