multiple independent CPU processes using data that is in Device Memory

[crroush]

I have a situation where I have large volumes of data to process on the GPU, however I need to be able to have different independent processes on the CPU have access to the data. I want to avoid each process having to transfer the same data to the GPU.

In a perfect world I would push the data to the GPU, process 1 would process with one set of kernels, then process 2 would process the data using different kernels on the same set of data. Much like what you can do with shared memory on a CPU where multiple processes can grab data from the shared memory space, but instead on the GPU.

Is this supported or even possible under the current CUDA framework?

[seibert]

Separate processes definitely cannot share data in device memory. With CUDA 4.0, several host threads in the same process can share a device context. Although I have not tried to use the API in this way, it sounds like it could be done with multiple threads.