GPUDirect and OpenCL?

Is it possible to use the NVIDIA GPUDirect features also with OpenCL on a Tesla GPU? I want to develop a HPC application using OpenCL because I am more familiar with it but haven’t been able to find information about how to use GPUDirect with OpenCL (only CUDA information is available). Any suggestions?

I fear that top-of-the-line features such as GPUDirect are not implemented in OpenCL. If it were, it would be a platform extension, as it requires a corresponding CUDA function call. It is either default method and the runtime tries to utilize it where possible, or it would have a seperate host-side function.

There is a really strong feel to NV’s OpenCL support, that it is intentionally being held far behind CUDA. This is simply a defensive mechanism to prevent customers shifting to OpenCL, as more and more OpenCL enabled products hit the market, it opens more and more opportunities for customers to shift to other vendor’s products as viable alternatives to NV HW. Therefore NV tries to prolongue proper OpenCL support as long as possible. NV software team is extremely capable and implement kick@ss features in practically no time. nvcc is an extremely capable compiler, whereas NV OCL compiler crashes at every second corner and cannot give understandable error messages either. I cannot imagine that this is by accident. Porting a far more capable API to a somewhat more humble API cannot be this difficult and take this long time.

Although should someone know of a way to get it working, I would be interested in it also.

Thanks for your comment! I will let you know if I can find out more about that topic.