Using CUDA to obtain physical address of texture?

Using CUDA, is it possible to obtain the physical address of a texture?

Specifically I’m trying to check if it’s possible to DMA data directly from a device on the PCI bus to the GPU on the same bus. So far I’ve only seen examples where the data is handled by the CPU (which involves higher CPU utilization).

Unfortunately not. We also don’t expose the internal format of textures in memory, so even if you did know the address you wouldn’t be able to write to it in the correct layout.

It is possible for PCIe devices to DMA into each other’s address spaces and we are investigating ways to expose this functionality.