DMA'ing into GPU card from another device

Accelerated Computing CUDA CUDA Programming and Performance
1

Does anyone know if it is possible to DMA directly into GPU memory from another PCI card? I guess that to be able to do this we’d need to allocate a block of GPU device memory and then get a handle/address for it which could be passed to the DMA master. Is this possible via the CUDA API or some other means? I assume that the GPU acts as a DMA master for the pinned memory transfer ops (?) but can it be pushed into? Has anyone tried this before?

regards
Mark

2

There is nothing in the CUDA API to facilitate direct DMA transfer between different PCI devices. However, this feature is requested every 6 months, so hopefully someone at NVIDIA is thinking about it.