Add API to get the physical address of the peer GPU. This should allow some NIC hardware to directly use the physical address for DMA instead of the CUDA Unified Memory provided by current mem_map calls.
A dependent code change has been pushed to the gdrcopy tree on github. The pull request is open here. https://github.com/NVIDIA/gdrcopy/pull/304 Please review. Thanks Ajit Khaparde (2): gpudev: add API to get GPU physical address gpu/cuda: extend cuda code to get PA of GPU drivers/gpu/cuda/common.h | 2 + drivers/gpu/cuda/cuda.c | 117 ++++++++++++++++++++++++++++++------- drivers/gpu/cuda/gdrcopy.c | 50 ++++++++++++++++ lib/gpudev/gpudev.c | 61 +++++++++++++++++++ lib/gpudev/gpudev_driver.h | 6 ++ lib/gpudev/rte_gpudev.h | 49 ++++++++++++++++ lib/gpudev/version.map | 2 + 7 files changed, 265 insertions(+), 22 deletions(-) -- 2.39.3 (Apple Git-146)
