Announcement

Collapse
No announcement yet.

Peer-To-Peer DMA-BUF Support Being Ironed Out, Patches Pending For AMDGPU

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • Peer-To-Peer DMA-BUF Support Being Ironed Out, Patches Pending For AMDGPU

    Phoronix: Peer-To-Peer DMA-BUF Support Being Ironed Out, Patches Pending For AMDGPU

    One of the latest patch series being worked on by long-time open-source AMD Linux graphics driver developer Christian König is P2P DMA-BUF support...

    http://www.phoronix.com/scan.php?pag...DMA-BUF-AMDGPU

  • #2
    What use-cases does this feature optimize? Is it relevant for single GPU, or for offloading to dedicated GPU from APU?
    ​​​​​​

    Comment


    • #3
      This patch set allows devices to directly access device specific memory on another device (peer to peer access) using the dma-buf infrastructure. E.g., one GPU accessing another GPU's vram directly rather than going through system memory. Or a v4l device to stream directly to an nvme device, etc.
      Last edited by agd5f; 03-12-2020, 09:58 AM. Reason: clarify

      Comment


      • #4
        Originally posted by agd5f View Post
        This patch set allows devices to directly access device specific memory on another device (peer to peer access) using the dma-buf infrastructure. E.g., one GPU accessing another GPU's vram directly rather than going through system memory. Or a v4l device to stream directly to an nvme device, etc.
        Will that speed up PRIME setups?

        Comment


        • #5
          Originally posted by FireBurn View Post

          Will that speed up PRIME setups?
          Yes, it could be used avoid extra copies in some cases.

          Comment


          • #6
            Originally posted by agd5f View Post
            This patch set allows devices to directly access device specific memory on another device (peer to peer access) using the dma-buf infrastructure. E.g., one GPU accessing another GPU's vram directly rather than going through system memory. Or a v4l device to stream directly to an nvme device, etc.
            Blender and other 3D Modeling systems will love it.

            Comment


            • #7
              Originally posted by agd5f View Post
              This patch set allows devices to directly access device specific memory on another device (peer to peer access) using the dma-buf infrastructure. E.g., one GPU accessing another GPU's vram directly rather than going through system memory. Or a v4l device to stream directly to an nvme device, etc.
              You've forgotten all your supercomputer wins...

              BTW Sady, Europe (Germany) is near to Shutdown (Austria is still there, now).

              Comment

              Working...
              X