Passthrough/Headless GPU Gets Ahead
summary
Describes a working PV-based path to share guest scanout buffers from a passthrough GPU to a host display without CPU copies by exporting dma-bufs through a virtio-based dma-buf exporter and importing them into the passthrough driver, with metadata passed to expose a host dmabuf for display. Adds blob resources and huge‑page support to eliminate CPU memcpy, uses EGL/native sync objects for producer/consumer fencing, and demonstrates a proof‑of‑concept that largely achieves zero copy.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:50Agenda
- 1:13MultiGPU Platform Introduction
- 3:23Challenges
- 4:17Proposed Solution
- 8:58Background
- 10:56Blob Resources
- 13:06Sync Objects
- 15:36GPU Copies
- 17:08Advantages Cons
- 19:09Summary
- 19:52Thank you
related talks
17mZero-copy display of guest framebuffers using GEM
33mGPU Para-Virtualization on Xen
19mSpeeding Up VM’s I/O Sharing Host's io_uring Queues With Guests
20mSpeed Up Creation of a VM With Pass Through GPU
15mThe Practice Method to Speed Up 10x Boot-up Time for Guest in Alibaba Cloud
23mScalable Work Submission in Device Virtualization
36mWhy Are GPUs (Not) Fast - A Trip Through the Driver Stack
20mMemory Harvesting in Multi-GPU Systems with Hierarchical Unified Virtual Memory