Speaker
Description
Shared Virtual Memory (SVM) enables a unified virtual address space for heterogeneous computing, but existing implementations remain outside the upstream DRM framework and face scalability challenges. In particular, KFD SVM relies on mmap read/write locks for synchronization, which tightly couples driver-level behavior with the system memory management and increases overall design complexity.
At XDC 2025, an initial DRM SVM framework was introduced based on Intel Xe. In this talk, we build that foundation by evolving SVM into a unified DRM-based abstraction and demonstrating it on AMDGPU.
We highlight several key architectural improvements. First, DRM SVM adopts a significantly simplified synchronization model, avoiding heavy mmap locking. Second, it is designed with native large page support, improving efficiency without invasive changes. Third, it introduces a cleaner abstraction by separating memory attributes from the physical layer, combined with a range-based model (up to 2MB granularity).
We also demonstrate both fault-driven (XNACK-on)[1] and non-fault-driven (XNACK-off) [2][3] execution models and discuss ongoing work to extend DRM SVM beyond its current single-device scope toward multi-device support[4].
This work advances DRM SVM toward a scalable, vendor-neutral, and upstream-aligned memory model.
[1] https://lore.kernel.org/dri-devel/20260612090928.29682-1-ray.huang@amd.com/
[2] https://lore.kernel.org/amd-gfx/20260519092313.3115848-1-honglei1.huang@amd.com/
[3] https://lore.kernel.org/amd-gfx/20260605075340.20199-1-Junhua.Shen@amd.com/
[4] https://lore.kernel.org/dri-devel/20260603065620.2555316-1-honglei1.huang@amd.com/
| Code of Conduct | Yes |
|---|---|
| In-person or virtual presentation | In-person |
