Speaker
Description
Heterogeneous Memory Management (HMM) has been part of the Linux kernel for some time, with prior GPU-focused implementations in AMDKFD, Nouveau, and OpenNV. However, integrating HMM correctly—ensuring proper API usage, race handling, and locking—is notoriously complex. This talk introduces GPU SVM, a new abstraction layer that streamlines interactions between HMM and the core memory management (MM) subsystem, offering a more accessible and safer API for DRM driver developers. We will walk through the design of GPU SVM, its initial deployment in the Xe driver, key challenges and learnings from the integration process, and discuss future directions for shared virtual memory support in the DRM ecosystem.
[1] https://www.kernel.org/doc/html/v5.0/vm/hmm.html
| GSoC, EVoC or Outreachy | No |
|---|---|
| Code of Conduct | Yes |
| In-person or virtual presentation | In-person |
