Correction: full documentation and testing resources are available at:
https://gitlab.freedesktop.org/virgl/virglrenderer/-/merge_requests/1568
On 2025/12/8 15:50, Honglei Huang wrote:
Dear all,
I would like to share some progress that has been made in exploring GPU
compute capabilities within virtio-gpu. Work has been undertaken to
extend virtio-gpu beyond traditional graphics workloads to support
compute applications, and I wanted to update the community on the
current status and gather feedback on this effort.
Virtio-gpu has the potential to become a unified solution for both
graphics and compute, providing efficient GPU access in virtualized
environments. This work aims to enable modern GPU compute stacks like
ROCm, allowing AI/ML workloads and scientific computing to run in
virtual machines with minimal performance penalty.
The current implementation is functional with essential features enabled
and operational. The implementation has been validated with production
AI workloads. For ComfyUI image/video generation on AMD W7900, the
observed overhead ranges from 0.4% to 8.3% across different tasks. For
LLM inference using deepseek-r1-distill-qwen-1.5b, the implementation
achieved approximately 86% efficiency with 14% virtualization overhead.
For OpenCL, most CTS tests have also been validated, and on certain
hardware platforms, full CTS can be fully validated.
As someone approaching this work as a learner, I would greatly
appreciate feedback from the community. The current functionality and
performance are not bad, there are certainly areas that could benefit
from improvement. I am fully committed to addressing any feedback
received and working with the community to refine the implementation to
meet the expected standards.
Full documentation and testing resources are available at: https://
gitlab.freedesktop.org/virgl/virglrenderer/-/merge_requests/1475
Thank you for your time and consideration. I welcome all feedback,
questions, and suggestions.
Best regards,
Honglei Huang
On 2025/11/12 15:54, Honglei Huang wrote:
This patch series introduces three key enhancements to virtio-gpu to
improve
memory management and GPU virtualization capabilities:
1. VIRTIO_GPU_BLOB_FLAG_USE_USERPTR support: Enables user pointer mapping
for blob resources, allowing guest applications to use user-allocated
memory for GPU resources more efficiently.
2. Configurable HSAKMT capset support: Provides better control over
HSAKMT
functionality with a new device property "hsakmt=on" to avoid
exposing
unsupported capabilities to guests.
3. VIRTIO_GPU_F_RESOURCE_USERPTR feature support: Introduces a new
virtio-gpu
feature flag with configurable "userptr=on" device property to enable
user pointer resources for enhanced memory management.
These patches work together to provide more flexible and efficient memory
management between guest and host in GPU virtualization scenarios. The
changes are backward compatible and controlled by new device properties.
Usage examples:
-device virtio-gpu-gl,hsakmt=on,userptr=on
The series has been tested with GPU workloads requiring advanced memory
management capabilities.
Honglei Huang (3):
virtio-gpu: Add support for VIRTIO_GPU_BLOB_FLAG_USE_USERPTR flag
virtio-gpu: add configurable HSAKMT capset support
virtio-gpu: Add VIRTIO_GPU_F_RESOURCE_USERPTR feature support
hw/display/virtio-gpu-base.c | 3 +++
hw/display/virtio-gpu-gl.c | 2 ++
hw/display/virtio-gpu-virgl.c | 30 ++++++++++++++++-----
hw/display/virtio-gpu.c | 9 ++-----
include/hw/virtio/virtio-gpu.h | 6 +++++
include/standard-headers/linux/virtio_gpu.h | 4 +++
6 files changed, 41 insertions(+), 13 deletions(-)
