virgl - exposes GLES3.1/3.2 and GL4.3

I'd had a bit of a break from adding feature to virgl while I was working on radv, but recently Google and Collabora have started to invest in virgl as a solution. A number of developers from both companies have joined the project.

This meant trying to get virgl to pass their dEQP suite and adding support for newer GL/GLES feature levels. They also have a goal for the renderer to run on a host GLES implementation whereas it currently only ran on a host GL.

Over the past few months I've worked with the group to add support for all the necessary features needed for guest GLES3.1 support (for them) and GL4.3 (for me).

The feature list was roughly:
tessellation shaders
fp64 support
ARB_gpu_shader5 support
Shader buffer objects
Shader image objects
Compute shaders
Copy Image
Texture views

With this list implemented we achieved GL4.3 and GLES3.1.

However Marek@AMD did some work on exposing ASTC for gallium drivers,
and with some extra work on EXT_shader_framebuffer_fetch, we now expose GLES3.2.

There was also plenty of work done on avoiding crashes from rogue guests (rewrote the whole feature/capability bit handling), and lots of bug fixes. There are still ongoing fixes to finish the dEQP tests, but it looks like all the feature work should be landed now.

What next?

Well there is one big problem facing virgl in exposing GL 4.4. GL_ARB_buffer_storage requires exposing coherent buffer memory, and with the virgl architecture, we currently don't have a way to map the pages behind a host GL buffer mapping into a guest GL buffer mapping in order to achieve coherency. This is going to require some thought and it may even require exposing some new GL extensions to export a buffer to a dma-buf.

There has also been a GSoC student Nathan working on vulkan/virgl support, he's made some iniital progess, however vulkan also has requirements on coherent memory so that tricky problems needs to be solved.

Thanks again to all the contributors to the virgl project.

Comments

Popular posts from this blog

Fedora 38 LLVM vs Team Fortress 2 (TF2)

tinygrad + rusticl + aco: why not?

nvk: the kernel changes needed