Last week I was at the X Developers Conference, where I gave a talk about the vc4 and vc5 projects (video).
Probably the best result of that talk was a developer mentioning to me that the AMDGPU kernel driver’s execbuf model resolved my concerns about the VC5 submit ioctl I was building: All private buffers are implicitly included in an exec, the kernel tracks a list of evicted buffers to be brought back in (probably empty), and it uses a single reservation object for all the non-shared buffers in the address space, so that there are no O(n) walks in the kernel!
As far as code goes, I’m continuing to make progress on the VC5 Vulkan port. This so far involves copying code from the Intel (“anv”) driver trying to get the symbols all implemented, and replacing the core structures I think I need to (batchbuffers and relocation lists turned into a set of command lists and buffers referenced, but without relocations). I’ve also picked a few bits out of the AMD (“radv”) driver, where they’ve done things I need that Intel didn’t have.
Next up in userspace is to get the VC5 code merged (I’m still blocked on review of the NIR alphatest pass, and the patch where I touch common build system stuff to hook the vc5 driver into it), get bcmv to the point where it links (buffer layout is the next big chunk of code to write), finish off state emission, and then start trying to get my first Vulkan triangle rendered.