2017-10-18: VC5 GL merge, GL_MESA_tile_raster_order, GL_OES_required_internalformat

Something went incredibly right, and review feedback poured in last week and I got to merge a lot of code.

My VC5 GL driver’s patches for core Mesa got reviewed (thanks Rob Clark, Adam Jackson, and Emil Velikov), so I got to merge it to Mesa. It’s so nice to finally be able to work in tree instead of on a rebasing branch that breaks most weeks.

My GL_OES_required_internalformat got reviewed by Nicolai Hähnle, so I gave it another test run on the Intel CI farm (thanks, Mark Janes!) and merged. VC4 and VC5 now have proper 5551 texture format support, and VC4 conformance test failures with 565 are fixed.

My GL_MESA_tile_raster_order extension for overlapping blit support on VC4 got merged to Khronos’s git tree. Nicolai reviewed my Mesa implementation of the extension, so I’ve merged it. All that’s left for that is merging the X Server usage of it and pushing it on downstream to Raspbian.

I tested the fast mutex patch series for Mesa, and found a 4.3% (+/- .9%) improvement in 10x10 copywinwin on my Intel hardware. Hopefully this lands soon, since those performance improvements should show up on ARM as well.

On the VC5 front, I fixed VPM setup on actual HW (the simulator’s restrictions didn’t catch one of the HW requirements), getting a lot of tests that do gl_ModelViewProject * gl_Vertex to work. I played around with the new GPU reset code a bit, and it looks like the next step is to implement binner overflow handling.

I’ve been doing some more review feedback with Boris. We’re getting closer to merge on MADVISE, for sure. I respun my DSI transactions fix based on Boris’s feedback, and it’s even nicer now.

Next week: VC5 binner overflow handling, merging MADVISE, and hopefully putting together some Raspbian backports.