2018-10-30: tinydrm, vc4 vs SDL2, V3D

This last week, I worked on the oldest issue I had in my rpi kernel repo: Adding support for tinydrm displays (it actually predates tinydrm). I wrote a tinydrm driver for the panel I had on hand and got it upstreamed. I wrote a Mesa series enabling vc4 to talk to this new hx8357d tinydrm driver. Once the Mesa side is agreed on, then we can extend it to the other tinydrm drivers and have a solution for X with 3D with limited updates (as opposed to full screen refreshing like fbdev did) on these panels.

I also fixed a longstanding vc4 bug that slightly rotated SDL2 output. They were running all of their coordinates through a rotation matrix they built on each VS invocation, and in the common case of no rotation the cos(0)/sin(0) was inaccurate enough to rotate things by a pixel near the window edges. The sin()/cos() in Mesa is now more accurate so this problem shouldn’t happen with old SDL2, we have a piglit test to make sure other drivers don’t break with SDL2, and upstream SDL2 has stopped building a rotation matrix in the shader. (yay!)

Small stuff from the V3D driver since my last post:

  • Fixed depth writes (or depth/color non-writes) on V3D 4.2.
  • Reduced shader recompiles on V3D 4.2.
  • Reduced VPM consumption by vertex shaders using shared NIR optimizations.
  • Took advantage of hardware half-float pack/unpack operations.
  • Switched from FLUSH_ALL_STATE to FLUSH (less overhead, tested by HW team).
  • Merged a fix for use-after-free of sched fences.
  • Merged a fix for debugfs oops on V3D 4.1
  • Merged the kernel V3D clock measurement debugfs entry.
  • Merged more documentation of the SUBMIT_CL.

Also in V3D, I’ve been experimenting with other ways of running the driver using the software simulator. Right now I have a frankenstein mode in vc4 and v3d where the driver sits on top of the i915 GEM driver, and does BO mapping and interactions with the window system on i915 and then copies those BOs into the simulator to run vc4/v3d ioctls on it. This is gross, and it’s only really usable on my desktop with the i915 driver.

Talking with one of the HW engineers who’s interested in running Mesa against the simulator, there are a few ways we could do simulation. One would be to have the actual kernel driver call back up into a userspace daemon linking the simulator, and provide mappings of BOs and proxied register writes to that userspace dameon. Another would be to implement an actual V3D device in QEMU by linking in the simulator, but we can’t do that due to the GPL. Another would be to use my current simulation code, but let it work on top of vkms or vgem (so that a CI system wouldn’t need i915). This one seems easy. However, I started on a project mimicing what Intel did for their software-only CI setup: intercept libc calls to simulate having a V3D driver that’s not actually present in the kernel. I’ve got the code to the point of trying to submit CLs, at which point the simulator hangs for reasons I haven’t debugged yet. I don’t know if I’ll want to do this long term, but it seems like potentially useful code to other driver developers who have a software simulator.

Finally, in the X world, this month I’ve reviewed and merged more patches than I normally would (thanks, gitlab!), and wrote an extensible testcase for the damage extension due to a bug report/fix that a user had been submitted.