This week I mostly spent on the 7268, getting the GL driver stabilized on the actual HW.
First, I implemented basic overflow memory allocation, so we wouldn’t just hang and reset when that condition triggers. This let me complete an entire piglit run on the HW, which is a big milestone.
I also ended up debugging why the GPU reset on overflow wasn’t working before – when we reset we would put the current bin job onto the tail of the binner job list, so we’d just try it again later and hang again if that was the bad job. The intent of the original code had been to move it to the “done” list so it would get cleaned up without executing any more. However, that was also a problem – you’d end up behind by one in your seqnos completed, so BO idle would never work. Instead, I now just move it to the next stage of the execution pipeline with a “hung” flag to say “don’t actually execute anything from this”. This is a bug in vc4 as well, and I need to backport the fix.
Once I had reliable piglit, I found that there was an alignment requirement for default vertex attributes that I hadn’t uncovered in the simulator. By moving them to CSO time, I reduced the draw overhead in the driver and implicitly got the buffer aligned like we needed.
Additionally, I had implemented discards using conditional tile buffer writes, same as vc4. This worked fine on the simulator, but had no effect on the HW. Replacing that with SETMSF usage made the discards work, and probably reduced instruction count.
On the vc4 front, I merged the MADVISE code from Boris just in time for the last pull request for 4.15. I also got in the DSI transactions sleeping fix, so the code should now be ready for people to try hooking up random DSI panels to vc4.