Weekly reports for 2016

November 28th - December 4th

Parts for new system finally arrived, spent a few days putting together, getting os installed and setting up network, setting up dev environment.
Took two whole days to get AMD drivers working, ended up getting them working on Arch rather than Ubuntu.

Next week: update cpu split kernel diff and start on memory / work scheduling

Updated cpu split kernel diff with requested changes. (https://developer.blender.org/D2348)
Tried to investigate some OpenCL issues in the tracker
Discovered that issues seen by users seem to depend on which series of card they have, which will make diagnosing and fixing issues much harder. (I was unable to reproduce any of the issues I looked at, which is a bit worrisome)
Started work on new memory management / work scheduling. Unfortunately driver issues got in the way.
Ended up spending a whole day trying to reproduce the issue reliably and gathering logs. Also tried to work around the issue by adding more rigorous error handling to Cycles in hope that it was something Cycles was doing wrong, but this turned out to be insufficient.

Got tile updates working, so users get feedback while a tile is rendering rather than only seeing completed tiles.
Spent two days tracking down what I thought was a mistake in my code. Turned out to be an issue with drivers again.
Started to port the split kernel to cuda, so we have another device type to test against and ensure we don't have issues doing so later. This is progressing very quickly due to prior work, will be finished early next week.

Next week: finish porting to cuda, bring back automatic tile splitting.

Ported split kernel to CUDA
Most of the week was spent on tile splitting. While I've got it working I've decided not to push anything yet, as I'm not happy with the result. Getting things to work nicely on different devices and in low memory environments is quite difficult, and will require more careful design. Will do more work on this next week.

Did some code clean up
Tile splitting wasn't working out, so came up with a few alternative approaches that would give better performance and versatility
Made some changes to how render times are reported in debug info, so that kernel build times are reported separately, and fixed a few issues in the benchmark script (still need to submit for review).
Ran benchmarks for both release and split kernel branch