利用者:DingTo/GSoC 2014/Weekly Reports/Week7
< 利用者:DingTo | GSoC 2014 | Weekly Reports
Week 7
What I did this week
I did some cleanup and fixes for CUDA:
- Code cleanup: (6ac06523955e, 4800c52700b4)
- Updated CUDA error messages, so we get more meaningful errors instead of "Unknown CUDA error value". (5898abe99d5d)
- Volumes can be compiled on GPU again, yet this is still quite fragile. (5aec61f8493f)
- Distance Sampling works, Equi-Angular and MIS fail with memory/address CUDA errors, issue probably in kernel_volume_decoupled_scatter().
- Sample All Lights Indirect causes errors too. Not sure why Direct works though, it's the same code in kernel_branched_path_volume_connect_light().
- In general I think Decoupled Ray Marching needs some fixes for GPU, or we completely #ifdef that and only use distance sampling on the GPU. The latter is a viable option for the 2.72 release, *if* the performance drop is not too big. Have not benchmarked it yet.
Other work:
- Looked a bit into the Smooth normal optimization of my original proposal, but no real code there yet.
- The usual bug tracker maintenance (check on reports, close invalid/fixes ones...)
Next week
I have exams starting on Monday until the 17th July, so don't have time for GSoC. As said in Report #5, I will be back full time afterwards, which leave me full 4 weeks.
Questions
None.