> On 20 Aug 2017, at 00:18, Vasco Alexandre da Silva Costa > <vasco.co...@gmail.com> wrote: > > This patch took me a bit longer to read and evaluate because I haven't been > so well lately (sick with the flu) but here it is: > > Congratulations to Marco Domingues for competing this major GSoC milestone!
Thank you! I’m really happy to see my work finally on the trunk :) but there is still some work to do to make the code more efficient! > > This code has enough performance that it can be used as an alternative to the > existing ANSI C code. Right now it has roughly the same order of performance > on scenes with small depth complexity. However the performance improvement we > expected with OpenCL hasn't materialized yet, the current iteration of the > OpenCL backend still lacks important optimizations which are done in the ANSI > C backend, like stopping ray traversal once the boolean evaluation detects a > solid hit. Currently we are processing all hits. This is obviously more > compute intensive. > > So we have talked about how to improve the performance by refactoring this > code and came up with some ideas. Marco will be working on performance > improvements over the remaining time we have left for the GSoC. Yes, I have been working really hard on this optimization, trying to get this to work by the end of the GSoC period (one week left). To give some context, this optimization is about merging the three OpenCl kernels (‘store_segs’, ‘rt_boolweave’ and ‘rt_boolfinal’) into a single kernel, to allow the weave of segments and evaluation of partitions to be done as soon as the first segments are calculated, which would avoid the unnecessary total ray traversal. After some intensive debugging, I found the bug that was causing the ray tracing to crash with this new rendering loop, and although there are still some bugs to be fixed, the results look very promising. The havoc scene now renders in 0.36 sec vs 0.78 sec previously. This when using the command ‘rt -z1 -s1024’ for the view: az 35 el 25. For reference, the ANSI C code renders this same scene in 0.63 sec, command ‘rt -s1024’, same view. As I mention before, there are still some bugs with this new system, i.e some pixels appear to be rendering the wrong object. My guess is that some partitions are being evaluated too early but I still need to investigate this issue with more detail. I will keep working on this and hopefully I can get this bug fixed in the next few days! Cheers! Marco > > What this patch does allow right now is to render CSG boolean operations, > like object intersections or subtractions, over either CPU or GPU platforms, > or even more exotic hardware like FPGAs which support OpenCL, which is > something that we couldn't do previously. The prior code only ran over > single-threaded CPUs with ANSI C. Another thing this work has allowed was to > clean up and understand better the behaviour of the current ANSI C bool code > which probably hadn't been touched in a really long time. > > I still hope we'll eventually get like a 3-4x performance improvement with > this vector OpenCL librt backend over the scalar ANSI C version running over > a modern CPU. > The performance could be even better than that if we further reduce thread > divergence and modify the data structure but it seems this will take more > effort. > > So we have achieved our initial goal with OpenCL boolean evaluation, but > there is still some more work to be done in this for sure, and hopefully also > in future GSoCs, > > Regards, > > -- > Vasco Alexandre da Silva Costa > PhD in Computer Engineering (Computer Graphics) > Instituto Superior Técnico/University of Lisbon, Portugal > ------------------------------------------------------------------------------ > Check out the vibrant tech community on one of the world's most > engaging tech sites, Slashdot.org! > http://sdm.link/slashdot_______________________________________________ > BRL-CAD Developer mailing list > brlcad-devel@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/brlcad-devel ------------------------------------------------------------------------------ Check out the vibrant tech community on one of the world's most engaging tech sites, Slashdot.org! http://sdm.link/slashdot _______________________________________________ BRL-CAD Developer mailing list brlcad-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/brlcad-devel
