Unlimited branches-per-level for fine GPU solver

Open bcumming opened this issue 7 years ago • 1 comments

The fine matrix solver on the GPU assigns one thread to each branch on a level. Matrices are fused together so that the number of threads per level does not exceed the thread block dimension.

Currently, a single matrix with more branches on a single level than the number of threads in a thread block will cause an exception to be thrown. Refactor to assign multiple branches per level in this case.

Nov 13 '18 10:11 bcumming