CUDA link errors on Pascal cards for Lagrangian and Streamline filters
This is related to the discussion in !2271 (merged) re: CUDA link errors. @kmorel, @ayenpure While CellLocators were being de-virtualized, there were CUDA link errors for Lagrangian and Streamline filters on Pascal cards. The issue appears to be that large amounts of code are generated and the Pascal cards can't handle this.
@cjy7117 had some suggestions: cuobjdump lib.so -res-usage | grep -A 1 $KERNEL' will show the register/shared/constant memory usage for a kernel.
@NAThompson said that he could take a look at this.