https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104916
--- Comment #3 from Tom de Vries <vries at gcc dot gnu.org> --- Anyway, having reread the volta architecture whitepaper again, I think it's ok to use the solution I already found that does work (see PR104783): add a warp sync at simt exit. The tricky bit is that we rely on the warps to stay uniform after that warp sync until the next simt region entry. Which AFAICT is not something enforced by specification, but rather by a assuming that the 'Convergence optimizer' has a sane implementation.
