https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80859
--- Comment #25 from Jakub Jelinek <jakub at gcc dot gnu.org> --- In the GCC implementation of offloading to PTX, all HW threads in a warp (i.e. 32 of them) are a single OpenMP thread, and one needs to use a simd region (effectively SIMT) to get useful work done by all all the threads of a warp rather than just one. Right now GCC doesn't do auto-SIMTization (but does auto-vectorization on the host or XeonPhi accelerator etc., but only with -O3 or -O2 -ftree-vectorize; while with simd constructs you get it even with just -O2 -fopenmp for those regions), so simd construct is important to get the right performance. Threads within a team are the warp groups of threads within a PTX CTA, and different teams are the CTAs in a CTA grid (in PTX terms).