1896e5da2SSatish Balay static char help[] = "Benchmarking device kernel launch time\n"; 2896e5da2SSatish Balay /* 3896e5da2SSatish Balay Running example on Summit at OLCF: 4896e5da2SSatish Balay # run with total 1 resource set (RS) (-n1), 1 RS per node (-r1), 1 MPI rank (-a1), 7 cores (-c7) and 1 GPU (-g1) per RS 5896e5da2SSatish Balay $ jsrun -n1 -a1 -c7 -g1 -r1 ./ex1k 6896e5da2SSatish Balay Average asynchronous device kernel launch time = 4.86 microseconds 7896e5da2SSatish Balay Average synchronous device kernel launch time = 12.83 microseconds 8896e5da2SSatish Balay 9896e5da2SSatish Balay Frontier@OLCF 10896e5da2SSatish Balay $ srun -n1 -c32 --cpu-bind=threads --gpus-per-node=8 --gpu-bind=closest ./ex1k 11896e5da2SSatish Balay Average asynchronous device kernel launch time = 1.88 microseconds 12896e5da2SSatish Balay Average synchronous device kernel launch time = 7.78 microseconds 13896e5da2SSatish Balay 14896e5da2SSatish Balay Aurora@ALCF 15*a8cf87e0SJunchao Zhang $ mpiexec -n 1 ./ex1k 16896e5da2SSatish Balay Average asynchronous device kernel launch time = 3.34 microseconds 17896e5da2SSatish Balay Average synchronous device kernel launch time = 6.24 microseconds 18896e5da2SSatish Balay 19896e5da2SSatish Balay Perlmutter@NERSC 20896e5da2SSatish Balay $ srun -n 1 --gpus-per-task=1 ./ex1k 21896e5da2SSatish Balay Average asynchronous device kernel launch time = 2.31 microseconds 22896e5da2SSatish Balay Average synchronous device kernel launch time = 7.13 microseconds 23896e5da2SSatish Balay */ 24896e5da2SSatish Balay 25896e5da2SSatish Balay #include <petscsys.h> 26896e5da2SSatish Balay #include <petsc_kokkos.hpp> 27896e5da2SSatish Balay 28896e5da2SSatish Balay int main(int argc, char **argv) 29896e5da2SSatish Balay { 30896e5da2SSatish Balay PetscInt i, n = 100000, N = 256; 31896e5da2SSatish Balay PetscLogDouble tstart, tend, time; 32896e5da2SSatish Balay 33896e5da2SSatish Balay PetscFunctionBeginUser; 34896e5da2SSatish Balay PetscCall(PetscInitialize(&argc, &argv, nullptr, help)); 35896e5da2SSatish Balay PetscCall(PetscOptionsGetInt(NULL, NULL, "-n", &n, NULL)); 36896e5da2SSatish Balay PetscCall(PetscKokkosInitializeCheck()); 37896e5da2SSatish Balay { 38896e5da2SSatish Balay Kokkos::DefaultExecutionSpace exec = PetscGetKokkosExecutionSpace(); 39896e5da2SSatish Balay Kokkos::RangePolicy<Kokkos::DefaultExecutionSpace> policy(exec, 0, N); 40896e5da2SSatish Balay 41896e5da2SSatish Balay PetscCallCXX(exec.fence()); // Initialize device runtime to get more accurate timing below 42896e5da2SSatish Balay // Launch a sequence of kernels asynchronously. Previous launched kernels do not need to be completed before launching a new one 43896e5da2SSatish Balay PetscCall(PetscTime(&tstart)); 44896e5da2SSatish Balay for (i = 0; i < n; i++) PetscCallCXX(Kokkos::parallel_for(policy, KOKKOS_LAMBDA(const PetscInt &i){})); 45896e5da2SSatish Balay PetscCall(PetscTime(&tend)); 46896e5da2SSatish Balay PetscCallCXX(exec.fence()); 47896e5da2SSatish Balay time = (tend - tstart) * 1e6 / n; 48896e5da2SSatish Balay PetscCall(PetscPrintf(PETSC_COMM_WORLD, "Average asynchronous device kernel launch time = %.2f microseconds\n", time)); 49896e5da2SSatish Balay 50896e5da2SSatish Balay // Launch a sequence of kernels synchronously. Only launch a new kernel after the one before it has been completed 51896e5da2SSatish Balay PetscCall(PetscTime(&tstart)); 52896e5da2SSatish Balay for (i = 0; i < n; i++) { 53896e5da2SSatish Balay PetscCallCXX(Kokkos::parallel_for(policy, KOKKOS_LAMBDA(const PetscInt &i){})); 54896e5da2SSatish Balay PetscCallCXX(exec.fence()); 55896e5da2SSatish Balay } 56896e5da2SSatish Balay PetscCall(PetscTime(&tend)); 57896e5da2SSatish Balay time = (tend - tstart) * 1e6 / n; 58896e5da2SSatish Balay PetscCall(PetscPrintf(PETSC_COMM_WORLD, "Average synchronous device kernel launch time = %.2f microseconds\n", time)); 59896e5da2SSatish Balay } 60896e5da2SSatish Balay 61896e5da2SSatish Balay PetscCall(PetscFinalize()); 62896e5da2SSatish Balay return 0; 63896e5da2SSatish Balay } 64896e5da2SSatish Balay 65896e5da2SSatish Balay /*TEST 66896e5da2SSatish Balay test: 67896e5da2SSatish Balay requires: kokkos 68896e5da2SSatish Balay args: -n 2 69896e5da2SSatish Balay output_file: output/empty.out 70896e5da2SSatish Balay filter: grep "DOES_NOT_EXIST" 71896e5da2SSatish Balay 72896e5da2SSatish Balay TEST*/ 73