1 static char help[] = "Benchmarking cudaPointerGetAttributes() time\n"; 2 /* 3 Running example on Summit at OLCF: 4 # 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 5 $ jsrun -n1 -a1 -c7 -g1 -r1 ./ex2cu 6 Average cudaPointerGetAttributes() time = 0.31 microseconds 7 */ 8 #include <petscsys.h> 9 #include <petscdevice.h> 10 11 int main(int argc,char **argv) 12 { 13 PetscInt i,n=4000; 14 cudaError_t cerr; 15 PetscScalar **ptrs; 16 PetscLogDouble tstart,tend,time; 17 struct cudaPointerAttributes attr; 18 19 PetscCall(PetscInitialize(&argc,&argv,(char*)0,help)); 20 PetscCall(PetscOptionsGetInt(NULL,NULL,"-n",&n,NULL)); 21 PetscCallCUDA(cudaStreamSynchronize(NULL)); /* Initialize CUDA runtime to get more accurate timing below */ 22 23 PetscCall(PetscMalloc1(n,&ptrs)); 24 for (i=0; i<n; i++) { 25 if (i%2) PetscCall(PetscMalloc1(i+16,&ptrs[i])); 26 else PetscCallCUDA(cudaMalloc((void**)&ptrs[i],(i+16)*sizeof(PetscScalar))); 27 } 28 29 PetscCall(PetscTime(&tstart)); 30 for (i=0; i<n; i++) { 31 cerr = cudaPointerGetAttributes(&attr,ptrs[i]); 32 if (cerr) cerr = cudaGetLastError(); 33 } 34 PetscCall(PetscTime(&tend)); 35 time = (tend-tstart)*1e6/n; 36 37 PetscCall(PetscPrintf(PETSC_COMM_WORLD,"Average cudaPointerGetAttributes() time = %.2f microseconds\n",time)); 38 39 for (i=0; i<n; i++) { 40 if (i%2) PetscCall(PetscFree(ptrs[i])); 41 else PetscCallCUDA(cudaFree(ptrs[i])); 42 } 43 PetscCall(PetscFree(ptrs)); 44 45 PetscCall(PetscFinalize()); 46 return 0; 47 } 48 49 /*TEST 50 build: 51 requires: cuda 52 53 test: 54 requires: cuda 55 args: -n 2 56 output_file: output/empty.out 57 filter: grep "DOES_NOT_EXIST" 58 59 TEST*/ 60