1 2 static char help[] = "Tests MATSEQDENSECUDA\n\n"; 3 4 #include <petscmat.h> 5 6 int main(int argc,char **argv) 7 { 8 Mat A,AC,B; 9 PetscErrorCode ierr; 10 PetscInt m = 10,n = 10; 11 PetscReal r,tol = 10*PETSC_SMALL; 12 13 ierr = PetscInitialize(&argc,&argv,(char*) 0,help);if (ierr) return ierr; 14 ierr = PetscOptionsGetInt(NULL,NULL,"-m",&m,NULL);CHKERRQ(ierr); 15 ierr = PetscOptionsGetInt(NULL,NULL,"-n",&n,NULL);CHKERRQ(ierr); 16 ierr = MatCreate(PETSC_COMM_SELF,&A);CHKERRQ(ierr); 17 ierr = MatSetSizes(A,PETSC_DECIDE,PETSC_DECIDE,m,n);CHKERRQ(ierr); 18 ierr = MatSetType(A,MATSEQDENSE);CHKERRQ(ierr); 19 ierr = MatSetFromOptions(A);CHKERRQ(ierr); 20 ierr = MatSeqDenseSetPreallocation(A,NULL);CHKERRQ(ierr); 21 ierr = MatSetRandom(A,NULL);CHKERRQ(ierr); 22 23 /* Create a CUDA version of A */ 24 ierr = MatConvert(A,MATSEQDENSECUDA,MAT_INITIAL_MATRIX,&AC);CHKERRQ(ierr); 25 ierr = MatDuplicate(AC,MAT_COPY_VALUES,&B);CHKERRQ(ierr); 26 27 /* full CUDA AXPY */ 28 ierr = MatAXPY(B,-1.0,AC,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 29 ierr = MatNorm(B,NORM_INFINITY,&r);CHKERRQ(ierr); 30 if (r != 0.0) SETERRQ1(PetscObjectComm((PetscObject)B),PETSC_ERR_PLIB,"Error MatDuplicate + MatCopy + MatAXPY %g",(double)r); 31 32 /* test Copy */ 33 ierr = MatCopy(AC,B,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 34 35 /* call MatAXPY_Basic since B is CUDA, A is CPU, */ 36 ierr = MatAXPY(B,-1.0,A,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 37 ierr = MatNorm(B,NORM_INFINITY,&r);CHKERRQ(ierr); 38 if (r != 0.0) SETERRQ1(PetscObjectComm((PetscObject)B),PETSC_ERR_PLIB,"Error MatDuplicate + MatCopy + MatAXPY_Basic %g",(double)r); 39 40 if (m == n) { 41 Mat B1,B2; 42 43 ierr = MatCopy(AC,B,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 44 /* full CUDA PtAP */ 45 ierr = MatPtAP(B,AC,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B1);CHKERRQ(ierr); 46 /* CPU PtAP since A is on the CPU only */ 47 ierr = MatPtAP(B,A,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B2);CHKERRQ(ierr); 48 ierr = MatAXPY(B2,-1.0,B1,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 49 ierr = MatNorm(B2,NORM_INFINITY,&r);CHKERRQ(ierr); 50 if (r > tol) SETERRQ1(PetscObjectComm((PetscObject)B),PETSC_ERR_PLIB,"Error MatPtAP %g",(double)r); 51 52 /* test reuse */ 53 ierr = MatPtAP(B,AC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B1);CHKERRQ(ierr); 54 ierr = MatPtAP(B,A,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B2);CHKERRQ(ierr); 55 ierr = MatAXPY(B2,-1.0,B1,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 56 ierr = MatNorm(B2,NORM_INFINITY,&r);CHKERRQ(ierr); 57 if (r > tol) SETERRQ1(PetscObjectComm((PetscObject)B),PETSC_ERR_PLIB,"Error MatPtAP %g",(double)r); 58 59 ierr = MatDestroy(&B1);CHKERRQ(ierr); 60 ierr = MatDestroy(&B2);CHKERRQ(ierr); 61 } 62 63 ierr = MatDestroy(&B);CHKERRQ(ierr); 64 ierr = MatDestroy(&AC);CHKERRQ(ierr); 65 ierr = MatDestroy(&A);CHKERRQ(ierr); 66 ierr = PetscFinalize(); 67 return ierr; 68 } 69 70 /*TEST 71 72 build: 73 requires: cuda 74 75 test: 76 output_file: output/ex32_1.out 77 args: -m {{3 5 12}} -n {{3 5 12}} 78 suffix: seqdensecuda 79 80 TEST*/ 81