Actual source code: ex32.c


  2: static char help[] = "Tests MATSEQDENSECUDA\n\n";

  4: #include <petscmat.h>

  6: int main(int argc,char **argv)
  7: {
  8:   Mat       A,AC,B;
  9:   PetscInt  m     = 10,n = 10;
 10:   PetscReal r,tol = 10*PETSC_SMALL;

 12:   PetscInitialize(&argc,&argv,(char*) 0,help);
 13:   PetscOptionsGetInt(NULL,NULL,"-m",&m,NULL);
 14:   PetscOptionsGetInt(NULL,NULL,"-n",&n,NULL);
 15:   MatCreate(PETSC_COMM_SELF,&A);
 16:   MatSetSizes(A,PETSC_DECIDE,PETSC_DECIDE,m,n);
 17:   MatSetType(A,MATSEQDENSE);
 18:   MatSetFromOptions(A);
 19:   MatSeqDenseSetPreallocation(A,NULL);
 20:   MatSetRandom(A,NULL);
 21: #if 0
 22:   PetscInt       i,j;
 23:   PetscScalar    val;
 24:   for (i=0; i<m; i++) {
 25:     for (j=0; j<n; j++) {
 26:       val = (PetscScalar)(i+j);
 27:       MatSetValues(A,1,&i,1,&j,&val,INSERT_VALUES);
 28:     }
 29:   }
 30:   MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
 31:   MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
 32: #endif

 34:   /* Create a CUDA version of A */
 35: #if defined(PETSC_HAVE_CUDA)
 36:   MatConvert(A,MATSEQDENSECUDA,MAT_INITIAL_MATRIX,&AC);
 37: #else
 38:   MatDuplicate(A,MAT_COPY_VALUES,&AC);
 39: #endif
 40:   MatDuplicate(AC,MAT_COPY_VALUES,&B);

 42:   /* full CUDA AXPY */
 43:   MatAXPY(B,-1.0,AC,SAME_NONZERO_PATTERN);
 44:   MatNorm(B,NORM_INFINITY,&r);

 47:   /* test Copy */
 48:   MatCopy(AC,B,SAME_NONZERO_PATTERN);

 50:   /* call MatAXPY_Basic since B is CUDA, A is CPU,  */
 51:   MatAXPY(B,-1.0,A,SAME_NONZERO_PATTERN);
 52:   MatNorm(B,NORM_INFINITY,&r);

 55:   if (m == n) {
 56:     Mat B1,B2;

 58:     MatCopy(AC,B,SAME_NONZERO_PATTERN);
 59:     /* full CUDA PtAP */
 60:     MatPtAP(B,AC,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B1);

 62:     /* CPU PtAP since A is on the CPU only */
 63:     MatPtAP(B,A,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B2);

 65:     MatAXPY(B2,-1.0,B1,SAME_NONZERO_PATTERN);
 66:     MatNorm(B2,NORM_INFINITY,&r);

 69:     /* test reuse */
 70:     MatPtAP(B,AC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B1);
 71:     MatPtAP(B,A,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B2);
 72:     MatAXPY(B2,-1.0,B1,SAME_NONZERO_PATTERN);
 73:     MatNorm(B2,NORM_INFINITY,&r);

 76:     MatDestroy(&B1);
 77:     MatDestroy(&B2);
 78:   }

 80:   MatDestroy(&B);
 81:   MatDestroy(&AC);
 82:   MatDestroy(&A);
 83:   PetscFinalize();
 84:   return 0;
 85: }

 87: /*TEST

 89:    build:
 90:      requires: cuda

 92:    test:
 93:      output_file: output/ex32_1.out
 94:      args: -m {{3 5 12}} -n {{3 5 12}}
 95:      suffix: seqdensecuda

 97: TEST*/