Actual source code: ex32.c

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

  3: #include <petscmat.h>

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

 11:   PetscFunctionBeginUser;
 12:   PetscCall(PetscInitialize(&argc, &argv, NULL, help));
 13:   PetscCall(PetscOptionsGetInt(NULL, NULL, "-m", &m, NULL));
 14:   PetscCall(PetscOptionsGetInt(NULL, NULL, "-n", &n, NULL));
 15:   PetscCall(MatCreate(PETSC_COMM_SELF, &A));
 16:   PetscCall(MatSetSizes(A, PETSC_DECIDE, PETSC_DECIDE, m, n));
 17:   PetscCall(MatSetType(A, MATSEQDENSE));
 18:   PetscCall(MatSetFromOptions(A));
 19:   PetscCall(MatSeqDenseSetPreallocation(A, NULL));
 20:   PetscCall(MatSetRandom(A, NULL));

 22:   /* Create a CUDA version of A */
 23: #if PetscDefined(HAVE_CUDA)
 24:   PetscCall(MatConvert(A, MATSEQDENSECUDA, MAT_INITIAL_MATRIX, &AC));
 25: #else
 26:   PetscCall(MatDuplicate(A, MAT_COPY_VALUES, &AC));
 27: #endif
 28:   PetscCall(MatDuplicate(AC, MAT_COPY_VALUES, &B));

 30:   /* full CUDA AXPY */
 31:   PetscCall(MatAXPY(B, -1.0, AC, SAME_NONZERO_PATTERN));
 32:   PetscCall(MatNorm(B, NORM_INFINITY, &r));
 33:   PetscCheck(r == 0.0, PetscObjectComm((PetscObject)B), PETSC_ERR_PLIB, "Error MatDuplicate + MatCopy + MatAXPY %g", (double)r);

 35:   /* test Copy */
 36:   PetscCall(MatCopy(AC, B, SAME_NONZERO_PATTERN));

 38:   /* call MatAXPY_Basic since B is CUDA, A is CPU,  */
 39:   PetscCall(MatAXPY(B, -1.0, A, SAME_NONZERO_PATTERN));
 40:   PetscCall(MatNorm(B, NORM_INFINITY, &r));
 41:   PetscCheck(r == 0.0, PetscObjectComm((PetscObject)B), PETSC_ERR_PLIB, "Error MatDuplicate + MatCopy + MatAXPY_Basic %g", (double)r);

 43:   if (m == n) {
 44:     Mat B1, B2;

 46:     PetscCall(MatCopy(AC, B, SAME_NONZERO_PATTERN));
 47:     /* full CUDA PtAP */
 48:     PetscCall(MatPtAP(B, AC, MAT_INITIAL_MATRIX, PETSC_DETERMINE, &B1));

 50:     /* CPU PtAP since A is on the CPU only */
 51:     PetscCall(MatPtAP(B, A, MAT_INITIAL_MATRIX, PETSC_DETERMINE, &B2));

 53:     PetscCall(MatAXPY(B2, -1.0, B1, SAME_NONZERO_PATTERN));
 54:     PetscCall(MatNorm(B2, NORM_INFINITY, &r));
 55:     PetscCheck(r <= tol, PetscObjectComm((PetscObject)B), PETSC_ERR_PLIB, "Error MatPtAP %g", (double)r);

 57:     /* test reuse */
 58:     PetscCall(MatPtAP(B, AC, MAT_REUSE_MATRIX, PETSC_DETERMINE, &B1));
 59:     PetscCall(MatPtAP(B, A, MAT_REUSE_MATRIX, PETSC_DETERMINE, &B2));
 60:     PetscCall(MatAXPY(B2, -1.0, B1, SAME_NONZERO_PATTERN));
 61:     PetscCall(MatNorm(B2, NORM_INFINITY, &r));
 62:     PetscCheck(r <= tol, PetscObjectComm((PetscObject)B), PETSC_ERR_PLIB, "Error MatPtAP %g", (double)r);

 64:     PetscCall(MatDestroy(&B1));
 65:     PetscCall(MatDestroy(&B2));
 66:   }

 68:   PetscCall(MatDestroy(&B));
 69:   PetscCall(MatDestroy(&AC));
 70:   PetscCall(MatDestroy(&A));
 71:   PetscCall(PetscFinalize());
 72:   return 0;
 73: }

 75: /*TEST

 77:    build:
 78:      requires: cuda

 80:    test:
 81:      output_file: output/empty.out
 82:      args: -m {{3 5 12}} -n {{3 5 12}}
 83:      suffix: seqdensecuda

 85: TEST*/