xref: /petsc/src/mat/tests/ex32.c (revision e19f88df7cd0bcfe73faf98683db6f77794e28aa)
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 #if 0
23   PetscInt       i,j;
24   PetscScalar    val;
25   for (i=0; i<m; i++) {
26     for (j=0; j<n; j++) {
27       val = (PetscScalar)(i+j);
28       ierr = MatSetValues(A,1,&i,1,&j,&val,INSERT_VALUES);CHKERRQ(ierr);
29     }
30   }
31   ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
32   ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
33 #endif
34 
35   /* Create a CUDA version of A */
36 #if defined(PETSC_HAVE_CUDA)
37   ierr = MatConvert(A,MATSEQDENSECUDA,MAT_INITIAL_MATRIX,&AC);CHKERRQ(ierr);
38 #else
39   ierr = MatDuplicate(A,MAT_COPY_VALUES,&AC);CHKERRQ(ierr);
40 #endif
41   ierr = MatDuplicate(AC,MAT_COPY_VALUES,&B);CHKERRQ(ierr);
42 
43   /* full CUDA AXPY */
44   ierr = MatAXPY(B,-1.0,AC,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
45   ierr = MatNorm(B,NORM_INFINITY,&r);CHKERRQ(ierr);
46   if (r != 0.0) SETERRQ1(PetscObjectComm((PetscObject)B),PETSC_ERR_PLIB,"Error MatDuplicate + MatCopy + MatAXPY %g",(double)r);
47 
48   /* test Copy */
49   ierr = MatCopy(AC,B,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
50 
51   /* call MatAXPY_Basic since B is CUDA, A is CPU,  */
52   ierr = MatAXPY(B,-1.0,A,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
53   ierr = MatNorm(B,NORM_INFINITY,&r);CHKERRQ(ierr);
54   if (r != 0.0) SETERRQ1(PetscObjectComm((PetscObject)B),PETSC_ERR_PLIB,"Error MatDuplicate + MatCopy + MatAXPY_Basic %g",(double)r);
55 
56   if (m == n) {
57     Mat B1,B2;
58 
59     ierr = MatCopy(AC,B,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
60     /* full CUDA PtAP */
61     ierr = MatPtAP(B,AC,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B1);CHKERRQ(ierr);
62 
63     /* CPU PtAP since A is on the CPU only */
64     ierr = MatPtAP(B,A,MAT_INITIAL_MATRIX,PETSC_DEFAULT,&B2);CHKERRQ(ierr);
65 
66     ierr = MatAXPY(B2,-1.0,B1,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
67     ierr = MatNorm(B2,NORM_INFINITY,&r);CHKERRQ(ierr);
68     if (r > tol) SETERRQ1(PetscObjectComm((PetscObject)B),PETSC_ERR_PLIB,"Error MatPtAP %g",(double)r);
69 
70     /* test reuse */
71     ierr = MatPtAP(B,AC,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B1);CHKERRQ(ierr);
72     ierr = MatPtAP(B,A,MAT_REUSE_MATRIX,PETSC_DEFAULT,&B2);CHKERRQ(ierr);
73     ierr = MatAXPY(B2,-1.0,B1,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
74     ierr = MatNorm(B2,NORM_INFINITY,&r);CHKERRQ(ierr);
75     if (r > tol) SETERRQ1(PetscObjectComm((PetscObject)B),PETSC_ERR_PLIB,"Error MatPtAP %g",(double)r);
76 
77     ierr = MatDestroy(&B1);CHKERRQ(ierr);
78     ierr = MatDestroy(&B2);CHKERRQ(ierr);
79   }
80 
81   ierr = MatDestroy(&B);CHKERRQ(ierr);
82   ierr = MatDestroy(&AC);CHKERRQ(ierr);
83   ierr = MatDestroy(&A);CHKERRQ(ierr);
84   ierr = PetscFinalize();
85   return ierr;
86 }
87 
88 /*TEST
89 
90    build:
91      requires: cuda
92 
93    test:
94      output_file: output/ex32_1.out
95      args: -m {{3 5 12}} -n {{3 5 12}}
96      suffix: seqdensecuda
97 
98 TEST*/
99