1*c4762a1bSJed Brown static char help[] = "Test CPU/GPU memory leaks, MatMult and MatMultTransposeAdd during successive matrix assemblies\n\n"; 2*c4762a1bSJed Brown 3*c4762a1bSJed Brown #include <petscmat.h> 4*c4762a1bSJed Brown 5*c4762a1bSJed Brown int main(int argc,char **argv) 6*c4762a1bSJed Brown { 7*c4762a1bSJed Brown PetscErrorCode ierr; 8*c4762a1bSJed Brown PetscMPIInt rank,size; 9*c4762a1bSJed Brown Mat A; 10*c4762a1bSJed Brown PetscInt i,j,k,n=3,vstart,rstart,rend,margin; 11*c4762a1bSJed Brown Vec x,y; 12*c4762a1bSJed Brown 13*c4762a1bSJed Brown ierr = PetscInitialize(&argc,&argv,(char*)0,help);if (ierr) return ierr; 14*c4762a1bSJed Brown ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr); 15*c4762a1bSJed Brown ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr); 16*c4762a1bSJed Brown 17*c4762a1bSJed Brown ierr = MatCreate(PETSC_COMM_WORLD,&A);CHKERRQ(ierr); 18*c4762a1bSJed Brown ierr = MatSetSizes(A,n,n,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr); 19*c4762a1bSJed Brown ierr = MatSetFromOptions(A);CHKERRQ(ierr); 20*c4762a1bSJed Brown 21*c4762a1bSJed Brown ierr = MatMPIAIJSetPreallocation(A,n,NULL,0,NULL);CHKERRQ(ierr); 22*c4762a1bSJed Brown ierr = MatSetOption(A,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 23*c4762a1bSJed Brown ierr = MatGetOwnershipRange(A,&rstart,&rend);CHKERRQ(ierr); 24*c4762a1bSJed Brown ierr = MatCreateVecs(A,&x,&y);CHKERRQ(ierr); 25*c4762a1bSJed Brown ierr = VecSet(x,1.0);CHKERRQ(ierr); 26*c4762a1bSJed Brown 27*c4762a1bSJed Brown /* 28*c4762a1bSJed Brown Matirx A only has nonzeros in the diagonal block, which is of size 3x3. 29*c4762a1bSJed Brown We do three successive assemblies on A. The first two have the same non-zero 30*c4762a1bSJed Brown pattern but different values, and the third breaks the non-zero pattern. The 31*c4762a1bSJed Brown first two assemblies have enough zero-rows that triggers compressed-row storage 32*c4762a1bSJed Brown in MATAIJ and MATAIJCUSPARSE. 33*c4762a1bSJed Brown 34*c4762a1bSJed Brown These settings are used to test memory management and correctness in MatMult 35*c4762a1bSJed Brown and MatMultTransposeAdd. 36*c4762a1bSJed Brown */ 37*c4762a1bSJed Brown 38*c4762a1bSJed Brown for (k=0; k<3; k++) { /* Three assemblies */ 39*c4762a1bSJed Brown vstart = (size*k + rank)*n*n+1; 40*c4762a1bSJed Brown margin = (k == 2)? 0 : 2; /* Create two zero-rows in the first two assemblies */ 41*c4762a1bSJed Brown for (i=rstart; i<rend-margin; i++) { 42*c4762a1bSJed Brown for (j=rstart; j<rend; j++) { 43*c4762a1bSJed Brown ierr = MatSetValue(A,i,j,(PetscScalar)vstart,INSERT_VALUES);CHKERRQ(ierr); 44*c4762a1bSJed Brown vstart++; 45*c4762a1bSJed Brown } 46*c4762a1bSJed Brown } 47*c4762a1bSJed Brown ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 48*c4762a1bSJed Brown ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 49*c4762a1bSJed Brown ierr = MatMult(A,x,y);CHKERRQ(ierr); 50*c4762a1bSJed Brown ierr = MatMultTransposeAdd(A,x,y,y);CHKERRQ(ierr); /* y[i] = sum of row i and column i of A */ 51*c4762a1bSJed Brown ierr = VecView(y,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr); 52*c4762a1bSJed Brown } 53*c4762a1bSJed Brown 54*c4762a1bSJed Brown ierr = MatDestroy(&A);CHKERRQ(ierr); 55*c4762a1bSJed Brown ierr = VecDestroy(&x);CHKERRQ(ierr); 56*c4762a1bSJed Brown ierr = VecDestroy(&y);CHKERRQ(ierr); 57*c4762a1bSJed Brown ierr = PetscFinalize(); 58*c4762a1bSJed Brown 59*c4762a1bSJed Brown /* Uncomment this line if you want to use "cuda-memcheck --leaf-check full" to check this program */ 60*c4762a1bSJed Brown /*cudaDeviceReset();*/ 61*c4762a1bSJed Brown return ierr; 62*c4762a1bSJed Brown } 63*c4762a1bSJed Brown 64*c4762a1bSJed Brown /*TEST 65*c4762a1bSJed Brown 66*c4762a1bSJed Brown testset: 67*c4762a1bSJed Brown nsize: 2 68*c4762a1bSJed Brown output_file: output/ex236_1.out 69*c4762a1bSJed Brown filter: grep -v type 70*c4762a1bSJed Brown 71*c4762a1bSJed Brown test: 72*c4762a1bSJed Brown args: -mat_type aij 73*c4762a1bSJed Brown 74*c4762a1bSJed Brown test: 75*c4762a1bSJed Brown requires: cuda 76*c4762a1bSJed Brown suffix: cuda 77*c4762a1bSJed Brown args: -mat_type aijcusparse 78*c4762a1bSJed Brown TEST*/ 79