1 static char help[] = "Scalable algorithm for Connected Components problem.\n\ 2 Entails changing the MatMult() for this matrix.\n\n\n"; 3 4 #include <petscmat.h> 5 6 PETSC_EXTERN PetscErrorCode MatMultMax_SeqAIJ(Mat, Vec, Vec); 7 PETSC_EXTERN PetscErrorCode MatMultAddMax_SeqAIJ(Mat, Vec, Vec, Vec); 8 #include <../src/mat/impls/aij/mpi/mpiaij.h> 9 10 /* 11 Paper with Ananth: Frbenius norm of band was good proxy, but really want to know the rank outside 12 13 LU for diagonal blocks must do shifting instead of pivoting, preferably shifting individual rows (like Pardiso) 14 15 Draw picture of flow of reordering 16 17 Measure Forbenius norm of the blocks being dropped by Truncated SPIKE (might be contaminated by pivoting in LU) 18 19 Report on using Florida matrices (Maxim, Murat) 20 */ 21 22 /* 23 I have thought about how to do this. Here is a prototype algorithm. Let A be 24 the adjacency matrix (0 or 1), and let each component be identified by the 25 lowest numbered vertex in it. We initialize a vector c so that each vertex is 26 a component, c_i = i. Now we act on c with A, using a special product 27 28 c = A * c 29 30 where we replace addition with min. The fixed point of this operation is a vector 31 c which is the component for each vertex. The number of iterates is 32 33 max_{components} depth of BFS tree for component 34 35 We can accelerate this algorithm by preprocessing all locals domains using the 36 same algorithm. Then the number of iterations is bounded the depth of the BFS 37 tree for the graph on supervertices defined over local components, which is 38 bounded by p. In practice, this should be very fast. 39 */ 40 41 /* Only isolated vertices get a 1 on the diagonal */ 42 PetscErrorCode CreateGraph(MPI_Comm comm, PetscInt testnum, Mat *A) { 43 Mat G; 44 45 PetscFunctionBegin; 46 PetscCall(MatCreate(comm, &G)); 47 /* The identity matrix */ 48 switch (testnum) { 49 case 0: { 50 Vec D; 51 52 PetscCall(MatSetSizes(G, PETSC_DETERMINE, PETSC_DETERMINE, 5, 5)); 53 PetscCall(MatSetUp(G)); 54 PetscCall(MatCreateVecs(G, &D, NULL)); 55 PetscCall(VecSet(D, 1.0)); 56 PetscCall(MatDiagonalSet(G, D, INSERT_VALUES)); 57 PetscCall(VecDestroy(&D)); 58 } break; 59 case 1: { 60 PetscScalar vals[3] = {1.0, 1.0, 1.0}; 61 PetscInt cols[3]; 62 PetscInt rStart, rEnd, row; 63 64 PetscCall(MatSetSizes(G, PETSC_DETERMINE, PETSC_DETERMINE, 5, 5)); 65 PetscCall(MatSetFromOptions(G)); 66 PetscCall(MatSeqAIJSetPreallocation(G, 2, NULL)); 67 PetscCall(MatSetUp(G)); 68 PetscCall(MatGetOwnershipRange(G, &rStart, &rEnd)); 69 row = 0; 70 cols[0] = 0; 71 cols[1] = 1; 72 if ((row >= rStart) && (row < rEnd)) PetscCall(MatSetValues(G, 1, &row, 2, cols, vals, INSERT_VALUES)); 73 row = 1; 74 cols[0] = 0; 75 cols[1] = 1; 76 if ((row >= rStart) && (row < rEnd)) PetscCall(MatSetValues(G, 1, &row, 2, cols, vals, INSERT_VALUES)); 77 row = 2; 78 cols[0] = 2; 79 cols[1] = 3; 80 if ((row >= rStart) && (row < rEnd)) PetscCall(MatSetValues(G, 1, &row, 2, cols, vals, INSERT_VALUES)); 81 row = 3; 82 cols[0] = 3; 83 cols[1] = 4; 84 if ((row >= rStart) && (row < rEnd)) PetscCall(MatSetValues(G, 1, &row, 2, cols, vals, INSERT_VALUES)); 85 row = 4; 86 cols[0] = 4; 87 cols[1] = 2; 88 if ((row >= rStart) && (row < rEnd)) PetscCall(MatSetValues(G, 1, &row, 2, cols, vals, INSERT_VALUES)); 89 PetscCall(MatAssemblyBegin(G, MAT_FINAL_ASSEMBLY)); 90 PetscCall(MatAssemblyEnd(G, MAT_FINAL_ASSEMBLY)); 91 } break; 92 case 2: { 93 PetscScalar vals[3] = {1.0, 1.0, 1.0}; 94 PetscInt cols[3]; 95 PetscInt rStart, rEnd, row; 96 97 PetscCall(MatSetSizes(G, PETSC_DETERMINE, PETSC_DETERMINE, 5, 5)); 98 PetscCall(MatSetFromOptions(G)); 99 PetscCall(MatSeqAIJSetPreallocation(G, 2, NULL)); 100 PetscCall(MatSetUp(G)); 101 PetscCall(MatGetOwnershipRange(G, &rStart, &rEnd)); 102 row = 0; 103 cols[0] = 0; 104 cols[1] = 4; 105 if ((row >= rStart) && (row < rEnd)) PetscCall(MatSetValues(G, 1, &row, 2, cols, vals, INSERT_VALUES)); 106 row = 1; 107 cols[0] = 1; 108 cols[1] = 2; 109 if ((row >= rStart) && (row < rEnd)) PetscCall(MatSetValues(G, 1, &row, 2, cols, vals, INSERT_VALUES)); 110 row = 2; 111 cols[0] = 2; 112 cols[1] = 3; 113 if ((row >= rStart) && (row < rEnd)) PetscCall(MatSetValues(G, 1, &row, 2, cols, vals, INSERT_VALUES)); 114 row = 3; 115 cols[0] = 3; 116 cols[1] = 1; 117 if ((row >= rStart) && (row < rEnd)) PetscCall(MatSetValues(G, 1, &row, 2, cols, vals, INSERT_VALUES)); 118 row = 4; 119 cols[0] = 0; 120 cols[1] = 4; 121 if ((row >= rStart) && (row < rEnd)) PetscCall(MatSetValues(G, 1, &row, 2, cols, vals, INSERT_VALUES)); 122 PetscCall(MatAssemblyBegin(G, MAT_FINAL_ASSEMBLY)); 123 PetscCall(MatAssemblyEnd(G, MAT_FINAL_ASSEMBLY)); 124 } break; 125 default: SETERRQ(comm, PETSC_ERR_PLIB, "Unknown test %d", testnum); 126 } 127 *A = G; 128 PetscFunctionReturn(0); 129 } 130 131 int main(int argc, char **argv) { 132 MPI_Comm comm; 133 Mat A; /* A graph */ 134 Vec c; /* A vector giving the component of each vertex */ 135 Vec cold; /* The vector c from the last iteration */ 136 PetscScalar *carray; 137 PetscInt testnum = 0; 138 PetscInt V, vStart, vEnd, v, n; 139 PetscMPIInt size; 140 141 PetscFunctionBeginUser; 142 PetscCall(PetscInitialize(&argc, &argv, NULL, help)); 143 comm = PETSC_COMM_WORLD; 144 PetscCallMPI(MPI_Comm_size(comm, &size)); 145 /* Use matrix to encode a graph */ 146 PetscCall(PetscOptionsGetInt(NULL, NULL, "-testnum", &testnum, NULL)); 147 PetscCall(CreateGraph(comm, testnum, &A)); 148 PetscCall(MatGetSize(A, &V, NULL)); 149 /* Replace matrix-vector multiplication with one that calculates the minimum rather than the sum */ 150 if (size == 1) { 151 PetscCall(MatShellSetOperation(A, MATOP_MULT, (void(*))MatMultMax_SeqAIJ)); 152 } else { 153 Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data; 154 155 PetscCall(MatShellSetOperation(a->A, MATOP_MULT, (void(*))MatMultMax_SeqAIJ)); 156 PetscCall(MatShellSetOperation(a->B, MATOP_MULT, (void(*))MatMultMax_SeqAIJ)); 157 PetscCall(MatShellSetOperation(a->B, MATOP_MULT_ADD, (void(*))MatMultAddMax_SeqAIJ)); 158 } 159 /* Initialize each vertex as a separate component */ 160 PetscCall(MatCreateVecs(A, &c, NULL)); 161 PetscCall(MatGetOwnershipRange(A, &vStart, &vEnd)); 162 PetscCall(VecGetArray(c, &carray)); 163 for (v = vStart; v < vEnd; ++v) { carray[v - vStart] = v; } 164 PetscCall(VecRestoreArray(c, &carray)); 165 /* Preprocess in parallel to find local components */ 166 /* Multiply until c does not change */ 167 PetscCall(VecDuplicate(c, &cold)); 168 for (v = 0; v < V; ++v) { 169 Vec cnew = cold; 170 PetscBool stop; 171 172 PetscCall(MatMult(A, c, cnew)); 173 PetscCall(VecEqual(c, cnew, &stop)); 174 if (stop) break; 175 cold = c; 176 c = cnew; 177 } 178 /* Report */ 179 PetscCall(VecUniqueEntries(c, &n, NULL)); 180 PetscCall(PetscPrintf(comm, "Components: %d Iterations: %d\n", n, v)); 181 PetscCall(VecView(c, PETSC_VIEWER_STDOUT_WORLD)); 182 /* Cleanup */ 183 PetscCall(VecDestroy(&c)); 184 PetscCall(VecDestroy(&cold)); 185 PetscCall(PetscFinalize()); 186 return 0; 187 } 188