1c4762a1bSJed Brown 2c4762a1bSJed Brown static char help[] = "Tests MatMult(), MatMultAdd(), MatMultTranspose().\n\ 3c4762a1bSJed Brown Also MatMultTransposeAdd(), MatScale(), MatGetDiagonal(), and MatDiagonalScale().\n\n"; 4c4762a1bSJed Brown 5c4762a1bSJed Brown #include <petscmat.h> 6c4762a1bSJed Brown 79371c9d4SSatish Balay int main(int argc, char **args) { 8c4762a1bSJed Brown Mat C; 9c4762a1bSJed Brown Vec s, u, w, x, y, z; 10c4762a1bSJed Brown PetscInt i, j, m = 8, n, rstart, rend, vstart, vend; 11c4762a1bSJed Brown PetscScalar one = 1.0, negone = -1.0, v, alpha = 0.1; 12c4762a1bSJed Brown PetscReal norm, tol = PETSC_SQRT_MACHINE_EPSILON; 13c4762a1bSJed Brown PetscBool flg; 14c4762a1bSJed Brown 15327415f7SBarry Smith PetscFunctionBeginUser; 169566063dSJacob Faibussowitsch PetscCall(PetscInitialize(&argc, &args, (char *)0, help)); 179566063dSJacob Faibussowitsch PetscCall(PetscViewerPushFormat(PETSC_VIEWER_STDOUT_WORLD, PETSC_VIEWER_ASCII_COMMON)); 189566063dSJacob Faibussowitsch PetscCall(PetscOptionsGetInt(NULL, NULL, "-m", &m, NULL)); 19c4762a1bSJed Brown n = m; 209566063dSJacob Faibussowitsch PetscCall(PetscOptionsHasName(NULL, NULL, "-rectA", &flg)); 21c4762a1bSJed Brown if (flg) n += 2; 229566063dSJacob Faibussowitsch PetscCall(PetscOptionsHasName(NULL, NULL, "-rectB", &flg)); 23c4762a1bSJed Brown if (flg) n -= 2; 24c4762a1bSJed Brown 25c4762a1bSJed Brown /* ---------- Assemble matrix and vectors ----------- */ 26c4762a1bSJed Brown 279566063dSJacob Faibussowitsch PetscCall(MatCreate(PETSC_COMM_WORLD, &C)); 289566063dSJacob Faibussowitsch PetscCall(MatSetSizes(C, PETSC_DECIDE, PETSC_DECIDE, m, n)); 299566063dSJacob Faibussowitsch PetscCall(MatSetFromOptions(C)); 309566063dSJacob Faibussowitsch PetscCall(MatSetUp(C)); 319566063dSJacob Faibussowitsch PetscCall(MatGetOwnershipRange(C, &rstart, &rend)); 329566063dSJacob Faibussowitsch PetscCall(VecCreate(PETSC_COMM_WORLD, &x)); 339566063dSJacob Faibussowitsch PetscCall(VecSetSizes(x, PETSC_DECIDE, m)); 349566063dSJacob Faibussowitsch PetscCall(VecSetFromOptions(x)); 359566063dSJacob Faibussowitsch PetscCall(VecDuplicate(x, &z)); 369566063dSJacob Faibussowitsch PetscCall(VecDuplicate(x, &w)); 379566063dSJacob Faibussowitsch PetscCall(VecCreate(PETSC_COMM_WORLD, &y)); 389566063dSJacob Faibussowitsch PetscCall(VecSetSizes(y, PETSC_DECIDE, n)); 399566063dSJacob Faibussowitsch PetscCall(VecSetFromOptions(y)); 409566063dSJacob Faibussowitsch PetscCall(VecDuplicate(y, &u)); 419566063dSJacob Faibussowitsch PetscCall(VecDuplicate(y, &s)); 429566063dSJacob Faibussowitsch PetscCall(VecGetOwnershipRange(y, &vstart, &vend)); 43c4762a1bSJed Brown 44c4762a1bSJed Brown /* Assembly */ 45c4762a1bSJed Brown for (i = rstart; i < rend; i++) { 46c4762a1bSJed Brown v = 100 * (i + 1); 479566063dSJacob Faibussowitsch PetscCall(VecSetValues(z, 1, &i, &v, INSERT_VALUES)); 48c4762a1bSJed Brown for (j = 0; j < n; j++) { 49c4762a1bSJed Brown v = 10 * (i + 1) + j + 1; 509566063dSJacob Faibussowitsch PetscCall(MatSetValues(C, 1, &i, 1, &j, &v, INSERT_VALUES)); 51c4762a1bSJed Brown } 52c4762a1bSJed Brown } 53c4762a1bSJed Brown 54c4762a1bSJed Brown /* Flush off proc Vec values and do more assembly */ 559566063dSJacob Faibussowitsch PetscCall(VecAssemblyBegin(z)); 56c4762a1bSJed Brown for (i = vstart; i < vend; i++) { 57c4762a1bSJed Brown v = one * ((PetscReal)i); 589566063dSJacob Faibussowitsch PetscCall(VecSetValues(y, 1, &i, &v, INSERT_VALUES)); 59c4762a1bSJed Brown v = 100.0 * i; 609566063dSJacob Faibussowitsch PetscCall(VecSetValues(u, 1, &i, &v, INSERT_VALUES)); 61c4762a1bSJed Brown } 62c4762a1bSJed Brown 63c4762a1bSJed Brown /* Flush off proc Mat values and do more assembly */ 649566063dSJacob Faibussowitsch PetscCall(MatAssemblyBegin(C, MAT_FLUSH_ASSEMBLY)); 65c4762a1bSJed Brown for (i = rstart; i < rend; i++) { 66c4762a1bSJed Brown for (j = 0; j < n; j++) { 67c4762a1bSJed Brown v = 10 * (i + 1) + j + 1; 689566063dSJacob Faibussowitsch PetscCall(MatSetValues(C, 1, &i, 1, &j, &v, INSERT_VALUES)); 69c4762a1bSJed Brown } 70c4762a1bSJed Brown } 71c4762a1bSJed Brown /* Try overlap Coomunication with the next stage XXXSetValues */ 729566063dSJacob Faibussowitsch PetscCall(VecAssemblyEnd(z)); 73c4762a1bSJed Brown 749566063dSJacob Faibussowitsch PetscCall(MatAssemblyEnd(C, MAT_FLUSH_ASSEMBLY)); 75c4762a1bSJed Brown CHKMEMQ; 76c4762a1bSJed Brown /* The Assembly for the second Stage */ 779566063dSJacob Faibussowitsch PetscCall(MatAssemblyBegin(C, MAT_FINAL_ASSEMBLY)); 789566063dSJacob Faibussowitsch PetscCall(MatAssemblyEnd(C, MAT_FINAL_ASSEMBLY)); 799566063dSJacob Faibussowitsch PetscCall(VecAssemblyBegin(y)); 809566063dSJacob Faibussowitsch PetscCall(VecAssemblyEnd(y)); 819566063dSJacob Faibussowitsch PetscCall(MatScale(C, alpha)); 829566063dSJacob Faibussowitsch PetscCall(VecAssemblyBegin(u)); 839566063dSJacob Faibussowitsch PetscCall(VecAssemblyEnd(u)); 849566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "testing MatMult()\n")); 85c4762a1bSJed Brown CHKMEMQ; 869566063dSJacob Faibussowitsch PetscCall(MatMult(C, y, x)); 87c4762a1bSJed Brown CHKMEMQ; 889566063dSJacob Faibussowitsch PetscCall(VecView(x, PETSC_VIEWER_STDOUT_WORLD)); 899566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "testing MatMultAdd()\n")); 909566063dSJacob Faibussowitsch PetscCall(MatMultAdd(C, y, z, w)); 919566063dSJacob Faibussowitsch PetscCall(VecAXPY(x, one, z)); 929566063dSJacob Faibussowitsch PetscCall(VecAXPY(x, negone, w)); 939566063dSJacob Faibussowitsch PetscCall(VecNorm(x, NORM_2, &norm)); 949371c9d4SSatish Balay if (norm > tol) { PetscCall(PetscPrintf(PETSC_COMM_WORLD, "Norm of error difference = %g\n", (double)norm)); } 95c4762a1bSJed Brown 96c4762a1bSJed Brown /* ------- Test MatMultTranspose(), MatMultTransposeAdd() ------- */ 97c4762a1bSJed Brown 98c4762a1bSJed Brown for (i = rstart; i < rend; i++) { 99c4762a1bSJed Brown v = one * ((PetscReal)i); 1009566063dSJacob Faibussowitsch PetscCall(VecSetValues(x, 1, &i, &v, INSERT_VALUES)); 101c4762a1bSJed Brown } 1029566063dSJacob Faibussowitsch PetscCall(VecAssemblyBegin(x)); 1039566063dSJacob Faibussowitsch PetscCall(VecAssemblyEnd(x)); 1049566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "testing MatMultTranspose()\n")); 1059566063dSJacob Faibussowitsch PetscCall(MatMultTranspose(C, x, y)); 1069566063dSJacob Faibussowitsch PetscCall(VecView(y, PETSC_VIEWER_STDOUT_WORLD)); 107c4762a1bSJed Brown 1089566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "testing MatMultTransposeAdd()\n")); 1099566063dSJacob Faibussowitsch PetscCall(MatMultTransposeAdd(C, x, u, s)); 1109566063dSJacob Faibussowitsch PetscCall(VecAXPY(y, one, u)); 1119566063dSJacob Faibussowitsch PetscCall(VecAXPY(y, negone, s)); 1129566063dSJacob Faibussowitsch PetscCall(VecNorm(y, NORM_2, &norm)); 1139371c9d4SSatish Balay if (norm > tol) { PetscCall(PetscPrintf(PETSC_COMM_WORLD, "Norm of error difference = %g\n", (double)norm)); } 114c4762a1bSJed Brown 115c4762a1bSJed Brown /* -------------------- Test MatGetDiagonal() ------------------ */ 116c4762a1bSJed Brown 1179566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "testing MatGetDiagonal(), MatDiagonalScale()\n")); 1189566063dSJacob Faibussowitsch PetscCall(MatView(C, PETSC_VIEWER_STDOUT_WORLD)); 1199566063dSJacob Faibussowitsch PetscCall(VecSet(x, one)); 1209566063dSJacob Faibussowitsch PetscCall(MatGetDiagonal(C, x)); 1219566063dSJacob Faibussowitsch PetscCall(VecView(x, PETSC_VIEWER_STDOUT_WORLD)); 122c4762a1bSJed Brown for (i = vstart; i < vend; i++) { 123c4762a1bSJed Brown v = one * ((PetscReal)(i + 1)); 1249566063dSJacob Faibussowitsch PetscCall(VecSetValues(y, 1, &i, &v, INSERT_VALUES)); 125c4762a1bSJed Brown } 126c4762a1bSJed Brown 127c4762a1bSJed Brown /* -------------------- Test () MatDiagonalScale ------------------ */ 1289566063dSJacob Faibussowitsch PetscCall(PetscOptionsHasName(NULL, NULL, "-test_diagonalscale", &flg)); 129c4762a1bSJed Brown if (flg) { 1309566063dSJacob Faibussowitsch PetscCall(MatDiagonalScale(C, x, y)); 1319566063dSJacob Faibussowitsch PetscCall(MatView(C, PETSC_VIEWER_STDOUT_WORLD)); 132c4762a1bSJed Brown } 133c4762a1bSJed Brown /* Free data structures */ 1349371c9d4SSatish Balay PetscCall(VecDestroy(&u)); 1359371c9d4SSatish Balay PetscCall(VecDestroy(&s)); 1369371c9d4SSatish Balay PetscCall(VecDestroy(&w)); 1379371c9d4SSatish Balay PetscCall(VecDestroy(&x)); 1389371c9d4SSatish Balay PetscCall(VecDestroy(&y)); 1399371c9d4SSatish Balay PetscCall(VecDestroy(&z)); 1409566063dSJacob Faibussowitsch PetscCall(MatDestroy(&C)); 141c4762a1bSJed Brown 1429566063dSJacob Faibussowitsch PetscCall(PetscFinalize()); 143b122ec5aSJacob Faibussowitsch return 0; 144c4762a1bSJed Brown } 145c4762a1bSJed Brown 146c4762a1bSJed Brown /*TEST 147c4762a1bSJed Brown 148c4762a1bSJed Brown test: 149c4762a1bSJed Brown suffix: 11_A 150c4762a1bSJed Brown args: -mat_type seqaij -rectA 151c4762a1bSJed Brown filter: grep -v type 152c4762a1bSJed Brown 153c4762a1bSJed Brown test: 154c4762a1bSJed Brown args: -mat_type seqdense -rectA 155c4762a1bSJed Brown suffix: 12_A 156c4762a1bSJed Brown 157c4762a1bSJed Brown test: 158c4762a1bSJed Brown args: -mat_type seqaij -rectB 159c4762a1bSJed Brown suffix: 11_B 160c4762a1bSJed Brown filter: grep -v type 161c4762a1bSJed Brown 162c4762a1bSJed Brown test: 163c4762a1bSJed Brown args: -mat_type seqdense -rectB 164c4762a1bSJed Brown suffix: 12_B 165c4762a1bSJed Brown 166c4762a1bSJed Brown test: 167c4762a1bSJed Brown suffix: 21 168c4762a1bSJed Brown args: -mat_type mpiaij 169c4762a1bSJed Brown filter: grep -v type 170c4762a1bSJed Brown 171c4762a1bSJed Brown test: 172c4762a1bSJed Brown suffix: 22 173c4762a1bSJed Brown args: -mat_type mpidense 174c4762a1bSJed Brown 175c4762a1bSJed Brown test: 176c4762a1bSJed Brown suffix: 23 177c4762a1bSJed Brown nsize: 3 178c4762a1bSJed Brown args: -mat_type mpiaij 179c4762a1bSJed Brown filter: grep -v type 180c4762a1bSJed Brown 181c4762a1bSJed Brown test: 182c4762a1bSJed Brown suffix: 24 183c4762a1bSJed Brown nsize: 3 184c4762a1bSJed Brown args: -mat_type mpidense 185c4762a1bSJed Brown 186c4762a1bSJed Brown test: 187c4762a1bSJed Brown suffix: 2_aijcusparse_1 188c4762a1bSJed Brown args: -mat_type mpiaijcusparse -vec_type cuda 189c4762a1bSJed Brown filter: grep -v type 190c4762a1bSJed Brown output_file: output/ex5_21.out 191c4762a1bSJed Brown requires: cuda 192c4762a1bSJed Brown 193c4762a1bSJed Brown test: 194c4762a1bSJed Brown nsize: 3 195c4762a1bSJed Brown suffix: 2_aijcusparse_2 196c4762a1bSJed Brown filter: grep -v type 197c4762a1bSJed Brown args: -mat_type mpiaijcusparse -vec_type cuda 198bd46da1dSJunchao Zhang args: -sf_type {{basic neighbor}} 199c4762a1bSJed Brown output_file: output/ex5_23.out 200c4762a1bSJed Brown requires: cuda 201c4762a1bSJed Brown 202c4762a1bSJed Brown test: 203c4762a1bSJed Brown nsize: 3 204c4762a1bSJed Brown suffix: 2_aijcusparse_3 205c4762a1bSJed Brown filter: grep -v type 206c4762a1bSJed Brown args: -mat_type mpiaijcusparse -vec_type cuda 207c20d7725SJed Brown args: -sf_type {{basic neighbor}} 208c4762a1bSJed Brown output_file: output/ex5_23.out 209dfd57a17SPierre Jolivet requires: cuda defined(PETSC_HAVE_MPI_GPU_AWARE) 210c4762a1bSJed Brown 211c4762a1bSJed Brown test: 212c4762a1bSJed Brown suffix: 31 213c4762a1bSJed Brown args: -mat_type mpiaij -test_diagonalscale 214c4762a1bSJed Brown filter: grep -v type 215c4762a1bSJed Brown 216c4762a1bSJed Brown test: 217c4762a1bSJed Brown suffix: 32 218c4762a1bSJed Brown args: -mat_type mpibaij -test_diagonalscale 219c4762a1bSJed Brown filter: grep -v Mat_ 220c4762a1bSJed Brown 221c4762a1bSJed Brown test: 222c4762a1bSJed Brown suffix: 33 223c4762a1bSJed Brown nsize: 3 224c4762a1bSJed Brown args: -mat_type mpiaij -test_diagonalscale 225c4762a1bSJed Brown filter: grep -v type 226c4762a1bSJed Brown 227c4762a1bSJed Brown test: 228c4762a1bSJed Brown suffix: 34 229c4762a1bSJed Brown nsize: 3 230c4762a1bSJed Brown args: -mat_type mpibaij -test_diagonalscale 231c4762a1bSJed Brown filter: grep -v Mat_ 232c4762a1bSJed Brown 233c4762a1bSJed Brown test: 234c4762a1bSJed Brown suffix: 3_aijcusparse_1 235c4762a1bSJed Brown args: -mat_type mpiaijcusparse -vec_type cuda -test_diagonalscale 236c4762a1bSJed Brown filter: grep -v type 237c4762a1bSJed Brown output_file: output/ex5_31.out 238c4762a1bSJed Brown requires: cuda 239c4762a1bSJed Brown 240c4762a1bSJed Brown test: 241c4762a1bSJed Brown suffix: 3_aijcusparse_2 242c4762a1bSJed Brown nsize: 3 243c4762a1bSJed Brown args: -mat_type mpiaijcusparse -vec_type cuda -test_diagonalscale 244c4762a1bSJed Brown filter: grep -v type 245c4762a1bSJed Brown output_file: output/ex5_33.out 246c4762a1bSJed Brown requires: cuda 247c4762a1bSJed Brown 248c4762a1bSJed Brown test: 24935990778SJunchao Zhang suffix: 3_kokkos 25035990778SJunchao Zhang nsize: 3 25135990778SJunchao Zhang args: -mat_type mpiaijkokkos -vec_type kokkos -test_diagonalscale 25235990778SJunchao Zhang filter: grep -v type 25335990778SJunchao Zhang output_file: output/ex5_33.out 254*dcfd994dSJunchao Zhang requires: kokkos_kernels 25535990778SJunchao Zhang 25635990778SJunchao Zhang test: 257c4762a1bSJed Brown suffix: aijcusparse_1 258c4762a1bSJed Brown args: -mat_type seqaijcusparse -vec_type cuda -rectA 259c4762a1bSJed Brown filter: grep -v type 260c4762a1bSJed Brown output_file: output/ex5_11_A.out 261c4762a1bSJed Brown requires: cuda 262c4762a1bSJed Brown 263c4762a1bSJed Brown test: 264c4762a1bSJed Brown suffix: aijcusparse_2 265c4762a1bSJed Brown args: -mat_type seqaijcusparse -vec_type cuda -rectB 266c4762a1bSJed Brown filter: grep -v type 267c4762a1bSJed Brown output_file: output/ex5_11_B.out 268c4762a1bSJed Brown requires: cuda 269c4762a1bSJed Brown 270c4762a1bSJed Brown test: 271c4762a1bSJed Brown suffix: sell_1 272c4762a1bSJed Brown args: -mat_type sell 273c4762a1bSJed Brown output_file: output/ex5_41.out 274c4762a1bSJed Brown 275c4762a1bSJed Brown test: 276c4762a1bSJed Brown suffix: sell_2 277c4762a1bSJed Brown nsize: 3 278c4762a1bSJed Brown args: -mat_type sell 279c4762a1bSJed Brown output_file: output/ex5_43.out 280c4762a1bSJed Brown 281c4762a1bSJed Brown test: 282c4762a1bSJed Brown suffix: sell_3 283c4762a1bSJed Brown args: -mat_type sell -test_diagonalscale 284c4762a1bSJed Brown output_file: output/ex5_51.out 285c4762a1bSJed Brown 286c4762a1bSJed Brown test: 287c4762a1bSJed Brown suffix: sell_4 288c4762a1bSJed Brown nsize: 3 289c4762a1bSJed Brown args: -mat_type sell -test_diagonalscale 290c4762a1bSJed Brown output_file: output/ex5_53.out 291c4762a1bSJed Brown 292c4762a1bSJed Brown TEST*/ 293