1c4762a1bSJed Brown static char help[] = "Test sequential USFFT interface on a 3-dof field over a uniform DMDA and compares to the result of FFTW acting on a split version of the field\n\n"; 2c4762a1bSJed Brown 3c4762a1bSJed Brown /* 4c4762a1bSJed Brown Compiling the code: 5c4762a1bSJed Brown This code uses the complex numbers version of PETSc and the FFTW package, so configure 6c4762a1bSJed Brown must be run to enable these. 7c4762a1bSJed Brown 8c4762a1bSJed Brown */ 9c4762a1bSJed Brown 10c4762a1bSJed Brown #define DOF 3 11c4762a1bSJed Brown 12c4762a1bSJed Brown #include <petscmat.h> 13c4762a1bSJed Brown #include <petscdm.h> 14c4762a1bSJed Brown #include <petscdmda.h> 15d71ae5a4SJacob Faibussowitsch int main(int argc, char **args) 16d71ae5a4SJacob Faibussowitsch { 179371c9d4SSatish Balay typedef enum { 189371c9d4SSatish Balay RANDOM, 199371c9d4SSatish Balay CONSTANT, 209371c9d4SSatish Balay TANH, 219371c9d4SSatish Balay NUM_FUNCS 229371c9d4SSatish Balay } FuncType; 23c4762a1bSJed Brown const char *funcNames[NUM_FUNCS] = {"random", "constant", "tanh"}; 24c4762a1bSJed Brown Mat A, AA; 25c4762a1bSJed Brown PetscMPIInt size; 26c4762a1bSJed Brown PetscInt N, i, stencil = 1, dof = 3; 27c4762a1bSJed Brown PetscInt dim[3] = {10, 10, 10}, ndim = 3; 28c4762a1bSJed Brown Vec coords, x, y, z, xx, yy, zz; 29c4762a1bSJed Brown Vec xxsplit[DOF], yysplit[DOF], zzsplit[DOF]; 30c4762a1bSJed Brown PetscReal h[3]; 31c4762a1bSJed Brown PetscScalar s; 32c4762a1bSJed Brown PetscRandom rdm; 33c4762a1bSJed Brown PetscReal norm, enorm; 34c4762a1bSJed Brown PetscInt func, ii; 35c4762a1bSJed Brown FuncType function = TANH; 36c4762a1bSJed Brown DM da, da1, coordsda; 37c4762a1bSJed Brown PetscBool view_x = PETSC_FALSE, view_y = PETSC_FALSE, view_z = PETSC_FALSE; 38c4762a1bSJed Brown 39327415f7SBarry Smith PetscFunctionBeginUser; 40c8025a54SPierre Jolivet PetscCall(PetscInitialize(&argc, &args, NULL, help)); 419566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(PETSC_COMM_WORLD, &size)); 42*bd158744SPierre Jolivet PetscCheck(size == 1, PETSC_COMM_WORLD, PETSC_ERR_WRONG_MPI_SIZE, "This is a uniprocessor example only!"); 43d0609cedSBarry Smith PetscOptionsBegin(PETSC_COMM_WORLD, NULL, "USFFT Options", "ex27"); 449566063dSJacob Faibussowitsch PetscCall(PetscOptionsEList("-function", "Function type", "ex27", funcNames, NUM_FUNCS, funcNames[function], &func, NULL)); 45c4762a1bSJed Brown function = (FuncType)func; 46d0609cedSBarry Smith PetscOptionsEnd(); 479566063dSJacob Faibussowitsch PetscCall(PetscOptionsGetBool(NULL, NULL, "-view_x", &view_x, NULL)); 489566063dSJacob Faibussowitsch PetscCall(PetscOptionsGetBool(NULL, NULL, "-view_y", &view_y, NULL)); 499566063dSJacob Faibussowitsch PetscCall(PetscOptionsGetBool(NULL, NULL, "-view_z", &view_z, NULL)); 509566063dSJacob Faibussowitsch PetscCall(PetscOptionsGetIntArray(NULL, NULL, "-dim", dim, &ndim, NULL)); 51c4762a1bSJed Brown 52c4762a1bSJed Brown /* DMDA with the correct fiber dimension */ 539371c9d4SSatish Balay PetscCall(DMDACreate3d(PETSC_COMM_SELF, DM_BOUNDARY_NONE, DM_BOUNDARY_NONE, DM_BOUNDARY_NONE, DMDA_STENCIL_STAR, dim[0], dim[1], dim[2], PETSC_DECIDE, PETSC_DECIDE, PETSC_DECIDE, dof, stencil, NULL, NULL, NULL, &da)); 549566063dSJacob Faibussowitsch PetscCall(DMSetFromOptions(da)); 559566063dSJacob Faibussowitsch PetscCall(DMSetUp(da)); 56c4762a1bSJed Brown /* DMDA with fiber dimension 1 for split fields */ 579371c9d4SSatish Balay PetscCall(DMDACreate3d(PETSC_COMM_SELF, DM_BOUNDARY_NONE, DM_BOUNDARY_NONE, DM_BOUNDARY_NONE, DMDA_STENCIL_STAR, dim[0], dim[1], dim[2], PETSC_DECIDE, PETSC_DECIDE, PETSC_DECIDE, 1, stencil, NULL, NULL, NULL, &da1)); 589566063dSJacob Faibussowitsch PetscCall(DMSetFromOptions(da1)); 599566063dSJacob Faibussowitsch PetscCall(DMSetUp(da1)); 60c4762a1bSJed Brown 61c4762a1bSJed Brown /* Coordinates */ 629566063dSJacob Faibussowitsch PetscCall(DMGetCoordinateDM(da, &coordsda)); 639566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(coordsda, &coords)); 649566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject)coords, "Grid coordinates")); 65c4762a1bSJed Brown for (i = 0, N = 1; i < 3; i++) { 66c4762a1bSJed Brown h[i] = 1.0 / dim[i]; 67c4762a1bSJed Brown PetscScalar *a; 689566063dSJacob Faibussowitsch PetscCall(VecGetArray(coords, &a)); 69c4762a1bSJed Brown PetscInt j, k, n = 0; 70c4762a1bSJed Brown for (i = 0; i < 3; ++i) { 71c4762a1bSJed Brown for (j = 0; j < dim[i]; ++j) { 72c4762a1bSJed Brown for (k = 0; k < 3; ++k) { 73c4762a1bSJed Brown a[n] = j * h[i]; /* coordinate along the j-th point in the i-th dimension */ 74c4762a1bSJed Brown ++n; 75c4762a1bSJed Brown } 76c4762a1bSJed Brown } 77c4762a1bSJed Brown } 789566063dSJacob Faibussowitsch PetscCall(VecRestoreArray(coords, &a)); 79c4762a1bSJed Brown } 809566063dSJacob Faibussowitsch PetscCall(DMSetCoordinates(da, coords)); 819566063dSJacob Faibussowitsch PetscCall(VecDestroy(&coords)); 82c4762a1bSJed Brown 83c4762a1bSJed Brown /* Work vectors */ 849566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(da, &x)); 859566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject)x, "Real space vector")); 869566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(da, &xx)); 879566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject)xx, "Real space vector")); 889566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(da, &y)); 899566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject)y, "USFFT frequency space vector")); 909566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(da, &yy)); 919566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject)yy, "FFTW frequency space vector")); 929566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(da, &z)); 939566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject)z, "USFFT reconstructed vector")); 949566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(da, &zz)); 959566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject)zz, "FFTW reconstructed vector")); 96c4762a1bSJed Brown /* Split vectors for FFTW */ 97c4762a1bSJed Brown for (ii = 0; ii < 3; ++ii) { 989566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(da1, &xxsplit[ii])); 999566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject)xxsplit[ii], "Real space split vector")); 1009566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(da1, &yysplit[ii])); 1019566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject)yysplit[ii], "FFTW frequency space split vector")); 1029566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(da1, &zzsplit[ii])); 1039566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject)zzsplit[ii], "FFTW reconstructed split vector")); 104c4762a1bSJed Brown } 105c4762a1bSJed Brown 10663a3b9bcSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "%3-" PetscInt_FMT ": USFFT on vector of ")); 107c4762a1bSJed Brown for (i = 0, N = 1; i < 3; i++) { 1089566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "dim[%d] = %d ", i, dim[i])); 109c4762a1bSJed Brown N *= dim[i]; 110c4762a1bSJed Brown } 1119566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "; total size %d \n", N)); 112c4762a1bSJed Brown 113c4762a1bSJed Brown if (function == RANDOM) { 1149566063dSJacob Faibussowitsch PetscCall(PetscRandomCreate(PETSC_COMM_SELF, &rdm)); 1159566063dSJacob Faibussowitsch PetscCall(PetscRandomSetFromOptions(rdm)); 1169566063dSJacob Faibussowitsch PetscCall(VecSetRandom(x, rdm)); 1179566063dSJacob Faibussowitsch PetscCall(PetscRandomDestroy(&rdm)); 118c4762a1bSJed Brown } else if (function == CONSTANT) { 1199566063dSJacob Faibussowitsch PetscCall(VecSet(x, 1.0)); 120c4762a1bSJed Brown } else if (function == TANH) { 121c4762a1bSJed Brown PetscScalar *a; 1229566063dSJacob Faibussowitsch PetscCall(VecGetArray(x, &a)); 123c4762a1bSJed Brown PetscInt j, k = 0; 124c4762a1bSJed Brown for (i = 0; i < 3; ++i) { 125c4762a1bSJed Brown for (j = 0; j < dim[i]; ++j) { 126c4762a1bSJed Brown a[k] = tanh((j - dim[i] / 2.0) * (10.0 / dim[i])); 127c4762a1bSJed Brown ++k; 128c4762a1bSJed Brown } 129c4762a1bSJed Brown } 1309566063dSJacob Faibussowitsch PetscCall(VecRestoreArray(x, &a)); 131c4762a1bSJed Brown } 1321baa6e33SBarry Smith if (view_x) PetscCall(VecView(x, PETSC_VIEWER_STDOUT_WORLD)); 1339566063dSJacob Faibussowitsch PetscCall(VecCopy(x, xx)); 134c4762a1bSJed Brown /* Split xx */ 1359566063dSJacob Faibussowitsch PetscCall(VecStrideGatherAll(xx, xxsplit, INSERT_VALUES)); /*YES! 'Gather' means 'split' (or maybe 'scatter'?)! */ 136c4762a1bSJed Brown 1379566063dSJacob Faibussowitsch PetscCall(VecNorm(x, NORM_2, &norm)); 1389566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "|x|_2 = %g\n", norm)); 139c4762a1bSJed Brown 140c4762a1bSJed Brown /* create USFFT object */ 1419566063dSJacob Faibussowitsch PetscCall(MatCreateSeqUSFFT(da, da, &A)); 142c4762a1bSJed Brown /* create FFTW object */ 1439566063dSJacob Faibussowitsch PetscCall(MatCreateSeqFFTW(PETSC_COMM_SELF, 3, dim, &AA)); 144c4762a1bSJed Brown 145c4762a1bSJed Brown /* apply USFFT and FFTW FORWARD "preemptively", so the fftw_plans can be reused on different vectors */ 1469566063dSJacob Faibussowitsch PetscCall(MatMult(A, x, z)); 14748a46eb9SPierre Jolivet for (ii = 0; ii < 3; ++ii) PetscCall(MatMult(AA, xxsplit[ii], zzsplit[ii])); 148c4762a1bSJed Brown /* Now apply USFFT and FFTW forward several (3) times */ 149c4762a1bSJed Brown for (i = 0; i < 3; ++i) { 1509566063dSJacob Faibussowitsch PetscCall(MatMult(A, x, y)); 15148a46eb9SPierre Jolivet for (ii = 0; ii < 3; ++ii) PetscCall(MatMult(AA, xxsplit[ii], yysplit[ii])); 1529566063dSJacob Faibussowitsch PetscCall(MatMultTranspose(A, y, z)); 15348a46eb9SPierre Jolivet for (ii = 0; ii < 3; ++ii) PetscCall(MatMult(AA, yysplit[ii], zzsplit[ii])); 154c4762a1bSJed Brown } 155c4762a1bSJed Brown /* Unsplit yy */ 1569566063dSJacob Faibussowitsch PetscCall(VecStrideScatterAll(yysplit, yy, INSERT_VALUES)); /*YES! 'Scatter' means 'collect' (or maybe 'gather'?)! */ 157c4762a1bSJed Brown /* Unsplit zz */ 1589566063dSJacob Faibussowitsch PetscCall(VecStrideScatterAll(zzsplit, zz, INSERT_VALUES)); /*YES! 'Scatter' means 'collect' (or maybe 'gather'?)! */ 159c4762a1bSJed Brown 160c4762a1bSJed Brown if (view_y) { 1619566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "y = \n")); 1629566063dSJacob Faibussowitsch PetscCall(VecView(y, PETSC_VIEWER_STDOUT_WORLD)); 1639566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "yy = \n")); 1649566063dSJacob Faibussowitsch PetscCall(VecView(yy, PETSC_VIEWER_STDOUT_WORLD)); 165c4762a1bSJed Brown } 166c4762a1bSJed Brown 167c4762a1bSJed Brown if (view_z) { 1689566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "z = \n")); 1699566063dSJacob Faibussowitsch PetscCall(VecView(z, PETSC_VIEWER_STDOUT_WORLD)); 1709566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "zz = \n")); 1719566063dSJacob Faibussowitsch PetscCall(VecView(zz, PETSC_VIEWER_STDOUT_WORLD)); 172c4762a1bSJed Brown } 173c4762a1bSJed Brown 174c4762a1bSJed Brown /* compare x and z. USFFT computes an unnormalized DFT, thus z = N*x */ 175c4762a1bSJed Brown s = 1.0 / (PetscReal)N; 1769566063dSJacob Faibussowitsch PetscCall(VecScale(z, s)); 1779566063dSJacob Faibussowitsch PetscCall(VecAXPY(x, -1.0, z)); 1789566063dSJacob Faibussowitsch PetscCall(VecNorm(x, NORM_1, &enorm)); 1799566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "|x-z| = %g\n", enorm)); 180c4762a1bSJed Brown 181c4762a1bSJed Brown /* compare xx and zz. FFTW computes an unnormalized DFT, thus zz = N*x */ 182c4762a1bSJed Brown s = 1.0 / (PetscReal)N; 1839566063dSJacob Faibussowitsch PetscCall(VecScale(zz, s)); 1849566063dSJacob Faibussowitsch PetscCall(VecAXPY(xx, -1.0, zz)); 1859566063dSJacob Faibussowitsch PetscCall(VecNorm(xx, NORM_1, &enorm)); 1869566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "|xx-zz| = %g\n", enorm)); 187c4762a1bSJed Brown 188c4762a1bSJed Brown /* compare y and yy: USFFT and FFTW results*/ 1899566063dSJacob Faibussowitsch PetscCall(VecNorm(y, NORM_2, &norm)); 1909566063dSJacob Faibussowitsch PetscCall(VecAXPY(y, -1.0, yy)); 1919566063dSJacob Faibussowitsch PetscCall(VecNorm(y, NORM_1, &enorm)); 1929566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "|y|_2 = %g\n", norm)); 1939566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "|y-yy| = %g\n", enorm)); 194c4762a1bSJed Brown 195c4762a1bSJed Brown /* compare z and zz: USFFT and FFTW results*/ 1969566063dSJacob Faibussowitsch PetscCall(VecNorm(z, NORM_2, &norm)); 1979566063dSJacob Faibussowitsch PetscCall(VecAXPY(z, -1.0, zz)); 1989566063dSJacob Faibussowitsch PetscCall(VecNorm(z, NORM_1, &enorm)); 1999566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "|z|_2 = %g\n", norm)); 2009566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "|z-zz| = %g\n", enorm)); 201c4762a1bSJed Brown 202c4762a1bSJed Brown /* free spaces */ 2039566063dSJacob Faibussowitsch PetscCall(DMRestoreGlobalVector(da, &x)); 2049566063dSJacob Faibussowitsch PetscCall(DMRestoreGlobalVector(da, &xx)); 2059566063dSJacob Faibussowitsch PetscCall(DMRestoreGlobalVector(da, &y)); 2069566063dSJacob Faibussowitsch PetscCall(DMRestoreGlobalVector(da, &yy)); 2079566063dSJacob Faibussowitsch PetscCall(DMRestoreGlobalVector(da, &z)); 2089566063dSJacob Faibussowitsch PetscCall(DMRestoreGlobalVector(da, &zz)); 209c4762a1bSJed Brown 2109566063dSJacob Faibussowitsch PetscCall(PetscFinalize()); 211b122ec5aSJacob Faibussowitsch return 0; 212c4762a1bSJed Brown } 213