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> 15*ad0e0ea3SStefano Zampini int main(int argc,char **args) 16c4762a1bSJed Brown { 17c4762a1bSJed Brown typedef enum {RANDOM, CONSTANT, TANH, NUM_FUNCS} FuncType; 18c4762a1bSJed Brown const char *funcNames[NUM_FUNCS] = {"random", "constant", "tanh"}; 19c4762a1bSJed Brown Mat A, AA; 20c4762a1bSJed Brown PetscMPIInt size; 21c4762a1bSJed Brown PetscInt N,i, stencil=1,dof=3; 22c4762a1bSJed Brown PetscInt dim[3] = {10,10,10}, ndim = 3; 23c4762a1bSJed Brown Vec coords,x,y,z,xx, yy, zz; 24c4762a1bSJed Brown Vec xxsplit[DOF], yysplit[DOF], zzsplit[DOF]; 25c4762a1bSJed Brown PetscReal h[3]; 26c4762a1bSJed Brown PetscScalar s; 27c4762a1bSJed Brown PetscRandom rdm; 28c4762a1bSJed Brown PetscReal norm, enorm; 29c4762a1bSJed Brown PetscInt func,ii; 30c4762a1bSJed Brown FuncType function = TANH; 31c4762a1bSJed Brown DM da, da1, coordsda; 32c4762a1bSJed Brown PetscBool view_x = PETSC_FALSE, view_y = PETSC_FALSE, view_z = PETSC_FALSE; 33c4762a1bSJed Brown PetscErrorCode ierr; 34c4762a1bSJed Brown 35c4762a1bSJed Brown ierr = PetscInitialize(&argc,&args,(char*)0,help);if (ierr) return ierr; 36c4762a1bSJed Brown ierr = MPI_Comm_size(PETSC_COMM_WORLD, &size);CHKERRQ(ierr); 37c4762a1bSJed Brown if (size != 1) SETERRQ(PETSC_COMM_WORLD,PETSC_ERR_SUP, "This is a uniprocessor example only!"); 38c4762a1bSJed Brown ierr = PetscOptionsBegin(PETSC_COMM_WORLD, NULL, "USFFT Options", "ex27");CHKERRQ(ierr); 39c4762a1bSJed Brown ierr = PetscOptionsEList("-function", "Function type", "ex27", funcNames, NUM_FUNCS, funcNames[function], &func, NULL);CHKERRQ(ierr); 40c4762a1bSJed Brown function = (FuncType) func; 41c4762a1bSJed Brown ierr = PetscOptionsEnd();CHKERRQ(ierr); 42c4762a1bSJed Brown ierr = PetscOptionsGetBool(NULL,NULL,"-view_x",&view_x,NULL);CHKERRQ(ierr); 43c4762a1bSJed Brown ierr = PetscOptionsGetBool(NULL,NULL,"-view_y",&view_y,NULL);CHKERRQ(ierr); 44c4762a1bSJed Brown ierr = PetscOptionsGetBool(NULL,NULL,"-view_z",&view_z,NULL);CHKERRQ(ierr); 45c4762a1bSJed Brown ierr = PetscOptionsGetIntArray(NULL,NULL,"-dim",dim,&ndim,NULL);CHKERRQ(ierr); 46c4762a1bSJed Brown 47c4762a1bSJed Brown /* DMDA with the correct fiber dimension */ 48c4762a1bSJed Brown ierr = 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, 49c4762a1bSJed Brown dof,stencil,NULL,NULL,NULL,&da);CHKERRQ(ierr); 50c4762a1bSJed Brown ierr = DMSetFromOptions(da);CHKERRQ(ierr); 51c4762a1bSJed Brown ierr = DMSetUp(da);CHKERRQ(ierr); 52c4762a1bSJed Brown /* DMDA with fiber dimension 1 for split fields */ 53c4762a1bSJed Brown ierr = 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, 54c4762a1bSJed Brown 1,stencil,NULL,NULL,NULL,&da1);CHKERRQ(ierr); 55c4762a1bSJed Brown ierr = DMSetFromOptions(da1);CHKERRQ(ierr); 56c4762a1bSJed Brown ierr = DMSetUp(da1);CHKERRQ(ierr); 57c4762a1bSJed Brown 58c4762a1bSJed Brown /* Coordinates */ 59c4762a1bSJed Brown ierr = DMGetCoordinateDM(da,&coordsda); 60c4762a1bSJed Brown ierr = DMGetGlobalVector(coordsda,&coords);CHKERRQ(ierr); 61c4762a1bSJed Brown ierr = PetscObjectSetName((PetscObject) coords,"Grid coordinates");CHKERRQ(ierr); 62c4762a1bSJed Brown for (i = 0, N = 1; i < 3; i++) { 63c4762a1bSJed Brown h[i] = 1.0/dim[i]; 64c4762a1bSJed Brown PetscScalar *a; 65c4762a1bSJed Brown ierr = VecGetArray(coords, &a);CHKERRQ(ierr); 66c4762a1bSJed Brown PetscInt j,k,n = 0; 67c4762a1bSJed Brown for (i = 0; i < 3; ++i) { 68c4762a1bSJed Brown for (j = 0; j < dim[i]; ++j) { 69c4762a1bSJed Brown for (k = 0; k < 3; ++k) { 70c4762a1bSJed Brown a[n] = j*h[i]; /* coordinate along the j-th point in the i-th dimension */ 71c4762a1bSJed Brown ++n; 72c4762a1bSJed Brown } 73c4762a1bSJed Brown } 74c4762a1bSJed Brown } 75c4762a1bSJed Brown ierr = VecRestoreArray(coords, &a);CHKERRQ(ierr); 76c4762a1bSJed Brown 77c4762a1bSJed Brown } 78c4762a1bSJed Brown ierr = DMSetCoordinates(da, coords);CHKERRQ(ierr); 79c4762a1bSJed Brown ierr = VecDestroy(&coords);CHKERRQ(ierr); 80c4762a1bSJed Brown 81c4762a1bSJed Brown /* Work vectors */ 82c4762a1bSJed Brown ierr = DMGetGlobalVector(da, &x);CHKERRQ(ierr); 83c4762a1bSJed Brown ierr = PetscObjectSetName((PetscObject) x, "Real space vector");CHKERRQ(ierr); 84c4762a1bSJed Brown ierr = DMGetGlobalVector(da, &xx);CHKERRQ(ierr); 85c4762a1bSJed Brown ierr = PetscObjectSetName((PetscObject) xx, "Real space vector");CHKERRQ(ierr); 86c4762a1bSJed Brown ierr = DMGetGlobalVector(da, &y);CHKERRQ(ierr); 87c4762a1bSJed Brown ierr = PetscObjectSetName((PetscObject) y, "USFFT frequency space vector");CHKERRQ(ierr); 88c4762a1bSJed Brown ierr = DMGetGlobalVector(da, &yy);CHKERRQ(ierr); 89c4762a1bSJed Brown ierr = PetscObjectSetName((PetscObject) yy, "FFTW frequency space vector");CHKERRQ(ierr); 90c4762a1bSJed Brown ierr = DMGetGlobalVector(da, &z);CHKERRQ(ierr); 91c4762a1bSJed Brown ierr = PetscObjectSetName((PetscObject) z, "USFFT reconstructed vector");CHKERRQ(ierr); 92c4762a1bSJed Brown ierr = DMGetGlobalVector(da, &zz);CHKERRQ(ierr); 93c4762a1bSJed Brown ierr = PetscObjectSetName((PetscObject) zz, "FFTW reconstructed vector");CHKERRQ(ierr); 94c4762a1bSJed Brown /* Split vectors for FFTW */ 95c4762a1bSJed Brown for (ii = 0; ii < 3; ++ii) { 96c4762a1bSJed Brown ierr = DMGetGlobalVector(da1, &xxsplit[ii]);CHKERRQ(ierr); 97c4762a1bSJed Brown ierr = PetscObjectSetName((PetscObject) xxsplit[ii], "Real space split vector");CHKERRQ(ierr); 98c4762a1bSJed Brown ierr = DMGetGlobalVector(da1, &yysplit[ii]);CHKERRQ(ierr); 99c4762a1bSJed Brown ierr = PetscObjectSetName((PetscObject) yysplit[ii], "FFTW frequency space split vector");CHKERRQ(ierr); 100c4762a1bSJed Brown ierr = DMGetGlobalVector(da1, &zzsplit[ii]);CHKERRQ(ierr); 101c4762a1bSJed Brown ierr = PetscObjectSetName((PetscObject) zzsplit[ii], "FFTW reconstructed split vector");CHKERRQ(ierr); 102c4762a1bSJed Brown } 103c4762a1bSJed Brown 104c4762a1bSJed Brown 105c4762a1bSJed Brown ierr = PetscPrintf(PETSC_COMM_SELF, "%3-D: USFFT on vector of ");CHKERRQ(ierr); 106c4762a1bSJed Brown for (i = 0, N = 1; i < 3; i++) { 107c4762a1bSJed Brown ierr = PetscPrintf(PETSC_COMM_SELF, "dim[%d] = %d ",i,dim[i]);CHKERRQ(ierr); 108c4762a1bSJed Brown N *= dim[i]; 109c4762a1bSJed Brown } 110c4762a1bSJed Brown ierr = PetscPrintf(PETSC_COMM_SELF, "; total size %d \n",N);CHKERRQ(ierr); 111c4762a1bSJed Brown 112c4762a1bSJed Brown 113c4762a1bSJed Brown if (function == RANDOM) { 114c4762a1bSJed Brown ierr = PetscRandomCreate(PETSC_COMM_SELF, &rdm);CHKERRQ(ierr); 115c4762a1bSJed Brown ierr = PetscRandomSetFromOptions(rdm);CHKERRQ(ierr); 116c4762a1bSJed Brown ierr = VecSetRandom(x, rdm);CHKERRQ(ierr); 117c4762a1bSJed Brown ierr = PetscRandomDestroy(&rdm);CHKERRQ(ierr); 118c4762a1bSJed Brown } else if (function == CONSTANT) { 119c4762a1bSJed Brown ierr = VecSet(x, 1.0);CHKERRQ(ierr); 120c4762a1bSJed Brown } else if (function == TANH) { 121c4762a1bSJed Brown PetscScalar *a; 122c4762a1bSJed Brown ierr = VecGetArray(x, &a);CHKERRQ(ierr); 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 } 130c4762a1bSJed Brown ierr = VecRestoreArray(x, &a);CHKERRQ(ierr); 131c4762a1bSJed Brown } 132c4762a1bSJed Brown if (view_x) { 133c4762a1bSJed Brown ierr = VecView(x, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr); 134c4762a1bSJed Brown } 135c4762a1bSJed Brown ierr = VecCopy(x,xx);CHKERRQ(ierr); 136c4762a1bSJed Brown /* Split xx */ 137c4762a1bSJed Brown ierr = VecStrideGatherAll(xx,xxsplit, INSERT_VALUES);CHKERRQ(ierr); /*YES! 'Gather' means 'split' (or maybe 'scatter'?)! */ 138c4762a1bSJed Brown 139c4762a1bSJed Brown ierr = VecNorm(x,NORM_2,&norm);CHKERRQ(ierr); 140c4762a1bSJed Brown ierr = PetscPrintf(PETSC_COMM_SELF, "|x|_2 = %g\n",norm);CHKERRQ(ierr); 141c4762a1bSJed Brown 142c4762a1bSJed Brown /* create USFFT object */ 143c4762a1bSJed Brown ierr = MatCreateSeqUSFFT(da,da,&A);CHKERRQ(ierr); 144c4762a1bSJed Brown /* create FFTW object */ 145c4762a1bSJed Brown ierr = MatCreateSeqFFTW(PETSC_COMM_SELF,3,dim,&AA);CHKERRQ(ierr); 146c4762a1bSJed Brown 147c4762a1bSJed Brown /* apply USFFT and FFTW FORWARD "preemptively", so the fftw_plans can be reused on different vectors */ 148c4762a1bSJed Brown ierr = MatMult(A,x,z);CHKERRQ(ierr); 149c4762a1bSJed Brown for (ii = 0; ii < 3; ++ii) { 150c4762a1bSJed Brown ierr = MatMult(AA,xxsplit[ii],zzsplit[ii]);CHKERRQ(ierr); 151c4762a1bSJed Brown } 152c4762a1bSJed Brown /* Now apply USFFT and FFTW forward several (3) times */ 153c4762a1bSJed Brown for (i=0; i<3; ++i) { 154c4762a1bSJed Brown ierr = MatMult(A,x,y);CHKERRQ(ierr); 155c4762a1bSJed Brown for (ii = 0; ii < 3; ++ii) { 156c4762a1bSJed Brown ierr = MatMult(AA,xxsplit[ii],yysplit[ii]);CHKERRQ(ierr); 157c4762a1bSJed Brown } 158c4762a1bSJed Brown ierr = MatMultTranspose(A,y,z);CHKERRQ(ierr); 159c4762a1bSJed Brown for (ii = 0; ii < 3; ++ii) { 160c4762a1bSJed Brown ierr = MatMult(AA,yysplit[ii],zzsplit[ii]);CHKERRQ(ierr); 161c4762a1bSJed Brown } 162c4762a1bSJed Brown } 163c4762a1bSJed Brown /* Unsplit yy */ 164c4762a1bSJed Brown ierr = VecStrideScatterAll(yysplit, yy, INSERT_VALUES);CHKERRQ(ierr); /*YES! 'Scatter' means 'collect' (or maybe 'gather'?)! */ 165c4762a1bSJed Brown /* Unsplit zz */ 166c4762a1bSJed Brown ierr = VecStrideScatterAll(zzsplit, zz, INSERT_VALUES);CHKERRQ(ierr); /*YES! 'Scatter' means 'collect' (or maybe 'gather'?)! */ 167c4762a1bSJed Brown 168c4762a1bSJed Brown if (view_y) { 169c4762a1bSJed Brown ierr = PetscPrintf(PETSC_COMM_WORLD, "y = \n");CHKERRQ(ierr); 170c4762a1bSJed Brown ierr = VecView(y, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr); 171c4762a1bSJed Brown ierr = PetscPrintf(PETSC_COMM_WORLD, "yy = \n");CHKERRQ(ierr); 172c4762a1bSJed Brown ierr = VecView(yy, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr); 173c4762a1bSJed Brown } 174c4762a1bSJed Brown 175c4762a1bSJed Brown if (view_z) { 176c4762a1bSJed Brown ierr = PetscPrintf(PETSC_COMM_WORLD, "z = \n");CHKERRQ(ierr); 177c4762a1bSJed Brown ierr = VecView(z, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr); 178c4762a1bSJed Brown ierr = PetscPrintf(PETSC_COMM_WORLD, "zz = \n");CHKERRQ(ierr); 179c4762a1bSJed Brown ierr = VecView(zz, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr); 180c4762a1bSJed Brown } 181c4762a1bSJed Brown 182c4762a1bSJed Brown /* compare x and z. USFFT computes an unnormalized DFT, thus z = N*x */ 183c4762a1bSJed Brown s = 1.0/(PetscReal)N; 184c4762a1bSJed Brown ierr = VecScale(z,s);CHKERRQ(ierr); 185c4762a1bSJed Brown ierr = VecAXPY(x,-1.0,z);CHKERRQ(ierr); 186c4762a1bSJed Brown ierr = VecNorm(x,NORM_1,&enorm);CHKERRQ(ierr); 187c4762a1bSJed Brown ierr = PetscPrintf(PETSC_COMM_SELF, "|x-z| = %g\n",enorm);CHKERRQ(ierr); 188c4762a1bSJed Brown 189c4762a1bSJed Brown /* compare xx and zz. FFTW computes an unnormalized DFT, thus zz = N*x */ 190c4762a1bSJed Brown s = 1.0/(PetscReal)N; 191c4762a1bSJed Brown ierr = VecScale(zz,s);CHKERRQ(ierr); 192c4762a1bSJed Brown ierr = VecAXPY(xx,-1.0,zz);CHKERRQ(ierr); 193c4762a1bSJed Brown ierr = VecNorm(xx,NORM_1,&enorm);CHKERRQ(ierr); 194c4762a1bSJed Brown ierr = PetscPrintf(PETSC_COMM_SELF, "|xx-zz| = %g\n",enorm);CHKERRQ(ierr); 195c4762a1bSJed Brown 196c4762a1bSJed Brown /* compare y and yy: USFFT and FFTW results*/ 197c4762a1bSJed Brown ierr = VecNorm(y,NORM_2,&norm);CHKERRQ(ierr); 198c4762a1bSJed Brown ierr = VecAXPY(y,-1.0,yy);CHKERRQ(ierr); 199c4762a1bSJed Brown ierr = VecNorm(y,NORM_1,&enorm);CHKERRQ(ierr); 200c4762a1bSJed Brown ierr = PetscPrintf(PETSC_COMM_SELF, "|y|_2 = %g\n",norm);CHKERRQ(ierr); 201c4762a1bSJed Brown ierr = PetscPrintf(PETSC_COMM_SELF, "|y-yy| = %g\n",enorm);CHKERRQ(ierr); 202c4762a1bSJed Brown 203c4762a1bSJed Brown /* compare z and zz: USFFT and FFTW results*/ 204c4762a1bSJed Brown ierr = VecNorm(z,NORM_2,&norm);CHKERRQ(ierr); 205c4762a1bSJed Brown ierr = VecAXPY(z,-1.0,zz);CHKERRQ(ierr); 206c4762a1bSJed Brown ierr = VecNorm(z,NORM_1,&enorm);CHKERRQ(ierr); 207c4762a1bSJed Brown ierr = PetscPrintf(PETSC_COMM_SELF, "|z|_2 = %g\n",norm);CHKERRQ(ierr); 208c4762a1bSJed Brown ierr = PetscPrintf(PETSC_COMM_SELF, "|z-zz| = %g\n",enorm);CHKERRQ(ierr); 209c4762a1bSJed Brown 210c4762a1bSJed Brown 211c4762a1bSJed Brown /* free spaces */ 212c4762a1bSJed Brown ierr = DMRestoreGlobalVector(da,&x);CHKERRQ(ierr); 213c4762a1bSJed Brown ierr = DMRestoreGlobalVector(da,&xx);CHKERRQ(ierr); 214c4762a1bSJed Brown ierr = DMRestoreGlobalVector(da,&y);CHKERRQ(ierr); 215c4762a1bSJed Brown ierr = DMRestoreGlobalVector(da,&yy);CHKERRQ(ierr); 216c4762a1bSJed Brown ierr = DMRestoreGlobalVector(da,&z);CHKERRQ(ierr); 217c4762a1bSJed Brown ierr = DMRestoreGlobalVector(da,&zz);CHKERRQ(ierr); 218c4762a1bSJed Brown 219c4762a1bSJed Brown ierr = PetscFinalize(); 220c4762a1bSJed Brown return ierr; 221c4762a1bSJed Brown } 222