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> 15ad0e0ea3SStefano 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 349566063dSJacob Faibussowitsch PetscCall(PetscInitialize(&argc,&args,(char*)0,help)); 359566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(PETSC_COMM_WORLD, &size)); 3608401ef6SPierre Jolivet PetscCheck(size == 1,PETSC_COMM_WORLD,PETSC_ERR_SUP, "This is a uniprocessor example only!"); 37d0609cedSBarry Smith PetscOptionsBegin(PETSC_COMM_WORLD, NULL, "USFFT Options", "ex27"); 389566063dSJacob Faibussowitsch PetscCall(PetscOptionsEList("-function", "Function type", "ex27", funcNames, NUM_FUNCS, funcNames[function], &func, NULL)); 39c4762a1bSJed Brown function = (FuncType) func; 40d0609cedSBarry Smith PetscOptionsEnd(); 419566063dSJacob Faibussowitsch PetscCall(PetscOptionsGetBool(NULL,NULL,"-view_x",&view_x,NULL)); 429566063dSJacob Faibussowitsch PetscCall(PetscOptionsGetBool(NULL,NULL,"-view_y",&view_y,NULL)); 439566063dSJacob Faibussowitsch PetscCall(PetscOptionsGetBool(NULL,NULL,"-view_z",&view_z,NULL)); 449566063dSJacob Faibussowitsch PetscCall(PetscOptionsGetIntArray(NULL,NULL,"-dim",dim,&ndim,NULL)); 45c4762a1bSJed Brown 46c4762a1bSJed Brown /* DMDA with the correct fiber dimension */ 47d0609cedSBarry Smith 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, 48d0609cedSBarry Smith dof,stencil,NULL,NULL,NULL,&da)); 499566063dSJacob Faibussowitsch PetscCall(DMSetFromOptions(da)); 509566063dSJacob Faibussowitsch PetscCall(DMSetUp(da)); 51c4762a1bSJed Brown /* DMDA with fiber dimension 1 for split fields */ 52d0609cedSBarry Smith 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, 53d0609cedSBarry Smith 1,stencil,NULL,NULL,NULL,&da1)); 549566063dSJacob Faibussowitsch PetscCall(DMSetFromOptions(da1)); 559566063dSJacob Faibussowitsch PetscCall(DMSetUp(da1)); 56c4762a1bSJed Brown 57c4762a1bSJed Brown /* Coordinates */ 589566063dSJacob Faibussowitsch PetscCall(DMGetCoordinateDM(da,&coordsda)); 599566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(coordsda,&coords)); 609566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject) coords,"Grid coordinates")); 61c4762a1bSJed Brown for (i = 0, N = 1; i < 3; i++) { 62c4762a1bSJed Brown h[i] = 1.0/dim[i]; 63c4762a1bSJed Brown PetscScalar *a; 649566063dSJacob Faibussowitsch PetscCall(VecGetArray(coords, &a)); 65c4762a1bSJed Brown PetscInt j,k,n = 0; 66c4762a1bSJed Brown for (i = 0; i < 3; ++i) { 67c4762a1bSJed Brown for (j = 0; j < dim[i]; ++j) { 68c4762a1bSJed Brown for (k = 0; k < 3; ++k) { 69c4762a1bSJed Brown a[n] = j*h[i]; /* coordinate along the j-th point in the i-th dimension */ 70c4762a1bSJed Brown ++n; 71c4762a1bSJed Brown } 72c4762a1bSJed Brown } 73c4762a1bSJed Brown } 749566063dSJacob Faibussowitsch PetscCall(VecRestoreArray(coords, &a)); 75c4762a1bSJed Brown 76c4762a1bSJed Brown } 779566063dSJacob Faibussowitsch PetscCall(DMSetCoordinates(da, coords)); 789566063dSJacob Faibussowitsch PetscCall(VecDestroy(&coords)); 79c4762a1bSJed Brown 80c4762a1bSJed Brown /* Work vectors */ 819566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(da, &x)); 829566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject) x, "Real space vector")); 839566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(da, &xx)); 849566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject) xx, "Real space vector")); 859566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(da, &y)); 869566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject) y, "USFFT frequency space vector")); 879566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(da, &yy)); 889566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject) yy, "FFTW frequency space vector")); 899566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(da, &z)); 909566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject) z, "USFFT reconstructed vector")); 919566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(da, &zz)); 929566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject) zz, "FFTW reconstructed vector")); 93c4762a1bSJed Brown /* Split vectors for FFTW */ 94c4762a1bSJed Brown for (ii = 0; ii < 3; ++ii) { 959566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(da1, &xxsplit[ii])); 969566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject) xxsplit[ii], "Real space split vector")); 979566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(da1, &yysplit[ii])); 989566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject) yysplit[ii], "FFTW frequency space split vector")); 999566063dSJacob Faibussowitsch PetscCall(DMGetGlobalVector(da1, &zzsplit[ii])); 1009566063dSJacob Faibussowitsch PetscCall(PetscObjectSetName((PetscObject) zzsplit[ii], "FFTW reconstructed split vector")); 101c4762a1bSJed Brown } 102c4762a1bSJed Brown 103*63a3b9bcSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "%3-" PetscInt_FMT ": USFFT on vector of ")); 104c4762a1bSJed Brown for (i = 0, N = 1; i < 3; i++) { 1059566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "dim[%d] = %d ",i,dim[i])); 106c4762a1bSJed Brown N *= dim[i]; 107c4762a1bSJed Brown } 1089566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "; total size %d \n",N)); 109c4762a1bSJed Brown 110c4762a1bSJed Brown if (function == RANDOM) { 1119566063dSJacob Faibussowitsch PetscCall(PetscRandomCreate(PETSC_COMM_SELF, &rdm)); 1129566063dSJacob Faibussowitsch PetscCall(PetscRandomSetFromOptions(rdm)); 1139566063dSJacob Faibussowitsch PetscCall(VecSetRandom(x, rdm)); 1149566063dSJacob Faibussowitsch PetscCall(PetscRandomDestroy(&rdm)); 115c4762a1bSJed Brown } else if (function == CONSTANT) { 1169566063dSJacob Faibussowitsch PetscCall(VecSet(x, 1.0)); 117c4762a1bSJed Brown } else if (function == TANH) { 118c4762a1bSJed Brown PetscScalar *a; 1199566063dSJacob Faibussowitsch PetscCall(VecGetArray(x, &a)); 120c4762a1bSJed Brown PetscInt j,k = 0; 121c4762a1bSJed Brown for (i = 0; i < 3; ++i) { 122c4762a1bSJed Brown for (j = 0; j < dim[i]; ++j) { 123c4762a1bSJed Brown a[k] = tanh((j - dim[i]/2.0)*(10.0/dim[i])); 124c4762a1bSJed Brown ++k; 125c4762a1bSJed Brown } 126c4762a1bSJed Brown } 1279566063dSJacob Faibussowitsch PetscCall(VecRestoreArray(x, &a)); 128c4762a1bSJed Brown } 129c4762a1bSJed Brown if (view_x) { 1309566063dSJacob Faibussowitsch PetscCall(VecView(x, PETSC_VIEWER_STDOUT_WORLD)); 131c4762a1bSJed Brown } 1329566063dSJacob Faibussowitsch PetscCall(VecCopy(x,xx)); 133c4762a1bSJed Brown /* Split xx */ 1349566063dSJacob Faibussowitsch PetscCall(VecStrideGatherAll(xx,xxsplit, INSERT_VALUES)); /*YES! 'Gather' means 'split' (or maybe 'scatter'?)! */ 135c4762a1bSJed Brown 1369566063dSJacob Faibussowitsch PetscCall(VecNorm(x,NORM_2,&norm)); 1379566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "|x|_2 = %g\n",norm)); 138c4762a1bSJed Brown 139c4762a1bSJed Brown /* create USFFT object */ 1409566063dSJacob Faibussowitsch PetscCall(MatCreateSeqUSFFT(da,da,&A)); 141c4762a1bSJed Brown /* create FFTW object */ 1429566063dSJacob Faibussowitsch PetscCall(MatCreateSeqFFTW(PETSC_COMM_SELF,3,dim,&AA)); 143c4762a1bSJed Brown 144c4762a1bSJed Brown /* apply USFFT and FFTW FORWARD "preemptively", so the fftw_plans can be reused on different vectors */ 1459566063dSJacob Faibussowitsch PetscCall(MatMult(A,x,z)); 146c4762a1bSJed Brown for (ii = 0; ii < 3; ++ii) { 1479566063dSJacob Faibussowitsch PetscCall(MatMult(AA,xxsplit[ii],zzsplit[ii])); 148c4762a1bSJed Brown } 149c4762a1bSJed Brown /* Now apply USFFT and FFTW forward several (3) times */ 150c4762a1bSJed Brown for (i=0; i<3; ++i) { 1519566063dSJacob Faibussowitsch PetscCall(MatMult(A,x,y)); 152c4762a1bSJed Brown for (ii = 0; ii < 3; ++ii) { 1539566063dSJacob Faibussowitsch PetscCall(MatMult(AA,xxsplit[ii],yysplit[ii])); 154c4762a1bSJed Brown } 1559566063dSJacob Faibussowitsch PetscCall(MatMultTranspose(A,y,z)); 156c4762a1bSJed Brown for (ii = 0; ii < 3; ++ii) { 1579566063dSJacob Faibussowitsch PetscCall(MatMult(AA,yysplit[ii],zzsplit[ii])); 158c4762a1bSJed Brown } 159c4762a1bSJed Brown } 160c4762a1bSJed Brown /* Unsplit yy */ 1619566063dSJacob Faibussowitsch PetscCall(VecStrideScatterAll(yysplit, yy, INSERT_VALUES)); /*YES! 'Scatter' means 'collect' (or maybe 'gather'?)! */ 162c4762a1bSJed Brown /* Unsplit zz */ 1639566063dSJacob Faibussowitsch PetscCall(VecStrideScatterAll(zzsplit, zz, INSERT_VALUES)); /*YES! 'Scatter' means 'collect' (or maybe 'gather'?)! */ 164c4762a1bSJed Brown 165c4762a1bSJed Brown if (view_y) { 1669566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "y = \n")); 1679566063dSJacob Faibussowitsch PetscCall(VecView(y, PETSC_VIEWER_STDOUT_WORLD)); 1689566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "yy = \n")); 1699566063dSJacob Faibussowitsch PetscCall(VecView(yy, PETSC_VIEWER_STDOUT_WORLD)); 170c4762a1bSJed Brown } 171c4762a1bSJed Brown 172c4762a1bSJed Brown if (view_z) { 1739566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "z = \n")); 1749566063dSJacob Faibussowitsch PetscCall(VecView(z, PETSC_VIEWER_STDOUT_WORLD)); 1759566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_WORLD, "zz = \n")); 1769566063dSJacob Faibussowitsch PetscCall(VecView(zz, PETSC_VIEWER_STDOUT_WORLD)); 177c4762a1bSJed Brown } 178c4762a1bSJed Brown 179c4762a1bSJed Brown /* compare x and z. USFFT computes an unnormalized DFT, thus z = N*x */ 180c4762a1bSJed Brown s = 1.0/(PetscReal)N; 1819566063dSJacob Faibussowitsch PetscCall(VecScale(z,s)); 1829566063dSJacob Faibussowitsch PetscCall(VecAXPY(x,-1.0,z)); 1839566063dSJacob Faibussowitsch PetscCall(VecNorm(x,NORM_1,&enorm)); 1849566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "|x-z| = %g\n",enorm)); 185c4762a1bSJed Brown 186c4762a1bSJed Brown /* compare xx and zz. FFTW computes an unnormalized DFT, thus zz = N*x */ 187c4762a1bSJed Brown s = 1.0/(PetscReal)N; 1889566063dSJacob Faibussowitsch PetscCall(VecScale(zz,s)); 1899566063dSJacob Faibussowitsch PetscCall(VecAXPY(xx,-1.0,zz)); 1909566063dSJacob Faibussowitsch PetscCall(VecNorm(xx,NORM_1,&enorm)); 1919566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "|xx-zz| = %g\n",enorm)); 192c4762a1bSJed Brown 193c4762a1bSJed Brown /* compare y and yy: USFFT and FFTW results*/ 1949566063dSJacob Faibussowitsch PetscCall(VecNorm(y,NORM_2,&norm)); 1959566063dSJacob Faibussowitsch PetscCall(VecAXPY(y,-1.0,yy)); 1969566063dSJacob Faibussowitsch PetscCall(VecNorm(y,NORM_1,&enorm)); 1979566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "|y|_2 = %g\n",norm)); 1989566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "|y-yy| = %g\n",enorm)); 199c4762a1bSJed Brown 200c4762a1bSJed Brown /* compare z and zz: USFFT and FFTW results*/ 2019566063dSJacob Faibussowitsch PetscCall(VecNorm(z,NORM_2,&norm)); 2029566063dSJacob Faibussowitsch PetscCall(VecAXPY(z,-1.0,zz)); 2039566063dSJacob Faibussowitsch PetscCall(VecNorm(z,NORM_1,&enorm)); 2049566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "|z|_2 = %g\n",norm)); 2059566063dSJacob Faibussowitsch PetscCall(PetscPrintf(PETSC_COMM_SELF, "|z-zz| = %g\n",enorm)); 206c4762a1bSJed Brown 207c4762a1bSJed Brown /* free spaces */ 2089566063dSJacob Faibussowitsch PetscCall(DMRestoreGlobalVector(da,&x)); 2099566063dSJacob Faibussowitsch PetscCall(DMRestoreGlobalVector(da,&xx)); 2109566063dSJacob Faibussowitsch PetscCall(DMRestoreGlobalVector(da,&y)); 2119566063dSJacob Faibussowitsch PetscCall(DMRestoreGlobalVector(da,&yy)); 2129566063dSJacob Faibussowitsch PetscCall(DMRestoreGlobalVector(da,&z)); 2139566063dSJacob Faibussowitsch PetscCall(DMRestoreGlobalVector(da,&zz)); 214c4762a1bSJed Brown 2159566063dSJacob Faibussowitsch PetscCall(PetscFinalize()); 216b122ec5aSJacob Faibussowitsch return 0; 217c4762a1bSJed Brown } 218