1c4762a1bSJed Brown /* This program illustrates use of paralllel real FFT*/ 2c4762a1bSJed Brown static char help[]="This program illustrates the use of parallel real 3D fftw (without PETSc interface)"; 3c4762a1bSJed Brown #include <petscmat.h> 4c4762a1bSJed Brown #include <fftw3.h> 5c4762a1bSJed Brown #include <fftw3-mpi.h> 6c4762a1bSJed Brown 7ad0e0ea3SStefano Zampini int main(int argc,char **args) 8c4762a1bSJed Brown { 9c4762a1bSJed Brown ptrdiff_t N0=256,N1=256,N2=256,N3=2,dim[4]; 10c4762a1bSJed Brown fftw_plan bplan,fplan; 11c4762a1bSJed Brown fftw_complex *out; 12c4762a1bSJed Brown double *in1,*in2; 13c4762a1bSJed Brown ptrdiff_t alloc_local,local_n0,local_0_start; 14c4762a1bSJed Brown ptrdiff_t local_n1,local_1_start; 15c4762a1bSJed Brown PetscInt i,j,indx,n1; 16c4762a1bSJed Brown PetscInt size,rank,n,N,*in,N_factor,NM; 17c4762a1bSJed Brown PetscScalar *data_fin,value1,one=1.57,zero=0.0; 18c4762a1bSJed Brown PetscScalar a,*x_arr,*y_arr,*z_arr,enorm; 19c4762a1bSJed Brown Vec fin,fout,fout1,ini,final; 20c4762a1bSJed Brown PetscRandom rnd; 21c4762a1bSJed Brown PetscErrorCode ierr; 22c4762a1bSJed Brown VecScatter vecscat,vecscat1; 23c4762a1bSJed Brown IS indx1,indx2; 24c4762a1bSJed Brown PetscInt *indx3,k,l,*indx4; 25c4762a1bSJed Brown PetscInt low,tempindx,tempindx1; 26c4762a1bSJed Brown 27c4762a1bSJed Brown ierr = PetscInitialize(&argc,&args,(char*)0,help);if (ierr) return ierr; 28c4762a1bSJed Brown #if defined(PETSC_USE_COMPLEX) 29c4762a1bSJed Brown SETERRQ(PETSC_COMM_WORLD,PETSC_ERR_SUP, "This example requires real numbers. Your current scalar type is complex"); 30c4762a1bSJed Brown #endif 31*5f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Comm_size(PETSC_COMM_WORLD, &size)); 32*5f80ce2aSJacob Faibussowitsch CHKERRMPI(MPI_Comm_rank(PETSC_COMM_WORLD, &rank)); 33c4762a1bSJed Brown 34c4762a1bSJed Brown PetscRandomCreate(PETSC_COMM_WORLD,&rnd); 35c4762a1bSJed Brown 36c4762a1bSJed Brown alloc_local = fftw_mpi_local_size_3d_transposed(N0,N1,N2/2+1,PETSC_COMM_WORLD,&local_n0,&local_0_start,&local_n1,&local_1_start); 37c4762a1bSJed Brown 38c4762a1bSJed Brown /* printf("The value alloc_local is %ld from process %d\n",alloc_local,rank); */ 39c4762a1bSJed Brown printf("The value local_n0 is %ld from process %d\n",local_n0,rank); 40c4762a1bSJed Brown /* printf("The value local_0_start is %ld from process %d\n",local_0_start,rank);*/ 41c4762a1bSJed Brown /* printf("The value local_n1 is %ld from process %d\n",local_n1,rank); */ 42c4762a1bSJed Brown /* printf("The value local_1_start is %ld from process %d\n",local_1_start,rank);*/ 43c4762a1bSJed Brown 44c4762a1bSJed Brown /* Allocate space for input and output arrays */ 45c4762a1bSJed Brown 46c4762a1bSJed Brown in1=(double*)fftw_malloc(sizeof(double)*alloc_local*2); 47c4762a1bSJed Brown in2=(double*)fftw_malloc(sizeof(double)*alloc_local*2); 48c4762a1bSJed Brown out=(fftw_complex*)fftw_malloc(sizeof(fftw_complex)*alloc_local); 49c4762a1bSJed Brown 50c4762a1bSJed Brown N=2*N0*N1*(N2/2+1);N_factor=N0*N1*N2; 51c4762a1bSJed Brown n=2*local_n0*N1*(N2/2+1);n1=local_n1*N0*2*N1; 52c4762a1bSJed Brown 53c4762a1bSJed Brown /* printf("The value N is %d from process %d\n",N,rank); */ 54c4762a1bSJed Brown /* printf("The value n is %d from process %d\n",n,rank); */ 55c4762a1bSJed Brown /* printf("The value n1 is %d from process %d\n",n1,rank); */ 56c4762a1bSJed Brown /* Creating data vector and accompanying array with VeccreateMPIWithArray */ 57*5f80ce2aSJacob Faibussowitsch CHKERRQ(VecCreateMPIWithArray(PETSC_COMM_WORLD,1,n,N,(PetscScalar*)in1,&fin)); 58*5f80ce2aSJacob Faibussowitsch CHKERRQ(VecCreateMPIWithArray(PETSC_COMM_WORLD,1,n,N,(PetscScalar*)out,&fout)); 59*5f80ce2aSJacob Faibussowitsch CHKERRQ(VecCreateMPIWithArray(PETSC_COMM_WORLD,1,n,N,(PetscScalar*)in2,&fout1)); 60c4762a1bSJed Brown 61c4762a1bSJed Brown /* VecGetSize(fin,&size); */ 62c4762a1bSJed Brown /* printf("The size is %d\n",size); */ 63c4762a1bSJed Brown 64c4762a1bSJed Brown VecSet(fin,one); 65c4762a1bSJed Brown VecSet(fout,zero); 66c4762a1bSJed Brown VecSet(fout1,zero); 67c4762a1bSJed Brown 68c4762a1bSJed Brown VecAssemblyBegin(fin); 69c4762a1bSJed Brown VecAssemblyEnd(fin); 70c4762a1bSJed Brown /* VecView(fin,PETSC_VIEWER_STDOUT_WORLD); */ 71c4762a1bSJed Brown 72c4762a1bSJed Brown VecGetArray(fin,&x_arr); 73c4762a1bSJed Brown VecGetArray(fout1,&z_arr); 74c4762a1bSJed Brown VecGetArray(fout,&y_arr); 75c4762a1bSJed Brown 76c4762a1bSJed Brown fplan=fftw_mpi_plan_dft_r2c_3d(N0,N1,N2,(double*)x_arr,(fftw_complex*)y_arr,PETSC_COMM_WORLD,FFTW_ESTIMATE); 77c4762a1bSJed Brown bplan=fftw_mpi_plan_dft_c2r_3d(N0,N1,N2,(fftw_complex*)y_arr,(double*)z_arr,PETSC_COMM_WORLD,FFTW_ESTIMATE); 78c4762a1bSJed Brown 79c4762a1bSJed Brown fftw_execute(fplan); 80c4762a1bSJed Brown fftw_execute(bplan); 81c4762a1bSJed Brown 82c4762a1bSJed Brown VecRestoreArray(fin,&x_arr); 83c4762a1bSJed Brown VecRestoreArray(fout1,&z_arr); 84c4762a1bSJed Brown VecRestoreArray(fout,&y_arr); 85c4762a1bSJed Brown 86c4762a1bSJed Brown /* a = 1.0/(PetscReal)N_factor; */ 87*5f80ce2aSJacob Faibussowitsch /* CHKERRQ(VecScale(fout1,a)); */ 88c4762a1bSJed Brown VecCreate(PETSC_COMM_WORLD,&ini); 89c4762a1bSJed Brown VecCreate(PETSC_COMM_WORLD,&final); 90c4762a1bSJed Brown VecSetSizes(ini,local_n0*N1*N2,N_factor); 91c4762a1bSJed Brown VecSetSizes(final,local_n0*N1*N2,N_factor); 92c4762a1bSJed Brown /* VecSetSizes(ini,PETSC_DECIDE,N_factor); */ 93c4762a1bSJed Brown /* VecSetSizes(final,PETSC_DECIDE,N_factor); */ 94c4762a1bSJed Brown VecSetFromOptions(ini); 95c4762a1bSJed Brown VecSetFromOptions(final); 96c4762a1bSJed Brown 97c4762a1bSJed Brown if (N2%2==0) NM=N2+2; 98c4762a1bSJed Brown else NM=N2+1; 99c4762a1bSJed Brown 100*5f80ce2aSJacob Faibussowitsch CHKERRQ(VecGetOwnershipRange(fin,&low,NULL)); 101c4762a1bSJed Brown printf("The local index is %d from %d\n",low,rank); 102*5f80ce2aSJacob Faibussowitsch CHKERRQ(PetscMalloc1(local_n0*N1*N2,&indx3)); 103*5f80ce2aSJacob Faibussowitsch CHKERRQ(PetscMalloc1(local_n0*N1*N2,&indx4)); 104c4762a1bSJed Brown for (i=0; i<local_n0; i++) { 105c4762a1bSJed Brown for (j=0;j<N1;j++) { 106c4762a1bSJed Brown for (k=0;k<N2;k++) { 107c4762a1bSJed Brown tempindx = i*N1*N2 + j*N2 + k; 108c4762a1bSJed Brown tempindx1 = i*N1*NM + j*NM + k; 109c4762a1bSJed Brown 110c4762a1bSJed Brown indx3[tempindx]=local_0_start*N1*N2+tempindx; 111c4762a1bSJed Brown indx4[tempindx]=low+tempindx1; 112c4762a1bSJed Brown } 113c4762a1bSJed Brown /* printf("index3 %d from proc %d is \n",indx3[tempindx],rank); */ 114c4762a1bSJed Brown /* printf("index4 %d from proc %d is \n",indx4[tempindx],rank); */ 115c4762a1bSJed Brown } 116c4762a1bSJed Brown } 117c4762a1bSJed Brown VecGetValues(fin,local_n0*N1*N2,indx4,x_arr); 118c4762a1bSJed Brown VecSetValues(ini,local_n0*N1*N2,indx3,x_arr,INSERT_VALUES); 119c4762a1bSJed Brown VecAssemblyBegin(ini); 120c4762a1bSJed Brown VecAssemblyEnd(ini); 121c4762a1bSJed Brown 122c4762a1bSJed Brown VecGetValues(fout1,local_n0*N1*N2,indx4,y_arr); 123c4762a1bSJed Brown VecSetValues(final,local_n0*N1*N2,indx3,y_arr,INSERT_VALUES); 124c4762a1bSJed Brown VecAssemblyBegin(final); 125c4762a1bSJed Brown VecAssemblyEnd(final); 126c4762a1bSJed Brown 127c4762a1bSJed Brown printf("The local index value is %ld from %d",local_n0*N1*N2,rank); 128c4762a1bSJed Brown /* 129c4762a1bSJed Brown for (i=0;i<N0;i++) { 130c4762a1bSJed Brown for (j=0;j<N1;j++) { 131c4762a1bSJed Brown indx=i*N1*NM+j*NM; 132c4762a1bSJed Brown ISCreateStride(PETSC_COMM_WORLD,N2,indx,1,&indx1); 133c4762a1bSJed Brown indx=i*N1*N2+j*N2; 134c4762a1bSJed Brown ISCreateStride(PETSC_COMM_WORLD,N2,indx,1,&indx2); 135c4762a1bSJed Brown VecScatterCreate(fin,indx1,ini,indx2,&vecscat); 136c4762a1bSJed Brown VecScatterBegin(vecscat,fin,ini,INSERT_VALUES,SCATTER_FORWARD); 137c4762a1bSJed Brown VecScatterEnd(vecscat,fin,ini,INSERT_VALUES,SCATTER_FORWARD); 138c4762a1bSJed Brown VecScatterCreate(fout1,indx1,final,indx2,&vecscat1); 139c4762a1bSJed Brown VecScatterBegin(vecscat1,fout1,final,INSERT_VALUES,SCATTER_FORWARD); 140c4762a1bSJed Brown VecScatterEnd(vecscat1,fout1,final,INSERT_VALUES,SCATTER_FORWARD); 141c4762a1bSJed Brown } 142c4762a1bSJed Brown } 143c4762a1bSJed Brown */ 144c4762a1bSJed Brown a = 1.0/(PetscReal)N_factor; 145*5f80ce2aSJacob Faibussowitsch CHKERRQ(VecScale(fout1,a)); 146*5f80ce2aSJacob Faibussowitsch CHKERRQ(VecScale(final,a)); 147c4762a1bSJed Brown 148c4762a1bSJed Brown VecAssemblyBegin(ini); 149c4762a1bSJed Brown VecAssemblyEnd(ini); 150c4762a1bSJed Brown 151c4762a1bSJed Brown VecAssemblyBegin(final); 152c4762a1bSJed Brown VecAssemblyEnd(final); 153c4762a1bSJed Brown 154c4762a1bSJed Brown /* VecView(final,PETSC_VIEWER_STDOUT_WORLD); */ 155*5f80ce2aSJacob Faibussowitsch CHKERRQ(VecAXPY(final,-1.0,ini)); 156*5f80ce2aSJacob Faibussowitsch CHKERRQ(VecNorm(final,NORM_1,&enorm)); 157*5f80ce2aSJacob Faibussowitsch CHKERRQ(PetscPrintf(PETSC_COMM_WORLD," Error norm of |x - z| = %e\n",enorm)); 158c4762a1bSJed Brown fftw_destroy_plan(fplan); 159c4762a1bSJed Brown fftw_destroy_plan(bplan); 160*5f80ce2aSJacob Faibussowitsch fftw_free(in1); CHKERRQ(VecDestroy(&fin)); 161*5f80ce2aSJacob Faibussowitsch fftw_free(out); CHKERRQ(VecDestroy(&fout)); 162*5f80ce2aSJacob Faibussowitsch fftw_free(in2); CHKERRQ(VecDestroy(&fout1)); 163c4762a1bSJed Brown 164c4762a1bSJed Brown ierr = PetscFinalize(); 165c4762a1bSJed Brown return ierr; 166c4762a1bSJed Brown } 167