void SymmetricEuclideanDistanceMatrix(El::UpperOrLower uplo, direction_t dir,
T alpha, const El::ElementalMatrix<T> &A,
T beta, El::ElementalMatrix<T> &C) {
T *c = C.Buffer();
int ldC = C.LDim();
if (dir == base::COLUMNS) {
El::Herk(uplo, El::ADJOINT, -2.0 * alpha, A, beta, C);
//El::Gemm(El::ADJOINT, El::NORMAL, T(-2.0) * alpha, A, A, beta, C);
El::DistMatrix<El::Base<T>, El::STAR, El::STAR > N;
ColumnNrm2(A, N);
El::Base<T> *nn = N.Buffer();;
El::Int n = C.LocalWidth();
El::Int m = C.LocalHeight();
for(El::Int j = 0; j < n; j++)
for(El::Int i =
((uplo == El::UPPER) ? 0 : C.LocalRowOffset(A.GlobalCol(j)));
i < ((uplo == El::UPPER) ? C.LocalRowOffset(A.GlobalCol(j) + 1) : m); i++) {
El::Base<T> a = nn[C.GlobalRow(i)];
El::Base<T> b = nn[C.GlobalCol(j)];
c[j * ldC + i] += alpha * (a * a + b * b);
}
}
// TODO the rest of the cases.
}
/*
* Convert a std::vector to an elemental STAR, STAR distributed vector
*/
int vec2elstar(const std::vector<double> &vec, El::DistMatrix<El::Complex<double>,El::STAR,El::STAR > &Y){
El::Complex<double> *Y_ptr = Y.Buffer();
int sz = vec.size()/2;
#pragma omp parallel for
for(int i=0;i<sz; i++){
double real = vec[2*i];
double imag = vec[2*i+1];
El::SetRealPart(Y_ptr[i],real);
El::SetImagPart(Y_ptr[i],imag);
}
return 0;
}
inline void Gemv(El::Orientation oA,
T alpha, const El::DistMatrix<T, El::VC, El::STAR>& A,
const El::DistMatrix<T, El::VC, El::STAR>& x,
T beta, El::DistMatrix<T, El::STAR, El::STAR>& y) {
// TODO verify sizes etc.
// TODO verify matching grids.
if (oA == El::TRANSPOSE) {
boost::mpi::communicator comm(y.Grid().Comm().comm,
boost::mpi::comm_attach);
El::Matrix<T> ylocal(y.Matrix());
El::Gemv(El::TRANSPOSE,
alpha, A.LockedMatrix(), x.LockedMatrix(),
beta / T(comm.size()), ylocal);
boost::mpi::all_reduce(comm,
ylocal.Buffer(), ylocal.MemorySize(), y.Buffer(),
std::plus<T>());
} else {
SKYLARK_THROW_EXCEPTION(base::unsupported_base_operation());
}
}
void L1DistanceMatrixTU(El::UpperOrLower uplo,
direction_t dirA, direction_t dirB, T alpha,
const El::DistMatrix<T, El::STAR, El::MC> &A,
const El::DistMatrix<T, El::STAR, El::MR> &B,
T beta, El::DistMatrix<T> &C) {
// TODO verify sizes
const T *a = A.LockedBuffer();
El::Int ldA = A.LDim();
const T *b = B.LockedBuffer();
El::Int ldB = B.LDim();
T *c = C.Buffer();
El::Int ldC = C.LDim();
El::Int d = A.Height();
/* Not the most efficient way... but mimicking BLAS is too much work! */
if (dirA == base::COLUMNS && dirB == base::COLUMNS) {
El::Int n = C.LocalWidth();
El::Int m = C.LocalHeight();
for (El::Int j = 0; j < n; j++)
for(El::Int i =
((uplo == El::UPPER) ? 0 : C.LocalRowOffset(A.GlobalCol(j)));
i < ((uplo == El::UPPER) ? C.LocalRowOffset(A.GlobalCol(j) + 1) : m); i++) {
T v = 0.0;
for (El::Int k = 0; k < d; k++)
v += std::abs(b[j * ldB + k] - a[i * ldA + k]);
c[j * ldC + i] = beta * c[j * ldC + i] + alpha * v;
}
}
// TODO the rest of the cases.
}
inline void outer_panel_mixed_gemm_impl_tn(
const double alpha,
const SpParMat<index_type, value_type, SpDCCols<index_type, value_type> > &A,
const El::DistMatrix<value_type, col_d, El::STAR> &S,
const double beta,
El::DistMatrix<value_type, El::STAR, El::STAR> &C) {
El::DistMatrix<value_type, El::STAR, El::STAR>
tmp_C(C.Height(), C.Width());
El::Zero(tmp_C);
utility::combblas_slab_view_t<index_type, value_type> cbview(A, false);
//FIXME: factor
size_t slab_size = 2 * S.Grid().Height();
for(size_t cur_row_idx = 0; cur_row_idx < cbview.ncols();
cur_row_idx += slab_size) {
size_t cur_slab_size =
std::min(slab_size, cbview.ncols() - cur_row_idx);
// get the next slab_size columns of B
El::DistMatrix<value_type, El::STAR, El::STAR>
A_row(cur_slab_size, S.Height());
// transpose is column
//cbview.extract_elemental_column_slab_view(A_row, cur_slab_size);
cbview.extract_full_slab_view(cur_slab_size);
// matrix mult (FIXME only iter nz)
for(size_t l_row_idx = 0; l_row_idx < A_row.LocalHeight();
++l_row_idx) {
size_t g_row_idx = l_row_idx * A_row.ColStride()
+ A_row.ColShift() + cur_row_idx;
for(size_t l_col_idx = 0; l_col_idx < A_row.LocalWidth();
l_col_idx++) {
//XXX: should be the same as l_col_idx
size_t g_col_idx = l_col_idx * A_row.RowStride()
+ A_row.RowShift();
// continue if we don't own values in S in this row
if(!S.IsLocalRow(g_col_idx))
continue;
//get transposed value
value_type val = alpha * cbview(g_col_idx, g_row_idx);
for(size_t s_col_idx = 0; s_col_idx < S.LocalWidth();
s_col_idx++) {
tmp_C.UpdateLocal(g_row_idx, s_col_idx,
val * S.GetLocal(S.LocalRow(g_col_idx), s_col_idx));
}
}
}
}
//FIXME: scaling
if(A.getcommgrid()->GetRank() == 0) {
for(size_t col_idx = 0; col_idx < C.Width(); col_idx++)
for(size_t row_idx = 0; row_idx < C.Height(); row_idx++)
tmp_C.UpdateLocal(row_idx, col_idx,
beta * C.GetLocal(row_idx, col_idx));
}
//FIXME: Use utility getter
boost::mpi::communicator world(
A.getcommgrid()->GetWorld(), boost::mpi::comm_duplicate);
boost::mpi::all_reduce (world,
tmp_C.LockedBuffer(),
C.Height() * C.Width(),
C.Buffer(),
std::plus<value_type>());
}
int vec2elemental(const std::vector<double> &vec, El::DistMatrix<El::Complex<double>,El::VC,El::STAR> &Y){
int data_dof=2;
int SCAL_EXP = 1;
int nlocal,gsize; //local elements, start p_id, global size
double *pt_array; // will hold local array
int r,q,rq; //Grid sizes
int nbigs; //Number of large sends (i.e. send 1 extra data point)
int pstart; // p_id of nstart
int rank = El::mpi::WorldRank(); //p_id
int send_size; // base send size
bool print = rank == -1;
// Get Grid and associated params
const El::Grid* g = &(Y.Grid());
r = g->Height();
q = g->Width();
MPI_Comm comm = (g->Comm()).comm;
// Get sizes, array in petsc
nlocal = vec.size()/data_dof;
int nstart = 0;
MPI_Exscan(&nlocal,&nstart,1,MPI_INT,MPI_SUM,comm);
//VecGetOwnershipRange(pt_vec,&nstart,NULL);
//Find processor that nstart belongs to, number of larger sends
rq = r * q;
pstart = nstart % rq; //int div
nbigs = nlocal % rq;
send_size = nlocal/rq;
if(print){
std::cout << "r: " << r << " q: " << q <<std::endl;
std::cout << "nstart: " << nstart << std::endl;
std::cout << "ps: " << pstart << std::endl;
std::cout << "nbigs: " << nbigs << std::endl;
std::cout << "send_size: " << send_size << std::endl;
}
// Make send_lengths
std::vector<int> send_lengths(rq);
std::fill(send_lengths.begin(),send_lengths.end(),send_size);
if(nbigs >0){
for(int j=0;j<nbigs;j++){
send_lengths[(pstart + j) % rq] += 1;
}
}
// Make send_disps
std::vector<int> send_disps = exscan(send_lengths);
std::vector<El::Complex<double>> indata(nlocal);
// copy the data from an ffm tree to into a local vec of complex data for sending #pragma omp parallel for
El::Complex<double> val;
for(int i=0;i<nlocal;i++){
El::SetRealPart(val,vec[2*i+0]);
El::SetImagPart(val,vec[2*i+1]);
indata[i] = val;
}
// Make send_dataA, i.e. reorder the data
std::vector<El::Complex<double>> send_data(nlocal);
for(int proc=0;proc<rq;proc++){
int offset = send_disps[proc];
int base_idx = (proc - pstart + rq) % rq;
for(int j=0; j<send_lengths[proc]; j++){
int idx = base_idx + (j * rq);
send_data[offset + j] = indata[idx];
}
}
// Do all2all to get recv_lengths
std::vector<int> recv_lengths(rq);
MPI_Alltoall(&send_lengths[0], 1, MPI_INT, &recv_lengths[0], 1, MPI_INT,comm);
// Scan to get recv_disps
std::vector<int> recv_disps = exscan(recv_lengths);
// Do all2allv to get data on correct processor
El::Complex<double> * recv_data = Y.Buffer();
//MPI_Alltoallv(&send_data[0],&send_lengths[0],&send_disps[0],MPI_DOUBLE, \
// &recv_data[0],&recv_lengths[0],&recv_disps[0],MPI_DOUBLE,comm);
El::mpi::AllToAll(&send_data[0], &send_lengths[0], &send_disps[0], recv_data,&recv_lengths[0],&recv_disps[0],comm);
if(print){
std::cout << "Send data: " <<std::endl << send_data <<std::endl;
std::cout << "Send lengths: " <<std::endl << send_lengths <<std::endl;
std::cout << "Send disps: " <<std::endl << send_disps <<std::endl;
std::cout << "Recv data: " <<std::endl << recv_data <<std::endl;
std::cout << "Recv lengths: " <<std::endl << recv_lengths <<std::endl;
std::cout << "Recv disps: " <<std::endl << recv_disps <<std::endl;
}
return 0;
}
int tree2elemental(InvMedTree<FMM_Mat_t> *tree, El::DistMatrix<T,El::VC,El::STAR> &Y){
int data_dof=2;
int SCAL_EXP = 1;
int nlocal,gsize; //local elements, start p_id, global size
double *pt_array; // will hold local array
int r,q,rq; //Grid sizes
int nbigs; //Number of large sends (i.e. send 1 extra data point)
int pstart; // p_id of nstart
int rank = El::mpi::WorldRank(); //p_id
int send_size; // base send size
bool print = rank == -1;
// Get Grid and associated params
const El::Grid* g = &(Y.Grid());
r = g->Height();
q = g->Width();
MPI_Comm comm = (g->Comm()).comm;
std::vector<FMMNode_t*> nlist = tree->GetNGLNodes();
int cheb_deg = InvMedTree<FMM_Mat_t>::cheb_deg;
int omp_p=omp_get_max_threads();
size_t n_coeff3=(cheb_deg+1)*(cheb_deg+2)*(cheb_deg+3)/6;
// Get sizes, array in petsc
//VecGetSize(pt_vec,&gsize);
gsize = tree->M/data_dof;
nlocal = tree->m/data_dof;
//VecGetLocalSize(pt_vec,&nlocal);
//VecGetArray(pt_vec,&pt_array);
int nstart = 0;
MPI_Exscan(&nlocal,&nstart,1,MPI_INT,MPI_SUM,comm);
//VecGetOwnershipRange(pt_vec,&nstart,NULL);
//Find processor that nstart belongs to, number of larger sends
rq = r * q;
pstart = nstart % rq; //int div
nbigs = nlocal % rq;
send_size = nlocal/rq;
if(print){
std::cout << "r: " << r << " q: " << q <<std::endl;
std::cout << "nstart: " << nstart << std::endl;
std::cout << "ps: " << pstart << std::endl;
std::cout << "nbigs: " << nbigs << std::endl;
std::cout << "send_size: " << send_size << std::endl;
}
// Make send_lengths
std::vector<int> send_lengths(rq);
std::fill(send_lengths.begin(),send_lengths.end(),send_size);
if(nbigs >0){
for(int j=0;j<nbigs;j++){
send_lengths[(pstart + j) % rq] += 1;
}
}
// Make send_disps
std::vector<int> send_disps = exscan(send_lengths);
std::vector<El::Complex<double>> indata(nlocal);
// copy the data from an ffm tree to into a local vec of complex data for sending #pragma omp parallel for
for(size_t tid=0;tid<omp_p;tid++){
size_t i_start=(nlist.size()* tid )/omp_p;
size_t i_end =(nlist.size()*(tid+1))/omp_p;
for(size_t i=i_start;i<i_end;i++){
pvfmm::Vector<double>& coeff_vec=nlist[i]->ChebData();
double s=std::pow(0.5,COORD_DIM*nlist[i]->Depth()*0.5*SCAL_EXP);
size_t offset=i*n_coeff3;
for(size_t j=0;j<n_coeff3;j++){
double real = coeff_vec[j]*s; // local indices as in the pvfmm trees
double imag = coeff_vec[j+n_coeff3]*s;
El::Complex<double> coeff;
El::SetRealPart(coeff,real);
El::SetImagPart(coeff,imag);
indata[offset+j] = coeff;
}
}
}
// Make send_data
std::vector<El::Complex<double>> send_data(nlocal);
for(int proc=0;proc<rq;proc++){
int offset = send_disps[proc];
int base_idx = (proc - pstart + rq) % rq;
for(int j=0; j<send_lengths[proc]; j++){
int idx = base_idx + (j * rq);
send_data[offset + j] = indata[idx];
}
}
// Do all2all to get recv_lengths
std::vector<int> recv_lengths(rq);
MPI_Alltoall(&send_lengths[0], 1, MPI_INT, &recv_lengths[0], 1, MPI_INT,comm);
// Scan to get recv_disps
std::vector<int> recv_disps = exscan(recv_lengths);
// Do all2allv to get data on correct processor
El::Complex<double> * recv_data = Y.Buffer();
//MPI_Alltoallv(&send_data[0],&send_lengths[0],&send_disps[0],MPI_DOUBLE, \
// &recv_data[0],&recv_lengths[0],&recv_disps[0],MPI_DOUBLE,comm);
El::mpi::AllToAll(&send_data[0], &send_lengths[0], &send_disps[0], recv_data,&recv_lengths[0],&recv_disps[0],comm);
if(print){
std::cout << "Send data: " <<std::endl << send_data <<std::endl;
std::cout << "Send lengths: " <<std::endl << send_lengths <<std::endl;
std::cout << "Send disps: " <<std::endl << send_disps <<std::endl;
std::cout << "Recv data: " <<std::endl << recv_data <<std::endl;
std::cout << "Recv lengths: " <<std::endl << recv_lengths <<std::endl;
std::cout << "Recv disps: " <<std::endl << recv_disps <<std::endl;
}
return 0;
}
void El2Petsc_vec(El::DistMatrix<double,VC,STAR>& el_vec,Vec& pt_vec){
PetscInt nlocal, nstart; // petsc vec info
PetscScalar *pt_array,*pt_perm_array;
int r,q,ll,rq; // el vec info
int nbigs; //Number of large recv (i.e. recv 1 extra data point)
int pstart; // p_id of nstart
int p = El::mpi::WorldRank(); //p_id
int recv_size; // base recv size
bool print = p == -1;
// Get el vec info
ll = el_vec.Height();
const El::Grid* g = &(el_vec.Grid());
r = g->Height();
q = g->Width();
MPI_Comm comm = (g->Comm()).comm;
// Get petsc vec params
VecGetLocalSize(pt_vec,&nlocal);
VecGetArray(pt_vec,&pt_array);
VecGetOwnershipRange(pt_vec,&nstart,NULL);
// Determine who owns the first element we want
rq = r * q;
pstart = nstart % rq;
nbigs = nlocal % rq;
recv_size = nlocal / rq;
if(print){
std::cout << "r: " << r << " q: " << q <<std::endl;
std::cout << "nstart: " << nstart << std::endl;
std::cout << "ps: " << pstart << std::endl;
std::cout << "nbigs: " << nbigs << std::endl;
std::cout << "recv_size: " << recv_size << std::endl;
}
// Make recv sizes
std::vector<int> recv_lengths(rq);
std::fill(recv_lengths.begin(),recv_lengths.end(),recv_size);
if(nbigs >0){
for(int i=0;i<nbigs;i++){
recv_lengths[(pstart + i) % rq] += 1;
}
}
// Make recv disps
std::vector<int> recv_disps = exscan(recv_lengths);
// All2all to get send sizes
std::vector<int> send_lengths(rq);
MPI_Alltoall(&recv_lengths[0], 1, MPI_INT, &send_lengths[0], 1, MPI_INT,comm);
// Scan to get send_disps
std::vector<int> send_disps = exscan(send_lengths);
// Do all2allv to get data on correct processor
std::vector<double> recv_data(nlocal);
MPI_Alltoallv(el_vec.Buffer(),&send_lengths[0],&send_disps[0],MPI_DOUBLE, \
&recv_data[0],&recv_lengths[0],&recv_disps[0],MPI_DOUBLE,comm);
if(print){
//std::cout << "Send data: " <<std::endl << *el_vec.Buffer() <<std::endl;
std::cout << "Send lengths: " <<std::endl << send_lengths <<std::endl;
std::cout << "Send disps: " <<std::endl << send_disps <<std::endl;
std::cout << "Recv data: " <<std::endl << recv_data <<std::endl;
std::cout << "Recv lengths: " <<std::endl << recv_lengths <<std::endl;
std::cout << "Recv disps: " <<std::endl << recv_disps <<std::endl;
}
// Reorder for petsc
for(int p=0;p<rq;p++){
int base_idx = (p - pstart + rq) % rq;
int offset = recv_disps[p];
for(int i=0;i<recv_lengths[p];i++){
pt_array[base_idx + rq*i] = recv_data[offset + i];
}
}
// Copy into array
VecRestoreArray(pt_vec,&pt_array);
}
void Petsc2El_vec(Vec& pt_vec,El::DistMatrix<double,VC,STAR>& el_vec){
PetscInt nlocal,nstart,gsize; //local elements, start p_id, global size
PetscScalar *pt_array; // will hold local array
int r,q,rq; //Grid sizes
int nbigs; //Number of large sends (i.e. send 1 extra data point)
int pstart; // p_id of nstart
int p = El::mpi::WorldRank(); //p_id
int send_size; // base send size
bool print = p == -1;
// Get Grid and associated params
const El::Grid* g = &(el_vec.Grid());
r = g->Height();
q = g->Width();
MPI_Comm comm = (g->Comm()).comm;
// Get sizes, array in petsc
VecGetSize(pt_vec,&gsize);
VecGetLocalSize(pt_vec,&nlocal);
VecGetArray(pt_vec,&pt_array);
VecGetOwnershipRange(pt_vec,&nstart,NULL);
//Find processor that nstart belongs to, number of larger sends
rq = r * q;
pstart = nstart % rq; //int div
nbigs = nlocal % rq;
send_size = nlocal/rq;
if(print){
std::cout << "r: " << r << " q: " << q <<std::endl;
std::cout << "nstart: " << nstart << std::endl;
std::cout << "ps: " << pstart << std::endl;
std::cout << "nbigs: " << nbigs << std::endl;
std::cout << "send_size: " << send_size << std::endl;
}
// Make send_lengths
std::vector<int> send_lengths(rq);
std::fill(send_lengths.begin(),send_lengths.end(),send_size);
if(nbigs >0){
for(int j=0;j<nbigs;j++){
send_lengths[(pstart + j) % rq] += 1;
}
}
// Make send_disps
std::vector<int> send_disps = exscan(send_lengths);
// Make send_data
std::vector<double> send_data(nlocal);
for(int proc=0;proc<rq;proc++){
int offset = send_disps[proc];
int base_idx = (proc - pstart + rq) % rq;
for(int j=0; j<send_lengths[proc]; j++){
int idx = base_idx + (j * rq);
send_data[offset + j] = pt_array[idx];
}
}
// Do all2all to get recv_lengths
std::vector<int> recv_lengths(rq);
MPI_Alltoall(&send_lengths[0], 1, MPI_INT, &recv_lengths[0], 1, MPI_INT,comm);
// Scan to get recv_disps
std::vector<int> recv_disps = exscan(recv_lengths);
// Do all2allv to get data on correct processor
double * recv_data = el_vec.Buffer();
MPI_Alltoallv(&send_data[0],&send_lengths[0],&send_disps[0],MPI_DOUBLE, \
&recv_data[0],&recv_lengths[0],&recv_disps[0],MPI_DOUBLE,comm);
if(print){
std::cout << "Send data: " <<std::endl << send_data <<std::endl;
std::cout << "Send lengths: " <<std::endl << send_lengths <<std::endl;
std::cout << "Send disps: " <<std::endl << send_disps <<std::endl;
std::cout << "Recv data: " <<std::endl << recv_data <<std::endl;
std::cout << "Recv lengths: " <<std::endl << recv_lengths <<std::endl;
std::cout << "Recv disps: " <<std::endl << recv_disps <<std::endl;
}
}