void HostMatrixCOO<ValueType>::AllocateCOO(const int nnz, const int nrow, const int ncol) {
assert( nnz >= 0);
assert( ncol >= 0);
assert( nrow >= 0);
if (this->get_nnz() > 0)
this->Clear();
if (nnz > 0) {
allocate_host(nnz, &this->mat_.row);
allocate_host(nnz, &this->mat_.col);
allocate_host(nnz, &this->mat_.val);
set_to_zero_host(nnz, this->mat_.row);
set_to_zero_host(nnz, this->mat_.col);
set_to_zero_host(nnz, this->mat_.val);
this->nrow_ = nrow;
this->ncol_ = ncol;
this->nnz_ = nnz;
}
}
void HostMatrixELL<ValueType>::AllocateELL(const int nnz, const int nrow, const int ncol, const int max_row) {
assert( nnz >= 0);
assert( ncol >= 0);
assert( nrow >= 0);
assert( max_row >= 0);
if (this->nnz_ > 0)
this->Clear();
if (nnz > 0) {
assert(nnz == max_row * nrow);
allocate_host(nnz, &this->mat_.val);
allocate_host(nnz, &this->mat_.col);
set_to_zero_host(nnz, this->mat_.val);
set_to_zero_host(nnz, this->mat_.col);
this->mat_.max_row = max_row;
this->nrow_ = nrow;
this->ncol_ = ncol;
this->nnz_ = nnz;
}
}
void HostVector<ValueType>::Allocate(const int n) {
assert(n >= 0);
if (this->size_ >0)
this->Clear();
if (n > 0) {
allocate_host(n, &this->vec_);
set_to_zero_host(n, this->vec_);
this->size_ = n;
}
}
void HostVector<ValueType>::Assemble(const int *ii, const ValueType *v,
int size, const int n) {
assert(ii != NULL);
assert(v != NULL);
assert(size > 0);
assert(n >= 0);
_set_omp_backend_threads(this->local_backend_, this->size_);
const int nThreads = omp_get_max_threads();
int N = n;
if (N == 0) {
#pragma omp parallel for
for (int i=0; i<size; ++i) {
assert(ii[i] >= 0);
int val = ii[i]+1;
if (val > N) {
#pragma omp critical
{
N = val;
}
}
}
this->Clear();
this->Allocate(N);
}
if (nThreads <= 2) {
// serial
for (int i=0; i<size; ++i)
this->vec_[ ii[i] ] += v[i];
} else {
// parallel
ValueType **v_red;
v_red = (ValueType **) malloc(nThreads*sizeof(ValueType*));
for (int k=0; k<nThreads; ++k) {
v_red[k] = NULL;
allocate_host(N, &v_red[k]);
set_to_zero_host(N, v_red[k]);
}
#pragma omp parallel
{
const int me = omp_get_thread_num();
const int istart = size*me/nThreads;
const int iend = size*(me+1)/nThreads;
for (int i = istart; i < iend; i++)
v_red[me][ii[i]] += v[i];
}
#pragma omp parallel
{
const int me = omp_get_thread_num();
const int istart = N*me/nThreads;
const int iend = N*(me+1)/nThreads;
for (int i = istart; i < iend; ++i)
for (int k=0; k<nThreads; ++k)
this->vec_[i] += v_red[k][i];
}
for (int k=0; k<nThreads; ++k)
free_host(&v_red[k]);
free(v_red);
}
}
