static void task_axpby_work_steal( void * arg , TPI_ThreadPool pool )
{
enum { BLOCK = UNROLL * 128 };
int p_size ;
int p_rank ;
if ( ! TPI_Rank( pool , & p_rank , & p_size ) ) {
struct TaskXY * const t = (struct TaskXY *) arg ;
const double a = t->alpha ;
const double b = t->beta ;
const unsigned n = t->number ;
const double * const x = t->x_beg ;
double * const y = t->y_beg ;
unsigned * const all_iter = t->iter ;
unsigned * const my_iter = all_iter + p_size ;
{
unsigned i ;
for ( i = 0 ; i < n ; ) {
TPI_Lock( pool , p_rank );
i = *my_iter * BLOCK ; *my_iter += p_size ;
TPI_Unlock( pool , p_rank );
if ( i < n ) {
const unsigned len = BLOCK < n - i ? BLOCK : n - i ;
daxpby_work( len, a, x + i, b, y + i );
}
}
}
/* Finished my work, steal work from someone else */
{
int working ;
int p = 0 ;
for ( working = 1 ; working ; ) {
working = 0 ;
for ( p = 0 ; p < p_size ; ++p ) {
if ( all_iter[p] * BLOCK < n ) {
if ( ! TPI_Trylock( pool , p ) ) {
const unsigned i = all_iter[p] * BLOCK ;
all_iter[p] += p_size ;
TPI_Unlock( pool , p );
if ( i < n ) {
const unsigned len = BLOCK < n - i ? BLOCK : n - i ;
daxpby_work( len, a, x + i, b, y + i );
}
}
working = 1 ;
}
}
}
}
}
}
inline
int Trylock( ThreadPool pool , int n ) { return TPI_Trylock( pool , n ); }
