void txddot( double * s , unsigned n , const double * x , const double * y )
{
int p_size ;
if ( ! TPI_Size( & p_size ) ) {
double* tmp;
const int ntmp = 4 * p_size ;
tmp = malloc(ntmp * sizeof(double));
{ struct TaskXY data = { tmp , x , y , n , BLOCKING_SIZE };
int i ;
for ( i = 0 ; i < ntmp ; ++i ) {
tmp[i] = 0 ;
}
if ( data.block ) {
TPI_Run( & task_xddot_xy_work_blocking , & data , 0 );
}
else {
TPI_Run( & task_xddot_xy_work , & data , 0 );
}
for ( i = 0 ; i < p_size ; ++i ) {
xdsum_add_dsum( s , tmp + 4 * i );
}
}
free(tmp);
}
}
void tdaxpby( unsigned n ,
double a , const double * x ,
double b , double * y ,
int block )
{
int p_size ;
TPI_Size( & p_size );
{
unsigned *tmp = malloc( p_size );
struct TaskXY data = { a , b , x , y , n , tmp };
int i ;
for ( i = 0 ; i < p_size ; ++i ) { tmp[i] = i ; }
if ( 0 < block ) {
TPI_Run( & task_axpby_work_block , & data , 0 );
}
else if ( block < 0 ) {
TPI_Set_lock_size( p_size );
TPI_Run( & task_axpby_work_steal , & data , 0 );
}
else {
TPI_Run( & task_axpby_work , & data , 0 );
}
free(tmp);
}
}
void txddot1( double * s , unsigned n , const double * x )
{
struct TaskX data ;
data.x_sum = s ;
data.x_beg = x ;
data.number = n ;
TPI_Set_lock_size( 1 );
TPI_Run( & task_xddot_x_work , & data , 0 );
}
inline
int Run( Worker & worker, void (Worker::*method)(ThreadPool) , int n )
{
typedef WorkerMethodHelper<Worker> WM ;
WM tmp( worker , method );
return TPI_Run( reinterpret_cast<TPI_parallel_subprogram>(& WM::run),&tmp,n);
}
inline
int Run( Worker & worker, void (Worker::*method)(Work &) ,
int work_count , int lock_count )
{
typedef WorkerMethodHelper<Worker> WM ;
WM tmp( worker , method );
return TPI_Run( reinterpret_cast<TPI_work_subprogram>(& WM::run),&tmp,work_count,lock_count);
}
void txblas_cr_mxv(
const unsigned nr /* Number rows */ ,
const unsigned pc[] ,
const unsigned ia[] ,
const double a[] ,
const double x[] , /* Input vector */
double y[] ) /* Output vector */
{
txblasTask_cr_Matrix data = { nr , pc , ia , a , x , y };
TPI_Run( & txblas_task_cr_mxv , & data , 0 );
}
void tddot( double * s , unsigned n , const double * x , const double * y )
{
int p_size ;
if ( ! TPI_Size( & p_size ) ) {
double* tmp = malloc( p_size * sizeof(double));
struct TaskXY data = { tmp , x , y , n , BLOCKING_SIZE };
int i ;
for ( i = 0 ; i < p_size ; ++i ) {
tmp[i] = 0 ;
}
if ( data.block ) {
TPI_Run( & task_ddot_xy_work_blocking , & data , 0 );
}
else {
TPI_Run( & task_ddot_xy_work , & data , 0 );
}
for ( i = 1 ; i < p_size ; ++i ) {
tmp[0] += tmp[i] ;
}
*s = tmp[0] ;
free(tmp);
}
}
void test_tpi_work( const int ntest , const int nthread[] , const int nwork ,
const int ntrial )
{
int * const flags = (int *) malloc( sizeof(int) * nwork );
int j ;
fprintf( stdout , "\n\"TEST TPI_Run / TPI_Run_reduce\"\n" );
fprintf( stdout , "\"#Thread\" , \"#Work\" , \"#Trial\" , \"TPI_Run(avg-msec)\" , \"TPI_Run(stddev-msec)\" , \"TPI_Run_reduce(avg-msec)\" , \"TPI_Run_reduce(stddev-msec)\"\n");
for ( j = 0 ; j < ntest ; ++j ) {
const int nth = nthread[j];
double dt_work_total = 0.0 ;
double dt_work_total_2 = 0.0 ;
double dt_reduce_total = 0.0 ;
double dt_reduce_total_2 = 0.0 ;
int i , k ;
int result = TPI_Init( nth );
if ( result != nth ) {
fprintf(stderr,"%d != TPI_Init(%d) : FAILED\n", result , nth );
}
for ( i = 0 ; i < ntrial ; ++i ) {
double t , dt ;
int value = 0 ;
for ( k = 0 ; k < nwork ; ++k ) { flags[k] = 0 ; }
t = TPI_Walltime();
TPI_Run( test_work , & flags , nwork , 0 );
dt = TPI_Walltime() - t ;
dt_work_total += dt ;
dt_work_total_2 += dt * dt ;
for ( k = 0 ; k < nwork && flags[k] ; ++k );
if ( k < nwork ) {
fprintf(stderr, "TPI_Run(...) : FAILED at trial %d\n", i );
abort();
}
for ( k = 0 ; k < nwork ; ++k ) { flags[k] = 0 ; }
t = TPI_Walltime();
TPI_Run_reduce( test_reduce_work , & flags , nwork ,
test_reduce_join , test_reduce_init ,
sizeof(value) , & value );
dt = TPI_Walltime() - t ;
dt_reduce_total += dt ;
dt_reduce_total_2 += dt * dt ;
for ( k = 0 ; k < nwork && flags[k] ; ++k );
if ( value != nwork || k < nwork ) {
fprintf(stderr, "TPI_Run_reduce(...) : FAILED at trial %d\n", i );
abort();
}
}
TPI_Finalize();
if ( 1 < ntrial ) {
const double work_mean = 1.0e6 * dt_work_total / ntrial ;
const double work_sdev = 1.0e6 * sqrt( ( ntrial * dt_work_total_2 -
dt_work_total * dt_work_total ) /
( ntrial * ( ntrial - 1 ) ) );
const double reduce_mean = 1.0e6 * dt_reduce_total / ntrial ;
const double reduce_sdev = 1.0e6 * sqrt( ( ntrial * dt_reduce_total_2 -
dt_reduce_total * dt_reduce_total) /
( ntrial * ( ntrial - 1 ) ) );
fprintf(stdout,"%d , %d , %d , %10g , %10g , %10g , %10g\n",
nth, ntrial, nwork, work_mean, work_sdev, reduce_mean, reduce_sdev);
}
}
free( flags );
}
inline
int Run( void (*func)( void * , ThreadPool ) , void * arg , int n )
{
return TPI_Run( reinterpret_cast< TPI_parallel_subprogram >(func), arg , n );
}