void NANOS_parallel_end( void )
{
nanos_err_t err = nanos_omp_barrier( );
if( err != NANOS_OK )
nanos_handle_error( err );
err = nanos_leave_team( );
if( err != NANOS_OK )
nanos_handle_error( err );
}
void NANOS_parallel_init( void )
{
nanos_err_t err = nanos_omp_set_implicit( nanos_current_wd( ) );
if( err != NANOS_OK )
nanos_handle_error( err );
err = nanos_enter_team( );
if( err != NANOS_OK )
nanos_handle_error( err );
}
void NANOS_taskwait( void )
{
void * wg = nanos_current_wd( );
nanos_err_t err = nanos_wg_wait_completion( wg, 0 );
if( err != NANOS_OK )
nanos_handle_error( err );
}
bool NANOS_single( void )
{
bool single_guard;
nanos_err_t err = nanos_single_guard( &single_guard );
if( err != NANOS_OK )
nanos_handle_error(err);
return single_guard;
}
static void nanos_xlate_fun_fibompmemoc_1(struct nanos_args_1_t *const arg, void *wd)
{
{
void *device_base_address;
nanos_err_t err;
device_base_address = 0;
err = nanos_get_addr(0, &device_base_address, wd);
if (err != NANOS_OK)
{
nanos_handle_error(err);
}
(*arg).res = (int *)device_base_address;
}
}
void NANOS_sections( void ( * func ) ( void * section_data, nanos_ws_desc_t * wsd ), void * data,
long data_size, long ( * get_data_align )( void ), void * empty_data, void ( * init_func ) ( void *, void * ),
int n_sections, bool wait )
{
// Get scheduling policy
void * ws_policy = nanos_omp_find_worksharing( omp_sched_static );
if( ws_policy == 0 )
nanos_handle_error( NANOS_UNIMPLEMENTED );
char * sections_name;
asprintf( §ions_name, "sections_%d", sections_id++ );
NANOS_worksharing( /*lb*/ 0, /*ub*/ n_sections - 1, /*step*/ 1 , /*chunk*/ 1, sections_name,
func, data, data_size, get_data_align, empty_data, init_func,
ws_policy, wait );
}
void NANOS_loop( void ( * func ) ( void * loop_data, nanos_ws_desc_t * wsd ), void * data, long data_size, long ( * get_data_align )( void ),
void* empty_data, void ( * init_func ) ( void *, void * ), int policy,
int lower_bound, int upper_bound, int step, int chunk, bool wait )
{
// Get scheduling policy
void * ws_policy = nanos_omp_find_worksharing( nanos_get_scheduling( policy ) );
if( ws_policy == 0 )
nanos_handle_error( NANOS_UNIMPLEMENTED );
char * loop_name;
asprintf( &loop_name, "loop_%d", loop_id++ );
NANOS_worksharing( lower_bound, upper_bound, step , chunk, loop_name,
func, data, data_size, get_data_align, empty_data, init_func,
ws_policy, wait );
}
void fib(int n, int *res)
{
if (n < 2)
{
*res = n;
}
else
{
int res1 = 0;
int res2 = 0;
{
int mcc_arg_0 = n - 1;
int *mcc_arg_1 = &res1;
{
_Bool mcc_is_in_final;
nanos_err_t mcc_err_in_final = nanos_in_final(&mcc_is_in_final);
if (mcc_err_in_final != NANOS_OK)
{
nanos_handle_error(mcc_err_in_final);
}
if (mcc_is_in_final)
{
fib_mcc_serial(n - 1, &res1);
}
else
{
{
nanos_wd_dyn_props_t nanos_wd_dyn_props;
int memo_dimensions[1];
struct nanos_args_0_t *ol_args;
nanos_err_t err;
struct nanos_args_0_t imm_args;
static nanos_smp_args_t smp_ol_fib_1_args = {.outline = (void (*)(void *))(void (*)(struct nanos_args_0_t *))&smp_ol_fib_1};
static struct nanos_const_wd_definition_1 nanos_wd_const_data = {.base = {.props = {.mandatory_creation = 0, .tied = 0, .clear_chunk = 0, .reserved0 = 0, .reserved1 = 0, .reserved2 = 0, .reserved3 = 0, .reserved4 = 0}, .data_alignment = __alignof__(struct nanos_args_0_t), .num_copies = 1, .num_devices = 1, .num_dimensions = 1, .description = 0}, .devices = {[0] = {.factory = &nanos_smp_factory, .arg = &smp_ol_fib_1_args}}};
nanos_wd_dyn_props.tie_to = 0;
nanos_wd_dyn_props.priority = 0;
nanos_wd_dyn_props.flags.is_final = 1;
memo_dimensions[0] = mcc_arg_0 + 1;
nanos_wd_dyn_props.memo.num_dimensions = 1;
nanos_wd_dyn_props.memo.dimensions = memo_dimensions;
ol_args = (struct nanos_args_0_t *)0;
nanos_wd_t nanos_wd_ = (void *)0;
nanos_copy_data_t *ol_copy_data = (nanos_copy_data_t *)0;
nanos_region_dimension_internal_t *ol_copy_dimensions = (nanos_region_dimension_internal_t *)0;
err = nanos_create_wd_compact(&nanos_wd_, &nanos_wd_const_data.base, &nanos_wd_dyn_props, sizeof(struct nanos_args_0_t), (void **)&ol_args, nanos_current_wd(), &ol_copy_data, &ol_copy_dimensions);
if (err != NANOS_OK)
{
nanos_handle_error(err);
}
nanos_region_dimension_t dimensions_0[1] = {[0] = {.size = sizeof(int), .lower_bound = 0, .accessed_length = sizeof(int)}};
nanos_data_access_t dependences[1] = {[0] = {.address = (void *)mcc_arg_1, .flags = {.input = 0, .output = 1, .can_rename = 0, .concurrent = 0, .commutative = 0}, .dimension_count = (short int)1, .dimensions = dimensions_0, .offset = 0}};
;
if (nanos_wd_ != (void *)0)
{
(*ol_args).n = mcc_arg_0;
(*ol_args).res = mcc_arg_1;
ol_copy_dimensions[0].size = 1 * sizeof(int);
ol_copy_dimensions[0].lower_bound = 0 * sizeof(int);
ol_copy_dimensions[0].accessed_length = (0 - 0 + 1) * sizeof(int);
ol_copy_data[0].sharing = NANOS_SHARED;
ol_copy_data[0].address = (void *)mcc_arg_1;
ol_copy_data[0].flags.input = 0;
ol_copy_data[0].flags.output = 1;
ol_copy_data[0].dimension_count = (short int)1;
ol_copy_data[0].dimensions = &ol_copy_dimensions[0];
ol_copy_data[0].offset = 0;
err = nanos_set_translate_function(nanos_wd_, (void (*)(void *, nanos_wd_t))nanos_xlate_fun_fibompmemoc_0);
if (err != NANOS_OK)
{
nanos_handle_error(err);
}
err = nanos_submit(nanos_wd_, 1, dependences, (void *)0);
if (err != NANOS_OK)
{
nanos_handle_error(err);
}
}
else
{
nanos_region_dimension_internal_t imm_copy_dimensions[1];
nanos_copy_data_t imm_copy_data[1];
imm_args.n = mcc_arg_0;
imm_args.res = mcc_arg_1;
imm_copy_dimensions[0].size = 1 * sizeof(int);
imm_copy_dimensions[0].lower_bound = 0 * sizeof(int);
imm_copy_dimensions[0].accessed_length = (0 - 0 + 1) * sizeof(int);
imm_copy_data[0].sharing = NANOS_SHARED;
imm_copy_data[0].address = (void *)mcc_arg_1;
imm_copy_data[0].flags.input = 0;
imm_copy_data[0].flags.output = 1;
imm_copy_data[0].dimension_count = (short int)1;
imm_copy_data[0].dimensions = &imm_copy_dimensions[0];
imm_copy_data[0].offset = 0;
err = nanos_create_wd_and_run_compact(&nanos_wd_const_data.base, &nanos_wd_dyn_props, sizeof(struct nanos_args_0_t), &imm_args, 1, dependences, imm_copy_data, imm_copy_dimensions, (void (*)(void *, nanos_wd_t))nanos_xlate_fun_fibompmemoc_0);
if (err != NANOS_OK)
{
nanos_handle_error(err);
}
}
}
}
}
}
int main(int argc, char * argv[])
{
int * * l_array_of_arrays;
int * l_partial_sums;
int l_num_procs;
int l_total;
int l_i, l_j;
if (argc != 2)
{
printf("Usage: %s number_of_processors\n", argv[0]);
return 0;
}
l_num_procs = atoi(argv[1]);
if (l_num_procs < 1 && l_num_procs > 16)
{
printf("The number of processors must be between 1 and 16\n");
return 0;
}
l_partial_sums = (int *) malloc(l_num_procs * sizeof(int));
l_array_of_arrays = (int **) malloc(l_num_procs * sizeof(int *));
for (l_i = 0;
l_i < l_num_procs;
l_i++)
{
l_array_of_arrays[l_i] = (int *) malloc(16834 * sizeof(int));
for (l_j = 0;
l_j < 16834;
l_j++)
{
if ((l_j % 2) == 0)
l_array_of_arrays[l_i][l_j] = 1;
else
l_array_of_arrays[l_i][l_j] = 0;
}
}
for (l_i = 0;
l_i < l_num_procs;
l_i++)
{
{
nanos_smp_args_t _ol_main_0_smp_args = {
(void (*)(void *)) _smp__ol_main_0
};
_nx_data_env_0_t * ol_args = (_nx_data_env_0_t *) 0;
nanos_wd_t wd = (nanos_wd_t) 0;
const_data1.data_alignment = __alignof__(_nx_data_env_0_t);
const_data1.devices[0].arg = &_ol_main_0_smp_args;
nanos_wd_dyn_props_t dyn_data1 = { 0 };
nanos_err_t err;
err = nanos_create_wd_compact(&wd, (nanos_const_wd_definition_t *) &const_data1, &dyn_data1, sizeof(_nx_data_env_0_t), (void **) &ol_args, nanos_current_wd(), (nanos_copy_data_t **) 0, NULL);
if (err != NANOS_OK)
nanos_handle_error(err);
if (wd != (nanos_wd_t) 0)
{
ol_args->l_array_of_arrays_0 = l_array_of_arrays;
ol_args->l_partial_sums_0 = l_partial_sums;
ol_args->l_i_0 = l_i;
err = nanos_submit(wd, 0, (nanos_data_access_t *) 0, (nanos_team_t) 0);
if (err != NANOS_OK)
nanos_handle_error(err);
}
else
{
_nx_data_env_0_t imm_args;
imm_args.l_array_of_arrays_0 = l_array_of_arrays;
imm_args.l_partial_sums_0 = l_partial_sums;
imm_args.l_i_0 = l_i;
err = nanos_create_wd_and_run_compact((nanos_const_wd_definition_t *) &const_data1, &dyn_data1, sizeof(_nx_data_env_0_t),
&imm_args, 0, (nanos_data_access_t *) 0, (nanos_copy_data_t *) 0, 0, NULL);
if (err != NANOS_OK)
nanos_handle_error(err);
}
}
}
nanos_wg_wait_completion( nanos_current_wd(), 0 );
l_total = 0;
for (l_i = 0;
l_i < l_num_procs;
l_i++)
{
printf("%d -> %d\n", l_i, l_partial_sums[l_i]);
l_total += l_partial_sums[l_i];
}
printf("Result = %d\n", l_total);
return 0;
}
int main(int argc, char **argv)
{
int i;
int a[16];
int it;
int rv = 0;
for (i = 0; i < 16; i++)
{
a[i] = 0;
}
for (it = 0; it < 4; it++)
{
{
nanos_err_t nanos_err;
nanos_wd_dyn_props_t dyn_props;
unsigned int nth_i;
struct nanos_args_1_t imm_args;
nanos_data_access_t dependences[1];
static nanos_smp_args_t smp_ol_main_1_args = {.outline = (void (*)(void *))(void (*)(struct nanos_args_1_t *))&smp_ol_main_1};
static struct nanos_const_wd_definition_1 nanos_wd_const_data = {.base = {.props = {.mandatory_creation = 1, .tied = 1, .clear_chunk = 0, .reserved0 = 0, .reserved1 = 0, .reserved2 = 0, .reserved3 = 0, .reserved4 = 0}, .data_alignment = __alignof__(struct nanos_args_1_t), .num_copies = 0, .num_devices = 1, .num_dimensions = 0, .description = 0}, .devices = {[0] = {.factory = &nanos_smp_factory, .arg = &smp_ol_main_1_args}}};
unsigned int nanos_num_threads = nanos_omp_get_num_threads_next_parallel(0);
nanos_team_t nanos_team = (void *)0;
nanos_thread_t nanos_team_threads[nanos_num_threads];
nanos_err = nanos_create_team(&nanos_team, (void *)0, &nanos_num_threads, (nanos_constraint_t *)0, 1, nanos_team_threads, &nanos_wd_const_data.base);
if (nanos_err != NANOS_OK)
{
nanos_handle_error(nanos_err);
}
dyn_props.tie_to = (void *)0;
dyn_props.priority = 0;
dyn_props.flags.is_final = 0;
for (nth_i = 1; nth_i < nanos_num_threads; nth_i = nth_i + 1)
{
dyn_props.tie_to = nanos_team_threads[nth_i];
struct nanos_args_1_t *ol_args = 0;
nanos_wd_t nanos_wd_ = (void *)0;
nanos_err = nanos_create_wd_compact(&nanos_wd_, &nanos_wd_const_data.base, &dyn_props, sizeof(struct nanos_args_1_t), (void **)&ol_args, nanos_current_wd(), (nanos_copy_data_t **)0, (nanos_region_dimension_internal_t **)0);
if (nanos_err != NANOS_OK)
{
nanos_handle_error(nanos_err);
}
(*ol_args).i = &i;
(*ol_args).a = &a;
nanos_err = nanos_submit(nanos_wd_, 0, (nanos_data_access_t *)0, (void *)0);
if (nanos_err != NANOS_OK)
{
nanos_handle_error(nanos_err);
}
}
dyn_props.tie_to = nanos_team_threads[0];
imm_args.i = &i;
imm_args.a = &a;
nanos_err = nanos_create_wd_and_run_compact(&nanos_wd_const_data.base, &dyn_props, sizeof(struct nanos_args_1_t), &imm_args, 0, dependences, (nanos_copy_data_t *)0, (nanos_region_dimension_internal_t *)0, (void (*)(void *, nanos_wd_t))0);
if (nanos_err != NANOS_OK)
{
nanos_handle_error(nanos_err);
}
nanos_err = nanos_end_team(nanos_team);
if (nanos_err != NANOS_OK)
{
nanos_handle_error(nanos_err);
}
}
}
void NANOS_critical_end( void )
{
nanos_err_t err = nanos_unset_lock( &nanos_default_critical_lock );
if( err != NANOS_OK )
nanos_handle_error( err );
}
void NANOS_parallel( void ( * func ) ( void * ), void * data, unsigned numThreads, long data_size, long ( *get_data_align )( void ),
void * ( * get_empty_data )( void ), void ( * init_func ) ( void *, void * ) )
{
nanos_err_t err;
// Compute copy data (For SMP devices there are no copies. Just CUDA device requires copy data)
int num_copies = 0;
// TODO Compute dimensions
int num_dimensions = 0;
// Compute device descriptor (at the moment, only SMP is supported)
int num_devices = 1;
// TODO No dependencies for parallel construct in SMP devices
int num_data_accesses = 0;
nanos_data_access_t dependences[1];
// Create the Device descriptor (at the moment, only SMP is supported)
nanos_smp_args_t _smp_args = { func };
char * parallel_name;
asprintf( ¶llel_name, "parallel_%d", parallel_id++ );
struct nanos_const_wd_definition nanos_wd_const_data = {
{ { 1, // mandatory creation
1, // tied
0, 0, 0, 0, 0, 0 }, // properties
( *get_data_align )( ), // data alignment
num_copies, num_devices, num_dimensions,
parallel_name // description
},
{ { &nanos_smp_factory, // device description
&_smp_args } // outlined function
}
};
// Compute properties of the WD: mandatory creation, priority, tiedness, real-time info and copy declarations
nanos_wd_dyn_props_t dyn_props;
dyn_props.tie_to = ( void * ) 0;
dyn_props.priority = 0;
dyn_props.flags.is_final = 0;
// Create the working team
if( numThreads == 0 )
numThreads = nanos_omp_get_num_threads_next_parallel( 0 );
void * nanos_team = ( void * ) 0;
const unsigned int nthreads_vla = numThreads;
void * team_threads[nthreads_vla];
err = nanos_create_team( &nanos_team, ( void * ) 0, &numThreads,
(nanos_constraint_t *) 0, /*reuse current*/ 1, team_threads );
if( err != NANOS_OK )
nanos_handle_error( err );
// Create a wd tied to each thread
unsigned nth_i;
for( nth_i = 1; nth_i < numThreads; nth_i++ )
{
// Set properties to the current wd of the team
dyn_props.tie_to = team_threads[nth_i];
// Create the current WD of the team
void * empty_data = ( *get_empty_data )( );
void * wd = ( void * ) 0;
err = nanos_create_wd_compact( &wd, &nanos_wd_const_data.base, &dyn_props,
data_size, ( void** ) &empty_data,
nanos_current_wd( ), ( nanos_copy_data_t ** ) 0,
( nanos_region_dimension_internal_t ** ) 0 );
if (err != NANOS_OK)
nanos_handle_error(err);
// Initialize outlined data
( *init_func )( empty_data, data );
// Submit work to the WD
err = nanos_submit( wd, num_data_accesses, ( nanos_data_access_t * ) 0, ( void * ) 0 );
if( err != NANOS_OK )
nanos_handle_error( err );
}
// Create the wd for the master thread, which will run the team
dyn_props.tie_to = team_threads[0];
err = nanos_create_wd_and_run_compact( &nanos_wd_const_data.base, &dyn_props, data_size, data,
num_data_accesses, dependences, ( nanos_copy_data_t * ) 0,
( nanos_region_dimension_internal_t * ) 0,
( void ( * )( void *, void * ) ) 0 );
if( err != NANOS_OK )
nanos_handle_error( err );
// End the team
err = nanos_end_team( nanos_team );
if( err != NANOS_OK )
nanos_handle_error( err );
}
void NANOS_reduction( int n_reductions,
void ( ** all_threads_reduction )( void * out, void * in, int num_scalars ),
void ( * func )( void * data, /*void** globals, */nanos_ws_desc_t * wsd ), void * data,
void ( ** copy_back )( int team_size, void * original, void * privates ),
void ( ** set_privates )( void * nanos_private, void ** global_data, int reduction_id, int thread ),
void ** global_th_data, void ** global_data, long * global_data_size,
nanos_ws_desc_t * wsd, const char * filename, int fileline )
{
nanos_err_t err;
err = nanos_omp_set_implicit( nanos_current_wd( ) );
if( err != NANOS_OK )
nanos_handle_error( err );
bool red_single_guard;
err = nanos_enter_sync_init( &red_single_guard );
if( err != NANOS_OK )
nanos_handle_error( err );
nanos_reduction_t* result[n_reductions];
int nanos_n_threads = nanos_omp_get_num_threads( );
// void * _global_[ nanos_n_threads ];
if( red_single_guard )
{
int i;
for( i = 0; i < n_reductions; i++ )
{
err = nanos_malloc( ( void ** ) &result, sizeof( nanos_reduction_t ), filename, fileline );
if( err != NANOS_OK )
nanos_handle_error( err );
( * ( result[i] ) ).original = global_data[i];
err = nanos_malloc( &( * ( result[i] ) ).privates, global_data_size[i] * nanos_n_threads, filename, fileline );
if( err != NANOS_OK )
nanos_handle_error( err );
( * ( result[i] ) ).descriptor = ( * ( result[i] ) ).privates; // Fortran only
// _global_[i] = (int *)( * ( result[i] ) ).privates;
( * ( result[i] ) ).vop = copy_back[i];
( * ( result[i] ) ).bop = all_threads_reduction[i];
( * ( result[i] ) ).element_size = global_data_size[i];
( * ( result[i] ) ).num_scalars = 1;
( * ( result[i] ) ).cleanup = nanos_free0;
err = nanos_register_reduction( result[i] );
if( err != NANOS_OK )
nanos_handle_error( err );
}
err = nanos_release_sync_init( );
if( err != NANOS_OK )
nanos_handle_error( err );
}
else
{
err = nanos_wait_sync_init( );
if( err != NANOS_OK )
nanos_handle_error( err );
int i;
for( i = 0; i < n_reductions; i++ )
{
err = nanos_reduction_get( & ( result[i] ), global_data[i] );
if( err != NANOS_OK )
nanos_handle_error( err );
// _global_[i] = (int *)( * ( result[i] ) ).privates;
}
}
// Execute the function containing the reduction
( * func )( data, /*_global_, */wsd );
// Point the 'privates' member to the actual private value computed in the reduction code
// FIXME copy back data cannot be made at the end because
// the privates member is used before this copy back is performed
int i;
for( i = 0; i < n_reductions; i++ )
{
( * ( set_privates[i] ) )( ( * ( result[i] ) ).privates, global_th_data, i, omp_get_thread_num( ) );
}
}
static void NANOS_worksharing( int lb, int ub, int step, int chunk, char * description,
void ( * func ) ( void * data, nanos_ws_desc_t * wsd ), void * data, long data_size, long ( * get_data_align )( void ),
void * empty_data, void ( * init_func ) ( void *, void * ), void * ws_policy, bool wait )
{
nanos_err_t err;
// Create the Worksharing
bool single_guard;
nanos_ws_desc_t * wsd;
nanos_ws_info_loop_t ws_info_loop;
ws_info_loop.lower_bound = lb;
ws_info_loop.upper_bound = ub;
ws_info_loop.loop_step = step;
ws_info_loop.chunk_size = chunk;
err = nanos_worksharing_create( &wsd, ws_policy, ( void ** ) &ws_info_loop, &single_guard );
if( err != NANOS_OK )
nanos_handle_error( err );
if( single_guard )
{
int sup_threads;
err = nanos_team_get_num_supporting_threads( &sup_threads );
if( err != NANOS_OK )
nanos_handle_error( err );
if( sup_threads > 0 )
{
// Configure the Worksahring
err = nanos_malloc( ( void ** ) &( *wsd ).threads, sizeof( void * ) * sup_threads, /*filename*/"", /*fileline*/0 );
if( err != NANOS_OK )
nanos_handle_error( err );
err = nanos_team_get_supporting_threads( &( *wsd ).nths, ( *wsd ).threads );
if( err != NANOS_OK )
nanos_handle_error( err );
// Create the WD and its properties
void * wd = ( void * ) 0;
nanos_wd_dyn_props_t props;
props.tie_to = ( void * ) 0;
props.priority = 0;
props.flags.is_final = 0;
// Compute copy data (For SMP devices there are no copies. Just CUDA device requires copy data)
int num_copies = 0;
// Compute dependencies (ROSE is not currently supporting dependencies among the tasks)
int num_data_accesses = 0;
// TODO Compute dimensions
int num_dimensions = 0;
// Compute device descriptor (at the moment, only SMP is supported)
int num_devices = 1;
// Create the slicer
nanos_smp_args_t _smp_args = { func };
struct nanos_const_wd_definition nanos_wd_const_data = {
{ { 1, // mandatory creation
1, // tied
0, 0, 0, 0, 0, 0 }, // properties
( *get_data_align )( ), // data alignment
num_copies, num_devices, num_dimensions,
description // description
},
{ { &nanos_smp_factory, // device description
&_smp_args } // outlined function
}
};
void * slicer = nanos_find_slicer( "replicate" );
if( slicer == (void *)0 )
nanos_handle_error( NANOS_UNIMPLEMENTED );
struct sections_data_t* empty_data = ( struct sections_data_t * ) 0;
err = nanos_create_sliced_wd( &wd, nanos_wd_const_data.base.num_devices, nanos_wd_const_data.devices,
data_size, nanos_wd_const_data.base.data_alignment, ( void ** ) &empty_data,
( void ** ) 0, slicer, &nanos_wd_const_data.base.props, &props,
num_copies, ( nanos_copy_data_t ** ) 0,
num_dimensions, ( nanos_region_dimension_internal_t ** ) 0 );
if( err != NANOS_OK )
nanos_handle_error( err );
// Initialize outlined data
( *init_func )( empty_data, data );
// Submit the work to the runtime system
err = nanos_submit( wd, num_data_accesses, ( nanos_data_access_t * ) 0, ( nanos_team_t ) 0 );
if( err != NANOS_OK )
nanos_handle_error( err );
err = nanos_free( ( * wsd ).threads );
if( err != NANOS_OK )
nanos_handle_error( err );
}
}
( * func )( data, wsd );
// Wait in case it is necessary
if( wait )
{
err = nanos_omp_barrier( );
if( err != NANOS_OK )
nanos_handle_error( err );
}
}
void NANOS_task( void ( * func ) ( void * ), void *data,
long data_size, long ( * get_data_align ) ( void ),
void * empty_data, void ( * init_func ) ( void *, void * ),
bool if_clause, unsigned untied,
int num_deps, int * deps_dir, void ** deps_data,
int * deps_n_dims, nanos_region_dimension_t ** deps_dims,
long int * deps_offset )
{
nanos_err_t err;
bool nanos_is_in_final;
err = nanos_in_final( &nanos_is_in_final );
if( nanos_is_in_final )
{
( *func )( data );
}
else
{
// Compute copy data (For SMP devices there are no copies. Just CUDA device requires copy data)
int num_copies = 0;
// TODO Compute dimensions (for devices other than SMP)
int num_dimensions = 0;
// Compute device descriptor (at the moment, only SMP is supported)
int num_devices = 1;
// Compute dependencies
const unsigned int num_data_accesses = num_deps;
nanos_data_access_t dependences[num_data_accesses];
int i;
for( i = 0; i < num_data_accesses; ++i )
{
int in = ( deps_dir[i] & ( e_dep_dir_in | e_dep_dir_inout ) );
int out = ( deps_dir[i] & ( e_dep_dir_out | e_dep_dir_inout ) );
nanos_access_type_internal_t flags = {
( in != 0 ), // input
( out != 0 ), // output
0 , // can rename
0 , // concurrent
0 , // commutative
};
nanos_data_access_t dep = { deps_data[i], flags, deps_n_dims[i], deps_dims[i], deps_offset[i] };
dependences[i] = dep;
}
// Create the Device descriptor (at the moment, only SMP is supported)
nanos_smp_args_t _smp_args = { func };
char * task_name;
asprintf( &task_name, "task_%d", task_id++ );
struct nanos_const_wd_definition nanos_wd_const_data = {
{ { 0, // mandatory creation
!untied, // tied
0, 0, 0, 0, 0, 0 }, // properties
( *get_data_align )( ), // data alignment
num_copies, num_devices, num_dimensions,
task_name // description
},
{ { &nanos_smp_factory, // device description
&_smp_args } // outlined function
}
};
// Compute properties of the WD: mandatory creation, priority, tiedness, real-time info and copy declarations
nanos_wd_dyn_props_t dyn_props;
dyn_props.tie_to = 0;
dyn_props.priority = 0;
dyn_props.flags.is_final = 0;
// Create the WD
nanos_wd_t wd = (nanos_wd_t) 0;
err = nanos_create_wd_compact( &wd, &nanos_wd_const_data.base, &dyn_props,
data_size, ( void ** ) &empty_data,
nanos_current_wd( ), ( nanos_copy_data_t ** ) 0,
( nanos_region_dimension_internal_t ** ) 0 );
if( err != NANOS_OK )
nanos_handle_error( err );
if( wd != ( void * ) 0 )
{ // Submit the task to the existing actual working group
// Initialize outlined data
( *init_func )( empty_data, data );
err = nanos_submit( wd, num_data_accesses, dependences, ( void * ) 0 );
if( err != NANOS_OK )
nanos_handle_error( err );
}
else
{ // The task must be run immediately
err = nanos_create_wd_and_run_compact( &nanos_wd_const_data.base, &dyn_props,
data_size, data, num_data_accesses,
dependences, ( nanos_copy_data_t * ) 0,
( nanos_region_dimension_internal_t * ) 0,
( void ( * )( void *, void * ) ) 0 );
if( err != NANOS_OK )
nanos_handle_error( err );
}
}
}