void LoweringVisitor::loop_spawn_worksharing(OutlineInfo& outline_info,
Nodecl::NodeclBase construct,
Nodecl::List distribute_environment,
Nodecl::RangeLoopControl& range,
const std::string& outline_name,
TL::Symbol structure_symbol,
TL::Symbol slicer_descriptor,
Nodecl::NodeclBase task_label)
{
Symbol enclosing_function = Nodecl::Utils::get_enclosing_function(construct);
Nodecl::OpenMP::Schedule schedule = distribute_environment.find_first<Nodecl::OpenMP::Schedule>();
ERROR_CONDITION(schedule.is_null(), "Schedule tree is missing", 0);
Nodecl::NodeclBase lower = range.get_lower();
Nodecl::NodeclBase upper = range.get_upper();
Nodecl::NodeclBase step = range.get_step();
Source struct_size, dynamic_size, struct_arg_type_name;
struct_arg_type_name
<< ((structure_symbol.get_type().is_template_specialized_type()
&& structure_symbol.get_type().is_dependent()) ? "typename " : "")
<< structure_symbol.get_qualified_name(enclosing_function.get_scope())
;
struct_size << "sizeof( " << struct_arg_type_name << " )" << dynamic_size;
Source immediate_decl;
allocate_immediate_structure(
structure_symbol.get_user_defined_type(),
outline_info,
struct_arg_type_name,
struct_size,
// out
immediate_decl,
dynamic_size);
Source call_outline_function;
Source schedule_setup;
schedule_setup
<< "int nanos_chunk;"
;
if (schedule.get_text() == "runtime")
{
schedule_setup
<< "nanos_omp_sched_t nanos_runtime_sched;"
<< "nanos_err = nanos_omp_get_schedule(&nanos_runtime_sched, &nanos_chunk);"
<< "if (nanos_err != NANOS_OK)"
<< "nanos_handle_error(nanos_err);"
<< "nanos_ws_t current_ws_policy = nanos_omp_find_worksharing(nanos_runtime_sched);"
;
}
else
{
Source schedule_name;
if (Nanos::Version::interface_is_at_least("openmp", 8))
{
schedule_name << "nanos_omp_sched_" << schedule.get_text();
}
else
{
// We used nanos_omp_sched in versions prior to 8
schedule_name << "omp_sched_" << schedule.get_text();
}
schedule_setup
<< "nanos_ws_t current_ws_policy = nanos_omp_find_worksharing(" << schedule_name << ");"
<< "if (current_ws_policy == 0)"
<< "nanos_handle_error(NANOS_UNIMPLEMENTED);"
<< "nanos_chunk = " << as_expression(schedule.get_chunk()) << ";"
;
}
Source worksharing_creation;
if (IS_CXX_LANGUAGE)
{
worksharing_creation
<< as_statement(Nodecl::CxxDef::make(Nodecl::NodeclBase::null(), slicer_descriptor));
}
worksharing_creation
<< "nanos_err = nanos_worksharing_create("
<< "&" << as_symbol(slicer_descriptor) << ","
<< "current_ws_policy,"
<< "(void**)&nanos_setup_info_loop,"
<< "&single_guard);"
<< "if (nanos_err != NANOS_OK)"
<< "nanos_handle_error(nanos_err);"
;
Nodecl::NodeclBase fill_outline_arguments_tree, fill_immediate_arguments_tree;
TL::Source pm_specific_code;
if (!_lowering->in_ompss_mode())
{
// OpenMP
pm_specific_code
<< immediate_decl
<< statement_placeholder(fill_immediate_arguments_tree)
<< "smp_" << outline_name << "(imm_args);"
;
}
else
{
// OmpSs
std::string wd_description =
(!task_label.is_null()) ? task_label.get_text() : enclosing_function.get_name();
Source const_wd_info;
const_wd_info
<< fill_const_wd_info(struct_arg_type_name,
/* is_untied */ false,
/* mandatory_creation */ true,
/* is_function_task */ false,
wd_description,
outline_info,
construct);
std::string dyn_props_var = "nanos_wd_dyn_props";
Source dynamic_wd_info;
dynamic_wd_info << "nanos_wd_dyn_props_t " << dyn_props_var << ";";
fill_dynamic_properties(dyn_props_var,
/* priority_expr */ nodecl_null(), /* final_expr */ nodecl_null(), /* is_implicit */ 0, dynamic_wd_info);
pm_specific_code
<< struct_arg_type_name << " *ol_args = (" << struct_arg_type_name <<"*) 0;"
<< const_wd_info
<< "nanos_wd_t nanos_wd_ = (nanos_wd_t) 0;"
<< dynamic_wd_info
<< "static nanos_slicer_t replicate = (nanos_slicer_t)0;"
<< "if (replicate == (nanos_slicer_t)0)"
<< "replicate = nanos_find_slicer(\"replicate\");"
<< "if (replicate == (nanos_slicer_t)0)"
<< "nanos_handle_error(NANOS_UNIMPLEMENTED);"
<< "nanos_err = nanos_create_sliced_wd(&nanos_wd_, "
<< "nanos_wd_const_data.base.num_devices, nanos_wd_const_data.devices, "
<< "(size_t)" << struct_size << ", nanos_wd_const_data.base.data_alignment, "
<< "(void**)&ol_args, nanos_current_wd(), replicate,"
<< "&nanos_wd_const_data.base.props, &" << dyn_props_var << ", 0, (nanos_copy_data_t**)0,"
<< "0, (nanos_region_dimension_internal_t**)0"
<< ");"
<< "if (nanos_err != NANOS_OK)"
<< "nanos_handle_error(nanos_err);"
<< statement_placeholder(fill_outline_arguments_tree)
<< "nanos_err = nanos_submit(nanos_wd_, 0, (nanos_data_access_t *) 0, (nanos_team_t) 0);"
<< "if (nanos_err != NANOS_OK)"
<< "nanos_handle_error(nanos_err);"
;
}
TL::Source implicit_barrier_or_tw;
if (!distribute_environment.find_first<Nodecl::OpenMP::BarrierAtEnd>().is_null())
{
implicit_barrier_or_tw << get_implicit_sync_end_construct_source();
}
Source spawn_code;
spawn_code
<< "{"
<< as_type(get_bool_type()) << " single_guard;"
<< "nanos_err_t nanos_err;"
<< schedule_setup
<< "nanos_ws_info_loop_t nanos_setup_info_loop;"
<< "nanos_setup_info_loop.lower_bound = " << as_expression(lower) << ";"
<< "nanos_setup_info_loop.upper_bound = " << as_expression(upper) << ";"
<< "nanos_setup_info_loop.loop_step = " << as_expression(step) << ";"
<< "nanos_setup_info_loop.chunk_size = nanos_chunk;"
<< worksharing_creation
<< pm_specific_code
<< implicit_barrier_or_tw
<< "}"
;
Source fill_outline_arguments, fill_immediate_arguments;
fill_arguments(construct, outline_info, fill_outline_arguments, fill_immediate_arguments);
if (IS_FORTRAN_LANGUAGE)
Source::source_language = SourceLanguage::C;
Nodecl::NodeclBase spawn_code_tree = spawn_code.parse_statement(construct);
if (IS_FORTRAN_LANGUAGE)
Source::source_language = SourceLanguage::Current;
Nodecl::NodeclBase arguments_tree;
TL::Source *fill_arguments;
if (!_lowering->in_ompss_mode())
{
// OpenMP
arguments_tree = fill_immediate_arguments_tree;
fill_arguments = &fill_immediate_arguments;
}
else
{
// OmpSs
arguments_tree = fill_outline_arguments_tree;
fill_arguments = &fill_outline_arguments;
}
// Now attach the slicer symbol to its final scope (see tl-lower-for-worksharing.cpp)
const decl_context_t* spawn_inner_context = arguments_tree.retrieve_context().get_decl_context();
slicer_descriptor.get_internal_symbol()->decl_context = spawn_inner_context;
::insert_entry(spawn_inner_context->current_scope, slicer_descriptor.get_internal_symbol());
// Parse the arguments
Nodecl::NodeclBase new_tree = fill_arguments->parse_statement(arguments_tree);
arguments_tree.replace(new_tree);
// Finally, replace the construct by the tree that represents the spawn code
construct.replace(spawn_code_tree);
}
bool LoweringVisitor::handle_reductions_on_task(
Nodecl::NodeclBase construct,
OutlineInfo& outline_info,
Nodecl::NodeclBase statements,
bool generate_final_stmts,
Nodecl::NodeclBase& final_statements)
{
int num_reductions = 0;
TL::Source
reductions_stuff,
final_clause_stuff,
// This source represents an expression which is used to check if
// we can do an optimization in the final code. This optimization
// consists on calling the original code (with a serial closure) if
// we are in a final context and the reduction variables that we
// are using have not been registered previously
final_clause_opt_expr,
extra_array_red_memcpy;
std::map<TL::Symbol, std::string> reduction_symbols_map;
TL::ObjectList<OutlineDataItem*> data_items = outline_info.get_data_items();
for (TL::ObjectList<OutlineDataItem*>::iterator it = data_items.begin();
it != data_items.end();
it++)
{
if (!(*it)->is_reduction())
continue;
std::pair<TL::OpenMP::Reduction*, TL::Type> red_info_pair = (*it)->get_reduction_info();
TL::OpenMP::Reduction* reduction_info = red_info_pair.first;
TL::Type reduction_type = red_info_pair.second.no_ref();
TL::Symbol reduction_item = (*it)->get_symbol();
TL::Type reduction_item_type = reduction_item.get_type().no_ref();
std::string storage_var_name = (*it)->get_field_name() + "_storage";
TL::Type storage_var_type = reduction_type.get_pointer_to();
TL::Symbol reduction_function, reduction_function_original_var, initializer_function;
// Checking if the current reduction type has been treated before
// Note that if that happens we can reuse the combiner and
// initializer function.
//
// C/C++: note that if the type of the list item is an array type,
// we regiter the reduction over its element type
TL::Type registered_reduction_type = reduction_type;
while (!IS_FORTRAN_LANGUAGE
&& registered_reduction_type.is_array())
{
registered_reduction_type = registered_reduction_type.array_element();
}
LoweringVisitor::reduction_task_map_t::iterator task_red_info =
_task_reductions_map.find(std::make_pair(reduction_info, registered_reduction_type));
if (task_red_info != _task_reductions_map.end())
{
reduction_function = task_red_info->second._reducer;
reduction_function_original_var = task_red_info->second._reducer_orig_var;
initializer_function = task_red_info->second._initializer;
}
else
{
create_reduction_functions(reduction_info,
construct,
registered_reduction_type,
reduction_item,
reduction_function,
reduction_function_original_var);
create_initializer_function(reduction_info,
construct,
registered_reduction_type,
initializer_function);
_task_reductions_map.insert(
std::make_pair(
std::make_pair(reduction_info, registered_reduction_type),
TaskReductionsInfo(reduction_function, reduction_function_original_var, initializer_function)
));
}
// Mandatory TL::Sources to be filled by any reduction
TL::Source
orig_address, // address of the original reduction variable
storage_var; // variable which holds the address of the storage
// Specific TL::Sources to be filled only by Fortran array reduction
TL::Source extra_array_red_decl;
if (IS_C_LANGUAGE || IS_CXX_LANGUAGE)
{
storage_var << storage_var_name;
orig_address << (reduction_item_type.is_pointer() ? "" : "&") << (*it)->get_field_name();
final_clause_stuff
<< "if (" << storage_var_name << " == 0)"
<< "{"
<< storage_var_name << " = "
<< "(" << as_type(storage_var_type) << ")" << orig_address << ";"
<< "}"
;
}
else
{
orig_address << "&" << (*it)->get_field_name();
if (reduction_item_type.is_array())
{
size_t size_of_array_descriptor =
fortran_size_of_array_descriptor(
fortran_get_rank0_type(reduction_item_type.get_internal_type()),
fortran_get_rank_of_type(reduction_item_type.get_internal_type()));
storage_var << storage_var_name << "_indirect";
extra_array_red_decl << "void *" << storage_var << ";";
extra_array_red_memcpy
<< "nanos_err = nanos_memcpy("
<< "(void **) &" << storage_var_name << ","
<< storage_var << ","
<< size_of_array_descriptor << ");"
;
final_clause_stuff
<< "if (" << storage_var << " == 0)"
<< "{"
<< "nanos_err = nanos_memcpy("
<< "(void **) &" << storage_var_name << ","
<< "(void *) "<< orig_address << ","
<< size_of_array_descriptor << ");"
<< "}"
<< "else"
<< "{"
<< extra_array_red_memcpy
<< "}"
;
}
else
{
// We need to convert a void* type into a pointer to the reduction type.
// As a void* in FORTRAN is represented as an INTEGER(8), we cannot do this
// conversion directly in the FORTRAN source. For this reason we introduce
// a new function that will be defined in a C file.
TL::Symbol func = TL::Nanox::get_function_ptr_conversion(
// Destination
reduction_item_type.get_pointer_to(),
// Origin
TL::Type::get_void_type().get_pointer_to(),
construct.retrieve_context());
storage_var << storage_var_name;
final_clause_stuff
<< "if (" << storage_var << " == 0)"
<< "{"
<< storage_var_name << " = " << func.get_name() << "(" << orig_address << ");"
<< "}"
;
}
}
if (num_reductions > 0)
final_clause_opt_expr << " && ";
final_clause_opt_expr << storage_var << " == 0 ";
num_reductions++;
reductions_stuff
<< extra_array_red_decl
<< as_type(storage_var_type) << " " << storage_var_name << ";"
<< "nanos_err = nanos_task_reduction_get_thread_storage("
<< "(void *)" << orig_address << ","
<< "(void **) &" << storage_var << ");"
;
reduction_symbols_map[reduction_item] = storage_var_name;
}
if (num_reductions != 0)
{
// Generating the final code if needed
if (generate_final_stmts)
{
std::map<Nodecl::NodeclBase, Nodecl::NodeclBase>::iterator it4 = _final_stmts_map.find(construct);
ERROR_CONDITION(it4 == _final_stmts_map.end(), "Unreachable code", 0);
Nodecl::NodeclBase placeholder;
TL::Source new_statements_src;
new_statements_src
<< "{"
<< "nanos_err_t nanos_err;"
<< reductions_stuff
<< "if (" << final_clause_opt_expr << ")"
<< "{"
<< as_statement(it4->second)
<< "}"
<< "else"
<< "{"
<< final_clause_stuff
<< statement_placeholder(placeholder)
<< "}"
<< "}"
;
final_statements = handle_task_statements(
construct, statements, placeholder, new_statements_src, reduction_symbols_map);
}
// Generating the task code
{
TL::Source new_statements_src;
Nodecl::NodeclBase placeholder;
new_statements_src
<< "{"
<< "nanos_err_t nanos_err;"
<< reductions_stuff
<< extra_array_red_memcpy
<< statement_placeholder(placeholder)
<< "}"
;
Nodecl::NodeclBase new_statements = handle_task_statements(
construct, statements, placeholder, new_statements_src, reduction_symbols_map);
statements.replace(new_statements);
}
}
ERROR_CONDITION(num_reductions != 0 &&
!Nanos::Version::interface_is_at_least("task_reduction", 1001),
"The version of the runtime begin used does not support task reductions", 0);
return (num_reductions != 0);
}
void LoweringVisitor::visit(const Nodecl::OpenMP::TaskExpression& task_expr)
{
Nodecl::NodeclBase join_task = task_expr.get_join_task();
Nodecl::List task_calls = task_expr.get_task_calls().as<Nodecl::List>();
if (!_lowering->final_clause_transformation_disabled()
&& Nanos::Version::interface_is_at_least("master", 5024))
{
ERROR_CONDITION(!task_expr.get_parent().is<Nodecl::ExpressionStatement>(), "Unexpected node", 0);
Nodecl::NodeclBase expr_stmt = task_expr.get_parent();
TL::Source code;
code
<< "{"
<< as_type(TL::Type::get_bool_type()) << "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)"
<< "{"
<< as_statement(task_expr.get_sequential_code())
<< "}"
<< "else"
<< "{"
<< as_statement(Nodecl::ExpressionStatement::make(task_expr))
<< "}"
<< "}"
;
std::cout << "In expression\n";
if (IS_FORTRAN_LANGUAGE)
Source::source_language = SourceLanguage::C;
Nodecl::NodeclBase if_else_tree = code.parse_statement(expr_stmt);
if (IS_FORTRAN_LANGUAGE)
Source::source_language = SourceLanguage::Current;
expr_stmt.replace(if_else_tree);
}
Nodecl::NodeclBase placeholder_task_expr_transformation;
if (join_task.is<Nodecl::OpenMP::Task>())
{
// Note: don't walk over the OpenMP::Task node because Its visitor ignores
// the placeholder and sets to false the 'inside_task_expression' boolean
visit_task(
join_task.as<Nodecl::OpenMP::Task>(),
/* inside_task_expression */ true,
&placeholder_task_expr_transformation);
}
else if (join_task.is<Nodecl::ExpressionStatement>() &&
join_task.as<Nodecl::ExpressionStatement>().get_nest().is<Nodecl::OpenMP::TaskCall>())
{
visit_task_call(
join_task.as<Nodecl::ExpressionStatement>().get_nest().as<Nodecl::OpenMP::TaskCall>(),
/* inside_task_expression */ true,
&placeholder_task_expr_transformation);
}
else
{
internal_error("Unreachable code", 0);
}
// Note: don't walk over the OpenMP::TaskCall because It's visitor sets to
// false the 'inside_task_expression' boolean
for (Nodecl::List::iterator it = task_calls.begin();
it != task_calls.end();
++it)
{
Nodecl::ExpressionStatement current_expr_stmt = it->as<Nodecl::ExpressionStatement>();
Nodecl::OpenMP::TaskCall current_task_call = current_expr_stmt.get_nest().as<Nodecl::OpenMP::TaskCall>();
visit_task_call(current_task_call,
/* inside_task_expression */ true,
/* placeholder_task_expr_transformation */ NULL);
Nodecl::Utils::prepend_items_before(placeholder_task_expr_transformation, *it);
}
ERROR_CONDITION(!task_expr.get_parent().is<Nodecl::ExpressionStatement>(), "Unexpected node", 0);
Nodecl::NodeclBase expr_stmt = task_expr.get_parent();
Nodecl::Utils::prepend_items_before(expr_stmt, task_expr.get_join_task());
// Finally, remove from the tree the TaskExpression node
Nodecl::Utils::remove_from_enclosing_list(expr_stmt);
}