static void handle_ompss_opencl_deallocate_intrinsic( Nodecl::FunctionCall function_call, Nodecl::NodeclBase expr_stmt) { Nodecl::List arguments = function_call.get_arguments().as<Nodecl::List>(); ERROR_CONDITION(arguments.size() != 1, "More than one argument in ompss_opencl_deallocate call", 0); Nodecl::NodeclBase actual_argument = arguments[0]; ERROR_CONDITION(!actual_argument.is<Nodecl::FortranActualArgument>(), "Unexpected tree", 0); Nodecl::NodeclBase arg = actual_argument.as<Nodecl::FortranActualArgument>().get_argument(); TL::Symbol array_sym = ::fortran_data_ref_get_symbol(arg.get_internal_nodecl()); ERROR_CONDITION( !(array_sym.get_type().is_fortran_array() && array_sym.is_allocatable()) && !(array_sym.get_type().is_pointer() && array_sym.get_type().points_to().is_fortran_array()), "The argument of 'ompss_opencl_deallocate' intrinsic must be " "an allocatable array or a pointer to an array\n", 0); // Replace the current intrinsic call by a call to the Nanos++ API TL::Symbol ptr_of_arr_sym = get_function_ptr_of(array_sym, expr_stmt.retrieve_context()); TL::Source new_function_call; new_function_call << "CALL NANOS_OPENCL_DEALLOCATE_FORTRAN(" << ptr_of_arr_sym.get_name() << "("<< as_expression(arg) << "))\n" ; expr_stmt.replace(new_function_call.parse_statement(expr_stmt)); }
static TL::Symbol create_initializer_function_fortran( OpenMP::Reduction* red, TL::Type reduction_type, Nodecl::NodeclBase construct) { std::string fun_name; { std::stringstream ss; ss << "nanos_ini_" << red << "_" << reduction_type.get_internal_type() << "_" << simple_hash_str(construct.get_filename().c_str()); fun_name = ss.str(); } Nodecl::NodeclBase initializer = red->get_initializer().shallow_copy(); TL::Type omp_out_type = reduction_type, omp_ori_type = reduction_type; // These sources are only used in array reductions TL::Source omp_out_extra_attributes, extra_stuff_array_red; if (reduction_type.is_array()) { Source dims_descr; TL::Type t = reduction_type; int rank = 0; if (t.is_fortran_array()) { rank = t.fortran_rank(); } dims_descr << "("; omp_out_extra_attributes << ", POINTER, DIMENSION("; int i; for (i = 0; i < rank; i++) { if (i != 0) { dims_descr << ","; omp_out_extra_attributes << ","; } dims_descr << "LBOUND(omp_orig, DIM = " << (rank - i) << ")" << ":" << "UBOUND(omp_orig, DIM = " << (rank - i) << ")" ; omp_out_extra_attributes << ":"; t = t.array_element(); } dims_descr << ")"; omp_out_extra_attributes << ")"; omp_out_type = t; extra_stuff_array_red << "ALLOCATE(omp_out" << dims_descr <<")\n"; } Source src; src << "SUBROUTINE " << fun_name << "(omp_out, omp_orig)\n" << "IMPLICIT NONE\n" << as_type(omp_out_type) << omp_out_extra_attributes << " :: omp_out\n" << as_type(omp_ori_type) << " :: omp_orig\n" << extra_stuff_array_red << "omp_out = " << as_expression(initializer) << "\n" << "END SUBROUTINE " << fun_name << "\n" ; TL::Scope global_scope = construct.retrieve_context().get_global_scope(); Nodecl::NodeclBase function_code = src.parse_global(global_scope); TL::Symbol function_sym = global_scope.get_symbol_from_name(fun_name); ERROR_CONDITION(!function_sym.is_valid(), "Symbol %s not found", fun_name.c_str()); // As the initializer function is needed during the instantiation of // the task, this function should be inserted before the construct Nodecl::Utils::prepend_to_enclosing_top_level_location(construct, function_code); return function_sym; }
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); }
static void handle_ompss_opencl_allocate_intrinsic( Nodecl::FunctionCall function_call, std::map<std::pair<TL::Type, std::pair<int, bool> > , Symbol> &declared_ocl_allocate_functions, Nodecl::NodeclBase expr_stmt) { Nodecl::List arguments = function_call.get_arguments().as<Nodecl::List>(); ERROR_CONDITION(arguments.size() != 1, "More than one argument in 'ompss_opencl_allocate' call\n", 0); Nodecl::NodeclBase actual_argument = arguments[0]; ERROR_CONDITION(!actual_argument.is<Nodecl::FortranActualArgument>(), "Unexpected tree\n", 0); Nodecl::NodeclBase arg = actual_argument.as<Nodecl::FortranActualArgument>().get_argument(); ERROR_CONDITION(!arg.is<Nodecl::ArraySubscript>(), "Unreachable code\n", 0); Nodecl::NodeclBase subscripted = arg.as<Nodecl::ArraySubscript>().get_subscripted(); TL::Symbol subscripted_symbol = ::fortran_data_ref_get_symbol(subscripted.get_internal_nodecl()); ERROR_CONDITION( !(subscripted_symbol.get_type().is_fortran_array() && subscripted_symbol.is_allocatable()) && !(subscripted_symbol.get_type().is_pointer() && subscripted_symbol.get_type().points_to().is_fortran_array()), "The argument of 'ompss_opencl_allocate' intrinsic must be " "an allocatable array or a pointer to an array with all its bounds specified\n", 0); TL::Type array_type; int num_dimensions; bool is_allocatable; if (subscripted_symbol.is_allocatable()) { array_type = subscripted_symbol.get_type(); num_dimensions = subscripted_symbol.get_type().get_num_dimensions(); is_allocatable = true; } else { array_type = subscripted_symbol.get_type().points_to(); num_dimensions = array_type.get_num_dimensions(); is_allocatable = false; } TL::Type element_type = array_type; while (element_type.is_array()) { element_type = element_type.array_element(); } ERROR_CONDITION(!array_type.is_array(), "This type should be an array type", 0); std::pair<TL::Type, std::pair<int, bool> > key = std::make_pair(element_type, std::make_pair(num_dimensions, is_allocatable)); std::map<std::pair<TL::Type, std::pair<int, bool> > , Symbol>::iterator it_new_fun = declared_ocl_allocate_functions.find(key); // Reuse the auxiliar function if it already exists Symbol new_function_sym; if (it_new_fun != declared_ocl_allocate_functions.end()) { new_function_sym = it_new_fun->second; } else { new_function_sym = create_new_function_opencl_allocate( expr_stmt, subscripted_symbol, element_type, num_dimensions, is_allocatable); declared_ocl_allocate_functions[key] = new_function_sym; } // Replace the current intrinsic call by a call to the new function TL::Source actual_arg_array; Nodecl::NodeclBase subscripted_lvalue = subscripted.shallow_copy(); subscripted_lvalue.set_type(subscripted_symbol.get_type().no_ref().get_lvalue_reference_to()); actual_arg_array << as_expression(subscripted_lvalue); TL::Source actual_arg_bounds; Nodecl::List subscripts = arg.as<Nodecl::ArraySubscript>().get_subscripts().as<Nodecl::List>(); for (Nodecl::List::reverse_iterator it = subscripts.rbegin(); it != subscripts.rend(); it++) { Nodecl::NodeclBase subscript = *it, lower, upper; if (it != subscripts.rbegin()) actual_arg_bounds << ", "; if (subscript.is<Nodecl::Range>()) { lower = subscript.as<Nodecl::Range>().get_lower(); upper = subscript.as<Nodecl::Range>().get_upper(); } else { lower = nodecl_make_integer_literal( fortran_get_default_integer_type(), const_value_get_signed_int(1), make_locus("", 0, 0)); upper = subscript; } actual_arg_bounds << as_expression(lower) << "," << as_expression(upper); } TL::Source new_function_call; new_function_call << "CALL " << as_symbol(new_function_sym) << "(" << actual_arg_array << ", " << actual_arg_bounds << ")\n" ; expr_stmt.replace(new_function_call.parse_statement(expr_stmt)); }
void LoweringVisitor::reduction_initialization_code( OutlineInfo& outline_info, Nodecl::NodeclBase ref_tree, Nodecl::NodeclBase construct) { ERROR_CONDITION(ref_tree.is_null(), "Invalid tree", 0); if (!Nanos::Version::interface_is_at_least("master", 5023)) { running_error("%s: error: a newer version of Nanos++ (>=5023) is required for reductions support\n", construct.get_locus_str().c_str()); } TL::ObjectList<OutlineDataItem*> reduction_items = outline_info.get_data_items().filter( predicate(lift_pointer(functor(&OutlineDataItem::is_reduction)))); ERROR_CONDITION (reduction_items.empty(), "No reductions to process", 0); Source result; Source reduction_declaration, thread_initializing_reduction_info, thread_fetching_reduction_info; result << reduction_declaration << "{" << as_type(get_bool_type()) << " red_single_guard;" << "nanos_err_t err;" << "err = nanos_enter_sync_init(&red_single_guard);" << "if (err != NANOS_OK)" << "nanos_handle_error(err);" << "if (red_single_guard)" << "{" << "int nanos_num_threads = nanos_omp_get_num_threads();" << thread_initializing_reduction_info << "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);" << thread_fetching_reduction_info << "}" << "}" ; for (TL::ObjectList<OutlineDataItem*>::iterator it = reduction_items.begin(); it != reduction_items.end(); it++) { std::string nanos_red_name = "nanos_red_" + (*it)->get_symbol().get_name(); std::pair<OpenMP::Reduction*, TL::Type> reduction_info = (*it)->get_reduction_info(); OpenMP::Reduction* reduction = reduction_info.first; TL::Type reduction_type = reduction_info.second; if (reduction_type.is_any_reference()) reduction_type = reduction_type.references_to(); TL::Type reduction_element_type = reduction_type; if (IS_FORTRAN_LANGUAGE) { while (reduction_element_type.is_fortran_array()) reduction_element_type = reduction_element_type.array_element(); } else { while (reduction_element_type.is_array()) reduction_element_type = reduction_element_type.array_element(); } Source element_size; if (IS_FORTRAN_LANGUAGE) { if (reduction_type.is_fortran_array()) { // We need to parse this bit in Fortran Source number_of_bytes; number_of_bytes << "SIZE(" << (*it)->get_symbol().get_name() << ") * " << reduction_element_type.get_size(); element_size << as_expression(number_of_bytes.parse_expression(construct)); } else { element_size << "sizeof(" << as_type(reduction_type) << ")"; } } else { element_size << "sizeof(" << as_type(reduction_type) << ")"; } reduction_declaration << "nanos_reduction_t* " << nanos_red_name << ";" ; Source allocate_private_buffer, cleanup_code; Source num_scalars; TL::Symbol basic_reduction_function, vector_reduction_function; create_reduction_function(reduction, construct, reduction_type, basic_reduction_function, vector_reduction_function); (*it)->reduction_set_basic_function(basic_reduction_function); thread_initializing_reduction_info << "err = nanos_malloc((void**)&" << nanos_red_name << ", sizeof(nanos_reduction_t), " << "\"" << construct.get_filename() << "\", " << construct.get_line() << ");" << "if (err != NANOS_OK)" << "nanos_handle_error(err);" << nanos_red_name << "->original = (void*)" << (reduction_type.is_array() ? "" : "&") << (*it)->get_symbol().get_name() << ";" << allocate_private_buffer << nanos_red_name << "->vop = " << (vector_reduction_function.is_valid() ? as_symbol(vector_reduction_function) : "0") << ";" << nanos_red_name << "->bop = (void(*)(void*,void*,int))" << as_symbol(basic_reduction_function) << ";" << nanos_red_name << "->element_size = " << element_size << ";" << nanos_red_name << "->num_scalars = " << num_scalars << ";" << cleanup_code << "err = nanos_register_reduction(" << nanos_red_name << ");" << "if (err != NANOS_OK)" << "nanos_handle_error(err);" ; if (IS_C_LANGUAGE || IS_CXX_LANGUAGE) { if (reduction_type.is_array()) { num_scalars << "sizeof(" << as_type(reduction_type) << ") / sizeof(" << as_type(reduction_element_type) <<")"; } else { num_scalars << "1"; } allocate_private_buffer << "err = nanos_malloc(&" << nanos_red_name << "->privates, sizeof(" << as_type(reduction_type) << ") * nanos_num_threads, " << "\"" << construct.get_filename() << "\", " << construct.get_line() << ");" << "if (err != NANOS_OK)" << "nanos_handle_error(err);" << nanos_red_name << "->descriptor = " << nanos_red_name << "->privates;" << "rdv_" << (*it)->get_field_name() << " = (" << as_type( (*it)->get_private_type().get_pointer_to() ) << ")" << nanos_red_name << "->privates;" ; thread_fetching_reduction_info << "err = nanos_reduction_get(&" << nanos_red_name << ", " << (reduction_type.is_array() ? "" : "&") << (*it)->get_symbol().get_name() << ");" << "if (err != NANOS_OK)" << "nanos_handle_error(err);" << "rdv_" << (*it)->get_field_name() << " = (" << as_type( (*it)->get_private_type().get_pointer_to() ) << ")" << nanos_red_name << "->privates;" ; cleanup_code << nanos_red_name << "->cleanup = nanos_free0;" ; } else if (IS_FORTRAN_LANGUAGE) { Type private_reduction_vector_type; Source extra_dims; { TL::Type t = (*it)->get_symbol().get_type().no_ref(); int rank = 0; if (t.is_fortran_array()) { rank = t.fortran_rank(); } if (rank != 0) { // We need to parse this bit in Fortran Source size_call; size_call << "SIZE(" << (*it)->get_symbol().get_name() << ")"; num_scalars << as_expression(size_call.parse_expression(construct)); } else { num_scalars << "1"; } private_reduction_vector_type = fortran_get_n_ranked_type_with_descriptor( get_void_type(), rank + 1, construct.retrieve_context().get_decl_context()); int i; for (i = 0; i < rank; i++) { Source lbound_src; lbound_src << "LBOUND(" << (*it)->get_symbol().get_name() << ", DIM = " << (rank - i) << ")"; Source ubound_src; ubound_src << "UBOUND(" << (*it)->get_symbol().get_name() << ", DIM = " << (rank - i) << ")"; extra_dims << "[" << as_expression(lbound_src.parse_expression(construct)) << ":" << as_expression(ubound_src.parse_expression(construct)) << "]"; t = t.array_element(); } } allocate_private_buffer << "@FORTRAN_ALLOCATE@((*rdv_" << (*it)->get_field_name() << ")[0:(nanos_num_threads-1)]" << extra_dims <<");" << nanos_red_name << "->privates = &(*rdv_" << (*it)->get_field_name() << ");" << "err = nanos_malloc(&" << nanos_red_name << "->descriptor, sizeof(" << as_type(private_reduction_vector_type) << "), " << "\"" << construct.get_filename() << "\", " << construct.get_line() << ");" << "if (err != NANOS_OK)" << "nanos_handle_error(err);" << "err = nanos_memcpy(" << nanos_red_name << "->descriptor, " "&rdv_" << (*it)->get_field_name() << ", sizeof(" << as_type(private_reduction_vector_type) << "));" << "if (err != NANOS_OK)" << "nanos_handle_error(err);" ; thread_fetching_reduction_info << "err = nanos_reduction_get(&" << nanos_red_name << ", &" << (*it)->get_symbol().get_name() << ");" << "if (err != NANOS_OK)" << "nanos_handle_error(err);" << "err = nanos_memcpy(" << "&rdv_" << (*it)->get_field_name() << "," << nanos_red_name << "->descriptor, " << "sizeof(" << as_type(private_reduction_vector_type) << "));" << "if (err != NANOS_OK)" << "nanos_handle_error(err);" ; TL::Symbol reduction_cleanup = create_reduction_cleanup_function(reduction, construct); cleanup_code << nanos_red_name << "->cleanup = " << as_symbol(reduction_cleanup) << ";" ; } else { internal_error("Code unreachable", 0); } } FORTRAN_LANGUAGE() { Source::source_language = SourceLanguage::C; } ref_tree.replace(result.parse_statement(ref_tree)); FORTRAN_LANGUAGE() { Source::source_language = SourceLanguage::Current; } }