Nodecl::NodeclVisitor<void>::Ret AVX2StrideVisitorConv::unhandled_node(const Nodecl::NodeclBase& node)
{
//printf("Unsupported %d: %s\n", _vector_num_elements, node.prettyprint().c_str());
if (node.get_type().is_vector())
{
Nodecl::NodeclBase new_node = node.shallow_copy().as<Nodecl::NodeclBase>();
new_node.set_type(TL::Type::get_int_type().get_vector_of_elements(
_vector_num_elements));
// TODO better
node.replace(new_node);
Nodecl::NodeclBase::Children children = node.children();
for(Nodecl::NodeclBase::Children::iterator it = children.begin();
it != children.end();
it ++)
{
walk(*it);
}
}
return Ret();
}
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));
}
Nodecl::NodeclBase handle_task_statements(
Nodecl::NodeclBase construct,
Nodecl::NodeclBase task_statements,
Nodecl::NodeclBase& task_placeholder, // Do not remove the reference
TL::Source &new_stmts_src, // It should be a const reference
const std::map<TL::Symbol, std::string> &reduction_symbols_map)
{
if (IS_FORTRAN_LANGUAGE)
Source::source_language = SourceLanguage::C;
Nodecl::NodeclBase new_statements = new_stmts_src.parse_statement(construct);
if (IS_FORTRAN_LANGUAGE)
Source::source_language = SourceLanguage::Current;
TL::Scope new_scope = ReferenceScope(task_placeholder).get_scope();
std::map<TL::Symbol, Nodecl::NodeclBase> reduction_symbol_to_nodecl_map;
for (std::map<TL::Symbol, std::string>::const_iterator it = reduction_symbols_map.begin();
it != reduction_symbols_map.end();
++it)
{
TL::Symbol reduction_sym = it->first;
std::string storage_name = it->second;
TL::Symbol storage_sym = new_scope.get_symbol_from_name(storage_name);
ERROR_CONDITION(!storage_sym.is_valid(), "This symbol is not valid", 0);
Nodecl::NodeclBase deref_storage = Nodecl::Dereference::make(
storage_sym.make_nodecl(/* set_ref_type */ true, storage_sym.get_locus()),
storage_sym.get_type().points_to());
reduction_symbol_to_nodecl_map[reduction_sym] = deref_storage;
}
ReplaceReductionSymbols visitor(reduction_symbol_to_nodecl_map);
Nodecl::NodeclBase copied_statements = task_statements.shallow_copy();
visitor.walk(copied_statements);
task_placeholder.replace(copied_statements);
return new_statements;
}
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::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::perform_partial_reduction(OutlineInfo& outline_info, Nodecl::NodeclBase ref_tree)
{
ERROR_CONDITION(ref_tree.is_null(), "Invalid tree", 0);
Source reduction_code;
TL::ObjectList<OutlineDataItem*> reduction_items = outline_info.get_data_items().filter(
predicate(lift_pointer(functor(&OutlineDataItem::is_reduction))));
if (!reduction_items.empty())
{
for (TL::ObjectList<OutlineDataItem*>::iterator it = reduction_items.begin();
it != reduction_items.end();
it++)
{
if (IS_C_LANGUAGE || IS_CXX_LANGUAGE)
{
if ((*it)->get_private_type().is_array())
{
reduction_code
<< "__builtin_memcpy(rdv_" << (*it)->get_field_name() << "[nanos_omp_get_thread_num()],"
<< "rdp_" << (*it)->get_symbol().get_name() << ","
<< " sizeof(" << as_type((*it)->get_private_type()) << "));"
;
}
else
{
reduction_code
<< "rdv_" << (*it)->get_field_name() << "[nanos_omp_get_thread_num()] "
<< "= rdp_" << (*it)->get_symbol().get_name() << ";"
;
}
}
else if (IS_FORTRAN_LANGUAGE)
{
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();
}
int i;
for (i = 0; i < rank; i++)
{
extra_dims << ":,";
}
}
reduction_code
<< "rdv_" << (*it)->get_field_name() << "( " << extra_dims << "nanos_omp_get_thread_num() ) = rdp_" << (*it)->get_symbol().get_name() << "\n"
;
}
else
{
internal_error("Code unreachable", 0);
}
}
}
ref_tree.replace(reduction_code.parse_statement(ref_tree));
}
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;
}
}
void LoweringVisitor::perform_partial_reduction_slicer(OutlineInfo& outline_info,
Nodecl::NodeclBase ref_tree,
Nodecl::Utils::SimpleSymbolMap*& symbol_map)
{
ERROR_CONDITION(ref_tree.is_null(), "Invalid tree", 0);
TL::ObjectList<OutlineDataItem*> reduction_items = outline_info.get_data_items().filter(
lift_pointer<bool, OutlineDataItem>(&OutlineDataItem::is_reduction));
if (!reduction_items.empty())
{
TL::ObjectList<Nodecl::NodeclBase> reduction_stmts;
Nodecl::Utils::SimpleSymbolMap* simple_symbol_map = new Nodecl::Utils::SimpleSymbolMap(symbol_map);
symbol_map = simple_symbol_map;
for (TL::ObjectList<OutlineDataItem*>::iterator it = reduction_items.begin();
it != reduction_items.end();
it++)
{
scope_entry_t* shared_symbol = (*it)->get_symbol().get_internal_symbol();
// We need this to avoid the original symbol be replaced
// incorrectly
scope_entry_t* shared_symbol_proxy = NEW0(scope_entry_t);
shared_symbol_proxy->symbol_name = UNIQUESTR_LITERAL("<<reduction-variable>>"); // Crude way to ensure it is replaced
shared_symbol_proxy->kind = shared_symbol->kind;
symbol_entity_specs_copy_from(shared_symbol_proxy, shared_symbol);
shared_symbol_proxy->decl_context = shared_symbol->decl_context;
shared_symbol_proxy->type_information = shared_symbol->type_information;
shared_symbol_proxy->locus = shared_symbol->locus;
simple_symbol_map->add_map( shared_symbol_proxy,
(*it)->reduction_get_shared_symbol_in_outline() );
Source reduction_code;
Nodecl::NodeclBase partial_reduction_code;
reduction_code
<< "{"
<< "nanos_lock_t* red_lock;"
<< "nanos_err_t nanos_err;"
<< "nanos_err = nanos_get_lock_address("
<< ((*it)->get_private_type().is_array() ? "" : "&")
<< as_symbol( shared_symbol_proxy ) << ", &red_lock);"
<< "if (nanos_err != NANOS_OK) nanos_handle_error(nanos_err);"
<< "nanos_err = nanos_set_lock(red_lock);"
<< "if (nanos_err != NANOS_OK) nanos_handle_error(nanos_err);"
<< statement_placeholder(partial_reduction_code)
<< "nanos_err = nanos_unset_lock(red_lock);"
<< "if (nanos_err != NANOS_OK) nanos_handle_error(nanos_err);"
<< "}"
;
FORTRAN_LANGUAGE()
{
Source::source_language = SourceLanguage::C;
}
Nodecl::NodeclBase statement = reduction_code.parse_statement(ref_tree);
FORTRAN_LANGUAGE()
{
Source::source_language = SourceLanguage::Current;
}
ERROR_CONDITION(!statement.is<Nodecl::List>(), "Expecting a list", 0);
reduction_stmts.append(statement.as<Nodecl::List>()[0]);
TL::Type elemental_type = (*it)->get_private_type();
while (elemental_type.is_array())
elemental_type = elemental_type.array_element();
Source partial_reduction_code_src;
if (IS_C_LANGUAGE || IS_CXX_LANGUAGE)
{
partial_reduction_code_src
<< as_symbol( (*it)->reduction_get_basic_function() ) << "("
// This will be the reduction shared
<< ((*it)->get_private_type().is_array() ? "" : "&")
<< as_symbol( shared_symbol_proxy ) << ", "
// This will be the reduction private var
<< ((*it)->get_private_type().is_array() ? "" : "&")
<< as_symbol( (*it)->get_symbol() ) << ", "
<< ((*it)->get_private_type().is_array() ?
(
"sizeof(" + as_type( (*it)->get_private_type()) + ")"
"/ sizeof(" + as_type(elemental_type) + ")"
)
: "1")
<< ");"
;
}
else if (IS_FORTRAN_LANGUAGE)
{
// We use an ELEMENTAL call here
partial_reduction_code_src
<< "CALL " << as_symbol ( (*it)->reduction_get_basic_function() ) << "("
// This will be the reduction shared
<< as_symbol( shared_symbol_proxy ) << ", "
// This will be the reduction private var
<< as_symbol( (*it)->get_symbol() )
<< ")"
;
}
else
{
internal_error("Code unreachable", 0);
}
partial_reduction_code.replace(
partial_reduction_code_src.parse_statement(partial_reduction_code));
}
ref_tree.replace(
Nodecl::CompoundStatement::make(
Nodecl::List::make(reduction_stmts),
Nodecl::NodeclBase::null()
)
);
}
