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));
}
void Core::collapse_check_loop(TL::PragmaCustomStatement construct)
{
TL::PragmaCustomClause collapse = construct.get_pragma_line().get_clause("collapse");
if (!collapse.is_defined())
return;
TL::ObjectList<Nodecl::NodeclBase> expr_list = collapse.get_arguments_as_expressions(construct);
if (expr_list.size() != 1)
{
error_printf("%s: error: collapse clause needs exactly one argument\n",
locus_to_str(construct.get_locus()));
return;
}
Nodecl::NodeclBase expr = expr_list[0];
if (!expr.is_constant()
|| !is_any_int_type(expr.get_type().get_internal_type()))
{
error_printf("%s: error: collapse clause requires an integer constant expression\n",
locus_to_str(construct.get_locus()));
return;
}
const_value_t* cval = expr.get_constant();
if (!const_value_is_one(cval))
{
error_printf("%s: error: only collapse(1) is supported\n",
locus_to_str(construct.get_locus()));
return;
}
}
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();
}
bool ArrayAccessInfoVisitor::unhandled_node( const Nodecl::NodeclBase& n )
{
std::cerr << "Unhandled node while parsing Array Subscript '"
<< codegen_to_str( n.get_internal_nodecl( ),
nodecl_retrieve_context( n.get_internal_nodecl( ) ) )
<< "' of type '" << ast_print_node_type( n.get_kind( ) ) << "'" << std::endl;
return false;
}
Nodecl::NodeclVisitor<void>::Ret NeonVectorBackend::unhandled_node(const Nodecl::NodeclBase& n)
{
internal_error("NEON Backend: Unknown node %s at %s.",
ast_print_node_type(n.get_kind()),
locus_to_str(n.get_locus()));
return Ret();
}
void VectorizerVisitorExpression::visit(const Nodecl::ArraySubscript& n)
{
// Computing new vector type
TL::Type vector_type = n.get_type();
if (vector_type.is_lvalue_reference())
{
vector_type = vector_type.references_to();
}
vector_type = get_qualified_vector_to(vector_type, _vector_length);
TL::Type basic_type = n.get_type();
if (basic_type.is_lvalue_reference())
{
basic_type = basic_type.references_to();
}
// Vector Load
if (Vectorizer::_analysis_info->is_adjacent_access(
Vectorizer::_analysis_scopes->back(),
n))
{
const Nodecl::VectorLoad vector_load =
Nodecl::VectorLoad::make(
Nodecl::Reference::make(
Nodecl::ParenthesizedExpression::make(
n.shallow_copy(),
basic_type,
n.get_locus()),
basic_type.get_pointer_to(),
n.get_locus()),
vector_type,
n.get_locus());
n.replace(vector_load);
}
else // Vector Gather
{
const Nodecl::NodeclBase base = n.get_subscripted();
const Nodecl::List subscripts = n.get_subscripts().as<Nodecl::List>();
ERROR_CONDITION(subscripts.size() > 1,
"Vectorizer: Gather on multidimensional array is not supported yet!", 0);
std::cerr << "Gather: " << n.prettyprint() << "\n";
Nodecl::NodeclBase strides = *subscripts.begin();
walk(strides);
const Nodecl::VectorGather vector_gather =
Nodecl::VectorGather::make(
base.shallow_copy(),
strides,
vector_type,
n.get_locus());
n.replace(vector_gather);
}
}
Nodecl::NodeclVisitor<void>::Ret VectorizerVisitorFunction::unhandled_node(const Nodecl::NodeclBase& n)
{
std::cerr << "Function Visitor: Unknown node "
<< ast_print_node_type(n.get_kind())
<< " at " << n.get_locus()
<< std::endl;
return Ret();
}
void VectorizerVisitorPostprocessor::visit(const Nodecl::ObjectInit& n)
{
TL::Symbol sym = n.get_symbol();
Nodecl::NodeclBase init = sym.get_value();
if(!init.is_null())
{
walk(init);
}
}
Type Type::get_array_to_with_descriptor(Nodecl::NodeclBase lower_bound, Nodecl::NodeclBase upper_bound, Scope sc)
{
type_t* result_type = this->_type_info;
const decl_context_t* decl_context = sc.get_decl_context();
type_t* array_to = get_array_type_bounds_with_descriptor(result_type,
lower_bound.get_internal_nodecl(),
upper_bound.get_internal_nodecl(),
decl_context);
return Type(array_to);
}
void SSEVectorLegalization::visit(const Nodecl::ObjectInit& node)
{
TL::Source intrin_src;
TL::Symbol sym = node.get_symbol();
fix_mask_symbol(sym);
// Vectorizing initialization
Nodecl::NodeclBase init = sym.get_value();
if (!init.is_null())
{
walk(init);
}
}
void NeonVectorBackend::visit(const Nodecl::ObjectInit& n)
{
TL::Source intrin_src;
if(n.has_symbol())
{
TL::Symbol sym = n.get_symbol();
// Vectorizing initialization
Nodecl::NodeclBase init = sym.get_value();
if(!init.is_null())
{
walk(init);
}
}
}
void SimdVisitor::visit(const Nodecl::OpenMP::Simd& simd_node)
{
Nodecl::NodeclBase for_statement = simd_node.get_statement();
// Vectorize for
Nodecl::NodeclBase epilog =
_vectorizer.vectorize(for_statement.as<Nodecl::ForStatement>(),
_device_name, _vector_length, NULL);
// Add epilog
if (!epilog.is_null())
{
simd_node.append_sibling(epilog);
}
// Remove Simd node
simd_node.replace(for_statement);
}
Type Type::get_array_to(Nodecl::NodeclBase array_expr, Scope sc)
{
type_t* result_type = this->_type_info;
const decl_context_t* decl_context = sc.get_decl_context();
type_t* array_to = get_array_type(result_type, array_expr.get_internal_nodecl(), decl_context);
return Type(array_to);
}
std::string as_statement(const Nodecl::NodeclBase& n)
{
std::stringstream ss;
ss << nodecl_stmt_to_source(n.get_internal_nodecl());
if (IS_FORTRAN_LANGUAGE)
ss << "\n";
return ss.str();
}
std::string statement_placeholder(Nodecl::NodeclBase& placeholder)
{
std::stringstream ss;
ss << "@STATEMENT-PH::" << placeholder.get_internal_tree_address() << "@";
if (IS_FORTRAN_LANGUAGE)
ss << "\n";
return ss.str();
}
TL::Symbol LoweringVisitor::create_reduction_cleanup_function(OpenMP::Reduction* red, Nodecl::NodeclBase construct)
{
if (IS_C_LANGUAGE || IS_CXX_LANGUAGE)
{
internal_error("Currently only valid in Fortran", 0);
}
reduction_map_t::iterator it = _reduction_cleanup_map.find(red);
if (it != _reduction_cleanup_map.end())
{
return it->second;
}
std::string fun_name;
{
std::stringstream ss;
ss << "nanos_cleanup_" << red << "_" << simple_hash_str(construct.get_filename().c_str());
fun_name = ss.str();
}
Source src;
src << "SUBROUTINE " << fun_name << "(X)\n"
<< as_type(red->get_type()) << ", POINTER :: X(:)\n"
<< "DEALLOCATE(X)\n"
<< "END SUBROUTINE\n"
;
Nodecl::NodeclBase function_code = src.parse_global(construct);
TL::Symbol function_sym = ReferenceScope(construct).get_scope().get_symbol_from_name(fun_name);
ERROR_CONDITION(!function_sym.is_valid(), "Symbol %s not found", fun_name.c_str());
_reduction_cleanup_map[red] = function_sym;
Nodecl::Utils::append_to_enclosing_top_level_location(construct, function_code);
Nodecl::Utils::Fortran::append_used_modules(construct.retrieve_context(), function_sym.get_related_scope());
return function_sym;
}
Type Type::get_array_to_with_region(Nodecl::NodeclBase lower_bound,
Nodecl::NodeclBase upper_bound,
Nodecl::NodeclBase region_lower_bound,
Nodecl::NodeclBase region_upper_bound,
Scope sc)
{
type_t* result_type = this->_type_info;
const decl_context_t* decl_context = sc.get_decl_context();
// Make the range of the region
Nodecl::NodeclBase range = Nodecl::Range::make(
region_lower_bound,
region_upper_bound,
const_value_to_nodecl(const_value_get_one(4, 1)),
region_lower_bound.get_type(),
region_lower_bound.get_locus());
type_t* array_to = get_array_type_bounds_with_regions(
result_type,
lower_bound.get_internal_nodecl(),
upper_bound.get_internal_nodecl(),
decl_context,
range.get_internal_nodecl(),
decl_context);
return array_to;
}
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;
}
void KNCVectorLegalization::visit(const Nodecl::VectorLoad& n)
{
const Nodecl::NodeclBase rhs = n.get_rhs();
const Nodecl::NodeclBase mask = n.get_mask();
const Nodecl::List flags = n.get_flags().as<Nodecl::List>();
walk(rhs);
walk(mask);
walk(flags);
bool explicitly_aligned = !flags.find_first<Nodecl::AlignedFlag>().is_null();
// Turn unaligned load into gather
if (!explicitly_aligned
&& (_prefer_gather_scatter ||
(_prefer_mask_gather_scatter && !mask.is_null())))
{
VECTORIZATION_DEBUG()
{
fprintf(stderr, "KNC Legalization: Turn unaligned load '%s' into "\
"adjacent gather\n", rhs.prettyprint().c_str());
}
bool NodeclStaticInfo::is_simd_aligned_access( const Nodecl::NodeclBase& n,
const TL::ObjectList<Nodecl::NodeclBase>* suitable_expressions,
int unroll_factor, int alignment ) const
{
if( !n.is<Nodecl::ArraySubscript>( ) )
{
std::cerr << "warning: returning false for is_simd_aligned_access when asking for nodecl '"
<< n.prettyprint( ) << "' which is not an array subscript" << std::endl;
return false;
}
bool result = false;
Nodecl::NodeclBase subscripted = n.as<Nodecl::ArraySubscript>( ).get_subscripted( );
int type_size = subscripted.get_type().basic_type().get_size();
SuitableAlignmentVisitor sa_v( _induction_variables, suitable_expressions, unroll_factor, type_size, alignment );
int subscript_alignment = sa_v.walk( n );
if( (subscript_alignment % alignment) == 0 )
result = true;
return result;
}
void VectorizerVectorReduction::vectorize_reduction(const TL::Symbol& scalar_symbol,
TL::Symbol& vector_symbol,
const Nodecl::NodeclBase& reduction_initializer,
const std::string& reduction_name,
const TL::Type& reduction_type,
Nodecl::List& pre_nodecls,
Nodecl::List& post_nodecls)
{
// Step1: ADD REDUCTION SYMBOLS
vector_symbol.set_value(Nodecl::VectorPromotion::make(
reduction_initializer.shallow_copy(),
vector_symbol.get_type()));
// Add new ObjectInit with the initialization
Nodecl::ObjectInit reduction_object_init =
Nodecl::ObjectInit::make(vector_symbol);
pre_nodecls.append(reduction_object_init);
// Step2: ADD VECTOR REDUCTION INSTRUCTIONS
if(reduction_name.compare("+") == 0)
{
Nodecl::ExpressionStatement post_reduction_stmt =
Nodecl::ExpressionStatement::make(
Nodecl::VectorReductionAdd::make(
scalar_symbol.make_nodecl(true),
vector_symbol.make_nodecl(true),
scalar_symbol.get_type()));
post_nodecls.append(post_reduction_stmt);
}
else if (reduction_name.compare("-") == 0)
{
Nodecl::ExpressionStatement post_reduction_stmt =
Nodecl::ExpressionStatement::make(
Nodecl::VectorReductionMinus::make(
scalar_symbol.make_nodecl(true),
vector_symbol.make_nodecl(true),
scalar_symbol.get_type()));
post_nodecls.append(post_reduction_stmt);
}
}
void build_empty_body_for_function(
TL::Symbol function_symbol,
Nodecl::NodeclBase &function_code,
Nodecl::NodeclBase &empty_stmt)
{
empty_stmt = Nodecl::EmptyStatement::make(make_locus("", 0, 0));
Nodecl::List stmt_list = Nodecl::List::make(empty_stmt);
if (IS_C_LANGUAGE || IS_CXX_LANGUAGE)
{
Nodecl::CompoundStatement compound_statement =
Nodecl::CompoundStatement::make(stmt_list,
/* destructors */ Nodecl::NodeclBase::null(),
make_locus("", 0, 0));
stmt_list = Nodecl::List::make(compound_statement);
}
Nodecl::NodeclBase context = Nodecl::Context::make(
stmt_list,
function_symbol.get_related_scope(), make_locus("", 0, 0));
function_symbol.get_internal_symbol()->defined = 1;
if (function_symbol.is_dependent_function())
{
function_code = Nodecl::TemplateFunctionCode::make(context,
// Initializers
Nodecl::NodeclBase::null(),
function_symbol,
make_locus("", 0, 0));
}
else
{
function_code = Nodecl::FunctionCode::make(context,
// Initializers
Nodecl::NodeclBase::null(),
function_symbol,
make_locus("", 0, 0));
}
function_symbol.get_internal_symbol()->entity_specs.function_code = function_code.get_internal_nodecl();
}
bool NodeclStaticInfo::is_suitable_expression( const Nodecl::NodeclBase& n,
const TL::ObjectList<Nodecl::NodeclBase>* suitable_expressions,
int unroll_factor, int alignment, int& vector_size_module ) const
{
bool result = false;
int type_size = n.get_type().basic_type().get_size();
SuitableAlignmentVisitor sa_v( _induction_variables, suitable_expressions, unroll_factor, type_size, alignment );
int subscript_alignment = sa_v.walk( n );
printf("SUBSCRIPT ALIGNMENT %d\n", subscript_alignment);
vector_size_module = ( ( subscript_alignment == -1 ) ? subscript_alignment :
subscript_alignment % alignment );
if( vector_size_module == 0 )
result = true;
return result;
}
int SuitableAlignmentVisitor::visit( const Nodecl::Symbol& n )
{
if (is_suitable_expression(n))
{
return 0;
}
else if( n.is_constant( ) )
{
int value = const_value_cast_to_signed_int( n.get_constant( )) * _type_size;
if(is_suitable_constant(value))
return 0;
else
return value;
}
else if( Utils::induction_variable_list_contains_variable( _induction_variables, n ) )
{
Utils::InductionVariableData* iv = Utils::get_induction_variable_from_list( _induction_variables, n );
Nodecl::Utils::ReduceExpressionVisitor v;
Nodecl::NodeclBase lb = iv->get_lb( ).shallow_copy( );
v.walk( lb );
if( lb.is_constant( ) )
{
Nodecl::NodeclBase incr = iv->get_increment( ).shallow_copy( );
v.walk( incr );
if( incr.is_constant( ) )
{
return (const_value_cast_to_signed_int( lb.get_constant( ) )
+ ( const_value_cast_to_signed_int( incr.get_constant( ) )
* _unroll_factor)) * _type_size;
}
}
}
return -1;
}
void Symbol::set_value(Nodecl::NodeclBase n)
{
_symbol->value = n.get_internal_nodecl();
}
void LoweringVisitor::register_reductions(
Nodecl::NodeclBase construct, OutlineInfo& outline_info, TL::Source& src)
{
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();
ERROR_CONDITION(!Nanos::Version::interface_is_at_least("task_reduction", 1001),
"The version of the runtime being used does not support task reductions", 0);
TL::Symbol reduction_item = (*it)->get_symbol();
ERROR_CONDITION(reduction_type.is_array()
&& (IS_C_LANGUAGE || IS_CXX_LANGUAGE)
&& !Nanos::Version::interface_is_at_least("task_reduction", 1002),
"The version of the runtime being used does not support array reductions in C/C++", 0);
// Note that at this point all the reduction must be registered.
// For C/C++ array reductions, the registered_reduction type is the
// 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));
ERROR_CONDITION(task_red_info == _task_reductions_map.end(),
"Unregistered task reduction\n", 0);
TL::Symbol reduction_function = task_red_info->second._reducer;
TL::Symbol reduction_function_original_var = task_red_info->second._reducer_orig_var;
TL::Symbol initializer_function = task_red_info->second._initializer;
// Common case: the runtime will host the private copies of the list item
if (!(IS_FORTRAN_LANGUAGE && reduction_type.is_array()))
{
// Array Reductions in C/C++ are defined over the elements of the array
TL::Source reduction_size_src_opt;
TL::Type element_type = registered_reduction_type;
reduction_size_src_opt << "sizeof(" << as_type(reduction_type) <<"),";
TL::Source item_address =
(reduction_item.get_type().is_pointer() ? "" : "&") + (*it)->get_field_name();
src
<< "nanos_err = nanos_task_reduction_register("
<< "(void *) " << item_address << "," // object address
<< reduction_size_src_opt // whole reduction size
<< "sizeof(" << as_type(element_type) << ")," // element size
<< "(void (*)(void *, void *)) &"
<< initializer_function.get_name() << "," // initializer
<< "(void (*)(void *, void *)) &"
<< reduction_function.get_name() << ");" // reducer
;
}
else
{
// Specific case for Fortran Array Reductions: the runtime will
// host a private array descriptor for each thread. Later, in
// the initializer function, this array descriptors will be
// initialized and their array storage will be allocated
TL::Source target_address;
size_t size_array_descriptor =
fortran_size_of_array_descriptor(
fortran_get_rank0_type(reduction_type.get_internal_type()),
fortran_get_rank_of_type(reduction_type.get_internal_type()));
if (reduction_type.array_requires_descriptor())
{
TL::Symbol ptr_of_sym = get_function_ptr_of(reduction_item, construct.retrieve_context());
target_address << ptr_of_sym.get_name() << "( " << (*it)->get_symbol().get_name() << ")";
}
else
{
target_address << "(void *) &" << (*it)->get_field_name();
}
src
<< "nanos_err = nanos_task_fortran_array_reduction_register("
<< target_address << "," // Address to the array descriptor
<< "(void *) & " << (*it)->get_field_name() << "," // Address to the storage
<< size_array_descriptor << "," // size
<< "(void (*)(void *, void *)) &"
<< initializer_function.get_name() << "," // initializer
<< "(void (*)(void *, void *)) &"
<< reduction_function.get_name() << "," // reducer
<< "(void (*)(void *, void *)) &"
<< reduction_function_original_var.get_name() << ");" // reducer ori
;
}
}
}
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);
}
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;
}
static TL::Symbol create_initializer_function_c(
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 function_body;
Source src;
src << "void " << fun_name << "("
<< as_type(reduction_type.no_ref().get_lvalue_reference_to()) << " omp_priv,"
<< as_type(reduction_type.no_ref().get_lvalue_reference_to()) << " omp_orig)"
<< "{"
<< statement_placeholder(function_body)
<< "}"
;
Nodecl::NodeclBase function_code = src.parse_global(construct.retrieve_context().get_global_scope());
TL::Scope inside_function = ReferenceScope(function_body).get_scope();
TL::Symbol param_omp_priv = inside_function.get_symbol_from_name("omp_priv");
ERROR_CONDITION(!param_omp_priv.is_valid(), "Symbol omp_priv not found", 0);
TL::Symbol param_omp_orig = inside_function.get_symbol_from_name("omp_orig");
ERROR_CONDITION(!param_omp_orig.is_valid(), "Symbol omp_orig not found", 0);
TL::Symbol function_sym = inside_function.get_symbol_from_name(fun_name);
ERROR_CONDITION(!function_sym.is_valid(), "Symbol %s not found", fun_name.c_str());
Nodecl::NodeclBase initializer = red->get_initializer().shallow_copy();
if (initializer.is<Nodecl::StructuredValue>())
{
Nodecl::StructuredValue structured_value = initializer.as<Nodecl::StructuredValue>();
if (structured_value.get_form().is<Nodecl::StructuredValueBracedImplicit>())
{
structured_value.set_form(Nodecl::StructuredValueCompoundLiteral::make());
}
}
Nodecl::Utils::SimpleSymbolMap translation_map;
translation_map.add_map(red->get_omp_priv(), param_omp_priv);
translation_map.add_map(red->get_omp_orig(), param_omp_orig);
Nodecl::NodeclBase new_initializer = Nodecl::Utils::deep_copy(initializer, inside_function, translation_map);
if (red->get_is_initialization())
{
// The original initializer was something like 'omp_priv = expr1', but the
// new_initializer only represents the lhs expression (in our example, expr1).
// For this reason we create manually an assignment expression.
Nodecl::NodeclBase param_omp_priv_ref = Nodecl::Symbol::make(param_omp_priv);
param_omp_priv_ref.set_type(param_omp_priv.get_type());
function_body.replace(
Nodecl::List::make(
Nodecl::ExpressionStatement::make(
Nodecl::Assignment::make(
param_omp_priv_ref,
new_initializer,
param_omp_priv_ref.get_type().no_ref())
)));
}
else
{
function_body.replace(
Nodecl::List::make(Nodecl::ExpressionStatement::make(new_initializer)));
}
// 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;
}
std::string as_expression(const Nodecl::NodeclBase& n)
{
ERROR_CONDITION (n.is_null(), "Cannot create a literal expression from a null node", 0);
return nodecl_expr_to_source(n.get_internal_nodecl());
}