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));
}
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;
}
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;
}
ReferenceScope::ReferenceScope(Nodecl::NodeclBase n)
: _scope(n.retrieve_context())
{
}
static Symbol create_new_function_opencl_allocate(
Nodecl::NodeclBase expr_stmt,
Symbol subscripted_symbol,
Type element_type,
int num_dimensions,
bool is_allocatable)
{
std::string alloca_or_pointer = is_allocatable ? "ALLOCATABLE" : "POINTER";
TL::Source dummy_arguments_bounds, dimension_attr, allocate_dims;
dimension_attr << "DIMENSION(";
for (int i = 1; i <= num_dimensions; ++i)
{
if (i != 1)
{
allocate_dims << ", ";
dummy_arguments_bounds <<", ";
dimension_attr << ", ";
}
dummy_arguments_bounds <<"LB" << i <<", " << "UB" << i;
dimension_attr << ":";
allocate_dims << "LB" << i << ":" << "UB" << i;
}
dimension_attr << ")";
size_t size_of_array_descriptor =
fortran_size_of_array_descriptor(
fortran_get_rank0_type(subscripted_symbol.get_type().get_internal_type()),
fortran_get_rank_of_type(subscripted_symbol.get_type().get_internal_type()));
TL::Source new_function_name;
new_function_name
<< "NANOX_OPENCL_ALLOCATE_INTERNAL_"
<< (ptrdiff_t) subscripted_symbol.get_internal_symbol()
;
Nodecl::NodeclBase nodecl_body;
TL::Source new_function;
new_function
<< "SUBROUTINE " << new_function_name << "(ARR, " << dummy_arguments_bounds << ")\n"
<< as_type(element_type) << ", " << dimension_attr << ", " << alloca_or_pointer << " :: ARR\n"
<< as_type(element_type) << ", " << dimension_attr << ", ALLOCATABLE :: TMP\n"
<< "INTEGER :: " << dummy_arguments_bounds << "\n"
<< "INTEGER :: ERR \n"
<< "ALLOCATE(TMP(" << allocate_dims << "))\n"
<< statement_placeholder(nodecl_body)
<< "DEALLOCATE(TMP)\n"
<< "END SUBROUTINE " << new_function_name << "\n"
;
Nodecl::NodeclBase function_code = new_function.parse_global(expr_stmt.retrieve_context().get_global_scope());
TL::Scope inside_function = ReferenceScope(nodecl_body).get_scope();
TL::Symbol new_function_sym = inside_function.get_symbol_from_name(strtolower(new_function_name.get_source().c_str()));
TL::Symbol arr_sym = inside_function.get_symbol_from_name("arr");
TL::Symbol tmp_sym = inside_function.get_symbol_from_name("tmp");
TL::Symbol ptr_of_arr_sym = get_function_ptr_of(arr_sym, inside_function);
TL::Symbol ptr_of_tmp_sym = get_function_ptr_of(tmp_sym, inside_function);
TL::Source aux;
aux
<< "ERR = NANOS_MEMCPY("
<< ptr_of_arr_sym.get_name() << "(ARR),"
<< ptr_of_tmp_sym.get_name() << "(TMP),"
<< "INT(" << size_of_array_descriptor << "," << type_get_size(get_ptrdiff_t_type()) << "))\n"
<< "CALL NANOS_OPENCL_ALLOCATE_FORTRAN("
<< "SIZEOF(TMP),"
<< ptr_of_arr_sym.get_name() << "(ARR))\n"
;
nodecl_body.replace(aux.parse_statement(inside_function));
Nodecl::Utils::prepend_to_enclosing_top_level_location(expr_stmt, function_code);
return new_function_sym;
}
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;
}
}
TL::Symbol LoweringVisitor::create_basic_reduction_function_fortran(OpenMP::Reduction* red, Nodecl::NodeclBase construct)
{
reduction_map_t::iterator it = _basic_reduction_map_openmp.find(red);
if (it != _basic_reduction_map_openmp.end())
{
return it->second;
}
std::string fun_name;
{
std::stringstream ss;
ss << "nanos_red_" << red << "_" << simple_hash_str(construct.get_filename().c_str());
fun_name = ss.str();
}
Nodecl::NodeclBase function_body;
Source src;
src << "SUBROUTINE " << fun_name << "(omp_out, omp_in, num_scalars)\n"
<< "IMPLICIT NONE\n"
<< as_type(red->get_type()) << " :: omp_out(num_scalars)\n"
<< as_type(red->get_type()) << " :: omp_in(num_scalars)\n"
<< "INTEGER, VALUE :: num_scalars\n"
<< "INTEGER :: I\n"
<< statement_placeholder(function_body) << "\n"
<< "END SUBROUTINE " << fun_name << "\n";
;
Nodecl::NodeclBase function_code = src.parse_global(construct);
TL::Scope inside_function = ReferenceScope(function_body).get_scope();
TL::Symbol param_omp_in = inside_function.get_symbol_from_name("omp_in");
ERROR_CONDITION(!param_omp_in.is_valid(), "Symbol omp_in not found", 0);
TL::Symbol param_omp_out = inside_function.get_symbol_from_name("omp_out");
ERROR_CONDITION(!param_omp_out.is_valid(), "Symbol omp_out 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());
TL::Symbol index = inside_function.get_symbol_from_name("i");
ERROR_CONDITION(!index.is_valid(), "Symbol %s not found", "i");
TL::Symbol num_scalars = inside_function.get_symbol_from_name("num_scalars");
ERROR_CONDITION(!num_scalars.is_valid(), "Symbol %s not found", "num_scalars");
Nodecl::NodeclBase num_scalars_ref = Nodecl::Symbol::make(num_scalars);
num_scalars_ref.set_type(num_scalars.get_type().no_ref().get_lvalue_reference_to());
Nodecl::Symbol nodecl_index = Nodecl::Symbol::make(index);
nodecl_index.set_type(index.get_type().get_lvalue_reference_to());
Nodecl::NodeclBase loop_header = Nodecl::RangeLoopControl::make(
nodecl_index,
const_value_to_nodecl(const_value_get_signed_int(1)),
num_scalars_ref,
Nodecl::NodeclBase::null());
Nodecl::NodeclBase expanded_combiner =
red->get_combiner().shallow_copy();
BasicReductionExpandVisitor expander_visitor(
red->get_omp_in(),
param_omp_in,
red->get_omp_out(),
param_omp_out,
index);
expander_visitor.walk(expanded_combiner);
function_body.replace(
Nodecl::ForStatement::make(loop_header,
Nodecl::List::make(
Nodecl::ExpressionStatement::make(
expanded_combiner)),
Nodecl::NodeclBase::null()));
_basic_reduction_map_openmp[red] = function_sym;
if (IS_FORTRAN_LANGUAGE)
{
Nodecl::Utils::Fortran::append_used_modules(construct.retrieve_context(),
function_sym.get_related_scope());
}
Nodecl::Utils::append_to_enclosing_top_level_location(construct, function_code);
return function_sym;
}
