TL::Type get_array_of_vector(TL::Type t)
{
ERROR_CONDITION(!t.is_vector(), "Invalid type", 0);
return t.vector_element().get_array_to(
const_value_to_nodecl(
const_value_get_signed_int(t.vector_num_elements())
),
TL::Scope::get_global_scope());
}
TL::Type get_array_of_mask(TL::Type t)
{
ERROR_CONDITION(!t.is_mask(), "Invalid type", 0);
int n = (t.get_mask_num_elements() / 8)
+ !!(t.get_mask_num_elements() % 8);
return TL::Type::get_unsigned_char_type().get_array_to(
const_value_to_nodecl(
const_value_get_signed_int(n)
),
TL::Scope::get_global_scope());
}
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 LoopNormalize::normalize()
{
Nodecl::ForStatement loop = this->_loop.as<Nodecl::ForStatement>();
TL::Scope orig_loop_scope = loop.retrieve_context();
TL::ForStatement for_stmt(loop);
TL::Symbol induction_var = for_stmt.get_induction_variable();
Nodecl::NodeclBase orig_loop_lower = for_stmt.get_lower_bound();
Nodecl::NodeclBase orig_loop_upper = for_stmt.get_upper_bound();
Nodecl::NodeclBase orig_loop_step = for_stmt.get_step();
// Do nothing if the loop is already a normalized loop
if ( orig_loop_lower.is_constant()
&& const_value_is_zero(orig_loop_lower.get_constant())
&& orig_loop_step.is_constant()
&& const_value_is_one(orig_loop_step.get_constant()))
{
_transformation = _loop.shallow_copy();
return;
}
// DO I = L, U, S
// A(I)
// becomes
// DO I1 = 0, (U-L)/S, 1
// A(S * I1 + L)
//
Nodecl::NodeclBase new_upper;
if (orig_loop_upper.is_constant() && orig_loop_lower.is_constant())
{
new_upper = const_value_to_nodecl(
const_value_sub(
orig_loop_upper.get_constant(),
orig_loop_lower.get_constant()));
if (orig_loop_step.is_constant())
{
new_upper = const_value_to_nodecl(
const_value_div(
new_upper.get_constant(),
orig_loop_step.get_constant()));
}
else
{
new_upper = Nodecl::Div::make(
new_upper,
orig_loop_step.shallow_copy(),
new_upper.get_type());
}
}
else
{
new_upper = Nodecl::Div::make(
Nodecl::Minus::make(
orig_loop_upper.shallow_copy(),
orig_loop_lower.shallow_copy(),
orig_loop_upper.get_type()),
orig_loop_step.shallow_copy(),
orig_loop_upper.get_type());
}
Nodecl::NodeclBase normalized_loop_body = loop.get_statement().shallow_copy();
ReplaceInductionVar replace_induction_var(induction_var, orig_loop_lower, orig_loop_step);
replace_induction_var.walk(normalized_loop_body);
Nodecl::NodeclBase normalized_loop_control;
if (IS_FORTRAN_LANGUAGE)
{
normalized_loop_control =
Nodecl::RangeLoopControl::make(
induction_var.make_nodecl(),
const_value_to_nodecl(const_value_get_signed_int(0)),
new_upper,
const_value_to_nodecl(const_value_get_signed_int(1)));
}
else // IS_C_LANGUAGE || IS_CXX_LANGUAGE
{
Nodecl::NodeclBase init;
// i = 0
if (for_stmt.induction_variable_in_separate_scope())
{
induction_var.set_value(const_value_to_nodecl(const_value_get_signed_int(0)));
init = Nodecl::ObjectInit::make(induction_var);
}
else
{
init =
Nodecl::Assignment::make(
induction_var.make_nodecl(/* ref */ true),
const_value_to_nodecl(const_value_get_signed_int(0)),
induction_var.get_type().no_ref().get_lvalue_reference_to());
}
// i <= new_upper
Nodecl::NodeclBase cond =
Nodecl::LowerOrEqualThan::make(
induction_var.make_nodecl(/* set_ref_type */ true),
new_upper,
::get_bool_type());
// i = i + 1
Nodecl::NodeclBase next =
Nodecl::Assignment::make(
induction_var.make_nodecl(/* set_ref_type */ true),
Nodecl::Add::make(
induction_var.make_nodecl(/* set_ref_type */ true),
const_value_to_nodecl(const_value_get_signed_int(1)),
induction_var.get_type().no_ref()),
induction_var.get_type().no_ref().get_lvalue_reference_to());
normalized_loop_control =
Nodecl::LoopControl::make(
Nodecl::List::make(init),
cond,
next);
}
Nodecl::NodeclBase normalized_for =
Nodecl::ForStatement::make(
normalized_loop_control,
normalized_loop_body,
/* loop-name */ Nodecl::NodeclBase::null());
_transformation = normalized_for;
// Compute what value the induction variable should take after the loop
if (!for_stmt.induction_variable_in_separate_scope())
{
Nodecl::NodeclBase induction_variable =
for_stmt.get_induction_variable().make_nodecl(/* set_ref_type */ true);
Nodecl::NodeclBase expr =
HLT::Utils::compute_induction_variable_final_expr(_loop);
_post_transformation_stmts.append(
Nodecl::ExpressionStatement::make(
Nodecl::Assignment::make(
induction_variable,
expr,
induction_variable.get_type())));
}
}
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;
}
