void visit(const Nodecl::ObjectInit& node)
{
TL::Symbol sym = node.get_symbol();
if (sym.get_value().is_null())
return;
walk(sym.get_value());
}
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);
}
}
void visit(const Nodecl::Symbol& node)
{
TL::Symbol sym = node.get_symbol();
if ((_data_sharing.get_data_sharing(sym, /* check_enclosing */ false) & ~DS_IMPLICIT)
== DS_UNDEFINED)
{
// Mark this as an implicit firstprivate
_data_sharing.set_data_sharing(sym, TL::OpenMP::DataSharingAttribute( DS_FIRSTPRIVATE | DS_IMPLICIT) );
std::cerr << node.get_locus_str() << ": warning: assuming '" << sym.get_qualified_name() << "' as firstprivate" << std::endl;
}
}
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 Fortran::append_module_to_scope(TL::Symbol module,
TL::Scope scope)
{
ERROR_CONDITION(!module.is_valid() || !module.is_fortran_module(), "Symbol must be a Fortran module", 0);
scope_entry_t* used_modules_info
= ::get_or_create_used_modules_symbol_info(scope.get_decl_context());
P_LIST_ADD_ONCE(used_modules_info->entity_specs.related_symbols,
used_modules_info->entity_specs.num_related_symbols,
module.get_internal_symbol());
if (!module.get_internal_symbol()->entity_specs.is_builtin)
fortran_load_module(module.get_internal_symbol()->symbol_name, /* intrinsic */ 0, make_locus("", 0, 0));
}
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 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();
}
TL::Symbol new_function_symbol(TL::Symbol function)
{
TL::ObjectList<TL::Type> parameter_types = function.get_type().parameters();
TL::ObjectList<std::string> parameter_names;
TL::ObjectList<TL::Symbol> function_related_symbols = function.get_related_symbols();
for (TL::ObjectList<TL::Symbol>::iterator it = function_related_symbols.begin();
it != function_related_symbols.end();
it++)
{
parameter_names.append(it->get_name());
}
TL::Symbol new_function = SymbolUtils::new_function_symbol(
function,
function.get_name(),
function.get_type().returns(),
parameter_names,
parameter_types);
return new_function;
}
void DeviceFPGA::copy_stuff_to_device_file(
const TL::ObjectList<Nodecl::NodeclBase>& stuff_to_be_copied)
{
for (TL::ObjectList<Nodecl::NodeclBase>::const_iterator it = stuff_to_be_copied.begin();
it != stuff_to_be_copied.end();
++it)
{
if (it->is<Nodecl::FunctionCode>()
|| it->is<Nodecl::TemplateFunctionCode>())
{
TL::Symbol function = it->get_symbol();
TL::Symbol new_function = SymbolUtils::new_function_symbol(function, function.get_name() + "_hls");
Nodecl::Utils::SimpleSymbolMap symbol_map;
symbol_map.add_map(function, new_function);
_fpga_file_code.append(Nodecl::Utils::deep_copy(*it, *it, symbol_map));
}
else
{
_fpga_file_code.append(Nodecl::Utils::deep_copy(*it, *it));
}
}
}
void SimdVisitor::visit(const Nodecl::OpenMP::SimdFunction& simd_node)
{
Nodecl::FunctionCode function_code = simd_node.get_statement()
.as<Nodecl::FunctionCode>();
// Remove SimdFunction node
simd_node.replace(function_code);
TL::Symbol sym = function_code.get_symbol();
Nodecl::FunctionCode vectorized_func_code =
Nodecl::Utils::deep_copy(function_code, function_code).as<Nodecl::FunctionCode>();
// Vectorize function
_vectorizer.vectorize(vectorized_func_code,
_device_name, _vector_length, NULL);
// Set new name
std::stringstream vectorized_func_name;
vectorized_func_name <<"__"
<< sym.get_name()
<< "_"
<< _device_name
<< "_"
<< _vector_length;
vectorized_func_code.get_symbol().set_name(vectorized_func_name.str());
// Add SIMD version to vector function versioning
_vectorizer.add_vector_function_version(sym.get_name(), vectorized_func_code,
_device_name, _vector_length, NULL, TL::Vectorization::SIMD_FUNC_PRIORITY);
// Append vectorized function code to scalar function
simd_node.append_sibling(vectorized_func_code);
}
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 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);
}
TL::Source LoopBlocking::do_blocking()
{
Source result, block_loops;
result
<< block_loops
;
ObjectList<ForStatement> nest_loops = _for_nest_info.get_nest_list();
_nesting = std::min(_nest_factors.size(), nest_loops.size());
TL::Source *current_innermost_part = &block_loops;
// For every loop declare its block loop variable and the inter-block loop
ObjectList<TL::Expression>::iterator current_factor = _nest_factors.begin();
ObjectList<TL::ForStatement>::iterator current_for = nest_loops.begin();
for (int current_nest = 0;
current_nest < _nesting;
current_nest++, current_for++, current_factor++)
{
TL::IdExpression induction_var = current_for->get_induction_variable();
TL::Symbol sym = induction_var.get_symbol();
TL::Type type = sym.get_type();
std::string var = "_blk_" + sym.get_name();
TL::Source *new_innermost_part = new TL::Source();
(*current_innermost_part)
<< "for(" << type.get_declaration(sym.get_scope(), var) << " = " << current_for->get_lower_bound() << ";"
<< var << current_for->get_bound_operator() << current_for->get_upper_bound() << ";"
<< var << "+= ( " << current_for->get_step() << ") * " << current_factor->prettyprint() << ")"
<< (*new_innermost_part)
;
current_innermost_part = new_innermost_part;
}
// Now for every loop, declare the intra-loop
current_factor = _nest_factors.begin();
current_for = nest_loops.begin();
for (int current_nest = 0;
current_nest < _nesting;
current_nest++, current_for++, current_factor++)
{
TL::IdExpression induction_var = current_for->get_induction_variable();
TL::Symbol sym = induction_var.get_symbol();
TL::Type type = sym.get_type();
std::string var = induction_var.prettyprint();
std::string init_var = var;
// If the loop declares the iterator in the for statement
// declare it again
AST_t loop_init = current_for->get_iterating_init();
if (Declaration::predicate(loop_init))
{
// Fix init_var to be a declaration
init_var = type.get_declaration(sym.get_scope(), var);
}
std::string blk_var = "_blk_" + sym.get_name();
TL::Source min_code;
TL::Source *new_innermost_part = new TL::Source();
(*current_innermost_part)
<< "for(" << init_var << " = " << blk_var << ";"
<< var << current_for->get_bound_operator() << min_code << ";"
<< var << "+= ( " << current_for->get_step() << "))"
<< (*new_innermost_part)
;
TL::Source a, b;
min_code
<< "((" << a << ") < (" << b << ") ? (" << a << ") : (" << b << "))"
;
a << blk_var << " + (" << current_for->get_step() << ") * (" << current_factor->prettyprint() << " - 1 )";
b << current_for->get_upper_bound();
current_innermost_part = new_innermost_part;
}
// And now the innermost loop
(*current_innermost_part)
<< nest_loops[_nesting - 1].get_loop_body()
;
return result;
}
// Old version - Deprecated. Kept here for compatibility with old runtimes (Nanos++ 0.7)
void DeviceOpenCL::old_generate_ndrange_code(
const TL::Symbol& called_task,
const TL::Symbol& unpacked_function,
const TargetInformation& target_info,
const std::string filename,
const std::string kernel_name,
const TL::ObjectList<OutlineDataItem*>& data_items,
Nodecl::Utils::SimpleSymbolMap* called_fun_to_outline_data_map,
Nodecl::Utils::SimpleSymbolMap* outline_data_to_unpacked_fun_map,
// Out
TL::Source& code_ndrange)
{
// The arguments of the clauses 'ndrange' and 'shmem' must be updated because
// they are not expressed in terms of the unpacked function parameters
TL::ObjectList<Nodecl::NodeclBase> new_ndrange, new_shmem;
update_ndrange_and_shmem_expressions(
unpacked_function.get_related_scope(),
target_info,
outline_data_to_unpacked_fun_map,
new_ndrange,
new_shmem);
int num_args_ndrange = new_ndrange.size();
TL::Source code_ndrange_aux;
Nodecl::Utils::SimpleSymbolMap called_fun_to_unpacked_fun_map;
const std::map<TL::Symbol, TL::Symbol>* called_fun_to_outline_data_map_simple =
called_fun_to_outline_data_map->get_simple_symbol_map();
for (std::map<TL::Symbol, TL::Symbol>::const_iterator it = called_fun_to_outline_data_map_simple->begin();
it != called_fun_to_outline_data_map_simple->end();
it++)
{
TL::Symbol key = it->first;
TL::Symbol value = outline_data_to_unpacked_fun_map->map(it->second.get_internal_symbol());
called_fun_to_unpacked_fun_map.add_map(key, value);
}
bool dim_const = new_ndrange[0].is_constant();
char is_null_ended = 0;
bool check_dim = !(new_ndrange[num_args_ndrange - 1].is_constant()
&& const_value_is_string(new_ndrange[num_args_ndrange - 1].get_constant())
&& (strcmp(const_value_string_unpack_to_string(new_ndrange[num_args_ndrange-1].get_constant(), &is_null_ended),
"noCheckDim") == 0));
int num_dim = 0;
if (dim_const)
{
num_dim = const_value_cast_to_4(new_ndrange[0].get_constant());
ERROR_CONDITION(num_dim < 1 || num_dim > 3,
"invalid number of dimensions for 'ndrange' clause. Valid values: 1, 2 and 3." , 0);
ERROR_CONDITION((((num_dim * 3) + 1 + !check_dim) != num_args_ndrange)
&& (((num_dim * 2) + 1 + !check_dim) != num_args_ndrange),
"invalid number of arguments for 'ndrange' clause", 0);
}
std::string compiler_opts;
if (CURRENT_CONFIGURATION->opencl_build_options != NULL)
{
compiler_opts = std::string(CURRENT_CONFIGURATION->opencl_build_options);
}
//Create OCL Kernel
code_ndrange_aux << "nanos_err_t nanos_err;"
<< "void* ompss_kernel_ocl = nanos_create_current_kernel(\""
<< kernel_name << "\",\""
<< filename << "\",\""
<< compiler_opts << "\");";
//Prepare setArgs
unsigned int index_local = 0;
TL::ObjectList<TL::Symbol> parameters_called = called_task.get_function_parameters();
for (unsigned int i = 0; i < parameters_called.size(); ++i)
{
TL::Symbol unpacked_argument = called_fun_to_unpacked_fun_map.map(parameters_called[i]);
// The attribute __global is deduced: the current argument will be __global if it has any copies
bool is_global = false;
if (unpacked_argument.get_type().no_ref().is_pointer()
|| unpacked_argument.get_type().no_ref().is_array())
{
for (TL::ObjectList<OutlineDataItem*>::const_iterator it = data_items.begin();
it != data_items.end() && !is_global;
++it)
{
TL::Symbol outline_data_item_sym = (*it)->get_symbol();
// If the outline data item has not a valid symbol, skip it
if (!outline_data_item_sym.is_valid())
continue;
// If the symbol of the current outline data item is not the
// same as the unpacked_argument, skip it
if(outline_data_to_unpacked_fun_map->map(outline_data_item_sym.get_internal_symbol()) != unpacked_argument)
continue;
is_global = !((*it)->get_copies().empty());
}
}
bool is_local = !is_global && unpacked_argument.get_type().no_ref().is_pointer();
if (is_global)
{
code_ndrange_aux
<< "nanos_err = nanos_opencl_set_bufferarg("
<< "ompss_kernel_ocl, "
<< i << ", "
<< as_symbol(unpacked_argument) <<");";
}
else if (is_local)
{
TL::Source sizeof_arg;
if (index_local >= new_shmem.size())
{
warn_printf_at(called_task.get_locus(),
"the size of the local symbol '%s' has not been specified in the 'shmem' clause, assuming zero\n",
unpacked_argument.get_name().c_str());
sizeof_arg << "0";
}
else
{
sizeof_arg << as_expression(new_shmem[index_local]);
}
code_ndrange_aux << "nanos_err = nanos_opencl_set_arg("
<< "ompss_kernel_ocl, "
<< i << ", "
<< sizeof_arg << ", "
<< "0);";
++index_local;
}
else
{
code_ndrange_aux << "nanos_err = nanos_opencl_set_arg("
<< "ompss_kernel_ocl, "
<< i << ", "
<< "sizeof(" << as_type(unpacked_argument.get_type().no_ref()) << "), "
<< "&" << as_symbol(unpacked_argument) <<");";
}
}
//Build arrays with information from ndrange clause or pointing to the ndrange pointers
if (!dim_const)
{
if (IS_FORTRAN_LANGUAGE)
{
internal_error("The number of dimensions is non-constant. This feature is not implemented yet in Fortran.", 0);
}
//Prepare ndrange calc pointers and arrays
code_ndrange_aux
<< "int num_dim = " << as_expression(new_ndrange[0]) <<";"
<< "size_t offset_tmp[num_dim];"
<< "size_t offset_arr[num_dim];"
<< "size_t local_size_arr[num_dim];"
<< "size_t global_size_arr[num_dim];"
<< "size_t* local_size_ptr;"
<< "size_t* offset_ptr;"
<< "size_t* global_size_ptr;"
<< "size_t* final_local_size_ptr;"
<< as_type(TL::Type::get_bool_type()) << " local_size_zero = 0;"
<< "int i = 0;"
;
if (num_args_ndrange == 3)
{
code_ndrange_aux
<< "for (i = 0; i < num_dim; ++i)"
<< "{"
<< "offset_tmp[i] = 0;"
<< "}"
<< "offset_ptr = offset_tmp;"
<< "global_size_ptr = " << as_expression(new_ndrange[1]) << ";"
<< "local_size_ptr = " << as_expression(new_ndrange[2]) << ";"
;
}
else if (num_args_ndrange == 4)
{
code_ndrange_aux
<< "offset_ptr = " << as_expression(new_ndrange[1]) << ";"
<< "global_size_ptr = " << as_expression(new_ndrange[2]) << ";"
<< "local_size_ptr = " << as_expression(new_ndrange[3]) << ";"
;
}
else
{
WARNING_MESSAGE("Invalid number of parameters for ndrange, when number of dimensions is not const, it must be 3 or 4",0);
}
//Check if local_size has zeros
code_ndrange_aux
<< "for (i = 0; i < num_dim; ++i)"
<< "{"
<< "if (local_size_ptr[i] == 0)"
<< "{"
<< "local_size_zero = 1;"
<< "}"
<< " }"
<< "if (local_size_zero)"
<< "{"
<< "for (i = 0; i < num_dim; ++i)"
<< "{"
<< "local_size_ptr[i] = 1;"
<< "}"
<< "}"
;
//Now do the rounding
if (check_dim)
{
code_ndrange_aux
<< "for (i = 0; i < num_dim; ++i)"
<< "{"
<< "offset_arr[i] = offset_ptr[i];"
<< "local_size_arr[i] = (global_size_ptr[i] < local_size_ptr[i]) ? "
<< "global_size_ptr[i] : local_size_ptr[i];"
<< "global_size_arr[i] = (global_size_ptr[i] < local_size_ptr[i]) ? "
<< "global_size_ptr[i] : global_size_ptr[i] + ("
<< "(global_size_ptr[i] % local_size_ptr[i] == 0) ? "
<< "0 : (local_size_ptr[i] - global_size_ptr[i] % local_size_ptr[i]));"
<< "}"
;
}
if (check_dim)
{
code_ndrange_aux
<< "if (local_size_zero)"
<< "{"
<< "final_local_size_ptr = 0;"
<< "}"
<< "else"
<< "{"
<< "final_local_size_ptr = local_size_arr;"
<< "}"
//Launch kernel/ it will be freed inside, with ndrange calculated inside the checkDim loop
<< "nanos_exec_kernel(ompss_kernel_ocl, num_dim, offset_arr, final_local_size_ptr, global_size_arr);";
;
}
else
{
code_ndrange_aux
<< "if (local_size_zero)"
<< "{"
<< "final_local_size_ptr = 0;"
<< "}"
<< "else"
<< "{"
<< "final_local_size_ptr = local_size_ptr;"
<< "}"
<< "nanos_exec_kernel(ompss_kernel_ocl, num_dim, offset_ptr, final_local_size_ptr, global_size_ptr);"
;
}
}
else
{
int num_dim_offset = num_dim;
//Prepare ndrange calc pointers and arrays
code_ndrange_aux
<< "int num_dim = " << as_expression(new_ndrange[0]) <<";"
<< "size_t offset_arr[num_dim];"
<< "size_t local_size_arr[num_dim];"
<< "size_t global_size_arr[num_dim];"
<< as_type(TL::Type::get_bool_type()) << " local_size_zero;"
<< "local_size_zero = 0;"
;
for (int i = 1; i <= num_dim; ++i)
{
if (((num_dim * 3) + 1 + !check_dim) != num_args_ndrange)
{
num_dim_offset = 0;
code_ndrange_aux << "offset_arr[" << i-1 << "] = 0;";
}
else
{
code_ndrange_aux
<< "offset_arr[" << i-1 << "] = " << as_expression(new_ndrange[i]) << ";";
}
code_ndrange_aux
<< "local_size_arr[" << i-1 << "] = " << as_expression(new_ndrange[num_dim + num_dim_offset + i]) << ";"
<< "if (local_size_arr[" << i - 1 << "] == 0)"
<< "{"
<< "local_size_zero = 1;"
<< "}"
<< "global_size_arr[" << i-1 << "] = " << as_expression(new_ndrange[num_dim_offset + i]) << ";"
;
}
//Now do the rounding
if (check_dim)
{
code_ndrange_aux
<< "if (!local_size_zero)"
<< "{"
<< "int i;"
<< "for (i = 0; i < num_dim; i = i + 1)"
<< "{"
<< "if (global_size_arr[i] < local_size_arr[i])"
<< "{"
<< "local_size_arr[i] = global_size_arr[i];"
<< "}"
<< "else"
<< "{"
<< "if (global_size_arr[i] % local_size_arr[i] != 0)"
<< "{"
<< "global_size_arr[i] = global_size_arr[i]"
<< " + (local_size_arr[i] - global_size_arr[i] % local_size_arr[i]);"
<< "}"
<< "}"
<< "}"
<< "}"
;
}
code_ndrange_aux
<< "if (local_size_zero)"
<< "{"
//Launch kernel/ it will be freed inside, with ndrange calculated inside the checkDim loop
<< "nanos_err = nanos_exec_kernel(ompss_kernel_ocl, num_dim, offset_arr, 0, global_size_arr);"
<< "}"
<< "else"
<< "{"
//Launch kernel/ it will be freed inside, with ndrange calculated inside the checkDim loop
<< "nanos_err = nanos_exec_kernel(ompss_kernel_ocl, num_dim, offset_arr, local_size_arr, global_size_arr);"
<< "}"
;
}
if (IS_FORTRAN_LANGUAGE)
{
Source::source_language = SourceLanguage::C;
Nodecl::NodeclBase code_ndrange_tree = code_ndrange_aux.parse_statement(unpacked_function.get_related_scope());
Source::source_language = SourceLanguage::Current;
code_ndrange << as_statement(code_ndrange_tree);
}
else
{
code_ndrange << code_ndrange_aux;
}
}
TL::Symbol new_function_symbol(
TL::Symbol current_function,
const std::string& name,
TL::Type return_type,
TL::ObjectList<std::string> parameter_names,
TL::ObjectList<TL::Type> parameter_types)
{
if (IS_FORTRAN_LANGUAGE && current_function.is_nested_function())
{
// Get the enclosing function
current_function = current_function.get_scope().get_related_symbol();
}
decl_context_t decl_context = current_function.get_scope().get_decl_context();
ERROR_CONDITION(parameter_names.size() != parameter_types.size(), "Mismatch between names and types", 0);
decl_context_t function_context;
if (IS_FORTRAN_LANGUAGE)
{
function_context = new_program_unit_context(decl_context);
}
else
{
function_context = new_function_context(decl_context);
function_context = new_block_context(function_context);
}
// Build the function type
int num_parameters = 0;
scope_entry_t** parameter_list = NULL;
parameter_info_t* p_types = new parameter_info_t[parameter_types.size()];
parameter_info_t* it_ptypes = &(p_types[0]);
TL::ObjectList<TL::Type>::iterator type_it = parameter_types.begin();
for (TL::ObjectList<std::string>::iterator it = parameter_names.begin();
it != parameter_names.end();
it++, it_ptypes++, type_it++)
{
scope_entry_t* param = new_symbol(function_context, function_context.current_scope, it->c_str());
param->entity_specs.is_user_declared = 1;
param->kind = SK_VARIABLE;
param->locus = make_locus("", 0, 0);
param->defined = 1;
param->type_information = get_unqualified_type(type_it->get_internal_type());
P_LIST_ADD(parameter_list, num_parameters, param);
it_ptypes->is_ellipsis = 0;
it_ptypes->nonadjusted_type_info = NULL;
it_ptypes->type_info = get_indirect_type(param);
}
type_t *function_type = get_new_function_type(
return_type.get_internal_type(),
p_types,
parameter_types.size());
delete[] p_types;
// Now, we can create the new function symbol
scope_entry_t* new_function_sym = NULL;
if (!current_function.get_type().is_template_specialized_type())
{
new_function_sym = new_symbol(decl_context, decl_context.current_scope, name.c_str());
new_function_sym->entity_specs.is_user_declared = 1;
new_function_sym->kind = SK_FUNCTION;
new_function_sym->locus = make_locus("", 0, 0);
new_function_sym->type_information = function_type;
}
else
{
scope_entry_t* new_template_sym = new_symbol(
decl_context, decl_context.current_scope, name.c_str());
new_template_sym->kind = SK_TEMPLATE;
new_template_sym->locus = make_locus("", 0, 0);
new_template_sym->type_information = get_new_template_type(
decl_context.template_parameters,
function_type,
uniquestr(name.c_str()),
decl_context, make_locus("", 0, 0));
template_type_set_related_symbol(new_template_sym->type_information, new_template_sym);
// The new function is the primary template specialization
new_function_sym = named_type_get_symbol(
template_type_get_primary_type(
new_template_sym->type_information));
}
function_context.function_scope->related_entry = new_function_sym;
function_context.block_scope->related_entry = new_function_sym;
new_function_sym->related_decl_context = function_context;
new_function_sym->entity_specs.related_symbols = parameter_list;
new_function_sym->entity_specs.num_related_symbols = num_parameters;
for (int i = 0; i < new_function_sym->entity_specs.num_related_symbols; ++i)
{
symbol_set_as_parameter_of_function(
new_function_sym->entity_specs.related_symbols[i], new_function_sym, /* parameter position */ i);
}
// Make it static
new_function_sym->entity_specs.is_static = 1;
// Make it member if the enclosing function is member
if (current_function.is_member())
{
new_function_sym->entity_specs.is_member = 1;
new_function_sym->entity_specs.class_type = current_function.get_class_type().get_internal_type();
new_function_sym->entity_specs.access = AS_PUBLIC;
::class_type_add_member(new_function_sym->entity_specs.class_type, new_function_sym);
}
if (current_function.is_inline())
new_function_sym->entity_specs.is_inline = 1;
// new_function_sym->entity_specs.is_defined_inside_class_specifier =
// current_function.get_internal_symbol()->entity_specs.is_defined_inside_class_specifier;
if (IS_FORTRAN_LANGUAGE && current_function.is_in_module())
{
scope_entry_t* module_sym = current_function.in_module().get_internal_symbol();
new_function_sym->entity_specs.in_module = module_sym;
P_LIST_ADD(
module_sym->entity_specs.related_symbols,
module_sym->entity_specs.num_related_symbols,
new_function_sym);
new_function_sym->entity_specs.is_module_procedure = 1;
}
return new_function_sym;
}
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;
}
std::string as_symbol(TL::Symbol s)
{
return symbol_to_source(s.get_internal_symbol());
}
void LoweringVisitor::visit(const Nodecl::ExpressionStatement& expr_stmt)
{
Nodecl::NodeclBase nest = expr_stmt.get_nest();
if (IS_FORTRAN_LANGUAGE
&& nest.is<Nodecl::FunctionCall>())
{
Nodecl::FunctionCall function_call = nest.as<Nodecl::FunctionCall>();
if (function_call.get_called().is<Nodecl::Symbol>())
{
TL::Symbol sym = function_call.get_called().as<Nodecl::Symbol>().get_symbol();
// We are only interested in two intrinsic symbols
if (sym.is_intrinsic())
{
if(sym.get_name() == "ompss_opencl_allocate")
{
// We replace the intrinsic call by a call to a new function which:
// - allocates a new temporary array with descriptor
// - copies the array descriptor to the address of the array
// - calls to the Nanos++ API to allocate the buffer in the shared memory
// - deallocates the temporary array with descriptor
//
// Example:
//
// ...
// INTEGER, ALLOCATABLE :: V(:)
// OMPSS_OPENCL_ALLOCATE(V(10))
// ...
//
// Is transformed into:
//
// SUBROUTINE NANOX_OPENCL_ALLOCATE_INTERNAL(ARR, LB1, UB1)
// INTEGER, ALLOCATABLE :: ARR(:)
// INTEGER :: LB1, UB1, ERR
// INTEGER, ALLOCATABLE :: TMP(:)
//
// ALLOCATE(TMP(LB1:UB1))
//
// ERR = NANOS_MEMCPY(
// MERCURIUM_GET_ADDRESS_OF(ARR),
// MERCURIUM_GET_ADDRESS_OF(TMP),
// 48)
//
// CALL NANOS_OPENCL_ALLOCATE_FORTRAN(
// SIZEOF(TMP),
// MERCURIUM_GET_ADDRESS_OF(ARR))
//
// DEALLOCATE(TMP)
// END SUBROUTINE NANOX_OPENCL_ALLOCATE_INTERNAL
//
// ...
// INTEGER, ALLOCATABLE :: V(:)
// CALL NANOX_OPENCL_ALLOCATE_INTERNAL(V, 1, 10)
// ...
//
// For more information: https://pm.bsc.es/projects/mcxx/ticket/1994
handle_ompss_opencl_allocate_intrinsic(
function_call,
_declared_ocl_allocate_functions,
expr_stmt);
}
else if (sym.get_name() == "ompss_opencl_deallocate")
{
// The transformation applied to this intrinsic is more
// simple than the other one, we only need to replace
// the call to the intrinsic by a call to the Nanos++
// API:
//
// ...
// INTEGER, ALLOCATABLE :: V(:)
// ...
// OMPSS_OPENCL_DEALLOCATE(V)
// ...
//
// Is transformed into:
//
// ...
// INTEGER, ALLOCATABLE :: V(:)
// ...
// CALL NANOS_OPENCL_ALLOCATE_FORTRAN(MERCURIUM_GET_ADDRESS_OF(V))
// ...
handle_ompss_opencl_deallocate_intrinsic(function_call, expr_stmt);
}
}
}
}
walk(expr_stmt.get_nest());
}
void DeviceOpenCL::generate_ndrange_code(
const TL::Symbol& called_task,
const TL::Symbol& unpacked_function,
const TargetInformation& target_info,
const std::string filename,
const std::string kernel_name,
const TL::ObjectList<OutlineDataItem*>& data_items,
Nodecl::Utils::SimpleSymbolMap* called_fun_to_outline_data_map,
Nodecl::Utils::SimpleSymbolMap* outline_data_to_unpacked_fun_map,
// Out
TL::Source& code_ndrange)
{
if (!Nanos::Version::interface_is_at_least("opencl", 1003))
{
return old_generate_ndrange_code(
called_task,
unpacked_function,
target_info,
filename,
kernel_name,
data_items,
called_fun_to_outline_data_map,
outline_data_to_unpacked_fun_map,
code_ndrange);
}
// The arguments of the clauses 'ndrange' and 'shmem' must be updated because
// they are not expressed in terms of the unpacked function parameters
TL::ObjectList<Nodecl::NodeclBase> new_ndrange, new_shmem;
update_ndrange_and_shmem_expressions(
unpacked_function.get_related_scope(),
target_info,
outline_data_to_unpacked_fun_map,
new_ndrange,
new_shmem);
// Prepare mapping for the call to the kernel
TL::Source code_ndrange_aux;
Nodecl::Utils::SimpleSymbolMap called_fun_to_unpacked_fun_map;
const std::map<TL::Symbol, TL::Symbol>* called_fun_to_outline_data_map_simple =
called_fun_to_outline_data_map->get_simple_symbol_map();
for (std::map<TL::Symbol, TL::Symbol>::const_iterator it = called_fun_to_outline_data_map_simple->begin();
it != called_fun_to_outline_data_map_simple->end();
it++)
{
TL::Symbol key = it->first;
TL::Symbol value = outline_data_to_unpacked_fun_map->map(it->second.get_internal_symbol());
called_fun_to_unpacked_fun_map.add_map(key, value);
}
// The syntax of ndrange is
//
// ndrange(N, global-list [, local-list])
//
// Each X-list has as much as N elements
Nodecl::NodeclBase num_dims_expr = new_ndrange[0];
// N must be a constant
if (!num_dims_expr.is_constant())
{
fatal_printf_at(num_dims_expr.get_locus(),
"first argument in 'ndrange' clause must be constant\n");
}
// At this point we can remove "N" from the new_ndrange list (pop_front)
new_ndrange.erase(new_ndrange.begin());
// N must be between 1 and 3
int num_dims = const_value_cast_to_signed_int(num_dims_expr.get_constant());
if (num_dims < 1 || num_dims > 3)
{
fatal_printf_at(num_dims_expr.get_locus(),
"number of dimensions for 'ndrange' clause is not 1, 2 or 3\n");
}
// Checking the number of remaining expressions in the new_ndrange list
if (num_dims != (int)new_ndrange.size()
&& (num_dims * 2) != (int)new_ndrange.size())
{
fatal_printf_at(num_dims_expr.get_locus(),
"a 'ndrange(%d, argument-list)' clause requires %d or %d arguments in argument-list\n",
num_dims,
num_dims ,
num_dims * 2);
}
std::string compiler_options;
if (CURRENT_CONFIGURATION->opencl_build_options != NULL)
{
compiler_options = std::string(CURRENT_CONFIGURATION->opencl_build_options);
}
// Create OpenCL kernel
code_ndrange_aux << "nanos_err_t nanos_err;"
<< "void* ompss_kernel_ocl = nanos_create_current_kernel(\""
<< kernel_name << "\",\""
<< filename << "\","
<< "\"" << compiler_options << "\");";
// Prepare setArgs
TL::ObjectList<Nodecl::NodeclBase> global_list;
TL::ObjectList<Nodecl::NodeclBase> local_list;
unsigned int index_local = 0;
TL::ObjectList<TL::Symbol> parameters_called = called_task.get_function_parameters();
for (unsigned int i = 0; i < parameters_called.size(); ++i)
{
TL::Symbol unpacked_argument = called_fun_to_unpacked_fun_map.map(parameters_called[i]);
// The attribute __global is deduced: the current argument will be __global if it has any copies
bool is_global = false;
if (unpacked_argument.get_type().no_ref().is_pointer()
|| unpacked_argument.get_type().no_ref().is_array())
{
for (TL::ObjectList<OutlineDataItem*>::const_iterator it = data_items.begin();
it != data_items.end() && !is_global;
++it)
{
TL::Symbol outline_data_item_sym = (*it)->get_symbol();
// If the outline data item has not a valid symbol, skip it
if (!outline_data_item_sym.is_valid())
continue;
// If the symbol of the current outline data item is not the
// same as the unpacked_argument, skip it
if(outline_data_to_unpacked_fun_map->map(outline_data_item_sym.get_internal_symbol()) != unpacked_argument)
continue;
is_global = !((*it)->get_copies().empty());
}
}
bool is_local = !is_global && unpacked_argument.get_type().no_ref().is_pointer();
if (is_global)
{
code_ndrange_aux
<< "nanos_err = nanos_opencl_set_bufferarg("
<< "ompss_kernel_ocl, "
<< i << ", "
<< as_symbol(unpacked_argument) <<");";
}
else if (is_local)
{
TL::Source sizeof_arg;
if (index_local >= new_shmem.size())
{
warn_printf_at(called_task.get_locus(),
"the size of the local symbol '%s' has not been specified in the 'shmem' clause, assuming zero\n",
unpacked_argument.get_name().c_str());
sizeof_arg << "0";
}
else
{
sizeof_arg << as_expression(new_shmem[index_local]);
}
code_ndrange_aux << "nanos_err = nanos_opencl_set_arg("
<< "ompss_kernel_ocl, "
<< i << ", "
<< sizeof_arg << ", "
<< "0);";
++index_local;
}
else
{
code_ndrange_aux << "nanos_err = nanos_opencl_set_arg("
<< "ompss_kernel_ocl, "
<< i << ", "
<< "sizeof(" << as_type(unpacked_argument.get_type().no_ref()) << "), "
<< "&" << as_symbol(unpacked_argument) <<");";
}
}
//Build arrays with information from ndrange clause or pointing to the ndrange pointers
if (num_dims * 2 == (int)new_ndrange.size())
{
// ndrange(global-list, local-list)
int i = 0;
for (; i < num_dims; i++)
{
global_list.append(new_ndrange[i]);
}
for (; i < num_dims*2; i++)
{
local_list.append(new_ndrange[i]);
}
}
// locals are not specified here
else if (num_dims == (int)new_ndrange.size())
{
// ndrange(global-list)
int i = 0;
for (int k = 0; k < num_dims; k++, i++)
{
global_list.append(new_ndrange[i]);
}
}
else
{
internal_error("Code unreachable", 0);
}
bool there_is_local_size = !local_list.empty();
// Prepare ndrange calc pointers and arrays
if (there_is_local_size)
{
code_ndrange_aux
<< "size_t local_size_arr[" << num_dims << "];"
;
}
code_ndrange_aux
<< "size_t global_size_arr[" << num_dims << "];"
;
for (int i = 0; i < num_dims; i++)
{
if (there_is_local_size)
{
code_ndrange_aux
<< "local_size_arr[" << i << "] = " << as_expression(local_list[i]) << ";"
;
}
code_ndrange_aux
<< "global_size_arr[" << i << "] = " << as_expression(global_list[i]) << ";"
;
}
if (there_is_local_size)
{
// Launch kernel/ it will be freed inside, with ndrange calculated inside the checkDim loop
code_ndrange_aux << "nanos_err = nanos_exec_kernel(ompss_kernel_ocl, " << num_dims << ", local_size_arr, global_size_arr);"
;
}
else
{
// Let the runtime choose the best local size
code_ndrange_aux << "nanos_err = nanos_profile_exec_kernel(ompss_kernel_ocl, " << num_dims << ", global_size_arr);"
;
}
if (IS_FORTRAN_LANGUAGE)
{
Source::source_language = SourceLanguage::C;
Nodecl::NodeclBase code_ndrange_tree = code_ndrange_aux.parse_statement(unpacked_function.get_related_scope());
Source::source_language = SourceLanguage::Current;
code_ndrange << as_statement(code_ndrange_tree);
}
else
{
code_ndrange << code_ndrange_aux;
}
}
TL::Source LoopUnroll::do_unroll()
{
if (!_for_stmt.regular_loop())
{
return silly_unroll();
}
// Get parts of the loop
IdExpression induction_var = _for_stmt.get_induction_variable();
Expression lower_bound = _for_stmt.get_lower_bound();
Expression upper_bound = _for_stmt.get_upper_bound();
Expression step = _for_stmt.get_step();
TL::Source operator_bound = _for_stmt.get_bound_operator();
Statement loop_body = _for_stmt.get_loop_body();
TL::Source result, epilogue,
main, induction_var_decl,
before_main, after_main;
std::stringstream ss;
ss << _factor;
result
<< "{"
<< induction_var_decl
<< before_main
<< main
<< after_main
<< epilogue
<< "}"
;
Source replicated_body;
Source epilogue_body;
if (_factor > 1)
{
AST_t init = _for_stmt.get_iterating_init();
if (Declaration::predicate(init))
{
TL::Symbol sym = induction_var.get_symbol();
TL::Type type = sym.get_type();
// Declare it since it will have local scope
induction_var_decl
<< type.get_declaration(sym.get_scope(), sym.get_name()) << ";"
;
}
main
<< "for (" << induction_var << " = " << lower_bound << ";"
<< induction_var << operator_bound << "((" << upper_bound << ") - (" << _factor << " - 1)* (" << step << "));"
<< induction_var << "+= (" << step << ") * " << _factor << ")"
<< "{"
<< replicated_body
<< "}"
;
// FIXME - It could help to initialize here another variable and make both loops independent
epilogue
<< "for ( ; " // No initialization, keep using the old induction var
<< induction_var << operator_bound << upper_bound << ";"
<< induction_var << "+= (" << step << "))"
<< epilogue_body
;
if (!_remove_tasks)
{
epilogue_body << loop_body;
}
else
{
std::cerr << "Do not create task " << __FILE__ << ":" << __LINE__ << std::endl;
running_error("Path not supported yet", 0);
// epilogue_body << loop_body.get_ast().prettyprint_with_callback(functor(ignore_tasks));
}
}
else
{
// Leave it as is
main << "for(" << _for_stmt.get_iterating_init().prettyprint()
<< _for_stmt.get_iterating_condition() << ";"
<< _for_stmt.get_iterating_expression() << ")"
<< "{"
<< replicated_body
<< "}"
;
}
// Replicate the body
bool consider_omp = false;
if (TaskAggregation::contains_relevant_openmp(loop_body))
{
consider_omp = true;
}
if (_ignore_omp || !consider_omp)
{
simple_replication(_factor, replicated_body, epilogue_body,
induction_var, loop_body);
}
else
{
omp_replication(_factor, replicated_body, epilogue_body,
induction_var, loop_body, before_main, after_main);
}
return result;
}
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 bool check_symbol(TL::Symbol s)
{
return CheckIfInCudaCompiler::check(s.get_filename());
}