static nodecl_t make_list_helper(const TL::ObjectList<NodeclBase>::const_iterator& first,
const TL::ObjectList<NodeclBase>::const_iterator& last)
{
if (first == last)
{
if (first->is<Nodecl::List>())
{
return first->get_internal_nodecl();
}
else
{
return nodecl_make_list_1(first->get_internal_nodecl());
}
}
else
{
nodecl_t previous_list = make_list_helper(first, last - 1);
if (last->is<Nodecl::List>())
{
return nodecl_concat_lists(previous_list,
last->get_internal_nodecl());
}
else
{
return nodecl_append_to_list(previous_list,
last->get_internal_nodecl());
}
}
}
TL::ObjectList<NodeclBase> List::to_object_list() const
{
TL::ObjectList<NodeclBase> result;
for (List::const_iterator it = this->begin(); it != this->end(); ++it)
{
result.append(*it);
}
return result;
}
void TL::Optimizations::strength_reduce(
TL::ObjectList<Nodecl::NodeclBase>& list,
bool fast_math)
{
for(TL::ObjectList<Nodecl::NodeclBase>::iterator it = list.begin();
it != list.end();
it++)
{
strength_reduce(*it, fast_math);
}
}
void TL::Optimizations::canonicalize_and_fold(
const TL::ObjectList<Nodecl::NodeclBase>& list,
bool fast_math)
{
ReduceExpressionVisitor exp_reducer;
for(TL::ObjectList<Nodecl::NodeclBase>::const_iterator it = list.begin();
it != list.end();
it++)
{
canonicalize_and_fold(*it, fast_math);
}
}
bool DeviceFPGA::task_has_scalars(TL::ObjectList<OutlineDataItem*> & dataitems)
{
for (ObjectList<OutlineDataItem*>::iterator it = dataitems.begin();
it != dataitems.end();
it++)
{
/*
* We happily assume that everything that does not need a copy is a scalar
* Which is true as long everything that is not a scalar is going to be copied
* We could also use DataReference to check this
* FIXME structs should be treated
*/
if((*it)->get_copies().empty())
{
return true;
}
}
return false;
}
ObjectList<MemberDeclarationInfo> Type::get_member_declarations() const
{
TL::ObjectList<MemberDeclarationInfo> result;
int num_decls = 0;
member_declaration_info_t* mdi = ::class_type_get_member_declarations(_type_info, &num_decls);
int i;
for (i = 0; i < num_decls; i++)
{
result.push_back(
MemberDeclarationInfo(mdi[i].entry, mdi[i].decl_context, mdi[i].is_definition)
);
}
DELETE(mdi);
return result;
}
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 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;
}
}
// 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;
}
}
static void add_copy_items(PragmaCustomLine construct,
DataEnvironment& data_sharing_environment,
const ObjectList<Nodecl::NodeclBase>& list,
TL::OmpSs::CopyDirection copy_direction,
TL::OmpSs::TargetInfo& target_info,
bool in_ompss_mode)
{
TL::ObjectList<TL::OmpSs::CopyItem> items;
for (ObjectList<Nodecl::NodeclBase>::const_iterator it = list.begin();
it != list.end();
it++)
{
DataReference expr(*it);
std::string warning;
if (!expr.is_valid())
{
expr.commit_diagnostic();
warn_printf_at(construct.get_locus(), "'%s' is not a valid copy data-reference, skipping\n",
expr.prettyprint().c_str());
continue;
}
// In OmpSs copies we may fix the data-sharing to something more natural
if (in_ompss_mode)
{
Symbol sym = expr.get_base_symbol();
// In OmpSs, the storage of a copy is always SHARED. Note that with this
// definition we aren't defining the data-sharings of the variables involved
// in that expression.
//
// About the data-sharings of the variables involved in the copy expression:
// - Fortran: the base symbol of the copy expression is always SHARED
// - C/C++:
// * The base symbol of a trivial copy (i.e the expression is a symbol) must always be SHARED:
// int x, a[10];
// copy_inout(x) -> shared(x)
// copy_inout(a) -> shared(a)
// * The base symbol of an array expression or a reference to an array must be SHARED too:
// copy_int a[10];
// copy_inout(a[4]) -> shared(a)
// copy_inout(a[1:2]) -> shared(a)
// * The base symbol of a class member access must be shared too:
// struct C { int z; } c;
// copy_inout(c.z) -> shared(c)
// * Otherwise, the data-sharing of the base symbol is FIRSTPRIVATE:
// int* p;
// copy_inout(*p) -> firstprivate(p)
// copy_inout(p[10]) -> firstprivate(p)
// copy_inout(p[1:2]) -> firstprivate(p)
// copy_inout([10][20] p) -> firstprivate(p)
if (IS_FORTRAN_LANGUAGE)
{
data_sharing_environment.set_data_sharing(sym, DS_SHARED, DSK_IMPLICIT,
"the variable is mentioned in a copy and it did not have an explicit data-sharing");
}
else if (expr.is<Nodecl::Symbol>())
{
data_sharing_environment.set_data_sharing(sym, DS_SHARED, DSK_IMPLICIT,
"the variable is mentioned in a copy and it did not have an explicit data-sharing");
}
else if (sym.get_type().is_array()
|| (sym.get_type().is_any_reference()
&& sym.get_type().references_to().is_array()))
{
data_sharing_environment.set_data_sharing(sym, DS_SHARED, DSK_IMPLICIT,
"the variable is an array mentioned in a non-trivial copy "
"and it did not have an explicit data-sharing");
}
else if (sym.get_type().is_class())
{
data_sharing_environment.set_data_sharing(sym, DS_SHARED, DSK_IMPLICIT,
"the variable is an object mentioned in a non-trivial dependence "
"and it did not have an explicit data-sharing");
}
else
{
data_sharing_environment.set_data_sharing(sym, DS_FIRSTPRIVATE, DSK_IMPLICIT,
"the variable is a non-array mentioned in a non-trivial copy "
"and it did not have an explicit data-sharing");
}
}
TL::OmpSs::CopyItem copy_item(expr, copy_direction);
items.append(copy_item);
}
switch (copy_direction)
{
case TL::OmpSs::COPY_DIR_IN:
{
target_info.append_to_copy_in(items);
break;
}
case TL::OmpSs::COPY_DIR_OUT:
{
target_info.append_to_copy_out(items);
break;
}
case TL::OmpSs::COPY_DIR_INOUT:
{
target_info.append_to_copy_inout(items);
break;
}
default:
{
internal_error("Unreachable code", 0);
}
}
}
void VectorizerVisitorFunction::visit(const Nodecl::FunctionCode& function_code)
{
// Set up enviroment
_environment._external_scope =
function_code.retrieve_context();
_environment._local_scope_list.push_back(
function_code.get_statements().retrieve_context());
// Get analysis info
Vectorizer::initialize_analysis(function_code);
// Push FunctionCode as scope for analysis
_environment._analysis_scopes.push_back(function_code);
//Vectorize function type and parameters
TL::Symbol vect_func_sym = function_code.get_symbol();
TL::Type func_type = vect_func_sym.get_type();
TL::ObjectList<TL::Symbol> parameters = vect_func_sym.get_function_parameters();
TL::ObjectList<TL::Type> parameters_type = func_type.parameters();
TL::ObjectList<TL::Type> parameters_vector_type;
TL::ObjectList<TL::Type>::iterator it_type;
TL::ObjectList<TL::Symbol>::iterator it_param_sym;
for(it_param_sym = parameters.begin(), it_type = parameters_type.begin();
it_type != parameters_type.end();
it_param_sym++, it_type++)
{
TL::Type sym_type = get_qualified_vector_to((*it_type), _environment._vector_length);
// Set type to parameter TL::Symbol
(*it_param_sym).set_type(sym_type);
parameters_vector_type.append(sym_type);
}
if(_masked_version)
{
TL::Scope scope = vect_func_sym.get_related_scope();
// Create mask parameter
TL::Symbol mask_sym = scope.new_symbol("__mask_param");
mask_sym.get_internal_symbol()->kind = SK_VARIABLE;
mask_sym.get_internal_symbol()->entity_specs.is_user_declared = 1;
mask_sym.set_type(TL::Type::get_mask_type(_environment._mask_size));
symbol_set_as_parameter_of_function(mask_sym.get_internal_symbol(),
vect_func_sym.get_internal_symbol(),
parameters.size());
// Add mask symbol and type to parameters
parameters.append(mask_sym);
parameters_vector_type.append(mask_sym.get_type());
vect_func_sym.set_related_symbols(parameters);
// Take care of default_argument_info_t*
//TODO: Move this into a function
{
int num_parameters =
vect_func_sym.get_internal_symbol()->entity_specs.num_parameters;
default_argument_info_t** default_argument_info =
vect_func_sym.get_internal_symbol()->entity_specs.default_argument_info;
num_parameters++;
default_argument_info = (default_argument_info_t**)xrealloc(default_argument_info,
num_parameters * sizeof(*default_argument_info));
default_argument_info[num_parameters-1] = NULL;
vect_func_sym.get_internal_symbol()->entity_specs.default_argument_info = default_argument_info;
vect_func_sym.get_internal_symbol()->entity_specs.num_parameters = num_parameters;
}
Nodecl::Symbol mask_nodecl_sym =
mask_sym.make_nodecl(true, function_code.get_locus());
_environment._mask_list.push_back(mask_nodecl_sym);
}
else // Add MaskLiteral to mask_list
{
Nodecl::MaskLiteral all_one_mask =
Nodecl::MaskLiteral::make(
TL::Type::get_mask_type(_environment._mask_size),
const_value_get_minus_one(_environment._mask_size, 1),
make_locus("", 0, 0));
_environment._mask_list.push_back(all_one_mask);
}
vect_func_sym.set_type(get_qualified_vector_to(func_type.returns(),
_environment._vector_length).get_function_returning(parameters_vector_type));
// Vectorize function statements
VectorizerVisitorStatement visitor_stmt(_environment);
visitor_stmt.walk(function_code.get_statements());
// Add final return if multi-return function
if (_environment._function_return.is_valid())
{
// Return value
Nodecl::Symbol return_value= _environment._function_return.make_nodecl(
false, function_code.get_locus());
// VectorizerVisitorExpression visitor_sym(_environment);
// visitor_sym.walk(return_value);
// Return value at the end of the Compound Statement
Nodecl::ReturnStatement return_stmt =
Nodecl::ReturnStatement::make(return_value, function_code.get_locus());
function_code.get_statements().as<Nodecl::Context>().get_in_context().as<Nodecl::List>()
.front().as<Nodecl::CompoundStatement>().get_statements()
.as<Nodecl::List>().append(return_stmt);
}
_environment._analysis_scopes.pop_back();
_environment._mask_list.pop_back();
// Analysis in functions won't be reused anywhere so it must be freed
Vectorizer::finalize_analysis();
}
void DeviceFPGA::add_hls_pragmas(
Nodecl::NodeclBase &task,
TL::ObjectList<OutlineDataItem*> &data_items
)
{
/*
* Insert hls pragmas in order to denerate input/output connections
* Every parameter needs a directive:
* scalar: create plain wire connections:
* #pragma HLS INTERFACE ap_none port=VAR
* #pragma AP resource core=AXI_SLAVE variable=VAR metadata="-bus_bundle AXIlite"
*
* Array; create fifo port to be handled by axi stream
* #pragma HLS stream variable=VAR <-- NOT NEEDED
* #pragma HLS resource core=AXI4Stream variable=VAR
* #pragma HLS interface ap_fifo port=VAR
*
* For every task there is a control bus defined to kick the accelerator off:
*
* #pragma AP resource core=AXI_SLAVE variable=return metadata="-bus_bundle AXIlite" \
* port_map={{ap_start START} {ap_done DONE} {ap_idle IDLE} {ap_return RETURN}}
*
* All of this stuff must be inside the function body i.e.
*
* void foo(...)
* {
* pragma stuff
* function body
* }
*
*/
//see what kind of ast it really is
std::cerr << ast_node_type_name(task.get_kind())
<< " in_list: " << task.is_in_list()
<< " locus: " << task.get_locus()
<< std::endl;
//Dig into the tree and find where the function statements are
ObjectList<Nodecl::NodeclBase> tchildren = task.children();
Nodecl::NodeclBase& context = tchildren.front();
ObjectList<Nodecl::NodeclBase> cchildren = context.children();
Nodecl::List list(cchildren.front().get_internal_nodecl());
Nodecl::List stlist(list.begin()->children().front().get_internal_nodecl());
Nodecl::UnknownPragma ctrl_bus = Nodecl::UnknownPragma::make(
"AP resource core=AXI_SLAVE variable=return metadata=\"-bus_bundle AXIlite\" port_map={{ap_start START} {ap_done DONE} {ap_idle IDLE} {ap_return RETURN}}");
stlist.prepend(ctrl_bus);
//since we are using prepend, everything is going to appar in reverse order
//but this may not be a real issue
// TL::ObjectList<OutlineDataItem*> data_items = outline_info.get_data_items();
for (TL::ObjectList<OutlineDataItem*>::iterator it = data_items.begin();
it != data_items.end();
it++)
{
std::string field_name = (*it)->get_field_name();
Nodecl::UnknownPragma pragma_node;
if ((*it)->get_copies().empty())
{
//set scalar argumenit pragmas
pragma_node = Nodecl::UnknownPragma::make("HLS INTERFACE ap_none port=" + field_name);
stlist.prepend(pragma_node);
pragma_node = Nodecl::UnknownPragma::make("AP resource core=AXI_SLAVE variable="
+ field_name
+ " metadata=\"-bus_bundle AXIlite\"");
stlist.prepend(pragma_node);
}
else
{
//set array/stream pragmas
pragma_node = Nodecl::UnknownPragma::make(
"HLS resource core=AXI4Stream variable=" + field_name);
stlist.prepend(pragma_node);
pragma_node = Nodecl::UnknownPragma::make(
"HLS interface ap_fifo port=" + field_name);
stlist.prepend(pragma_node);
}
}
}
/*
* We may need to set scalar arguments here, but not transfers
*/
Source DeviceFPGA::fpga_param_code(
TL::ObjectList<OutlineDataItem*> &data_items,
Nodecl::Utils::SymbolMap *symbol_map,//we may not need it
Scope sc
)
{
//Nodecl::Utils::SimpleSymbolMap *ssmap = (Nodecl::Utils::SimpleSymbolMap*)symbol_map;
//TL::ObjectList<OutlineDataItem*> data_items = outline_info.get_data_items();
Source args_src;
/*
* Get the fpga handle to write the data that we need.
*
* TODO: Make sure mmap + set arg does not break when we don't have scalar arguments
*/
args_src
<< "int fd = open(\"/dev/mem\", NANOS_O_RDWR);" //2=O_RDWR
// << "unsigned int acc_addr = NANOS_AXI_BASE_ADDRESS;"
// << "printf(\"address: %x\\n\", acc_addr);"
<< "unsigned int *acc_handle = "
<< " (unsigned int *) mmap(0, NANOS_MMAP_SIZE," //0=NULL
<< " NANOS_PROT_READ|NANOS_PROT_WRITE, NANOS_MAP_SHARED," //" PROT_READ | PROT_WRITE, MAP_SHARED"
<< " fd, NANOS_AXI_BASE_ADDRESS);"
;
//set scalar arguments
/* FIXME
* We assume that the base address to set scalar parameters
* (which appears to be true)
* In fact all of this is defined in the
* impl/pcores/foo_top_v1_00_a/include/xfoo_AXIlite.h
* where foo is the name of the function and top_v1_00_a is the version name
* This path may change so we are assuming base addres does not
*/
/*
* Parameter have an offset of 8 bytes with the preceding one (except for 64bit ones)
* If any parameter is smaller than 32bit (4byte), padding is added in between
* If a paramater is 64bit(aka long long int) another 32bit of padding are added
*/
int argIndex = 0x14/4; //base address/(sizeof(int)=4)
for (TL::ObjectList<OutlineDataItem*>::iterator it = data_items.begin();
it != data_items.end();
it++)
{
// print_dataItem_info(*it, sc);
Symbol outline_symbol = symbol_map->map((*it)->get_symbol());
const TL::ObjectList<OutlineDataItem::CopyItem> &copies = (*it)->get_copies();
//if copies are empty, we need to set the scalar value
if (copies.empty())
{
const Type & type = (*it)->get_field_type();
args_src
<< "acc_handle[" << argIndex << "] = "
<< outline_symbol.get_name() << ";"
;
//+1 for field +1 for padding
//we are adding an index to int[] => addresses are 4x
argIndex+=2;
if (type.get_size() >= 8)
{
argIndex ++;
}
}
}
/*
* There should be a control bus mapped at 0x40440000
* This is true if function has non-scalar parameters
* To start the device we must set the first bt to 1
*/
args_src << "acc_handle[0] = 1;"
<< "munmap(acc_handle, NANOS_MMAP_SIZE);"
<< "close(fd);"
;
return args_src;
}
List List::make(const TL::ObjectList<NodeclBase>& list)
{
if (list.empty())
return nodecl_null();
return make_list_helper(list.begin(), list.end() - 1);
}
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;
}
