HLS allocation constraints fix (ferrandi#118)

* HLS allocation constraints fix

* Dominator allocation instance count guard

* Added relative connection offset tag to NG-ULTRA

* HLS backend default dependence fix

* Code refactoring

* Inlined functions need -O2 for the testbench co-simulation.

* STG info computation update

* Fixed a scheduling problem.

* Fixed a possible scheduling conflict between load/store and reading/writing functions.

* STG optimization conditions fix

* Technology manager get_fu overload
get_fu version to search for both functional unit and library name

* Time model unused field

* Added multiple top functions allocation support

* Functions to library components binding fix

* Function dominator allocation update

* Function scope memory interface
Memory allocation policy and channels are now function scope;
memory dominator allocation update

* Register binding refectoring

---------

Co-authored-by: Fabrizio Ferrandi <[email protected]>
Co-authored-by: Claudio Barone <[email protected]>

QTCreator-PandA-GitHub.files

@@ -1663,6 +1663,8 @@ src/HLS/hls_constraints.cpp
 src/HLS/hls_constraints.hpp
 src/HLS/hls_flow/classical_synthesis_flow.cpp
 src/HLS/hls_flow/classical_synthesis_flow.hpp
+src/HLS/hls_flow/dominator_allocation.cpp
+src/HLS/hls_flow/dominator_allocation.hpp
 src/HLS/hls_flow/initialize_hls.cpp
 src/HLS/hls_flow/initialize_hls.hpp
 src/HLS/hls_flow/synthesis/hls_synthesis_flow.cpp

etc/devices/NanoXplore_devices/nx2h540tsc-seed.xml

@@ -14,6 +14,7 @@
     <DSPs_ry_sizes value="16" />
     <max_lut_size value="4" />
     <OutputDSPConnectionRatio value="3" />
+    <RelativeConnectionOffset value="2.0" />
     <OutputCarryConnectionRatio value="3" />
     <use_soft_64_mul value="1" />
   </device>

etc/devices/NanoXplore_devices/nx2h540tsc.xml

@@ -13,6 +13,7 @@
     <DSPs_x_sizes value="18"/>
     <DSPs_y_sizes value="24"/>
     <OutputCarryConnectionRatio value="3"/>
+    <RelativeConnectionOffset value="2.0" />
     <OutputDSPConnectionRatio value="3"/>
     <max_lut_size value="4"/>
     <use_soft_64_mul value="1"/>

panda_regressions/hls/gcc_regression_simple/builtins/printf.c

@@ -18,14 +18,16 @@ main_test (void)
   printf ("%s\n", "hello");
   printf ("%s\n", *s2);
   s3 = s2;
-  printf ("%s\n", *s3++);
+  printf ("%s\n", *s3);
+  s3++;
   if (s3 != s2+1 || *s3 != 0)
     abort();
 
   printf ("%c", '\n');
   printf ("%c", **s2);
   s3 = s2;
-  printf ("%c", **s3++);
+  printf ("%c", **s3);
+  s3++;
   if (s3 != s2+1 || *s3 != 0)
     abort();
 

src/HLS/architecture_creation/TopEntityMemoryMapped.cpp

@@ -101,6 +101,9 @@ void TopEntityMemoryMapped::Initialize()
    const auto is_addressed_fun = HLSMgr->hasToBeInterfaced(funId) && !is_root_function;
    needMemoryMappedRegisters = is_wb4_root || is_addressed_fun || parameters->getOption<bool>(OPT_memory_mapped_top);
    AddedComponents.clear();
+   const auto FB = HLSMgr->CGetFunctionBehavior(funId);
+   _channels_number = FB->GetChannelsNumber();
+   _channels_type = FB->GetChannelsType();
 }
 
 DesignFlowStep_Status TopEntityMemoryMapped::InternalExec()
@@ -256,9 +259,8 @@ void TopEntityMemoryMapped::insertMemoryMappedRegister(structural_managerRef SM_
    const auto FB = HLSMgr->CGetFunctionBehavior(funId);
    const auto BH = FB->CGetBehavioralHelper();
 
-   structural_objectRef controlSignal = interfaceObj->find_member("ControlSignal", signal_o_K, interfaceObj);
-   bool multi_channel_bus = parameters->getOption<MemoryAllocation_ChannelsType>(OPT_channels_type) ==
-                            MemoryAllocation_ChannelsType::MEM_ACC_NN;
+   const auto controlSignal = interfaceObj->find_member("ControlSignal", signal_o_K, interfaceObj);
+   const auto multi_channel_bus = _channels_type == MemoryAllocation_ChannelsType::MEM_ACC_NN;
 
    for(const auto& function_parameter : function_parameters)
    {
@@ -277,8 +279,7 @@ void TopEntityMemoryMapped::insertMemoryMappedRegister(structural_managerRef SM_
                                                     interfaceObj, HLS->HLS_T->get_technology_manager());
       if(multi_channel_bus)
       {
-         resizing_IO(GetPointer<module>(memoryMappedRegister),
-                     parameters->getOption<unsigned int>(OPT_channels_number));
+         resizing_IO(GetPointerS<module>(memoryMappedRegister), _channels_number);
       }
       GetPointerS<module>(memoryMappedRegister)
           ->SetParameter("ALLOCATED_ADDRESS",
@@ -339,7 +340,7 @@ void TopEntityMemoryMapped::insertMemoryMappedRegister(structural_managerRef SM_
                                                  interfaceObj, HLS->HLS_T->get_technology_manager());
    if(multi_channel_bus)
    {
-      resizing_IO(GetPointer<module>(returnRegister), parameters->getOption<unsigned int>(OPT_channels_number));
+      resizing_IO(GetPointer<module>(returnRegister), _channels_number);
    }
    GetPointer<module>(returnRegister)
        ->SetParameter("ALLOCATED_ADDRESS", HLSMgr->Rmem->get_symbol(returnType, HLS->functionId)->get_symbol_name());
@@ -401,15 +402,14 @@ void TopEntityMemoryMapped::insertStartDoneLogic(structural_managerRef SM_mm, st
 
    if(HLSMgr->CGetCallGraphManager()->ExistsAddressedFunction())
    {
-      bool multi_channel_bus = parameters->getOption<MemoryAllocation_ChannelsType>(OPT_channels_type) ==
-                               MemoryAllocation_ChannelsType::MEM_ACC_NN;
-      std::string component_name = multi_channel_bus ? NOTYFY_CALLER_MINIMALN_FU : NOTYFY_CALLER_MINIMAL_FU;
-      structural_objectRef notifyCaller = SM_mm->add_module_from_technology_library(
+      const auto multi_channel_bus = _channels_type == MemoryAllocation_ChannelsType::MEM_ACC_NN;
+      const auto component_name = multi_channel_bus ? NOTYFY_CALLER_MINIMALN_FU : NOTYFY_CALLER_MINIMAL_FU;
+      const auto notifyCaller = SM_mm->add_module_from_technology_library(
           "notifyCaller", component_name, HLS->HLS_T->get_technology_manager()->get_library(component_name),
           interfaceObj, HLS->HLS_T->get_technology_manager());
       if(multi_channel_bus)
       {
-         resizing_IO(GetPointer<module>(notifyCaller), parameters->getOption<unsigned int>(OPT_channels_number));
+         resizing_IO(GetPointerS<module>(notifyCaller), _channels_number);
       }
       setBusSizes(notifyCaller, HLSMgr);
       connectClockAndReset(SM_mm, interfaceObj, notifyCaller);
@@ -428,9 +428,8 @@ void TopEntityMemoryMapped::insertStartDoneLogic(structural_managerRef SM_mm, st
 
 void TopEntityMemoryMapped::insertStatusRegister(structural_managerRef SM_mm, structural_objectRef wrappedObj)
 {
-   bool multi_channel_bus = parameters->getOption<MemoryAllocation_ChannelsType>(OPT_channels_type) ==
-                            MemoryAllocation_ChannelsType::MEM_ACC_NN;
-   structural_objectRef interfaceObj = SM_mm->get_circ();
+   const auto multi_channel_bus = _channels_type == MemoryAllocation_ChannelsType::MEM_ACC_NN;
+   const auto interfaceObj = SM_mm->get_circ();
    if(HLSMgr->CGetCallGraphManager()->ExistsAddressedFunction())
    {
       const auto component_name = multi_channel_bus ? STATUS_REGISTERN_FU : STATUS_REGISTER_FU;
@@ -439,7 +438,7 @@ void TopEntityMemoryMapped::insertStatusRegister(structural_managerRef SM_mm, st
           interfaceObj, HLS->HLS_T->get_technology_manager());
       if(multi_channel_bus)
       {
-         resizing_IO(GetPointer<module>(statusRegister), parameters->getOption<unsigned int>(OPT_channels_number));
+         resizing_IO(GetPointer<module>(statusRegister), _channels_number);
       }
       GetPointerS<module>(statusRegister)
           ->SetParameter("ALLOCATED_ADDRESS",
@@ -482,7 +481,7 @@ void TopEntityMemoryMapped::insertStatusRegister(structural_managerRef SM_mm, st
           interfaceObj, HLS->HLS_T->get_technology_manager());
       if(multi_channel_bus)
       {
-         resizing_IO(GetPointer<module>(statusRegister), parameters->getOption<unsigned int>(OPT_channels_number));
+         resizing_IO(GetPointer<module>(statusRegister), _channels_number);
       }
       GetPointerS<module>(statusRegister)
           ->SetParameter("ALLOCATED_ADDRESS",
@@ -553,7 +552,7 @@ static void propagateInterface(structural_managerRef SM, structural_objectRef wr
          continue;
       }
 
-      SM->add_port(portObj->get_id(), portObj->get_port_direction(), interfaceObj, portObj->get_typeRef());
+      SM->add_port(portID, portObj->get_port_direction(), interfaceObj, portObj->get_typeRef());
    }
 
    connectClockAndReset(SM, interfaceObj, wrappedObj);

src/HLS/architecture_creation/TopEntityMemoryMapped.hpp

@@ -43,6 +43,7 @@
 #include <list>
 #include <string>
 class module;
+enum class MemoryAllocation_ChannelsType;
 
 /**
  * @brief Build a wrapper layer on the top entity implementing the
@@ -92,6 +93,12 @@ class TopEntityMemoryMapped : public top_entity
    /// true when the module is a root function
    bool is_root_function{false};
 
+   /// Function scope channels number
+   unsigned int _channels_number;
+
+   /// Function scope channels type
+   MemoryAllocation_ChannelsType _channels_type;
+
    /**
     * Allocates the in/out parameters of the module as internal registers
     */

src/HLS/architecture_creation/datapath_creation/datapath_parallel_cs.cpp

@@ -89,8 +89,8 @@ datapath_parallel_cs::ComputeHLSRelationships(const DesignFlowStep::Relationship
    {
       case DEPENDENCE_RELATIONSHIP:
       {
-         ret.insert(std::make_tuple(HLSFlowStep_Type::INITIALIZE_HLS, HLSFlowStepSpecializationConstRef(),
-                                    HLSFlowStep_Relationship::SAME_FUNCTION));
+         ret.insert(std::make_tuple(HLSFlowStep_Type::DOMINATOR_ALLOCATION, HLSFlowStepSpecializationConstRef(),
+                                    HLSFlowStep_Relationship::WHOLE_APPLICATION));
          ret.insert(std::make_tuple(HLSFlowStep_Type::OMP_BODY_LOOP_SYNTHESIS_FLOW, HLSFlowStepSpecializationConstRef(),
                                     HLSFlowStep_Relationship::CALLED_FUNCTIONS));
          break;

src/HLS/architecture_creation/top_entity.cpp

@@ -404,7 +404,8 @@ void top_entity::add_ports(structural_objectRef circuit, structural_objectRef cl
          top_obj = SM->add_port(FB->CGetBehavioralHelper()->PrintVariable(function_parameter), port_o::IN, circuit,
                                 port_type);
       }
-      if(has_registered_inputs)
+      const auto is_pipelined = HLSMgr->CGetFunctionBehavior(funId)->is_simple_pipeline();
+      if(has_registered_inputs && !is_pipelined)
       {
          const auto port_prefix = GetPointerS<port_o>(in_obj)->get_id();
          if(in_obj->get_kind() == port_vector_o_K)
@@ -657,8 +658,8 @@ void top_entity::add_ports(structural_objectRef circuit, structural_objectRef cl
 
 void top_entity::add_command_signals(structural_objectRef circuit)
 {
-   structural_objectRef Datapath = HLS->datapath->get_circ();
-   structural_objectRef Controller = HLS->controller->get_circ();
+   const auto Datapath = HLS->datapath->get_circ();
+   const auto Controller = HLS->controller->get_circ();
 
    for(const auto& selector : HLS->Rconn->GetSelectors())
    {

src/HLS/backend/generate_hdl.cpp

@@ -87,7 +87,7 @@ generate_hdl::ComputeHLSRelationships(const DesignFlowStep::RelationshipType rel
       {
          ret.insert(std::make_tuple(HLSFlowStep_Type::HLS_SYNTHESIS_FLOW, HLSFlowStepSpecializationConstRef(),
                                     HLSFlowStep_Relationship::ALL_FUNCTIONS));
-         if(parameters->isOption(OPT_discrepancy_hw) and parameters->getOption<bool>(OPT_discrepancy_hw))
+         if(parameters->isOption(OPT_discrepancy_hw) && parameters->getOption<bool>(OPT_discrepancy_hw))
          {
             ret.insert(std::make_tuple(HLSFlowStep_Type::HW_DISCREPANCY_ANALYSIS, HLSFlowStepSpecializationConstRef(),
                                        HLSFlowStep_Relationship::TOP_FUNCTION));
@@ -110,20 +110,19 @@ generate_hdl::ComputeHLSRelationships(const DesignFlowStep::RelationshipType rel
 
 generate_hdl::~generate_hdl() = default;
 
+bool generate_hdl::HasToBeExecuted() const
+{
+   return true;
+}
+
 DesignFlowStep_Status generate_hdl::Exec()
 {
-   HDL_managerRef HM =
-       HDL_managerRef(new HDL_manager(HLSMgr, HLSMgr->get_HLS_target()->get_target_device(), parameters));
-   auto file_name = parameters->getOption<std::string>(OPT_top_file);
+   HDL_manager HM(HLSMgr, HLSMgr->get_HLS_target()->get_target_device(), parameters);
+   const auto file_name = parameters->getOption<std::string>(OPT_top_file);
    std::list<structural_objectRef> top_circuits;
    const auto root_functions = HLSMgr->CGetCallGraphManager()->GetRootFunctions();
    std::transform(root_functions.begin(), root_functions.end(), std::back_inserter(top_circuits),
                   [&](unsigned int top_function) { return HLSMgr->get_HLS(top_function)->top->get_circ(); });
-   HM->hdl_gen(file_name, top_circuits, false, HLSMgr->hdl_files, HLSMgr->aux_files);
+   HM.hdl_gen(file_name, top_circuits, false, HLSMgr->hdl_files, HLSMgr->aux_files);
    return DesignFlowStep_Status::SUCCESS;
 }
-
-bool generate_hdl::HasToBeExecuted() const
-{
-   return true;
-}

src/HLS/binding/interconnection/conn_binding_creator.cpp

@@ -126,12 +126,12 @@ conn_binding_creator::ComputeHLSRelationships(const DesignFlowStep::Relationship
 
 void conn_binding_creator::add_parameter_ports()
 {
-   const FunctionBehaviorConstRef FB = HLSMgr->CGetFunctionBehavior(funId);
-   const BehavioralHelperConstRef BH = FB->CGetBehavioralHelper();
+   const auto FB = HLSMgr->CGetFunctionBehavior(funId);
+   const auto BH = FB->CGetBehavioralHelper();
    input_ports.clear();
    output_ports.clear();
    /// list containing the parameters of the original function (representing input and output values)
-   const std::list<unsigned int>& function_parameters = BH->get_parameters();
+   const auto function_parameters = BH->get_parameters();
    PRINT_DBG_STRING(DEBUG_LEVEL_PEDANTIC, debug_level, "Parameters values: ");
    for(auto function_parameter : function_parameters)
    {

src/HLS/binding/interconnection/conn_binding_cs.cpp

@@ -100,7 +100,7 @@ void conn_binding_cs::instantiate_suspension_component(const HLS_managerRef HLSM
       {
          structural_objectRef portStart = curr_gate->find_member(STR(START_PORT_NAME), port_o_K, curr_gate);
          structural_objectRef startMemOp = GetPointer<port_o>(portStart)->find_bounded_object();
-         THROW_ASSERT(startMemOp != NULL, "No start port for mem_ctrl_found");
+         THROW_ASSERT(startMemOp != nullptr, "No start port for mem_ctrl_found");
          andStartMemOp_required = true;
          break;
       }
@@ -188,7 +188,7 @@ void conn_binding_cs::instantiate_suspension_component(const HLS_managerRef HLSM
       {
          structural_objectRef portStart = curr_gate->find_member(STR(START_PORT_NAME), port_o_K, curr_gate);
          structural_objectRef startMemOp = GetPointer<port_o>(portStart)->find_bounded_object();
-         THROW_ASSERT(startMemOp != NULL, "No start port for mem_ctrl_found");
+         THROW_ASSERT(startMemOp != nullptr, "No start port for mem_ctrl_found");
          SM->add_connection(startMemOp, GetPointer<port_o>(port_in_and)->get_port(1));
          break;
       }

src/HLS/binding/module/cdfc_module_binding.cpp

@@ -1950,26 +1950,30 @@ DesignFlowStep_Status cdfc_module_binding::InternalExec()
             total_area_muxes_prev = total_area_muxes;
             total_area_muxes = total_area_muxes_initial;
 
-            DesignFlowStepRef regb;
-            // if(iteration%2)
-            if(parameters->getOption<HLSFlowStep_Type>(OPT_register_allocation_algorithm) ==
-               HLSFlowStep_Type::WEIGHTED_CLIQUE_REGISTER_BINDING)
             {
-               regb = GetPointer<const HLSFlowStepFactory>(design_flow_manager.lock()->CGetDesignFlowStepFactory("HLS"))
+               DesignFlowStepRef regb;
+               // if(iteration%2)
+               if(parameters->getOption<HLSFlowStep_Type>(OPT_register_allocation_algorithm) ==
+                  HLSFlowStep_Type::WEIGHTED_CLIQUE_REGISTER_BINDING)
+               {
+                  regb =
+                      GetPointer<const HLSFlowStepFactory>(design_flow_manager.lock()->CGetDesignFlowStepFactory("HLS"))
                           ->CreateHLSFlowStep(
                               HLSFlowStep_Type::WEIGHTED_CLIQUE_REGISTER_BINDING, funId,
                               HLSFlowStepSpecializationConstRef(new WeightedCliqueRegisterBindingSpecialization(
                                   parameters->getOption<CliqueCovering_Algorithm>(
                                       OPT_weighted_clique_register_algorithm))));
-            }
-            else
-            {
-               regb = GetPointer<const HLSFlowStepFactory>(design_flow_manager.lock()->CGetDesignFlowStepFactory("HLS"))
+               }
+               else
+               {
+                  regb =
+                      GetPointer<const HLSFlowStepFactory>(design_flow_manager.lock()->CGetDesignFlowStepFactory("HLS"))
                           ->CreateHLSFlowStep(
                               parameters->getOption<HLSFlowStep_Type>(OPT_register_allocation_algorithm), funId);
+               }
+               regb->Initialize();
+               regb->Exec();
             }
-            regb->Initialize();
-            regb->Exec();
          }
 
          if(output_level >= OUTPUT_LEVEL_VERBOSE)

src/HLS/binding/module/easy_module_binding.cpp

@@ -113,8 +113,8 @@ easy_module_binding::ComputeHLSRelationships(const DesignFlowStep::RelationshipT
    {
       case DEPENDENCE_RELATIONSHIP:
       {
-         ret.insert(std::make_tuple(HLSFlowStep_Type::INITIALIZE_HLS, HLSFlowStepSpecializationConstRef(),
-                                    HLSFlowStep_Relationship::SAME_FUNCTION));
+         ret.insert(std::make_tuple(HLSFlowStep_Type::DOMINATOR_ALLOCATION, HLSFlowStepSpecializationConstRef(),
+                                    HLSFlowStep_Relationship::WHOLE_APPLICATION));
          if(HLSMgr->get_HLS(funId))
          {
             ret.insert(std::make_tuple(HLSMgr->get_HLS(funId)->chaining_algorithm, HLSFlowStepSpecializationConstRef(),
@@ -146,18 +146,18 @@ DesignFlowStep_Status easy_module_binding::InternalExec()
    const auto TM = HLSMgr->get_tree_manager();
    // resource binding and allocation  info
    fu_binding& fu = *(HLS->Rfu);
-   const AllocationInformationConstRef allocation_information = HLS->allocation_information;
+   const auto allocation_information = HLS->allocation_information;
    // pointer to a Control, Data dependence and antidependence graph graph
-   const FunctionBehaviorConstRef FB = HLSMgr->CGetFunctionBehavior(funId);
-   const OpGraphConstRef sdg = FB->CGetOpGraph(FunctionBehavior::SDG);
+   const auto FB = HLSMgr->CGetFunctionBehavior(funId);
+   const auto sdg = FB->CGetOpGraph(FunctionBehavior::SDG);
 
    unsigned int fu_unit;
    /// compute unshared resources
    std::map<unsigned int, unsigned int> n_shared_fu;
    for(const auto operation : sdg->CGetOperations())
    {
       const auto id = sdg->CGetOpNodeInfo(operation)->GetNodeId();
-      if(id == ENTRY_ID or id == EXIT_ID)
+      if(id == ENTRY_ID || id == EXIT_ID)
       {
          continue;
       }
@@ -175,7 +175,7 @@ DesignFlowStep_Status easy_module_binding::InternalExec()
          n_shared_fu[fu_unit] = 1 + n_shared_fu[fu_unit];
       }
    }
-   if(output_level >= OUTPUT_LEVEL_MINIMUM and output_level <= OUTPUT_LEVEL_PEDANTIC)
+   if(output_level >= OUTPUT_LEVEL_MINIMUM && output_level <= OUTPUT_LEVEL_PEDANTIC)
    {
       INDENT_OUT_MEX(OUTPUT_LEVEL_MINIMUM, output_level, "");
    }
@@ -247,7 +247,7 @@ DesignFlowStep_Status easy_module_binding::InternalExec()
       INDENT_OUT_MEX(OUTPUT_LEVEL_MINIMUM, output_level,
                      "---Bound operations:" + STR(easy_bound_vertices.size()) + "/" + STR(boost::num_vertices(*sdg)));
    }
-   if(output_level >= OUTPUT_LEVEL_MINIMUM and output_level <= OUTPUT_LEVEL_PEDANTIC)
+   if(output_level >= OUTPUT_LEVEL_MINIMUM && output_level <= OUTPUT_LEVEL_PEDANTIC)
    {
       STOP_TIME(step_time);
       INDENT_OUT_MEX(OUTPUT_LEVEL_MINIMUM, output_level,