Skip to content

Latest commit

 

History

History
361 lines (288 loc) · 15.4 KB

DeveloperGuide.md

File metadata and controls

361 lines (288 loc) · 15.4 KB

Developer Guide

Tool Philosophy

OpenROAD is a tool to build a chip from synthesizable RTL (Verilog) to completed physical layout (manufacturable, tapeout-clean GDSII).

The unifying principle behind the design of OpenROAD is for all of the tools to reside in one tool, with one process, and one database. All tools in the flow should use Tcl commands exclusively to control them instead of external "configuration files". File-based communication between tools and forking processes is strongly discouraged. This architecture streamlines the construction of a flexible tool flow and minimizes the overhead of invoking each tool in the flow.

Tool File Organization

Every tool follows the following file structure, grouping sources, tests and headers together.

  • src/ This folder contains the source files for individual tools.
src Purpose
CMakeLists.txt add_subdirectory for each tool
tool/src sources and private headers
tool/src/CMakeLists.txt tool specific CMake file
tool/include/tool exported headers
tool/test tool tests
tool/regression tool unit tests
  • OpenROAD repository: This folder contains the top-level files for overall compilation. OpenROAD uses swig that acts as a wrapper for C/C++ programs to be callable in higher-level languages, such as Python and Tcl.
OpenROAD Purpose
CMakeLists.txt top-level CMake file
src/Main.cc main file
src/OpenROAD.cc OpenROAD class functions
src/OpenROAD.i top-level swig, includes, tool swig files
src/OpenROAD.tcl basic read/write lef/def/db commands
include/ord/OpenROAD.hh OpenROAD top-level class, has instances of tools

Some tools such as OpenSTA are submodules, which are simply subdirectories in src/ that are pointers to the git submodule. They are intentionally not segregated into a separate module.

The use of submodules for new code integrated into OpenROAD is strongly discouraged. Submodules make changes to the underlying infrastructure (e.g., OpenSTA) difficult to propagate across the dependent submodule repositories.

Where external/third-party code that a tool depends on should be placed depends on the nature of the dependency.

  • Libraries - code packaged as a linkable library. Examples are tcl, boost, zlib, eigen, lemon, spdlog.

These should be installed in the build environment and linked by OpenROAD. Document these dependencies in the top-level README.md file. The Dockerfile should be updated to illustrate where to find the library and how to install it. Adding libraries to the build environment requires coordination with system administrators, so that continuous integration hosts ensure that environments include the dependency. Advance notification should also be given to the development team so that their private build environments can be updated.

Each tool CMake file builds a library that is linked by the OpenROAD application. The tools should not define a main() function. If the tool is Tcl only and has no C++ code, it does not need to have a CMake file. Tool CMake files should not include the following:

  • cmake_minimum_required
  • GCC_COVERAGE_COMPILE_FLAGS
  • GCC_COVERAGE_LINK_FLAGS
  • CMAKE_CXX_FLAGS
  • CMAKE_EXE_LINKER_FLAGS

None of the tools have commands to read or write LEF, DEF, Verilog or database files. For consistency, these functions are all provided by the OpenROAD framework.

Tools should package all of their state in a single class. An instance of each tool class resides in the top-level OpenROAD object. This allows multiple tools to exist at the same time. If any tool keeps state in global variables (even static), then only one tool can exist at a time. Many of the tools being integrated were not built with this goal in mind and will only work on one design at a time.

Each tool should use a unique namespace for all of its code. The same namespace should be used for Tcl functions, including those defined by a swig interface file. Internal Tcl commands stay inside the namespace, and user visible Tcl commands should be defined in the global namespace. User commands should be simple Tcl commands such as global_placement that do not create tool instances that must be based to the commands. Defining Tcl commands for a tool class is fine for internal commands, but not for user visible commands. Commands have an implicit argument of the current OpenROAD class object. Functions to get individual tools from the OpenROAD object can be defined.

Initialization (C++ tools only)

The OpenROAD class has pointers to each tool, with functions to get each tool. Each tool has (at a minimum) a function to make an instance of the tool class, an initialization function that is called after all of the tools have been made, and a function to delete the tool. This small header does not include the class definition for the tool so that the OpenROAD framework does not have to know anything about the tool internals or include a gigantic header file.

MakeTool.hh defines the following:

Tool *makeTool();
void initTool(OpenRoad *openroad);
void deleteTool(Tool *tool);

The OpenRoad::init() function calls all of the makeTool functions and then all of the initTool() functions. The init functions are called from the bottom of the tool dependencies. Each init function grabs the state it needs out of the OpenRoad instance.

Commands

Tools should provide Tcl commands to control them. Tcl object based tool interfaces are not user-friendly. Define Tcl procedures that take keyword arguments that reference the OpenRoad object to get tool state. OpenSTA has Tcl utilities to parse keyword arguments (sta::parse_keyword_args). See OpenSTA/tcl/*.tcl for examples. Use swig to define internal functions to C++ functionality.

Tcl files can be included by encoding them in CMake into a string that is evaluated at run time (See Resizer::init()).

:::{Note} Please refer to the top-level Tcl formatting guide. Our top-level Tcl files, in particular, have to be formatted in this specific manner because of the automatic parsing used to convert the READMEs into manpages. :::

Errors

Tools should report errors to the user using the ord::error function defined in include/openroad/Error.hh. ord::error throws ord::Exception. The variables ord::exit_on_error and ord::file_continue_on_error control how the error is handled. If ord::exit_on_error is true then OpenROAD reports the error and exits. If the error is encountered while reading a file with the source or read_sdc commands and ord::file_continue_on_error is false then no other commands are read from the file. The default value is false for both variables.

Test

Each "tool" has a /test directory containing a script named regression to run "unit" tests. With no arguments it should run default unit tests.

No database files should be in tests. Read LEF/DEF/Verilog to make a database.

The regression script should not depend on the current working directory. It should be able to be run from any directory. Use filenames relative to the script name rather the current working directory.

Regression scripts should print a concise summary of test failures. The regression script should return an exit code of 0 if there are no errors and 1 if there are errors. The script should not print thousands of lines of internal tool information.

Regression scripts should pass the -no_init option to openroad so that a user's init file is not sourced before the tests runs.

Regression scripts should add output files or directories to .gitignore so that running does not leave the source repository "dirty".

The Nangate45 open-source library data used by many tests is in test/Nangate45. Use the following command to add a link in the tool command:

cd src/<tool>/test
ln -s ../../../test/Nangate45

After the link is installed, the test script can read the Liberty file with the command shown below.

read_liberty Nangate45/Nangate45_typ.lib

Building

Instructions for building are available here.

Example of Adding a Tool to OpenROAD

The patch file "AddTool.patch" illustrates how to add a tool to OpenROAD. Use the following commands to add a sample tool:

# first, update existing config files
patch -p1 < docs/misc/AddTool.patch

# next, create the additional source files of the tool using this command
patch -p1 < docs/misc/AddToolFiles.patch

# finally, create the regression tests as follows
cd src/tool/test
ln -s ../../../test/regression.tcl regression.tcl

This adds a directory OpenRoad/src/tool that illustrates a tool named "Tool" that uses the file structure described above and defines a command to run the tool with keyword and flag arguments as illustrated below:

> toolize foo
Helping 23/6
Gotta positional_argument1 foo
Gotta param1 0.000000
Gotta flag1 false

> toolize -flag1 -key1 2.0 bar
Helping 23/6
Gotta positional_argument2 bar
Gotta param1 2.000000
Gotta flag1 true

> help toolize
toolize [-key1 key1] [-flag1] positional_argument1

Documentation

Tool commands should be documented in the top-level OpenROAD README.md file. Detailed documentation should be the tool/README.md file.

:::{Note} Please refer to the README formatting guide. Our top-level READMEs, in particular, have to be formatted in this specific manner because of the automatic parsing used to convert the READMEs into manpages. :::

Tool Flow Namespace

Tool namespaces are usually three-lettered lowercase letters.

  • Verilog to DB (dbSTA)
  • OpenDB: Open Database (odb)
  • TritonPart: constraints-driven paritioner (par)
  • Floorplan Initialization (ifp)
  • ICeWall chip-level connections� (pad)
  • I/O Placement (ppl)
  • PDN Generation (pdn)
  • Tapcell and Welltie Insertion (tap)
  • Triton Macro Placer (mpl)
  • Hierarchical Automatic Macro Placer (mpl2)
  • RePlAce Global Placer (gpl)
  • Gate resizing and buffering (rsz)
  • Detailed placement (dpl)
  • Clock tree synthesis (cts)
  • FastRoute Global routing (grt)
  • Antenna check and diode insertion (ant)
  • TritonRoute Detailed routing (drt)
  • Metal fill insertion (fin)
  • Design for Test (dft)
  • OpenRCX Parasitic Extraction (rcx)
  • OpenSTA timing/power analyzer (sta
  • Graphical User Interface (gui)
  • Static IR analyzer (psm)

Tool Checklist

Tools should make every attempt to minimize external dependencies. Linking libraries other than those currently in use complicates the builds and sacrifices the portability of OpenROAD. OpenROAD should be portable to many different compiler/operating system versions and dependencies make this vastly more complicated.

  1. OpenROAD submodules reference tool openroad branch head. No git develop, openroad_app, or openroad_build branches.
  2. Submodules used by more than one tool belong in src/, not duplicated in each tool repo.
  3. CMakeLists.txt does not use add_compile_options include_directories link_directories link_libraries. Use target_ versions instead. See tips here.
  4. CMakeLists.txt does not use glob. Use explicit lists of source files and headers instead.
  5. CMakeLists.txt does not define CFLAGS CMAKE_CXX_FLAGS CMAKE_CXX_FLAGS_DEBUG CMAKE_CXX_FLAGS_RELEASE. Let the top level and defaults control these.
  6. No main.cpp or main procedure.
  7. No compiler warnings for GCC or Clang with optimization enabled.
  8. Does not call flute::readLUT (called once by openroad).
  9. Tcl command(s) documented in top level README.md in flow order.
  10. Command line tool documentation in tool README.
  11. Conforms to Tcl command naming standards (no camel case).
  12. Does not read configuration files. Use command arguments or support commands.
  13. .clang-format at tool root directory to aid foreign programmers.
  14. No jenkins/, Jenkinsfile, Dockerfile in tool directory.
  15. regression script named test/regression with no arguments that runs tests. Not tests/regression-tcl.sh, not test/run_tests.py etc.
  16. regression script should run independent of current directory. For example, ../test/regression should work.
  17. regression should only print test results or summary, not belch 1000s of lines of output.
  18. Test scripts use OpenROAD tcl commands (not itcl, not internal accessors).
  19. regression script should only write files in a directory that is in the tool's .gitignore so the hierarchy does not have modified files in it as a result or running the regressions.
  20. Regressions report no memory errors with valgrind (stretch goal).
  21. Regressions report no memory leaks with valgrind (difficult).
  22. Ensure the top-level README and Tcl format are compliant.

Code Linting and Formatting

OpenROAD uses both clang-tidy and clang-format to perform automatic linting and formatting whenever a pull request is submitted. To run these locally, please first setup Clang Tooling using this guide. Thereafter, you may run these commands:

cmake . -B build  # generate build files
# typically only run these commands on files you changed.
clang-tidy -p ./build source_file.cpp
clang-format -i -style=file:.clang-format source_file.cpp

Doxygen

OpenROAD uses Doxygen style comments to generate documentation. See the generated documentation here. Our preferred syntax for Doxygen comments can be found in this file. Also, do refer to the official Doxygen documentation for more information on what you can include in your Doxygen comments here.

Below shows an example snippet taken from ./src/odb/include/odb/db.h:

///
/// dbProperty - Int property.
///
class dbIntProperty : public dbProperty
{
 public:
  /// Get the value of this property.
  int getValue();

  /// Set the value of this property.
  void setValue(int value);

  /// Create a int property. Returns nullptr if a property with the same name
  /// already exists.
  static dbIntProperty* create(dbObject* object, const char* name, int value);

  /// Find the named property of type int. Returns nullptr if the property does
  /// not exist.
  static dbIntProperty* find(dbObject* object, const char* name);
};

Guidelines

  1. Internally, the code should use int for all database units and int64_t for all area calculations. Refer to this link for a more detailed writeup on the reasons why this approach is preferred. The only place that the database distance units should appear in any program should be in the user interface, as microns are easier for humans than DBUs.