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.
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.
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.
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()
).
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.
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
Instructions for building are available here.
The patch file "AddTool.patch" illustrates how to add a tool to OpenROAD. Use the following commands to add a sample tool:
patch -p < docs/misc/AddTool.patch
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
Tool commands should be documented in the top-level OpenROAD README.md
file. Detailed documentation should be the tool/README.md
file.
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 (dst)
- OpenRCX Parasitic Extraction (rcx)
- OpenSTA timing/power report (sta)
- Graphical User Interface (gui)
- Static IR analyser (psm)
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.
- OpenROAD submodules reference tool
openroad
branch head. No gitdevelop
,openroad_app
, oropenroad_build
branches. - Submodules used by more than one tool belong in
src/
, not duplicated in each tool repo. CMakeLists.txt
does not use add_compile_options include_directories link_directories link_libraries. Use target_ versions instead. See tips here.CMakeLists.txt
does not use glob. Use explicit lists of source files and headers instead.CMakeLists.txt
does not defineCFLAGS
CMAKE_CXX_FLAGS
CMAKE_CXX_FLAGS_DEBUG
CMAKE_CXX_FLAGS_RELEASE
. Let the top level and defaults control these.- No
main.cpp
or main procedure. - No compiler warnings for GCC or Clang with optimization enabled.
- Does not call
flute::readLUT
(called once byopenroad
). - Tcl command(s) documented in top level
README.md
in flow order. - Command line tool documentation in tool README.
- Conforms to Tcl command naming standards (no camel case).
- Does not read configuration files. Use command arguments or support commands.
.clang-format
at tool root directory to aid foreign programmers.- No
jenkins/
,Jenkinsfile
,Dockerfile
in tool directory. regression
script namedtest/regression
with no arguments that runs tests. Nottests/regression-tcl.sh
, nottest/run_tests.py
etc.regression
script should run independent of current directory. For example,../test/regression
should work.regression
should only print test results or summary, not belch 1000s of lines of output.- Test scripts use OpenROAD tcl commands (not
itcl
, not internal accessors). 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.- Regressions report no memory errors with
valgrind
(stretch goal). - Regressions report no memory leaks with
valgrind
(difficult).
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