Skip to content
This repository has been archived by the owner on Jan 26, 2021. It is now read-only.

The-OpenROAD-Project/OpenDB

Repository files navigation

Introduction

OpenDB is a design database to support tools for physical chip design. It was originally developed by Athena Design Systems. Nefelus, Inc acquired the rights to the code and open sourced it in 2019 to support the DARPA OpenROAD project.

The structure of OpenDB is based on the Cadence Design Systems text file formats LEF (library) and DEF (design) formats version 5.6. OpenDB supports a binary file format to save and load the design much faster than using LEF and DEF.

OpenDB is written in C++ 98 with standard library style iterators. The classes are designed to be fast enough to base an application on without having to copy them into application specific structures.

Installation

git clone https://github.com/The-OpenROAD-Project/OpenDB.git
cd OpenDB
mkdir build
cd build
cmake ..
make

The original Athena code found in /zrouter is built using Makefiles. See zrouter/src/BUILD/compilation_package_dependencies.

Directory structure

include/opendb/db.h - public header for all database classes
src/db - private/internal database representations
src/lefin - LEF reader
src/lefout - LEF writer
src/defin - DEF reader
src/defout - DEF writer

Database API

‼️ We are still working on documenting the APIs. We have over 1,800 objects and functions that we are still documenting (for both TCL and Python). Contributions are very welcome in this effort. Find starting points below.

TCL

After building successfully, run OpenDB tcl shell using ./build/src/swig/tcl/opendbtcl. An example usage:

set db [dbDatabase_create]
set lef_parser [new_lefin $db true]
set tech [lefin_createTech $lef_parser ./OpenDB/tests/data/gscl45nm.lef]

You can find examples on using the API from TCL under tests/tcl/ directory.

The full set of the tcl commands exposed can be found under ./build/src/swig/tcl/opendb_wrapper.cpp. Search for SWIG_prefix.

Python

After building successfully, run a Python shell using python3. Load `opendbpy module using:

import importlib.util
spec = importlib.util.spec_from_file_location("opendbpy", "./build/src/swig/python/opendbpy.py")
odb = importlib.util.module_from_spec(spec)
spec.loader.exec_module(odb)

# use it as following
odb.[class_name]

You can find examples on using the API from Python under tests/python/ directory.

The full set of the Python classes exposed can be found under ./build/src/swig/python/opendbpy.py.

C++

All public database classes are defined in db.h. These class definitions provide all functions for examining and modifying the database objects. The database is an object itself so multiple database objects can exist simultaineously (no global state).

dbTypes.h defines types returned by database class member functions.

All database objects are in the odb namespace.

dbChip dbBlock dbTech dbLib

All database objects have a 32bit object identifier accessed with the dbObject::getOID base class member function that returns a uint. This identifier is preserved across save/restores of the database so it should be used to reference database object by data structures instead of pointers if the reference lifetime is across database save/restores. OIDs allow the database to have exactly the same layout across save/restores.

The database distance units are nanometers and use the type uint.

Tests

There are a set of regression tests in /tests.

./tests/regression-tcl.sh
./tests/regression-py.sh

Database Internals

The internal description included here is paraphrased from Lukas van Ginneken by James Cherry.

The database separates the implementation from the interface, and as a result, each class becomes two classes, a public one and a private one. For instance, dbInst has the public API functions, while class _dbInst has the private data fields.

The objects allocated in dynamically resizable tables the implementation of which is in dbTable.hpp. Each table consists of a number of pages, each containing 128 objects. The table contains the body of the struct, not a set of pointers. This eliminates most of the pointer overhead while iteration is accomplished by stepping through the table. Thus grouping these objects does not require a double linked list and saves 16 bytes per object (at the cost of some table overhead). Each object has an id, which is the index into the table. The lowest 7 bits are the index in the page, while the higher bits are the page number. Object id's are persistent when saving and reading the datamodel to disk, even as pointer addresses may change.

Everything in the data model can be stored on disk and restored from disk exactly the way it was. An extensive set of equality tests and diff functions make it possible to check for even the smallest deviation. The capability to save an exact copy of the state of the system makes it possible to create a checkpoint. This is a necessary capability for debugging complex systems.

The code follows the definition of LEF and DEF closely and reflects many of the idiosyncrasies of LEF and DEF. The code defines many types of objects to reflect LEF and DEF constructs although it sometimes uses different terminology, for instance, the object to represent a library cell is called dbMaster while the LEF keyword is MACRO.

The data model supports the EEQ and LEQ keywords, (electrically equivalent and logically equivalent Masters) which could be used for sizing. However, it does not support any logic function representation. In general, there is very limited support for synthesis specific information: no way to represent busses, no way to represent logic function, very limited understanding of signal flow, limited support of timing information, no support for high level synthesis or test insertion.

The db represents routing as in DEF, representing a trace from point to point with a given width. The layout for a net is stored in a class named dbWire and it requires a special dbWireDecoder (which works like an iterator) to unpack the data and another dbWireEncoder to pack it. The data model does not support a region query and objects that are in the same layer are scattered about the data model and are of different classes.

This means that whatever tool is using the layout information will have to build it's own data structures that are suitable to the layout operations of that tool. For instance, the router, the extractor, and the DRC engine would each have to build their unique data structures. This encourages batch mode operation (route the whole chip, extract the whole chip, run DRC on the whole chip).

LICENSE

BSD 3-Clause License. See LICENSE file