On Libraries, Reuse, and the Value of EDA Software Igor Markov Univ. of Michigan & Synplicity
Outline The challenge Extrapolating from past experiences What undermines the value of SW? What can we do ?
The Challenge Are EDA companies undervalued ? –Very sophisticated software –Highly educated employees –But stock does not grow ! Little is said about creating value in EDA software development process –Are we spinning our wheels ? –Are we undermining the value of EDA ? –Are there deficiencies in our eco-system?
Efficiency, Success, Adoption How fast should EDA tools be developed ? Should they be maintained or rewritten ? How to ensure that they do their job well ? How to ensure/evaluate adoption? How to improve value of EDA tools?
Personal Experiences Developing several academic tools adopted in companies –UCLApack UMpack –Capo, MLPart, infrastructure, etc –Very liberal license Interaction with adopters –We get 2-3 requests per week Measurements of popularity in academia (surprising conclusions)
UCLApack / UMpack Developed mostly at UCLA by Andrew Caldwell ( Simplex Cadence Tabula) and Igor Markov ( U.Michigan) –supervised by Andrew Kahng Initial release at DAC 2000 –120K lines in C++ Currently over 200K lines
What’s Available in UMpack? (1) Most popular: the Capo placer –Originally written in , maintained and extended at Michigan –Uses min-cut partitioning, works well for <100K std. cells –Routability-driven (beats most of the academic tools, some commercial tools) –Robust, well-tested, >100 tape-outs –All source code is available
What’s Available in UMpack? (2) UCLA DB (written in ) –An object-oriented database that maps most of LEF/DEF syntax to in-memory data structures –Includes two parsers (one written at UCLA, one released by Cadence) –Highly modular, reasonably efficient –Not entirely up-to-date, but all source is available
What’s Available in UMpack? (3) MLPart (written in ) –A multi-level min-cut partitioner –Used in Capo has been tested extremely well –Used by several companies: for prototyping logic synthesis tools, for verification (production code) –Results are usually a little worse than hMetis, but MLPart is available in source code
What’s Available in UMpack? (4) Parquet floorplanner (written in ) –Now a component of Capo –Helped Capo outperform Cadence by 70% at ISPD 2002 Extensive infrastructure in two dozen packages –Generic data structures, statistics –Built-in debugging tools –Geometry primitives, hierarchy mgmt, etc –Utilities, e.g., LEFDEF our formats
What’s Available in UMpack? (5) OpenAccess compatibility –Michigan + Cadence Labs UMpack/Capo is recommended for all OA Gear downloads –Used to visualize circuits MLPart is compatible with hMetis –C-API (Synplicity) + hMetis wrapper Works with g and above on Linux & Solaris Works with MSVC++ on Windows Synplicity contrib’d a 64-bit port
What’s Available in UMpack? (5) Simplified data formats –The Capo input format is now supported by 20+ academic placers –Intel, IBM and others have converters + LEF/DEF converter –A good number of examples given as regression tests Documentation –Web-based + included + “self-documented code”
Adoption of Our Tools (1) The license allows any use for free (the MIT X Window license) –No restrictions for academic use –No notification requirement Dozens of papers report modifying Capo Start-ups asked for a list of people who know Capo source code
Adoption of Our Tools (2) Synplicity used Capo in Amplify RC for LSI Logic Rapid-chip architecture –100s tape-outs over two years –Suddenly discontinued when LSI quit the fab business Several start-ups are still using Capo (are sending bug reports) MLPart is used in Certify
Observations Surprise: Capo adoption 10x greater than MLPart adoption –MLPart has only one competitor (hMetis, unavailable in source code, unavailable for commercial use) –There are about 10 academic placers claim better results than Capo on large netlists (but none are available in source code) UCLA DB adoption – non-existent Parquet adoption - huge
Explanations ? Source-code availability does wonders EDA industry & EDA research is tool-oriented –To force people think about infrastructure, we need the scale of OpenAccess –A good library can be overlooked b/c its value is not clearly seen Best combination: lightweight tool with a clear functionality
Personal Experiences Superficial familiarity with commercial EDA software –Talking to developers –Listening to invited talks –Occasionally looking at source code 8 EDA companies, names starting with –A, C, I, M, S
EDA Industry SW is Old Several companies limit g++ to very old versions –Perceived stability –At least 20% lost in tool runtime –Old versions may not support many language features Several companies ban C++ –Main argument: developers shoot themselves in the foot
Compare to UCLApack Written with heavy use of C++ Relies on the Standard Template Library (STL) for data structures –Abundant online documentation –Undergraduate students know it (vs. homegrown data structures in companies) –Very efficient –“Clean” and elegant interface UCLApack: practically no pointers
Compare to UCLApack Use of STL –More compact, conceptual code –Less documentation –Less unit testing However… –Using STL was a nightmare before ~2002 –Now g++ and MSVC++ are stable
Takeaways To improve productivity –Must use C++ with STL –Must develop reusable software libraries with clean interfaces (as is done by OpenAccess coalition) Obstacles? –Maturity level of SW developers
Personal Experiences (3) Coaching Michigan students participating in ICCAD CADathlon –Three wins for Michigan in 5-6 years –Two 2 nd places Participating in ISPD contests –Won the routing contest last year Where did the best coders go? (are they still interested in EDA ?)
Observations Of CADathlon prize-winners –One went to Microsoft, one to LM –Two quit EDA –One became an EDA faculty –Two are working for EDA companies Big questions –Do we need to attract best coders? –Is there much room improving SW?
ISPD P&R contests Dramatic year-to-year improvements in results In 2006 and 2007, the 1 st place team was last the year before ! In most cases, the winning entries were written from scratch (APlace, Kraftwerk2, MaizeRoute, FGR) Academic tools better than industry
Efficiency, Success, Adoption How fast should EDA tools be developed ? Should they be maintained or rewritten ? How to ensure that they do their job well ? Is EDA research at fault ? How to improve value of EDA tools?
Conclusions Existing EDA code-bases are old and inefficient –Rely on outdated SW development infrastructure –There is room for improvement in core tools + new tools are needed Need to ensure better code reuse New SW development methods more efficient Need to attract best coders and keep them
Riddle for you … The greatest threat to the EDA industry –Six letters _ _ _ _ _ _ Letters: T