High Level Synthesis: from Algorithm to Digital Circuit- P2 pdf

Contributors xi Rom´an Hermida Facultad de Inform´atica, Universidad Complutense de Madrid, c/Prof. Jos´eGarc´ıa Santesmases s/n, 28040 Madrid, Spain, rhermida@dacya.ucm.es Niraj K. Jha Department of Electrical and Engineering, Princeton University, Princeton, NJ 08544, USA, jha@princeton.edu Wei Jiang AutoESL Design Technolgoies, Inc., 12100 Wilshire Blvd, Los Angeles, CA 90025, USA, wjiang@autoesl.com Ryan Kastner Department of Electrical and Computer Engineering, University of California, Santa Barbara, CA 93106, USA, kastner@ucsd.edu Vinod Kathail Synfora, Inc., 2465 Latham Street, Suite # 300, Mountain View, CA 94040, USA, vinod.kathail@synfora.com Hyukmin Kwon Samsung Electronics Co., Suwon, Kyunggi Province, South Korea, hm25.kwon@samsung.com Eric Martin European University of Brittany – UBS, Lab-STICC, BP 92116, 56321 Lorient Cedex, France, eric.martin@univ-ubs.fr Jos´e Manuel Mend´ıas Facultad de Inform´atica, Universidad Complutense de Madrid, c/Prof. Jos´eGarc´ıa Santesmases s/n, 28040 Madrid, Spain, mendias@dacya.ucm.es Michael Meredith VP Technical Marketing, Forte Design Systems, San Jose, CA 95112, USA, mmeredith@ForteDS.com Mar´ıa Carmen Molina Facultad de Inform´atica, Universidad Complutense de Madrid, c/Prof. Jos´eGarc´ıa Santesmases s/n, 28040 Madrid, Spain, cmolinap@dacya.ucm.es Rishiyur S. Nikhil Bluespec, Inc., 14 Spring Street, Waltham, MA 02451, USA, nikhil@bluespec.com Fr´ed´eric P´etrot INPG-TIMA/SLS, 46 Avenue F´elix Viallet, 38031 Grenoble Cedex, France, Frederic.Petrot@imag.fr Patrice Quinton ENS de Cachan, antenne de Bretagne, Campus de Ker Lann, 35 170 Bruz Cedex, France, patrice.quinton@irisa.fr xii Contributors Sanjay Rajopadhye Department of Computer Science, Colorado State University, 601 S. Howes St. USC Bldg., Fort Collins, CO 80523-1873, USA, Sanjay.Rajopadhye@colostate.edu Tanguy Risset CITI – INSA Lyon, 20 avenue Albert Einstein, 69621, Villeurbanne, France, tanguy.risset@insa-lyon.fr Rafael Ruiz-Sautua Facultad de Inform´atica, Universidad Complutense de Madrid, c/Prof. Jos´eGarc´ıa Santesmases s/n, 28040 Madrid, Spain, rsautua@fdi.ucm.es Benjamin Carrion Schafer EDA R&D Center, Central Research Laboratories, NEC Corp., Kawasaki, Japan, schaferb@bq.jp.nec.com Eric Senn European University of Brittany – UBS, Lab-STICC, BP 92116, 56321 Lorient Cedex, France, eric.senn@univ-ubs.fr Li Shang Department of Electrical and Computer Engineering, Queen’s University, Kingston, ON, Canada K7L 3N6, li.shang@queensu.ca Pascal Urard STMicroelectronics, Crolles, France, pascal.urard@st.com Kazutoshi Wakabayashi EDA R&D Center, Central Research Laboratories, NEC Corp., Kawasaki, Japan, wakaba@bl.jp.nec.com Gang Wang Technology Innovation Architect, Intuit, Inc., 7535 Torrey Santa Fe Road, San Diego, CA 92129, USA, Gang Wang@intuit.com Changqi Yang AutoESL Design Technolgoies, Inc., 12100 Wilshire Blvd, Los Angeles, CA 90025, USA, charles@autoesl.com Joonhwan Yi Samsung Electronics Co., Suwon, Kyunggi Province, South Korea, joonhwan.yi@samsung.com, joonhwan.yi@gmail.com Zhiru Zhang AutoESL Design Technolgoies, Inc., 12100 Wilshire Blvd, Los Angeles, CA 90025, USA, zhiruz@autoesl.com List of Web sites Chapter 2 related to system level design, synthesis and verification. Our recent projects include the SPARK parallelizing synthesis framework, SATYA verification framework. Ear- lier work from the laboratory formed the technical basis for the SystemC initiative. http://mesl.ucsd.edu/ Chapter 3 Catapult Synthesis product information page The home page for Catapult Synthesis on www.mentor.com, with links to product datasheets, free software evaluation, technical publications, success stories, testimo- nials and related ESL product information. http://www.mentor.com/products/esl/high level synthesis/ Algorithmic C datatypes download page The Algorithmic C arbitrary-length bit-accurate integer and fixed-point data types allow designers to easily model bit-accurate behavior in their designs. The data types were designed to approach the speed of plain C integers. It is no longer necessary to compromise on bit-accuracy for the sake of speed or to explicitly code fixed-point behavior using integers in combination with shifts and bit masking. http://www.mentor.com/products/esl/high level synthesis/ac datatypes Chapter 4 Synfora, Inc. is the premier provider of PICO family of algorithmic synthesis tools to design complex application engines for SoCs and FPGAs. Synfora’s technology helps to reduce design costs, dramatically speed IP development and verification, xiii MicroelectronicEmbedded Systrems Laboratory at UCSD hosts a number of projects xiv List of Web sites and reduce time-to-market. For the latest information on Synfora and PICO prod- ucts, please visit http://www.synfora.com Chapter 5 More information on Cynthesizer from Forte Design Systems can be found at http://www.ForteDS.com Chapter 6 More information on AutoPilotTM from AutoESL Design Technologies can be found at http://www.autoesl.com and http://cadlab.cs.ucla.edu/soc/ Chapter 7 Home Page for CyberWorkBench from NEC http://www.cyberworkbench.com Chapter 8 More information on Bluespec can be found at http://www.bluespec.com Documentation, training materials, discussion forums, inquiries about Bluespec SystemVerilog. http://csg.csail.mit.edu/oshd/ Open source hardware designs done by MIT and Nokia in Bluespec SystemVer- ilog for H.264 decoder (baseline profile), OFDM transmitter and receiver, 802.11a transmitter, and more. Chapter 9 GAUT is an open source project at UEB-Lab-STICC. The software for this project is freely available for download. It is provided with a graphical user interface, a quick start guide, a user manual and several design examples. GAUT is currently supported on Linux and Windows. GAUT has already been downloaded more than 200 times by people from industry and academia in 36 different countries. For more information, please visit: http://web.univ-ubs.fr/gaut/ List of Web sites xv Chapter 10 More information can be found on UGH from at UPMC-LIP6/SoC and INPG- TIMA/SLS at http://www-asim.lip6.fr/recherche/disydent/ This web site contains introduction text, source code and tutorials (through CVS) of the opensource Dysident framework that includes the UGH HLS tool. Chapter 11 More information on Chapter 11 can be found at http://cas.ee.ic.ac.uk/ Chapter 12 More information on MMAlpha can be found at http://www.irisa.fr/cosi/ALPHA/ Chapter 13 More information Chapter 13 can be found on at http://www.cse.ucsd.edu/∼kastner/research/aco/ Chapter 14 More information on Chapter 14 can be found at http://atc.dacya.ucm.es/ Chapter 15 More information on Chapter 15 can be found at http://www.princeton.edu/∼jha Chapter 1 User Needs Pascal Urard, Joonhwan Yi, Hyukmin Kwon, and Alexandre Gouraud Abstract One can see successful adoption in industry of innovative technologies mainly in the cases where they provide acceptable solution to very concrete prob- lems that this industry is facing. High-level synthesis promises to be one of the solutions to cope with the significant increase in the demand for design productivity beyond the state-of-the-art methodsand flows. It also offers an unparalleled possibil- ity to explore the design space in an efficient way by dealing with higher abstraction levels and fast implementation ways to prove the feasibility of algorithms and enables optimisation of performances. Beyond the productivityimprovement, which is of course very pertinent in the design practice, the system and SoC companies are more and more concerned with their overall capability to design highly com- plex systems providing sophisticated functions and services. High-level synthesis may considerably contribute to maintain such a design capability in the context of continuously increasing chip manufacturing capacities and ever growing customer demand for function-rich products. In this chapter three leading industrial users present their expectations with regard to the high-level synthesis technology and the results of their experiments in practical application of currently available HLS tools and flows. The users also draw conclusions on the future directions in which they wish to see the high-level synthesis evolves like multi-clock domain support, block interface synthesis, joint optimisation of the datapath and control logic, integration of automated testing to the generated hardware or efficient taking into account of the target implementation technology for ASICs and FPGAs in the synthesis process. Pascal Urard STMicroelectronics Joonhwan Yi and Hyukmin Kwon Telecommunication R&D, Samsung Electronics Co., South Korea Alexandre Gouraud France Telecom R&D P. Coussy and A. Morawiec (eds.) High-Level Synthesis. c  Springer Science + Business Media B.V. 2008 1 2 P. Ur ar d et al . Keywords: High-level synthesis, Productivity, ESL, ASIC, SoC, FPGA, RTL, ANSI C, C++, SystemC, VHDL, Verilog, Design, Verification, IP, TLM, Design space exploration, Memory, Parallelism, Simulation, Prototyping 1.1 System Level Design Evolution and Needs for an IDM Point of View: STMicroelectronics 1 Pascal Urard, STMicroelectronics The complexity of digital integrated circuits has always increased from a technol- ogy node to another. The designers often had to adapt to the challenge of providing commercially acceptable solution with a reasonable effort. Many evolutions (and sometimes revolutions) occurred in the past: back-end automation or logical syn- thesis were part of those, enabling new area of innovation. Thanks to the increasing integration factor offered by technology nodes, the complexity in latest SoC has reached tens of millions of gates. Starting with 90 nm and bellow, RTL design flow (Fig. 1.1) now shows its limits. The gap between the productivity per designer and per year and the increasing complexity of the SoC, even taking into account some really conservative number of gates per technology node, lead to an explosion of the manpower for SoCs in the coming technology node (Fig. 1.2). There is a tremendous need for productivity improvement at design level. This creates an outstanding opportunity for new design techniques to be adopted: design- ers, facing this challenge, are hunger to progress and open to raise the level of abstraction of the golden reference model they trust. A new step is needed in productivity. Part of this step could be offered by ESLD: Electronics System Level Design. This includes HW/SW co-design and High-Level Synthesis (HLS). HW/SW co-design deployment has occurred few years ago, thanks to SystemC and TLM coding. HLS however is new and just starting to be deployed. Figure 1.3 shows the basis of STMicroelectronics C-level design methodology. A bit-accurate reference model is described at functional level in C/C++ using SystemC or equiv- alent datatypes. In the ideal case, this C-level description has to be extensively validated using a C-level testbench, in the functional environment, in order to become the golden model of the implementation flow. This is facilitated by the sim- ulation speed of this C model, usually faster than other kinds of description. Then, taking into account technology constraints, the HLS tool produces an RTL represen- tation, compatible with RTL-to-GDS2 flow. Verification between C-level model and RTL is done either thanks to sequential equivalence checking tools, or by extensive simulations. Started in 2001 with selected CAD-vendors, the research on new flows 1 (C) Pascal Urard, STMicroelectronics Nov. 2006. Extracted for P. Urard presentation at ICCAD, Nov. 2006, San Jos´e, California, USA. 1 User Needs 3 Gates P&R + Layout System System Analysis Analysis Algorithm GDS2 RTL code Design model Target Target Asic Logic Synthesis Technology files (Standard Cells + RAM cuts) Formal proof (equivalence checking) Fig. 1.1 RTL Level design flow ~300~150~75~60~43~40~40~80~40~10 200k200k200k125k91k56k40k9k6k4k 60M30M15M7.5M4M2.2M1.5M750k250K50K 1.2M600k300k150k80k45k30k15k5k1k 324565900.130.180.250.350.50.7 Men / Years per 50 mm2 Die #Gates per Designer per year #Gates / Die (50mm2) conservative numbers 2010200820062004200220001998100619941991  It is urgent to win some productivity Fig. 1.2 Design challenges for 65 nm and below Fig. 1.3 High level synthesis flow 4 P. Ur ar d et al . Design Productivity vs Manual RTL (base 1) 1X 5X 1/2X t % Behavioral IP Reuse, further improves design productivity 10X Fig. 1.4 Learning curve has allowed some deployment of HLS tools within STMicroelectronics starting in 2004, with early division adopters. We clearly see in 2007 an acceleration of the demand from designers. Those designers report to win a factor ×5to×10 in terms of productivity when using C-level design methodology depending on the way they reuse in design their IPs (Fig. 1.4). More promising: designers that moved to C-level design usually don’t want to come back to RTL level to create their IPs Side benefit of these C-level design automation, the IP reuse of signal processing IP is now becoming reality. The flow automation allows to have C-IPs quite indepen- dent of implementation constraints (technology, throughput, parameters), described at functional level, easy to modify to cope with new specification and easy to re- synthesize. Another benefit: the size of the manual description (C instead of RTL) is reduced by roughly a factor 10. This reduces the time to modification (ECO) as well as the number of functional bugs manually introduced in the final silicon. The link with Transactional Level Modelling (TLM) platform has to be enhanced. Prior to HLS flow, both TLM and RTL descriptions where done manually (Fig. 1.5). HLS tools would be able to produce the TLM view needed for platform vali- dation. However, the slowing-down of TLM standardization did not allow in 2006 neither H1-2007 to have a common agreement of what should be TLM 2.0 interface. This lack of standardization has penalized the convergence of TLM platform flow and C-level HW design flow. Designer community would benefit of such a common agreement between major players of the SystemC TLM community.More and more, we need CAD community to think in terms of flows in their global environment, and not in terms of tools alone. Another benefit of HLS tools automation is the micro-architecture exploration. Figure 1.6 basically describes a change of paradigm: clock frequency can be partially de-correlated from throughput constraints. This means that, focusing on the functional constraints (throughput/latency), designer can explore several solutions fulfilling the specifications, but using various clock frequencies. Thanks to released clock constraint, the low-speed design will not have the area penalty of the high-speed solution. Combining this exploration 1 User Needs 5 Spec description High level algorithmic description C/TLM model RTL model TLM TLM Reference Platform RTL Verification Platform HLS tool Compatible thanks to TLM 2.0 Fig. 1.5 Convergence of TLM and design flows Fig. 1.6 One benefit of automation: exploration to memory partitioning and management exploration leads to some very interesting solutions. As an example, Fig. 1.7 shows that combining double buffering of an LDPC encoder to a division by 2 of the clock speed, produces a ×0.63 lower power solution for a 27% area penalty. The time-to-solution is dramatically reduced thanks to automation. The designer can then take the most appropriated solution depend- ing on application constraints (area/power). Currently, power is estimated at RTL level, on automatically produced RTL, thanks to some specialized tools. Experience shows that power savings can be greatly improved at architectural level, compared to back-end design level. There is currently no real power-driven synthesis solution known to us. This is one of the major needs we have for the future. Power driven synthesis will have to be much more than purely based on signals activity monitoring in the SoC buses. It will need also to take into account leakage current, memory consumption and will have to be compliant with multi-power-modes solutions (voltage and frequency scaling). There are many parameters to take into account to determine a power optimized solution, the ideal tool would have to take care of all these parameters in order to . C -level design methodology depending on the way they reuse in design their IPs (Fig. 1.4). More promising: designers that moved to C -level design usually don’t want to come back to RTL level to. RTL level, on automatically produced RTL, thanks to some specialized tools. Experience shows that power savings can be greatly improved at architectural level, compared to back-end design level. There. masking. http://www.mentor.com/products/esl /high level synthesis/ac datatypes Chapter 4 Synfora, Inc. is the premier provider of PICO family of algorithmic synthesis tools to design complex application