Behavioral synthesis is the process of compiling an Electronic System Level (ESL) design to a register-transfer level (RTL) implementation. ESL specifications define the design functionality at a high level of abstraction (e.g., with C/C++ or SystemC), and thus provide a promising approach to address the exacting demands to develop feature-rich, optimized, and complex hardware systems within aggressive time-to-market schedules. Behavioral synthesis entails application of complex and error-prone transformations during the compilation process. Therefore, the adoption of behavioral synthesis highly depends on our ability to ensure that the synthesized RTL conforms to the ESL description. This dissertation provides an end-to-end scalable equivalence checking support for behavioral synthesis. The major challenge of this research is to bridge the huge semantic gap between the ESL and RTL descriptions, which makes the direct comparison of designs in ESL and RTL difficult. Moreover, a large number and a wide variety of aggressive transformations from front-end to back-end require an end-to-end scalable checking framework.
This dissertation provides an end-to-end scalable equivalence checking support for behavioral synthesis. The major challenge of this research is to bridge the huge semantic gap between the ESL and RTL descriptions, which makes the direct comparison of designs in ESL and RTL difficult. Moreover, a large number and a wide variety of aggressive transformations from front-end to back-end require an end-to-end scalable checking framework.
A behavioral synthesis flow can be divided into three major phases, including 1) front-end : compiler transformations, 2) scheduling: assigning each operation a clock cycle and satisfying the user-specified constraints, and 3) back-end : local optimizations and RTL generation. In our end-to-end and incremental equivalence checking framework, we check each of the three phases one by one. Firstly, we check the front-end that consists of a sequence of compiler transformations by decomposing it into a series of checks, one for each transformation applied. We symbolically explore paths in the input and output programs of each transformation, and check whether the input and output programs have the same observable behavior under the same path condition. Secondly, we validate the scheduling transformation by checking the preservation of control and data dependencies, and the preservation of I/O timing in the user-specified scheduling mode. Thirdly, we symbolically simulate the scheduled design and the generated RTL cycle by cycle, and check the equivalence of each mapped variables. We also develop several key optimizations to make our back-end checker scale to real industrial-strength designs. In addition to the equivalence checking framework, we also present an approach to detecting deadlocks introduced by parallelization of RTL blocks that are connected by synthesized interfaces with handshaking protocols.
To demonstrate the efficiency and scalability of our framework, we evaluated it on transformations applied by a behavioral synthesis tool to designs from the C-based CHStone and SystemC-based S2CBench benchmarks. Based on the evaluation results, our front-end checker can efficiently validate more than 75 percent of the total of 1008 compiler transformations applied to designs from the CHStone benchmark, taking an average time of 1.5 seconds per transformation. Our scheduling checker can validate control-data dependencies and I/O timing of all designs from S2CBench benchmark. Our back-end checker can handle designs with more than 32K lines of synthesized RTL from the CHStone benchmark, which demonstrates the scalability of the checker. Furthermore, our checker found several bugs in a commercial tool, underlining both the importance of formal equivalence checking and the effectiveness of our approach.
Identifer | oai:union.ndltd.org:pdx.edu/oai:pdxscholar.library.pdx.edu:open_access_etds-3467 |
Date | 05 August 2015 |
Creators | Yang, Zhenkun |
Publisher | PDXScholar |
Source Sets | Portland State University |
Detected Language | English |
Type | text |
Format | application/pdf |
Source | Dissertations and Theses |
Page generated in 0.0015 seconds