Research using Tensilica's Technology
We are primarily using Tensilica tools for a research project "Flares: Flexible Architectures for Embedded systems".
Instruction-set extensible processors, such as Xtensa, allow a processor core to be extended with application-specific custom instructions. However, choosing an appropriate set of custom instructions for an application is a difficult problem. We have previously developed fast and effective algorithms for automatic selection of custom instructions under various constraints (3, 4, 5). Recently, we used Xtensa toolchain to identify and select custom instructions in the context of real-time systems. This work will be presented at DATE 2007. In addition, we are using Tensilica tools for the final year projects of undergraduate students to design customized processors for security applications.
In the future, we will focus on previously unexplored issues in the context of extensible processors. In particular, we plan to address the following:
- Scheduling of custom instructions in compiler: existing techniques perform custom instruction selection and scheduling in two separate phases leading to inefficient design. We plan to integrate the selection and scheduling in two separate phases leading to inefficient design. We plan to integrate he selection and scheduling to obtain the optimal design point.
- Exploiting custom instruction spanning control flow boundaries: Previous research have focused exclusively on selecting custom instructions within basic blocks. However, our study (2) suggests that significant speedup is possible if we relax this constraint.
From a broader perspective, we would like to extend the custom instruction generation methodology to multiprocessor SoCs. Existing custom instruction design techniques are restricted to a single task running on a processing element. However, in the context of multiprocessor SoCs, the deign space to be explored is more complex. First, custom instruction generation should be integrated with task assignment and scheduling so as to select the optimal design point. Second, we have to consider both fine-grain reconfigurability (as in custom instructions) as well as coarse-grain reconfigurability (such as hardware implementation of entire FFT core). We have previously proposed separate approaches for these two levels of reconfigurability. Now, we plan to explore an integrated solution.
We have so far used Simplescalar simulator as our experimental framework. In the future, we are keen to use Tensilica tools so that we can have more realistic validations of our proposed techniques.
More information can be obtained from the project webpage:
http://www.comp.nus.edu.sg/~rpembed/flares/
Publications
- “Instruction-Set Customization for Real-Time Systems”
Huynh Phung Huynh and Tulika Mitra
Design Automation and Test in Europe (DATE), April 2007
http://www.comp.nus.edu.sg/~tulika/date07.pdf
- “Characterizing Embedded Applications for Instruction-Set Extensible Processors”
Pan Yu and Tulika Mitra
41st ACM/IEEE Design Automation Conference (DAC), June 2004
http://www.comp.nus.edu.sg/~tulika/dac04.pdf
- “Satisfying Real-Time Constraints with Custom Instructions”
Pan Yu and Tulika Mitra
ACM International Conference on Hardware/Software Codesign and System Synthesis (CODES+ISSS), September 2005
http://www.comp.nus.edu.sg/~tulika/codes05.pdf
- “Scalable Custom Instructions Custom Instructions Identification for Instruction-Set Extensible Processors”
Pan Yu and Tulika Mitra
ACM/IEEE International Conference on Compilers, Architecture, and Synthesis for Embedded Systems (CASES), September 2004
http://www.comp.nus.edu.sg/~tulika/cases04.pdf
- “Handling Constraints in Multi objective GA for Embedded System Design”
Biman Chakraborty, Ting Chen, Tulika Mitra, Abhik Roychoudhury
IEEE 19th International Conference on VLSI Design (VLSI), January 2006
http://www.comp.nus.edu.sg/~tulika/mitrat_GA.pdf
Outline of Curriculum and Courses
Student/Assistants Outreach
- 5 graduate students using Tensilica tools per year
- 45 graduate students may use Tensilica tools per year per coursework
- 5 undergraduate students using Tensilica tools per year for final year project
- Very interested in participating in college recruiting programs with Tensilica
National University of Singapore's link:
http://www.nus.edu.sg
National University of Singapore's department link:
http://www.comp.nus.edu.sg
National University of Singapore's FLARES (Flexible Architectures for Embedded Systems) project link:
http://www.comp.nus.edu.sg/~rpembed/flares/
|
