Eric TOULLEC |
IRISA
Campus de Beaulieu
35042 RENNES Cedex
FRANCE
phone: +33 2 99 84 25 61
fax: +33 2 99 84 25 28
e-mail: Eric.Toullec@irisa.fr
I'm a PhD student since October 2001 within the CAPS project. My supervisor is André SEZNEC.
The path for performance on superscalar processors should combine instruction level parallelism with high clock frequency. However, constructors are facing obstacles for implementing both features. For instance, a very high clock frequency is enabled through very deep pipeline on Intel Pentium 4 while Alpha EV8 will feature a very wide instruction level parallelism and SMT parallelism with a less aggressive clock.
The register file is one of the piece which limits clock frequency on a wide issue superscalar processor. This register file is central in a superscalar processor. Assuming N instructions or micro-operations per cycle, each consuming (up to) two operands and producing (up to) one result, the register file should support 2N reads and N writes per cycle. Deep pipelines and wide instruction issuing enlarges the depth of speculation in processors. Then, large number of physical registers are needed. Unfortunately the silicon area, the power consumption and the access time of the register file all increase with the number of write and read ports on the register file as well as with the number of physical registers.
Research propositions to improve access time, silicon area and power consumption on the register file includes the virtual-physical register file (limits the number of physical registers) , register caching (caching the critical register) and the use of two function units clusters (duplicates the register file, but halves the number of read ports).
We propose to investigate new solutions to reduce the number of ports in a register file for a wide-issue superscalar (and SMT) processor , to increase the possible instruction level parallelim exploited with a fixed number of ports in a register file and/or to decrease the number of physical registers needed.