Electrical Engineering
      and Computer Sciences

Electrical Engineering and Computer Sciences

COLLEGE OF ENGINEERING

UC Berkeley

Empirical Study of Opportunities for Bit-Level Specialization in Word-Based Programs

Eylon Caspi

EECS Department
University of California, Berkeley
Technical Report No. UCB/CSD-00-1126
January 2001

http://www.eecs.berkeley.edu/Pubs/TechRpts/2000/CSD-00-1126.pdf

A majority of programs in use today are written for word-based computing architectures, such as the microprocessor, using word-based programming languages. The word model, while convenient, typically provides quantized word widths that are a mismatch for many applications. Consequently, many bits of a word may go unused and contribute no useful information to the computation. Removing these bits from the computation, e.g. using specialized hardware data-paths, may provide the implementation with significant savings in run-time, area, and/or power. In this project, we analyze and quantify this bit-level waste using a model of bit constancy and binding-time. Applying the model to the UCLA MediaBench suite of C programs, we find that some 70% of bit-level read operations are to easily identified constant data, much of it in unused, high-order bits. These findings suggest that there is significant opportunity for bit-level specialization of these programs by relatively simple means such as narrower data-paths.


BibTeX citation:

@techreport{Caspi:CSD-00-1126,
    Author = {Caspi, Eylon},
    Title = {Empirical Study of Opportunities for Bit-Level Specialization in Word-Based Programs},
    Institution = {EECS Department, University of California, Berkeley},
    Year = {2001},
    Month = {Jan},
    URL = {http://www.eecs.berkeley.edu/Pubs/TechRpts/2001/6431.html},
    Number = {UCB/CSD-00-1126},
    Abstract = {A majority of programs in use today are written for word-based computing architectures, such as the microprocessor, using word-based programming languages. The word model, while convenient, typically provides quantized word widths that are a mismatch for many applications. Consequently, many bits of a word may go unused and contribute no useful information to the computation. Removing these bits from the computation, e.g. using specialized hardware data-paths, may provide the implementation with significant savings in run-time, area, and/or power. In this project, we analyze and quantify this bit-level waste using a model of bit constancy and binding-time. Applying the model to the UCLA MediaBench suite of C programs, we find that some 70% of bit-level read operations are to easily identified constant data, much of it in unused, high-order bits. These findings suggest that there is significant opportunity for bit-level specialization of these programs by relatively simple means such as narrower data-paths.}
}

EndNote citation:

%0 Report
%A Caspi, Eylon
%T Empirical Study of Opportunities for Bit-Level Specialization in Word-Based Programs
%I EECS Department, University of California, Berkeley
%D 2001
%@ UCB/CSD-00-1126
%U http://www.eecs.berkeley.edu/Pubs/TechRpts/2001/6431.html
%F Caspi:CSD-00-1126