Bit-width optimization for configurable DSP's by multi-interval analysis

Optimizing hardware design for Human Action Recognition

“…The study of fixed-point implementation in image and video processing were primarily focused on the data range analysis and precision/errors associated with reduced bit-width [27], [28], [29]. It was also shown that machine learning model has certain tolerance on the reduced data precision and has lead to various FPGA implementation using fixed-point data [30], [31], [32].…”

Optimum wordlength allocation

“…an error-free Y, whereas a bounded error power of up to ½¼ ¾ has been allowed for Cr and Cb. We believe such optimum results, even for small circuits, to be highly valuable as a benchmark against which many new and existing heuristics [1,3,4,5,8,9, 10] may be compared. For this reason we are making several optimum wordlength benchmarks publicly available at http://infoeng.ee.ic.ac.uk/ gac1/optimumWL.…”

“…Bendetti and Perona [8] have proposed an analytic method for wordlength optimization based on interval arithmetic. The authors propose a 'multi-interval' approach, and demonstrate that the addition, subtraction, multiplication and division of the proposed intervals result in similar intervals, which may be propagated through a loop-free data-flow graph in order to estimate the wordlength required for a computation without losing any precision.…”

Perturbation analysis for word-length optimization

“…Benedetti and Perona [10] have proposed an analytic method for word-length optimization of general systems based on interval arithmetic. The authors concentrate on error-free word-length determination in loop-free dataflow graphs, and thus do not consider the problem of allowing a trade-off between accuracy and other figures of merit.…”

“…Some use an analytic approach to scaling and/or error estimation [1,2,7,9,10,11], some use simulation [8], and some use a hybrid of the two [12,13]. Analytic techniques often have the advantage of speed over simulation-based approaches, however they tend to be less general (applicable only to linear systems [1,2] or only to systems with loop-free data-flow graphs [7,9,10,11]). …”