Accelerating error-tolerant applications with approximate function reuse

Brandalero, Marcelo; Silveira, Leonardo Almeida da; Souza, Jeckson Dellagostin; Beck, Antonio Carlos Schneider

doi:10.1016/j.scico.2017.05.003

Cited by 6 publications

(9 citation statements)

References 7 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The corresponding result is returned if an exact match is found; otherwise, the actual function call is made, which returns the result and updates the LUT 21 . Figure 5B shows the second scenario, which is commonly referred to as approximate function memoization 15,20,30 . The only difference between these two approaches is in its matching rule: the former searches for an exact match, and the latter search for an inexact (approximate) match.…”

Section: Preliminaries and Discussionmentioning

confidence: 99%

“…Still, we categorized both techniques according to the terminologies used in describing the article. Many works of literature combine both these terminologies and use the term interchangeably 10‐12,15 . Initially, MZ is described as a software technique implemented through simple data structures, 23 and CR is a hardware technique that stores the result(s) of a previously executed instruction in a hardware structure called reuse buffer 7 .…”

Section: Prior Workmentioning

confidence: 99%

“…The approximate scheme's primary focus is to provide a better solution to the issues of function memoization described in the introduction section. Several researchers implemented approximate computation techniques to avoid such issues 15,25,27,33 . They either follow a hardware or software approach and work on either of the three granularity levels: instruction‐level, function‐level, or block‐level.…”

Section: Prior Workmentioning

confidence: 99%

“…However, this technique requires a method to quantify the magnitude of output fluctuation for consecutive function calls. In Reference 15, the author proposes a hashing technique combined with a quantization method. The primary idea is to reduce the LUT by collapsing similar input contexts.…”

Section: Prior Workmentioning

confidence: 99%

“…We categorize these works into three groups based on either a hardware or a software approach, as shown in Figure 1. Value prediction 2‐5 is a hardware approaches, but computation reuse 6‐16 and memoization 17‐30 can be implemented using hardware as well as software. A standard procedure to execute these techniques follows some simple steps described below: Analyze and identify the hot code (hc) of a program to concentrate your optimization effort.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Approximate function memoization

Arundhati

Jena

Pani

2022

Concurrency and Computation

View full text Add to dashboard Cite

Summary Function memoization is an optimization technique that reduces a function call overhead when the same input appears again. A table that stores the previous result is searched and used to skip the repeated computation. This way, it increases the performance of the function call. In this article, we propose a software approach of function memoization to improve computing efficiency by bypassing the execution of the function implemented using approximate computing techniques. Searching overhead is a primary concern in any memoization technique proposed so far. In traditional function memoization, the input arguments are first searched in the look‐up table (LUT) for an exact match, and the corresponding result is extracted for further use. But, in this article, a decision‐making rule is proposed to help us decide whether to search the LUT or go for the actual computation. This decision‐making model is implemented through Bloom filter and Cantor's pairing function. Because Bloom filter sometimes produces false‐positive results, we suggest a simple approximation technique that searches the LUT for an approximate match rather than an exact match. The proposed model also contains a bypass algorithm implemented through C++ code that identifies the trivial computations from the input argument of the candidate function. By this, we can avoid the actual calculation and generate the result directly. Here, trivial computation identifies one or more input arguments that are either 0 or prefix±1$$ \pm 1 $$. To analyze the effectiveness of our proposed technique, we conducted several experiments using the benchmarks from the AxBench suite. We found that our result outperforms some of the methods proposed so far in terms of energy consumption and quality of results, particularly in image processing applications.

show abstract

Section: Preliminaries and Discussionmentioning

confidence: 99%

Section: Prior Workmentioning

confidence: 99%

Section: Prior Workmentioning

confidence: 99%

Section: Prior Workmentioning

confidence: 99%