Chisel

Abstract

The accuracy of an approximate computation is the distance between the result that the computation produces and the corresponding fully accurate result. The reliability of the computation is the probability that it will produce an acceptably accurate result. Emerging approximate hardware platforms provide approximate operations that, in return for reduced energy consumption and/or increased performance, exhibit reduced reliability and/or accuracy.We present Chisel, a system for reliability-and accuracyaware optimization of approximate computational kernels that run on approximate hardware platforms. Given a combined reliability and/or accuracy specification, Chisel automatically selects approximate kernel operations to synthesize an approximate computation that minimizes energy consumption while satisfying its reliability and accuracy specification.We evaluate Chisel on five applications from the image processing, scientific computing, and financial analysis domains. The experimental results show that our implemented optimization algorithm enables Chisel to optimize our set of benchmark kernels to obtain energy savings from 8.7% to 19.8% compared to the original (exact) kernel implementations while preserving important reliability guarantees.

ApproxHPVM: a portable compiler IR for accuracy-aware optimizations

“…Rely [Carbin et al 2013] and Chisel [Misailovic et al 2014] introduced the idea of quantifiable reliability and absolute error at the program level. They define function-level specifications that express the maximum probability with which the function can produce an inaccurate result.…”

A Framework for Automated and Controlled Floating-Point Accuracy Reduction in Graphics Applications on GPUs

“…Input: Approximation levels, a (sorted with highest precision first), All registers in kernel, reдisters, Quality threshold, qt Globals: A mapping between registers and approximation levels, precision_o f _reдisters Output: precision_o f _reдisters 1 Algorithm setPrecisions(a, registers, qt) Set unlocked_reдisters in precision_o f _reдisters to current_precision 13 Execute kernel with precision_o f _reдisters 14 if output quality < qt then 15 if size of unlocked_registers is 1 then 16 Set unlocked_reдisters in precision_o f _reдisters to previous_precision In each iteration, all unlocked registers, that is, registers not locked to an approximation level, are considered (lines [11][12][13][14][15][16][17][18][19][20][21][22][23][24]. The unlocked registers are adjusted to the current level of approximation, the kernel is executed, and the output quality is tested against the quality threshold (lines 12-14).…”

“…Misailovic et al [13] proposes a reliability-and accuracy-aware framework for optimizing approximate computational kernels, including interval analysis of source-code variables to bound the variables to minimum-required precisions. In contrast, our approach optimizes on actual data on assembly level, which makes it less likely to overestimate the required number of bits.…”

“…There are three limitations: First, the level of approximation is constrained to IEEE754-compliant data types, and, second, the optimizations target source-level floating-point variables instead of instruction-level floating-point values. Finally, the algorithm has limited scalability (O(an 2 ), where n is the number of analyzed floating-point values and a is the number of precision levels).Chisel [13] also supports automatic precision-tuning, but, like Precimonious, it does not target instruction-level registers. In addition, Chisel employs a static interval analysis, giving an upper bound on the required precision that yields conservative results.In this article, we first propose a method for automatically, and transparently to the programmer, classifying the amount of approximation of floating-point registers possible for individual low-level instructions.…”

“…Chisel [13] also supports automatic precision-tuning, but, like Precimonious, it does not target instruction-level registers. In addition, Chisel employs a static interval analysis, giving an upper bound on the required precision that yields conservative results.…”