Code Size and Accuracy-Aware Synthesis of Fixed-Point Programs for Matrix Multiplication

Matthieu Martel 1 Mohamed Amine Najahi 1 Guillaume Revy 1
1 DALI - Digits, Architectures et Logiciels Informatiques
LIRMM - Laboratoire d'Informatique de Robotique et de Microélectronique de Montpellier, UPVD - Université de Perpignan Via Domitia
Abstract : In digital signal processing, many primitives boil down to a matrix multiplication. In order to enable savings in time, energy consumption, and on-chip area, these primitives are often implemented in fixed-point arithmetic. Various conflicting goals undermine the process of writing fixed-point codes, such as numerical accuracy, run-time latency, and size of the codes. In this article, we introduce a new methodology to automate the synthesis of small and accurate codes for matrix multiplication in fixed-point arithmetic. Our approach relies on a heuristic to merge matrix rows or columns in order to reduce the synthesized code size, while guaranteeing a target accuracy. We suggest a merging strategy based on finding closest pairs of vectors, which makes it possible to address in a few seconds problems such as the synthesis of small and accurate codes for size-64 and more matrix multiplication. Finally, we illustrate its efficiency on a set of benchmarks, and we show that it allows to reduce the synthesized code size by more than 50% while maintaining good numerical properties.
Complete list of metadatas

Cited literature [21 references]  Display  Hide  Download
Contributor : Guillaume Revy <>
Submitted on : Wednesday, September 30, 2015 - 2:45:21 PM
Last modification on : Friday, July 20, 2018 - 11:36:03 AM
Long-term archiving on : Thursday, December 31, 2015 - 10:34:00 AM


Files produced by the author(s)




Matthieu Martel, Mohamed Amine Najahi, Guillaume Revy. Code Size and Accuracy-Aware Synthesis of Fixed-Point Programs for Matrix Multiplication. PECCS: Pervasive and Embedded Computing and Communication Systems, Jan 2014, Lisbonne, Portugal. ⟨10.5220/0004884802040214⟩. ⟨lirmm-00860383⟩



Record views


Files downloads