Acceleration Opportunities in Linear Algebra Applications via Idiom Recognition

Author(s)

de Carvalho, Joao P. L. ; Kuzma, Braedy ; Araujo, Guido ; Assoc Comp Machinery

• Document type: Journal article
• Source: ICPE'20: COMPANION OF THE ACM/SPEC INTERNATIONAL CONFERENCE ON PERFORMANCE ENGINEERING; v. N/A, p. 2-pg., 2020-01-01.

Abstract

General matrix-matrix multiplication (GEMM) is a critical operation in many application domains [1]. It is a central building block of deep learning and computer graphics algorithms and is also a core operation for most scientific applications based on the discretization of systems of differential equations. Due to this, GEMM has been extensively studied and optimized, resulting in libraries of exceptional quality such as BLAS, Eigen, and other platform specific implementations such as MKL (Intel's x86) and ESSL (IBM's PowerPC) [3, 5]. Despite these successes, the GEMM idiom continues to be reimplemented by programmers without consideration for the intricacies already accounted for by the aforementioned libraries. To this end, this project aims to provide transparent adoption of high-performance implementations of GEMM through a novel optimization pass implemented within the LLVM framework using idiom recognition techniques. (AU)

• FAPESP's process: 13/08293-7 - CCES - Center for Computational Engineering and Sciences
• Grantee: Munir Salomao Skaf
• Support Opportunities: Research Grants - Research, Innovation and Dissemination Centers - RIDC
• FAPESP's process: 16/15337-9 - Distributed Transactional Memories and Efficient Data Distribution Models to Speed-up Irregular Data Structure Intensive Applications
• Grantee: João Paulo Labegalini de Carvalho
• Support Opportunities: Scholarships in Brazil - Doctorate
• FAPESP's process: 19/01110-0 - Novel code analysis to identify constructs suitable for hardware primitives
• Grantee: João Paulo Labegalini de Carvalho
• Support Opportunities: Scholarships abroad - Research Internship - Doctorate