MIOpen: An Open Source Library For Deep Learning Primitives

• DOI: 10.51130/graphicon-2020-2-2-2
• Publication Date: 2020-12-24T23:01:48Z
• Authors: Jehandad Khan; Paul Fultz; Artem Tamazov; Daniel Lowell; Chao Liu; Michael Melesse; Murali Nandhimand...; Kamil Nasyrov; Ilya Perminov; Tejash Shah; Vasilii Filippov; Jing Zhang; Jing Zhou; Bragadeesh Natarajan; Mayank Daga
• Keywords: Implementation

Deep Learning has established itself to be a common occurrence in the business lexicon. The unprecedented success of deep learning recent years can attributed to: an abundance data, availability gargantuan compute capabilities offered by GPUs, and adoption open-source philosophy researchers industry. neural networks decomposed into series different operators. MIOpen, AMD's primitives library for provides highly optimized implementations such operators, shielding from internal implementation details hence, accelerating time discovery. This paper introduces MIOpen about workings supported features. innovates on several fronts, as implementing fusion optimize memory bandwidth GPU launch overheads, providing auto-tuning infrastructure overcome large design space problem configurations, algorithms convolutions filter input sizes. is one first libraries publicly support bfloat16 data-type convolutions, allowing efficient training at lower precision without loss accuracy.