A CUDA backend for Marrow and its Optimisation via Machine Learning

• Autor(a) principal: Valente, Pedro de Almeida Amaral Ramos
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Texto Completo: http://hdl.handle.net/10362/155512
• topic: code generation; Marrow; GPU; CUDA; machine learning; compilation; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

In the modern days, various industries like business and science deal with collecting, processing and storing massive amounts of data. Conventional CPUs, which are optimised for sequential performance, struggle to keep up with processing so much data, however GPUs, designed for parallel computations, are more than up for the task. Using GPUs for general processing has become more popular in recent years due to the need for fast parallel processing, but developing programs that execute on the GPU can be difficult and time consuming. Various high-level APIs that compile into GPU programs exist, however due to the abstraction of lower level concepts and lack of algorithm specific optimisations, it may not be possible to reach peak performance. Optimisation specifically is an interesting problem, optimisation patterns very rarely can be applied uniformly to different algorithms and manually tuning individual programs is extremely time consuming. Machine learning compilation is a concept that has gained some attention in recent years, with good reason. The idea is to have a model trained using a machine learning algorithm and have it make an estimate on how to optimise an input program. Predicting the best optimisations for a program is much faster than doing it manually, in works making use of this technique, it has shown to also provide even better optimisations. In this thesis, we will be working with the Marrow framework and develop a CUDA based backend for it, so that low-level GPU code may be generated. Additionally, we will be training a machine learning model and use it to automatically optimise the CUDA code generated from Marrow programs.