Talk Abstract:

This talk will show how state-of-the-art proposals to compute convolutions on architectures with CPUs supporting SIMD instructions deliver poor performance for long SIMD lengths due to frequent cache conflict misses. The talk will propose new algorithmic approaches to mitigate the limitation of state-of-the-art proposals via the adaptation of the amount of computation exposed to the microarchitecture to mitigate cache misses, and the redefinition of the activation memory layout to improve the memory access pattern. These algorithmic approaches will motivate the Matrix Tile Extension (MTE), a novel matrix Instruction-Set Architecture (ISA) that completely decouples the instruction set architecture from the microarchitecture and seamlessly interacts with existing vector ISAs. MTE incurs minimal implementation overhead since it only requires a few additional instructions and a 64-bit Control Status Register (CSR) to keep its state, and beats the best state-of-the-art matrix ISA by 1.20x.

 

Bio:

Marc Casas is a technical research lead at the Barcelona Supercomputing Center (BSC) and lecturer at the Universitat Politècnica de Catalunya (UPC). His research lays between computer architecture (e.g. memory address translation, vector architectures) and high-performance computing (e.g. sparse linear algebra, parallel deep learning). He is the technical lead of the SONAR (parallel SOftware and New ARchitectures) research group, composed of PhD students, engineers, and postdocs. Marc has lead BSC contributions to several european projects (Mont-Blanc2020, European Processor Initiative, etc.), and research collaborations with Intel and IBM. 

Marc has been at BSC since 2013. He was a postdoctoral research scholar at the Lawrence Livermore National Laboratory (LLNL) from 2010 to 2013. He received the Marie Curie and Ramón y Cajal Fellowships on 2014 and 2018, respectively. He obtained a 5-years degree in mathematics in 2004, and a PhD degree in Computer Science in 2010 from the Universitat Politècnica de Catalunya (UPC).