13h15 -- Lionel Lacassagne (LIP6, ALSOC): High Level Transforms for SIMD and Low-Level Computer Vision Algorithms, how to put HPC in an embedded system

14h15 -- Holger Fröning (Universität Heidelberg, Allemagne): "DeepChip: A Software Architecture for Deep Neural Networks on Resource-Constrained Systems"

15h15 -- Fabienne Jezequel (LIP6, Pequan): "Numerical validation of simulations with Discrete Stochastic Arithmetic"

16h10-16h30 -- break

16h30 -- Manfred Mücke (MCL, Autriche): "Report from Hipeac – ML end-to- end frameworks: Bonseyes, Lift, Dividiti (& TVM/Neo)"

17h00 -- Open discussion: "Applying precision estimation tools to parametric models of physical systems"

== Résumés ==

Holger Fröning, "DeepChip: A Software Architecture for Deep Neural Networks on Resource-Constrained Systems"

Abstract: While Deep Learning is becoming ubiquitous in computing, being deployed in systems ranging from data centers to mobile devices, its computational complexity is huge. Contrary, while resource-constrained systems steadily improve their processing power, there is a huge gap to the demands of deep neural networks. Furthermore, CMOS technology projections show that performance scaling will be increasingly difficult, due to reasons including power consumption and eventually limited component scaling. In the DeepChip collaboration, we gear to design software architectures that improve the mapping of DNNs to resource-constrained systems by removing redundancies found in many DNNs. In this talk, we will shortly review basics and recent related work, before we introduce the main concept of DeepChip, which is based on a reduce-and-scale architecture. We will conclude with a couple of anticipated research directions.

Bio: Holger Fröning is since 2018 an interim professor at the Faculty of Mathematics and Computer Science at the Ruprecht-Karls University of Heidelberg (Germany), and leads the Computer Engineering Group at the Institute of Computer Engineering. His research interests focus on machine learning and high-performance computing, and include hardware and software architectures, co-design, data movement optimizations, and associated power and energy aspects. Previously, he was associate professor at the same university. In 2016, he was with NVIDIA Research (Santa Clara, CA, US) as visiting scientist, sponsored by Bill Dally. Early 2015 he was visiting professor at the Graz University of Technology (Austria), sponsored by Gernot Kubin. From 2008 to 2011 he reported to Jose Duato from the Technical University of Valencia (Spain). He has received his PhD and MSc degrees 2007 respectively 2001 from the University of Mannheim, Germany. He was awarded a Google Faculty Research Award in 2014. Four of his publications have received a best paper award, and parts of his research results have been commercialized. He chaired tracks for EuroPar 2015 and International Supercomputer Conference 2017, and recently served as program committee member for IPDPS2019/18, CCGRID2019/18, SC2017, ICPP2018/17/16, CLUSTER2018/16, and Euro-Par2019. His recent sponsors include BMBF, DFG, FWF, NVIDIA, SAP, Xilinx, and Micron. For more information, visit his website: http://www.ziti.uni-heidelberg.de/compeng

Fabienne Jézéquel, Numerical validation of simulations with Discrete Stochastic Arithmetic

Discrete Stochastic Arithmetic (DSA) is an automatic method for rounding error analysis based on a probabilistic approach. DSA allows to estimate the number of exact significant digits in computed results by executing the user programs several times in a synchronous way using a random rounding mode. We present the CADNA library (http://cadna.lip6.fr) an implementation of DSA that controls the numerical quality of sequential or parallel programs and detects numerical instabilities generated during their execution. A particular version of CADNA which enables numerical validation in hybrid CPU-GPU environments is described. We also present the SAM library (Stochastic Arithmetic in Multiprecision, http://www-pequan.lip6.fr/~jezequel/SAM) that estimates rounding errors in arbitrary precision programs. Finally we describe PROMISE (PRecision OptiMISE, http://promise.lip6.fr), a tool for precision auto-tuning. Most numerical simulations are performed in double precision (IEEE754 binary64), and this can be costly in terms of computing time, memory transfer and energy consumption. The PROMISE tool, based on CADNA, aims at reducing in numerical programs the number of double precision variable declarations in favor of single precision ones, taking into account a requested accuracy of the results.