Skip to Main content Skip to Navigation
Habilitation à diriger des recherches

Contributions to Program Optimization and High-Level Synthesis

Christophe Alias 1, 2
1 CASH - CASH - Compilation and Analysis, Software and Hardware
Inria Grenoble - Rhône-Alpes, LIP - Laboratoire de l'Informatique du Parallélisme
Abstract : Since the end of Dennard scaling, power efficiency is the limiting factor for large-scale computing. Hardware accelerators such as reconfigurable circuits (FPGA, CGRA) or Graphics Processing Units (GPUs) were introduced to improve the performance under a limited energy budget, resulting into complex heterogeneous platforms. This document presents a synthetic description of my research activities over the last decade on compilers for high-performance computing and high-level synthesis of circuits (HLS) for FPGA accelerators. Specifically, my contributions covers both theoretical and practical aspects of automatic parallelization and HLS in a general theoretical framework called the polyhedral model. A first chapter describes our contributions to loop tiling, a key program transformation for automatic parallelization which splits the computation atomic blocks called tiles.We rephrase loop tiling in the polyhedral model to enable any polyhedral tile shape whose size depends on a single parameter (monoparametric tiling), and we present a tiling transformation for programs with reductions – accumulations w.r.t. an associative/commutative operator. Our results open the way for semantic program transformations ; program transformations which does not preserve the computation but still lead to an equivalent program. A second chapter describes our contributions to algorithm recognition. A compiler optimization will never replace a good algorithm, hence the idea to recognize algorithm instances in a program and to substitute them by a call to a performance library. In our PhD thesis, we have addressed the recognition of templates – functionswith first-order variables – into programs and its application to program optimization. We propose a complementary algorithm recognition framework which leverages our monoparametric tiling and our reduction tiling transformations. This automates semantic tiling, a new semantic program transformation which increases the grain of operators (scalar → matrix). A third chapter presents our contributions to the synthesis of communications with an off-chip memory in the context of high-level circuit synthesis (HLS). We propose an execution model based on loop tiling, a pipelined architecture and a source-level compilation algorithm which, connected to the C2H HLS tool from Altera, ends up to a FPGA configuration with minimized data transfers. Our compilation algorithm is optimal – the data are loaded as late as possible and stored as soon as possible with a maximal reuse. A fourth chapter presents our contributions to design a unified polyhedral compilation model for high-level circuit synthesis.We present the Data-aware Process Networks (DPN), a dataflow intermediate representation which leverages the ideas developed in chapter 3 to explicit the data transfers with an off-chip memory. We propose an algorithm to compile a DPN from a sequential program, and we present our contribution to the synthesis of DPN to a circuit. In particular, we present our algorithms to compile the control, the channels and the synchronizations of a DPN. These results are used in the production compiler of the Xtremlogic start-up.
Complete list of metadata

Cited literature [180 references]  Display  Hide  Download
Contributor : Christophe Alias <>
Submitted on : Monday, June 10, 2019 - 3:37:43 PM
Last modification on : Monday, October 19, 2020 - 11:09:14 AM


Files produced by the author(s)


  • HAL Id : tel-02151877, version 1



Christophe Alias. Contributions to Program Optimization and High-Level Synthesis. Hardware Architecture [cs.AR]. ENS de Lyon, 2019. ⟨tel-02151877⟩



Record views


Files downloads