Section: Overall Objectives

General Presentation

Classically, an embedded computer is a digital system that is part of a larger system and that is not directly accessible to the user. Examples are appliances like phones, TV sets, washing machines, game platforms, or even larger systems like radars and sonars. In particular, this computer is not programmable in the usual way. Its program, if it exists, is supplied as part of the manufacturing process and is seldom (or ever) modified thereafter. As the embedded systems market grows and evolves, this view of embedded systems is becoming obsolete and tends to be too restrictive. Many aspects of general-purpose computers apply to modern embedded platforms. Nevertheless, embedded systems remain characterized by a set of specialized application domains, rigid constraints (cost, power, efficiency, heterogeneity), and its market structure. The term embedded system has been used for naming a wide variety of objects. More precisely, there are two categories of so-called embedded systems: a) control-oriented and hard real-time embedded systems (automotive, plant control, airplanes, etc.); b) compute-intensive embedded systems (signal processing, multi-media, stream processing) processing large data sets with parallel and/or pipelined execution. Compsys is primarily concerned with this second type of embedded systems, now referred to as embedded computing systems.

Today, the industry sells many more embedded processors than general-purpose processors; the field of embedded systems is one of the few segments of the computer market where the European industry still has a substantial share, hence the importance of embedded system research in the European research initiatives. Our priority towards embedded software is motivated by the following observations: a) the embedded system market is expanding, among many factors, one can quote pervasive digitalization, low-cost products, appliances, etc.; b) research on software for embedded systems is poorly developed in France, especially if one considers the importance of actors like Alcatel, STMicroelectronics, Matra, Thales, etc.; c) since embedded systems increase in complexity, new problems are emerging: computer-aided design, shorter time-to-market, better reliability, modular design, and component reuse.

A specific aspect of embedded computing systems is the use of various kinds of processors, with many particularities (instruction sets, registers, data and instruction caches, now multiple cores) and constraints (code size, performance, storage). The development of compilers is crucial for this industry, as selling a platform without its programming environment and compiler would not be acceptable. To cope with such a range of different processors, the development of robust, generic (retargetable), though efficient compilers is mandatory. Unlike standard compilers for general-purpose processors, compilers for embedded processors and hardware accelerators can be more aggressive (i.e., take more time to optimize) for optimizing some important parts of applications. This opens a new range of optimizations. Another interesting aspect is the introduction of platform-independent intermediate languages, such as Java bytecode, that is compiled dynamically at runtime (aka just-in-time). Extreme lightweight compilation mechanisms that run faster and consume less memory have to be developed. The introduction of intermediate languages such as OpenCL is also a sign of the need for portability (as well as productivity) across diverse (if not heterogeneous) platforms. One of the initial objectives of Compsys was thus to revisit existing compilation techniques in the context of such embedded computing systems, to deconstruct some of these techniques, to improve them, and to develop new techniques taking constraints of embedded processors and platforms into account.

As for high-level synthesis (HLS), several compilers/systems have appeared, after some first unsuccessful industrial attempts in the past. These tools are mostly based on C or C++ as for example SystemC, VCC, CatapultC, Altera C2H, Pico-Express, Vivado HLS. Academic projects also exist (or existed) such as Flex and Raw at MIT, Piperench at Carnegie-Mellon University, Compaan at the University of Leiden, Ugh/Disydent at LIP6 (Paris), Gaut at Lester (Bretagne), MMAlpha (Insa-Lyon), and others. In general, the support for parallelism in HLS tools is minimal, especially in industrial tools. Also, the basic problem that these projects have to face is that the definition of performance is more complex than in classical systems. In fact, it is a multi-criteria optimization problem and one has to take into account the execution time, the size of the program, the size of the data structures, the power consumption, the manufacturing cost, etc. The impact of the compiler on these costs is difficult to assess and control. Success will be the consequence of a detailed knowledge of all steps of the design process, from a high-level specification to the chip layout. A strong cooperation of the compilation and chip design communities is needed. The main expertise in Compsys for this aspect is in the parallelization and optimization of regular computations. Hence, we target applications with a large potential parallelism, but we attempt to integrate our solutions into the big picture of CAD environments.

More generally, the aims of Compsys are to develop new compilation and optimization techniques for the field of embedded computing system design. This field is large, and Compsys does not intend to cover it in its entirety. As previously mentioned, we are mostly interested in the automatic design of accelerators, for example designing a VLSI or FPGA circuit for a digital filter, and in the development of new back-end compilation strategies for embedded processors. We study code transformations that optimize features such as execution time, power consumption, code and die size, memory constraints, and compiler reliability. These features are related to embedded systems but some are not specific to them. The code transformations we develop are both at source level and at assembly level. A specificity of Compsys is to mix a solid theoretical basis for all code optimizations we introduce with algorithmic/software developments. Within Inria, our project is related to the “architecture and compilation” theme, more precisely code optimization, as some of the research conducted in Parkas (previously known as Alchemy), Alf (previously known as Caps), Camus, and to high-level architectural synthesis, as some of the research in Cairn.

Most french researchers working on high-performance computing (automatic parallelization, languages, operating systems, networks) moved to grid computing at the end of the 90s. We thought that applications, industrial needs, and research problems were more interesting in the design of embedded platforms. Furthermore, we were convinced that our expertise on high-level code transformations could be more useful in this field. This is the reason why Tanguy Risset came to Lyon in 2002 to create the Compsys team with Anne Mignotte and Alain Darte, before Paul Feautrier, Antoine Fraboulet, Fabrice Rastello, and finally Christophe Alias joined the group. Then, Tanguy Risset left Compsys to become a professor at INSA Lyon, and Antoine Fraboulet and Anne Mignotte moved to other fields of research. As for Laure Gonnord, after a post-doc in Compsys, she obtained an assistant professor position in Lille but remained external collaborator of the team for the period 2009-2013 and finally obtained an assistant professor position in Lyon, and integrated officially the team. About the same time, Fabrice Rastello left while Tomofumi Yuki was hired as Inria researcher in 2014.

All present and past members of Compsys have a background in automatic parallelization and high-level program analyses and transformations. Paul Feautrier was the initiator of the polyhedral model for program transformations around 1990 and, before coming to Lyon, started to be more interested in programming models and optimizations for embedded applications, in particular through collaborations with Philips. Alain Darte worked on mathematical tools and algorithmic issues for parallelism extraction in programs. He became interested in the automatic generation of hardware accelerators, thanks to his stay at HP Labs in the Pico project in 2001. Antoine Fraboulet did a PhD with Anne Mignotte – who was working on high-level synthesis (HLS) – on code and memory optimizations for embedded applications. Fabrice Rastello did a PhD on tiling transformations for parallel machines, then was hired by STMicroelectronics where he worked on assembly code optimizations for embedded processors. Tanguy Risset worked for a long time on the synthesis of systolic arrays, being the main architect of the HLS tool MMAlpha. Christophe Alias did a PhD on algorithm recognition for program optimizations and parallelization. He first spent a year in Compsys working on array contraction, where he started to develop the tool Bee, then a year at Ohio State University with Prof. P. Sadayappan on memory optimizations. Laure Gonnord did a PhD on invariant generation and program analysis and became interested on application on compilation and code generation since her postdoc in the team. Finally, Tomofumi Yuki did a PhD on polyhedral programming environments and optimizations (in Colorado State University, with Prof. S. Rajopadhye) before a post-doc on polyhedral HLS in the Cairn team (Rennes).

To understand why we think automation in our field is highly important, it may be worth to quote Bob Rau and his colleagues (IEEE Computer, Sep. 2002):

"Engineering disciplines tend to go through fairly predictable phases: ad hoc, formal and rigorous, and automation. When the discipline is in its infancy and designers do not yet fully understand its potential problems and solutions, a rich diversity of poorly understood design techniques tends to flourish. As understanding grows, designers sacrifice the flexibility of wild and woolly design for more stylized and restrictive methodologies that have underpinnings in formalism and rigorous theory. Once the formalism and theory mature, the designers can automate the design process. This life cycle has played itself out in disciplines as diverse as PC board and chip layout and routing, machine language parsing, and logic synthesis.

We believe that the computer architecture discipline is ready to enter the automation phase. Although the gratification of inventing brave new architectures will always tempt us, for the most part the focus will shift to the automatic and speedy design of highly customized computer systems using well-understood architecture and compiler technologies.”

We share this view of the future of architecture and compilation. Without targeting too ambitious objectives, we were convinced of two complementary facts: a) the mathematical tools developed in the past for manipulating programs in automatic parallelization were lacking in high-level synthesis and embedded computing optimizations and, even more, they started to be rediscovered frequently in less mature forms, b) before being able to really use these techniques in HLS and embedded program optimizations, we needed to learn a lot from the application side, from the electrical engineering side, and from the embedded architecture side. Our primary goal was thus twofold: to increase our knowledge of embedded computing systems and to adapt/extend code optimization techniques, primarily designed for high performance computing, to the special case of embedded computing systems. In the initial Compsys proposal, we proposed four research directions, centered on compilation methods for embedded applications, both for software and accelerators design:

• Code optimization for specific processors (mainly DSP and VLIW processors);

• Platform-independent loop transformations (including memory optimization);

• Silicon compilation and hardware/software codesign;

• Development of polyhedral (but not only) optimization tools.

These research activities were primarily supported by a marked investment in polyhedra manipulation tools and, more generally, solid mathematical and algorithmic studies, with the aim of constructing operational software tools, not just theoretical results. Hence the fourth research theme was centered on the development of these tools.