Section: New Results

Compilation and Synthesis for Reconfigurable Platform

Adaptive dynamic compilation for low power embedded systems

Participants : Steven Derrien, Simon Rokicki.

Just-in-time (JIT) compilers have been introduced in the 1960s and became popular in the mid-1990s with the Java virtual machine. The use of JIT techniques for bytecode languages brings both portability and performance, making it an attractive solution for embedded systems, as evidenced by the Dalvik framework used by Android.

When targeting embedded systems, JIT compilation is even more challenging. First, because embedded systems are often based on architectures with an explicit use of Instruction- Level Parallelism (ILP), such as Very Long Instruction Word (VLIW) processors. Those architectures are highly dependent of the quality of the compilation, mainly because of the instruction scheduling phase performed by the compiler. The other challenge lies in the high constraints of the embedded system: the energy and execution time overhead due to the JIT compilation must be carefully kept under control. This is even more true if the JIT system is to be used in the context of a heterogeneous multi-core system with support dynamic task migration for heterogeneous ISA cores and/or support dynamically reconfigurable machines.

To address these challenges, we are currently studying how it is possible to take advantage of custom hardware to speed-up (and reduce the energy cost of) the JIT compilation stage. In this framework, basic optimizations and JIT management are performed in software, while the compilation back-end is implemented by means of specialized hardware. This back-end involves both instruction scheduling and register allocation, which are known to be the most time consuming stages of such a compiler. The first results are very encouraging, and we are finalizing an FPGA-based demonstration of the system.

Design Tools for Reconfigurable Video Coding

Participants : Emmanuel Casseau, Yaset Oliva.

In the field of multimedia coding, standardization recommendations are always evolving. To reduce design time taking benefit of available SW and HW designs, Reconfigurable Video Coding (RVC) standard allows defining new codec algorithms. The application is represented by a network of interconnected components (so called actors) defined in a modular library and the behaviour of each actor is described in the specific RVC-CAL language. Dataflow programming, such as RVC applications, express explicit parallelism within an application. However general purpose processors cannot cope with both high performance and low power consumption requirements embedded systems have to face. We have investigated the mapping of RVC applications onto a dedicated multiprocessor platform. Actually, our goal is to propose an automated co-design flow based on the RVC framework. The design flow starts with the Dynamic Dataflow and CAL descriptions of an application and goes up to the deployment of the system onto the hardware platform. We also propose a framework to explore dynamic mapping algorithms for multiprocessors systems. Such an algorithm should be capable of computing a more efficient workload repartition based on the current configuration and performances of the system. The targeted platform is composed of several Processing Elements (PE). They follow a hierarchical organization: one PE plays the role of master and the others are slaves. The master assigns tasks (actors) to the slaves. The slaves execute the application tasks. The system has been implemented on a Zynq platform. The mapping is computed at runtime on the ARM processor while two clusters of 8 Microblazes each play the role of slaves. The DDR memory is split into two sections: one is reserved to the Master and the other one is shared with the slaves. This later contains the actor’s code. On the FPGA, the Microblazes are connected to private memories through the Local Memory Bus (LMB) that store the runtime copy. A common shared memory is used for the data exchanges between the processors. It contains the FIFOs for token exchanges between actors. The dynamic mapping algorithm aims at increasing data throughput. It starts by gathering the performance metrics of the system. It then identifies the processor with the highest workload. The algorithm evaluates the gain when moving the actor to one of the other processors. The migration is only valuable if the overhead of moving the actor is less that the gain. The actor that would lead to the highest gain is selected for migration. As a use case, we implement an MPEG-4 decoder algorithm onto a multi-core heterogeneous system deployed onto the Zynq platform from Xilinx [61] [69] . This work is done in collaboration with Lab-STICC Lorient.

High-Level Synthesis Based Rapid Prototyping of Software Radio Waveforms

Participants : Emmanuel Casseau, Mai Thanh Tran.

Software Defined Radio (SDR) is now becoming a ubiquitous concept to describe and implement Physical Layers (PHYs) of wireless systems. In this context, FPGA (Field Programmable Gate Array) technology is expected to play a key role. To this aim, leveraging the nascent High-Level Synthesis (HLS) tools, a design flow from high-level specifications to Register-Transfer Level (RTL) description can be thought to generate processing blocks that can be reconfigured at run-time. We thus propose a methodology for the implementation of run-time reconfiguration in the context of FPGA-based SDR. The design flow allows the exploration between dynamic partial reconfiguration and control signal based multi-mode design. This architectural tradeoff relies upon HLS and its associated design optimizations. We apply the methodology to the architectural exploration of a Fast Fourier Transform (FFT) for Long Term Evolution (LTE) standard as a use case.

Optimization of loop kernels using software and memory information

Participant : Angeliki Kritikakou.

Compilers optimize the compilation sub-problems one after the other, following an order which leads to less efficient solutions because the different sub-problems are independently optimized taking into account only a part of the information available in the algorithms and the architecture. In [19] , we have presented an approach which applies loop transformations in order to increase the performance of loop kernels. The proposed approach focuses on reducing the L1, L2 data cache and main memory accesses and the addressing instructions. Our approach exploits the software information, such as the array subscript equations, and the memory architecture, such as the memory sizes. Then, it applies source-to-source transformations taking as input the C code of the loop kernels and producing a new C code which is compiled by the target compiler. We have applied our approach to five well-known loop kernels for both embedded processors and general purpose processors. From the obtained experimental results we observed speedup gains from 2 up to 18. [21] presents a new methodology for computing the Dense Matrix Vector Multiplication, for both embedded (processors without SIMD unit) and general purpose processors (single and multi-core processors with SIMD unit). The proposed methodology fully exploits the combination of the software (e.g., data reuse) and hardware parameters (e.g., data cache associativity) which are considered simultaneously giving a smaller search space and high-quality solutions. The proposed methodology produces a different schedule for different values of the (i) number of the levels of data cache; (ii) data cache sizes; (iii) data cache associativities; (iv) data cache and main memory latencies; (v) data array layout of the matrix and (vi) number of cores. With our experimental results we show that the proposed approach achieves increased performance than ATLAS state-of-the-art library with a speedup from 1.2 up to 1.45.

Leveraging Power Spectral Density for Scalable System-Level Accuracy Evaluation

Participants : Benjamin Barrois, Olivier Sentieys.

The choice of fixed-point word-lengths critically impacts the system performance by impacting the quality of computation, its energy, speed and area. Making a good choice of fixed-point word-length generally requires solving an NP-hard problem by exploring a vast search space. Therefore, the entire fixed-point refinement process becomes critically dependent on evaluating the effects of accuracy degradation. In [34] , a novel technique for the system-level evaluation of fixed-point systems, which is more scalable and that renders better accuracy, was proposed. This technique makes use of the information hidden in the power-spectral density of quantization noises. It is shown to be very effective in systems consisting of more than one frequency sensitive components. Compared to state-of-the-art hierarchical methods that are agnostic to the quantization noise spectrum, we show that the proposed approach is $5\times$ to $500\times$ more accurate on some representative signal processing kernels.