Machine learning is rapidly being integrated into all areas of society, however, that puts a lot of pressure on resource costraint hardware such as embedded systems. The company Ericsson is gradually integrating machine learning based on neural networks, so-called deep learning, into their radio products. One promising product is their vectorized Digital Signal Processor (DSP) that are based upon the machine learning suitable Single Instruction, Multiple Data (SIMD) paradigm and Very Long Instruction Word (VLIW) architecture. However, despite the suitability of the SIMD paradigm, the embedded system needs to efficiently execute a computation-intensive deep learning algorithm with proper use of its limited resources. Therefore commonly used methods of implementing each layer of the computation-intensive Convolutional Neural Network (CNN), a type of Deep Neural Network (DNN), have been used and evaluated its implementation on the hardware and to assess the vectorized DSP’s deep learning suitability and capabilities. Despite the suitability of the hardware, the implementation utilized less than half of the available resources at all times during the execution. The main limitations were identified to be the limited 16-bit element instructions. To enhance the performance and improve the utilization of the available resources, easy-to-implement hardware instructions have been suggested. This work has made the first steps of implementing an efficiently performing CNN implementation on the examined vectorized DSP. / Integreringen av maskininlärning in i alla samhällsområden sker idag i rusande fart, men det sätter stor press på begränsad hårdvara som inbyggda system. Företaget Ericsson integrerar successivt maskininlärning baserad på neurala nätverk, så kallad djupinlärning, i sina radioprodukter. En lovande produkt är deras vektoriserade DSP som är baserade på maskininlärningspasset SIMD-paradigm och VLIW-arkitektur. Men trots lämpligheten av SIMD-paradigmet, är den största utmaningen att utnyttja de begränsade resurserna i inbyggda systemet för att effektivt exekvera en beräkningsintensiv djupinlärningsalgoritm. Därför har vanligt använda metoder för att implementera varje lager av den beräkningsintensiva CNN, en typ av DNN, använts och utvärderats på hårdvaran för att bedöma den vektoriserade DSP:s djupinlärningslämplighet samt förmågor. Trots hårdvarans lämplighet använde alla implementeringar mindre än hälften av de tillgängliga resurserna vid alla tidpunkter under exekveringen. De huvudsakliga begränsningarna identifierades vara den begränsade tillgången på 16-bitars element instruktioner. För att förbättra prestandan för ett närmare fullt utnyttjande av tillgängliga resurser har hårdvaruinstruktioner som är enkla att implementera föreslagits. Detta arbete har tagit de första stegen för att implementera ett effektivt förformande CNN på den undersökta vekotriserade DSP.

Lattice-Boltzmann method(LBM), a promising new particle-based simulation technique for complex and multiscale fluid flows, has seen tremendous adoption in recent years in computational fluid dynamics. Even with a state-of-the-art LBM solver such as Palabos, a user still has to manually write his program using the library-supplied primitives. We propose an automated code generator for a class of LBM computations with the objective to achieve high performance on modern architectures. Tiling is a very important loop transformation used to improve the performance of stencil computations by exploiting locality and parallelism. In the first part of the work, we explore diamond tiling, a new tiling technique to exploit the inherent ability of most stencils to allow tile-wise concurrent start. This enables perfect load-balance during execution and reduces the frequency of synchronization required. Few studies have looked at time tiling for LBM codes. We exploit a key similarity between stencils and LBM to enable polyhedral optimizations and in turn time tiling for LBM. Besides polyhedral transformations, we also describe a number of other complementary transformations and post processing necessary to obtain good parallel and SIMD performance on modern architectures. We also characterize the performance of LBM with the Roofline performance model. Experimental results for standard LBM simulations like Lid Driven Cavity, Flow Past Cylinder, and Poiseuille Flow show that our scheme consistently outperforms Palabos–on average by3 x while running on 16 cores of a n Intel Xeon Sandy bridge system. We also obtain a very significant improvement of 2.47 x over the native production compiler on the SPECLBM benchmark.

Afin de limiter leur coût et/ou leur consommation électrique, certains processeurs embarqués sacrifient le support matériel de l'arithmétique à virgule flottante. Pourtant, pour des raisons de simplicité, les applications sont généralement spécifiées en utilisant l'arithmétique à virgule flottante. Porter ces applications sur des processeurs embarqués de ce genre nécessite une émulation logicielle de l'arithmétique à virgule flottante, qui peut sévèrement dégrader la performance. Pour éviter cela, l'application est converti pour utiliser l'arithmétique à virgule fixe, qui a l'avantage d'être plus efficace à implémenter sur des unités de calcul entier. La conversion de virgule flottante en virgule fixe est une procédure délicate qui implique des compromis subtils entre performance et précision de calcul. Elle permet, entre autre, de réduire la taille des données pour le coût de dégrader la précision de calcul. Par ailleurs, la plupart de ces processeurs fournissent un support pour le calcul vectoriel de type SIMD (Single Instruction Multiple Data) afin d'améliorer la performance. En effet, cela permet l'exécution d'une opération sur plusieurs données en parallèle, réduisant ainsi le temps d'exécution. Cependant, il est généralement nécessaire de transformer l'application pour exploiter les unités de calcul vectoriel. Cette transformation de vectorisation est sensible à la taille des données ; plus leurs tailles diminuent, plus le taux de vectorisation augmente. Il apparaît donc un compromis entre vectorisation et précision de calcul. Plusieurs travaux ont proposé des méthodologies permettant, d'une part la conversion automatique de virgule flottante en virgule fixe, et d'autre part la vectorisation automatique. Dans l'état de l'art, ces deux transformations sont considérées indépendamment, pourtant elles sont fortement liées. Dans ce contexte, nous étudions la relation entre ces deux transformations, dans le but d'exploiter efficacement le compromis entre performance et précision de calcul. Ainsi, nous proposons d'abord un algorithme amélioré pour l'extraction de parallélisme SLP (Superword Level Parallelism ; une technique de vectorisation). Puis, nous proposons une nouvelle méthodologie permettant l'application conjointe de la conversion de virgule flottante en virgule fixe et de l'exploitation du SLP. Enfin, nous implémentons cette approche sous forme d'un flot de compilation source-à-source complètement automatisé, afin de valider ces travaux. Les résultats montrent l'efficacité de cette approche, dans l'exploitation du compromis entre performance et précision, vis-à-vis d'une approche classique considérant ces deux transformations indépendamment. / In order to cut-down their cost and/or their power consumption, many embedded processors do not provide hardware support for floating-point arithmetic. However, applications in many domains, such as signal processing, are generally specified using floating-point arithmetic for the sake of simplicity. Porting these applications on such embedded processors requires a software emulation of floating-point arithmetic, which can greatly degrade performance. To avoid this, the application is converted to use fixed-point arithmetic instead. Floating-point to fixed-point conversion involves a subtle tradeoff between performance and precision ; it enables the use of narrower data word lengths at the cost of degrading the computation accuracy. Besides, most embedded processors provide support for SIMD (Single Instruction Multiple Data) as a mean to improve performance. In fact, this allows the execution of one operation on multiple data in parallel, thus ultimately reducing the execution time. However, the application should usually be transformed in order to take advantage of the SIMD instruction set. This transformation, known as Simdization, is affected by the data word lengths ; narrower word lengths enable a higher SIMD parallelism rate. Hence the tradeoff between precision and Simdization. Many existing work aimed at provide/improving methodologies for automatic floating-point to fixed-point conversion on the one side, and Simdization on the other. In the state-of-the-art, both transformations are considered separately even though they are strongly related. In this context, we study the interactions between these transformations in order to better exploit the performance/accuracy tradeoff. First, we propose an improved SLP (Superword Level Parallelism) extraction (an Simdization technique) algorithm. Then, we propose a new methodology to jointly perform floating-point to fixed-point conversion and SLP extraction. Finally, we implement this work as a fully automated source-to-source compiler flow. Experimental results, targeting four different embedded processors, show the validity of our approach in efficiently exploiting the performance/accuracy tradeoff compared to a typical approach, which considers both transformations independently.