Search

Simultaneous Localization And Mapping (SLAM) is crucial for autonomous navigation. ORB-SLAM is a state-of-the-art Visual SLAM system based on cameras used for self-driving cars. In this paper, we propose a high-performance, energy-efficient, and functionally accurate hardware accelerator for ORB-SLAM, focusing on its most time-consuming stage: Oriented FAST and Rotated BRIEF (ORB) feature extraction. The Rotated BRIEF (rBRIEF) descriptor generation is the main bottleneck in ORB computation, as it exhibits highly irregular access patterns to local on-chip memories causing a high-performance penalty due to bank conflicts. We introduce a technique to find an optimal static pattern to perform parallel accesses to banks based on a genetic algorithm. Furthermore, we propose the combination of an rBRIEF pixel duplication cache, selective ports replication, and pipelining to reduce latency without compromising cost. The accelerator achieves a reduction in energy consumption of 14597× and 9609×, with respect to high-end CPU and GPU platforms, respectively., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020- 113172RB-I00, the ICREA Academia program and the FPU grant FPU18/04413, Peer Reviewed, Postprint (published version)

GPGPU architectures have become the dominant platform for massively parallel workloads, delivering high performance and energy efficiency for popular applications such as machine learning, computer vision or self-driving cars. However, irregular applications, such as graph processing, fail to fully exploit GPGPU resources due to their divergent memory accesses that saturate the memory hierarchy. To reduce the pressure on the memory subsystem for divergent memory-intensive applications, programmers must take into account SIMT execution model and memory coalescing in GPGPUs, devoting significant efforts in complex optimization techniques. Despite these efforts, we show that irregular graph processing still suffers from low GPGPU performance. We observe that in many irregular applications the mapping of data to threads can be safely changed. In other words, it is possible to relax the strict relationship between thread and data processed to reduce memory divergence. Based on this observation, we propose the Irregular accesses Reorder Unit (IRU), a novel hardware extension tightly integrated in the GPGPU pipeline. The IRU reorders data processed by the threads on irregular accesses to improve memory coalescing, i.e., it tries to assign data elements to threads as to produce coalesced accesses in SIMT groups. Furthermore, the IRU is capable of filtering and merging duplicated accesses, significantly reducing the workload. Programmers can easily utilize the IRU with a simple API, or let the compiler issue instructions from our extended ISA. We evaluate our proposal for state-of-the-art graph-based algorithms and a wide selection of applications. Results show that the IRU achieves a memory coalescing improvement of 1.32x and a 46% reduction in the overall traffic in the memory hierarchy, which results in 1.33x speedup and 13% energy savings on average, while incurring in a small 5.6% area overhead., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00 and the ICREA Academia program., Peer Reviewed, Postprint (published version)

Binary Neural Networks (BNNs) are showing tremen-dous success on realistic image classification tasks. Notably, their accuracy is similar to the state-of-the-art accuracy obtained by full-precision models tailored to edge devices. In this regard, BNNs are very amenable to edge devices since they employ 1-bit to store the inputs and weights, and thus, their storage requirements are low. Moreover, BNNs computations are mainly done using xnor and pop-counts operations which are implemented very efficiently using simple hardware structures. Nonetheless, supporting BNNs efficiently on mobile CPUs is far from trivial since their benefits are hindered by frequent memory accesses to load weights and inputs. In BNNs, a weight or an input is stored using one bit, and aiming to increase storage and computation efficiency, several of them are packed together as a sequence of bits. In this work, we observe that the number of unique sequences representing a set of weights or inputs is typically low (i.e., 512). Also, we have seen that during the evaluation of a BNN layer, a small group of unique sequences is employed more frequently than others. Accordingly, we propose exploiting this observation by using Huffman Encoding to encode the bit sequences and then using an indirection table to decode them during the BNN evaluation. Also, we propose a clustering-based scheme to identify the most common sequences of bits and replace the less common ones with some similar common sequences. As a result, we decrease the storage requirements and memory accesses since the most common sequences are encoded with fewer bits. In this work, we extend a mobile CPU by adding a small hardware structure that can efficiently cache and decode the compressed sequence of bits. We evaluate our scheme using the ReAacNet model with the Imagenet dataset on an ARM CPU. Our experimental results show that our technique can reduce memory requirement by 1.32x and improve performance by 1.35x., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, and the ICREA Academia program., Peer Reviewed, Postprint (author's final draft)

Treball presentat a: 1st Workshop on Computer Architecture Modeling and Simulation (CAMS 2023), GPU architectures have become popular for executing generalpurpose programs. Their many-core architecture supports a large number of threads that run concurrently to hide the latency among dependent instructions. In modern GPU architectures, each SM/core is typically composed of several sub-cores, where each sub-core has its own independent pipeline. Simulators are a key tool for investigating novel concepts in computer architecture. They must be performance-accurate and have a proper model related to the target hardware to explore the different bottlenecks properly. This paper presents a wide analysis of different parts of Accelsim, a popular GPGPU simulator, and some improvements of its model. First, we focus on the front-end and developed a more realistic model. Then, we analyze the way the result bus works and develop a more realistic one. Next, we describe the current memory pipeline model and propose a model for a more cost-effective design. Finally, we discuss other areas of improvement of the simulator., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, and the ICREA Academia program., Preprint

With the rising need for efficient image processing in emerging applications such as Autonomous Driving (AD) and Augmented/Virtual Reality (AR/VR), many existing solutions do not meet their performance and energy efficiency requirements or are domain-specific and lack generality. In this work, we introduce SLIDEX, a novel ISA extension that leverages Sliding Window Processing (SWP) to bridge this gap in the image processing domain. SWP is a novel SIMD model that exposes to the programmer and natively manipulates vector registers as groups of overlapped windows of pixels to exploit the sliding-window dataflow found in convolutions and other stencil operations. SWP amplifies the available Data Level Parallelism (DLP) and reduces memory and register file accesses. We evaluated SLIDEX benefits in the critical image processing task of a state-of-the-art visual localization system widely used in AD and AR/VR. SLIDEX obtains an ~1.2× overall speedup and 22% energy reduction., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, the ICREA Academia program, and the FPU grant FPU18/04413., Peer Reviewed, Postprint (author's final draft)

Literature is plentiful in works exploiting cache locality for GPUs. A majority of them explore replacement or bypassing policies. In this paper, however, we surpass this exploration by fabricating a formal proof for a no-overhead quasi-optimal caching technique for caching textures in graphics workloads. Textures make up a significant part of main memory traffic in mobile GPUs, which contributes to the total GPU energy consumption. Since texture accesses use a shared L2 cache, improving the L2 texture caching efficiency would decrease main memory traffic, thus improving energy efficiency, which is crucial for mobile GPUs. Our proposal reaches quasi-optimality by exploiting the frame-to-frame reuse of textures in graphics. We do this by traversing frames in a boustrophedonic 1 1 Boustrophedon is a style of writing in which alternate lines of writing are reversed in order. This is in contrast to most modern languages, where the order of lines is the same, usually left-to-right. manner w.r.t. the frame-to-frame tile order. We first approximate the texture access trace to a circular trace and then forge a formal proof for our proposal being optimal for such traces. We also complement the proof with empirical data that demonstrates the quasi-optimality of our no-cost proposal., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, the ICREA Academia program and the AGAUR grant 2020-FISDU-00287., Peer Reviewed, Postprint (author's final draft)

Treball presentat a: RoboARCH, Workshop on Robotics Acceleration with Computing Hardware, an MICRO 2023 Workshop., Modern robots collect and process point clouds to perform accurate registration and segmentation. The most time-consuming kernel within point cloud processing -namely neighbor search- relies on appropriate data structures, such as k-d trees to prune the search space. In this work, we exploit similarity in subsequent k-d tree traversals to improve search on CPUs leveraging vector units. We show the potential for vectorizing k-d tree search using modern vector hardware, and performance benefits with a software-only implementation. Further, we discuss limitations in current architectures and present in-progress ideas to specialize the CPU vector hardware for an efficient k-d tree vector search., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, the ICREA Academia program, and a FPI-UPC 2020 grant., Preprint

The constant growth of DNNs makes them challenging to implement and run efficiently on traditional computecentric architectures. Some works have attempted to enhance accelerators by adding more compute units and on-chip buffers, but they often worsen the memory issue due to increased bandwidth demands. Memory-centric designs based on Near-Data Processing (NDP) have been proposed to mitigate this problem by moving computations closer to the memory hierarchy. Leveraging 3D-stacked memory for its storage density and near-memory processing capabilities, this paper introduces QeiHaN, a hardware accelerator that optimizes DNN inference efficiency. QeiHaN employs a 3D-stacked memory-centric weight storage scheme combined with a logarithmic quantization of activations, resulting in reduced memory accesses by 25%. Evaluation demonstrates significant speedup and energy savings compared to a Neurocube-like accelerator across various DNNs., QeiHaN has been supported by the CoCoUnit ERC Advanced Grant of EU’s Horizon 2020 (grant No 833057), Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, and ICREA Academia program., Peer Reviewed, Postprint (author's final draft)

Modern GPUs benefit from a sizable Register File (RF) to provide fine-grained thread switching. As the RF is huge and accessed frequently, it consumes a considerable share of the dynamic energy of the GPU. Designing a large, high-throughput RF with low energy consumption and area for GPUs is challenging. In this paper, an energy-efficient hierarchical RF design for GPUs, called Malekeh, is introduced. Malekeh keeps registers in energy-efficient small caches and maximizes cache efficacy by using lightweight policies and supporting adaptive algorithms. The policies’ effectiveness is improved by leveraging register reuse distance information provided by the compiler as a hint. Malekeh reduces the RF reads by 48.5% and dynamic energy by 29.1%. It also improves performance by 9.6% with a negligible overhead of 0.04% in the area., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020- 113172RB-I00, and the ICREA Academia program., Peer Reviewed, Postprint (author's final draft)

GPU architectures have become popular for executing general-purpose programs which rely on having a large number of threads that run concurrently to hide the latency among dependent instructions. This approach has an important cost/overhead in terms of low data locality due to the increased pressure on the memory hierarchy of the many threads being run concurrently and the extra cost of storing and managing the on-chip state of those many threads. This paper presents SOCGPU (Simple Out-of-order Core for GPU), a simple out-of-order execution mechanism that does not require register renaming nor scoreboards. It uses a small Instruction Buffer and a tiny Dependence matrix to keep track of dependencies among instructions and avoid data hazards. Evaluations for an Nvidia Tesla V100-like GPU show that SOCGPU provides a speed-up of up to 2.3 in some machine learning programs and 1.38 on average for a variety of benchmarks, while it reduces energy consumption by 6.5%, with only 2.4% area overhead., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, and the ICREA Academia program., Peer Reviewed, Postprint (author's final draft)

The effectiveness of Recurrent Neural Networks (RNNs) for tasks such as Automatic Speech Recognition has fostered interest in RNN inference acceleration. Due to the recurrent nature and data dependencies of RNN computations, prior work has designed customized architectures specifically tailored to the computation pattern of RNN, getting high computation efficiency for certain chosen model sizes. However, given that the dimensionality of RNNs varies a lot for different tasks, it is crucial to generalize this efficiency to diverse configurations. In this work, we identify adaptiveness as a key feature that is missing from today’s RNN accelerators. In particular, we first show the problem of low resource utilization and low adaptiveness for the state-of-the-art RNN implementations on GPU, FPGA, and ASIC architectures. To solve these issues, we propose an intelligent tiled-based dispatching mechanism for increasing the adaptiveness of RNN computation, in order to efficiently handle the data dependencies. To do so, we propose Sharp as a hardware accelerator, which pipelines RNN computation using an effective scheduling scheme to hide most of the dependent serialization. Furthermore, Sharp employs dynamic reconfigurable architecture to adapt to the model’s characteristics. Sharp achieves 2×, 2.8×, and 82× speedups on average, considering different RNN models and resource budgets, compared to the state-of-the-art ASIC, FPGA, and GPU implementations, respectively. Furthermore, we provide significant energy reduction with respect to the previous solutions, due to the low power dissipation of Sharp (321 GFLOPS/Watt)., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, and the ICREA Academia program., Peer Reviewed, Postprint (author's final draft)

The need for increased application performance in high-integrity systems like those in avionics is on the rise as software continues to implement more complex functionalities. The prevalent computing solution for future high-integrity embedded products are multi-processors systems-on-chip (MPSoC) processors. MPSoCs include CPU multicores that enable improving performance via thread-level parallelism. MPSoCs also include generic accelerators (GPUs) and application-specific accelerators. However, the data processing approach (DPA) required to exploit each of these underlying parallel hardware blocks carries several open challenges to enable the safe deployment in high-integrity domains. The main challenges include the qualification of its associated runtime system and the difficulties in analyzing programs deploying the DPA with out-of-the-box timing analysis and code coverage tools. In this work, we perform a thorough analysis of vector extensions (VExt) in current COTS processors for high-integrity systems. We show that VExt prevent many of the challenges arising with parallel programming models and GPUs. Unlike other DPAs, VExt require no runtime support, prevent by design race conditions that might arise with parallel programming models, and have minimum impact on the software ecosystem enabling the use of existing code coverage and timing analysis tools. We develop vectorized versions of neural network kernels and show that the NVIDIA Xavier VExt provide a reasonable increase in guaranteed application performance of up to 2.7x. Our analysis contends that VExt are the DPA approach with arguably the fastest path for adoption in high-integrity systems., This work has received funding from the the European Research Council (ERC) grant agreement No. 772773 (SuPerCom) and the Spanish Ministry of Science and Innovation (AEI/10.13039/501100011033) under grants PID2019-107255GB-C21 and IJC2020-045931-I., Peer Reviewed, Postprint (author's final draft)

Modern mobile GPUs integrate an increasing number of shader cores to speedup the execution of graphics workloads. Each core integrates a private Texture Cache to apply texturing effects on objects, which is backed-up by a shared L2 cache. However, as in any other memory hierarchy, such organization produces data replication in the upper levels (i.e., the private Texture Caches) to allow for faster accesses at the expense of reducing their overall effective capacity. E.g., in a mobile GPU with four shader cores, about 84.6% of the requested texture blocks are replicated in at least one of the other private Texture Caches. This paper proposes a novel dynamically-mapped NonUniform Cache Architecture (NUCA) organization for the private Texture Caches of a mobile GPU aimed at increasing their effective overall capacity and decreasing the overall access latency by attacking data replication. A block missing in a local Texture Cache may be serviced by a remote one at a cost smaller than a round trip to the shared L2. The proposed Dynamic Texture Mapping-NUCA (DTM-NUCA) features a lightweight mapping table, called Affinity Table, that is independent of the L2 cache size, unlike a traditional NUCA organization. The best owner for a given set of blocks is dynamically determined and stored in the Affinity Table to maximize local accesses. The mechanism also allows for a certain amount of replication to favor local accesses where appropriate, without hurting performance due to the small capacity loss resulting from the allowed replication. DTM-NUCA is presented in two flavors. One with a centralized Affinity Table, and another with a distributed Affinity Table. Experimental results show first that the L2 pressure is effectively reduced, eliminating 41.8% of the L2 accesses on average. As for the average latency, DTM-NUCA performs a very effective job at maximizing local over remote accesses, achieving 73.8% of local accesses on average. As a consequence, our novel DTM-NUCA orga, This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, the ICREA Academia program and a research fellowship from the University of Murcia’s “Plan Propio de Investigacion”., Peer Reviewed, Postprint (author's final draft)

Data-intensive workloads and applications, such as machine learning (ML), are fundamentally limited by traditional computing systems based on the von-Neumann architecture. As data movement operations and energy consumption become key bottlenecks in the design of computing systems, the interest in unconventional approaches such as Near-Data Processing (NDP), machine learning, and especially neural network (NN)-based accelerators has grown significantly. Emerging memory technologies, such as ReRAM and 3D-stacked, are promising for efficiently architecting NDP-based accelerators for NN due to their capabilities to work as both high-density/low-energy storage and in/near-memory computation/search engine. In this paper, we present a survey of techniques for designing NDP architectures for NN. By classifying the techniques based on the memory technology employed, we underscore their similarities and differences. Finally, we discuss open challenges and future perspectives that need to be explored in order to improve and extend the adoption of NDP architectures for future computing platforms. This paper will be valuable for computer architects, chip designers, and researchers in the area of machine learning., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, and the ICREA Academia program., Peer Reviewed, Postprint (published version)

This article proposes a novel micro-architecture approach for mobile GPUs aimed at early removing the occluded geometry in a scene by leveraging frame-to-frame coherence, thus reducing the overall energy consumption. Mobile GPUs commonly implement a Tile-Based Rendering (TBR) architecture that differentiates two main phases: the Geometry Pipeline, where all the geometry of a scene is processed; and the Raster Pipeline, where primitives are rendered in a framebuffer. After the Geometry Pipeline, only non-culled primitives inside the camera’s frustum are stored into the Parameter Buffer, a data structure stored in DRAM. However, among the non-culled primitives there is a significant amount that are rendered but non-visible at all, resulting in useless computations. On average, 60% of those primitives are completely occluded in our benchmarks. Despite TBR architectures use on-chip caches for the Parameter Buffer, about 46% of the DRAM traffic still comes from accesses to such buffer. The proposed Triangle Dropping technique leverages the visibility information computed along the Raster Pipeline to predict the primitives’ visibility in the next frame to early discard those that will be totally occluded, drastically reducing Parameter Buffer accesses. On average, our approach achieves overall 14.5% energy savings, 28.2% energy-delay product savings, and a speedup of 20.2%., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant no. 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00 (AEI/FEDER, EU), and the ICREA Academia program. D. Corbalán-Navarro has been also supported by a PhD research fellowship from the University of Murcia’s “Plan Propio de Investigación., Peer Reviewed, Postprint (author's final draft)

© 2022 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes,creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. http://dx.doi.org/10.1109/MICRO56248.2022.00028, Contemporary GPU architectures have multiple shader cores and a scheduler that distributes work (threads) among them, focusing on load balancing. These load balancing techniques favor thread distributions that are detrimental to texture memory locality for graphics applications in the L1 Texture Caches. Texture memory accesses make up the majority of the traffic to the memory hierarchy in typical low power graphics architectures. This paper focuses on improving the L1 Texture cache locality by focusing on a new workload scheduler by exploring various methods to group the threads, assign the groups to shader cores and also to reorder threads without violating the correctness of the pipeline. To overcome the resulting load imbalance, we also propose a minor modification in the GPU architecture that helps translate the improvement in cache locality to an improvement in the GPU’s performance. We propose DTexL that envelops these ideas and evaluate it over a benchmark suite of ten commercial games, to obtain a 46.8% decrease in L2 Accesses, a 19.3% increase in performance and a 6.3% decrease in total GPU energy. All this with a negligible overhead., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, the ICREA Academia program and the AGAUR grant 2020-FISDU-00287., Peer Reviewed, Postprint (author's final draft)

Deep Neural Networks (DNNs) have achieved tremendous success for cognitive applications. The core operation in a DNN is the dot product between quantized inputs and weights. Prior works exploit the weight/input repetition that arises due to quantization to avoid redundant computations in Convolutional Neural Networks (CNNs). However, in this paper we show that their effectiveness is severely limited when applied to FullyConnected (FC) layers, which are commonly used in state-of-the-art DNNs, as it is the case of modern Recurrent Neural Networks (RNNs) and Transformer models. To improve energy-efficiency of FC computation we present CREW, a hardware accelerator that implements Computation Reuse and an Efficient Weight Storage mechanism to exploit the large number of repeated weights in FC layers. CREW first performs the multiplications of the unique weights by their respective inputs and stores the results in an on-chip buffer. The storage requirements are modest due to the small number of unique weights and the relatively small size of the input compared to convolutional layers. Next, CREW computes each output by fetching and adding its required products. To this end, each weight is replaced offline by an index in the buffer of unique products. Indices are typically smaller than the quantized weights, since the number of unique weights for each input tends to be much lower than the range of quantized weights, which reduces storage and memory bandwidth requirements. Overall, CREW greatly reduces the number of multiplications and provides significant savings in model memory footprint and memory bandwidth usage. We evaluate CREW on a diverse set of modern DNNs. On average, CREW provides 2.61x speedup and 2.42x energy savings over a TPU-like accelerator. Compared to UCNN, a state-of-art computation reuse technique, CREW achieves 2.10x speedup and 2.08x energy savings on average., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, the ICREA Academia program, and the Spanish Ministry of Education under grant FPU15/02294., Peer Reviewed, Postprint (published version)

An important drawback of cycle-accurate microarchitectural simulators is that they are several orders of magnitude slower than the system they model. This becomes an important issue when simulations have to be repeated multiple times sweeping over the desired design space. In the specific context of graphics workloads, performing cycle-accurate simulations are even more demanding due to the high number of triangles that have to be shaded, lighted and textured to compose a single frame. As a result, simulating a few minutes of a video game sequence is extremely time-consuming.In this paper, we make the observation that collecting information about the vertices and primitives that are processed, along with the times that shader programs are invoked, allows us to characterize the activity performed on a given frame. Based on that, we propose a novel methodology for the efficient simulation of graphics workloads called MEGsim, an approach that is capable of accurately characterizing entire video sequences by using a small subset of selected frames which substantially drops the simulation time. For a set of popular Android games, we show that MEGsim achieves an average simulation speedup of 126×, achieving remarkably accurate results for the estimated final statistics, e.g., with average relative errors of just 0.84% for the total number of cycles, 0.99% for the number of DRAM accesses, 1.2% for the number of L2 cache accesses, and 0.86% for the number of L1 (tile cache) accesses., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, and the ICREA Academia program., Peer Reviewed, Postprint (author's final draft)

In real-time rendering, a 3D scene is modelled with meshes of triangles that the GPU projects to the screen. They are discretized by sampling each triangle at regular space intervals to generate fragments which are then added texture and lighting effects by a shader program. Realistic scenes require detailed geometric models, complex shaders, high-resolution displays and high screen refreshing rates, which all come at a great compute time and energy cost. This cost is often dominated by the fragment shader, which runs for each sampled fragment. Conventional GPUs sample the triangles once per pixel; however, there are many screen regions containing low variation that produce identical fragments and could be sampled at lower than pixel-rate with no loss in quality. Additionally, as temporal frame coherence makes consecutive frames very similar, such variations are usually maintained from frame to frame. This work proposes Dynamic Sampling Rate (DSR), a novel hardware mechanism to reduce redundancy and improve the energy efficiency in graphics applications. DSR analyzes the spatial frequencies of the scene once it has been rendered. Then, it leverages the temporal coherence in consecutive frames to decide, for each region of the screen, the lowest sampling rate to employ in the next frame that maintains image quality. We evaluate the performance of a state-of-the-art mobile GPU architecture extended with DSR for a wide variety of applications. Experimental results show that DSR is able to remove most of the redundancy inherent in the color computations at fragment granularity, which brings average speedups of 1.68x and energy savings of 40%., This work has been supported by the the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (Grant No. 833057), Spanish State Research Agency (MCIN/AEI) under Grant PID2020-113172RB-I00, the ICREA Academia program, and the Generalitat de Catalunya under Grant FI-DGR 2016. Funding was provided by Ministerio de Economía, Industria y Competitividad, Gobierno de España (Grant No. TIN2016-75344-R)., Peer Reviewed, Postprint (published version)

Graph-based applications are essential in emerging domains such as data analytics or machine learning. Data gathering in a knowledge-based society requires great data processing efficiency. High-throughput GPGPU architectures are key to enable efficient graph processing. Nonetheless, irregular and sparse memory access patterns present in graph-based applications induce high memory divergence and contention, which result in poor GPGPU efficiency for graph processing. Recent work has pointed out the importance of stream compaction operations, and has proposed a Stream Compaction Unit (SCU) to offload them to a specialized hardware. On the other hand, memory contention caused by high divergence has been tackled with the Irregular accesses Reorder Unit (IRU), delivering improved memory coalescing. In this paper, we propose a new unit, the IRU-enhanced SCU (ISCU), that leverages the strengths of both approaches. The ISCU employs the efficient mechanisms of the IRU to improve SCU stream compaction efficiency and throughput limitations, achieving a synergistic effect for graph processing. We evaluate the ISCU for a wide variety of state-of-the-art graph-based algorithms and applications. Results show that the ISCU achieves a performance speedup of 2.2x and 90% energy savings derived from a high reduction of 78% memory accesses, while incurring in 8.5% area overhead., Peer Reviewed, Postprint (author's final draft)

DNN pruning reduces memory footprint and computational work of DNN-based solutions to improve performance and energy-efficiency. An effective pruning scheme should be able to systematically remove connections and/or neurons that are unnecessary or redundant, reducing the DNN size without any loss in accuracy. In this paper we show that some of the most popular pruning schemes, such as the Near Zero Weights, require an extremely time-consuming iterative process that requires retraining the DNN many times to tune the pruning hyperparameters. Then, we propose a DNN pruning scheme based on Principal Component Analysis and relative importance of each neuron’s connection (PCA+DIRIE) that automatically finds the optimized DNN in one shot without requiring hand-tuning of multiple parameters. The experimental results show the effectiveness of our method on several benchmarks. Notably, on ImageNet, PCA+DIRIE can prune up to 60% of ResNet-50 with negligible impact on accuracy., This work has been supported by the CoCoUnit ERC Advanced Grantof the EU’s Horizon 2020 program (grant No 833057), the Spanish StateResearch Agency under grant PID2020-113172RB-I00 (MCIN/AEI), theICREA Academia program, and the Spanish Ministry of Education undergrant FPU15/02294., Peer Reviewed, Postprint (published version)

The most common task of GPUs is to render images in real time. When rendering a 3D scene, a key step is to determine which parts of every object are visible in the final image. There are different approaches to solve the visibility problem, the Z-Test being the most common. A main factor that significantly penalizes the energy efficiency of a GPU, especially in the mobile arena, is the so-called overdraw, which happens when a portion of an object is shaded and rendered but finally occluded by another object. This useless work results in a waste of energy; however, a conventional Z-Test only avoids a fraction of it. In this paper we present a novel microarchitectural technique, the Omega-Test, to drastically reduce the overdraw on a Tile-Based Rendering (TBR) architecture. Graphics applications have a great degree of inter-frame coherence, which makes the output of a frame very similar to the previous one. The proposed approach leverages the frame-to-frame coherence by using the resulting information of the Z-Test for a tile (a buffer containing all the calculated pixel depths for a tile), which is discarded by nowadays GPUs, to predict the visibility of the same tile in the next frame. As a result, the Omega-Test early identifies occluded parts of the scene and avoids the rendering of non-visible surfaces eliminating costly computations and off-chip memory accesses. Our experimental evaluation shows average EDP savings in the overall GPU/Memory system of 26.4% and an average speedup of 16.3% for the evaluated benchmarks., This work has been supported by the the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency under grant TIN2016-75344-R (AEI/FEDER, EU) and the ICREA Academia program. D. Corbalan-Navarro has been supported by a PhD research fellowship from the University of Murcia., Peer Reviewed, Postprint (author's final draft)

In this work, we extend the auto-tuning process of the state-of-the-art TVM framework with XFeatur; a tool that extracts new meaningful hardware-related features that improve the quality of the representation of the search space and consequently improve the accuracy of its prediction algorithm. These new features provide information about the amount of thread-level parallelism, shared memory usage, register usage, dynamic instruction count and memory access dependencies. Optimizing ResNet-18 with the proposed features improves the quality of the search space representation by 63% on average and a maximum of 2× for certain tasks, while it reduces the tuning time by 9% (approximately 1.1 hours) and produces configurations that have equal or better performance (up to 92.7%) than the baseline., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, and the ICREA Academia program and the FPU grant 2019-FPU-998758., Peer Reviewed, Postprint (author's final draft)

Recurrent Neural Network (RNN) inference exhibits low hardware utilization due to the strict data dependencies across time-steps. Batching multiple requests can increase throughput. However, RNN batching requires a large amount of padding since the batched input sequences may vastly differ in length. Schemes that dynamically update the batch every few time-steps avoid padding. However, they require executing different RNN layers in a short time span, decreasing energy efficiency. Hence, we propose E-BATCH, a low-latency and energy-efficient batching scheme tailored to RNN accelerators. It consists of a runtime system and effective hardware support. The runtime concatenates multiple sequences to create large batches, resulting in substantial energy savings. Furthermore, the accelerator notifies it when the evaluation of an input sequence is done. Hence, a new input sequence can be immediately added to a batch, thus largely reducing the amount of padding. E-BATCH dynamically controls the number of time-steps evaluated per batch to achieve the best trade-off between latency and energy efficiency for the given hardware platform. We evaluate E-BATCH on top of E-PUR and TPU. E-BATCH improves throughput by 1.8× and energy efficiency by 3.6× in E-PUR, whereas in TPU, it improves throughput by 2.1× and energy efficiency by 1.6×, over the state-of-the-art., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, and the ICREA Academia program., Peer Reviewed, Postprint (author's final draft)

Automatic Speech Recognition (ASR) is quickly becoming a mainstream technology, mainly driven by the outstanding accuracy achieved by modern systems based on machine learning. However, these systems often require billions of arithmetic operations to decode a second of audio and relying on cloud services for ASR is usually inconvenient. Even though deployment of ASR systems directly on the edge is highly desirable, the requirements for high performance and low energy consumption, combined with the fast pace of evolution and heterogeneity of existing ASR systems, result in challenges for effective deployment of ASR on edge devices. In this work, we propose a programmable accelerator to efficiently support a variety of ASR implementations. We estimate the performance of our system by implementing a recently proposed streaming ASR system and show that it can perform real-time streaming decoding with a tight power budget and low area footprint while offering great flexibility to implement a variety of different models., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, the ICREA Academia program and the Spanish MICINN Ministry under grant BES-2017-080605., Peer Reviewed, Postprint (published version)

Camera-based autonomous driving extensively ma-nipulates images for object detection, object tracking, or camera-based localization tasks. Therefore, efficient and fast image processing is crucial in those systems. Unfortunately, current solutions either do not meet AD’s constraints for real-time performance and energy efficiency or are domain-specific and, thus, not general [14]. In this work, we introduce Sliding Window Processing (SWP), a SIMD execution model that natively operates on sliding windows of image pixels. We illustrate the benefits of SWP through a novel ISA extension called SLIDEX that achieves high performance and energy efficiency while maintaining pro-grammability. We demonstrate the benefits of SLIDEX for the image processing tasks of ORB-SLAM [17] [18], a state-of-the-art camera-based localization system. SLIDEX achieves an average end-to-end speedup of ~1.65× and ~1.2× compared to equivalent scalar and vector baselines respectively. Compared with the vector implementation, our solution reduces the end-to-end energy consumption a 22% on average., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, the ICREA Academia program and the FPU grant FPU18/04413., Peer Reviewed, Postprint (published version)

Cache Replacement Policies are known to have an important impact on hit rates. The OPT replacement policy [27] has been formally proven as optimal for minimizing misses. Due to its need to look far ahead for future memory accesses, it is often reduced to a yardstick for measuring the efficacy of other practical caches. In this paper, we bring the OPT to life, in architectures for mobile GPUs, for which energy efficiency is of great consequence. We also mold other factors in the memory hierarchy to enhance its impact. The end results are a 13.8% decrease in the memory hierarchy energy consumption and an increased throughput in the Tiling Engine. We also observe a 5.5% decrease in the total GPU energy and a 3.7% increase in frames per second (FPS)., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, the ICREA Academia program and the AGAUR grant 2020-FISDU-00287. We would also like to thank the anonymous reviewers for their valuable comments., Peer Reviewed, Postprint (author's final draft)

Simultaneous Localization And Mapping (SLAM) is a key component for autonomous navigation. SLAM consists of building and creating a map of an unknown environment while keeping track of the exploring agent's location within it. An effective implementation of SLAM presents important challenges due to real-time inherent constraints and energy consumption. ORB-SLAM is a state-of-the-art Visual SLAM system based on cameras that can be used for self-driving cars. In this paper, we propose a high-performance, energy-efficient and functionally accurate hardware accelerator for ORB-SLAM, focusing on its most time-consuming stage: Oriented FAST and Rotated BRIEF (ORB) feature extraction. We identify the BRIEF descriptor generation as the main bottleneck, as it exhibits highly irregular access patterns to local on-chip memories, causing a high performance penalty due to bank conflicts. We propose a genetic algorithm to generate an optimal memory access pattern offline, which greatly simplifies the hardware while minimizing bank conflicts in the computation of the BRIEF descriptor. Compared with a CPU system, the accelerator achieves 8x speedup and 1957x reduction in power dissipation., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency (MCIN/AEI) under grant PID2020-113172RB-I00, the ICREA Academia program and the FPU grant FPU18/04413., Peer Reviewed, Postprint (author's final draft)

Autonomous Vehicles (AVs) have the potential to radically change the automotive industry. How- ever, computing solutions for AVs have to meet severe performance constraints to guarantee a safe driving experience. Current solutions either exhibit high cost or fail to meet the stringent latency constraints. Therefore, the popularization of AVs requires a low-cost yet effective computing sys- tem. Understanding the sources of latency is key in order to improve autonomous driving systems. Here, we present a detailed characterization of Autoware, a modern self-driving car system. We analyze the performance of the different components and leverage hardware counters to identify the main bottlenecks., This work has been supported by the the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency under grant PID2020-113172RB-I00 (AEI/FEDER, EU), the ICREA Academia program, and the grant 2020 FPI-UPC_033., Peer Reviewed, Postprint (published version)

The most common task of GPUs is to render images in real time. When rendering a 3D scene, a key step is determining which parts of every object are visible in the final image. There are different approaches to solve the visibility problem, the Z-Test being the most common in modern GPUs. A main factor that significantly penalizes the energy efficiency of a GPU, especially in the mobile arena, is the so-called overshading, which happens when a portion of an object is shaded and rendered but finally occluded by another object. This useless work results in a waste of energy, however, the conventional Z-Test only eliminates a fraction of it. In this paper we present a novel microarchitectural technique, the ¿-Test, to drastically reduce overshading on a Tile-Based Rendering (TBR) architecture. The proposed approach leverages frame-to-frame coherence by taking advantage of the costly and valuable calculations made in previous frames. In particular, we propose to reuse information from the Z-Buffer of the previous frame, which is currently discarded. We make the observation that due to the existing frame-to-frame coherence, the Z-Buffer of a frame will have a high similarity in many areas with that of the previous frame. As a result, the proposed technique avoids many costly computations and off-chip memory accesses. Our experimental evaluation shows that ¿-Test reduces the average energy consumption of the overall GPU/Memory system by 15.7 % and the runtime of the evaluated benchmarks by 10.6 % on average., This work has been supported by the the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057), the Spanish State Research Agency under grant TIN2016-75344-R (AEI/-FEDER, EU) and the ICREA Academia program. D. Corbal´an-Navarro has been supported by a PhD research fellowship from the University of Murcia., Peer Reviewed, Postprint (author's final draft)

Soft Error Rate (SER) estimation is an important challenge for integrated circuits because of the increased vulnerability brought by technology scaling. This paper presents a methodology to estimate in early stages of the design the susceptibility of combinational circuits to particle strikes. In the core of the framework lies MASkIt , a novel approach that combines signal probabilities with technology characterization to swiftly compute the logical, electrical, and timing masking effects of the circuit under study taking into account all input combinations and pulse widths at once. Signal probabilities are estimated applying a new hybrid approach that integrates heuristics along with selective simulation of reconvergent subnetworks. The experimental results validate our proposed technique, showing a speedup of two orders of magnitude in comparison with traditional fault injection estimation with an average estimation error of 5 percent. Finally, we analyze the vulnerability of the Decoder, Scheduler, ALU, and FPU of an out-of-order, superscalar processor design., This work has been partially supported by the Spanish Ministry of Economy and Competitiveness and Feder Funds under grant TIN2013-44375-R, by the Generalitat de Catalunya under grant FI-DGR 2016, and by the FP7 program of the EU under contract FP7-611404 (CLERECO)., Peer Reviewed, Postprint (author's final draft)

Automatic Speech Recognition (ASR) has experienced a dramatic evolution since pioneer development of Bell Lab’s single-digit recognizer more than 50 years ago. Current ASR systems have taken advantage of the tremendous improvements in AI during the past decade by incorporating Deep Neural Networks into the system and pushing their accuracy to levels comparable to that of humans. This article describes and characterizes a representative ASR system with state-of-the-art accuracy and proposes a hardware platform capable of decoding speech in real-time with a power dissipation close to 1 Watt. The software is based on the so-called hybrid approach with a vocabulary of 200K words and RNN-based language model re-scoring, whereas the hardware consists of a commercially available low-power processor along with two accelerators used for the most compute-intensive tasks. The article shows that high performance can be obtained with very low power, enabling the deployment of these systems in extremely power-constrained environments such as mobile and IoT devices., This work has been supported by the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No. 833057), the Spanish State Research Agency under grant TIN2016-75344-R (AEI/FEDER, EU), the ICREA Academia program, and the Spanish MICINN Ministry under grant BES-2017-080605., Peer Reviewed, Postprint (published version)

Temporary housing units (THUs), which are provided after disasters, are crucial in terms of sustainability pillars (economic, social, and environmental). In general, THUs, which are regular houses with minimum space and facilities, incorporate some negative aspects of the building industry. Additionally, as large numbers of THUs are usually provided in a short time and under emergency situations, some negative impacts of these units escalate. In this context, this study aims at reducing some negative impacts of THUs by applying a novel optimization model that maximises sustainability indexes by simulating the design of interior geometries for THUs. This method is based on the coupling of artificial intelligence and a multi-criteria decision-making model for sustainability assessment. The proposed model generates optimal solutions using a backtracking algorithm together with a binary search. To evaluate the sustainability indexes, an Integrated Value Model for Sustainability Assessment (MIVES) is applied. This novel method enables decision makers to automatically generate the most suitable alternative solutions for the early design stage of THUs. The results confirm that small changes in the interior geometric design can remarkably affect the sustainability indexes of THUs., The authors would like to acknowledge the kind support offered by professors and researchers from Shahid Beheshti University, Islamic Azad University, the Polytechnic University of Catalonia, and experts from the Housing Foundation of Islamic Republic of Iran (HFIR), who supported this paper for collecting and improving data. Additionally, the authors would like to thank UPC's Centre for Development Cooperation (CCD) for funding the authors (NO.: 2017-U006 and 2018-U012), which enabled the authors to follow and analyse the recovery program of Kermanshah earthquake., Peer Reviewed, Postprint (author's final draft)

©2020 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes,creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works., Autonomous Vehicles (AVs) have the potential to radically change the automotive industry. However, computing solutions for AVs have to meet severe performance and power constraints to guarantee a safe driving experience. Current solutions either exhibit high cost and power dissipation or fail to meet the stringent latency constraints. Therefore, the popularization of AVs requires a low-cost yet effective computing system. Understanding the sources of latency and energy consumption is key in order to improve autonomous driving systems. In this paper, we present a detailed characterization of Autoware, a modern self-driving car system. We analyze the performance and power of the different components and leverage hardware counters to identify the main bottlenecks. Our approach to AV characterization avoids pitfalls of previous works: profiling individual components in isolation and neglecting LiDAR-related components. We base our characterization on a rigorous methodology that considers the entire software stack. Profiling the end-to-end system accounts for interference and contention among different components that run in parallel, also including memory transfers to communicate data. We show that all these factors have a high impact on latency and cannot be measured by profiling isolated modules. Our characterization provides novel insights, some of the interesting findings are the following. First, contention among different modules drastically impacts latency and performance predictability. Second, LiDAR-related components are important contributors to the latency of the system. Finally, a modern platform with a high-end CPU and GPU cannot achieve real-time performance when considering the entire end-to-end system., This work has been supported by the the CoCoUnit ERC Advanced Grant of the EU’s Horizon 2020 program (grant No 833057) and the Spanish State Research Agency under grant TIN2016-75344-R (AEI/FEDER, EU)., Peer Reviewed, Postprint (author's final draft)

Automatic Speech Recognition (ASR) systems are changing the way people interact with different applications on mobile devices. Fulfilling such user-interactivity requires not only a highly accurate, large-vocabulary recognition system, but also a real-time, energy-efficient solution. However, these ASR systems need high memory bandwidth and power budget, which may be impractical for most of small form-factor battery-operated devices. In this article, we propose two combined techniques implemented on top of a state-of-the-art ASR accelerator in order to significantly reduce its energy consumption and memory requirements. First, by leveraging the locality among consecutive segments of the speech signal, we develop a Locality-AWare-Scheme (LAWS) which exploits the on-chip recently-explored data while removing most of the off-chip accesses during the ASR's decoding process. As a result, we remove up to 60 percent of ASR's workload. As the second step, we introduce an approach to improve LAWS's effectiveness by selectively adapting the amount of ASR's workload, based on run-time feedback. In particular, we exploit the fact that the confidence of the ASR system varies along the recognition process. When confidence is high, the ASR system can be more restrictive and reduce the amount of work. The end design including both techniques provides a saving of more than 87 percent in the memory requests and 2.3x reduction in energy consumption, and a speedup of 2.1x with respect to a state-of-the-art baseline design., This work has been supported in part by the CoCoUnit ERC Advanced Grant of the EUs Horizon 2020 program under Grant 833057, in part by the Spanish State Research Agency under Grant TIN2016-75344-R (AEI/FEDER, EU), and in part by the ICREA Academia program., Peer Reviewed, Postprint (author's final draft)

The use of low numerical precision is a fundamental optimization included in modern accelerators for Deep Neural Networks (DNNs). The number of bits of the numerical representation is set to the minimum precision that is able to retain accuracy based on an offline profiling, and it is kept constant for DNN inference. In this work, we explore the use of dynamic precision selection during DNN inference. We focus on Long Short Term Memory (LSTM) networks, which represent the state-of-the-art networks for applications such as machine translation and speech recognition. Unlike conventional DNNs, LSTM networks remember information from previous evaluations by storing data in the LSTM cell state. Our key observation is that the cell state determines the amount of precision required: time-steps where the cell state changes significantly require higher precision, whereas time-steps where the cell state is stable can be computed with lower precision without any loss in accuracy. We propose a novel hardware scheme that tracks the evolution of the elements in the LSTM cell state and dynamically selects the appropriate precision on each time-step. For a set of popular LSTM networks, it chooses the lowest precision for 57% of the time, outperforming systems that fix the precision statically. We evaluate our proposal on top of a modern highly-optimized LSTM accelerator, and show that it provides 1.46x speedup and 19.2% energy savings on average without degrading the model accuracy. Our scheme has an overhead of less than 8%., This work has been supported by the CoCoUnit ERC Advanced Grant of the ED's Horizon 2020 program (grant No 833057), the Spanish State Research Agency under grant TIN2016-75344-R (AEI/FEDER, EU), the ICREA Academia program, and the Fundación Carolina and PUCMM by a scholarship., Peer Reviewed, Postprint (author's final draft)

This chapter presents an overview of the main trends in processor architecture. It starts with an analysis of the past evolution of processors and the main driving forces behind it, and then it focuses on a description of the main architectural features of current processors. Finally, it presents a discussion on some promising directions for future evolution of processor architectures., Peer Reviewed, Postprint (author's final draft)

This paper summarizes our two-year study of corrected and uncor-rected errors on the MareNostrum 3 supercomputer, covering 2000 billion MB-hours of DRAM in the field. The study analyzes 4.5 million corrected and 71 uncorrected DRAM errors and it compares the reliability of DIMMs from all three major memory manufacturers, built in three different technologies. Our work has two sets of contributions. First, we illustrate the complexity of in-field DRAM error analysis and demonstrate the limitations of various widely-used methods and metrics. For example, we show that average error rates, errors per MB-hour and mean time between failures can provide volatile and unreliable results even after long periods of error logging, leading to incorrect conclusions about DRAM reliability. Second, we present formal statistical methods that overcome many of the limitations of the current approaches. The methods that we present are simple to understand and implement, reliable and widely accepted in the statistical community. Overall, our study alerts the community about the need to, firstly, question the current practice in quantifying DRAM reliability and, secondly, to select a proper analysis approach for future studies. Our strong recommendations are to focus on metrics with a practical value that could be easily related to system reliability, and to select methods that provide stable results, ideally supported with statistical significance., This work was supported by the Collaboration Agreement between Samsung Electronics Co., Ltd. and BSC, Spanish Government through Severo Ochoa programme (SEV-2015-0493), by the Spanish Ministry of Science and Technology through TIN2015-65316-P project and by the Generalitat de Catalunya (contracts 2014-SGR1051 and 2014-SGR-1272). This work has also received funding from the European Union’s Horizon 2020 research and innovation programme under EuroEXA project (grant agreement No 754337). Darko Zivanovic holds the Severo Ochoa grant (SVP-2014-068501) of the Ministry of Economy and Competitiveness of Spain., Postprint (author's final draft)

The final publication is available at ACM via http://dx.doi.org/10.1145/3352460.3358309, Recurrent Neural Networks (RNNs) are a key technology for applications such as automatic speech recognition or machine translation. Unlike conventional feed-forward DNNs, RNNs remember past information to improve the accuracy of future predictions and, therefore, they are very effective for sequence processing problems. For each application run, each recurrent layer is executed many times for processing a potentially large sequence of inputs (words, images, audio frames, etc.). In this paper, we make the observation that the output of a neuron exhibits small changes in consecutive invocations. We exploit this property to build a neuron-level fuzzy memoization scheme, which dynamically caches the output of each neuron and reuses it whenever it is predicted that the current output will be similar to a previously computed result, avoiding in this way the output computations. The main challenge in this scheme is determining whether the new neuron's output for the current input in the sequence will be similar to a recently computed result. To this end, we extend the recurrent layer with a much simpler Bitwise Neural Network (BNN), and show that the BNN and RNN outputs are highly correlated: if two BNN outputs are very similar, the corresponding outputs in the original RNN layer are likely to exhibit negligible changes. The BNN provides a low-cost and effective mechanism for deciding when fuzzy memoization can be applied with a small impact on accuracy. We evaluate our memoization scheme on top of a state-of-the-art accelerator for RNNs, for a variety of different neural networks from multiple application domains. We show that our technique avoids more than 24.2% of computations, resulting in 18.5% energy savings and 1.35x speedup on average., Peer Reviewed, Postprint (author's final draft)

© 2019 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes,creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works., In this work, we propose Locality-AWare-Scheme (LAWS) for an Automatic Speech Recognition (ASR) accelerator in order to significantly reduce its energy consumption and memory requirements, by leveraging the locality among consecutive segments of the speech signal. LAWS diminishes ASR's workload by up to 60% by removing most of the off-chip accesses during the ASR's decoding process. We furthermore improve LAWS's effectiveness by selectively adapting the amount of ASR's workload, based on run-time feedback. In particular, we exploit the fact that the confidence of the ASR system varies along the recognition process. When confidence is high, the ASR system can be more restrictive and reduce the amount of work. The end design provides a saving of 87% in memory requests, 2.3x reduction in energy consumption, and a speedup of 2.1x with respect to the state-of-the-art ASR accelerator., Peer Reviewed, Postprint (author's final draft)

© 2019 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes,creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works., Automatic Speech Recognition (ASR) is becoming increasingly ubiquitous, especially in the mobile segment. Fast and accurate ASR comes at high energy cost, not being affordable for the tiny power-budgeted mobile devices. Hardware acceleration reduces energy-consumption of ASR systems, while delivering high-performance. In this paper, we present an accelerator for largevocabulary, speaker-independent, continuous speech-recognition. It focuses on the Viterbi search algorithm representing the main bottleneck in an ASR system. The proposed design consists of innovative techniques to improve the memory subsystem, since memory is the main bottleneck for performance and power in these accelerators' design. It includes a prefetching scheme tailored to the needs of ASR systems that hides main memory latency for a large fraction of the memory accesses, negligibly impacting area. Additionally, we introduce a novel bandwidth-saving technique that removes off-chip memory accesses by 20 percent. Finally, we present a power saving technique that significantly reduces the leakage power of the accelerators scratchpad memories, providing between 8.5 and 29.2 percent reduction in entire power dissipation. Overall, the proposed design outperforms implementations running on the CPU by orders of magnitude, and achieves speedups between 1.7x and 5.9x for different speech decoders over a highly optimized CUDA implementation running on Geforce-GTX-980 GPU, while reducing the energy by 123-454x., Peer Reviewed, Postprint (author's final draft)

GPUs are one of the most energy-consuming components for real-time rendering applications, since a large number of fragment shading computations and memory accesses are involved. Main memory bandwidth is especially taxing battery-operated devices such as smart-phones. TileBased Rendering GPUs divide the screen space into multiple tiles that are independently rendered in on-chip buffers, thus reducing memory bandwidth and energy consumption. We have observed that, in many animated graphics workloads, a large number of screen tiles have the same color across adjacent frames. In this paper, we propose Rendering Elimination (RE), a novel micro-architectural technique that accurately determines if a tile will be identical to the same tile in the preceding frame before rasterization by means of comparing signatures. Since RE identifies redundant tiles early in the graphics pipeline, it completely avoids the computation and memory accesses of the most power consuming stages of the pipeline, which substantially reduces the execution time and the energy consumption of the GPU. For widely used Android applications, we show that RE achieves an average speedup of 1.74x and energy reduction of 43% for the GPU/Memory system, surpassing by far the benefits of Transaction Elimination, a state-of-the-art memory bandwidth reduction technique available in some commercial Tile-Based Rendering GPUs., Peer Reviewed, Postprint (published version)

© 2019 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes,creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works., Recurrent neural networks (RNNs) perform element-wise multiplications across the activations of gates. We show that a significant percentage of activations are saturated and propose coarse-grained pruning of activations (CGPA) to avoid the computation of entire neurons, based on the activation values of the gates. We show that CGPA can be easily implemented on top of a TPU-like architecture with negligible area overhead, resulting in 12% speedup and 12% energy savings on average for a set of widely used RNNs., Peer Reviewed, Postprint (author's final draft)

During real-time graphics rendering, objects are processed by the GPU in the order they are submitted by the CPU, and occluded surfaces are often processed even though they will end up not being part of the final image, thus wasting precious time and energy. To help discard occluded surfaces, most current GPUs include an Early-Depth test before the fragment processing stage. However, to be effective it requires that opaque objects are processed in a front-to-back order. Depth sorting and other occlusion culling techniques at the object level incur overheads that are only offset for applications having substantial depth and/or fragment shading complexity, which is often not the case in mobile workloads. We propose a novel architectural technique for mobile GPUs, Visibility Rendering Order (VRO), which reorders objects front-to-back entirely in hardware by exploiting the fact that the objects in graphics animated applications tend to keep its relative depth order across consecutive frames (temporal coherence). Since order relationships are already tested by the Depth Test, VRO incurs minimal energy overheads because it just requires adding a small hardware to capture that information and use it later to guide the rendering of the following frame. Moreover, unlike other approaches, this unit works in parallel with the graphics pipeline without any performance overhead. We illustrate the benefits of VRO using various unmodified commercial 3D applications for which VRO achieves 27% speed-up and 14.8% energy reduction on average over a state-of-the-art mobile GPU., Peer Reviewed, Postprint (author's final draft)

Graph processing algorithms are key in many emerging applications in areas such as machine learning and data analytics. Although the processing of large scale graphs exhibits a high degree of parallelism, the memory access pattern tend to be highly irregular, leading to poor GPGPU efficiency due to memory divergence. To ameliorate this issue, GPGPU applications perform a stream compaction operation each iteration of the algorithm to extract the subset of active nodes/edges, so subsequent steps work on compacted dataset. We show that GPGPU architectures are inefficient for stream compaction, and propose to offload this task to a programmable Stream Compaction Unit (SCU) tailored to the requirements of this kernel. The SCU is a small unit tightly integrated in the GPU that efficiently gathers the active nodes/edges into a compacted array in memory. Applications can make use of it through a simple API. The remaining steps of the graph-based algorithm are executed on the GPU cores taking benefit of the large amount of parallelism in the GPU, but they operate on the SCU-prepared data and achieve larger memory coalescing and, hence, much higher efficiency. Besides, the SCU performs filtering of repeated and already visited nodes during the compaction process, significantly reducing GPGPU workload, and writes the compacted nodes/edges in an order that improves memory coalescing by reducing memory divergence. We evaluate the performance of a state-of-the-art GPGPU architecture extended with our SCU for a wide variety of applications. Results show that for high-performance and for low-power GPU systems the SCU achieves speedups of 1.37x and 2.32x, 84.7% and 69% energy savings, and an area increase of 3.3% and 4.1% respectively., Peer Reviewed, Postprint (published version)

GPUs' main workload is real-time image rendering. These applications take a description of a (animated) scene and produce the corresponding image(s). An image is rendered by computing the colors of all its pixels. It is normal that multiple objects overlap at each pixel. Consequently, a significant amount of processing is devoted to objects that will not be visible in the final image, in spite of the widespread use of the Early Depth Test in modern GPUs, which attempts to discard computations related to occluded objects. Since animations are created by a sequence of similar images, visibility usually does not change much across consecutive frames. Based on this observation, we present Early Visibility Resolution (EVR), a mechanism that leverages the visibility information obtained in a frame to predict the visibility in the following one. Our proposal speculatively determines visibility much earlier in the pipeline than the Early Depth Test. We leverage this early visibility estimation to remove ineffectual computations at two different granularities: pixel-level and tile-level. Results show that such optimizations lead to 39% performance improvement and 43% energy savings for a set of commercial Android graphics applications running on stateof-the-art mobile GPUs., Peer Reviewed, Postprint (published version)

Recurrent Neural Networks (RNNs) are a key technology for emerging applications such as automatic speech recognition, machine translation or image description. Long Short Term Memory (LSTM) networks are the most successful RNN implementation, as they can learn long term dependencies to achieve high accuracy. Unfortunately, the recurrent nature of LSTM networks significantly constrains the amount of parallelism and, hence, multicore CPUs and many-core GPUs exhibit poor efficiency for RNN inference. In this paper, we present E-PUR, an energy-efficient processing unit tailored to the requirements of LSTM computation. The main goal of E-PUR is to support large recurrent neural networks for low-power mobile devices. E-PUR provides an efficient hardware implementation of LSTM networks that is flexible to support diverse applications. One of its main novelties is a technique that we call Maximizing Weight Locality (MWL), which improves the temporal locality of the memory accesses for fetching the synaptic weights, reducing the memory requirements by a large extent. Our experimental results show that E-PUR achieves real-time performance for different LSTM networks, while reducing energy consumption by orders of magnitude with respect to general-purpose processors and GPUs, and it requires a very small chip area. Compared to a modern mobile SoC, an NVIDIA Tegra X1, E-PUR provides an average energy reduction of 88x., Peer Reviewed, Postprint (published version)

DNN pruning has been recently proposed as an effective technique to improve the energy-efficiency of DNN-based solutions. It is claimed that by removing unimportant or redundant connections, the pruned DNN delivers higher performance and energy-efficiency with negligible impact on accuracy. However, DNN pruning has an important side effect: it May reduce the confidence of DNN predictions. We show that, although top-1 accuracy May be maintained with DNN pruning, the likelihood of the class in the top-1 is significantly reduced when using the pruned models. For applications such as Automatic Speech Recognition (ASR), where the DNN scores are consumed by a successive stage, the workload of this stage can be dramatically increased due to the loss of confidence in the DNN. An ASR system consists of a DNN for computing acoustic scores, followed by a Viterbi beam search to find the most likely sequence of words. We show that, when pruning the DNN model used for acoustic scoring, the Word Error Rate (WER) is maintained but the execution time of the ASR system is increased by 33%. Although pruning improves the efficiency of the DNN, it results in a huge increase of activity in the Viterbi search since the output scores of the pruned model are less reliable. Based on this observation, we propose a novel hardware-based ASR system that effectively integrates a DNN accelerator for pruned models with a Viterbi accelerator. In order to avoid the aforementioned increase in Viterbi search workload, our system loosely selects the N-best hypotheses at every time step, exploring only the N most likely paths. To avoid an expensive sort of the hypotheses based on their likelihoods, our accelerator employs a set-associative hash table to keep track of the best paths mapped to each set. In practice, this solution approaches the selection of N-best, but it requires much simpler hardware. Our approach manages to efficiently combine both DNN pruning and Viterbi search, and achieves 9x energy, Peer Reviewed, Postprint (published version)

In recent years, Deep Neural Networks (DNNs) have achieved tremendous success for diverse problems such as classification and decision making. Efficient support for DNNs on CPUs, GPUs and accelerators has become a prolific area of research, resulting in a plethora of techniques for energy-efficient DNN inference. However, previous proposals focus on a single execution of a DNN. Popular applications, such as speech recognition or video classification, require multiple back-to-back executions of a DNN to process a sequence of inputs (e.g., audio frames, images). In this paper, we show that consecutive inputs exhibit a high degree of similarity, causing the inputs/outputs of the different layers to be extremely similar for successive frames of speech or images of a video. Based on this observation, we propose a technique to reuse some results of the previous execution, instead of computing the entire DNN. Computations related to inputs with negligible changes can be avoided with minor impact on accuracy, saving a large percentage of computations and memory accesses. We propose an implementation of our reuse-based inference scheme on top of a state-of-the-art DNN accelerator. Results show that, on average, more than 60% of the inputs of any neural network layer tested exhibit negligible changes with respect to the previous execution. Avoiding the memory accesses and computations for these inputs results in 63% energy savings on average., Peer Reviewed, Postprint (published version)

The International Symposium on Computer Architecture (ISCA) recognizes every year the most influential paper published in this conference 15 years earlier, based on its impact on research, development, products or ideas. This award is sponsored by the IEEEComputer Society Technical Committee on Computer Architecture (IEEE-CS TCCA) and the ACM Special Interest Group on Computer Architecture (ACM SIGARCH). In this year’s edition, the candidate papers were those papers published in ISCA 2003 proceedings.The selection process was chaired by Antonio González. Candidate papers for the award were selected by the current year’s ISCA Pro-gram Committee. The final award selection was made by the Award Chair (Antonio González), the IEEE-CS TCCA Chair (Lieven Eeckhout) and the ACM SIGARCH Chair (Sarita Adve). The award includes an honorarium for the authors and a certificate.The 2018 award was presented to “Temperature-Aware Microarchitecture” by Kevin Skadron, Mircea R. Stan, Wei Huang, Sivakumar Velusamy, Karthik Sankaranarayanan and DavidTarjan., Peer Reviewed, Postprint (author's final draft)