To meet the extreme compute demands for deep learning across commercial and scientific applications, dataflow accelerators are becoming increasingly popular. While these "domain-specific" accelerators are not fully programmable like CPUs and GPUs, they retain varying levels of flexibility with respect to data orchestration, i.e., dataflow and tiling optimizations to enhance efficiency. There are several challenges when designing new algorithms and mapping approaches to execute the algorithms for a target problem on new hardware. Previous works have addressed these challenges individually. To address this challenge as a whole, in this work, we present a HW-SW co-design ecosystem for spatial accelerators called Union within the popular MLIR compiler infrastructure. Our framework allows exploring different algorithms and their mappings on several accelerator cost models. Union also includes a plug-and-play library of accelerator cost models and mappers which can easily be extended. The algorithms and accelerator cost models are connected via a novel mapping abstraction that captures the map space of spatial accelerators which can be systematically pruned based on constraints from the hardware, workload, and mapper. We demonstrate the value of Union for the community with several case studies which examine offloading different tensor operations(CONV/GEMM/Tensor Contraction) on diverse accelerator architectures using different mapping schemes.
相關內容
ACM/IEEE第23屆模型驅動工程語言和系統國際會議，是模型驅動軟件和系統工程的首要會議系列，由ACM-SIGSOFT和IEEE-TCSE支持組織。自1998年以來，模型涵蓋了建模的各個方面，從語言和方法到工具和應用程序。模特的參加者來自不同的背景，包括研究人員、學者、工程師和工業專業人士。MODELS 2019是一個論壇，參與者可以圍繞建模和模型驅動的軟件和系統交流前沿研究成果和創新實踐經驗。今年的版本將為建模社區提供進一步推進建模基礎的機會，并在網絡物理系統、嵌入式系統、社會技術系統、云計算、大數據、機器學習、安全、開源等新興領域提出建模的創新應用以及可持續性。
官網鏈接： ·
Performer
·
Neural Computation
·
卷積神經網絡
·
Neural Networks
·
Sparse convolutional neural networks (CNNs) have gained significant traction over the past few years as sparse CNNs can drastically decrease the model size and computations, if exploited befittingly, as compared to their dense counterparts. Sparse CNNs often introduce variations in the layer shapes and sizes, which can prevent dense accelerators from performing well on sparse CNN models. Recently proposed sparse accelerators like SCNN, Eyeriss v2, and SparTen, actively exploit the two-sided or full sparsity, that is, sparsity in both weights and activations, for performance gains. These accelerators, however, either have inefficient micro-architecture, which limits their performance, have no support for non-unit stride convolutions and fully-connected (FC) layers, or suffer massively from systematic load imbalance. To circumvent these issues and support both sparse and dense models, we propose Phantom, a multi-threaded, dynamic, and flexible neural computational core. Phantom uses sparse binary mask representation to actively lookahead into sparse computations, and dynamically schedule its computational threads to maximize the thread utilization and throughput. We also generate a two-dimensional (2D) mesh architecture of Phantom neural computational cores, which we refer to as Phantom-2D accelerator, and propose a novel dataflow that supports all layers of a CNN, including unit and non-unit stride convolutions, and FC layers. In addition, Phantom-2D uses a two-level load balancing strategy to minimize the computational idling, thereby, further improving the hardware utilization. To show support for different types of layers, we evaluate the performance of the Phantom architecture on VGG16 and MobileNet. Our simulations show that the Phantom-2D accelerator attains a performance gain of 12x, 4.1x, 1.98x, and 2.36x, over dense architectures, SCNN, SparTen, and Eyeriss v2, respectively.
The use of large-scale machine learning methods is becoming ubiquitous in many applications ranging from business intelligence to self-driving cars. These methods require a complex computation pipeline consisting of various types of operations, e.g., relational operations for pre-processing or post-processing the dataset, and matrix operations for core model computations. Many existing systems focus on efficiently processing matrix-only operations, and assume that the inputs to the relational operators are already pre-computed and are materialized as intermediate matrices. However, the input to a relational operator may be complex in machine learning pipelines, and may involve various combinations of matrix operators. Hence, it is critical to realize scalable and efficient relational query processors that directly operate on big matrix data. This paper presents new efficient and scalable relational query processing techniques on big matrix data for in-memory distributed clusters. The proposed techniques leverage algebraic transformation rules to rewrite query execution plans into ones with lower computation costs. A distributed query plan optimizer exploits the sparsity-inducing property of merge functions as well as Bloom join strategies for efficiently evaluating various flavors of the join operation. Furthermore, optimized partitioning schemes for the input matrices are developed to facilitate the performance of join operations based on a cost model that minimizes the communication overhead.The proposed relational query processing techniques are prototyped in Apache Spark. Experiments on both real and synthetic data demonstrate that the proposed techniques achieve up to two orders of magnitude performance improvement over state-of-the-art systems on a wide range of applications.
Python is rapidly becoming the lingua franca of machine learning and scientific computing. With the broad use of frameworks such as Numpy, SciPy, and TensorFlow, scientific computing and machine learning are seeing a productivity boost on systems without a requisite loss in performance. While high-performance libraries often provide adequate performance within a node, distributed computing is required to scale Python across nodes and make it genuinely competitive in large-scale high-performance computing. Many frameworks, such as Charm4Py, DaCe, Dask, Legate Numpy, mpi4py, and Ray, scale Python across nodes. However, little is known about these frameworks' relative strengths and weaknesses, leaving practitioners and scientists without enough information about which frameworks are suitable for their requirements. In this paper, we seek to narrow this knowledge gap by studying the relative performance of two such frameworks: Charm4Py and mpi4py. We perform a comparative performance analysis of Charm4Py and mpi4py using CPU and GPU-based microbenchmarks other representative mini-apps for scientific computing.
While discrete-event simulators are essential tools for architecture research, design, and development, their practicality is limited by an extremely long time-to-solution for realistic applications under investigation. This work describes a concerted effort, where machine learning (ML) is used to accelerate discrete-event simulation. First, an ML-based instruction latency prediction framework that accounts for both static instruction properties and dynamic processor states is constructed. Then, a GPU-accelerated parallel simulator is implemented based on the proposed instruction latency predictor, and its simulation accuracy and throughput are validated and evaluated against a state-of-the-art simulator. Leveraging modern GPUs, the ML-based simulator outperforms traditional simulators significantly.
Bidirectional reflectance distribution functions (BRDFs) are pervasively used in computer graphics to produce realistic physically-based appearance. In recent years, several works explored using neural networks to represent BRDFs, taking advantage of neural networks' high compression rate and their ability to fit highly complex functions. However, once represented, the BRDFs will be fixed and therefore lack flexibility to take part in follow-up operations. In this paper, we present a form of "Neural BRDF algebra", and focus on both representation and operations of BRDFs at the same time. We propose a representation neural network to compress BRDFs into latent vectors, which is able to represent BRDFs accurately. We further propose several operations that can be applied solely in the latent space, such as layering and interpolation. Spatial variation is straightforward to achieve by using textures of latent vectors. Furthermore, our representation can be efficiently evaluated and sampled, providing a competitive solution to more expensive Monte Carlo layering approaches.
In this essay, I present the advantages and, I dare say, the beauty of programming in a language with set-theoretic types, that is, types that include union, intersection, and negation type connectives. I show by several examples how set-theoretic types are necessary to type some common programming patterns, but also how they play a key role in typing several language constructs-from branching and pattern matching to function overloading and type-cases-very precisely. I start by presenting the theory of types known as semantic subtyping and extend it to include polymorphic types. Next, I discuss the design of languages that use these types. I start by defining a theoretical framework that covers all the examples given in the first part of the presentation. Since the system of the framework cannot be effectively implemented, I then describe three effective restrictions of this system: (i) a polymorphic language with explicitly-typed functions, (ii) an implicitly-typed polymorphic language{\`a} la Hindley-Milner, and (iii) a monomorphic language that, by implementing classic union-elimination, precisely reconstructs intersection types for functions and implements a very general form of occurrence typing. I conclude the presentation with a short overview of other aspects of these languages, such as pattern matching, gradual typing, and denotational semantics.
Graph convolutional networks (GCNs) have been introduced to effectively process non-euclidean graph data. However, GCNs incur large amounts of irregularity in computation and memory access, which prevents efficient use of traditional neural network accelerators. Moreover, existing dedicated GCN accelerators demand high memory volumes and are difficult to implement onto resource limited edge devices. In this work, we propose LW-GCN, a lightweight FPGA-based accelerator with a software-hardware co-designed process to tackle irregularity in computation and memory access in GCN inference. LW-GCN decomposes the main GCN operations into sparse-dense matrix multiplication (SDMM) and dense matrix multiplication (DMM). We propose a novel compression format to balance workload across PEs and prevent data hazards. Moreover, we apply data quantization and workload tiling, and map both SDMM and DMM of GCN inference onto a uniform architecture on resource limited hardware. Evaluation on GCN and GraphSAGE are performed on Xilinx Kintex-7 FPGA with three popular datasets. Compared to existing CPU, GPU, and state-of-the-art FPGA-based accelerator, LW-GCN reduces latency by up to 60x, 12x and 1.7x and increases power efficiency by up to 912x., 511x and 3.87x, respectively. Furthermore, compared with NVIDIA's latest edge GPU Jetson Xavier NX, LW-GCN achieves speedup and energy savings of 32x and 84x, respectively.
We present a holistic design for GPU-accelerated computation in TrustZone TEE. Without pulling the complex GPU software stack into the TEE, we follow a simple approach: record the CPU/GPU interactions ahead of time, and replay the interactions in the TEE at run time. This paper addresses the approach's key missing piece -- the recording environment, which needs both strong security and access to diverse mobile GPUs. To this end, we present a novel architecture called CODY, in which a mobile device (which possesses the GPU hardware) and a trustworthy cloud service (which runs the GPU software) exercise the GPU hardware/software in a collaborative, distributed fashion. To overcome numerous network round trips and long delays, CODY contributes optimizations specific to mobile GPUs: register access deferral, speculation, and metastate-only synchronization. With these optimizations, recording a compute workload takes only tens of seconds, which is up to 95% less than a naive approach; replay incurs 25% lower delays compared to insecure, native execution.
The task of detecting 3D objects in point cloud has a pivotal role in many real-world applications. However, 3D object detection performance is behind that of 2D object detection due to the lack of powerful 3D feature extraction methods. In order to address this issue, we propose to build a 3D backbone network to learn rich 3D feature maps by using sparse 3D CNN operations for 3D object detection in point cloud. The 3D backbone network can inherently learn 3D features from almost raw data without compressing point cloud into multiple 2D images and generate rich feature maps for object detection. The sparse 3D CNN takes full advantages of the sparsity in the 3D point cloud to accelerate computation and save memory, which makes the 3D backbone network achievable. Empirical experiments are conducted on the KITTI benchmark and results show that the proposed method can achieve state-of-the-art performance for 3D object detection.
In this paper, we propose a simple and general framework for training very tiny CNNs for object detection. Due to limited representation ability, it is challenging to train very tiny networks for complicated tasks like detection. To the best of our knowledge, our method, called Quantization Mimic, is the first one focusing on very tiny networks. We utilize two types of acceleration methods: mimic and quantization. Mimic improves the performance of a student network by transfering knowledge from a teacher network. Quantization converts a full-precision network to a quantized one without large degradation of performance. If the teacher network is quantized, the search scope of the student network will be smaller. Using this feature of the quantization, we propose Quantization Mimic. It first quantizes the large network, then mimic a quantized small network. The quantization operation can help student network to better match the feature maps from teacher network. To evaluate our approach, we carry out experiments on various popular CNNs including VGG and Resnet, as well as different detection frameworks including Faster R-CNN and R-FCN. Experiments on Pascal VOC and WIDER FACE verify that our Quantization Mimic algorithm can be applied on various settings and outperforms state-of-the-art model acceleration methods given limited computing resouces.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191