Error tolerant multimedia compression system

Unrestricted All existing digital system design approaches strive to provide error-free values at system outputs, which is achieved by using testing techniques, as well as defect and fault tolerant methods. In contrast, the focus of this thesis is on error tolerant systems, i.e., systems where certain types of errors at system outputs can be tolerated, provided their severities are below certain levels. The overall objective of our research in error tolerance is (i) to identify the inherent ability of some systems to tolerate errors, and (ii) to develop design and test approaches that exploit this ability.; Our target systems are multimedia applications, which represent a major workload on major hand-held devices such as cellular phones and laptops. Video coders (e.g., H.264/AVC and MPEG-4) and image coders (e.g., JPEG and JPEG2000) are the most complex part of multimedia applications. Within a typical video/image coder, we focus on motion estimation (ME) and linear transforms (e.g., discrete cosine transform) as those two consume a large percentage of resources. Achieving error tolerance in those two modules can lead to increased yield or lower power consumption.; In this thesis, we study two specific scenarios i) soft-error tolerance in matching metric computations within motion estimation, ii) hard-error tolerance in linear transforms (e.g., discrete cosine transform). While soft errors can also be introduced due to deep submicron (DSM) noise, we focus on voltage over scaling which introduces input-dependent errors. We show that soft errors within matching metric computation of motion estimation are tolerable to some extent. The tolerance to soft errors of the motion estimation module is exploited in a low power motion estimation system; i.e., the motion estimation module can operate (with someerror) at a lower voltage configuration, thus achieving power savings. We first explore one possible configuration which uses one voltage over scaled metric computation within ME. We then extend that work to a more general configuration using multiple low complexity metric computations (e.g., voltage over scaled and sub-sampled metrics) within ME.; Hard errors are introduced due to defects within hardware. By emulating the effect of those hard errors within linear transform hardware, we show that linear transforms have significant error tolerance. This error tolerance is exploited in order to achieve a higher yield rate by accepting systems with faults, while still operate with acceptable quality. To realize this, we introduce a systematic error tolerant testing method for this hardware.在无约束场景下，当前所有数字系统设计方法均致力于实现系统输出的无差错值，这一目标通过测试技术、缺陷容错与故障容错手段达成。与之相对，本论文的研究焦点为容错系统：即当系统输出端出现的特定类型错误的严重程度低于阈值时，可被容忍的系统。本研究在容错领域的总体目标有二：其一，甄别部分系统固有的错误容忍能力；其二，开发可利用该能力的设计与测试方法。 我们的目标系统为多媒体应用程序，此类应用是手机、笔记本电脑等主流手持设备的主要工作负载。视频编码器（如H.264/AVC与MPEG-4）与图像编码器（如JPEG及JPEG2000）是多媒体应用中复杂度最高的组成部分。在典型的视频/图像编码器中，我们聚焦于运动估计（Motion Estimation，ME）与线性变换，其中离散余弦变换（Discrete Cosine Transform）为典型代表，因为这两个模块占据了绝大部分系统资源。在这两个模块中实现容错，可提升芯片良率或降低功耗。 在本论文中，我们研究两类具体场景：其一为运动估计中匹配度量计算的软错误容错，其二为线性变换（如离散余弦变换）的硬错误容错。尽管软错误也可由深亚微米（Deep Submicron，DSM）噪声引发，但本研究聚焦于会引入输入相关错误的电压过度缩放（Voltage Over Scaling）。我们证实，运动估计中的匹配度量计算模块所产生的软错误，在一定程度上是可被容忍的。运动估计模块对软错误的容忍能力，可被应用于低功耗运动估计系统中：即该模块可在部分存在错误的状态下，以更低的电压配置运行，从而实现功耗节省。我们首先探索了一种在运动估计（ME）中采用单组电压过度缩放过的度量计算的可行配置；随后，我们将该研究拓展至更通用的配置：即在ME中采用多组低复杂度度量计算（如电压过度缩放与亚采样度量计算）。 硬错误由硬件内部的缺陷引发。通过在线性变换硬件中模拟此类硬错误的影响，我们证实线性变换具备显著的错误容忍能力。我们可利用该容错能力，通过接纳存在故障的系统来提升良率，同时仍能保证可接受的运行质量。为实现这一目标，我们针对该硬件提出了一种系统化的容错测试方法。