RISC/CISC纵横谈

在外行人看来,计算机好象很聪明,其实,它那“聪明”,归根到底,是人(特别是内行的计算机专家)赋予它的.它每做一件事,哪怕很简单的微小动作,都有赖于人的指示.人所用以指示计算机做各种各样事的指令,集合起来,就是该台计算机的指令集.计算机只认这些指令.为解决某一特定问题所要用到的所有指令一条一条地按顺序排列起来,便是所谓“程序”(或“软件”).不同的计算机有不同的指令集,不同的计算机,指令条数多少不一.在计算机研制的早期,人们以为,计算机所能辨识和执行的指令越多,功能便越强,编制程序便越方便,编出的程序功能便越强.因此,计算机的指令集越来越大,条数越来越多,从最初的二、三十条到几百上千条.而     

基于宏指令集的专用汇编器的研究与实现

针对将复杂算法转化为硬件电路难度很大的问题,设计一种专用汇编器,可将由宏指令组成的汇编语言自动转化为二进制代码表示的机器语言,并根据宏指令集计算机理论的系统结构,将二进制代码组织成VHDL语言描述的状态机输出,便于进一步的硬件优化设计。性能测试结果证明,该汇编器支持从汇编语言到硬件电路设计语言的自动转化,可较大地提高用硬件电路实现复杂算法的研发效率。     

基于集束式整数线性规划模型的专用指令集自动定制

提出集束式整数线性规划形式化模型,利用指令间的功能依赖性解决专用指令集处理器中指令集自动定制的指数性空间问题.在此基础上,针对其前端和后端分别提出了相应的指令定制实现策略.实验结果表明,该指令定制方法可以有效地实现专用指令集的自动设计,并使最终处理器的运算性能得到优化.     

虚拟指令集的构建及翻译技术研究

利用虚拟指令作为中间语言来构建可重用指令集模拟器是解决模拟器可重用性的重要技术.介绍了可重用指令集模拟器的工作原理,提出了虚拟指令的构建原则和方法,描述了对汇编指令的语义规则,最后举例说明如何从汇编指令的语义描述规则出发,生成与目标指令语义等价的虚拟指令.     

基于RPU的TTI程序设计和分析方法

实时嵌入式系统不仅要保证计算结果的逻辑正确性,还要确保与外界交互的时序正确性,所以底层程序要能精确表达上层模型中的时间行为. TTI指令集（time-triggered instruction set）的提出尝试解决计算机指令集体系结构层次缺少时间语义的问题,并且基于TTI指令集实现的实时处理单元（real-time processing unit,RPU）证明了TTI指令集的可行性和有效性.但是目前的工作缺少对于TTI程序设计和分析方法的研究.所以,基于TTI指令集和RPU,提出了TTI指令集可以表达的4种时间语义,给出了TTI程序的设计范式.并且构建了TTI程序时间行为的表示方法——TFG+,TFG+是对TFG的扩展,TFG+区分了TTI程序中时间语义指令和普通代码段,可以表示TTI程序的控制流信息、用户规定的时间行为和TTI程序平台相关的时间属性.最后,提出了TTI程序的时间分析方法以及时间安全性检查方法,为TTI程序的设计和部署提供了依据.     

可逆乘除法指令的设计与仿真

为解决可逆指令集PISA缺少乘法指令和除法指令的问题，提出一种通过可逆子过程实现乘除法指令的方法。研究PISA指令集中的现有指令，总结并概括使用可逆指令进行编程的方法和原则，分析设计可逆乘除法指令须遵守的约束，给出可逆乘除法指令的格式，分别使用可逆的原码一位乘和恢复余数法实现乘法指令和除法指令。在可逆指令集仿真平台上对实现的乘除法指令进行测试，测试结果表明，乘法指令和除法指令均逻辑可逆，在满足特定约束条件时互为逆指令。     

利用CSP求解极小碰集的方法

基于模型诊断是人工智能领域中具有挑战性的问题,包含了很多人工智能中的关键问题,其研究对整个人工智能领域起着重要推动作用。在基于模型诊断中,候选诊断结果通常由所有极小冲突集对应的所有极小碰集所描述,求出所有极小碰集是其核心问题之一。提出一种将极小碰集问题转换为约束满足问题的方法,该方法调用成熟的CSP求解器进行求解,扩展了约束可满足问题的应用领域。首次提出hard‐冲突集和sof t‐冲突集的概念,并给出利用所提的方法分别求解具有一些特征的极小碰集：小于固定长度、不含特定元素及包含hard‐冲突集和sof t‐冲突集。实验结果表明,提出的方法易于实现、扩展性强,对于特定类型极小碰集问题的求解效率较高。     

基于词序嵌入的二进制基本块相似性检测

神经机器翻译技术能够自动翻译多种语言的语义信息, 已被应用于跨指令集架构的二进制代码相似性检测, 并取得了较好的效果. 将汇编指令序列当作文本序列处理时, 指令顺序关系很重要. 进行二进制基本块级别相似性检测时, 神经网络使用位置嵌入来对指令位置进行建模. 然而, 这种位置嵌入未能捕获指令位置之间的邻接、优先等关系. 针对该问题, 本文使用指令位置的连续函数来建模汇编指令的全局绝对位置和顺序关系, 实现对词序嵌入的泛化. 首先使用Transformer训练源指令集架构编码器; 然后使用三元组损失训练目标指令集架构编码器, 并微调源指令集架构编码器; 最后使用嵌入向量之间欧氏距离的映射表示基本块之间的相似程度. 在公开数据集MISA上的实验表明, P@1评价指标达到69.5%, 比对比方法MIRROR提升了4.6%.     

RISC-V指令集架构研究综述

指令集作为软硬件之间的接口规范,是信息技术生态的起始原点.RISC-V是计算机体系结构走向开放的必然产物,其出现为系统研究领域带来了新的思路,即系统软件问题的研究深度可以进一步向下延伸至指令集架构,从而拓展甚至颠覆软件领域的“全栈”概念.对近年来RISC-V指令集架构相关的研究成果进行了综述.首先介绍了RISC-V指令集的发展现状,指出开展RISC-V研究应重点关注的指令集范围.然后分析了RISC-V处理器设计要点和适用范围.同时,围绕RISC-V系统设计问题,从指令集、功能实现、性能提升、安全策略这4个方面,论述了RISC-V处理器基本的研究思路,并分析了近年来的研究成果.最后借助具体的研究案例,阐述了RISC-V在领域应用的价值,并展望了RISC-V架构后续研究的可能切入点和未来发展方向.     

说说人工智能的那些事儿

人工智能包括“人类智能”与“人工打造”两部分,前者为强人工智能,后者为弱人工智能。人类智能有“大脑思维”与“智力行为”两种形式;与之对应,“人工打造”也有大脑思维仿真与人类智力行为机器替代两种方式。如今,无论是大脑思维仿真,还是智力行为替代,都是在通用计算机智力平台上或以嵌入式系统智力嵌入方式实现,因此,人工智能体现了现代计算机的智力革命。此外,无论是大脑思维,还是智力行为,都是人类知识基础上的能力表现,有人提出“人工智能是关于知识的学科”,因此,也有必要从知识学原理的另类视角来审视人工智能。     

基于LLVM的RISC-V自定义扩展指令支持方法

RISC-V指令集架构具有模块化、可扩展等特性.基于RISC-V架构的处理器,可以在整数指令集的基础之上,有选择地支持官方标准指令集扩展,以及非标准的用户自定义指令集扩展.这也意味着,对于每个新增的自定义扩展指令集,用户都需要自己在编译工具链中实现相应支持.通过分析LLVM编译框架,研究RISC-V自定义扩展指令支持的通用方法,并以玄铁C910自定义指令集为例进行实现和验证.为基于LLVM基础架构的RISC-V自定义指令集扩展研究与实现提供借鉴.     

Partially defined computer instructions and guards

We here extend our earlier work on the theory of computer instructions to consider instructions which are only partially defined. For every such instruction, we assume that it is defined whenever a certain Boolean expression is true; we refer to such a Boolean expression as a guard, following Dijkstra. This is a special case of a more general function on the set of states of a computer, representing an expression in a programming language. Many constructs for instructions now generalize to partially defined instructions; in particular, we define the notion of conditional input and output regions, as well as the relevant region of a more general expression. Fundamental theorems about instructions generalize to theorems about guards and about partially defined instructions. We also define the parallel execution of such instructions, which is useful in validating a generalized instruction commutativity criterion.     

基于C/S架构的教学系统中智能技术的应用研究

利用人工智能原理及认知理论建立ICAI模型,构建实用的教学系统是人工智能的一个重要应用.本文就ICAI的特点、教学系统的整体结构、功能模块及其关键技术进行了详细的分析,并对教学过程进行了定量分析.     

Have we witnessed a real-life Turing Test?

Did Deep Blue ace the Turing Test? Did it do much more? It seems that the IBM creation not only beat the reigning chess World Champion Gary Kasparov, but also took a large step, in some people's eyes, toward true artificial intelligence. For AI professionals, a computer defeating a human in chess is probably neither surprising nor really significant. After all, they contend, chess can be described in terms of a nondeterministic alternating Turing machine. Despite the enormous number of possible positions and available moves, the task does not present a challenging theoretical AI problem of NP completeness. There are many well-developed AI strategies that limit the search for the best move to an analysis of the most promising positions. Therefore, the progress in logical and numerical methods of AI and a computer's computational speed and available memory made the computer's victory inevitable. Deep Blue's victory, then, was attributable to its ability to analyze 200 million positions per second and a refined algorithm that accounted for positional-in addition to material-advantage. In summary, most AI professionals conclude that the computer won by brute force, rather than a sophisticated or original strategy. What most AI experts have overlooked, though, is another aspect of the match, which may signify a milestone in the history of computer science: for the first time, a computer seems to have passed the Turing Test     

动态二进制翻译中不对界问题的处理

复杂指令集计算机体系结构向精简指令集计算机体系结构的动态二进制翻译过程中经常出现地址不对界的问题。本文以I386到Alpha平台的动态二进制翻译为例,研究了内存映射时的不对界和数据存取时的不对界问题,提出了一种改进的内存映射方法以及在中间表示层处理不对界地址访存问题的方案,有效地解决了此类问题。经实验验证,该方法正确并有较高效率。     

Challenges and opportunities: from big data to knowledge in AI 2.0

In this paper, we review recent emerging theoretical and technological advances of artificial intelligence (AI) in the big data settings. We conclude that integrating data-driven machine learning with human knowledge (common priors or implicit intuitions) can effectively lead to explainable, robust, and general AI, as follows: from shallow computation to deep neural reasoning; from merely data-driven model to data-driven with structured logic rules models; from task-oriented (domain-specific) intelligence (adherence to explicit instructions) to artificial general intelligence in a general context (the capability to learn from experience). Motivated by such endeavors, the next generation of AI, namely AI 2.0, is positioned to reinvent computing itself, to transform big data into structured knowledge, and to enable better decision-making for our society.     

师范院校计算机人工智能双语教学的方案设计

双语教学作为一个新的教学模式在高校的教学中还处于探索、研究阶段。本文结合人工智能课程的双语教学实践,详细分析了师范院校双语教学的特点,提出了师范院校双语教学遵循的原则,最后设计出了人工智能双语教学的方案。     

Deus est machina [artificial intelligence]

What happens if the artificial intelligence community, in its quest to build intelligent systems, succeeds too well and creates an AI whose intelligence exceeds the threshold marked out by our own? Up to now, it is humans who develop the software and hardware and who drive all progress in capability. After crossing the threshold, however, the AI itself will rapidly augment its own capabilities. What's the intuition here? Although we use technology to help us conceptualize, design, and build today's computers and software there's no doubt that we remain in the driver's seat. But imagine the software design process reaching a level of complexity at which human designers exert only executive oversight.     

Hubert L. Dreyfus’s Critique of Classical AI and its Rationalist Assumptions

This paper deals with the rationalist assumptions behind researches of artificial intelligence (AI) on the basis of Hubert Dreyfus’s critique. Dreyfus is a leading American philosopher known for his rigorous critique on the underlying assumptions of the field of artificial intelligence. Artificial intelligence specialists, especially those whose view is commonly dubbed as “classical AI,” assume that creating a thinking machine like the human brain is not a too far away project because they believe that human intelligence works on the basis of formalized rules of logic. In contradistinction to classical AI specialists, Dreyfus contends that it is impossible to create intelligent computer programs analogous to the human brain because the workings of human intelligence is entirely different from that of computing machines. For Dreyfus, the human mind functions intuitively and not formally. Following Dreyfus, this paper aims to pinpointing the major flaws classical AI suffers from. The author of this paper believes that pinpointing these flaws would inform inquiries on and about artificial intelligence. Over and beyond this, this paper contributes something indisputably original. It strongly argues that classical AI research programs have, though inadvertently, falsified an entire epistemological enterprise of the rationalists not in theory as philosophers do but in practice. When AI workers were trying hard in order to produce a machine that can think like human minds, they have in a way been testing—and testing it up to the last point—the rationalist assumption that the workings of the human mind depend on logical rules. Result: No computers actually function like the human mind. Reason: the human mind does not depend on the formal or logical rules ascribed to computers. Thus, symbolic AI research has falsified the rationalist assumption that ‘the human mind reaches certainty by functioning formally’ by virtue of its failure to create a thinking machine.