Defining interfaces between hardware and software - Quality and performance

Alastair Reid
School of Computing Science
University of Glasgow
[pdf]

Glasgow, Scotland
March 2019

Abstract

One of the most important interfaces in a computer system is the interface between hardware and software. This interface is the contract between the hardware designer and the programmer that defines the functional behaviour of the hardware. This thesis examines two critical aspects of defining the hardware-software interface: quality and performance.

The first aspect is creating a high quality specification of the interface as conventionally defined in an instruction set architecture. The majority of this thesis is concerned with creating a specification that covers the full scope of the interface; that is applicable to all current implementations of the architecture; and that can be trusted to accurately describe the behaviour of implementations of the architecture. We describe the development of a formal specification of the two major types of Arm processors: A-class (for mobile devices such as phones and tablets) and M-class (for micro-controllers). These specifications are unparalleled in their scope, applicability and trustworthiness. This thesis identifies and illustrates what we consider the key ingredient in achieving this goal: creating a specification that is used by many different user groups. Supporting many different groups leads to improved quality as each group finds different problems in the specification; and, by providing value to each different group, it helps justify the considerable effort required to create a high quality specification of a major processor architecture. The work described in this thesis led to a step change in Arm's ability to use formal verification techniques to detect errors in their processors; enabled extensive testing of the specification against Arm's official architecture conformance suite; improved the quality of Arm's architecture conformance suite based on measuring the architectural coverage of the tests; supported earlier, faster development of architecture extensions by enabling animation of changes as they are being made; and enabled early detection of problems created from architecture extensions by performing formal validation of the specification against semi-structured natural language specifications. As far as we are aware, no other mainstream processor architecture has this capability. The formal specifications are included in Arm's publicly released architecture reference manuals and the A-class specification is also released in machine-readable form.

The second aspect is creating a high performance interface by defining the hardware-software interface of a software-defined radio subsystem using a programming language. That is, an interface that allows software to exploit the potential performance of the underlying hardware. While the hardware-software interface is normally defined in terms of machine code, peripheral control registers and memory maps, we define it using a programming language instead. This higher level interface provides the opportunity for compilers to hide some of the low-level differences between different systems from the programmer: a potentially very efficient way of providing a stable, portable interface without having to add hardware to provide portability between different hardware platforms. We describe the design and implementation of a set of extensions to the C programming language to support programming high performance, energy efficient, software defined radio systems. The language extensions enable the programmer to exploit the pipeline parallelism typically present in digital signal processing applications and to make efficient use of the asymmetric multiprocessor systems designed to support such applications. The extensions consist primarily of annotations that can be checked for consistency and that support annotation inference in order to reduce the number of annotations required. Reducing the number of annotations does not just save programmer effort, it also improves portability by reducing the number of annotations that need to be changed when porting an application from one platform to another. This work formed part of a project that developed a high-performance, energy-efficient, software defined radio capable of implementing the physical layers of the 4G cellphone standard (LTE), 802.11a WiFi and Digital Video Broadcast (DVB) with a power and silicon area budget that was competitive with a conventional custom ASIC solution.

The Arm architecture is the largest computer architecture by volume in the world. It behooves us to ensure that the interface it describes is appropriately defined.

BibTeX

@phdthesis{ReidPhDThesis2019 , abstract = {One of the most important interfaces in a computer system is the interface between hardware and software. This interface is the contract between the hardware designer and the programmer that defines the functional behaviour of the hardware. This thesis examines two critical aspects of defining the hardware-software interface: quality and performance.

The first aspect is creating a high quality specification of the interface as conventionally defined in an instruction set architecture. The majority of this thesis is concerned with creating a specification that covers the full scope of the interface; that is applicable to all current implementations of the architecture; and that can be trusted to accurately describe the behaviour of implementations of the architecture. We describe the development of a formal specification of the two major types of Arm processors: A-class (for mobile devices such as phones and tablets) and M-class (for micro-controllers). These specifications are unparalleled in their scope, applicability and trustworthiness. This thesis identifies and illustrates what we consider the key ingredient in achieving this goal: creating a specification that is used by many different user groups. Supporting many different groups leads to improved quality as each group finds different problems in the specification; and, by providing value to each different group, it helps justify the considerable effort required to create a high quality specification of a major processor architecture. The work described in this thesis led to a step change in Arm's ability to use formal verification techniques to detect errors in their processors; enabled extensive testing of the specification against Arm's official architecture conformance suite; improved the quality of Arm's architecture conformance suite based on measuring the architectural coverage of the tests; supported earlier, faster development of architecture extensions by enabling animation of changes as they are being made; and enabled early detection of problems created from architecture extensions by performing formal validation of the specification against semi-structured natural language specifications. As far as we are aware, no other mainstream processor architecture has this capability. The formal specifications are included in Arm's publicly released architecture reference manuals and the A-class specification is also released in machine-readable form.

The second aspect is creating a high performance interface by defining the hardware-software interface of a software-defined radio subsystem using a programming language. That is, an interface that allows software to exploit the potential performance of the underlying hardware. While the hardware-software interface is normally defined in terms of machine code, peripheral control registers and memory maps, we define it using a programming language instead. This higher level interface provides the opportunity for compilers to hide some of the low-level differences between different systems from the programmer: a potentially very efficient way of providing a stable, portable interface without having to add hardware to provide portability between different hardware platforms. We describe the design and implementation of a set of extensions to the C programming language to support programming high performance, energy efficient, software defined radio systems. The language extensions enable the programmer to exploit the pipeline parallelism typically present in digital signal processing applications and to make efficient use of the asymmetric multiprocessor systems designed to support such applications. The extensions consist primarily of annotations that can be checked for consistency and that support annotation inference in order to reduce the number of annotations required. Reducing the number of annotations does not just save programmer effort, it also improves portability by reducing the number of annotations that need to be changed when porting an application from one platform to another. This work formed part of a project that developed a high-performance, energy-efficient, software defined radio capable of implementing the physical layers of the 4G cellphone standard (LTE), 802.11a WiFi and Digital Video Broadcast (DVB) with a power and silicon area budget that was competitive with a conventional custom ASIC solution.

The Arm architecture is the largest computer architecture by volume in the world. It behooves us to ensure that the interface it describes is appropriately defined. } , affiliation = {School of Computing Science, University of Glasgow} , ar_file = {PhD_19} , ar_shortname = {PhD 19} , author = {Alastair Reid} , file = {2019reidphd.pdf} , location = {Glasgow, Scotland} , month = {March} , numpages = {161} , school = {School of Computing Science, University of Glasgow} , title = {{D}efining interfaces between hardware and software: {Q}uality and performance} , year = {2019} }