The Simulation Engine - SimEng¶
SimEng is a framework for building modern, cycle-accurate processor simulators. Its goals are to be:
Fast, typically 4-5X faster than gem5
Easy to use and modify to model desired microarchitecture configurations. New cores can be configured in just a few hours
Scalable, from simple scalar microarchitectures up to the most sophisticated, superscalar, out-of-order designs
Capable of supporting a wide range of instruction set architectures (ISAs), starting with AArch64 and RISC-V, but eventually including x86, POWER, etc.
Accurate, aiming for simulated cycle times being within 5-10% of real hardware
Open source, with a permissive license to enable collaboration across academia and industry
SimEng places an emphasis on performance and ease of use, whilst maintaining a clean, modern, simple and well-documented code base. For example, the current out-of-order (OoO) model is implemented in around 11,000 lines of simple C++, with another 18,000 and 5,000 lines or so implementing the specifics of the AArch64 and RISC-V ISAs respectively, and around 41,000 lines of code in the accompanying test suite. SimEng should be simple to read and understand, making it ideal to modify to your requirements and include it in your projects.
Features¶
Currently, SimEng targets the Armv9.2-a ISA with support for the SVE, SVE2, and SME extensions as well as RISC-V rv64imafdc. SimEng has the ability to model up to out-of-order, superscalar, single-core processors, and to emulate a subset of Linux system-calls. It supports statically compiled C and Fortran binaries that run on real hardware, with additional support for single-threaded OpenMP binaries too. Internally, SimEng currently models memory as an infinite L1 cache, i.e. it assumes that all loads and stores hit the L1 cache. However, we have a tested integration with the Structural Simulation Toolkit (SST) allowing for a full memory model to be simulated; more information can be found in the SST Integration section.
The main component provided by the simulator is a discrete processor core model, shown in diagrammatic form below. This model accepts a clock signal and supports a memory access interface. A single YAML format configuration file can be passed to the simulation to specify models of existing microarchitectures, such as Marvell’s ThunderX2 or Fujitsu’s A64fx, or to model hypothetical core designs.
A future release of SimEng will support multi-core simulation by integrating with the SST.
Talks and presentations¶
SimEng was first presented by Professor Simon McIntosh-Smith at the 2019 Workshop on Modeling & Simulation of Systems and Applications (ModSim):
ModSim 2019 -
Enabling Processor Design Space Exploration with SimEng
Additionally, other works concerning SimEng and its use can be found below:
Second International workshop on RISC-V for HPC at SC2023 - An Empirical Comparison of the RISC-V and AArch64 Instruction Sets
ModSim 2023 -
Leveraging Arm's Scalable Matrix Extension to Accelerate Matrix Multiplication Kernels
PMBS 2022 - An Initial Evaluation of Arm’s Scalable Matrix Extension
ModSim 2022 -
A design space exploration for optimal vector unit composition
Release¶
This is SimEng’s eighth release, and should be considered beta level software (version 0.9.6). We expect you to find issues, primarily in unimplemented instructions or unimplemented system calls. Please let us know when you hit these, either by submitting a pull request (PR), or by filing an issue on the Github repo. You can find all the code and associated test suites for SimEng in the GitHub repository. The file RELEASE_NOTES.txt, found in the root of the project, explains the status of the project and includes other relevant information from the SimEng development team.
SimEng is released under the same license as LLVM, the permissive Apache 2.0 license. We are passionate about enabling experimentation with computer architectures, and want users and developers in academic and industry to have complete freedom to use SimEng anyway they wish, including using it in commercial settings.
External project usage¶
While we have tried to minimise SimEng’s dependencies to keep it as simple as possible, it does make use of a small number of libraries and frameworks to provide crucial capabilities:
Capstone disassembly engine - Provides instruction decoding for AArch64, RISC-V, x86 and other important ISAs
Rapid YAML - Parsing YAML configuration files
GoogleTest - Framework for the test suites
LLVM - Generation of binaries for use in the regression test suite
Contributors¶
Major contributors to SimEng to date include:
Project leader: Simon McIntosh-Smith
Current development team:
Jack Jones (lead developer)
Finn Wilkinson
Rahat Muneeb
Dan Weaver
Alex Cockrean
Joseph Moore
Original SimEng design and implementation:
Hal Jones
James Price
Additional Contributors:
Ainsley Rutterford
Andrei Poenaru
Harry Waugh
Mutalib Mohammed
Seunghun Lee
Tom Hepworth
Tom Lin
Will Robinson
Funding¶
The SimEng development team is grateful for the funding which has made this project possible, which to date has been from the UKRI/EPSRC ASiMoV project (Advanced Simulation and Modelling of Virtual systems), number EP/S005072/1, and from Arm via the Arm Centre of Excellence in HPC at the University of Bristol.