NOP: A Strong Baseline for LoongArch Out-of-order Processor
[ Video ]
Authors: Mingdao Liu, Huan-ang Gao, Bowen Wang and Jiacheng Hua.
本项目是第七届“龙芯杯”全国大学生计算机系统能力培养大赛 (NSCSCC 2023) 的参赛作品。项目成功地开发了一款基于龙芯架构 32 位精简版 (LoongArch32-Reduced, LA32R) 指令集的 CPU，命名为 NOP。NOP 属于乱序多发射微架构，整体基于 Tomasulo 动态调度算法的思路 + 重排序缓存实现，并实现了分支预测、指令/数据缓存、数据旁路、推测唤醒等特性。从功能的角度上，NOP 实现了基础的算术指令、分支指令、访存指令，支持精确异常处理、虚实地址转换与 LA32R 指令集规定的各类中断，通过使用 PMON 引导程序，NOP 可以稳定地启动Linux 操作系统。从性能上看，NOP 作为一款乱序多发射处理器，在大赛提供的 FPGA 实验平台与性能测试程序 / SoC 上达到了 107.69 MHz 的主频与 1.02 的 IPC。相比于基线 openLA500 处理器，NOP 核的整体加速比达到了 3.00，IPC 加速比达到了 1.402.
Usage
Java development kit (JDK) is required to run the project. We use sbt as our build tool. To build the project, you need to install sbt first. Then, you can run the following command to build the project:
sbt run
Then you can see the build result in build directory. Simply copy mycpu_top.v to your chiplab FPGA project and you can simulate / run it on FPGA.
Project Structure
There are several subfolders in the project:
.ci-scripts: Scripts used in CI.docs: Documentation of the project.src: Source code of the project.xilinx_ip: Xilinx IP cores used in the project.
In src folder, there are several subfolders:
blackbox: Blackbox modules used in the project. (e.g. Multiplier, xpm_memory)builder: Taken from VexRiscv project. Contains the definition of the class of Stage, Plugin and Pipeline.constants: Contains the definition of constants used in the project, mainly for the LA32R ISA.debug: Contains the definition of debug module. (Not used. We use difftest signals instead.)peripheral: Contains the definition of peripheral modules. (e.g. AxiBuffer, AxiCrossbar)pipeline: Contains the definition of pipeline modules.utils: Contains the definition of utility modules.
The src/pipeline folder is the heart of our NOP processor’s pipeline structure. We dissect the pipeline of NOP CPU into various stages and components, implementing each as Scala modules.
-
core: This folder houses the central architectural files for the NOP-Core pipeline. Modules here handle key tasks like bypassing network operations (BypassNetworkPlugin.scala), managing commit operations and traits (CommitPlugin.scala, CommitTraits.scala), managing exceptions (ExceptionMuxPlugin.scala), and operating on the Reorder Buffer (ROBFIFOPlugin.scala, ReorderBufferModel.scala), among others.
-
decode: The decode phase of a CPU pipeline is where instructions fetched from memory are translated into actions the CPU can understand. This directory appears to deal with decoding micro-operations (DecodeMicroOP.scala), the actual decoding pipeline (DecodePipeline.scala), signals for decoding (DecodeSignals.scala), and renaming of registers (RenameModel.scala, RenamePlugin.scala).
-
exe: Short for “execute,” this directory concerns itself with the CPU’s execution stage. Files here manage ALU (Arithmetic Logic Unit) operations (ALU.scala), branch unit operations (BRU.scala), comparisons (Comparator.scala), integer executions (IntExecutePlugin.scala), and multiplication and division tasks (MulDivExecutePlugin.scala).
-
fetch: The fetch phase is the first step in a CPU’s operation cycle where instructions are retrieved from memory. Modules in this directory cater to functionalities like branch prediction (BranchPredictModel.scala), buffering of fetched data (FetchBufferPlugin.scala), and instruction caching (ICachePlugin.scala).
-
mem: As the name suggests, this folder focuses on memory operations within the CPU pipeline. This includes address generation (AddressGenerationPlugin.scala), managing data cache (DCachePlugin.scala), and ensuring the correct retrieval of data and the in-order commit for store / uncached load instructions with modules like LoadPostProcessPlugin.scala and StoreBufferPlugin.scala.
-
privilege: This directory seems to center around privileged operations and states of the CPU. It features modules for Control Status Registers (CSRPlugin.scala), exception handling (ExceptionHandlerPlugin.scala), interrupt handling (InterruptHandlerPlugin.scala), and the Memory Management Unit (MMUPlugin.scala).
Supplementary Resources
Also see our:
Acknowledgement
We are grateful to the following open-source projects:
评论