DSWP - Decoupled Software Pipelining

This is a class project for class EECS 583. We reimplemented the DSWP-plus repository (https://github.com/djsutherland/dswp-plus) with LLVM version 15.0.6 and improve its functionality. This repository can divide loops into multithreading and potentially speed up.

In this repository, we complete the following task,

• We fix several logic bugs including

  • not correctly replacing the uses of registers obtained from consume
  • copy only part of the codes between different threads instead of all the instructions
  • keep branch instructions

• We also allow multithreading instead of 2 threading in the original repository

  • You can change the number of threads by changing the DSWP header file (DSWP.h) (note that you need to make clean)

• We successfully compile and optimize the demo including

  • simple.c
  • problem1.c

Dependencies:

- LLVM 15.0.6

How to use it:

cd Example/
sh try.sh

Parameters:

• You can change the MAX_THREAD in the DSWP.h file. The minimum value is 2. When you change the maximum thread allowed, you need to do a remake.
• In the first line of Examples/problem1.sh, you are able to try running the file you like by clang -O0 -c -fPIE yourfile.c -o obj.o -emit-llvm.

DSWP structures:

• To view the structures of DSWP, there are 4 files in the Example folder:
  • dag outlines the SCCs and their dependencies
  • showgraph shows the instructions
  • partition presents how SCCs are organized into each partition/threads. This is determined by the heuristic based on the latencies of each SCC. If you see one empty partition, reduce MAX_THREAD
  • out.o.ll is a human readable file that displays the whole structures in each thread.

Contributing

For major changes, please open an issue first to discuss what you would like to change.

Please make sure to update tests as appropriate.

License

MIT