yingj4 / audio_pipeline_llvm-tracer

CS 598 JT Course Project

This is the project for CS 598 JT in Fall 2020. It aims to analyze the power-performance tradeoff of the audio component in ILLIXR. This project is based on ILLIXR, LLVM-Tracer, and Aladdin.

Source Code

The output files generated from LLVM-Tracer are those end with llvm and s. The files in ./audio_pipeline/llvm-tracer-label-1/ and ./audio_pipeline/llvm-tracer-no-label/ are the stale ones we generated when doing this project.
The binary executable is also an output from the LLVM-Tracer. It is ./audio_pipeline/main-single.

Experiment Steps

The binary executable requires inputs for the number of input blocks and the operating mode. The operating mode includes encode and decode. For example, if you want to run audio encoding with 500 blocks, you should run

./audio_pipeline/main-single 500 encode

If your third input argument is not encode, it will do audio decoding automatically. If you do not have a third argument, it will perform encoding and then decoding.
The output of the above command is dynamic_trace.gz. This trace file should be further fed into Aladdin with configuration parameters. Then you should be able to see the power and timing values.
Since the function performance for each block is the same from Aladdin, we recommend to use small block sizes (e.g. 1 ~ 20). The corresponding trace sizes will also be small. If the trace size is very large (e.g. 1GB), the computer memory may be filled when Aladdin is generating the dynamic data dependence graph (DDDG).

Configuration Setup

The configuration files for Aladdin can be found in ./audio_pipeline/loopLabels.md and ./audio_pipeline/arraySizes.md.
In ./audio_pipeline/loopLabels.md, we give the loop informations on which function the loop belongs to and whether the loop should be unrolled or pipelined.
In ./audio_pipeline/arraySizes.md, we give the total sizes and the wordsize of the arrays. These can be used to do partition in the Aladdin configuration.