A collection of three programs developed to demonstrate the usefulness of Thread-Data Remapping for inverting a matrix.

cpu-inverse [input matrix] [inverse matrix]
inverse     [input matrix] [inverse matrix]
tdr-inverse [input matrix] [inverse matrix]

Matrices may either be in comma separated CSV format or Matrix Market Coordinate Format.

The code in developed in service of a masters research project. The related documents can be found here.

There are three programs included in this repository:

• cpu-inverse: Inverts a matrix on-CPU.
• inverse: Inverts a matrix on-GPU utilizing a standard CUDA approach.
• tdr-inverse: Inverts a matrix on-GPU utilizing an asynchronous approach, see Harmonize.

In addition, a few subdirectories contain tools to aid in comparing each implementation:

• matrix_gen/: Python scripts for generating matrices and generating inverses for matrices in Matrix Market format.
• results/: Scripts for running benchmarks against all implementations and interpreting results.
• tests/: Some basic matrices to confirm programs are working as intended.

This repository contains submodules, please checkout recursively:

git clone --recursive https://github.com/scrufulufugus/tdr-inverse.git

The only hard dependencies outside of submodules are CUDA and gnumake. Both should be installed though the correct repositories for your distribution of choice.

The matrix_gen/ and results/ subdirectories additionally require anaconda or a python environment with the dependencies detailed in their respective environment files. See the README in each subdirectory for more information.

All three inverse programs can be built by running:

make

Once built, the binaries can be found in bin/. Each program takes two arguments: an input matrix and its inverse. The program will then perform an inverse and return the error from the second argument.

./bin/tdr-inverse ./tests/3x3.csv ./tests/3x3_soln.csv