ACMI's a Convergent Matrix Inverter.

ACMI implements M. Altman's cubically-convergent, iterative algorithm to compute matrix inverses. GPU-hosted computation is currently supported using CUDA. ACMI is being developed on the Pascal architecture.

The research behind this work may be found here.

Dependencies:

• CUDA SDK 8.0
• GCC
• GNU Make
• OpenBLAS
• CMocka for unit tests

To build:
make

To build with debugging symbols:
make debug

To run unit tests:
make test

For usage instructions, run (after building):
acmi

To invert a 10000x10000 random, diagonally-dominant integer matrix using CUDA without writing any matrix files, run:
acmi @10000 -D

Large, sparse matrices can be found here.

This project uses Matrix Market I/O routines from:
http://math.nist.gov/MatrixMarket/mmio-c.html

TODO:

• More unit tests!
• Make reported iteration count consistent
• OpenCL/Vulkan support
• Library build
• Fix makefile quirks
• Sprout CLI handler out of main() and handle conflicting options
• Portable PRNG seeding
• Optimized multiplication of symmetric or sparse matrices during the first iteration
• Double double support
• Add support for high-precision MATLAB-format matrix I/O