Skip to content

jamesETsmith/mtxio

Repository files navigation

Faster Parsing for Matrix Market

build

There are several great C++ projects (PIGO and fast_matrix_market) for reading matrix market files faster in serial and parallel. However, GraphBLAS is written in ANSI C and I wanted to try and write a matrix market reader (and eventually writer) in C so it could be used side-by-side with GraphBLAS. This is based heavily on the great C++ projects mentioned above and several of the utility functions are lifted almost verbatim from those projects.

⚠️ This project is pre-alpha and is currently unstable. Currently only reading matrix market coordinate files with doubles is supported. Please understand there will be breaking changes.

TODOs

Features required to meet specification:

  • Object

    • matrix
    • vector
  • Format

    • coordinate
    • array
  • Fields

    • real
    • complex
    • integer
    • pattern (i.e. no values)
  • Symmetry

    • general
    • symmetric
    • skew-symmetric
    • hermitian
  • Feature: Add sanitizer checks to CI

  • Feature: handle inf, nan, etc

  • Feature: gracefully handle non-conformant/non-mtx files

  • Feature: add parallel write

  • Feature: Add proper CMake aliases (like mtxio::mtxio) to make it easier to integrate with other CMake projects

  • Feature: Clean up CMake default flags (maybe use CMake configurations?)

  • Performance: Improve parallel read performance, still slower than PIGO by 30-50%.

Prerequisites

  • C/C++ compiler
  • Python3 and SciPy (Optional, needed to make benchmarking data)

Build

[email protected]:jamesETsmith/mtxio.git
cd mtxio
export CPM_SOURCE_CACHE=$HOME/.cache/CPM # (optional)
cmake -B build
cmake --build build
ctest --test-dir build

Benchmarking

cd data && python generate_data.py # May take a while
cd ..
perf stat -- env OMP_NUM_THREADS=12 ./build/bench/pigo_read data/n_15_0.3.mtx
perf stat -- env OMP_NUM_THREADS=12 ./build/bench/mtxio_read data/n_15_0.3.mtx