https://github.com/queracomputing/parallelmergecsr.jl
Merge-based Parallel Sparse Matrix-Vector Multiplication in Julia
https://github.com/queracomputing/parallelmergecsr.jl
julia matrix-multiplication
Last synced: 4 months ago
JSON representation
Merge-based Parallel Sparse Matrix-Vector Multiplication in Julia
- Host: GitHub
- URL: https://github.com/queracomputing/parallelmergecsr.jl
- Owner: QuEraComputing
- License: apache-2.0
- Created: 2022-11-18T18:44:07.000Z (almost 3 years ago)
- Default Branch: main
- Last Pushed: 2024-07-11T14:46:11.000Z (about 1 year ago)
- Last Synced: 2025-06-15T01:08:19.906Z (4 months ago)
- Topics: julia, matrix-multiplication
- Language: Julia
- Homepage:
- Size: 945 KB
- Stars: 8
- Watchers: 2
- Forks: 2
- Open Issues: 2
-
Metadata Files:
- Readme: README.md
- License: LICENSE
- Citation: CITATION.cff
Awesome Lists containing this project
README
# ParallelMergeCSR.jl
[](https://github.com/QuEraComputing/ParallelMergeCSR.jl/actions)
[](https://codecov.io/gh/QuEraComputing/ParallelMergeCSR.jl)
An implementation/port of Merrill and Garland's Merge-based Parallel Sparse Matrix-Vector Multiplication (10.1109/SC.2016.57) paper in
the
Julia Programming Language
ParallelMergeCSR allows you to perform *multithreaded* [CSC formatted sparse Matrix](https://en.wikipedia.org/wiki/Sparse_matrix#Compressed_sparse_column_.28CSC_or_CCS.29) multiplication against dense vectors and matrices as long as the sparse Matrix has had a **transpose** or **adjoint** operation applied to it via `LinearAlgebra`, built-in to Julia Base. The reason for this is the original algorithm was restricted to [CSR formatted sparse Matrices](https://en.wikipedia.org/wiki/Sparse_matrix#Compressed_sparse_row_(CSR,_CRS_or_Yale_format)) but by taking the transpose of a CSC matrix you've created a CSR representation of the same matrix.
ParallelMergeCSR only has one function intended for use: `mul!`, which is used for both Sparse Matrix - Dense Vector and Sparse Matrix - Dense Matrix multiplication. The function is not exported by default to avoid conflict with `LinearAlgebra.mul!` and must be invoked through its fully qualified name `ParallelMergeCSR.mul!`.
## Supported Platforms
ParallelMergeCSR takes advantage of the [Polyester.jl](https://github.com/JuliaSIMD/Polyester.jl) package for multithreading support. Unfortunately, Polyester has known issues with working on Mac OS and ParallelMergeCSR should be used on Linux platforms only.
## Installation
You can install ParallelMergeCSR through the Julia package manager interface (just hit `]` when you're in the Julia prompt) by typing the following:
```
add ParallelMergeCSR
```
## Usage
Start Julia with the desired number of threads by launching it with the `-t`/`--threads` argument:
```
julia --threads
```
or setting `JULIA_NUM_THREADS` in your environment and then running Julia.
You can confirm Julia is using the specified number of threads via:
```julia
julia> Threads.nthreads()
```
You can then use `ParallelMergeCSR` in a similar fashion to the example below:
```julia
julia> using ParallelMergeCSR, SparseArrays
# create a 20x20 transposed CSC-formatted Sparse matrix with a 30% chance of values appearing
julia> A = transpose(sprand(20, 20, 0.3));
# dense vector (can be a matrix too)
julia> B = rand(size(A, 2));
# output
julia> C = rand(size(A, 2));
# coefficients
julia> α = -1.0; β = 2.9;
# perform the operation C = ABα + Cβ, mutating C to store the answer
julia> ParallelMergeCSR.mul!(C, A, B, α, β)
```