matvec

A domain-specific language for fast graph shift operations. This implements mathematical fields on numbers, n-dimensional column vectors, and n-by-n sparse matrices.

License: Apache Software License
Author: Emmanouil (Manios) Krasanakis
Dependencies: numpy

:zap: Quickstart

Creating a 5-dimensional vector (can use numpy arrays as inputs interchangeably with lists everywhere):

from matvec import Vector
x = Vector([1, 2, 3, 4, 5])

Creating a 5x5 sparse matrix A in coo-format with non-zero elements A[1,2]=9 and A[3,0]=21

from matvec import Matrix
A = Matrix([1, 2],
           [3, 0],
           [9, 21],
           5)

Print the outcome of matrix-vector multiplication:

print(A*x)

Print the outcome of left-multiplying the transpose of x with A:

print(x*A)

:fire: Features

:rocket: Parallelized matrix-vector multiplication.
:chart_with_downwards_trend: Memory reuse optimization.
:mag: numpy compatibility.
:factory: Common arithmetic operations.

:volcano: Benchmark

Benchmarks tested on a machine with 2.6 GHz CPU base clock and up to 4.4 GHz turbo boost, 12 logical cores, and 16GB DDR3 RAM. They span vectors of 1.E4 to 1.E6 elements and matrices with up to 100x the number of non-zeroes (that is, if the matrices represented graphs, node degrees would be up to 100).

For a lot of non-zeroes paralellization may be worse than scipy.

Task	numpy/scipy	matvec
Create new vector or array	0.019 sec	0.019 sec
1000 temp. additions of 1.E6 vectors	1.897 sec	1.321 sec
Create matrix	0.505 sec	0.183 sec
Sparse matrix with vector multiplication	0.269 sec	0.103 sec

:memo: List of Operations

• Full arithmetic operations * + - / == < > <= >= between vectors and other vectors or scalars.
• Matrix-vector multiplication * (both left and right).
• Element access and assignment for vectors with [].
• Masking, such as y = x[x>0].
• matvec.clear() clears cache.