jblas – Linear Algebra for Java

If you are really impatient, I’d suggest you read the Classes Overview below and otherwise stick to the API Documentation for the classes like DoubleMatrix.

The main goals of jblas were to provide very high performance, close to what you get from state-of-the-art BLAS and LAPACK libraries, and easy of use, which means that in the ideal case, you can just mechanically translate a matrix expression from formulas to Java code.

In all brevity, here is what you need to know to get started:

• There exist four classes: FloatMatrix, DoubleMatrix, ComplexFloatMatrix and ComplexDoubleMatrix in the package org.jblas which represent real and complex matrices in single and double precision.
• Higher-level routines for solving equations, or computing eigenvalues are grouped in classes like Eigen, Solve, or Geometry.
• To construct a new matrix, you can either use the constructor, or one of the factory methods ones (constructs a matrix of all ones), zeros, rand (entries uniformly distributed between 0 and 1), randn (entries normally distributed), eye (unit matrix), diag (matrix with given diagonal). Dimensions are specified in the order “row”, “column”. The number of columns defaults to 1 if omitted (meaning that you construct a row vector, if you supply just one dimension).
• To access elements, you use put and get. Methods also exist for reading or writing a whole column, row, or submatrix.
• There are only two-dimensional matrices. Vectors are matrices whose number of columns or rows are 1. This has turned out to be much more convenient than having separated classes.
• Every math operator maps to a short mnemonic name. For example, + becomes add, – becomes sub, * becomes mul, / becomes div, and so on.
• Often, you can pass a double or float value, or a matrix with only one element as the argument to a method, for example, to add the same value to all elements of the matrix.
• mul is element-wise multiplication. Matrix-matrix multiplication is called mmul.
• Often, you can add an “i” to a method to have it work “in-place”. For example, addi is like +=.

What is missing right now:

• Right now, the four classes more or less exist next to each other, with no abstract superclass. This makes the classes pretty straightforward, but the downside is that you cannot have a function which works with any kind of matrices.
• No support for sparse matrices.
• Not all of LAPACK is covered, only things I’m using myself. In principle, there is little overhead in adding further functions as the generation of wrappers is automatic, but I’d rather include a function from LAPACK only after I’m sure it does what it’s supposed to do. In other words, I’ll happily add anything somebody needs as long as he can check whether the method works as it should.
• jblas uses double and float arrays to store the matrix. Whenever you call a native function, the array is first copied. This means that it doesn’t make much sense to call a native routine if its computation is linear in the size of the data, but this includes most of BLAS Level 1 and Level 2. jblas therefore uses the Java implementation for things like vector addition, or even matrix-vector multiplication and is therefore not as fast as native BLAS. Currently, I’m contemplating some caching schemes to improve performance here.