Join GitHub today
GitHub is home to over 28 million developers working together to host and review code, manage projects, and build software together.Sign up
Clone this wiki locally
Implementation guide for core.matrix
This guide is for people who want to create a core.matrix implementation. i.e. extend core.matrix to work with a new underlying matrix library
We encourage you to extend the clojure.core.matrix protocols to different types of matrices / multi-dimensional data formats. This is relatively easy and is referred to as a clojure.core.matrix "implementation". You can write an implementation for a wide variety of purposes:
- Java matrix libaries that you want to wrap for use with clojure.core.matrix
- Custom Clojure data structures
- "array-like" objects that can be viewed as matrices (e.g. bitmap images)
- Tabular "result-set" objects
Depending on the requirements, you may choose to implement support for some or all core.matrix APIs. Writing a implementation that supports just the essential protocols is pretty simple. If you want to add extra features or obtain faster performance, you can implement the other protocols as needed at a later date.
If you are interested in creating a clojure.core.matrix implementation, please read:
Also please be aware that prior to version 1.0.0, the API is still in flux and you can expect some breaking changes.
1. Create a project for your matrix implementation
This can be any project type (leiningen or Maven).
You want to have a separate project for your core.matrix implementation, which can pull in any dependencies it needs. Typically you might import any .jar libraries that are needed for your underlying matrix implementation.
e.g. if you are wrapping the UJMP Java matrix library, you would add UJMP as a dependency.
2. Add clojure.core.matrix as a dependency to your project
This is necessary to get access to the key namespaces you need:
- clojure.core.matrix : contains the user-facing API
- clojure.core.matrix.protocols : contains protocols that must be implemented
- clojure.core.matrix.implementations : contains code to register / manage your implementation
- clojure.core.matrix.compliance-tester: test code to verify your implementation is correct
3. Implement the mandatory protocols
These are documented in
4. Register your implementation
You should call
clojure.core.matrix.implementations/register-implementation with an instance of your matrix library.
You can call this at any point during the loading of your implementation after you have implemented the mandatory protocols for the given instance.
5. Run compliance tests
In your test suite you should call into the compliance tester tool to verify that you have correctly implemented the core.matrix API.
Code to call the compliance test should look something like this:
(deftest compliance-test (clojure.core.matrix.compliance-tester/compliance-test my-matrix-instance))
Any error in the compliance test mean that there is a bug somewhere - either your implementation doesn't conform to the clojure.core.matrix API, or we have a bug in the compliance test assumptions :-)
1. Implement the optional protocols
You only need the mandatory protocols to have a working clojure.core.matrix implementation.
Typically however, you will want to implement some or all of the optional protocols to take advantage of special features of your matrix library - using these features may offer users much better performance than the default core.matrix implementations, which are written for flexibility and a generic design rather than for speed.
2. Send a patch for "KNOWN-IMPLEMENTATIONS"
This will potentially enable us to do more clever stuff with implementation metadata in the future.
It will also help users discover your library!
Handling interop with other implementations
When control enters your implementation via one of the protocol functions, you don't know much about the type of the other arguments. This is particularly important when the other argument may be a matrix from a different core.matrix implementation.
Such operations in clojure.core.matrix are left to the discretion of the matrix implementation. You are required to either:
- Perform the operation and return a valid result.
- Throw an exception.
Here are your options, in rough order of preference:
Coerce the other matrix to your format - via calling
(coerce your-matrix other-matrix). This may be an expensive operation (likely to require constructing a whole new matrix in your format) but should work effectively as a general approach
Defer to a generic mutimethod - core.matrix provides some generic multimethods that should perform
the necessary operations, e.g.
clojure.core.matrix.multimethods/mul. Generic implementations are likely to be slow but correct. In some cases the generic method may be able to exploit optimisations, e.g. multiplication of two diagonal matrices from different implementations.
- Explicitly recognise and work with the other implementation - this requires hard-coding and is a lot of work. Probably not recommended except for special cases where you really need to work well with another specific implementation
- Give a warning - useful in combination with one of the other methods if your library is focused on performance, to let the user know that they are taking a performance hit while still attempting to produce a correct result.
- Throw an exception - This is not ideal, but may be acceptable if you are willing to label your implementation clearly as incomplete / not compatible with other implementations.