Overview

Galley is a system for declarative sparse tensor algebra. It combines techniques from database query optimization with the flexible formats and execution of sparse tensor compilers, specifically the Finch compiler. Details about the theory and system design can be found in our paper here. There are two main ways to interact with Galley:

• Galley can be used as an optimizer for the array interface of Finch.jl.
• Users can write programs directly in Galley's notation.

Note: In both of these cases, Finch needs to be imported because Galley relies on it for both data loading and for compiling sparse tensor functions.

Galley: An Optimizer for Finch.jl

Using Galley with Finch's array API just requires setting the ctx parameter to galley_scheduler() in the compute function!

using Finch
using Galley

A = lazy(Tensor(Dense(SparseList(Element(0.0))), fsprand(100, 100, .1)))
B = lazy(Tensor(Dense(SparseList(Element(0.0))), fsprand(100, 100, .1)))
C = compute(A * B, ctx = galley_scheduler())

Alternatively, you can specify that Galley should be used at the top of the file:

using Finch
using Galley

Finch.set_scheduler!(galley_scheduler())
A = lazy(Tensor(Dense(SparseList(Element(0.0))), fsprand(100, 100, .1)))
B = lazy(Tensor(Dense(SparseList(Element(0.0))), fsprand(100, 100, .1)))
C = compute(A * B)

This is the preferred usage pattern for Galley.

Galley: Direct Usage

Instead of going through the array API, you can specify your computation directly with Galley. This uses the grammar defined in src/plan.jl. For example, we can specify the same matrix multiplication as above using:

using Finch 
using Galley

A = Tensor(Dense(SparseList(Element(0.0))), fsprand(100, 100, .1))
B = Tensor(Dense(SparseList(Element(0.0))), fsprand(100, 100, .1))
C = galley(Mat(:i, :k, Agg(+, :j, MapJoin(*, Input(A, :i, :j), Input(B, :j, :k))))).value[1]
C == compute(lazy(A) * B, ctx=galley_scheduler()) # True