[ANN] Finch.jl: Sparse and Structured Array Fusion

Package Announcements
, ,
16

Thanks for your feedback! I’ll add a TODO for migrating from SparseArrays.

On the performance: The Finch high-level scheduler (which handles operations like + and *) is experimental, and built for generality. We’re currently working on optimizing the high-level scheduler to use more specialized techniques as well as adding parallelism. In the meantime, it may be better to write Finch IR directly:

julia> function faster_mul(A, B)
           w = Tensor(SparseByteMap(Element(0.0)))
           C = Tensor(CSCFormat())
           @finch begin
               C .= 0
               for j=_
                   w .= 0
                   for k=_, i=_; w[i] += A[i, k] * B[k, j] end
                   for i=_; C[i, j] = w[i] end
               end
               return C
           end
           return C
       end
@time x * y; @time X * Y; @time faster_mul(X, Y);
  2.599353 seconds (13 allocations: 1.663 GiB, 3.37% gc time)
 14.993908 seconds (601 allocations: 20.312 GiB, 3.12% gc time, 0.00% compilation time)
  4.000148 seconds (2.21 M allocations: 5.690 GiB, 2.28% gc time)
6 Likes