Fixed performance issues in consistent/assemble, new and improved distributed sparse matrix multiplication algorithms. #170
Conversation
…d versions of repartition(A,rows,cols) and centralize that support non-default sparse method.
…improved cosistent further by reusing own_ghost_block as much as possible through new sparse utils expand_sparse_matrix _columns
…atency hiding to PartitionedArrays. Additional tests for these functions have been added. Sparse utils has been extended with some new functions required by SpMM, etc. Serial kernels for the methods are provided in a new file 'sequential implementations'.
jop611
changed the title
…plications algorithms
| @@ -948,9 +948,9 @@ parallel implementations. | |||
|
|
|||
| # Properties | |||
|
|
|||
| - `matrix_partition::A` | |||
There was a problem hiding this comment.
The first line of the doc string says struct PSparseMatrix{V,B,C,D,T}, and the first property used to say matrix_partition::A, but A it not an existing field, so to me it feels like my changes are correct, and the existing documentation is not.
src/sequential_implementations.jl
Outdated
There was a problem hiding this comment.
There is a lot of "type piracy in this file". One cannot define methods for functions and types in other packages.
If you want to work with your own versions of the functions, then you need to consider different function names.
There was a problem hiding this comment.
I believe I resolved this now, but this probably means that *,\, +, - do not behave as expected for SparseMatrixCSR structs. mul! is replaced by matmul!, which I believe is not exported by any library
Performance changes to consistent and assemble.