Store the broadcast style of the blocks in BlockSparseArrayStyle. This would help with generic code, for example if a block sparse array has GPU blocks, BlockSparseArrayStyle could keep track of that by storing the block style as the corresponding GPU array broadcasting style. Then, we can more accurately determine the destination type of a broadcast expression from all of the arguments that were input.
This is analogous to how ArrayLayouts.MemoryLayout definitions for BlockArrays.jlBlockedArray store information about the memory layout of the blocks: https://github.com/JuliaArrays/BlockArrays.jl/blob/v1.4.0/src/blocklinalg.jl#L82-L83
Currently, we determine the output type just based on the first argument:
|
return similar(first(m.args), elt, combine_axes(axes.(m.args)...)) |
which is not ideal and could make some mistakes (like making a destination with sparse blocks that should have dense blocks).
Store the broadcast style of the blocks in
BlockSparseArrayStyle. This would help with generic code, for example if a block sparse array has GPU blocks,BlockSparseArrayStylecould keep track of that by storing the block style as the corresponding GPU array broadcasting style. Then, we can more accurately determine the destination type of a broadcast expression from all of the arguments that were input.This is analogous to how
ArrayLayouts.MemoryLayoutdefinitions forBlockArrays.jlBlockedArraystore information about the memory layout of the blocks: https://github.com/JuliaArrays/BlockArrays.jl/blob/v1.4.0/src/blocklinalg.jl#L82-L83Currently, we determine the output type just based on the first argument:
BlockSparseArrays.jl/src/blocksparsearrayinterface/broadcast.jl
Line 37 in 6211e57