Defer BroadcastStyle construction to back-ends. #506
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
maleadt
force-pushed
the
tb/defer_broadcaststyle
branch
3 times, most recently
from
9ec1077
to
f20efe2
Compare
maleadt
force-pushed
the
tb/defer_broadcaststyle
branch
14 times, most recently
from
852447f
to
abdac33
Compare
maleadt
force-pushed
the
tb/defer_broadcaststyle
branch
from
abdac33
to
6a017b6
Compare
maleadt
force-pushed
the
tb/defer_broadcaststyle
branch
from
6b6eed4
to
0359514
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
To fix JuliaGPU/CUDA.jl#2191, where broadcasting a CuArray backed by unified memory results in a CuArray in device memory (i.e. the buffer type is lost), I want to make the
CuArrayStylebroadcast style include the buffer type so that we can preserve it. That's currently impossible, as the style is constructed by GPUArrays, by doing the very cursedAdapt.parent(W){Adapt.eltype(W), Adapt.ndims(W)}constructor call (Adapt.parentcurrently returns an unordained typename), which both does not know about the additional buffer typevar, and calls a specific constructor without setting the buffer typevar.Instead, in JuliaGPU/Adapt.jl#75 I make it so that
Adapt.parent(renamed toAdapt.parent_typeto avoid confusion withBase.parent, which doesn't work on types) returns the full type, including the buffer typevar. Then, in this PR, I make it so that the back-end is not responsible for providing theBroadcastStylemethods, so that additional information can be put in there.Changes to back-ends should be minimal, see the JLArrays diff in here, or JuliaGPU/CUDA.jl#2203. It is however a breaking change.
cc @vchuravy @jpsamaroo @pxl-th