Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Simplify conversions. #368

Merged
merged 1 commit into from
Jul 23, 2021
Merged

Simplify conversions. #368

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
67 changes: 9 additions & 58 deletions src/host/construction.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,14 @@
# convenience and indirect construction

# conversions from CPU arrays rely on constructors
Base.convert(::Type{T}, a::AbstractArray) where {T<:AbstractGPUArray} = a isa T ? a : T(a)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This looks like a potential for a lot of invalidations. Was that checked?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

No, I didn't check for that. We did used to have a Base.convert(AT::Type{<: AbstractGPUArray}, iter) method though, which is worse (right?), and I don't remember this convert method standing out when I had a quick look at invalidations a couple of weeks ago.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, this should at least be better. Interesting.

# TODO: can we implement constructors to and from ::AbstractArray here? by calling the undef
# constructor and doing a `copyto!`. this is tricky, due to ambiguities, and no easy
# way to go from <:AbstractGPUArray{T,N} to e.g. CuArray{S,N}


## convenience constructors

function Base.fill!(A::AnyGPUArray{T}, x) where T
length(A) == 0 && return A
gpu_call(A, convert(T, x)) do ctx, a, val
Expand Down Expand Up @@ -49,61 +58,3 @@ end

Base.one(x::AbstractGPUMatrix{T}) where {T} = _one(one(T), x)
Base.oneunit(x::AbstractGPUMatrix{T}) where {T} = _one(oneunit(T), x)


## collect & convert

function indexstyle(x::T) where T
style = try
Base.IndexStyle(x)
catch
nothing
end
style
end

function collect_kernel(ctx::AbstractKernelContext, A, iter, ::IndexCartesian)
idx = @cartesianidx(A)
@inbounds A[idx...] = iter[idx...]
return
end

function collect_kernel(ctx::AbstractKernelContext, A, iter, ::IndexLinear)
idx = linear_index(ctx)
@inbounds A[idx] = iter[idx]
return
end

eltype_or(::Type{<: AbstractGPUArray}, or) = or
eltype_or(::Type{<: AbstractGPUArray{T}}, or) where T = T
eltype_or(::Type{<: AbstractGPUArray{T, N}}, or) where {T, N} = T

function Base.convert(AT::Type{<: AbstractGPUArray}, iter)
isize = Base.IteratorSize(iter)
style = indexstyle(iter)
ettrait = Base.IteratorEltype(iter)
if isbits(iter) && isa(isize, Base.HasShape) && style != nothing && isa(ettrait, Base.HasEltype)
# We can collect on the GPU
A = similar(AT, eltype_or(AT, eltype(iter)), size(iter))
gpu_call(collect_kernel, A, iter, style)
A
else
convert(AT, collect(iter))
end
end

function Base.convert(AT::Type{<: AbstractGPUArray{T, N}}, A::DenseArray{T, N}) where {T, N}
copyto!(AT(undef, size(A)), A)
end

function Base.convert(AT::Type{<: AbstractGPUArray{T1}}, A::DenseArray{T2, N}) where {T1, T2, N}
copyto!(similar(AT, size(A)), convert(Array{T1, N}, A))
end

function Base.convert(AT::Type{<: AbstractGPUArray}, A::DenseArray{T2, N}) where {T2, N}
copyto!(similar(AT{T2}, size(A)), A)
end

function Base.convert(AT::Type{Array{T, N}}, A::AbstractGPUArray{CT, CN}) where {T, N, CT, CN}
convert(AT, copyto!(Array{CT, CN}(undef, size(A)), A))
end