Skip to content

Extending scatter! to work with CUDA sparse arrays (retry of #648) #672

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

Open
alonsoC1s wants to merge 7 commits into
base: master
Choose a base branch
from
Open

Extending scatter! to work with CUDA sparse arrays (retry of #648) #672

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
472122e
Expanded dispatch of scatter! to include AbstractCuSparseArray
alonsoC1s Nov 14, 2025
cfccdb3
Fix incorrect import
alonsoC1s Nov 14, 2025
1551922
restore formatting
alonsoC1s Nov 14, 2025
c749dfc
Typos
alonsoC1s Nov 14, 2025
94f18bb
Adding tests for scatter on CUSPARSE arrays
alonsoC1s Nov 25, 2025
098aca0
Fixing erroneously made tests for specialized sparse scatter kernels.
alonsoC1s Jan 27, 2026
465fea2
Merge branch 'master' into master
alonsoC1s Jan 27, 2026
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
18 changes: 9 additions & 9 deletions ext/NNlibCUDAExt/scatter.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,5 @@
# supported op: +, -, *, /, max, min, &, |, mean
import CUDA.CUSPARSE: AbstractCuSparseArray

## TODO support sparse dst/src/idx
## See issue https://github.com/FluxML/NNlib.jl/issues/647
Expand Down Expand Up @@ -54,10 +55,9 @@ function scatter_kernel!(op::OP, dst, src, idx::CUDA.CuDeviceArray{<:CartesianIn
return nothing
end


function NNlib.scatter!(op::OP, dst::AnyCuArray,
src::AnyCuArray,
idx::AnyCuArray) where OP
function NNlib.scatter!(op::OP, dst::Union{AnyCuArray,AbstractCuSparseArray},
src::Union{AnyCuArray,AbstractCuSparseArray},
idx::Union{AnyCuArray,AbstractCuSparseArray}) where OP
isempty(idx) && return dst
dims = NNlib.scatter_dims(dst, src, idx)
args = if dims == 0
Expand All @@ -78,9 +78,9 @@ function NNlib.scatter!(op::OP, dst::AnyCuArray,
return dst
end

function NNlib.scatter!(op::typeof(mean), dst::AnyCuArray,
src::AnyCuArray,
idx::AnyCuArray)
function NNlib.scatter!(op::typeof(mean), dst::Union{AnyCuArray,AbstractCuSparseArray},
src::Union{AnyCuArray,AbstractCuSparseArray},
idx::Union{AnyCuArray,AbstractCuSparseArray})
Ns = NNlib.scatter!(+, zero(dst), one.(src), idx)
dst_ = NNlib.scatter!(+, zero(dst), src, idx)
dst .+= NNlib.safe_div.(dst_, Ns)
Expand Down Expand Up @@ -177,8 +177,8 @@ function ∇scatter_src_kernel!(op::OP, Δsrc, src, idx::CUDA.CuDeviceArray{<:Ca
end

function NNlib.∇scatter_src(op::Union{typeof(*),typeof(/)}, Δ, dst,
src::AnyCuArray,
idx::AnyCuArray)
src::Union{AnyCuArray{Tsrc,Nsrc},AbstractCuSparseArray},
idx::Union{AnyCuArray{Tidx,Nidx},AbstractCuSparseArray}) where {Tsrc,Tidx,Nsrc,Nidx}
dims = ndims(src) - ndims(idx)
Δsrc = NNlib.modify_src(op, NNlib.gather(Δ, idx), src)
rev_idx = NNlib.reverse_indices(idx)
Expand Down
108 changes: 105 additions & 3 deletions test/ext_cuda/scatter.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,13 +1,13 @@
dsts = Dict(
0 => cu([3, 4, 5, 6, 7]),
1 => cu([3 3 4 4 5;
5 5 6 6 7]),
5 5 6 6 7]),
)
srcs = Dict(
(0, true) => cu(ones(Int, 3, 4)),
(0, false) => cu(ones(Int, 3) * collect(1:4)'),
(1, true) => cu(ones(Int, 2, 3, 4)),
(1, false) => cu([1, 2] .* reshape(ones(Int, 3) * collect(1:4)', 1,3,4)),
(1, false) => cu([1, 2] .* reshape(ones(Int, 3) * collect(1:4)', 1, 3, 4)),
)
idxs = [
cu([1 2 3 4;
Expand All @@ -21,7 +21,7 @@ idxs = [
(3,) (5,) (5,) (3,)])), # CartesianIndex index
]

types = [CuArray{Int32}, CuArray{Int64}, CuArray{Float32}, CuArray{Float64}]
types = [CuArray{Int32}, CuArray{Int64}, CuArray{Float32}, CuArray{Float64}, CuSparseMatrixCSC{Float32}, CuSparseMatrixCSR{Float32}, CuSparseMatrixCOO{Float32}]


@testset "scatter" begin
Expand Down Expand Up @@ -70,6 +70,108 @@ types = [CuArray{Int32}, CuArray{Int64}, CuArray{Float32}, CuArray{Float64}]
end


# Specialized sparse scatter kernels. Duplicated as test cases above do not cover sparse arrays.
dsts_sp = Dict(
0 => cu(sparse([3, 4, 5, 6, 7])),
1 => cu(sparse([3 3 4 4 5;
5 5 6 6 7])),
)
srcs_sp = Dict(
(0, true) => cu(sparse(ones(Int, 3, 4))),
(0, false) => cu(sparse(ones(Int, 3) * collect(1:4)')),
# No sparse equivalent for 3D arrays
)
types_sp = [
CuSparseMatrixCSC{Int32}, CuSparseMatrixCSC{Int64}, CuSparseMatrixCSC{Float32}, CuSparseMatrixCSC{Float64},
CuSparseMatrixCSR{Int32}, CuSparseMatrixCSR{Int64}, CuSparseMatrixCSR{Float32}, CuSparseMatrixCSR{Float64},
CuSparseMatrixCOO{Int32}, CuSparseMatrixCOO{Int64}, CuSparseMatrixCOO{Float32}, CuSparseMatrixCOO{Float64}
]

@testset "scatter sparse-specialized" begin
for T = types_sp
@testset "$(T)" begin
@testset "+" begin
# Dims is implicitly 0. No sparse equivant for multidimensional src/dst
for idx = idxs
mutated = true
gputest((dst, src) -> NNlib.scatter!(+, dst, src, idx), T(copy(dsts[0])), T(srcs[(0, mutated)]), checkgrad=true)

mutated = false
gputest(src -> NNlib.scatter(+, src, idx), T(srcs[(0, mutated)]), checkgrad=true)
end
end

@testset "-" begin
for idx = idxs
mutated = true
gputest((dst, src) -> NNlib.scatter!(-, dst, src, idx), T(copy(dsts[0])), T(srcs[(0, mutated)]), checkgrad=true)

mutated = false
gputest(src -> NNlib.scatter(-, src, idx), T(srcs[(0, mutated)]), checkgrad=true)
end
end

@testset "max" begin
for idx = idxs
mutated = true
gputest((dst, src) -> NNlib.scatter!(max, dst, src, idx), T(copy(dsts[0])), T(srcs[(0, mutated)]), checkgrad=true)

mutated = false
gputest(src -> NNlib.scatter(max, src, idx), T(srcs[(0, mutated)]), checkgrad=true)
end
end

@testset "min" begin
for idx = idxs
mutated = true
gputest((dst, src) -> NNlib.scatter!(min, dst, src, idx), T(copy(dsts[0])), T(srcs[(0, mutated)]), checkgrad=true)

mutated = false
gputest(src -> NNlib.scatter(min, src, idx), T(srcs[(0, mutated)]), checkgrad=true)
end
end
end
end

# Sparse-specialized for operations not tested on eltype <: Integer
for T = [CuSparseMatrixCSC{Float32}, CuSparseMatrixCSC{Float64}, CuSparseMatrixCSR{Float32}, CuSparseMatrixCSR{Float64}, CuSparseMatrixCOO{Float32}, CuSparseMatrixCOO{Float64}]
@testset "$(T)" begin
# Dims is implicitly 0. No sparse equivant for multidimensional src/dst
@testset "*" begin
for idx = idxs
mutated = true
gputest((dst, src) -> NNlib.scatter!(*, dst, src, idx), T(copy(dsts[0])), T(srcs[(0, mutated)]), checkgrad=true)

mutated = false
gputest(src -> NNlib.scatter(*, src, idx), T(srcs[(0, mutated)]), checkgrad=true)
end
end

@testset "/" begin
for idx = idxs, dims = [0, 1]
mutated = true
gputest((dst, src) -> NNlib.scatter!(/, dst, src, idx), T(copy(dsts[0])), T(srcs[(0, mutated)]), checkgrad=true)

mutated = false
gputest(src -> NNlib.scatter(/, src, idx), T(srcs[(0, mutated)]), checkgrad=true)
end
end

@testset "mean" begin
for idx = idxs, dims = [0, 1]
mutated = true
gputest((dst, src) -> NNlib.scatter!(mean, dst, src, idx), T(copy(dsts[0])), T(srcs[(0, mutated)]), checkgrad=true)

mutated = false
gputest(src -> NNlib.scatter(mean, src, idx), T(srcs[(0, mutated)]), checkgrad=true)
end
end
end
end
end



for T = [CuArray{Float32}, CuArray{Float64}]
@testset "$(T)" begin
@testset "*" begin
Expand Down
Loading