Skip to content

Extending scatter! to work with CUDA sparse arrays #648

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 5 commits into
base: master
Choose a base branch
from
+16 −9
Open

Extending scatter! to work with CUDA sparse arrays #648

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 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
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
20 changes: 13 additions & 7 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

function scatter_kernel!(op::OP, dst, src, idx) where OP
index = threadIdx().x + (blockIdx().x - 1) * blockDim().x
Expand Down Expand Up @@ -43,7 +44,10 @@ 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
dims = NNlib.scatter_dims(dst, src, idx)
args = if dims == 0
max_idx = length(idx)
Expand All @@ -63,7 +67,9 @@ function NNlib.scatter!(op::OP, dst::AnyCuArray, src::AnyCuArray, idx::AnyCuArra
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 All @@ -74,7 +80,7 @@ end
## Gradients

function ∇scatter_src_kernel!(op::OP, Δsrc, src, idx,
rev_idx, max_idx, T::Type{TT}) where {OP,TT}
rev_idx, max_idx, T::Type{TT}) where {OP,TT}
index = threadIdx().x + (blockIdx().x - 1) * blockDim().x

@inbounds if index <= max_idx
Expand All @@ -94,7 +100,7 @@ function ∇scatter_src_kernel!(op::OP, Δsrc, src, idx,
end

function ∇scatter_src_kernel!(op::OP, Δsrc, src, idx::CUDA.CuDeviceArray{<:CartesianIndex},
rev_idx, max_idx, T::Type{TT}) where {OP,TT}
rev_idx, max_idx, T::Type{TT}) where {OP,TT}
index = threadIdx().x + (blockIdx().x - 1) * blockDim().x

@inbounds if index <= max_idx
Expand All @@ -114,7 +120,7 @@ function ∇scatter_src_kernel!(op::OP, Δsrc, src, idx::CUDA.CuDeviceArray{<:Ca
end

function ∇scatter_src_kernel!(op::OP, Δsrc, src, idx,
rev_idx, pre_cart_idx, max_dims_idx, max_idx, T::Type{TT}) where {OP,TT}
rev_idx, pre_cart_idx, max_dims_idx, max_idx, T::Type{TT}) where {OP,TT}
index = threadIdx().x + (blockIdx().x - 1) * blockDim().x

@inbounds if index <= max_idx
Expand Down Expand Up @@ -160,8 +166,8 @@ 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{Tsrc,Nsrc},
idx::AnyCuArray{Tidx,Nidx}) where {Tsrc,Tidx,Nsrc,Nidx}
src::Union{AnyCuArray{Tsrc,Nsrc},AbstractCuSparseArray},
idx::Union{AnyCuArray{Tidx,Nidx},AbstractCuSparseArray}) where {Tsrc,Tidx,Nsrc,Nidx}
dims = Nsrc - Nidx
Δsrc = NNlib.modify_src(op, NNlib.gather(Δ, idx), src)
rev_idx = NNlib.reverse_indices(idx)
Expand Down
1 change: 1 addition & 0 deletions test/ext_cuda/runtests.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,6 +4,7 @@ using Zygote
using ForwardDiff: Dual
using Statistics: mean
using CUDA, cuDNN
import CUDA.CUSPARSE: CuSparseMatrixCSC, CuSparseMatrixCSR, CuSparseMatrixCOO
using NNlib: batchnorm, ∇batchnorm
CUDA.allowscalar(false)

Expand Down
4 changes: 2 additions & 2 deletions test/ext_cuda/scatter.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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 @@ -70,7 +70,7 @@ types = [CuArray{Int32}, CuArray{Int64}, CuArray{Float32}, CuArray{Float64}]
end


for T = [CuArray{Float32}, CuArray{Float64}]
for T = [CuArray{Float32}, CuArray{Float64}, Sparse, CuSparseMatrixCSC{Float32}, CuSparseMatrixCSR{Float32}, CuSparseMatrixCOO{Float32}]
@testset "$(T)" begin
@testset "*" begin
for idx = idxs, dims = [0, 1]
Expand Down
Loading