fix: correct subtyping for sparse matrices in cuSPARSE

[gdalle]

Fixes #3104

[gdalle]

Should I add tests?

[kshyatt]

Is this going to break the broadcasting tests, which rely on things being subtypes of the abstract GPUArrays types?

[gdalle]

No, this shouldn't be breaking because AbstractCuSparseMatrix itself subtypes AbstractGPUSparseMatrix

[kshyatt]

That wasn't what I asked. Is it going to break this logic?

[gdalle]

If I understand correctly, this line still applies to CuSparseMatrixCSC/CSR so all should be good?

https://github.com/JuliaGPU/GPUArrays.jl/blob/21d6bc6f3c3c8dc2a1d2dbd07c5eb177fdaf179d/src/host/sparse.jl#L290

[kshyatt]

I don't think everything will be fine because of this

[gdalle]

Oh right I understand now. Yeah, this will break.
I opened #3104 because a recent version of CUDA broke SparseMatrixColorings, where I assumed that all CUDA sparse matrices are subtypes of AbstractCuSparseMatrix (some even have the subtyping in their docstring by mistake). If that is no longer true because the CSC/CSR distinction happens at the GPUArrays level and we prefer safeguarding that hierarchy, what is AbstractCuSparseMatrix even for?

Missed that AbstractCuSparseMatrix doesn't have AbstractGPUSparseMatrixCSR in its hierarchy.

[kshyatt]

Right now it's kind of functioning as a dumping ground for things that aren't CSC and CSR 😅 . Ideally what we'd do is add broadcasting support for COO and BSR so we can fold them in to the wider AbstractGPUSparseMatrix*** system

[gdalle]

Okay then feel free to close this PR, and I'll make a note to not use AbstractCuSparseMatrix anymore (see JuliaDiff/SparseMatrixColorings.jl#314)

[kshyatt]

Sorry for the confusion here! It's kind of an ongoing process, as you can see...

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: 2c11d9a	Previous: af6961a	Ratio
array/accumulate/Float32/1d	101885 ns	101502.5 ns	1.00
array/accumulate/Float32/dims=1	77180 ns	76629.5 ns	1.01
array/accumulate/Float32/dims=1L	1586558 ns	1585691.5 ns	1.00
array/accumulate/Float32/dims=2	144533.5 ns	143840 ns	1.00
array/accumulate/Float32/dims=2L	658693.5 ns	657866.5 ns	1.00
array/accumulate/Int64/1d	119291 ns	119085 ns	1.00
array/accumulate/Int64/dims=1	80352 ns	79723 ns	1.01
array/accumulate/Int64/dims=1L	1695194.5 ns	1694532 ns	1.00
array/accumulate/Int64/dims=2	156647 ns	156114 ns	1.00
array/accumulate/Int64/dims=2L	962826 ns	961892 ns	1.00
array/broadcast	20872 ns	20629 ns	1.01
array/construct	1267 ns	1256.5 ns	1.01
array/copy	18257 ns	17886 ns	1.02
array/copyto!/cpu_to_gpu	218082 ns	212623 ns	1.03
array/copyto!/gpu_to_cpu	283779.5 ns	283142 ns	1.00
array/copyto!/gpu_to_gpu	10733 ns	10702 ns	1.00
array/iteration/findall/bool	135007 ns	134713 ns	1.00
array/iteration/findall/int	150383.5 ns	149574 ns	1.01
array/iteration/findfirst/bool	82540 ns	81484 ns	1.01
array/iteration/findfirst/int	84719 ns	83536 ns	1.01
array/iteration/findmin/1d	88840 ns	85967.5 ns	1.03
array/iteration/findmin/2d	117471 ns	116482 ns	1.01
array/iteration/logical	200391 ns	199753 ns	1.00
array/iteration/scalar	67244 ns	67105.5 ns	1.00
array/permutedims/2d	52813 ns	52568 ns	1.00
array/permutedims/3d	53243 ns	53057 ns	1.00
array/permutedims/4d	52187 ns	51868.5 ns	1.01
array/random/rand/Float32	13794 ns	12842 ns	1.07
array/random/rand/Int64	26019 ns	25396 ns	1.02
array/random/rand!/Float32	8899.666666666666 ns	8414.666666666666 ns	1.06
array/random/rand!/Int64	22129 ns	21981 ns	1.01
array/random/randn/Float32	37987 ns	37382.5 ns	1.02
array/random/randn!/Float32	30899 ns	30580 ns	1.01
array/reductions/mapreduce/Float32/1d	34669 ns	34346 ns	1.01
array/reductions/mapreduce/Float32/dims=1	39945 ns	39918.5 ns	1.00
array/reductions/mapreduce/Float32/dims=1L	51321 ns	51509 ns	1.00
array/reductions/mapreduce/Float32/dims=2	56515 ns	56563 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	69450.5 ns	69236 ns	1.00
array/reductions/mapreduce/Int64/1d	43066 ns	42783 ns	1.01
array/reductions/mapreduce/Int64/dims=1	51270 ns	42869 ns	1.20
array/reductions/mapreduce/Int64/dims=1L	87296 ns	87545 ns	1.00
array/reductions/mapreduce/Int64/dims=2	59402 ns	59325 ns	1.00
array/reductions/mapreduce/Int64/dims=2L	84700 ns	85131 ns	0.99
array/reductions/reduce/Float32/1d	34994 ns	34338 ns	1.02
array/reductions/reduce/Float32/dims=1	49664 ns	48895 ns	1.02
array/reductions/reduce/Float32/dims=1L	51712 ns	51567 ns	1.00
array/reductions/reduce/Float32/dims=2	56672 ns	56622 ns	1.00
array/reductions/reduce/Float32/dims=2L	69967 ns	69625 ns	1.00
array/reductions/reduce/Int64/1d	42893 ns	42697 ns	1.00
array/reductions/reduce/Int64/dims=1	42295 ns	47176.5 ns	0.90
array/reductions/reduce/Int64/dims=1L	87293 ns	87361 ns	1.00
array/reductions/reduce/Int64/dims=2	59586 ns	59337 ns	1.00
array/reductions/reduce/Int64/dims=2L	84606 ns	84717 ns	1.00
array/reverse/1d	17873 ns	18051 ns	0.99
array/reverse/1dL	68581 ns	68628 ns	1.00
array/reverse/1dL_inplace	65762 ns	65835 ns	1.00
array/reverse/1d_inplace	10334 ns	8676.333333333334 ns	1.19
array/reverse/2d	20742 ns	20541 ns	1.01
array/reverse/2dL	72870 ns	72496 ns	1.01
array/reverse/2dL_inplace	65853 ns	65804 ns	1.00
array/reverse/2d_inplace	9986 ns	10104 ns	0.99
array/sorting/1d	2736964 ns	2734890.5 ns	1.00
array/sorting/2d	1070069 ns	1069333 ns	1.00
array/sorting/by	3305529.5 ns	3305442 ns	1.00
cuda/synchronization/context/auto	1143.3 ns	1142.4 ns	1.00
cuda/synchronization/context/blocking	923.5405405405405 ns	918.9428571428572 ns	1.01
cuda/synchronization/context/nonblocking	7128.8 ns	7080.9 ns	1.01
cuda/synchronization/stream/auto	1001.5714285714286 ns	983.3846153846154 ns	1.02
cuda/synchronization/stream/blocking	823.6436781609195 ns	819.060606060606 ns	1.01
cuda/synchronization/stream/nonblocking	7053.8 ns	7951.1 ns	0.89
integration/byval/reference	143725 ns	143938 ns	1.00
integration/byval/slices=1	145783 ns	145713 ns	1.00
integration/byval/slices=2	284767 ns	284653 ns	1.00
integration/byval/slices=3	422924 ns	422978 ns	1.00
integration/cudadevrt	102332 ns	102487 ns	1.00
integration/volumerhs	23447510 ns	23490453 ns	1.00
kernel/indexing	13255 ns	13409 ns	0.99
kernel/indexing_checked	13922 ns	14040 ns	0.99
kernel/launch	2139.4444444444443 ns	2267.4444444444443 ns	0.94
kernel/occupancy	706.8611111111111 ns	828.1463414634146 ns	0.85
kernel/rand	14503 ns	14466 ns	1.00
latency/import	3805709559 ns	3809615751 ns	1.00
latency/precompile	4569868964 ns	4560767959.5 ns	1.00
latency/ttfp	4380765383 ns	4401798424 ns	1.00

This comment was automatically generated by workflow using github-action-benchmark.