[CUSPARSE] create slices of sparse matrices using boolean masks

[hexaeder]

Slicing into a sparse matrix by providing Vector{Bool} per axis to select a subset of rows and cols is quite usefull. It is implemented for several types, for example "normal" matrices and sparse matrices.

N = 10
A = sprand(N, N, 0.5)
rowmask = [rand([true, false]) for _ in 1:size(A, 1)]
colmask = [rand([true, false]) for _ in 1:size(A, 2)]
S = A[rowmask, colmask]

This PR implements this functionality for CuSparseMatrixCSR and CuSparseMatrixCSC.

AI Disclaimer: Claude code and Chat GPT helped me to implement the kernels. But I think they are straight forward and it is tested against the slicing results of CPU code.

[github-actions]

Your PR no longer requires formatting changes. Thank you for your contribution!

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: aabf559	Previous: 01a0795	Ratio
array/accumulate/Float32/1d	100441 ns	101155 ns	0.99
array/accumulate/Float32/dims=1	76384 ns	76811 ns	0.99
array/accumulate/Float32/dims=1L	1583662.5 ns	1585181 ns	1.00
array/accumulate/Float32/dims=2	143158.5 ns	143761.5 ns	1.00
array/accumulate/Float32/dims=2L	656653.5 ns	657146 ns	1.00
array/accumulate/Int64/1d	118378 ns	118686 ns	1.00
array/accumulate/Int64/dims=1	79220 ns	79898.5 ns	0.99
array/accumulate/Int64/dims=1L	1694073 ns	1693836 ns	1.00
array/accumulate/Int64/dims=2	155363 ns	155970.5 ns	1.00
array/accumulate/Int64/dims=2L	961384 ns	961435.5 ns	1.00
array/broadcast	20447 ns	20595 ns	0.99
array/construct	1300.35 ns	1360 ns	0.96
array/copy	18749 ns	18770 ns	1.00
array/copyto!/cpu_to_gpu	213448 ns	214074.5 ns	1.00
array/copyto!/gpu_to_cpu	281725 ns	282538.5 ns	1.00
array/copyto!/gpu_to_gpu	11352 ns	11352 ns	1
array/iteration/findall/bool	131262 ns	131993 ns	0.99
array/iteration/findall/int	147834 ns	148953 ns	0.99
array/iteration/findfirst/bool	81393 ns	81399 ns	1.00
array/iteration/findfirst/int	83708 ns	83388 ns	1.00
array/iteration/findmin/1d	86996 ns	87034 ns	1.00
array/iteration/findmin/2d	117120 ns	116982 ns	1.00
array/iteration/logical	199699 ns	200300.5 ns	1.00
array/iteration/scalar	67377 ns	66902 ns	1.01
array/permutedims/2d	52720 ns	52557 ns	1.00
array/permutedims/3d	53191.5 ns	52858 ns	1.01
array/permutedims/4d	52327 ns	51891 ns	1.01
array/random/rand/Float32	12947 ns	12938 ns	1.00
array/random/rand/Int64	37236 ns	30232 ns	1.23
array/random/rand!/Float32	8517 ns	8486.333333333334 ns	1.00
array/random/rand!/Int64	26764 ns	34141 ns	0.78
array/random/randn/Float32	41380 ns	37762.5 ns	1.10
array/random/randn!/Float32	26645 ns	31561 ns	0.84
array/reductions/mapreduce/Float32/1d	35411 ns	34932 ns	1.01
array/reductions/mapreduce/Float32/dims=1	40167 ns	39395 ns	1.02
array/reductions/mapreduce/Float32/dims=1L	51692 ns	52264 ns	0.99
array/reductions/mapreduce/Float32/dims=2	56570 ns	56993 ns	0.99
array/reductions/mapreduce/Float32/dims=2L	69206 ns	69821.5 ns	0.99
array/reductions/mapreduce/Int64/1d	42851 ns	42986 ns	1.00
array/reductions/mapreduce/Int64/dims=1	48347 ns	42873 ns	1.13
array/reductions/mapreduce/Int64/dims=1L	87721 ns	87741 ns	1.00
array/reductions/mapreduce/Int64/dims=2	59691 ns	60047 ns	0.99
array/reductions/mapreduce/Int64/dims=2L	84689.5 ns	85349 ns	0.99
array/reductions/reduce/Float32/1d	35715.5 ns	35002.5 ns	1.02
array/reductions/reduce/Float32/dims=1	49584 ns	39743 ns	1.25
array/reductions/reduce/Float32/dims=1L	52037 ns	51984 ns	1.00
array/reductions/reduce/Float32/dims=2	56907 ns	57086 ns	1.00
array/reductions/reduce/Float32/dims=2L	70059 ns	69836.5 ns	1.00
array/reductions/reduce/Int64/1d	42850.5 ns	43090 ns	0.99
array/reductions/reduce/Int64/dims=1	42436 ns	42912 ns	0.99
array/reductions/reduce/Int64/dims=1L	87624 ns	87746 ns	1.00
array/reductions/reduce/Int64/dims=2	59926 ns	59769 ns	1.00
array/reductions/reduce/Int64/dims=2L	84711 ns	84712 ns	1.00
array/reverse/1d	18236 ns	18240 ns	1.00
array/reverse/1dL	68886 ns	68915 ns	1.00
array/reverse/1dL_inplace	65913 ns	65875 ns	1.00
array/reverse/1d_inplace	8529 ns	10243 ns	0.83
array/reverse/2d	20748 ns	20594 ns	1.01
array/reverse/2dL	73039 ns	72723 ns	1.00
array/reverse/2dL_inplace	65994 ns	65965 ns	1.00
array/reverse/2d_inplace	10220 ns	10404 ns	0.98
array/sorting/1d	2735313 ns	2734392 ns	1.00
array/sorting/2d	1072679 ns	1075205 ns	1.00
array/sorting/by	3305049 ns	3315166 ns	1.00
cuda/synchronization/context/auto	1174.7 ns	1168.9 ns	1.00
cuda/synchronization/context/blocking	930.6888888888889 ns	934.3225806451613 ns	1.00
cuda/synchronization/context/nonblocking	7234.700000000001 ns	8363.400000000001 ns	0.87
cuda/synchronization/stream/auto	1021.5882352941177 ns	1040.1333333333334 ns	0.98
cuda/synchronization/stream/blocking	787.747572815534 ns	841.1473684210526 ns	0.94
cuda/synchronization/stream/nonblocking	7399.6 ns	8156.4 ns	0.91
integration/byval/reference	143949 ns	144104 ns	1.00
integration/byval/slices=1	145864 ns	145896 ns	1.00
integration/byval/slices=2	284558 ns	284543 ns	1.00
integration/byval/slices=3	423124 ns	423283 ns	1.00
integration/cudadevrt	102623 ns	102484 ns	1.00
integration/volumerhs	23467460.5 ns	23421671.5 ns	1.00
kernel/indexing	13254 ns	13338 ns	0.99
kernel/indexing_checked	14088.5 ns	14091 ns	1.00
kernel/launch	2191.1111111111113 ns	2181.166666666667 ns	1.00
kernel/occupancy	722.158940397351 ns	705.1118421052631 ns	1.02
kernel/rand	17967 ns	15255 ns	1.18
latency/import	3812112212 ns	3822749737.5 ns	1.00
latency/precompile	4595809169 ns	4597359574 ns	1.00
latency/ttfp	4400844606.5 ns	4411959980 ns	1.00

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

[hexaeder]

The new tests uncovered a CSC to CSR conversion error in CUDA runtime 12.0 (fixed in 12.1 and above).
The colVal stored in the CSR matrix is just wrong, which messes printout, indexing and everything.
Should this be tracked or is it not usefull since it appears to be an upstream bug of that CUDA version?

I updated the tests to go through the COO format which hopefully avoids the buggy codepath.

MWE on Runtime 12.0

# version 12.0
julia> using CUDA, SparseArrays, CUDA.CUSPARSE

julia> A = sparse([1,1,1,1,2], [1,2,4,5,3], [1.0,2.0,3.0,4.0,5.0], 2, 5)
2×5 SparseMatrixCSC{Float64, Int64} with 5 stored entries:
 1.0  2.0   ⋅   3.0  4.0
  ⋅    ⋅   5.0   ⋅    ⋅

julia> dA = CuSparseMatrixCSR(A) # wrong!
2×5 CuSparseMatrixCSR{Float64, Int32} with 5 stored entries:
 1.0  2.0  5.0  3.0   ⋅
  ⋅    ⋅    ⋅    ⋅   4.0

julia> CUDA.@allowscalar dA[1,1] # wrong!
0.0

MWE on Runtime 12.1

# version 12.1
julia> using CUDA, SparseArrays, CUDA.CUSPARSE

julia> A = sparse([1,1,1,1,2], [1,2,4,5,3], [1.0,2.0,3.0,4.0,5.0], 2, 5)
2×5 SparseMatrixCSC{Float64, Int64} with 5 stored entries:
 1.0  2.0   ⋅   3.0  4.0
  ⋅    ⋅   5.0   ⋅    ⋅

julia> dA = CuSparseMatrixCSR(A) # correct!
2×5 CuSparseMatrixCSR{Float64, Int32} with 5 stored entries:
 1.0  2.0   ⋅   3.0  4.0
  ⋅    ⋅   5.0   ⋅    ⋅

julia> CUDA.@allowscalar dA[1,1] # correct!
1.0

[codecov]

Codecov Report

❌ Patch coverage is 80.71429% with 27 lines in your changes missing coverage. Please review.
✅ Project coverage is 90.32%. Comparing base (01a0795) to head (aabf559).
⚠️ Report is 1 commits behind head on master.

Files with missing lines	Patch %	Lines
lib/cusparse/src/array.jl	49.05%	27 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #3032      +/-   ##
==========================================
- Coverage   90.43%   90.32%   -0.12%     
==========================================
  Files         141      141              
  Lines       12025    12165     +140     
==========================================
+ Hits        10875    10988     +113     
- Misses       1150     1177      +27     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[hexaeder]

Gentle bump on this. Is there something I can do to support review?

EDIT: I rebased, github UI was showing some apparent merge conflicts without showing my where the problems lie.

[kshyatt]

Looks like this might need some TLC to get up to date with the (unreleased) CUDA.jl 6 and package splitting

[hexaeder]

I think the problems related to the Library split are fixed (just some CUDA -> CUDACore renamings). There are three remaning test fails I don't quite understand but seem unrelated?

• Multi GPU Tests failed on core/array tests
• Julia Nightly fails because of out of bounds indexing in core/array test
• Enzyme tests fail

[maleadt]

Nightly and Enzyme are allowed to fail, and the multiGPU one is intermittent. So no relevant CI failures.

[maleadt]

LGTM. Too bad about the scalar getindexes that essentially synchronize the GPU before doing any work, but I'm not familiar with the domain to suggest alternatives here.