Tests for descriptors

[kshyatt]

No description provided.

[github-actions]

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

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: f0f26be	Previous: d223d48	Ratio
latency/precompile	45875773921 ns	45641120785 ns	1.01
latency/ttfp	6494221270.5 ns	6558560140 ns	0.99
latency/import	3171040795 ns	3191866733 ns	0.99
integration/volumerhs	9628852.5 ns	9611544 ns	1.00
integration/byval/slices=1	147055 ns	146796 ns	1.00
integration/byval/slices=3	425122 ns	425150 ns	1.00
integration/byval/reference	145016 ns	145127 ns	1.00
integration/byval/slices=2	286279 ns	286038 ns	1.00
integration/cudadevrt	103422 ns	103471 ns	1.00
kernel/indexing	14256 ns	14169 ns	1.01
kernel/indexing_checked	14669 ns	14963 ns	0.98
kernel/occupancy	673.1834319526628 ns	657.8086419753087 ns	1.02
kernel/launch	2082.95 ns	2072.7 ns	1.00
kernel/rand	14702 ns	16123 ns	0.91
array/reverse/1d	19141 ns	19951 ns	0.96
array/reverse/2d	25088 ns	25142 ns	1.00
array/reverse/1d_inplace	11327 ns	11456 ns	0.99
array/reverse/2d_inplace	12925 ns	13247.5 ns	0.98
array/copy	20990 ns	21071 ns	1.00
array/iteration/findall/int	158173 ns	157380 ns	1.01
array/iteration/findall/bool	139025 ns	138848.5 ns	1.00
array/iteration/findfirst/int	154294.5 ns	154329 ns	1.00
array/iteration/findfirst/bool	155161 ns	154809.5 ns	1.00
array/iteration/scalar	72703 ns	70606 ns	1.03
array/iteration/logical	213855.5 ns	214919.5 ns	1.00
array/iteration/findmin/1d	41651 ns	41444 ns	1.00
array/iteration/findmin/2d	93829.5 ns	94472 ns	0.99
array/reductions/reduce/1d	43150.5 ns	35629 ns	1.21
array/reductions/reduce/2d	51433 ns	50066 ns	1.03
array/reductions/mapreduce/1d	39035 ns	33606 ns	1.16
array/reductions/mapreduce/2d	48288 ns	43968.5 ns	1.10
array/broadcast	20797 ns	20847 ns	1.00
array/copyto!/gpu_to_gpu	11879 ns	11929 ns	1.00
array/copyto!/cpu_to_gpu	208979 ns	209164.5 ns	1.00
array/copyto!/gpu_to_cpu	243581 ns	243460 ns	1.00
array/accumulate/1d	108404 ns	109349 ns	0.99
array/accumulate/2d	79788.5 ns	80109.5 ns	1.00
array/construct	1310.5 ns	1250 ns	1.05
array/random/randn/Float32	44854 ns	43560 ns	1.03
array/random/randn!/Float32	26417 ns	26514 ns	1.00
array/random/rand!/Int64	27264 ns	27315 ns	1.00
array/random/rand!/Float32	8727.666666666666 ns	8798 ns	0.99
array/random/rand/Int64	29920 ns	30086 ns	0.99
array/random/rand/Float32	13011 ns	13099 ns	0.99
array/permutedims/4d	60894 ns	61425.5 ns	0.99
array/permutedims/2d	55591 ns	55688 ns	1.00
array/permutedims/3d	56320 ns	56210.5 ns	1.00
array/sorting/1d	2776432 ns	2777197.5 ns	1.00
array/sorting/by	3367294 ns	3368328 ns	1.00
array/sorting/2d	1084782 ns	1084987 ns	1.00
cuda/synchronization/stream/auto	1044.6 ns	1036.1 ns	1.01
cuda/synchronization/stream/nonblocking	6568.8 ns	6554.8 ns	1.00
cuda/synchronization/stream/blocking	841.1627906976744 ns	807.505376344086 ns	1.04
cuda/synchronization/context/auto	1186.5 ns	1182.7 ns	1.00
cuda/synchronization/context/nonblocking	6806.8 ns	6792 ns	1.00
cuda/synchronization/context/blocking	928.1395348837209 ns	912.046511627907 ns	1.02

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

[maleadt]

LGTM, but CI failures are related.

[kshyatt]

Now that I look at it, I think these methods are just invalid. You can't supply a NULL pointer for the vector/matrix values. I think we should just remove these. I'll update the PR.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 83.60%. Comparing base (d223d48) to head (f0f26be).
Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2700      +/-   ##
==========================================
+ Coverage   83.51%   83.60%   +0.08%     
==========================================
  Files         153      153              
  Lines       13606    13592      -14     
==========================================
  Hits        11363    11363              
+ Misses       2243     2229      -14     

☔ 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.

[amontoison]

@kshyatt @maleadt
You broke my package KrylovPreconditioners.jl with these modifications. :(
-> JuliaSmoothOptimizers/KrylovPreconditioners.jl#68

I explained why I did that in a previous PR:
#1611

I can use these constructors to perform the symbolic analyis for SpMV / SPMM / SPSV and SPSM as well as the allocation of the buffers for a given number of right-hand sides.
CUSPARSE doesn't need the content of the right-hand sides to do the symbolic analysis.

[amontoison]

Update: It was these PR where I explained why I did that:
#2159

[amontoison]

It probably broke other of our optimization packages like CUSOLVERRF.jl because we preallocate the buffers for sparse triangular solves.

@maleadt Is it fine to add them back in CUDA.jl and do a new small minor release 5.7.2?

[kshyatt]

Sorry about that! Definitely add them back - is there a way we can add tests too?

[kshyatt]

It would also be good to add a comment inline with the restored methods explaining the above, sorry again, 100% my fault

[amontoison]

Thanks @kshyatt !!
Yes, we could perform the analysis of SPMV / SPMM / SPSV / SPSM and even the preprocess after the creation of the constructors.
We don't do that in CUDA.jl because we don't know to recycle the buffers for multiple mul! / ldiv! but it can be done in the tests.

Example: https://github.com/JuliaSmoothOptimizers/KrylovPreconditioners.jl/blob/79d8a350c9ce7ec3968f41fdf8f8bfcbf3276067/ext/CUDA/operators.jl#L24