Add a dispatch for LinearAlgebra.norm2

[sharanry]

norm(@view x[..], 2) was previously leading to a call of LinearAlgebra.generic_norm2 which led to a scalar indexing. This catches such cuda subarray norm2 calls earlier.

Inf-norm and p-norm with cuda subarrays still lead to the following dispatches:

LinearAlgebra.generic_normInf(x) = float(mapreduce(norm, max, x))
LinearAlgebra.generic_norm1(x) = mapreduce(float ∘ norm, +, x)

I am not sure if there is a better way to dispatch the above.

should resolve #2280

[maleadt]

What about generalizing the LinearAlgebra.norm method above to StridedCuArray? That seems cleaner than overriding an internal method.

[Azercoco]

Hi, what's the status of this PR ? This issue is troublemsome for one of my code and I would like to know if the fix will be implemented into CUDA.jl

[maleadt]

The PR fails CI, and there's an outstanding comment of mine, so it needs work I'd say. Feel free to take it up if you want.

[maleadt]

Rebased. Depends on JuliaGPU/GPUArrays.jl#720 now.

[codecov]

Codecov Report

❌ Patch coverage is 66.66667% with 1 line in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (main@e264549). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
lib/cublas/src/linalg.jl	66.66%	1 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2302   +/-   ##
=======================================
  Coverage        ?   16.39%           
=======================================
  Files           ?      124           
  Lines           ?     9827           
  Branches        ?        0           
=======================================
  Hits            ?     1611           
  Misses          ?     8216           
  Partials        ?        0           

☔ 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.

[github-actions]

CUDA.jl Benchmarks

Details

Benchmark suite	Current: 156a51a	Previous: 2ace55f	Ratio
array/accumulate/Float32/1d	101544 ns	100927 ns	1.01
array/accumulate/Float32/dims=1	76473 ns	77244 ns	0.99
array/accumulate/Float32/dims=1L	1585314 ns	1586192 ns	1.00
array/accumulate/Float32/dims=2	143576 ns	144286 ns	1.00
array/accumulate/Float32/dims=2L	657193.5 ns	658394 ns	1.00
array/accumulate/Int64/1d	118474 ns	118643 ns	1.00
array/accumulate/Int64/dims=1	79908 ns	79829 ns	1.00
array/accumulate/Int64/dims=1L	1706190.5 ns	1694383.5 ns	1.01
array/accumulate/Int64/dims=2	156617 ns	156179.5 ns	1.00
array/accumulate/Int64/dims=2L	961814 ns	961974 ns	1.00
array/broadcast	20524 ns	20478 ns	1.00
array/construct	1257.05 ns	1277.6 ns	0.98
array/copy	17916 ns	18044 ns	0.99
array/copyto!/cpu_to_gpu	213898 ns	215225 ns	0.99
array/copyto!/gpu_to_cpu	281245 ns	282660 ns	0.99
array/copyto!/gpu_to_gpu	10695 ns	10909 ns	0.98
array/iteration/findall/bool	134649 ns	134548 ns	1.00
array/iteration/findall/int	148368 ns	149686 ns	0.99
array/iteration/findfirst/bool	81151 ns	81331 ns	1.00
array/iteration/findfirst/int	83416.5 ns	83732 ns	1.00
array/iteration/findmin/1d	85341 ns	86429 ns	0.99
array/iteration/findmin/2d	114129.5 ns	117614.5 ns	0.97
array/iteration/logical	200491.5 ns	199724.5 ns	1.00
array/iteration/scalar	67792 ns	68306 ns	0.99
array/permutedims/2d	52319.5 ns	52646 ns	0.99
array/permutedims/3d	52444 ns	52872 ns	0.99
array/permutedims/4d	51338 ns	51484 ns	1.00
array/random/rand/Float32	13335 ns	12708 ns	1.05
array/random/rand/Int64	24379 ns	25198 ns	0.97
array/random/rand!/Float32	9851.666666666666 ns	8737.333333333334 ns	1.13
array/random/rand!/Int64	21090 ns	22022 ns	0.96
array/random/randn/Float32	43186 ns	42861 ns	1.01
array/random/randn!/Float32	30810 ns	30594 ns	1.01
array/reductions/mapreduce/Float32/1d	34712 ns	34442 ns	1.01
array/reductions/mapreduce/Float32/dims=1	40828 ns	39651 ns	1.03
array/reductions/mapreduce/Float32/dims=1L	51311 ns	51115 ns	1.00
array/reductions/mapreduce/Float32/dims=2	58180 ns	56288.5 ns	1.03
array/reductions/mapreduce/Float32/dims=2L	67774 ns	69317 ns	0.98
array/reductions/mapreduce/Int64/1d	42789.5 ns	42454.5 ns	1.01
array/reductions/mapreduce/Int64/dims=1	51811 ns	50760 ns	1.02
array/reductions/mapreduce/Int64/dims=1L	87260 ns	87043.5 ns	1.00
array/reductions/mapreduce/Int64/dims=2	60671.5 ns	59370.5 ns	1.02
array/reductions/mapreduce/Int64/dims=2L	84038 ns	84559 ns	0.99
array/reductions/reduce/Float32/1d	34945 ns	34787 ns	1.00
array/reductions/reduce/Float32/dims=1	39822.5 ns	39944 ns	1.00
array/reductions/reduce/Float32/dims=1L	51236 ns	51284 ns	1.00
array/reductions/reduce/Float32/dims=2	58238 ns	56294 ns	1.03
array/reductions/reduce/Float32/dims=2L	68142 ns	69595 ns	0.98
array/reductions/reduce/Int64/1d	42679 ns	42459 ns	1.01
array/reductions/reduce/Int64/dims=1	43466 ns	50053 ns	0.87
array/reductions/reduce/Int64/dims=1L	87194 ns	87114 ns	1.00
array/reductions/reduce/Int64/dims=2	60654 ns	59331.5 ns	1.02
array/reductions/reduce/Int64/dims=2L	84499.5 ns	84197 ns	1.00
array/reverse/1d	17902 ns	17694 ns	1.01
array/reverse/1dL	68542 ns	68202 ns	1.00
array/reverse/1dL_inplace	65769 ns	65756 ns	1.00
array/reverse/1d_inplace	8541.666666666666 ns	10156.166666666668 ns	0.84
array/reverse/2d	20901 ns	21001 ns	1.00
array/reverse/2dL	73023 ns	73024 ns	1.00
array/reverse/2dL_inplace	65914 ns	65634 ns	1.00
array/reverse/2d_inplace	9973 ns	11107 ns	0.90
array/sorting/1d	2734742 ns	2736491 ns	1.00
array/sorting/2d	1068254 ns	1070402 ns	1.00
array/sorting/by	3303672 ns	3304900 ns	1.00
cuda/synchronization/context/auto	1145.3 ns	1123.1 ns	1.02
cuda/synchronization/context/blocking	921.1923076923077 ns	902.5957446808511 ns	1.02
cuda/synchronization/context/nonblocking	7130.2 ns	7638.8 ns	0.93
cuda/synchronization/stream/auto	992.5625 ns	980.6666666666666 ns	1.01
cuda/synchronization/stream/blocking	833.6666666666666 ns	806.5 ns	1.03
cuda/synchronization/stream/nonblocking	7230.299999999999 ns	7187.8 ns	1.01
integration/byval/reference	143781 ns	143733 ns	1.00
integration/byval/slices=1	145763 ns	145971 ns	1.00
integration/byval/slices=2	284545 ns	284607 ns	1.00
integration/byval/slices=3	423071 ns	423028 ns	1.00
integration/cudadevrt	102317 ns	102291 ns	1.00
integration/volumerhs	23424198.5 ns	23455094 ns	1.00
kernel/indexing	13267 ns	13164 ns	1.01
kernel/indexing_checked	13841 ns	13822 ns	1.00
kernel/launch	2182.3333333333335 ns	2137 ns	1.02
kernel/occupancy	675.5125 ns	674.2788461538462 ns	1.00
kernel/rand	17207 ns	14157 ns	1.22
latency/import	3848898595 ns	3799013573 ns	1.01
latency/precompile	4628935725 ns	4593655026.5 ns	1.01
latency/ttfp	4456320918 ns	4367019984.5 ns	1.02

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