Fix randn/randexp in complex GPU kernels by replacing recursion with loops

[maleadt]

The Ziggurat rejection sampling for randn and randexp used recursive calls back to Random.randn/Random.randexp when retrying, creating call cycles that could exhaust the GPU's limited per-thread stack in complex kernels. Replace these with while loops, using a nothing return from the @noinline unlikely helpers to signal retry.

Fixes #3028

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 90.42%. Comparing base (a9a687c) to head (9d845b7).
⚠️ Report is 5 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #3086   +/-   ##
=======================================
  Coverage   90.41%   90.42%           
=======================================
  Files         141      141           
  Lines       11993    11993           
=======================================
+ Hits        10844    10845    +1     
+ Misses       1149     1148    -1     

☔ 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: 9d845b7	Previous: 5f45772	Ratio
array/accumulate/Float32/1d	101572 ns	101495 ns	1.00
array/accumulate/Float32/dims=1	76429 ns	76898 ns	0.99
array/accumulate/Float32/dims=1L	1583518.5 ns	1585143.5 ns	1.00
array/accumulate/Float32/dims=2	143396 ns	143801 ns	1.00
array/accumulate/Float32/dims=2L	657681 ns	657240 ns	1.00
array/accumulate/Int64/1d	118629 ns	118623 ns	1.00
array/accumulate/Int64/dims=1	79699 ns	80572.5 ns	0.99
array/accumulate/Int64/dims=1L	1694403.5 ns	1693852 ns	1.00
array/accumulate/Int64/dims=2	155719 ns	156484 ns	1.00
array/accumulate/Int64/dims=2L	961838 ns	961603 ns	1.00
array/broadcast	20460 ns	20294 ns	1.01
array/construct	1294.5 ns	1320.4 ns	0.98
array/copy	18764 ns	18780 ns	1.00
array/copyto!/cpu_to_gpu	215869 ns	214684 ns	1.01
array/copyto!/gpu_to_cpu	287011 ns	282072 ns	1.02
array/copyto!/gpu_to_gpu	11520 ns	11361 ns	1.01
array/iteration/findall/bool	132048 ns	131719.5 ns	1.00
array/iteration/findall/int	149637 ns	148883 ns	1.01
array/iteration/findfirst/bool	81505 ns	81470.5 ns	1.00
array/iteration/findfirst/int	83783 ns	83414 ns	1.00
array/iteration/findmin/1d	89907 ns	89419 ns	1.01
array/iteration/findmin/2d	117557 ns	117365 ns	1.00
array/iteration/logical	204362.5 ns	207612 ns	0.98
array/iteration/scalar	67784 ns	66780 ns	1.02
array/permutedims/2d	52396 ns	52471.5 ns	1.00
array/permutedims/3d	52862 ns	53137 ns	0.99
array/permutedims/4d	52673 ns	52429 ns	1.00
array/random/rand/Float32	13221 ns	13089 ns	1.01
array/random/rand/Int64	34713 ns	37236 ns	0.93
array/random/rand!/Float32	8426.333333333334 ns	8527.666666666666 ns	0.99
array/random/rand!/Int64	26934 ns	34109.5 ns	0.79
array/random/randn/Float32	39585 ns	38147 ns	1.04
array/random/randn!/Float32	31408.5 ns	31640 ns	0.99
array/reductions/mapreduce/Float32/1d	35584 ns	34735.5 ns	1.02
array/reductions/mapreduce/Float32/dims=1	40893.5 ns	40760 ns	1.00
array/reductions/mapreduce/Float32/dims=1L	51769.5 ns	51917 ns	1.00
array/reductions/mapreduce/Float32/dims=2	56681 ns	56503.5 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	69676.5 ns	69496.5 ns	1.00
array/reductions/mapreduce/Int64/1d	43300 ns	42820 ns	1.01
array/reductions/mapreduce/Int64/dims=1	44563 ns	44181 ns	1.01
array/reductions/mapreduce/Int64/dims=1L	87913.5 ns	87798 ns	1.00
array/reductions/mapreduce/Int64/dims=2	59744 ns	59808 ns	1.00
array/reductions/mapreduce/Int64/dims=2L	85226 ns	85232 ns	1.00
array/reductions/reduce/Float32/1d	35692 ns	34883 ns	1.02
array/reductions/reduce/Float32/dims=1	40448 ns	39758 ns	1.02
array/reductions/reduce/Float32/dims=1L	52032 ns	52166 ns	1.00
array/reductions/reduce/Float32/dims=2	57113 ns	56925 ns	1.00
array/reductions/reduce/Float32/dims=2L	70240 ns	69909 ns	1.00
array/reductions/reduce/Int64/1d	43041.5 ns	42673 ns	1.01
array/reductions/reduce/Int64/dims=1	46569.5 ns	42123 ns	1.11
array/reductions/reduce/Int64/dims=1L	87663 ns	87782 ns	1.00
array/reductions/reduce/Int64/dims=2	59601 ns	59551 ns	1.00
array/reductions/reduce/Int64/dims=2L	85041 ns	84796 ns	1.00
array/reverse/1d	18568 ns	18432.5 ns	1.01
array/reverse/1dL	69097 ns	69025 ns	1.00
array/reverse/1dL_inplace	65979 ns	65968 ns	1.00
array/reverse/1d_inplace	8624.666666666666 ns	10240.666666666666 ns	0.84
array/reverse/2d	20893 ns	20709 ns	1.01
array/reverse/2dL	73045 ns	72815 ns	1.00
array/reverse/2dL_inplace	66059 ns	65992 ns	1.00
array/reverse/2d_inplace	10249 ns	11117.5 ns	0.92
array/sorting/1d	2735784 ns	2754859 ns	0.99
array/sorting/2d	1069187 ns	1075967 ns	0.99
array/sorting/by	3304331.5 ns	3328240 ns	0.99
cuda/synchronization/context/auto	1202.6 ns	1192.4 ns	1.01
cuda/synchronization/context/blocking	927.060606060606 ns	947.7391304347826 ns	0.98
cuda/synchronization/context/nonblocking	6917.2 ns	7660.1 ns	0.90
cuda/synchronization/stream/auto	1018.2 ns	1032.5 ns	0.99
cuda/synchronization/stream/blocking	831.9883720930233 ns	841.5588235294117 ns	0.99
cuda/synchronization/stream/nonblocking	8341.2 ns	7189.6 ns	1.16
integration/byval/reference	144056 ns	143997 ns	1.00
integration/byval/slices=1	146082 ns	145776 ns	1.00
integration/byval/slices=2	284924 ns	284427 ns	1.00
integration/byval/slices=3	423388 ns	423129 ns	1.00
integration/cudadevrt	102628 ns	102598 ns	1.00
integration/volumerhs	9436436.5 ns	9429742.5 ns	1.00
kernel/indexing	13424 ns	13331 ns	1.01
kernel/indexing_checked	14112 ns	14116 ns	1.00
kernel/launch	2193 ns	2147 ns	1.02
kernel/occupancy	693.86875 ns	660.5723270440252 ns	1.05
kernel/rand	18175 ns	15598 ns	1.17
latency/import	3811113291 ns	3807044359.5 ns	1.00
latency/precompile	4590100515 ns	4590923492 ns	1.00
latency/ttfp	4399572655 ns	4392969126 ns	1.00

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