Use Metal.jl native rand by default

[christiangnrd]

Close #474

By my very unscientific tests (generating a bunch of values and plotting a histogram), quality seems similar to MPS, but without the NaN generation (#474).

In a future PR, if we really want to default to Apple-provided random number generation when supported, we could wrap the MPSGraph rng functionality but that may incur a performance hit.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 81.15%. Comparing base (67d668c) to head (c3fe826).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #727      +/-   ##
==========================================
+ Coverage   80.90%   81.15%   +0.25%     
==========================================
  Files          59       62       +3     
  Lines        2896     2892       -4     
==========================================
+ Hits         2343     2347       +4     
+ Misses        553      545       -8     

☔ 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]

Metal Benchmarks

Details

Benchmark suite	Current: c3fe826	Previous: 67d668c	Ratio
latency/precompile	24862918625 ns	24820843000 ns	1.00
latency/ttfp	2268736334 ns	2257593833 ns	1.00
latency/import	1438339834 ns	1431203750 ns	1.00
integration/metaldevrt	832083 ns	834875 ns	1.00
integration/byval/slices=1	1546583 ns	1525666.5 ns	1.01
integration/byval/slices=3	8758312.5 ns	8498958 ns	1.03
integration/byval/reference	1536791.5 ns	1538166 ns	1.00
integration/byval/slices=2	2611292 ns	2552562 ns	1.02
kernel/indexing	568792 ns	593833 ns	0.96
kernel/indexing_checked	579000 ns	575750 ns	1.01
kernel/launch	11625 ns	11250 ns	1.03
kernel/rand	548834 ns	557187.5 ns	0.99
array/construct	6166 ns	6000 ns	1.03
array/broadcast	585000 ns	591209 ns	0.99
array/random/randn/Float32	991229.5 ns	836917 ns	1.18
array/random/randn!/Float32	723750 ns	619542 ns	1.17
array/random/rand!/Int64	576625 ns	548834 ns	1.05
array/random/rand!/Float32	652750 ns	593333 ns	1.10
array/random/rand/Int64	835500 ns	735667 ns	1.14
array/random/rand/Float32	865021 ns	631792 ns	1.37
array/accumulate/Int64/1d	1241854.5 ns	1237125 ns	1.00
array/accumulate/Int64/dims=1	1799479 ns	1795625 ns	1.00
array/accumulate/Int64/dims=2	2185291.5 ns	2130458 ns	1.03
array/accumulate/Int64/dims=1L	11377104 ns	11609562.5 ns	0.98
array/accumulate/Int64/dims=2L	9754937.5 ns	9610834 ns	1.01
array/accumulate/Float32/1d	1121125 ns	1111187.5 ns	1.01
array/accumulate/Float32/dims=1	1523333 ns	1518146 ns	1.00
array/accumulate/Float32/dims=2	1852250 ns	1836167 ns	1.01
array/accumulate/Float32/dims=1L	9861979 ns	9757375 ns	1.01
array/accumulate/Float32/dims=2L	7196083 ns	7203562.5 ns	1.00
array/reductions/reduce/Int64/1d	1433416 ns	1498333 ns	0.96
array/reductions/reduce/Int64/dims=1	1072292 ns	1076542 ns	1.00
array/reductions/reduce/Int64/dims=2	1111667 ns	1129417 ns	0.98
array/reductions/reduce/Int64/dims=1L	1994917 ns	2002083.5 ns	1.00
array/reductions/reduce/Int64/dims=2L	4212604.5 ns	4214895.5 ns	1.00
array/reductions/reduce/Float32/1d	1006437.5 ns	991375 ns	1.02
array/reductions/reduce/Float32/dims=1	809292 ns	827000 ns	0.98
array/reductions/reduce/Float32/dims=2	842292 ns	833917 ns	1.01
array/reductions/reduce/Float32/dims=1L	1290833 ns	1305125 ns	0.99
array/reductions/reduce/Float32/dims=2L	1793750 ns	1788375 ns	1.00
array/reductions/mapreduce/Int64/1d	1526979 ns	1549292 ns	0.99
array/reductions/mapreduce/Int64/dims=1	1077917 ns	1085333 ns	0.99
array/reductions/mapreduce/Int64/dims=2	1116375 ns	1201959 ns	0.93
array/reductions/mapreduce/Int64/dims=1L	2008000 ns	2019583 ns	0.99
array/reductions/mapreduce/Int64/dims=2L	3606833.5 ns	3628521 ns	0.99
array/reductions/mapreduce/Float32/1d	962250 ns	1036542 ns	0.93
array/reductions/mapreduce/Float32/dims=1	827500 ns	819667 ns	1.01
array/reductions/mapreduce/Float32/dims=2	836708 ns	843917 ns	0.99
array/reductions/mapreduce/Float32/dims=1L	1296187.5 ns	1280500 ns	1.01
array/reductions/mapreduce/Float32/dims=2L	1808292 ns	1784500 ns	1.01
array/private/copyto!/gpu_to_gpu	627583 ns	635375 ns	0.99
array/private/copyto!/cpu_to_gpu	768459 ns	786625 ns	0.98
array/private/copyto!/gpu_to_cpu	784584 ns	773833 ns	1.01
array/private/iteration/findall/int	1572958.5 ns	1620458 ns	0.97
array/private/iteration/findall/bool	1425333.5 ns	1430125 ns	1.00
array/private/iteration/findfirst/int	2096209 ns	2024937.5 ns	1.04
array/private/iteration/findfirst/bool	2024396 ns	2010916 ns	1.01
array/private/iteration/scalar	3959334 ns	5600375 ns	0.71
array/private/iteration/logical	2637208 ns	2504521 ns	1.05
array/private/iteration/findmin/1d	2224729.5 ns	2209917 ns	1.01
array/private/iteration/findmin/2d	1514292 ns	1498584 ns	1.01
array/private/copy	569917 ns	558312.5 ns	1.02
array/shared/copyto!/gpu_to_gpu	83917 ns	82042 ns	1.02
array/shared/copyto!/cpu_to_gpu	81083 ns	79750 ns	1.02
array/shared/copyto!/gpu_to_cpu	81625 ns	82125 ns	0.99
array/shared/iteration/findall/int	1574833.5 ns	1600354 ns	0.98
array/shared/iteration/findall/bool	1427250 ns	1452458 ns	0.98
array/shared/iteration/findfirst/int	1633042 ns	1621520.5 ns	1.01
array/shared/iteration/findfirst/bool	1622396 ns	1607916.5 ns	1.01
array/shared/iteration/scalar	208792 ns	202916 ns	1.03
array/shared/iteration/logical	2241354 ns	2386416.5 ns	0.94
array/shared/iteration/findmin/1d	1823708 ns	1799396 ns	1.01
array/shared/iteration/findmin/2d	1520042 ns	1500416.5 ns	1.01
array/shared/copy	246375 ns	230791 ns	1.07
array/permutedims/4d	2368666 ns	2358000 ns	1.00
array/permutedims/2d	1137229.5 ns	1133208 ns	1.00
array/permutedims/3d	1664041 ns	1645604 ns	1.01
metal/synchronization/stream	18833 ns	18500 ns	1.02
metal/synchronization/context	19709 ns	19625 ns	1.00

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

[christiangnrd]

There seems to be a bit of a performance hit with uniform rand, maybe we keep using MPS for those since they weren't causing issues?