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 82.58%. Comparing base (6d1af96) to head (330289c).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #727      +/-   ##
==========================================
- Coverage   82.62%   82.58%   -0.04%     
==========================================
  Files          62       62              
  Lines        2866     2860       -6     
==========================================
- Hits         2368     2362       -6     
  Misses        498      498              

☔ 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: 330289c	Previous: 6d1af96	Ratio
latency/precompile	25075880292 ns	25160140042 ns	1.00
latency/ttfp	2279012750 ns	2277207000 ns	1.00
latency/import	1450336541 ns	1442078125 ns	1.01
integration/metaldevrt	872083 ns	867958 ns	1.00
integration/byval/slices=1	1569667 ns	1566500 ns	1.00
integration/byval/slices=3	8935333 ns	8820791.5 ns	1.01
integration/byval/reference	1559042 ns	1558125 ns	1.00
integration/byval/slices=2	2602916 ns	2640459 ns	0.99
kernel/indexing	618583 ns	612292 ns	1.01
kernel/indexing_checked	612896 ns	616562.5 ns	0.99
kernel/launch	11542 ns	13250 ns	0.87
kernel/rand	571833 ns	569542 ns	1.00
array/construct	6375 ns	6667 ns	0.96
array/broadcast	604666.5 ns	611708 ns	0.99
array/random/randn/Float32	941916 ns	846208 ns	1.11
array/random/randn!/Float32	747792 ns	627625 ns	1.19
array/random/rand!/Int64	555104.5 ns	564792 ns	0.98
array/random/rand!/Float32	579958 ns	588250 ns	0.99
array/random/rand/Int64	776584 ns	757521 ns	1.03
array/random/rand/Float32	606208 ns	621000 ns	0.98
array/accumulate/Int64/1d	1249416 ns	1217584 ns	1.03
array/accumulate/Int64/dims=1	1828583 ns	1838104 ns	0.99
array/accumulate/Int64/dims=2	2153083 ns	2176020.5 ns	0.99
array/accumulate/Int64/dims=1L	11661062.5 ns	11497375 ns	1.01
array/accumulate/Int64/dims=2L	9777520.5 ns	9787125 ns	1.00
array/accumulate/Float32/1d	1124791.5 ns	1135000 ns	0.99
array/accumulate/Float32/dims=1	1558667 ns	1574459 ns	0.99
array/accumulate/Float32/dims=2	1878250 ns	1885208.5 ns	1.00
array/accumulate/Float32/dims=1L	9815521 ns	9860583 ns	1.00
array/accumulate/Float32/dims=2L	7210958.5 ns	7186084 ns	1.00
array/reductions/reduce/Int64/1d	1347500 ns	1380521 ns	0.98
array/reductions/reduce/Int64/dims=1	1083791 ns	1100083 ns	0.99
array/reductions/reduce/Int64/dims=2	1137333.5 ns	1249709 ns	0.91
array/reductions/reduce/Int64/dims=1L	2028125 ns	2008333 ns	1.01
array/reductions/reduce/Int64/dims=2L	4237250 ns	4231042 ns	1.00
array/reductions/reduce/Float32/1d	1005333 ns	1068333 ns	0.94
array/reductions/reduce/Float32/dims=1	829625 ns	841958 ns	0.99
array/reductions/reduce/Float32/dims=2	836000 ns	863417 ns	0.97
array/reductions/reduce/Float32/dims=1L	1326166.5 ns	1293958 ns	1.02
array/reductions/reduce/Float32/dims=2L	1798917 ns	1858375 ns	0.97
array/reductions/mapreduce/Int64/1d	1547500 ns	1378416 ns	1.12
array/reductions/mapreduce/Int64/dims=1	1090375 ns	1097479 ns	0.99
array/reductions/mapreduce/Int64/dims=2	1302625 ns	1146146 ns	1.14
array/reductions/mapreduce/Int64/dims=1L	1998396 ns	2047937 ns	0.98
array/reductions/mapreduce/Int64/dims=2L	3610958 ns	3644979.5 ns	0.99
array/reductions/mapreduce/Float32/1d	1047709 ns	1002770.5 ns	1.04
array/reductions/mapreduce/Float32/dims=1	826459 ns	843292 ns	0.98
array/reductions/mapreduce/Float32/dims=2	846542 ns	863750 ns	0.98
array/reductions/mapreduce/Float32/dims=1L	1327042 ns	1323667 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	1805875 ns	1826250 ns	0.99
array/private/copyto!/gpu_to_gpu	636000 ns	631000 ns	1.01
array/private/copyto!/cpu_to_gpu	801479.5 ns	766666 ns	1.05
array/private/copyto!/gpu_to_cpu	796083 ns	796292 ns	1.00
array/private/iteration/findall/int	1567458.5 ns	1584104.5 ns	0.99
array/private/iteration/findall/bool	1408333 ns	1407375 ns	1.00
array/private/iteration/findfirst/int	2084542 ns	2078145.5 ns	1.00
array/private/iteration/findfirst/bool	2046458.5 ns	2053875 ns	1.00
array/private/iteration/scalar	4152458 ns	4242792 ns	0.98
array/private/iteration/logical	2652395.5 ns	2673833.5 ns	0.99
array/private/iteration/findmin/1d	2511833 ns	2539375 ns	0.99
array/private/iteration/findmin/2d	1818542 ns	1823250 ns	1.00
array/private/copy	563333 ns	600125 ns	0.94
array/shared/copyto!/gpu_to_gpu	81833 ns	83916.5 ns	0.98
array/shared/copyto!/cpu_to_gpu	81292 ns	83208 ns	0.98
array/shared/copyto!/gpu_to_cpu	82875 ns	82667 ns	1.00
array/shared/iteration/findall/int	1533416 ns	1586791.5 ns	0.97
array/shared/iteration/findall/bool	1389958 ns	1416458 ns	0.98
array/shared/iteration/findfirst/int	1635166 ns	1674583 ns	0.98
array/shared/iteration/findfirst/bool	1628208 ns	1655708 ns	0.98
array/shared/iteration/scalar	191458.5 ns	208208 ns	0.92
array/shared/iteration/logical	2150604.5 ns	2272208 ns	0.95
array/shared/iteration/findmin/1d	2105125 ns	2145542 ns	0.98
array/shared/iteration/findmin/2d	1804042 ns	1829583 ns	0.99
array/shared/copy	260229 ns	248416 ns	1.05
array/permutedims/4d	2392625 ns	2399458 ns	1.00
array/permutedims/2d	1159334 ns	1179542 ns	0.98
array/permutedims/3d	1671250 ns	1689042 ns	0.99
metal/synchronization/stream	19375 ns	19104.5 ns	1.01
metal/synchronization/context	20042 ns	20334 ns	0.99

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?

[rveltz]

Yes but the rand is buggy so is it a good comparison?

[christiangnrd]

Yes but the rand is buggy so is it a good comparison?

I was under the assumption that the issue (yours #474) this is attempting to fix only happens with the normally distributed randn and not uniformly distributed rand.

[rveltz]

you are right sorry.