Switch to the new GPUArrays RNG.

[maleadt]

Adapts to JuliaGPU/GPUArrays.jl#707. The new RNG is generally faster than the MPS one (and much faster than the native one), so switch over for all operations.

Analysis by 🤖 below.

Performance comparison

Apple M-series GPU, Metal.@sync timing, minimum of 30 samples per call.
"ratio" is MPS / GPUArrays — values below 1.0 mean MPS is faster.

Uniform rand!

type	n	ratio	GPUArrays	MPS	Kernel	GA GB/s	MPS GB/s	K GB/s
Float16	1024	—	138.7 μs	—	133.6 μs	0.01	—	0.02
Float16	1048576	—	250.4 μs	—	407.9 μs	8.37	—	5.14
Float16	16777216	—	464.2 μs	—	1.70 ms	72.28	—	19.70
Float32	1024	1.08×	138.7 μs	150.0 μs	128.9 μs	0.03	0.03	0.03
Float32	1048576	1.22×	276.4 μs	336.5 μs	437.6 μs	15.17	12.46	9.59
Float32	16777216	1.32×	600.6 μs	792.5 μs	1.70 ms	111.73	84.68	39.51
UInt8	1024	1.06×	135.4 μs	143.2 μs	134.0 μs	0.01	0.01	0.01
UInt8	1048576	0.88×	225.5 μs	199.4 μs	402.4 μs	4.65	5.26	2.61
UInt8	16777216	1.09×	456.6 μs	496.5 μs	1.69 ms	36.75	33.79	9.90
Int8	1024	1.28×	125.5 μs	160.3 μs	133.5 μs	0.01	0.01	0.01
Int8	1048576	0.85×	268.8 μs	227.6 μs	344.8 μs	3.90	4.61	3.04
Int8	16777216	0.67×	493.5 μs	329.9 μs	1.68 ms	33.99	50.86	9.96
UInt16	1024	0.98×	141.4 μs	139.2 μs	144.4 μs	0.01	0.01	0.01
UInt16	1048576	1.09×	257.8 μs	281.5 μs	428.0 μs	8.13	7.45	4.90
UInt16	16777216	0.77×	620.2 μs	475.2 μs	1.76 ms	54.10	70.62	19.04
Int16	1024	0.81×	166.5 μs	134.6 μs	157.3 μs	0.01	0.02	0.01
Int16	1048576	0.99×	282.3 μs	279.4 μs	445.7 μs	7.43	7.51	4.71
Int16	16777216	0.73×	612.5 μs	448.1 μs	1.69 ms	54.78	74.88	19.86
UInt32	1024	1.06×	135.4 μs	143.9 μs	137.2 μs	0.03	0.03	0.03
UInt32	1048576	1.92×	175.5 μs	337.1 μs	435.6 μs	23.90	12.44	9.63
UInt32	16777216	1.36×	678.2 μs	922.2 μs	1.72 ms	98.96	72.77	38.93
Int32	1024	1.00×	175.0 μs	175.5 μs	168.5 μs	0.02	0.02	0.02
Int32	1048576	1.19×	288.9 μs	342.5 μs	451.4 μs	14.52	12.25	9.29
Int32	16777216	1.25×	640.1 μs	802.3 μs	1.76 ms	104.84	83.64	38.22
UInt64	1024	1.03×	174.9 μs	180.6 μs	161.8 μs	0.05	0.05	0.05
UInt64	1048576	1.33×	394.3 μs	524.9 μs	442.4 μs	21.28	15.98	18.96
UInt64	16777216	1.27×	1.14 ms	1.45 ms	1.76 ms	117.64	92.55	76.09
Int64	1024	0.95×	147.6 μs	140.7 μs	133.8 μs	0.06	0.06	0.06
Int64	1048576	1.31×	372.4 μs	488.9 μs	451.3 μs	22.53	17.16	18.59
Int64	16777216	1.22×	1.18 ms	1.45 ms	1.73 ms	113.56	92.77	77.58

Normal randn!

type	n	ratio	GPUArrays	MPS	Kernel	GA GB/s	MPS GB/s	K GB/s
Float16	1024	—	148.2 μs	—	138.4 μs	0.01	—	0.01
Float16	1048576	—	280.7 μs	—	327.5 μs	7.47	—	6.40
Float16	16777216	—	583.0 μs	—	948.3 μs	57.55	—	35.38
Float32	1024	1.03×	142.7 μs	146.8 μs	135.3 μs	0.03	0.03	0.03
Float32	1048576	1.38×	298.7 μs	411.1 μs	372.0 μs	14.04	10.20	11.27
Float32	16777216	1.50×	658.4 μs	986.8 μs	936.2 μs	101.93	68.01	71.68
ComplexF32	1024	—	150.2 μs	—	175.0 μs	0.05	—	0.05
ComplexF32	1048576	—	430.7 μs	—	542.4 μs	19.48	—	15.47
ComplexF32	16777216	—	1.09 ms	—	1.45 ms	123.49	—	92.79

Summary

Looking at the 16M-element results (where launch overhead is negligible):

• GPUArrays wins on every type GPUArrays and MPS both support, except for the
  small signed integer types (Int8, Int16) where MPS is ~1.3–1.5× faster — and
  even there the absolute gap is sub-millisecond.
• GPUArrays is universally faster for randn! (1.08–1.50×) on top of fixing
  the NaN bug.
• GPUArrays handles types MPS can't: Float16, all complex types, Int128/UInt128.
• KernelRNG is consistently 2–4× slower than the GPUArrays RNG.

Routing all in-place / out-of-place calls through GPUArrays simplifies the code
(no per-type dispatch table) and makes the Metal randoms path consistent with
CUDA.jl. MPS stays available behind Metal.mps_rng() for users who want it.

[codecov]

Codecov Report

❌ Patch coverage is 27.77778% with 39 lines in your changes missing coverage. Please review.
✅ Project coverage is 80.70%. Comparing base (b94fd4b) to head (d1047e1).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
src/random.jl	27.77%	39 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #767      +/-   ##
==========================================
- Coverage   83.09%   80.70%   -2.39%     
==========================================
  Files          62       61       -1     
  Lines        2851     2846       -5     
==========================================
- Hits         2369     2297      -72     
- Misses        482      549      +67     

☔ 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: d1047e1	Previous: b94fd4b	Ratio
array/accumulate/Float32/1d	1119417 ns	1098958 ns	1.02
array/accumulate/Float32/dims=1	1563917 ns	1554708 ns	1.01
array/accumulate/Float32/dims=1L	9845771 ns	9848583.5 ns	1.00
array/accumulate/Float32/dims=2	1877792 ns	1886771 ns	1.00
array/accumulate/Float32/dims=2L	7236750 ns	7256459 ns	1.00
array/accumulate/Int64/1d	1245896 ns	1261958 ns	0.99
array/accumulate/Int64/dims=1	1841479 ns	1824291.5 ns	1.01
array/accumulate/Int64/dims=1L	11601292 ns	11664208.5 ns	0.99
array/accumulate/Int64/dims=2	2165917 ns	2170333.5 ns	1.00
array/accumulate/Int64/dims=2L	9751208 ns	10120062.5 ns	0.96
array/broadcast	608708 ns	605916 ns	1.00
array/construct	6334 ns	6292 ns	1.01
array/permutedims/2d	1170333 ns	1168125 ns	1.00
array/permutedims/3d	1677687 ns	1673084 ns	1.00
array/permutedims/4d	2388812.5 ns	2365959 ns	1.01
array/private/copy	565958 ns	545792 ns	1.04
array/private/copyto!/cpu_to_gpu	809375 ns	802916 ns	1.01
array/private/copyto!/gpu_to_cpu	811416 ns	801917 ns	1.01
array/private/copyto!/gpu_to_gpu	636333 ns	634458 ns	1.00
array/private/iteration/findall/bool	1413167 ns	1402750 ns	1.01
array/private/iteration/findall/int	1561625 ns	1564021 ns	1.00
array/private/iteration/findfirst/bool	2040000 ns	2055916 ns	0.99
array/private/iteration/findfirst/int	2066708 ns	2064479.5 ns	1.00
array/private/iteration/findmin/1d	2491959 ns	2499959 ns	1.00
array/private/iteration/findmin/2d	1775917 ns	1790791 ns	0.99
array/private/iteration/logical	2656688 ns	2631896 ns	1.01
array/private/iteration/scalar	4529875 ns	5047625 ns	0.90
array/random/rand/Float32	1164792 ns	582958 ns	2.00
array/random/rand/Int64	1326750 ns	775667 ns	1.71
array/random/rand!/Float32	924250 ns	574750 ns	1.61
array/random/rand!/Int64	874750 ns	550792 ns	1.59
array/random/randn/Float32	1068458.5 ns	1006937.5 ns	1.06
array/random/randn!/Float32	820146 ns	755666 ns	1.09
array/reductions/mapreduce/Float32/1d	1033375 ns	1029500 ns	1.00
array/reductions/mapreduce/Float32/dims=1	835917 ns	840875 ns	0.99
array/reductions/mapreduce/Float32/dims=1L	1618958.5 ns	1324000 ns	1.22
array/reductions/mapreduce/Float32/dims=2	805354 ns	860875 ns	0.94
array/reductions/mapreduce/Float32/dims=2L	1818229.5 ns	1799541 ns	1.01
array/reductions/mapreduce/Int64/1d	1552833 ns	1374875 ns	1.13
array/reductions/mapreduce/Int64/dims=1	1105583 ns	1097625 ns	1.01
array/reductions/mapreduce/Int64/dims=1L	2027500 ns	2002854 ns	1.01
array/reductions/mapreduce/Int64/dims=2	1177312 ns	1145000 ns	1.03
array/reductions/mapreduce/Int64/dims=2L	3619958 ns	3614000 ns	1.00
array/reductions/reduce/Float32/1d	1028125 ns	1028437.5 ns	1.00
array/reductions/reduce/Float32/dims=1	831125 ns	832667 ns	1.00
array/reductions/reduce/Float32/dims=1L	1316542 ns	1318416.5 ns	1.00
array/reductions/reduce/Float32/dims=2	853563 ns	853041.5 ns	1.00
array/reductions/reduce/Float32/dims=2L	1814042 ns	1810250 ns	1.00
array/reductions/reduce/Int64/1d	1496937.5 ns	1516958 ns	0.99
array/reductions/reduce/Int64/dims=1	1116833 ns	1095375 ns	1.02
array/reductions/reduce/Int64/dims=1L	2011937.5 ns	2023499.5 ns	0.99
array/reductions/reduce/Int64/dims=2	1157459 ns	1240750 ns	0.93
array/reductions/reduce/Int64/dims=2L	4242729 ns	4233875 ns	1.00
array/shared/copy	242125 ns	252417 ns	0.96
array/shared/copyto!/cpu_to_gpu	81750 ns	80750 ns	1.01
array/shared/copyto!/gpu_to_cpu	81709 ns	80667 ns	1.01
array/shared/copyto!/gpu_to_gpu	82584 ns	83083 ns	0.99
array/shared/iteration/findall/bool	1421500 ns	1427208.5 ns	1.00
array/shared/iteration/findall/int	1558458 ns	1559875 ns	1.00
array/shared/iteration/findfirst/bool	1622000 ns	1649000 ns	0.98
array/shared/iteration/findfirst/int	1635500 ns	1672458 ns	0.98
array/shared/iteration/findmin/1d	2093458 ns	2115583 ns	0.99
array/shared/iteration/findmin/2d	1783604.5 ns	1792625 ns	0.99
array/shared/iteration/logical	2427709 ns	2292167 ns	1.06
array/shared/iteration/scalar	201250 ns	199958 ns	1.01
integration/byval/reference	1582500 ns	1544250 ns	1.02
integration/byval/slices=1	1588625 ns	1560229.5 ns	1.02
integration/byval/slices=2	2614875 ns	2598333.5 ns	1.01
integration/byval/slices=3	7820666.5 ns	8092333 ns	0.97
integration/metaldevrt	878645.5 ns	868125 ns	1.01
kernel/indexing	621958 ns	592667 ns	1.05
kernel/indexing_checked	628125 ns	598292 ns	1.05
kernel/launch	11584 ns	11791.5 ns	0.98
kernel/rand	567917 ns	570709 ns	1.00
latency/import	1417687042 ns	1425597062.5 ns	0.99
latency/precompile	25459122750 ns	25453724708 ns	1.00
latency/ttfp	2335347354.5 ns	2341177208 ns	1.00
metal/synchronization/context	19792 ns	19667 ns	1.01
metal/synchronization/stream	18833 ns	18459 ns	1.02

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