Use CPU copy with SharedStorage

[christiangnrd]

Use CPU copy for shared storage arrays to avoid ObjectiveC.jl overhead.

Is this even a good idea?

Depends on #452

[github-actions]

Metal Benchmarks

Benchmark suite	Current: e9ac0d2	Previous: ff7c7eb	Ratio
private array/construct	27208.333333333332 ns	26687.5 ns	1.02
private array/broadcast	455584 ns	465979.5 ns	0.98
private array/random/randn/Float32	1011500 ns	993270.5 ns	1.02
private array/random/randn!/Float32	631583 ns	632166.5 ns	1.00
private array/random/rand!/Int64	577417 ns	568500 ns	1.02
private array/random/rand!/Float32	586000 ns	583500 ns	1.00
private array/random/rand/Int64	877125 ns	880458 ns	1.00
private array/random/rand/Float32	703750 ns	844333.5 ns	0.83
private array/copyto!/gpu_to_gpu	622250 ns	614333 ns	1.01
private array/copyto!/cpu_to_gpu	692250 ns	739479 ns	0.94
private array/copyto!/gpu_to_cpu	594083.5 ns	599208 ns	0.99
private array/accumulate/1d	1434083 ns	1447750.5 ns	0.99
private array/accumulate/2d	1479500 ns	1496375 ns	0.99
private array/iteration/findall/int	2218500 ns	2263917 ns	0.98
private array/iteration/findall/bool	2002187.5 ns	1989875 ns	1.01
private array/iteration/findfirst/int	1688250 ns	1678000 ns	1.01
private array/iteration/findfirst/bool	1650625 ns	1663625 ns	0.99
private array/iteration/scalar	2399750 ns	2393834 ns	1.00
private array/iteration/logical	3446416 ns	3431520.5 ns	1.00
private array/iteration/findmin/1d	1757084 ns	1794125 ns	0.98
private array/iteration/findmin/2d	1358875 ns	1403416 ns	0.97
private array/reductions/reduce/1d	800917 ns	805792 ns	0.99
private array/reductions/reduce/2d	700479.5 ns	704146 ns	0.99
private array/reductions/mapreduce/1d	811125 ns	815812.5 ns	0.99
private array/reductions/mapreduce/2d	701166.5 ns	716666.5 ns	0.98
private array/permutedims/4d	947645.5 ns	943959 ns	1.00
private array/permutedims/2d	950791 ns	938875 ns	1.01
private array/permutedims/3d	1007916 ns	1005416.5 ns	1.00
private array/copy	876354.5 ns	862875 ns	1.02
latency/precompile	4414162875 ns	4407793041 ns	1.00
latency/ttfp	6916084749.5 ns	6915521687.5 ns	1.00
latency/import	726415791.5 ns	726643917 ns	1.00
integration/metaldevrt	743792 ns	749270.5 ns	0.99
integration/byval/slices=1	1482750 ns	1557959 ns	0.95
integration/byval/slices=3	8832249.5 ns	8832020.5 ns	1.00
integration/byval/reference	1515979 ns	1611291 ns	0.94
integration/byval/slices=2	2747375 ns	2583750 ns	1.06
kernel/indexing	469583 ns	476584 ns	0.99
kernel/indexing_checked	444083 ns	441500 ns	1.01
kernel/launch	11125 ns	10875 ns	1.02
metal/synchronization/stream	19292 ns	19208 ns	1.00
metal/synchronization/context	19792 ns	19750 ns	1.00
shared array/construct	24017.416666666664 ns	23756.916666666664 ns	1.01
shared array/broadcast	466625 ns	469584 ns	0.99
shared array/random/randn/Float32	1024625 ns	1020166 ns	1.00
shared array/random/randn!/Float32	632917 ns	634458 ns	1.00
shared array/random/rand!/Int64	579292 ns	572000 ns	1.01
shared array/random/rand!/Float32	598750 ns	593208.5 ns	1.01
shared array/random/rand/Int64	862833 ns	742792 ns	1.16
shared array/random/rand/Float32	883625 ns	898812.5 ns	0.98
shared array/copyto!/gpu_to_gpu	97125 ns	659667 ns	0.15
shared array/copyto!/cpu_to_gpu	87542 ns	94458 ns	0.93
shared array/copyto!/gpu_to_cpu	82041 ns	84333 ns	0.97
shared array/accumulate/1d	1434500 ns	1418250 ns	1.01
shared array/accumulate/2d	1492917 ns	1500167 ns	1.00
shared array/iteration/findall/int	1972125 ns	1939666 ns	1.02
shared array/iteration/findall/bool	1780625 ns	1746333 ns	1.02
shared array/iteration/findfirst/int	1405208 ns	1413458 ns	0.99
shared array/iteration/findfirst/bool	1369834 ns	1374750 ns	1.00
shared array/iteration/scalar	187667 ns	189167 ns	0.99
shared array/iteration/logical	3193624.5 ns	3212770.5 ns	0.99
shared array/iteration/findmin/1d	1460500 ns	1481709 ns	0.99
shared array/iteration/findmin/2d	1374084 ns	1379250 ns	1.00
shared array/reductions/reduce/1d	673729 ns	659583 ns	1.02
shared array/reductions/reduce/2d	698209 ns	706354 ns	0.99
shared array/reductions/mapreduce/1d	631187 ns	620667 ns	1.02
shared array/reductions/mapreduce/2d	706416.5 ns	704958.5 ns	1.00
shared array/permutedims/4d	954291 ns	963438 ns	0.99
shared array/permutedims/2d	918604 ns	939020.5 ns	0.98
shared array/permutedims/3d	1013459 ns	1003520.5 ns	1.01
shared array/copy	239958.5 ns	880541 ns	0.27

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

[maleadt]

Is this even a good idea?

I think so; we have similar optimizations in CUDA.jl with unified memory. Copies from and to CPU memory are blocking anyway.