Move BFloat16 code out of extension

[christiangnrd]

This is the first half of #446.

Separated out partly to see the benchmarks. Update: No performance difference!

Impact to import/precompile time should be minimal and the decreased code complexity is probably worth the tradeoff.

[github-actions]

Your PR no longer requires formatting changes. Thank you for your contribution!

[github-actions]

Metal Benchmarks

Benchmark suite	Current: 91dca03	Previous: 924a130	Ratio
private array/construct	23527.833333333332 ns	25257 ns	0.93
private array/broadcast	458875 ns	463958 ns	0.99
private array/random/randn/Float32	795041 ns	760208 ns	1.05
private array/random/randn!/Float32	610500 ns	624687.5 ns	0.98
private array/random/rand!/Int64	551333 ns	563395.5 ns	0.98
private array/random/rand!/Float32	571167 ns	587854 ns	0.97
private array/random/rand/Int64	737833 ns	755083 ns	0.98
private array/random/rand/Float32	586750 ns	611333 ns	0.96
private array/copyto!/gpu_to_gpu	647146 ns	650916.5 ns	0.99
private array/copyto!/cpu_to_gpu	764624.5 ns	820417 ns	0.93
private array/copyto!/gpu_to_cpu	620375 ns	686687.5 ns	0.90
private array/accumulate/1d	1341334 ns	1362854.5 ns	0.98
private array/accumulate/2d	1381667 ns	1404375 ns	0.98
private array/iteration/findall/int	2081000 ns	2111541.5 ns	0.99
private array/iteration/findall/bool	1830750 ns	1849958 ns	0.99
private array/iteration/findfirst/int	1684896.5 ns	1715916 ns	0.98
private array/iteration/findfirst/bool	1657249.5 ns	1673542 ns	0.99
private array/iteration/scalar	3900291.5 ns	3935958 ns	0.99
private array/iteration/logical	3161333.5 ns	3227791 ns	0.98
private array/iteration/findmin/1d	1755916 ns	1778916.5 ns	0.99
private array/iteration/findmin/2d	1347625 ns	1352041 ns	1.00
private array/reductions/reduce/1d	1034083 ns	1043458 ns	0.99
private array/reductions/reduce/2d	664479.5 ns	669709 ns	0.99
private array/reductions/mapreduce/1d	1038270.5 ns	1052375 ns	0.99
private array/reductions/mapreduce/2d	660792 ns	675854.5 ns	0.98
private array/permutedims/4d	2491125 ns	2557249.5 ns	0.97
private array/permutedims/2d	1002084 ns	1026042 ns	0.98
private array/permutedims/3d	1582042 ns	1594916 ns	0.99
private array/copy	583917 ns	569146 ns	1.03
latency/precompile	8814246708 ns	8851374292 ns	1.00
latency/ttfp	3612261417 ns	3620472000 ns	1.00
latency/import	1237185791 ns	1235528208 ns	1.00
integration/metaldevrt	709250 ns	717959 ns	0.99
integration/byval/slices=1	1581250 ns	1539333 ns	1.03
integration/byval/slices=3	8888209 ns	9117125 ns	0.97
integration/byval/reference	1583000 ns	1560292 ns	1.01
integration/byval/slices=2	2645542 ns	2731458.5 ns	0.97
kernel/indexing	453708 ns	457750 ns	0.99
kernel/indexing_checked	448667 ns	455833 ns	0.98
kernel/launch	8125 ns	8166 ns	0.99
metal/synchronization/stream	14667 ns	14708 ns	1.00
metal/synchronization/context	14542 ns	14791 ns	0.98
shared array/construct	23527.75 ns	25149.333333333332 ns	0.94
shared array/broadcast	456520.5 ns	464959 ns	0.98
shared array/random/randn/Float32	834208 ns	835792 ns	1.00
shared array/random/randn!/Float32	616792 ns	635500 ns	0.97
shared array/random/rand!/Int64	550208 ns	560542 ns	0.98
shared array/random/rand!/Float32	586833 ns	598750 ns	0.98
shared array/random/rand/Int64	786708.5 ns	765583 ns	1.03
shared array/random/rand/Float32	593604.5 ns	604979 ns	0.98
shared array/copyto!/gpu_to_gpu	79291 ns	79167 ns	1.00
shared array/copyto!/cpu_to_gpu	87167 ns	82250 ns	1.06
shared array/copyto!/gpu_to_cpu	82500 ns	83000 ns	0.99
shared array/accumulate/1d	1346750 ns	1366937.5 ns	0.99
shared array/accumulate/2d	1385333 ns	1400458 ns	0.99
shared array/iteration/findall/int	1795209 ns	1854249.5 ns	0.97
shared array/iteration/findall/bool	1596417 ns	1619625 ns	0.99
shared array/iteration/findfirst/int	1400916 ns	1384771 ns	1.01
shared array/iteration/findfirst/bool	1365542 ns	1375770.5 ns	0.99
shared array/iteration/scalar	155000 ns	155000 ns	1
shared array/iteration/logical	2970604.5 ns	3012021 ns	0.99
shared array/iteration/findmin/1d	1449292 ns	1470417 ns	0.99
shared array/iteration/findmin/2d	1351458 ns	1355875 ns	1.00
shared array/reductions/reduce/1d	723333 ns	729437.5 ns	0.99
shared array/reductions/reduce/2d	660000 ns	674104.5 ns	0.98
shared array/reductions/mapreduce/1d	733125 ns	743833 ns	0.99
shared array/reductions/mapreduce/2d	669125 ns	658375 ns	1.02
shared array/permutedims/4d	2540729 ns	2575708 ns	0.99
shared array/permutedims/2d	1013854.5 ns	1011875 ns	1.00
shared array/permutedims/3d	1572250 ns	1573854 ns	1.00
shared array/copy	246417 ns	246417 ns	1

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