Fix global linear indexing (fill!)

[christiangnrd]

Implementation borrowed from CUDA.jl version.

Close #466

[github-actions]

Metal Benchmarks

Benchmark suite	Current: 28fe952	Previous: 634cec7	Ratio
private array/construct	27387 ns	26887 ns	1.02
private array/broadcast	460417 ns	462333 ns	1.00
private array/random/randn/Float32	806479.5 ns	865083 ns	0.93
private array/random/randn!/Float32	631125 ns	658875 ns	0.96
private array/random/rand!/Int64	566333 ns	548542 ns	1.03
private array/random/rand!/Float32	598083.5 ns	584292 ns	1.02
private array/random/rand/Int64	750791 ns	756500 ns	0.99
private array/random/rand/Float32	617292 ns	605542 ns	1.02
private array/copyto!/gpu_to_gpu	683708 ns	658041 ns	1.04
private array/copyto!/cpu_to_gpu	651791.5 ns	701083 ns	0.93
private array/copyto!/gpu_to_cpu	822896 ns	824854.5 ns	1.00
private array/accumulate/1d	1313500 ns	1325958.5 ns	0.99
private array/accumulate/2d	1381083 ns	1382792 ns	1.00
private array/iteration/findall/int	2062187 ns	2067791 ns	1.00
private array/iteration/findall/bool	1816000 ns	1825375 ns	0.99
private array/iteration/findfirst/int	1688250 ns	1674750 ns	1.01
private array/iteration/findfirst/bool	1645416 ns	1646437.5 ns	1.00
private array/iteration/scalar	3873833 ns	3884646 ns	1.00
private array/iteration/logical	3163875 ns	3164458 ns	1.00
private array/iteration/findmin/1d	1734833.5 ns	1740125 ns	1.00
private array/iteration/findmin/2d	1348875 ns	1346458 ns	1.00
private array/reductions/reduce/1d	1034000 ns	1020729 ns	1.01
private array/reductions/reduce/2d	651208.5 ns	664250 ns	0.98
private array/reductions/mapreduce/1d	1033791 ns	1032125 ns	1.00
private array/reductions/mapreduce/2d	658334 ns	659542 ns	1.00
private array/permutedims/4d	2540500 ns	2720542 ns	0.93
private array/permutedims/2d	1011000 ns	1011208 ns	1.00
private array/permutedims/3d	1579959 ns	1574854 ns	1.00
private array/copy	603542 ns	557250 ns	1.08
latency/precompile	5146918500 ns	5138248250 ns	1.00
latency/ttfp	6634638146 ns	6754936625 ns	0.98
latency/import	1162510500 ns	1151697916.5 ns	1.01
integration/metaldevrt	712750 ns	719667 ns	0.99
integration/byval/slices=1	1567645.5 ns	1560604.5 ns	1.00
integration/byval/slices=3	10250000 ns	10389833 ns	0.99
integration/byval/reference	1546208 ns	1566416.5 ns	0.99
integration/byval/slices=2	2583708 ns	2542084 ns	1.02
kernel/indexing	459667 ns	487187.5 ns	0.94
kernel/indexing_checked	451250 ns	469791.5 ns	0.96
kernel/launch	9895.833333333332 ns	8042 ns	1.23
metal/synchronization/stream	14708 ns	14125 ns	1.04
metal/synchronization/context	15250 ns	14542 ns	1.05
shared array/construct	26496.583333333336 ns	26607.14285714286 ns	1.00
shared array/broadcast	470917 ns	453208 ns	1.04
shared array/random/randn/Float32	820708 ns	790125 ns	1.04
shared array/random/randn!/Float32	666750 ns	668750 ns	1.00
shared array/random/rand!/Int64	564875 ns	573042 ns	0.99
shared array/random/rand!/Float32	590042 ns	596500 ns	0.99
shared array/random/rand/Int64	771042 ns	780209 ns	0.99
shared array/random/rand/Float32	590292 ns	618625 ns	0.95
shared array/copyto!/gpu_to_gpu	86583 ns	87334 ns	0.99
shared array/copyto!/cpu_to_gpu	88292 ns	98084 ns	0.90
shared array/copyto!/gpu_to_cpu	82375 ns	77041 ns	1.07
shared array/accumulate/1d	1325583.5 ns	1330458 ns	1.00
shared array/accumulate/2d	1384459 ns	1388291 ns	1.00
shared array/iteration/findall/int	1801250 ns	1768750 ns	1.02
shared array/iteration/findall/bool	1564583 ns	1573208 ns	0.99
shared array/iteration/findfirst/int	1384375 ns	1392333 ns	0.99
shared array/iteration/findfirst/bool	1364167 ns	1363479.5 ns	1.00
shared array/iteration/scalar	158270.5 ns	152500 ns	1.04
shared array/iteration/logical	2951000 ns	2956500 ns	1.00
shared array/iteration/findmin/1d	1471166.5 ns	1459687.5 ns	1.01
shared array/iteration/findmin/2d	1370666.5 ns	1350209 ns	1.02
shared array/reductions/reduce/1d	732833 ns	718916 ns	1.02
shared array/reductions/reduce/2d	653000 ns	670875 ns	0.97
shared array/reductions/mapreduce/1d	732583 ns	722583 ns	1.01
shared array/reductions/mapreduce/2d	667542 ns	669292 ns	1.00
shared array/permutedims/4d	2555500 ns	2722021 ns	0.94
shared array/permutedims/2d	1021812.5 ns	1005625 ns	1.02
shared array/permutedims/3d	1586041 ns	1573479 ns	1.01
shared array/copy	242416 ns	248062.5 ns	0.98

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

[maleadt]

What is the underlying issue here? The blame on CUDA.jl's implementation goes back to when we imported that code from KA.jl, so cc @vchuravy.

[christiangnrd]

I think the real fix may be to switch from dispatchThreadgroups to dispatchThreads by default since all the devices we support support nonuniform threadgroup sizes.

However, this would be a very big (potentially breaking) change.

[christiangnrd]

I opened #497 for discussion. In the meantime I think this PR should me merged as-is (assuming the code is sound)