Adapt to GPUArrays.jl transition to KernelAbstractions.jl.

[maleadt]

No description provided.

[github-actions]

Metal Benchmarks

Benchmark suite	Current: 1273856	Previous: 100f831	Ratio
private array/construct	26750 ns	26340.333333333332 ns	1.02
private array/broadcast	469291.5 ns	465791 ns	1.01
private array/random/randn/Float32	841854.5 ns	827937.5 ns	1.02
private array/random/randn!/Float32	632500 ns	635250 ns	1.00
private array/random/rand!/Int64	567958 ns	562291.5 ns	1.01
private array/random/rand!/Float32	574083 ns	594500 ns	0.97
private array/random/rand/Int64	788667 ns	801791 ns	0.98
private array/random/rand/Float32	608667 ns	591208 ns	1.03
private array/copyto!/gpu_to_gpu	668750 ns	645500 ns	1.04
private array/copyto!/cpu_to_gpu	619500 ns	618187.5 ns	1.00
private array/copyto!/gpu_to_cpu	812875.5 ns	798500 ns	1.02
private array/accumulate/1d	1396750 ns	1333083 ns	1.05
private array/accumulate/2d	1428000 ns	1424500 ns	1.00
private array/iteration/findall/int	2178958 ns	2100167 ns	1.04
private array/iteration/findall/bool	1914000 ns	1847000 ns	1.04
private array/iteration/findfirst/int	1727167 ns	1696166.5 ns	1.02
private array/iteration/findfirst/bool	1659145.5 ns	1651958.5 ns	1.00
private array/iteration/scalar	3045187 ns	3657771 ns	0.83
private array/iteration/logical	3330666 ns	3264437.5 ns	1.02
private array/iteration/findmin/1d	1579333 ns	1565166 ns	1.01
private array/iteration/findmin/2d	1336979 ns	1351333.5 ns	0.99
private array/reductions/reduce/1d	1063875 ns	1063291 ns	1.00
private array/reductions/reduce/2d	672791 ns	695645.5 ns	0.97
private array/reductions/mapreduce/1d	1085166.5 ns	1078084 ns	1.01
private array/reductions/mapreduce/2d	677500 ns	705166 ns	0.96
private array/permutedims/4d	2911500 ns	860084 ns	3.39
private array/permutedims/2d	1065021 ns	862229.5 ns	1.24
private array/permutedims/3d	1629229 ns	919520.5 ns	1.77
private array/copy	577083 ns	574854 ns	1.00
latency/precompile	4529354083 ns	4396587542 ns	1.03
latency/ttfp	6845760958.5 ns	6698494124.5 ns	1.02
latency/import	882184000 ns	722852834 ns	1.22
integration/metaldevrt	735937.5 ns	719875 ns	1.02
integration/byval/slices=1	1527250 ns	1530167 ns	1.00
integration/byval/slices=3	9243250.5 ns	9115541.5 ns	1.01
integration/byval/reference	1541021 ns	1520271 ns	1.01
integration/byval/slices=2	2508667 ns	2666416 ns	0.94
kernel/indexing	481583 ns	468541 ns	1.03
kernel/indexing_checked	464583 ns	461292 ns	1.01
kernel/launch	8875 ns	8834 ns	1.00
metal/synchronization/stream	14958 ns	14583 ns	1.03
metal/synchronization/context	15042 ns	15250 ns	0.99
shared array/construct	25541.666666666668 ns	26069.5 ns	0.98
shared array/broadcast	472500 ns	468333 ns	1.01
shared array/random/randn/Float32	772791 ns	785583 ns	0.98
shared array/random/randn!/Float32	660500 ns	626541.5 ns	1.05
shared array/random/rand!/Int64	566750 ns	564084 ns	1.00
shared array/random/rand!/Float32	590292 ns	598792 ns	0.99
shared array/random/rand/Int64	769583 ns	788666 ns	0.98
shared array/random/rand/Float32	628146 ns	629791 ns	1.00
shared array/copyto!/gpu_to_gpu	85708 ns	96916 ns	0.88
shared array/copyto!/cpu_to_gpu	89792 ns	88583 ns	1.01
shared array/copyto!/gpu_to_cpu	83667 ns	83458 ns	1.00
shared array/accumulate/1d	1417625.5 ns	1356667 ns	1.04
shared array/accumulate/2d	1432791.5 ns	1421333 ns	1.01
shared array/iteration/findall/int	1904333.5 ns	1792833 ns	1.06
shared array/iteration/findall/bool	1677792 ns	1620166.5 ns	1.04
shared array/iteration/findfirst/int	1394021 ns	1385791 ns	1.01
shared array/iteration/findfirst/bool	1374083.5 ns	1376291 ns	1.00
shared array/iteration/scalar	157917 ns	151458 ns	1.04
shared array/iteration/logical	3093958 ns	3042333 ns	1.02
shared array/iteration/findmin/1d	1288583 ns	1274875 ns	1.01
shared array/iteration/findmin/2d	1361333 ns	1346333 ns	1.01
shared array/reductions/reduce/1d	679937 ns	694458 ns	0.98
shared array/reductions/reduce/2d	690875 ns	702292 ns	0.98
shared array/reductions/mapreduce/1d	741979 ns	754229 ns	0.98
shared array/reductions/mapreduce/2d	700041 ns	705395.5 ns	0.99
shared array/permutedims/4d	2933000 ns	858875 ns	3.41
shared array/permutedims/2d	1054250 ns	862292 ns	1.22
shared array/permutedims/3d	1625958 ns	923916.5 ns	1.76
shared array/copy	250166.5 ns	246583 ns	1.01

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

[maleadt]

private array/permutedims/4d 	2929083 ns 	860084 ns 	3.41
shared array/permutedims/4d 	2931166.5 ns 	858875 ns 	3.41

That looks like a big regression. Let's re-start the run to be sure.

[maleadt]

OK yeah permutedims remains slow, will have a closer look.

[maleadt]

Let's track the slowdown here: JuliaGPU/GPUArrays.jl#565