Use GPUArrays 12's reductions, scans, sorting and findall

[maleadt]

Ports Metal.jl to GPUArrays 12, which implements Base's reductions, sorting, scans, findall, logical indexing and reverse once for every GPU array, on AcceleratedKernels 0.5 (JuliaGPU/GPUArrays.jl#790), and removes the GPUArrays.mapreducedim! hook.

• Deletes Metal's reduction kernels and the MPSGraph reductions (Metal.reduce_alg). GPUArrays 12 has no hook for vendor reductions; one can come in a 12.x minor if benchmarks show MPSGraph is worth it.
• Deletes the native scan, findall and logical indexing.
• Keeps MPSGraph as the fast path for sorting and scans (Metal.scan_alg); everything else goes to GPUArrays' methods instead of Base's scalar ones. Sorting now accepts AcceleratedKernels' algorithms, and sortperm is stable by default, as in Base: MPSGraph's argsort is only used when the algorithm allows an unstable result.

Intended as a minor release. Supersedes #874 (simd shuffles in the native scan, which is gone).

[github-actions]

Metal.jl Benchmarks

Details

Benchmark suite	Current: e3c5a11	Previous: bfe34d9	Ratio
array/accumulate/Float32/1d	396958 ns	391125 ns	1.01
array/accumulate/Float32/dims=1	352416 ns	365083 ns	0.97
array/accumulate/Float32/dims=1L	8865708 ns	8882000 ns	1.00
array/accumulate/Float32/dims=2	429125 ns	429250 ns	1.00
array/accumulate/Float32/dims=2L	2573625 ns	2203000 ns	1.17
array/accumulate/Int64/1d	776958 ns	847708 ns	0.92
array/accumulate/Int64/dims=1	878250 ns	912958 ns	0.96
array/accumulate/Int64/dims=1L	3780625 ns	9542583 ns	0.40
array/accumulate/Int64/dims=2	1079583 ns	1215708 ns	0.89
array/accumulate/Int64/dims=2L	14720708 ns	6558375 ns	2.24
array/broadcast	240042 ns	227250 ns	1.06
array/construct	2375 ns	2375 ns	1
array/permutedims/2d	454583 ns	460084 ns	0.99
array/permutedims/3d	1028000 ns	1025583 ns	1.00
array/permutedims/4d	1100334 ns	1119125 ns	0.98
array/private/copy	221958 ns	227333 ns	0.98
array/private/copyto!/cpu_to_gpu	223583 ns	212291 ns	1.05
array/private/copyto!/gpu_to_cpu	220375 ns	212625 ns	1.04
array/private/copyto!/gpu_to_gpu	218916 ns	219375 ns	1.00
array/private/iteration/findall/bool	539542 ns	1043625 ns	0.52
array/private/iteration/findall/int	617875 ns	1213625 ns	0.51
array/private/iteration/findfirst/bool	377625 ns	1015208 ns	0.37
array/private/iteration/findfirst/int	381375 ns	1021792 ns	0.37
array/private/iteration/findmin/1d	383333 ns	1092209 ns	0.35
array/private/iteration/findmin/2d	622791 ns	1036125 ns	0.60
array/private/iteration/logical	540750 ns	1540542 ns	0.35
array/private/iteration/scalar	1397000 ns	1391708 ns	1.00
array/random/rand/Float32	420125 ns	420583 ns	1.00
array/random/rand/Int64	498167 ns	498791 ns	1.00
array/random/rand!/Float32	405250 ns	403375 ns	1.00
array/random/rand!/Int64	428375 ns	429500 ns	1.00
array/random/randn/Float32	386833 ns	391167 ns	0.99
array/random/randn!/Float32	374250 ns	373917 ns	1.00
array/reductions/mapreduce/Float32/1d	246459 ns	274208 ns	0.90
array/reductions/mapreduce/Float32/dims=1	392292 ns	352959 ns	1.11
array/reductions/mapreduce/Float32/dims=1L	590458 ns	603125 ns	0.98
array/reductions/mapreduce/Float32/dims=2	329375 ns	358084 ns	0.92
array/reductions/mapreduce/Float32/dims=2L	458584 ns	1245500 ns	0.37
array/reductions/mapreduce/Int64/1d	255375 ns	458125 ns	0.56
array/reductions/mapreduce/Int64/dims=1	431917 ns	635584 ns	0.68
array/reductions/mapreduce/Int64/dims=1L	1008209 ns	1011667 ns	1.00
array/reductions/mapreduce/Int64/dims=2	325542 ns	788208 ns	0.41
array/reductions/mapreduce/Int64/dims=2L	631209 ns	2204250 ns	0.29
array/reductions/reduce/Float32/1d	234959 ns	268917 ns	0.87
array/reductions/reduce/Float32/dims=1	392875 ns	357458 ns	1.10
array/reductions/reduce/Float32/dims=1L	593125 ns	613250 ns	0.97
array/reductions/reduce/Float32/dims=2	333583 ns	248708 ns	1.34
array/reductions/reduce/Float32/dims=2L	446084 ns	468458 ns	0.95
array/reductions/reduce/Int64/1d	259166 ns	453750 ns	0.57
array/reductions/reduce/Int64/dims=1	428417 ns	641667 ns	0.67
array/reductions/reduce/Int64/dims=1L	1004083 ns	1004125 ns	1.00
array/reductions/reduce/Int64/dims=2	344709 ns	260208 ns	1.32
array/reductions/reduce/Int64/dims=2L	628541 ns	666041 ns	0.94
array/shared/copy	128833 ns	129375 ns	1.00
array/shared/copyto!/cpu_to_gpu	37291 ns	36541 ns	1.02
array/shared/copyto!/gpu_to_cpu	37125 ns	38125 ns	0.97
array/shared/copyto!/gpu_to_gpu	37333 ns	38500 ns	0.97
array/shared/iteration/findall/bool	543042 ns	1051583 ns	0.52
array/shared/iteration/findall/int	614791 ns	1215334 ns	0.51
array/shared/iteration/findfirst/bool	376459 ns	974709 ns	0.39
array/shared/iteration/findfirst/int	381833 ns	986416 ns	0.39
array/shared/iteration/findmin/1d	378333 ns	1063708 ns	0.36
array/shared/iteration/findmin/2d	504500 ns	1025750 ns	0.49
array/shared/iteration/logical	542833 ns	1524291 ns	0.36
array/shared/iteration/scalar	80.32955715756951 ns	77.34394250513347 ns	1.04
array/sorting/1d	1908209 ns	1916583 ns	1.00
array/sorting/2d	8292708 ns	8256666 ns	1.00
integration/byval/reference	1113625 ns	1121334 ns	0.99
integration/byval/slices=1	1106708 ns	1126292 ns	0.98
integration/byval/slices=2	2024250 ns	2029166 ns	1.00
integration/byval/slices=3	6768542 ns	6991667 ns	0.97
integration/metaldevrt	392083 ns	395792 ns	0.99
kernel/indexing	208458 ns	214541 ns	0.97
kernel/indexing_checked	396292 ns	394167 ns	1.01
kernel/launch	2138.8888888888887 ns	2111.1111111111113 ns	1.01
kernel/rand	408375 ns	408833 ns	1.00
latency/import	1967763958 ns	1608627167 ns	1.22
latency/precompile	40907675333 ns	32823841792 ns	1.25
latency/ttfp	2218190291 ns	2063787291 ns	1.07
metal/synchronization/context	522.6910994764398 ns	536.1587301587301 ns	0.97
metal/synchronization/stream	628.1882352941177 ns	620.6395348837209 ns	1.01

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