Add 64-bit atomic min/max support

[tgymnich]

Implement device-side / for by wiring Metal AIR i64 atomics, add compilation-time GPU family gating for Apple9+, and cover both supported and unsupported hardware paths in atomics tests.

fixes #477

Assisted by: Copilot

[christiangnrd]

Min and max are supported on Apple8 on macOS. It’s footnote 7 in the first set of footnotes

[codecov]

Codecov Report

❌ Patch coverage is 90.90909% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 86.66%. Comparing base (b8d46ce) to head (a291e76).

Files with missing lines	Patch %	Lines
src/compiler/compilation.jl	90.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #866      +/-   ##
==========================================
+ Coverage   86.62%   86.66%   +0.04%     
==========================================
  Files          76       76              
  Lines        5144     5152       +8     
==========================================
+ Hits         4456     4465       +9     
+ Misses        688      687       -1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

Metal Benchmarks

Details

Benchmark suite	Current: a291e76	Previous: b8d46ce	Ratio
array/accumulate/Float32/1d	525542 ns	515334 ns	1.02
array/accumulate/Float32/dims=1	517187.5 ns	506708 ns	1.02
array/accumulate/Float32/dims=1L	8947958 ns	8950000 ns	1.00
array/accumulate/Float32/dims=2	539000 ns	538958 ns	1.00
array/accumulate/Float32/dims=2L	2894416 ns	2913312.5 ns	0.99
array/accumulate/Int64/1d	930000 ns	931417 ns	1.00
array/accumulate/Int64/dims=1	1110125 ns	1114292 ns	1.00
array/accumulate/Int64/dims=1L	9839542 ns	9781937 ns	1.01
array/accumulate/Int64/dims=2	1447437.5 ns	1454062.5 ns	1.00
array/accumulate/Int64/dims=2L	6947125 ns	6944084 ns	1.00
array/broadcast	334792 ns	333167 ns	1.00
array/construct	3209 ns	3333 ns	0.96
array/permutedims/2d	619417 ns	619917 ns	1.00
array/permutedims/3d	1082625 ns	1099417 ns	0.98
array/permutedims/4d	1241875 ns	1231833 ns	1.01
array/private/copy	379500 ns	393104 ns	0.97
array/private/copyto!/cpu_to_gpu	366666 ns	368604.5 ns	0.99
array/private/copyto!/gpu_to_cpu	361145.5 ns	357542 ns	1.01
array/private/copyto!/gpu_to_gpu	324854.5 ns	325791 ns	1.00
array/private/iteration/findall/bool	1191583 ns	1192542 ns	1.00
array/private/iteration/findall/int	1322125 ns	1323145.5 ns	1.00
array/private/iteration/findfirst/bool	1325000 ns	1315083 ns	1.01
array/private/iteration/findfirst/int	1385708 ns	1375917 ns	1.01
array/private/iteration/findmin/1d	1517458 ns	1520792 ns	1.00
array/private/iteration/findmin/2d	1229062.5 ns	1229812.5 ns	1.00
array/private/iteration/logical	1984312.5 ns	1990958 ns	1.00
array/private/iteration/scalar	1731875 ns	1747333 ns	0.99
array/random/rand/Float32	598375 ns	603209 ns	0.99
array/random/rand/Int64	644084 ns	641812.5 ns	1.00
array/random/rand!/Float32	497333 ns	498917 ns	1.00
array/random/rand!/Int64	475458 ns	466875 ns	1.02
array/random/randn/Float32	557500 ns	561042 ns	0.99
array/random/randn!/Float32	452167 ns	454937.5 ns	0.99
array/reductions/mapreduce/Float32/1d	615208 ns	620042 ns	0.99
array/reductions/mapreduce/Float32/dims=1	447333.5 ns	442167 ns	1.01
array/reductions/mapreduce/Float32/dims=1L	699959 ns	701292 ns	1.00
array/reductions/mapreduce/Float32/dims=2	447354.5 ns	449542 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	1001375 ns	1001125 ns	1.00
array/reductions/mapreduce/Int64/1d	798292 ns	808375 ns	0.99
array/reductions/mapreduce/Int64/dims=1	749541.5 ns	758375 ns	0.99
array/reductions/mapreduce/Int64/dims=1L	1100250 ns	1102250 ns	1.00
array/reductions/mapreduce/Int64/dims=2	765625 ns	812583 ns	0.94
array/reductions/mapreduce/Int64/dims=2L	2214167 ns	2213417 ns	1.00
array/reductions/reduce/Float32/1d	606354 ns	610625 ns	0.99
array/reductions/reduce/Float32/dims=1	448625 ns	446333.5 ns	1.01
array/reductions/reduce/Float32/dims=1L	702709 ns	702500 ns	1.00
array/reductions/reduce/Float32/dims=2	355416 ns	356917 ns	1.00
array/reductions/reduce/Float32/dims=2L	494666 ns	497708 ns	0.99
array/reductions/reduce/Int64/1d	802709 ns	798708 ns	1.01
array/reductions/reduce/Int64/dims=1	740208 ns	760000 ns	0.97
array/reductions/reduce/Int64/dims=1L	1097542 ns	1097542 ns	1
array/reductions/reduce/Int64/dims=2	298042 ns	293833.5 ns	1.01
array/reductions/reduce/Int64/dims=2L	685459 ns	689500 ns	0.99
array/shared/copy	159958 ns	158167 ns	1.01
array/shared/copyto!/cpu_to_gpu	38875 ns	39584 ns	0.98
array/shared/copyto!/gpu_to_cpu	40375 ns	39917 ns	1.01
array/shared/copyto!/gpu_to_gpu	40916 ns	40292 ns	1.02
array/shared/iteration/findall/bool	1199125 ns	1199792 ns	1.00
array/shared/iteration/findall/int	1328333 ns	1330625 ns	1.00
array/shared/iteration/findfirst/bool	1008959 ns	994708 ns	1.01
array/shared/iteration/findfirst/int	1000000 ns	1020625 ns	0.98
array/shared/iteration/findmin/1d	1283750 ns	1281750 ns	1.00
array/shared/iteration/findmin/2d	1231875 ns	1231125 ns	1.00
array/shared/iteration/logical	1793208 ns	1796688 ns	1.00
array/shared/iteration/scalar	3468.75 ns	3437.5 ns	1.01
array/sorting/1d	1987562.5 ns	1993833.5 ns	1.00
array/sorting/2d	8453250 ns	8450062.5 ns	1.00
integration/byval/reference	1150709 ns	1150062.5 ns	1.00
integration/byval/slices=1	1152167 ns	1151750 ns	1.00
integration/byval/slices=2	2080792 ns	2079167 ns	1.00
integration/byval/slices=3	18378292 ns	18009271 ns	1.02
integration/metaldevrt	438083.5 ns	439250 ns	1.00
kernel/indexing	329958 ns	331500 ns	1.00
kernel/indexing_checked	487583 ns	492500 ns	0.99
kernel/launch	1987.5 ns	1954.1 ns	1.02
kernel/rand	489750 ns	497209 ns	0.98
latency/import	1809613541 ns	1812323292 ns	1.00
latency/precompile	40243004874.5 ns	39925435375 ns	1.01
latency/ttfp	2147904042 ns	2150481396 ns	1.00
metal/synchronization/context	518.5647668393782 ns	534.3264248704663 ns	0.97
metal/synchronization/stream	276.37987012987014 ns	274.8666666666667 ns	1.01

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

[maleadt]

Checking device() is wrong here, because it disallows even generating code (e.g. for tests, or debugging). Normally we check job.config.target as per above, but if we want a different compatibility axis we could encode those things in the job.config.params (which are controller by Metal.jl).

[tgymnich]

added gpufamily to params

[tgymnich]

Min and max are supported on Apple8 on macOS. It’s footnote 7 in the first set of footnotes

@christiangnrd good catch!

[christiangnrd]

The only question for me that remains is if the ::Union{Nothing, MTL.MTLGPUFamily} field affects performance and if we really need to guard against a GPUFamily query returning nothing

[tgymnich]

The only question for me that remains is if the ::Union{Nothing, MTL.MTLGPUFamily} field affects performance and if we really need to guard against a GPUFamily query returning nothing

@christiangnrd nothing is for virtualized metal devices. Falling back to Apple1 does not make any sense here. We could maybe introduce a fake Apple0, AppleVirtual, or NONE family.

[christiangnrd]

@christiangnrd nothing is for virtualized metal devices.

Don't they report Apple5?

[tgymnich]

maybe newer ones do. the ones in GitHub action report nothing.

[christiangnrd]

Maybe it's a bug in how you're checking? The Metal.versioninfo printout in the CI runs does show Apple5

[tgymnich]

@christiangnrd The problem seems to be with the intel runners

[christiangnrd]

I guess we can keep the nothing, but I've disabled the precompilation workload for non-aarch64 platforms since it's a waste to precompile on a platform we know doesn't work