[maleadt]

Device code emits barriers in their AIR 2.9 form, with a memory order, and relied on GPUCompiler (legalize_barrier_abi!, JuliaGPU/GPUCompiler.jl#1020) to drop the order for older AIR versions and to make it sequentially consistent before MSL 4.1. This moves that rewrite into finish_ir!, next to those of the simdgroup matrix and tensor intrinsics, following #1097: GPUCompiler then doesn't need to know the forms Metal.jl emits, and only legalizes the atomics and fences it emits itself.

With the GPUCompiler releases this currently allows, the rewrite there becomes a no-op, so this needs no compat change. It has to be merged before JuliaGPU/GPUCompiler.jl#1047 is released, which removes GPUCompiler's version; no released Metal.jl emits the AIR 2.9 form.

device/intrinsics/{synchronization,simd,atomics} pass on an M1 with macOS 27 against both GPUCompiler 2.16.0 and that branch. Without this change, the barrier ABI tests (Metal 3.2/AIR 2.7, Metal 4.0/AIR 2.9) fail against the branch.

[github-actions]

Metal.jl Benchmarks

Details

Benchmark suite	Current: d2bae92	Previous: fa8ed24	Ratio
array/accumulate/Float32/1d	417125 ns	409625 ns	1.02
array/accumulate/Float32/dims=1	384792 ns	391583 ns	0.98
array/accumulate/Float32/dims=1L	8944416 ns	8995750 ns	0.99
array/accumulate/Float32/dims=2	463083 ns	472458 ns	0.98
array/accumulate/Float32/dims=2L	3228792 ns	3206042 ns	1.01
array/accumulate/Int64/1d	563583 ns	566875 ns	0.99
array/accumulate/Int64/dims=1	560917 ns	562458 ns	1.00
array/accumulate/Int64/dims=1L	1860083 ns	1831541 ns	1.02
array/accumulate/Int64/dims=2	590709 ns	581583 ns	1.02
array/accumulate/Int64/dims=2L	6188917 ns	6260500 ns	0.99
array/broadcast	228666 ns	221625 ns	1.03
array/construct	2833 ns	2833 ns	1
array/permutedims/2d	373666 ns	383667 ns	0.97
array/permutedims/3d	502416 ns	492125 ns	1.02
array/permutedims/4d	674125 ns	656166 ns	1.03
array/private/copy	266292 ns	246958 ns	1.08
array/private/copyto!/cpu_to_gpu	249292 ns	245417 ns	1.02
array/private/copyto!/gpu_to_cpu	246625 ns	247334 ns	1.00
array/private/copyto!/gpu_to_gpu	196500 ns	188667 ns	1.04
array/private/iteration/findall/bool	545041 ns	537625 ns	1.01
array/private/iteration/findall/int	686250 ns	652167 ns	1.05
array/private/iteration/findfirst/bool	263625 ns	256000 ns	1.03
array/private/iteration/findfirst/int	266084 ns	268667 ns	0.99
array/private/iteration/findmin/1d	321042 ns	316209 ns	1.02
array/private/iteration/findmin/2d	812334 ns	809250 ns	1.00
array/private/iteration/logical	554541 ns	530167 ns	1.05
array/private/iteration/scalar	1816166 ns	1824917 ns	1.00
array/random/rand/Float32	487916 ns	480041 ns	1.02
array/random/rand/Int64	565625 ns	561167 ns	1.01
array/random/rand!/Float32	403083 ns	408292 ns	0.99
array/random/rand!/Int64	428666 ns	425958 ns	1.01
array/random/randn/Float32	403750 ns	393250 ns	1.03
array/random/randn!/Float32	337333 ns	338750 ns	1.00
array/reductions/mapreduce/Float32/1d	215417 ns	212541 ns	1.01
array/reductions/mapreduce/Float32/dims=1	211875 ns	213292 ns	0.99
array/reductions/mapreduce/Float32/dims=1L	665167 ns	652000 ns	1.02
array/reductions/mapreduce/Float32/dims=2	251750 ns	254625 ns	0.99
array/reductions/mapreduce/Float32/dims=2L	415375 ns	428334 ns	0.97
array/reductions/mapreduce/Int64/1d	209583 ns	230458 ns	0.91
array/reductions/mapreduce/Int64/dims=1	225084 ns	231125 ns	0.97
array/reductions/mapreduce/Int64/dims=1L	1085584 ns	1078916 ns	1.01
array/reductions/mapreduce/Int64/dims=2	274250 ns	260750 ns	1.05
array/reductions/mapreduce/Int64/dims=2L	631167 ns	640959 ns	0.98
array/reductions/reduce/Float32/1d	211625 ns	224042 ns	0.94
array/reductions/reduce/Float32/dims=1	195750 ns	195500 ns	1.00
array/reductions/reduce/Float32/dims=1L	653292 ns	655625 ns	1.00
array/reductions/reduce/Float32/dims=2	246458 ns	264167 ns	0.93
array/reductions/reduce/Float32/dims=2L	421750 ns	418958 ns	1.01
array/reductions/reduce/Int64/1d	232958 ns	231958 ns	1.00
array/reductions/reduce/Int64/dims=1	216208 ns	214334 ns	1.01
array/reductions/reduce/Int64/dims=1L	1085083 ns	1080583 ns	1.00
array/reductions/reduce/Int64/dims=2	263250 ns	275625 ns	0.96
array/reductions/reduce/Int64/dims=2L	644459 ns	638042 ns	1.01
array/shared/copy	131417 ns	130125 ns	1.01
array/shared/copyto!/cpu_to_gpu	36291 ns	37041 ns	0.98
array/shared/copyto!/gpu_to_cpu	36916 ns	36792 ns	1.00
array/shared/copyto!/gpu_to_gpu	37417 ns	37334 ns	1.00
array/shared/iteration/findall/bool	567417 ns	540375 ns	1.05
array/shared/iteration/findall/int	687750 ns	687083 ns	1.00
array/shared/iteration/findfirst/bool	277541 ns	283459 ns	0.98
array/shared/iteration/findfirst/int	284250 ns	278084 ns	1.02
array/shared/iteration/findmin/1d	326959 ns	338167 ns	0.97
array/shared/iteration/findmin/2d	817791 ns	813792 ns	1.00
array/shared/iteration/logical	577834 ns	579416 ns	1.00
array/shared/iteration/scalar	82.99180327868852 ns	83.29012345679013 ns	1.00
array/sorting/1d	2012791 ns	2032792 ns	0.99
array/sorting/2d	9012291 ns	9165833 ns	0.98
integration/byval/reference	1100709 ns	1127542 ns	0.98
integration/byval/slices=1	1129833 ns	1131208 ns	1.00
integration/byval/slices=2	2121209 ns	2129500 ns	1.00
integration/byval/slices=3	6860375 ns	6796084 ns	1.01
integration/metaldevrt	392916 ns	375000 ns	1.05
kernel/indexing	211791 ns	207333 ns	1.02
kernel/indexing_checked	382625 ns	372750 ns	1.03
kernel/launch	1579.1 ns	1587.5 ns	0.99
kernel/rand	397250 ns	380250 ns	1.04
latency/import	1520723625 ns	1518648625 ns	1.00
latency/precompile	42524629541 ns	42324049375 ns	1.00
latency/ttfp	1842728708 ns	1839527750 ns	1.00
metal/synchronization/context	594.9720670391062 ns	611.6685714285715 ns	0.97
metal/synchronization/stream	389.30348258706465 ns	396.97014925373134 ns	0.98

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

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 88.16%. Comparing base (fa8ed24) to head (d2bae92).
⚠️ Report is 2 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1099      +/-   ##
==========================================
- Coverage   88.18%   88.16%   -0.02%     
==========================================
  Files         105      105              
  Lines        9147     9194      +47     
==========================================
+ Hits         8066     8106      +40     
- Misses       1081     1088       +7     

☔ 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.