[maleadt]

The metal/synchronization/stream benchmark (an empty synchronize()) regressed from ~340 to ~640 ns over #995 and the #1001/#1002/#822 stack. Part of that is inherent to the extra locking those PRs need, but part comes from closure boxes: splitting synchronize into two @autoreleasepool blocks made their closures capture queue and submissions, which were reassigned afterwards. Both got boxed, which made the rest of the function type unstable.

Julia 1.14's Test.detect_closure_boxes found 9 more of these, mostly in @autoreleasepool functions rebinding an argument (that macro wraps the function body in a closure), including encode_copy! on the copyto! path. This PR removes all of them without introducing temporaries, and adds a test that runs detect_closure_boxes on Julia versions that have it.

Empty synchronize() on an M1, Julia 1.12, three interleaved runs:

	median	allocations
main	587–594 ns	5 (112 bytes)
this PR	474–488 ns	2 (48 bytes)

[github-actions]

Metal.jl Benchmarks

Details

Benchmark suite	Current: f05014c	Previous: fda32ae	Ratio
array/accumulate/Float32/1d	394167 ns	394000 ns	1.00
array/accumulate/Float32/dims=1	363833 ns	369166 ns	0.99
array/accumulate/Float32/dims=1L	8871041 ns	8869708 ns	1.00
array/accumulate/Float32/dims=2	431250 ns	434333 ns	0.99
array/accumulate/Float32/dims=2L	2971750 ns	2724458 ns	1.09
array/accumulate/Int64/1d	775416 ns	775291 ns	1.00
array/accumulate/Int64/dims=1	872209 ns	886250 ns	0.98
array/accumulate/Int64/dims=1L	3786708 ns	3729709 ns	1.02
array/accumulate/Int64/dims=2	1162334 ns	1163458 ns	1.00
array/accumulate/Int64/dims=2L	16149083 ns	16005042 ns	1.01
array/broadcast	237917 ns	232584 ns	1.02
array/construct	2375 ns	2292 ns	1.04
array/permutedims/2d	339292 ns	339750 ns	1.00
array/permutedims/3d	461792 ns	461417 ns	1.00
array/permutedims/4d	619958 ns	617417 ns	1.00
array/private/copy	228708 ns	221750 ns	1.03
array/private/copyto!/cpu_to_gpu	224250 ns	224250 ns	1
array/private/copyto!/gpu_to_cpu	222500 ns	221417 ns	1.00
array/private/copyto!/gpu_to_gpu	220584 ns	211959 ns	1.04
array/private/iteration/findall/bool	525458 ns	530292 ns	0.99
array/private/iteration/findall/int	601375 ns	605917 ns	0.99
array/private/iteration/findfirst/bool	375834 ns	374292 ns	1.00
array/private/iteration/findfirst/int	369958 ns	378458 ns	0.98
array/private/iteration/findmin/1d	373792 ns	375417 ns	1.00
array/private/iteration/findmin/2d	1109750 ns	881167 ns	1.26
array/private/iteration/logical	540375 ns	533709 ns	1.01
array/private/iteration/scalar	1384083 ns	1394709 ns	0.99
array/random/rand/Float32	436750 ns	441958 ns	0.99
array/random/rand/Int64	515875 ns	514375 ns	1.00
array/random/rand!/Float32	426833 ns	428250 ns	1.00
array/random/rand!/Int64	448166 ns	449958 ns	1.00
array/random/randn/Float32	350083 ns	366708 ns	0.95
array/random/randn!/Float32	354208 ns	354083 ns	1.00
array/reductions/mapreduce/Float32/1d	234458 ns	223417 ns	1.05
array/reductions/mapreduce/Float32/dims=1	389083 ns	388542 ns	1.00
array/reductions/mapreduce/Float32/dims=1L	584125 ns	586917 ns	1.00
array/reductions/mapreduce/Float32/dims=2	324125 ns	323083 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	457875 ns	451417 ns	1.01
array/reductions/mapreduce/Int64/1d	233708 ns	244208 ns	0.96
array/reductions/mapreduce/Int64/dims=1	406791 ns	406917 ns	1.00
array/reductions/mapreduce/Int64/dims=1L	1000875 ns	1004584 ns	1.00
array/reductions/mapreduce/Int64/dims=2	329625 ns	273167 ns	1.21
array/reductions/mapreduce/Int64/dims=2L	603584 ns	619750 ns	0.97
array/reductions/reduce/Float32/1d	228916 ns	235584 ns	0.97
array/reductions/reduce/Float32/dims=1	388125 ns	388125 ns	1
array/reductions/reduce/Float32/dims=1L	584875 ns	589709 ns	0.99
array/reductions/reduce/Float32/dims=2	301500 ns	316292 ns	0.95
array/reductions/reduce/Float32/dims=2L	456250 ns	442458 ns	1.03
array/reductions/reduce/Int64/1d	248792 ns	237000 ns	1.05
array/reductions/reduce/Int64/dims=1	326917 ns	405750 ns	0.81
array/reductions/reduce/Int64/dims=1L	1000667 ns	1004416 ns	1.00
array/reductions/reduce/Int64/dims=2	335125 ns	331500 ns	1.01
array/reductions/reduce/Int64/dims=2L	604666 ns	620583 ns	0.97
array/shared/copy	128583 ns	128875 ns	1.00
array/shared/copyto!/cpu_to_gpu	37625 ns	37583 ns	1.00
array/shared/copyto!/gpu_to_cpu	36583 ns	38125 ns	0.96
array/shared/copyto!/gpu_to_gpu	36625 ns	38000 ns	0.96
array/shared/iteration/findall/bool	530417 ns	528875 ns	1.00
array/shared/iteration/findall/int	603541 ns	607834 ns	0.99
array/shared/iteration/findfirst/bool	374875 ns	377291 ns	0.99
array/shared/iteration/findfirst/int	379041 ns	378333 ns	1.00
array/shared/iteration/findmin/1d	369417 ns	373375 ns	0.99
array/shared/iteration/findmin/2d	870042 ns	1113333 ns	0.78
array/shared/iteration/logical	536708 ns	547333 ns	0.98
array/shared/iteration/scalar	77.3076923076923 ns	82.34536082474227 ns	0.94
array/sorting/1d	1912167 ns	1919167 ns	1.00
array/sorting/2d	8354166 ns	8351750 ns	1.00
integration/byval/reference	1114875 ns	1121708 ns	0.99
integration/byval/slices=1	1115500 ns	1122125 ns	0.99
integration/byval/slices=2	2025917 ns	2028125 ns	1.00
integration/byval/slices=3	6790792 ns	6844167 ns	0.99
integration/metaldevrt	395875 ns	396500 ns	1.00
kernel/indexing	215625 ns	212083 ns	1.02
kernel/indexing_checked	391333 ns	395250 ns	0.99
kernel/launch	1537.5 ns	1545.9 ns	0.99
kernel/rand	384750 ns	406958 ns	0.95
latency/import	1995882417 ns	2011861958 ns	0.99
latency/precompile	30889583416 ns	30939510500 ns	1.00
latency/ttfp	2236765459 ns	2249344208 ns	0.99
metal/synchronization/context	527.8526315789474 ns	531.3578947368421 ns	0.99
metal/synchronization/stream	574.9071038251366 ns	660.1925465838509 ns	0.87

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 86.77%. Comparing base (fda32ae) to head (f05014c).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1026      +/-   ##
==========================================
+ Coverage   86.75%   86.77%   +0.01%     
==========================================
  Files          89       89              
  Lines        6403     6403              
==========================================
+ Hits         5555     5556       +1     
+ Misses        848      847       -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.