[maleadt]

Closes #1004. Rather than cleaning up after Metal's gpuarchiver-* leak with a private method, this removes the per-kernel MTLBinaryArchive cache from #930. Metal's own per-user shader cache ($(getconf DARWIN_USER_CACHE_DIR)com.apple.metal) is keyed on the library contents, and our metallibs are byte-stable across sessions, so that cache already handles the common case, and does so faster.

I measured this on the full test suite (--all, M1, macOS 27). I dumped every metallib it linked (2,805 kernels), then replayed only the back-end link, each case in a fresh process:

Case	ms/kernel (mean)	median
Archives off, system cache warm	0.14	0.12
Archives on, hit	0.85	0.77
Archives on, miss (harvest), system cache warm	22.6	21.7
Archives off, system cache cold	18.0	17.5

A harvest isn't served from the system cache, so every archive miss costs a second full compile, plus a leaked directory of ~350 KB. The system cache does evict: it keeps about two generations of ~280 MB, and one full test-suite run cycles through them. That is the only case where archives help, and even there they save ~50 s of CPU on a 116 CPU-minute run, which matches the absence of a CI speedup since #930. Keeping archives as an opt-in would mean maintaining ~600 lines of cache and tests that hardly anyone would enable.

The MTLBinaryArchive wrappers in lib/mtl stay. The binary_archives and binary_archives_max_size preferences, the JULIA_METAL_BINARY_ARCHIVES* environment variables, the versioninfo() kernel-cache section and the Scratch dependency go.

Leftovers from earlier releases are not cleaned up automatically:

• leaked Metal working directories: rm -rf "$(getconf DARWIN_USER_CACHE_DIR)com.apple.gpuarchiver"/gpuarchiver-* (while no Julia session is running);
• the archives themselves (normally pruned to 256 MB per device): rm -rf ~/.julia/scratchspaces/dde4c033-4e86-420c-a63e-0dd931031962/binary_archives.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 87.16%. Comparing base (c51cdc2) to head (e45dee0).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1009      +/-   ##
==========================================
- Coverage   87.29%   87.16%   -0.14%     
==========================================
  Files          92       91       -1     
  Lines        6816     6675     -141     
==========================================
- Hits         5950     5818     -132     
+ Misses        866      857       -9     

☔ 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: e45dee0	Previous: c51cdc2	Ratio
array/accumulate/Float32/1d	392042 ns	393209 ns	1.00
array/accumulate/Float32/dims=1	370083 ns	370125 ns	1.00
array/accumulate/Float32/dims=1L	8859084 ns	8891250 ns	1.00
array/accumulate/Float32/dims=2	432333 ns	437125 ns	0.99
array/accumulate/Float32/dims=2L	2791292 ns	2721125 ns	1.03
array/accumulate/Int64/1d	847042 ns	848000 ns	1.00
array/accumulate/Int64/dims=1	909083 ns	907708 ns	1.00
array/accumulate/Int64/dims=1L	9569583 ns	9559042 ns	1.00
array/accumulate/Int64/dims=2	1209750 ns	1205958 ns	1.00
array/accumulate/Int64/dims=2L	6558000 ns	6545792 ns	1.00
array/broadcast	234000 ns	240208 ns	0.97
array/construct	2333 ns	2375 ns	0.98
array/permutedims/2d	454958 ns	458750 ns	0.99
array/permutedims/3d	1033125 ns	1033750 ns	1.00
array/permutedims/4d	1020083 ns	1130792 ns	0.90
array/private/copy	224708 ns	224541 ns	1.00
array/private/copyto!/cpu_to_gpu	214875 ns	214917 ns	1.00
array/private/copyto!/gpu_to_cpu	213958 ns	216250 ns	0.99
array/private/copyto!/gpu_to_gpu	222708 ns	220167 ns	1.01
array/private/iteration/findall/bool	1053625 ns	1057500 ns	1.00
array/private/iteration/findall/int	1214500 ns	1212709 ns	1.00
array/private/iteration/findfirst/bool	1011959 ns	972583 ns	1.04
array/private/iteration/findfirst/int	1019750 ns	1020583 ns	1.00
array/private/iteration/findmin/1d	1080000 ns	1088625 ns	0.99
array/private/iteration/findmin/2d	1026250 ns	1021375 ns	1.00
array/private/iteration/logical	1547625 ns	1547084 ns	1.00
array/private/iteration/scalar	1398667 ns	1403375 ns	1.00
array/random/rand/Float32	420375 ns	417583 ns	1.01
array/random/rand/Int64	497500 ns	500375 ns	0.99
array/random/rand!/Float32	403542 ns	403625 ns	1.00
array/random/rand!/Int64	426250 ns	426250 ns	1
array/random/randn/Float32	391250 ns	390584 ns	1.00
array/random/randn!/Float32	376042 ns	375375 ns	1.00
array/reductions/mapreduce/Float32/1d	273125 ns	274667 ns	0.99
array/reductions/mapreduce/Float32/dims=1	349959 ns	354750 ns	0.99
array/reductions/mapreduce/Float32/dims=1L	614083 ns	609625 ns	1.01
array/reductions/mapreduce/Float32/dims=2	355917 ns	354292 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	1240000 ns	1240917 ns	1.00
array/reductions/mapreduce/Int64/1d	459167 ns	455583 ns	1.01
array/reductions/mapreduce/Int64/dims=1	634792 ns	635500 ns	1.00
array/reductions/mapreduce/Int64/dims=1L	1009750 ns	1015750 ns	0.99
array/reductions/mapreduce/Int64/dims=2	789500 ns	787291 ns	1.00
array/reductions/mapreduce/Int64/dims=2L	2206666 ns	2191209 ns	1.01
array/reductions/reduce/Float32/1d	273042 ns	274084 ns	1.00
array/reductions/reduce/Float32/dims=1	355083 ns	355625 ns	1.00
array/reductions/reduce/Float32/dims=1L	622208 ns	620375 ns	1.00
array/reductions/reduce/Float32/dims=2	247459 ns	251208 ns	0.99
array/reductions/reduce/Float32/dims=2L	488583 ns	471792 ns	1.04
array/reductions/reduce/Int64/1d	460667 ns	458333 ns	1.01
array/reductions/reduce/Int64/dims=1	641042 ns	641333 ns	1.00
array/reductions/reduce/Int64/dims=1L	1006792 ns	1010167 ns	1.00
array/reductions/reduce/Int64/dims=2	260375 ns	257625 ns	1.01
array/reductions/reduce/Int64/dims=2L	659417 ns	670834 ns	0.98
array/shared/copy	127500 ns	130250 ns	0.98
array/shared/copyto!/cpu_to_gpu	37208 ns	37291 ns	1.00
array/shared/copyto!/gpu_to_cpu	36459 ns	37875 ns	0.96
array/shared/copyto!/gpu_to_gpu	36750 ns	38209 ns	0.96
array/shared/iteration/findall/bool	1058084 ns	1048458 ns	1.01
array/shared/iteration/findall/int	1221583 ns	1223209 ns	1.00
array/shared/iteration/findfirst/bool	977333 ns	976875 ns	1.00
array/shared/iteration/findfirst/int	988250 ns	981250 ns	1.01
array/shared/iteration/findmin/1d	1049958 ns	1050875 ns	1.00
array/shared/iteration/findmin/2d	1022666 ns	1027416 ns	1.00
array/shared/iteration/logical	1512875 ns	1525334 ns	0.99
array/shared/iteration/scalar	75.35040983606558 ns	75.89014373716633 ns	0.99
array/sorting/1d	1916500 ns	1921125 ns	1.00
array/sorting/2d	8270958 ns	8319917 ns	0.99
integration/byval/reference	1120167 ns	1120625 ns	1.00
integration/byval/slices=1	1124208 ns	1122291 ns	1.00
integration/byval/slices=2	2028334 ns	2029542 ns	1.00
integration/byval/slices=3	6821750 ns	6806458 ns	1.00
integration/metaldevrt	396875 ns	396167 ns	1.00
kernel/indexing	213625 ns	207625 ns	1.03
kernel/indexing_checked	395458 ns	395208 ns	1.00
kernel/launch	2115.777777777778 ns	2101.8888888888887 ns	1.01
kernel/rand	318875 ns	408958 ns	0.78
latency/import	1609990875 ns	1862270958 ns	0.86
latency/precompile	30858224084 ns	31631620958 ns	0.98
latency/ttfp	2067661667 ns	2295776542 ns	0.90
metal/synchronization/context	520.0680628272252 ns	525.958115183246 ns	0.99
metal/synchronization/stream	627.9239766081871 ns	657.375796178344 ns	0.96

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