Reduce reliance on ObjC blocks

[maleadt]

ObjC blocks are executed on Metal's stack, which may have a more limited stack size that could explain the crashes seen in #822 (comment) when compilation is triggered on that stack. So instead, rely on a deferred mechanism that triggers when synchronizing. This should ideally also improve launch overhead because we don't need to allocate a block anymore.

[codecov]

Codecov Report

❌ Patch coverage is 94.73684% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 84.27%. Comparing base (623b6a9) to head (898b1fe).
⚠️ Report is 3 commits behind head on main.

Files with missing lines	Patch %	Lines
src/state.jl	92.85%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #845      +/-   ##
==========================================
+ Coverage   84.25%   84.27%   +0.01%     
==========================================
  Files          68       68              
  Lines        4110     4140      +30     
==========================================
+ Hits         3463     3489      +26     
- Misses        647      651       +4     

☔ 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: 898b1fe	Previous: 623b6a9	Ratio
array/accumulate/Float32/1d	830270.5 ns	860291 ns	0.97
array/accumulate/Float32/dims=1	968833 ns	962875 ns	1.01
array/accumulate/Float32/dims=1L	9986604 ns	9867583 ns	1.01
array/accumulate/Float32/dims=2	1239042 ns	1271313 ns	0.97
array/accumulate/Float32/dims=2L	6967375 ns	6945417 ns	1.00
array/accumulate/Int64/1d	925708 ns	955625 ns	0.97
array/accumulate/Int64/dims=1	1118750 ns	1074479 ns	1.04
array/accumulate/Int64/dims=1L	11600812.5 ns	11749458 ns	0.99
array/accumulate/Int64/dims=2	1420000 ns	1436813 ns	0.99
array/accumulate/Int64/dims=2L	9570417 ns	9559125 ns	1.00
array/broadcast	361333 ns	356666.5 ns	1.01
array/construct	3292 ns	3208 ns	1.03
array/permutedims/2d	544167 ns	615417 ns	0.88
array/permutedims/3d	1063375 ns	1087833 ns	0.98
array/permutedims/4d	1735500 ns	1937500 ns	0.90
array/private/copy	384854 ns	437209 ns	0.88
array/private/copyto!/cpu_to_gpu	377792 ns	380750 ns	0.99
array/private/copyto!/gpu_to_cpu	374667 ns	357625 ns	1.05
array/private/copyto!/gpu_to_gpu	354667 ns	349500 ns	1.01
array/private/iteration/findall/bool	1056583 ns	1054667 ns	1.00
array/private/iteration/findall/int	1178833 ns	1189125 ns	0.99
array/private/iteration/findfirst/bool	1418312 ns	1319958 ns	1.07
array/private/iteration/findfirst/int	1426375 ns	1501292 ns	0.95
array/private/iteration/findmin/1d	1545541 ns	1545624.5 ns	1.00
array/private/iteration/findmin/2d	1306125 ns	1302458 ns	1.00
array/private/iteration/logical	1575834 ns	1610583.5 ns	0.98
array/private/iteration/scalar	2495458 ns	2355875 ns	1.06
array/random/rand/Float32	562729 ns	555458 ns	1.01
array/random/rand/Int64	640500 ns	672541.5 ns	0.95
array/random/rand!/Float32	522250 ns	517208 ns	1.01
array/random/rand!/Int64	491250 ns	483750 ns	1.02
array/random/randn/Float32	483083 ns	567437.5 ns	0.85
array/random/randn!/Float32	476166 ns	473583 ns	1.01
array/reductions/mapreduce/Float32/1d	718708 ns	716833 ns	1.00
array/reductions/mapreduce/Float32/dims=1	473959 ns	473333.5 ns	1.00
array/reductions/mapreduce/Float32/dims=1L	754854 ns	742916 ns	1.02
array/reductions/mapreduce/Float32/dims=2	485916 ns	485000 ns	1.00
array/reductions/mapreduce/Float32/dims=2L	1313417 ns	1314167 ns	1.00
array/reductions/mapreduce/Int64/1d	968875 ns	924792 ns	1.05
array/reductions/mapreduce/Int64/dims=1	779729.5 ns	777479.5 ns	1.00
array/reductions/mapreduce/Int64/dims=1L	1310792 ns	1359917 ns	0.96
array/reductions/mapreduce/Int64/dims=2	964770.5 ns	956292 ns	1.01
array/reductions/mapreduce/Int64/dims=2L	2194250 ns	2194583 ns	1.00
array/reductions/reduce/Float32/1d	711250 ns	721042 ns	0.99
array/reductions/reduce/Float32/dims=1	473542 ns	471500 ns	1.00
array/reductions/reduce/Float32/dims=1L	725208 ns	729625 ns	0.99
array/reductions/reduce/Float32/dims=2	483270.5 ns	482458 ns	1.00
array/reductions/reduce/Float32/dims=2L	1292417 ns	1300937.5 ns	0.99
array/reductions/reduce/Int64/1d	911958 ns	943395.5 ns	0.97
array/reductions/reduce/Int64/dims=1	765750 ns	802312.5 ns	0.95
array/reductions/reduce/Int64/dims=1L	1294333 ns	1361771 ns	0.95
array/reductions/reduce/Int64/dims=2	925791 ns	963959 ns	0.96
array/reductions/reduce/Int64/dims=2L	2164834 ns	2175250 ns	1.00
array/shared/copy	234396 ns	236417 ns	0.99
array/shared/copyto!/cpu_to_gpu	40750 ns	39750 ns	1.03
array/shared/copyto!/gpu_to_cpu	38625 ns	41042 ns	0.94
array/shared/copyto!/gpu_to_gpu	39625 ns	41541 ns	0.95
array/shared/iteration/findall/bool	1044792 ns	1064750 ns	0.98
array/shared/iteration/findall/int	1181584 ns	1190167 ns	0.99
array/shared/iteration/findfirst/bool	1136000 ns	1131792 ns	1.00
array/shared/iteration/findfirst/int	1236750 ns	1225396 ns	1.01
array/shared/iteration/findmin/1d	1316042 ns	1323833 ns	0.99
array/shared/iteration/findmin/2d	1314333 ns	1316916.5 ns	1.00
array/shared/iteration/logical	1401291.5 ns	1459145.5 ns	0.96
array/shared/iteration/scalar	4904.714285714285 ns	4738.142857142857 ns	1.04
integration/byval/reference	1132416 ns	1138291 ns	0.99
integration/byval/slices=1	1134833 ns	1137584 ns	1.00
integration/byval/slices=2	2045250 ns	2055292 ns	1.00
integration/byval/slices=3	7496291 ns	8384542 ns	0.89
integration/metaldevrt	436875 ns	441542 ns	0.99
kernel/indexing	351895.5 ns	326042 ns	1.08
kernel/indexing_checked	529458.5 ns	521625 ns	1.02
kernel/launch	4458 ns	4833 ns	0.92
kernel/rand	526458 ns	515750 ns	1.02
latency/import	1646796458 ns	1649667041 ns	1.00
latency/precompile	36022232729 ns	35841764917 ns	1.01
latency/ttfp	1975584333 ns	1977516916 ns	1.00
metal/synchronization/context	620.9943502824859 ns	606.8121546961326 ns	1.02
metal/synchronization/stream	416.04 ns	402.7761194029851 ns	1.03

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