Default MtlArray storage to SharedStorage

[KaanKesginLW]

Summary

Change the default MtlArray storage mode from PrivateStorage to SharedStorage. On unified-memory (Apple Silicon) GPUs — essentially every supported Mac — SharedStorage is zero-copy between CPU and GPU and matches Apple's guidance. Discrete GPUs keep working (and get a one-time notice steering them to PrivateStorage).

using Metal
a = MtlArray(rand(Float32, 1024, 1024))   # SharedStorage by default now
w = unsafe_wrap(Array, a)                  # zero-copy CPU view — no Array() copy

This supersedes #717 with a clean, minimal diff (+31 / −10 across 4 files, vs #717's +655 / −606).

Motivation

• Apple's guidance. For unified memory, the Metal Best Practices Guide states "the Shared mode is usually the correct choice." The previous PrivateStorage default came from discrete-GPU guidance.
• All supported Macs are Apple Silicon (unified memory), so this is the right default for ~every user.
• Zero-copy CPU access. SharedStorage enables unsafe_wrap(Array, x), avoiding the allocate-and-copy of Array().

Benchmarks

No performance regression (M2 Max) — SharedStorage ≈ PrivateStorage:

size	op	Shared	Private
512²	broadcast	0.057 ms	0.056 ms
512²	matmul	0.286 ms	0.291 ms
1024²	broadcast	0.135 ms	0.132 ms
1024²	matmul	0.664 ms	0.788 ms

Consistent with the original 67-benchmark sweep: 78% ties, identical for copyto! / fill! / MPS-matmul, and all ties at ≥ 512 MB.

Faster CPU access via zero-copy unsafe_wrap(Array, x) instead of Array()
(M2 Max, best of 5; both paths sum() to force a full read):

size	Array() + use	unsafe_wrap + use	speedup
512 MB	17.1 ms	9.0 ms	1.9×
1 GB	37.5 ms	18.2 ms	2.1×
2 GB	249.7 ms	36.4 ms	6.9×
4 GB	539.2 ms	72.5 ms	7.4×

(The jump at ≥ 2 GB is Array()'s GPU→CPU blit crossing into chunked copies;
unsafe_wrap stays flat because it never copies.)

Implementation

• The default_storage preference now defaults to "shared". Every constructor (mtl, zeros, fill, similar, …) already resolves through DefaultStorageMode, so they follow automatically.
• The default stays a compile-time const — type-stable and with no device access during precompilation. (It is not a MTLDevice(…).hasUnifiedMemory call baked into a precompiled constant.)
• On the rare non-unified-memory (discrete) GPU, __init__ emits a one-time notice pointing the user to set_preferences!(Metal, "default_storage" => "private"). SharedStorage still works there; this just flags the faster option.
• Also fixes a latent bug in the old error message ($default_storage → $str).

Notes

Supersedes #717, closed as "this will probably eventually happen but it'll be easier to do it in a separate PR than to rebase and clean up this one." This is that clean PR: no formatting churn, no test rewrite, just the default plus the discrete-GPU notice, docs, and a focused test.

Related: #717

[codecov]

Codecov Report

❌ Patch coverage is 50.00000% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 81.30%. Comparing base (bfb9ba3) to head (12e9d2f).

Files with missing lines	Patch %	Lines
src/initialization.jl	50.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #820      +/-   ##
==========================================
- Coverage   81.43%   81.30%   -0.14%     
==========================================
  Files          66       66              
  Lines        3318     3321       +3     
==========================================
- Hits         2702     2700       -2     
- Misses        616      621       +5     

☔ 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: 12e9d2f	Previous: bfb9ba3	Ratio
array/accumulate/Float32/1d	818917 ns	813334 ns	1.01
array/accumulate/Float32/dims=1	1004917 ns	982750 ns	1.02
array/accumulate/Float32/dims=1L	10042458 ns	10003208 ns	1.00
array/accumulate/Float32/dims=2	1307166 ns	1259354 ns	1.04
array/accumulate/Float32/dims=2L	6166208.5 ns	5629958 ns	1.10
array/accumulate/Int64/1d	975125 ns	950125 ns	1.03
array/accumulate/Int64/dims=1	1107645.5 ns	1125625.5 ns	0.98
array/accumulate/Int64/dims=1L	11925167 ns	12141625 ns	0.98
array/accumulate/Int64/dims=2	1478124.5 ns	1476416 ns	1.00
array/accumulate/Int64/dims=2L	9481646 ns	9438083 ns	1.00
array/broadcast	376750 ns	374625 ns	1.01
array/construct	6125 ns	5666 ns	1.08
array/permutedims/2d	641312 ns	630750 ns	1.02
array/permutedims/3d	1129708 ns	1117000 ns	1.01
array/permutedims/4d	1992854 ns	1994209 ns	1.00
array/private/copy	440687.5 ns	412292 ns	1.07
array/private/copyto!/cpu_to_gpu	371417 ns	368583 ns	1.01
array/private/copyto!/gpu_to_cpu	368792 ns	358916 ns	1.03
array/private/copyto!/gpu_to_gpu	342917 ns	342666 ns	1.00
array/private/iteration/findall/bool	1076500 ns	1073250 ns	1.00
array/private/iteration/findall/int	1255209 ns	1252000 ns	1.00
array/private/iteration/findfirst/bool	1468708 ns	1458437 ns	1.01
array/private/iteration/findfirst/int	1508291.5 ns	1487958 ns	1.01
array/private/iteration/findmin/1d	1608750 ns	1592041 ns	1.01
array/private/iteration/findmin/2d	1310792 ns	1315875 ns	1.00
array/private/iteration/logical	1754354 ns	1743542 ns	1.01
array/private/iteration/scalar	2683833 ns	2638375.5 ns	1.02
array/random/rand/Float32	584458 ns	634917 ns	0.92
array/random/rand/Int64	697833 ns	669834 ns	1.04
array/random/rand!/Float32	587667 ns	580958 ns	1.01
array/random/rand!/Int64	508250 ns	509000 ns	1.00
array/random/randn/Float32	597416.5 ns	597958 ns	1.00
array/random/randn!/Float32	535542 ns	531209 ns	1.01
array/reductions/mapreduce/Float32/1d	342500 ns	750833 ns	0.46
array/reductions/mapreduce/Float32/dims=1	514250 ns	499041.5 ns	1.03
array/reductions/mapreduce/Float32/dims=1L	916395.5 ns	780791 ns	1.17
array/reductions/mapreduce/Float32/dims=2	511042 ns	502750 ns	1.02
array/reductions/mapreduce/Float32/dims=2L	1364583 ns	1356041 ns	1.01
array/reductions/mapreduce/Int64/1d	671750 ns	934917 ns	0.72
array/reductions/mapreduce/Int64/dims=1	795979.5 ns	786666 ns	1.01
array/reductions/mapreduce/Int64/dims=1L	1666167 ns	1712500 ns	0.97
array/reductions/mapreduce/Int64/dims=2	972209 ns	966667 ns	1.01
array/reductions/mapreduce/Int64/dims=2L	2271354.5 ns	2260917 ns	1.00
array/reductions/reduce/Float32/1d	340500 ns	743333 ns	0.46
array/reductions/reduce/Float32/dims=1	511542 ns	499625 ns	1.02
array/reductions/reduce/Float32/dims=1L	876250 ns	813625 ns	1.08
array/reductions/reduce/Float32/dims=2	509084 ns	505833 ns	1.01
array/reductions/reduce/Float32/dims=2L	1360042 ns	1346625 ns	1.01
array/reductions/reduce/Int64/1d	668500 ns	930625 ns	0.72
array/reductions/reduce/Int64/dims=1	800250 ns	783875 ns	1.02
array/reductions/reduce/Int64/dims=1L	1629125 ns	1680125 ns	0.97
array/reductions/reduce/Int64/dims=2	971542 ns	980563 ns	0.99
array/reductions/reduce/Int64/dims=2L	2272541.5 ns	2260250 ns	1.01
array/shared/copy	231917 ns	238375 ns	0.97
array/shared/copyto!/cpu_to_gpu	48083 ns	40667 ns	1.18
array/shared/copyto!/gpu_to_cpu	47250 ns	40667 ns	1.16
array/shared/copyto!/gpu_to_gpu	42333 ns	41292 ns	1.03
array/shared/iteration/findall/bool	1080083 ns	1079166 ns	1.00
array/shared/iteration/findall/int	1255333 ns	1250333 ns	1.00
array/shared/iteration/findfirst/bool	1199333 ns	1192416.5 ns	1.01
array/shared/iteration/findfirst/int	1235625 ns	1274104.5 ns	0.97
array/shared/iteration/findmin/1d	1343208 ns	1282291 ns	1.05
array/shared/iteration/findmin/2d	1326584 ns	1266125 ns	1.05
array/shared/iteration/logical	1601666 ns	1594459 ns	1.00
array/shared/iteration/scalar	8666.666666666666 ns	5868.166666666667 ns	1.48
integration/byval/reference	1162750 ns	1157708 ns	1.00
integration/byval/slices=1	1168479.5 ns	1159208 ns	1.01
integration/byval/slices=2	2092334 ns	2086791.5 ns	1.00
integration/byval/slices=3	7914208 ns	7931979 ns	1.00
integration/metaldevrt	489250 ns	468520.5 ns	1.04
kernel/indexing	370833 ns	366500 ns	1.01
kernel/indexing_checked	550208 ns	540834 ns	1.02
kernel/launch	14875 ns	13208 ns	1.13
kernel/rand	563916 ns	558042 ns	1.01
latency/import	1404497583 ns	1399005208 ns	1.00
latency/precompile	31277763959 ns	31215035583.5 ns	1.00
latency/ttfp	1710765542 ns	1710418937.5 ns	1.00
metal/synchronization/context	1045.8 ns	838.258064516129 ns	1.25
metal/synchronization/stream	696.7483443708609 ns	436.9748743718593 ns	1.59

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

[christiangnrd]

#822