Fix loading on unsupported platforms

[christiangnrd]

Closes #457
Closes #458

Backport to 1.4?

[christiangnrd]

@omlins This makes Metal precompile on Windows for me. Can you let me know it it fixes #457 for you?

[github-actions]

Metal Benchmarks

Benchmark suite	Current: a3ac9d1	Previous: ddca5c4	Ratio
private array/construct	26878.5 ns	26722.25 ns	1.01
private array/broadcast	463166 ns	467500 ns	0.99
private array/random/randn/Float32	770375 ns	751250 ns	1.03
private array/random/randn!/Float32	698833 ns	636125 ns	1.10
private array/random/rand!/Int64	568458 ns	578875 ns	0.98
private array/random/rand!/Float32	602000 ns	589417 ns	1.02
private array/random/rand/Int64	771708 ns	813250 ns	0.95
private array/random/rand/Float32	621375 ns	542500 ns	1.15
private array/copyto!/gpu_to_gpu	658895.5 ns	515812.5 ns	1.28
private array/copyto!/cpu_to_gpu	822875.5 ns	727666 ns	1.13
private array/copyto!/gpu_to_cpu	693916 ns	554625 ns	1.25
private array/accumulate/1d	1330854.5 ns	1444334 ns	0.92
private array/accumulate/2d	1408458 ns	1469333 ns	0.96
private array/iteration/findall/int	2077833 ns	2243208.5 ns	0.93
private array/iteration/findall/bool	1856416 ns	2115125 ns	0.88
private array/iteration/findfirst/int	1662875 ns	1690562.5 ns	0.98
private array/iteration/findfirst/bool	1658666 ns	1674250 ns	0.99
private array/iteration/scalar	3934708 ns	2412750 ns	1.63
private array/iteration/logical	3231791.5 ns	3443541 ns	0.94
private array/iteration/findmin/1d	1569958 ns	1771104 ns	0.89
private array/iteration/findmin/2d	1322791.5 ns	1362584 ns	0.97
private array/reductions/reduce/1d	1051104 ns	793250 ns	1.33
private array/reductions/reduce/2d	692917 ns	713896 ns	0.97
private array/reductions/mapreduce/1d	1064875 ns	864750 ns	1.23
private array/reductions/mapreduce/2d	692125 ns	712500 ns	0.97
private array/permutedims/4d	841916 ns	912916 ns	0.92
private array/permutedims/2d	933333 ns	931604.5 ns	1.00
private array/permutedims/3d	914749.5 ns	1000937.5 ns	0.91
private array/copy	512583 ns	509083 ns	1.01
latency/precompile	4389223125 ns	4397344834 ns	1.00
latency/ttfp	6674100542 ns	6908319354 ns	0.97
latency/import	720130042 ns	719607250 ns	1.00
integration/metaldevrt	739167 ns	756958 ns	0.98
integration/byval/slices=1	1598250 ns	1632792 ns	0.98
integration/byval/slices=3	8447604 ns	8919542 ns	0.95
integration/byval/reference	1628791 ns	1659354.5 ns	0.98
integration/byval/slices=2	2604958.5 ns	2684500 ns	0.97
kernel/indexing	453145.5 ns	478708 ns	0.95
kernel/indexing_checked	460354.5 ns	476750 ns	0.97
kernel/launch	8500 ns	12084 ns	0.70
metal/synchronization/stream	14750 ns	19916 ns	0.74
metal/synchronization/context	14916 ns	20209 ns	0.74
shared array/construct	25145.833333333332 ns	24899.25 ns	1.01
shared array/broadcast	464229.5 ns	470042 ns	0.99
shared array/random/randn/Float32	866125 ns	754416 ns	1.15
shared array/random/randn!/Float32	668562.5 ns	650708 ns	1.03
shared array/random/rand!/Int64	567458 ns	586083.5 ns	0.97
shared array/random/rand!/Float32	594250 ns	595750 ns	1.00
shared array/random/rand/Int64	798458 ns	809708.5 ns	0.99
shared array/random/rand/Float32	616562.5 ns	589562.5 ns	1.05
shared array/copyto!/gpu_to_gpu	87625 ns	93459 ns	0.94
shared array/copyto!/cpu_to_gpu	87125 ns	95291 ns	0.91
shared array/copyto!/gpu_to_cpu	83625 ns	84042 ns	1.00
shared array/accumulate/1d	1351500 ns	1440417 ns	0.94
shared array/accumulate/2d	1403583 ns	1482250 ns	0.95
shared array/iteration/findall/int	1838084 ns	1940666.5 ns	0.95
shared array/iteration/findall/bool	1611895.5 ns	1770833 ns	0.91
shared array/iteration/findfirst/int	1377875 ns	1395458 ns	0.99
shared array/iteration/findfirst/bool	1362541 ns	1371084 ns	0.99
shared array/iteration/scalar	156000 ns	195250 ns	0.80
shared array/iteration/logical	3048604 ns	3219395.5 ns	0.95
shared array/iteration/findmin/1d	1258541 ns	1497458 ns	0.84
shared array/iteration/findmin/2d	1326292 ns	1371541.5 ns	0.97
shared array/reductions/reduce/1d	737291.5 ns	638333.5 ns	1.16
shared array/reductions/reduce/2d	689833 ns	718750 ns	0.96
shared array/reductions/mapreduce/1d	751312.5 ns	708500 ns	1.06
shared array/reductions/mapreduce/2d	700083 ns	707208.5 ns	0.99
shared array/permutedims/4d	848958 ns	952146 ns	0.89
shared array/permutedims/2d	854625 ns	929292 ns	0.92
shared array/permutedims/3d	912542 ns	1039791 ns	0.88
shared array/copy	249375 ns	288396 ns	0.86

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

[omlins]

@omlins This makes Metal precompile on Windows for me. Can you let me know it it fixes #457 for you?

All the following runs should pass when the two issues #457 and #458 are resolved, merged and released:
https://github.com/omlins/CellArrays.jl/actions/runs/11330323307

I will surely test with your branch (once you have finished your discussions with @maleadt on the approach to take)

[christiangnrd]

@omlins You can go ahead and try out this branch. Once you confirm that it works, I'll merge this and backport to 1.4

[omlins]

@omlins You can go ahead and try out this branch. Once you confirm that it works, I'll merge this and backport to 1.4

I'm sorry, I just realized that you would need me to test it on windows. :) Unfortunately, I don't have a windows machine. Given that it works for you, I assume you can be confident enough to release it. Then, I can confirm you that it works in the CI before you do the backport.

[christiangnrd]

@omlins You can test directly on the release-1.4 branch if you prefer.

[omlins]

@christiangnrd : the CI has now passed as you have probably seen. So I can also confirm from my side that the fix worked.
(the CI with Julia 1.9 still fails, because Metal v1.2 is used:
https://github.com/omlins/CellArrays.jl/actions/runs/11380953251/job/32236749235?pr=42#step:6:53 . But this is not a real problem)