Generate MTL and MPS structs and enums with Clang.jl

[christiangnrd]

Currently works for structs and enums in Metal.framework with a few hacks and this branch of Clang devved: https://github.com/christiangnrd/Clang.jl/tree/objectiveC

Any advice, suggestions, or help welcome!

[maleadt]

Even if only for enums etc, this would be pretty useful! I imagined having to write our own Clang.jl-based generator in order to generate ObjC calls, but this would probably be a good first step.

[christiangnrd]

I imagined having to write our own Clang.jl-based generator in order to generate ObjC calls, but this would probably be a good first step.

I think most of it can be upstreamed to Clang.jl. The changes I've done on my branch are either bug fixes or definitions of the Generators types for Objective-C constructs.

[github-actions]

Metal Benchmarks

Benchmark suite	Current: a419a2b	Previous: 84447c4	Ratio
private array/construct	27380.85714285714 ns	26538.166666666668 ns	1.03
private array/broadcast	458833 ns	455500 ns	1.01
private array/random/randn/Float32	824917 ns	810750 ns	1.02
private array/random/randn!/Float32	650459 ns	641541 ns	1.01
private array/random/rand!/Int64	544250 ns	547500 ns	0.99
private array/random/rand!/Float32	581917 ns	581708 ns	1.00
private array/random/rand/Int64	784083.5 ns	754708 ns	1.04
private array/random/rand/Float32	622833.5 ns	576250 ns	1.08
private array/copyto!/gpu_to_gpu	642708 ns	678208 ns	0.95
private array/copyto!/cpu_to_gpu	817417 ns	821375 ns	1.00
private array/copyto!/gpu_to_cpu	624000 ns	674333 ns	0.93
private array/accumulate/1d	1354667 ns	1347708 ns	1.01
private array/accumulate/2d	1386542 ns	1384500 ns	1.00
private array/iteration/findall/int	2104209 ns	2102291.5 ns	1.00
private array/iteration/findall/bool	1809042 ns	1807917 ns	1.00
private array/iteration/findfirst/int	1693208 ns	1692020.5 ns	1.00
private array/iteration/findfirst/bool	1666500 ns	1656666 ns	1.01
private array/iteration/scalar	3546833 ns	3920667 ns	0.90
private array/iteration/logical	3217125 ns	3200708 ns	1.01
private array/iteration/findmin/1d	1767146 ns	1765458 ns	1.00
private array/iteration/findmin/2d	1354354 ns	1340875 ns	1.01
private array/reductions/reduce/1d	1036000 ns	1037000 ns	1.00
private array/reductions/reduce/2d	662500 ns	666542 ns	0.99
private array/reductions/mapreduce/1d	1034937.5 ns	1027813 ns	1.01
private array/reductions/mapreduce/2d	663792 ns	712062.5 ns	0.93
private array/permutedims/4d	2551354 ns	2564583 ns	0.99
private array/permutedims/2d	1020479 ns	1019958 ns	1.00
private array/permutedims/3d	1590396 ns	1593584 ns	1.00
private array/copy	548479.5 ns	541583 ns	1.01
latency/precompile	5775550229 ns	5246979375 ns	1.10
latency/ttfp	6740591145.5 ns	6653773250 ns	1.01
latency/import	1167900875 ns	1164822625 ns	1.00
integration/metaldevrt	718291.5 ns	713667 ns	1.01
integration/byval/slices=1	1565708.5 ns	1535999.5 ns	1.02
integration/byval/slices=3	8698562.5 ns	9724167 ns	0.89
integration/byval/reference	1566791.5 ns	1540854.5 ns	1.02
integration/byval/slices=2	2619229 ns	2638000 ns	0.99
kernel/indexing	461104 ns	476334 ns	0.97
kernel/indexing_checked	474312.5 ns	477542 ns	0.99
kernel/launch	8666 ns	8084 ns	1.07
metal/synchronization/stream	14166 ns	14625 ns	0.97
metal/synchronization/context	15042 ns	15041 ns	1.00
shared array/construct	27638.833333333332 ns	27142.333333333332 ns	1.02
shared array/broadcast	466937.5 ns	461000 ns	1.01
shared array/random/randn/Float32	810812 ns	817792 ns	0.99
shared array/random/randn!/Float32	647166 ns	636042 ns	1.02
shared array/random/rand!/Int64	546000 ns	542833 ns	1.01
shared array/random/rand!/Float32	583375 ns	577542 ns	1.01
shared array/random/rand/Int64	755354 ns	775874.5 ns	0.97
shared array/random/rand/Float32	617729 ns	573604 ns	1.08
shared array/copyto!/gpu_to_gpu	90667 ns	88375 ns	1.03
shared array/copyto!/cpu_to_gpu	87792 ns	85375 ns	1.03
shared array/copyto!/gpu_to_cpu	81209 ns	84084 ns	0.97
shared array/accumulate/1d	1338458 ns	1353166.5 ns	0.99
shared array/accumulate/2d	1386812 ns	1386916 ns	1.00
shared array/iteration/findall/int	1815750 ns	1831541.5 ns	0.99
shared array/iteration/findall/bool	1589542 ns	1577750 ns	1.01
shared array/iteration/findfirst/int	1391083 ns	1387770.5 ns	1.00
shared array/iteration/findfirst/bool	1361875 ns	1340584 ns	1.02
shared array/iteration/scalar	151584 ns	158625 ns	0.96
shared array/iteration/logical	2959875 ns	2945562.5 ns	1.00
shared array/iteration/findmin/1d	1464729.5 ns	1456666 ns	1.01
shared array/iteration/findmin/2d	1350417 ns	1357833 ns	0.99
shared array/reductions/reduce/1d	721500 ns	730167 ns	0.99
shared array/reductions/reduce/2d	662104.5 ns	669417 ns	0.99
shared array/reductions/mapreduce/1d	731042 ns	743312.5 ns	0.98
shared array/reductions/mapreduce/2d	661834 ns	667000 ns	0.99
shared array/permutedims/4d	2559146 ns	2570541.5 ns	1.00
shared array/permutedims/2d	1022750 ns	1026187.5 ns	1.00
shared array/permutedims/3d	1594438 ns	1596125 ns	1.00
shared array/copy	243604.5 ns	245812 ns	0.99

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

[maleadt]

Please ping when this is ready for review; you've marked it as such, but more commits keep appearing 🙂

Also, were the necessary changes to Clang.jl merged?

[christiangnrd]

Please ping when this is ready for review; you've marked it as such, but more commits keep appearing 🙂

I do have a terrible habit of refining my PRs after marking ready. I'll be more mindful of that in the future.

Also, were the necessary changes to Clang.jl merged?

Not yet, JuliaInterop/Clang.jl#519 and JuliaInterop/Clang.jl#522 are general improvements/bug fixes to handle the types of enums that Apple ObjectiveC headers love so much (elaborated and attributes). I just created JuliaInterop/Clang.jl#524 which once the other 2 are merged, shouldn't be too hard to review.

I just rebased and any further changes other than review changes will be for a future PR.

[maleadt]

LGTM. We could probably also remove some enums from compiler/library.jl, but that can happen in a different PR.