Enzyme reverse-mode autodiff extension

[danvinci]

Adds ext/EnzymeCoreExt.jl: Enzyme reverse-mode autodiff for Metal, mirroring CUDA.jl's and AMDGPU.jl's extensions. Implements the autodiff side of #414.

What differentiates on Metal:

• user-written @metal kernels - via device-side split-mode autodiff (meta_kernels.jl); the generic core, same shape as AMDGPU's AMDGPUEnzymeCoreExt
• sum reductions
• matrix multiply (A*B) - closed-form reverse, since MPS/MPSGraph is opaque to Enzyme

device / global_queue / synchronize are marked inactive to keep Enzyme out of Metal's Objective-C launch path; CUDA/AMDGPU need no equivalent.

Tested in test/enzyme.jl: sum, A*B, and a hand-written kernel match closed-form gradients.

Draft - open items:

• GPUArrays broadcast (y .= f.(x)) isn't covered yet; its reverse needs per-element autodiff of the fused broadcast. Maintainer input welcome.
• matmul needs set_runtime_activity.

Surfaced an unrelated Enzyme GPU tape_type bug while building this: EnzymeAD/Enzyme.jl#3265 (not required here - the launch rules pass explicit ModifiedBetween, as CUDA/AMDGPU do).

Drafted with Opus 4.8, review welcome.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 86.51%. Comparing base (59e377f) to head (458dd08).
⚠️ Report is 2 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #867      +/-   ##
==========================================
+ Coverage   86.25%   86.51%   +0.25%     
==========================================
  Files          76       76              
  Lines        5130     5132       +2     
==========================================
+ Hits         4425     4440      +15     
+ Misses        705      692      -13     

☔ 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: 458dd08	Previous: 59e377f	Ratio
array/accumulate/Float32/1d	568542 ns	573625 ns	0.99
array/accumulate/Float32/dims=1	568000 ns	571125 ns	0.99
array/accumulate/Float32/dims=1L	8858437.5 ns	8850292 ns	1.00
array/accumulate/Float32/dims=2	621500 ns	625125 ns	0.99
array/accumulate/Float32/dims=2L	3347250 ns	3358312 ns	1.00
array/accumulate/Int64/1d	895875 ns	894208 ns	1.00
array/accumulate/Int64/dims=1	1101958 ns	1095792 ns	1.01
array/accumulate/Int64/dims=1L	9576687 ns	9580896 ns	1.00
array/accumulate/Int64/dims=2	1463937.5 ns	1462208 ns	1.00
array/accumulate/Int64/dims=2L	6597416 ns	6584667 ns	1.00
array/broadcast	368959 ns	364209 ns	1.01
array/construct	3334 ns	3334 ns	1
array/permutedims/2d	630667 ns	633458 ns	1.00
array/permutedims/3d	1094709 ns	1098354 ns	1.00
array/permutedims/4d	1652250 ns	1590750.5 ns	1.04
array/private/copy	400854.5 ns	407917 ns	0.98
array/private/copyto!/cpu_to_gpu	378208 ns	382125 ns	0.99
array/private/copyto!/gpu_to_cpu	379770.5 ns	377459 ns	1.01
array/private/copyto!/gpu_to_gpu	355750 ns	357750 ns	0.99
array/private/iteration/findall/bool	1142541 ns	1140208 ns	1.00
array/private/iteration/findall/int	1280250 ns	1280458 ns	1.00
array/private/iteration/findfirst/bool	1382437.5 ns	1428334 ns	0.97
array/private/iteration/findfirst/int	1489687.5 ns	1466499.5 ns	1.02
array/private/iteration/findmin/1d	1576625 ns	1567125 ns	1.01
array/private/iteration/findmin/2d	1273125 ns	1270709 ns	1.00
array/private/iteration/logical	1892104.5 ns	1933750 ns	0.98
array/private/iteration/scalar	2802417 ns	2445416 ns	1.15
array/random/rand/Float32	622334 ns	616666 ns	1.01
array/random/rand/Int64	631291 ns	629708 ns	1.00
array/random/rand!/Float32	522479 ns	530625 ns	0.98
array/random/rand!/Int64	490500 ns	495291.5 ns	0.99
array/random/randn/Float32	565104 ns	570667 ns	0.99
array/random/randn!/Float32	481291 ns	477000 ns	1.01
array/reductions/mapreduce/Float32/1d	692833 ns	715520.5 ns	0.97
array/reductions/mapreduce/Float32/dims=1	487000 ns	492000 ns	0.99
array/reductions/mapreduce/Float32/dims=1L	689083 ns	689083 ns	1
array/reductions/mapreduce/Float32/dims=2	492270.5 ns	494792 ns	0.99
array/reductions/mapreduce/Float32/dims=2L	1339417 ns	1335459 ns	1.00
array/reductions/mapreduce/Int64/1d	911500 ns	943459 ns	0.97
array/reductions/mapreduce/Int64/dims=1	830208.5 ns	812625 ns	1.02
array/reductions/mapreduce/Int64/dims=1L	1075250 ns	1075667 ns	1.00
array/reductions/mapreduce/Int64/dims=2	965500 ns	987541 ns	0.98
array/reductions/mapreduce/Int64/dims=2L	2168395.5 ns	2168458 ns	1.00
array/reductions/reduce/Float32/1d	688375 ns	693083 ns	0.99
array/reductions/reduce/Float32/dims=1	478895.5 ns	496312.5 ns	0.96
array/reductions/reduce/Float32/dims=1L	687625 ns	687979.5 ns	1.00
array/reductions/reduce/Float32/dims=2	389792 ns	395834 ns	0.98
array/reductions/reduce/Float32/dims=2L	490375 ns	485542 ns	1.01
array/reductions/reduce/Int64/1d	915333 ns	962083 ns	0.95
array/reductions/reduce/Int64/dims=1	826458 ns	814541.5 ns	1.01
array/reductions/reduce/Int64/dims=1L	1071167 ns	1069500 ns	1.00
array/reductions/reduce/Int64/dims=2	978125 ns	973542 ns	1.00
array/reductions/reduce/Int64/dims=2L	2150417 ns	2149542 ns	1.00
array/shared/copy	154208 ns	154083 ns	1.00
array/shared/copyto!/cpu_to_gpu	39250 ns	40708 ns	0.96
array/shared/copyto!/gpu_to_cpu	40625 ns	41000 ns	0.99
array/shared/copyto!/gpu_to_gpu	41416 ns	41208 ns	1.01
array/shared/iteration/findall/bool	1148292 ns	1148500 ns	1.00
array/shared/iteration/findall/int	1287541 ns	1285750 ns	1.00
array/shared/iteration/findfirst/bool	1110791 ns	1119667 ns	0.99
array/shared/iteration/findfirst/int	1174709 ns	1207312.5 ns	0.97
array/shared/iteration/findmin/1d	1323166 ns	1317938 ns	1.00
array/shared/iteration/findmin/2d	1268583 ns	1269625 ns	1.00
array/shared/iteration/logical	1715042 ns	1709354 ns	1.00
array/shared/iteration/scalar	3406.25 ns	3614.625 ns	0.94
integration/byval/reference	1135542 ns	1141958 ns	0.99
integration/byval/slices=1	1134500 ns	1136875 ns	1.00
integration/byval/slices=2	2053541 ns	2055750 ns	1.00
integration/byval/slices=3	7327625 ns	8460958 ns	0.87
integration/metaldevrt	442333 ns	451500 ns	0.98
kernel/indexing	356167 ns	351792 ns	1.01
kernel/indexing_checked	539458 ns	534209 ns	1.01
kernel/launch	1970.8 ns	1975 ns	1.00
kernel/rand	518042 ns	528250 ns	0.98
latency/import	1782315667 ns	1791534292 ns	0.99
latency/precompile	39295008687.5 ns	39637308833.5 ns	0.99
latency/ttfp	2114608917 ns	2124080375 ns	1.00
metal/synchronization/context	535.4041450777202 ns	523.9637305699482 ns	1.02
metal/synchronization/stream	274.36102236421726 ns	277.5974025974026 ns	0.99

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