Add FFT support via AbstractFFTs interface

[KaanKesginLW]

Adds FFT support for MtlArray via the AbstractFFTs.jl interface.

HEAVILY based on CUDA.jl's AbstractFFTs.jl interface implementation using MPSGraph functionality.

using Metal

x = MtlArray(randn(ComplexF32, 2048, 2048))
y = fft(x)  # Just works!

Performance

Benchmarked on Apple M2 Max with 30-core GPU against FFTW.jl on CPU:

Size	CPU (FFTW)	GPU (Metal)
512×512	4.1ms	5.3ms
1024×1024	19.7ms	8.5ms
2048×2048	119.7ms	10.5ms
4096×4096	460.6ms	15.8ms

Example Usage

using Metal

# Complex FFT
x = MtlArray(randn(ComplexF32, 1024, 1024))
y = fft(x)
z = ifft(y)  # z ≈ x

# Real FFT  
r = MtlArray(randn(Float32, 1024, 1024))
c = rfft(r)           # Real → Complex
r2 = irfft(c, 1024)   # Complex → Real, r2 ≈ r

# FFT along specific dimensions
y = fft(x, 1)         # First dimension only
y = fft(x, (1, 2))    # Batched transform

# Plan reuse
x = MtlArray(randn(ComplexF32, 1024, 1024))
another_x = MtlArray(randn(ComplexF32, 1024, 1024))
p = plan_fft(x)
y1 = p * x
y2 = p * another_x    # Same plan, different data

Close #270

[codecov]

Codecov Report

❌ Patch coverage is 81.63265% with 27 lines in your changes missing coverage. Please review.
✅ Project coverage is 82.99%. Comparing base (b94fd4b) to head (d9d399a).

Files with missing lines	Patch %	Lines
src/fft.jl	79.54%	27 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #713      +/-   ##
==========================================
- Coverage   83.09%   82.99%   -0.10%     
==========================================
  Files          62       64       +2     
  Lines        2851     2999     +148     
==========================================
+ Hits         2369     2489     +120     
- Misses        482      510      +28     

☔ View full report in Codecov by Sentry.
📢 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: 9bd038f	Previous: b94fd4b	Ratio
array/accumulate/Float32/1d	1063959 ns	1098958 ns	0.97
array/accumulate/Float32/dims=1	1615375 ns	1554708 ns	1.04
array/accumulate/Float32/dims=1L	10382979 ns	9848583.5 ns	1.05
array/accumulate/Float32/dims=2	2040354 ns	1886771 ns	1.08
array/accumulate/Float32/dims=2L	7545979.5 ns	7256459 ns	1.04
array/accumulate/Int64/1d	1287958 ns	1261958 ns	1.02
array/accumulate/Int64/dims=1	1886709 ns	1824291.5 ns	1.03
array/accumulate/Int64/dims=1L	12201292 ns	11664208.5 ns	1.05
array/accumulate/Int64/dims=2	2271083.5 ns	2170333.5 ns	1.05
array/accumulate/Int64/dims=2L	9885208 ns	10120062.5 ns	0.98
array/broadcast	549000.5 ns	605916 ns	0.91
array/construct	6750 ns	6292 ns	1.07
array/permutedims/2d	1216916.5 ns	1168125 ns	1.04
array/permutedims/3d	1807000 ns	1673084 ns	1.08
array/permutedims/4d	2506313 ns	2365959 ns	1.06
array/private/copy	821834 ns	545792 ns	1.51
array/private/copyto!/cpu_to_gpu	724208 ns	802916 ns	0.90
array/private/copyto!/gpu_to_cpu	674583.5 ns	801917 ns	0.84
array/private/copyto!/gpu_to_gpu	522249.5 ns	634458 ns	0.82
array/private/iteration/findall/bool	1464854 ns	1402750 ns	1.04
array/private/iteration/findall/int	1588875 ns	1564021 ns	1.02
array/private/iteration/findfirst/bool	2058334 ns	2055916 ns	1.00
array/private/iteration/findfirst/int	2117375 ns	2064479.5 ns	1.03
array/private/iteration/findmin/1d	2597812.5 ns	2499959 ns	1.04
array/private/iteration/findmin/2d	1850812.5 ns	1790791 ns	1.03
array/private/iteration/logical	2697166 ns	2631896 ns	1.02
array/private/iteration/scalar	3268062 ns	5047625 ns	0.65
array/random/rand/Float32	840396 ns	582958 ns	1.44
array/random/rand/Int64	926000 ns	775667 ns	1.19
array/random/rand!/Float32	543354 ns	574750 ns	0.95
array/random/rand!/Int64	532709 ns	550792 ns	0.97
array/random/randn/Float32	1045959 ns	1006937.5 ns	1.04
array/random/randn!/Float32	721688 ns	755666 ns	0.96
array/reductions/mapreduce/Float32/1d	808125 ns	1029500 ns	0.78
array/reductions/mapreduce/Float32/dims=1	815875 ns	840875 ns	0.97
array/reductions/mapreduce/Float32/dims=1L	1360958 ns	1324000 ns	1.03
array/reductions/mapreduce/Float32/dims=2	848500 ns	860875 ns	0.99
array/reductions/mapreduce/Float32/dims=2L	1820166.5 ns	1799541 ns	1.01
array/reductions/mapreduce/Int64/1d	1311062.5 ns	1374875 ns	0.95
array/reductions/mapreduce/Int64/dims=1	1123083 ns	1097625 ns	1.02
array/reductions/mapreduce/Int64/dims=1L	2005208 ns	2002854 ns	1.00
array/reductions/mapreduce/Int64/dims=2	1162042 ns	1145000 ns	1.01
array/reductions/mapreduce/Int64/dims=2L	3647916 ns	3614000 ns	1.01
array/reductions/reduce/Float32/1d	832500 ns	1028437.5 ns	0.81
array/reductions/reduce/Float32/dims=1	808125 ns	832667 ns	0.97
array/reductions/reduce/Float32/dims=1L	1360333 ns	1318416.5 ns	1.03
array/reductions/reduce/Float32/dims=2	844270.5 ns	853041.5 ns	0.99
array/reductions/reduce/Float32/dims=2L	1816916.5 ns	1810250 ns	1.00
array/reductions/reduce/Int64/1d	1317166 ns	1516958 ns	0.87
array/reductions/reduce/Int64/dims=1	1130667 ns	1095375 ns	1.03
array/reductions/reduce/Int64/dims=1L	2042083 ns	2023499.5 ns	1.01
array/reductions/reduce/Int64/dims=2	1170625 ns	1240750 ns	0.94
array/reductions/reduce/Int64/dims=2L	4132354.5 ns	4233875 ns	0.98
array/shared/copy	215000 ns	252417 ns	0.85
array/shared/copyto!/cpu_to_gpu	83625 ns	80750 ns	1.04
array/shared/copyto!/gpu_to_cpu	84000 ns	80667 ns	1.04
array/shared/copyto!/gpu_to_gpu	84667 ns	83083 ns	1.02
array/shared/iteration/findall/bool	1478334 ns	1427208.5 ns	1.04
array/shared/iteration/findall/int	1586958 ns	1559875 ns	1.02
array/shared/iteration/findfirst/bool	1672604 ns	1649000 ns	1.01
array/shared/iteration/findfirst/int	1713729 ns	1672458 ns	1.02
array/shared/iteration/findmin/1d	2206812.5 ns	2115583 ns	1.04
array/shared/iteration/findmin/2d	1871021 ns	1792625 ns	1.04
array/shared/iteration/logical	2259416 ns	2292167 ns	0.99
array/shared/iteration/scalar	204562.5 ns	199958 ns	1.02
integration/byval/reference	1582563 ns	1544250 ns	1.02
integration/byval/slices=1	1595791 ns	1560229.5 ns	1.02
integration/byval/slices=2	2704083 ns	2598333.5 ns	1.04
integration/byval/slices=3	19409291.5 ns	8092333 ns	2.40
integration/metaldevrt	859791 ns	868125 ns	0.99
kernel/indexing	473937.5 ns	592667 ns	0.80
kernel/indexing_checked	500021 ns	598292 ns	0.84
kernel/launch	12625 ns	11791.5 ns	1.07
kernel/rand	538000 ns	570709 ns	0.94
latency/import	1446676250 ns	1425597062.5 ns	1.01
latency/precompile	25920507833 ns	25453724708 ns	1.02
latency/ttfp	2381656291.5 ns	2341177208 ns	1.02
metal/synchronization/context	19750 ns	19667 ns	1.00
metal/synchronization/stream	19375 ns	18459 ns	1.05

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

[aplavin]

each execution creates a new MPSGraph? it causes quite some latency, would be great to cache graph somehow

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Adds FFT support for MtlArray by implementing the AbstractFFTs.jl interface on top of MPSGraph FFT operations, along with a comprehensive test suite and required dependency wiring.

Changes:

• Implemented AbstractFFTs plans and execution for Metal-backed arrays using cached MPSGraph FFT graphs.
• Added MPSGraphs bindings for FFT descriptor creation and FFT operations.
• Added FFT-focused tests (including Float16 CPU shims) and updated project dependencies.

Reviewed changes

Copilot reviewed 6 out of 7 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
src/fft.jl	Implements AbstractFFTs plans/execution for MtlArray, including MPSGraph graph caching.
lib/mpsgraphs/fft.jl	Adds Objective-C bindings for MPSGraph FFT descriptor and FFT ops.
lib/mpsgraphs/MPSGraphs.jl	Includes the new FFT bindings module.
src/Metal.jl	Wires FFT module into the main package and adds Reexport.
Project.toml	Adds AbstractFFTs and Reexport deps/compat.
test/Project.toml	Adds AbstractFFTs and FFTW for testing.
test/fft.jl	Adds a large FFT test suite and Float16 CPU reference shims.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Reading from a Dict concurrently with writes is not thread-safe in Julia; the lockless get(graph_cache, ...) can race with the get! mutation and lead to memory corruption/crashes. Protect all Dict accesses (including reads) with the lock, or switch to a thread-safe cache strategy (e.g., always @lock around get! / get, or use a concurrency-safe map implementation if available in this codebase).

[christiangnrd]

@maleadt This failure happened because AppleAccelerate.jl was loaded in the flopscomp example test which heppened to be on the same runner as the fft tests. Should we just remove AppleAccelerate?

Also, Copilot seems to think that Dict reads while writes are potentially happening isn't thread safe. If that's true we'll need to fix the MPSGraph caching too.

[maleadt]

@maleadt This failure happened because AppleAccelerate.jl was loaded in the flopscomp example test which heppened to be on the same runner as the fft tests. Should we just remove AppleAccelerate?

Yeah I guess. It's unfortunate that AppleAccelerate breaks stuff.

Also, Copilot seems to think that Dict reads while writes are potentially happening isn't thread safe. If that's true we'll need to fix the MPSGraph caching too.

Yes, Dict isn't threadsafe. You'd want to wrap those concurrent accesses in a ReentrantLock.