Add FFT support via AbstractFFTs interface by KaanKesginLW · Pull Request #713 · JuliaGPU/Metal.jl

KaanKesginLW · 2025-12-03T09:42:08Z

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 · 2025-12-03T11:38:44Z

Codecov Report

❌ Patch coverage is 80.98592% with 27 lines in your changes missing coverage. Please review.
✅ Project coverage is 80.78%. Comparing base (706b87f) to head (d63bf16).

Files with missing lines	Patch %	Lines
src/fft.jl	78.74%	27 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff            @@
##             main     #713    +/-   ##
========================================
  Coverage   80.77%   80.78%            
========================================
  Files          61       63     +2     
  Lines        2866     3008   +142     
========================================
+ Hits         2315     2430   +115     
- Misses        551      578    +27

☔ 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: `d63bf16`	Previous: `706b87f`	Ratio
`array/accumulate/Float32/1d`	`1128208.5` ns	`1111917` ns	`1.01`
`array/accumulate/Float32/dims=1`	`1553459` ns	`1537333` ns	`1.01`
`array/accumulate/Float32/dims=1L`	`9811959` ns	`9801916.5` ns	`1.00`
`array/accumulate/Float32/dims=2`	`1861625` ns	`1836250` ns	`1.01`
`array/accumulate/Float32/dims=2L`	`7211937.5` ns	`7183625` ns	`1.00`
`array/accumulate/Int64/1d`	`1252666` ns	`1241334` ns	`1.01`
`array/accumulate/Int64/dims=1`	`1812667` ns	`1812541` ns	`1.00`
`array/accumulate/Int64/dims=1L`	`11791646` ns	`11645625` ns	`1.01`
`array/accumulate/Int64/dims=2`	`2152708` ns	`2147520.5` ns	`1.00`
`array/accumulate/Int64/dims=2L`	`9746042` ns	`9754042` ns	`1.00`
`array/broadcast`	`1017500` ns	`1010999.5` ns	`1.01`
`array/construct`	`5625` ns	`5708` ns	`0.99`
`array/permutedims/2d`	`1152354.5` ns	`1148187.5` ns	`1.00`
`array/permutedims/3d`	`1656583` ns	`1656083` ns	`1.00`
`array/permutedims/4d`	`2628000` ns	`2633729` ns	`1.00`
`array/private/copy`	`552833` ns	`543083` ns	`1.02`
`array/private/copyto!/cpu_to_gpu`	`783084` ns	`780167` ns	`1.00`
`array/private/copyto!/gpu_to_cpu`	`784042` ns	`778333` ns	`1.01`
`array/private/copyto!/gpu_to_gpu`	`609375` ns	`611292` ns	`1.00`
`array/private/iteration/findall/bool`	`1429958` ns	`1320854` ns	`1.08`
`array/private/iteration/findall/int`	`1597208` ns	`1546271` ns	`1.03`
`array/private/iteration/findfirst/bool`	`1972875` ns	`1983083` ns	`0.99`
`array/private/iteration/findfirst/int`	`2033083` ns	`2003667` ns	`1.01`
`array/private/iteration/findmin/1d`	`2247834` ns	`2254458` ns	`1.00`
`array/private/iteration/findmin/2d`	`2018270.5` ns	`2007458` ns	`1.01`
`array/private/iteration/logical`	`2519167` ns	`2527292` ns	`1.00`
`array/private/iteration/scalar`	`5097708` ns	`5570229.5` ns	`0.92`
`array/random/rand/Float32`	`1165709` ns	`1109542` ns	`1.05`
`array/random/rand/Int64`	`1313458` ns	`1291000` ns	`1.02`
`array/random/rand!/Float32`	`967208` ns	`940625` ns	`1.03`
`array/random/rand!/Int64`	`895292` ns	`877666.5` ns	`1.02`
`array/random/randn/Float32`	`1072021` ns	`1067875` ns	`1.00`
`array/random/randn!/Float32`	`842104` ns	`835541.5` ns	`1.01`
`array/reductions/mapreduce/Float32/1d`	`1114666` ns	`1124229.5` ns	`0.99`
`array/reductions/mapreduce/Float32/dims=1`	`827437.5` ns	`832167` ns	`0.99`
`array/reductions/mapreduce/Float32/dims=1L`	`1336417` ns	`1340166.5` ns	`1.00`
`array/reductions/mapreduce/Float32/dims=2`	`843625` ns	`851250` ns	`0.99`
`array/reductions/mapreduce/Float32/dims=2L`	`1772458` ns	`1763500` ns	`1.01`
`array/reductions/mapreduce/Int64/1d`	`1544104` ns	`1551104` ns	`1.00`
`array/reductions/mapreduce/Int64/dims=1`	`1130041` ns	`1126666` ns	`1.00`
`array/reductions/mapreduce/Int64/dims=1L`	`2033749.5` ns	`2021125` ns	`1.01`
`array/reductions/mapreduce/Int64/dims=2`	`1354583` ns	`1269083` ns	`1.07`
`array/reductions/mapreduce/Int64/dims=2L`	`3581208.5` ns	`3587375` ns	`1.00`
`array/reductions/reduce/Float32/1d`	`1023291.5` ns	`1048583.5` ns	`0.98`
`array/reductions/reduce/Float32/dims=1`	`828167` ns	`835563` ns	`0.99`
`array/reductions/reduce/Float32/dims=1L`	`1346125` ns	`1342834` ns	`1.00`
`array/reductions/reduce/Float32/dims=2`	`864750` ns	`854709` ns	`1.01`
`array/reductions/reduce/Float32/dims=2L`	`1769458` ns	`1793437.5` ns	`0.99`
`array/reductions/reduce/Int64/1d`	`1517666` ns	`1509583` ns	`1.01`
`array/reductions/reduce/Int64/dims=1`	`1116584` ns	`1109041` ns	`1.01`
`array/reductions/reduce/Int64/dims=1L`	`2013000` ns	`2022959` ns	`1.00`
`array/reductions/reduce/Int64/dims=2`	`1158979.5` ns	`1149458` ns	`1.01`
`array/reductions/reduce/Int64/dims=2L`	`4162375` ns	`4176771` ns	`1.00`
`array/shared/copy`	`241125` ns	`239875` ns	`1.01`
`array/shared/copyto!/cpu_to_gpu`	`75812.5` ns	`78959` ns	`0.96`
`array/shared/copyto!/gpu_to_cpu`	`80625` ns	`80041` ns	`1.01`
`array/shared/copyto!/gpu_to_gpu`	`80000` ns	`80500` ns	`0.99`
`array/shared/iteration/findall/bool`	`1447084` ns	`1443687.5` ns	`1.00`
`array/shared/iteration/findall/int`	`1541604.5` ns	`1597292` ns	`0.97`
`array/shared/iteration/findfirst/bool`	`1563542` ns	`1560312.5` ns	`1.00`
`array/shared/iteration/findfirst/int`	`1579896` ns	`1587125` ns	`1.00`
`array/shared/iteration/findmin/1d`	`1845562.5` ns	`1847104` ns	`1.00`
`array/shared/iteration/findmin/2d`	`2020209` ns	`2008542` ns	`1.01`
`array/shared/iteration/logical`	`2409604.5` ns	`2411709` ns	`1.00`
`array/shared/iteration/scalar`	`187125` ns	`186000` ns	`1.01`
`integration/byval/reference`	`1567334` ns	`1561833` ns	`1.00`
`integration/byval/slices=1`	`1555208` ns	`1557292` ns	`1.00`
`integration/byval/slices=2`	`2611333` ns	`2611166.5` ns	`1.00`
`integration/byval/slices=3`	`8553917` ns	`7720208.5` ns	`1.11`
`integration/metaldevrt`	`870958` ns	`861375` ns	`1.01`
`kernel/indexing`	`624875` ns	`637542` ns	`0.98`
`kernel/indexing_checked`	`614208.5` ns	`659708` ns	`0.93`
`kernel/launch`	`11375` ns	`11333` ns	`1.00`
`kernel/rand`	`586458` ns	`579167` ns	`1.01`
`latency/import`	`1380589000` ns	`1375283937.5` ns	`1.00`
`latency/precompile`	`28998263750` ns	`28780620500` ns	`1.01`
`latency/ttfp`	`1647912583` ns	`1640608042` ns	`1.00`
`metal/synchronization/context`	`19541` ns	`19084` ns	`1.02`
`metal/synchronization/stream`	`18208` ns	`17917` ns	`1.02`

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

They can be added in a different PR

Claude was instructed to implement this like the MPSGraphs matmul caching

KaanKesginLW mentioned this pull request Dec 3, 2025

FFT support #270

Open

github-actions Bot reviewed Dec 3, 2025

View reviewed changes

This comment was marked as outdated.

Sign in to view

This comment was marked as spam.

Sign in to view

christiangnrd reviewed Dec 4, 2025

View reviewed changes

Comment thread .gitignore Outdated

christiangnrd reviewed Dec 4, 2025

View reviewed changes

Comment thread lib/mpsgraphs/operations.jl

KaanKesginLW force-pushed the feature/fft-support branch from 40871e3 to b88d77f Compare December 4, 2025 17:05

This comment was marked as resolved.

Sign in to view

christiangnrd force-pushed the feature/fft-support branch 2 times, most recently from 130ed6a to e3aeeea Compare January 19, 2026 01:09

This comment was marked as outdated.

Sign in to view

This comment was marked as off-topic.

Sign in to view

This comment was marked as resolved.

Sign in to view

christiangnrd marked this pull request as draft February 2, 2026 01:36

christiangnrd force-pushed the feature/fft-support branch 2 times, most recently from e8d6b2c to ffdffe8 Compare February 4, 2026 02:59

christiangnrd marked this pull request as ready for review February 4, 2026 03:31

christiangnrd force-pushed the feature/fft-support branch from ffdffe8 to 6f6aa3c Compare February 4, 2026 03:32

gbene mentioned this pull request Feb 11, 2026

Other GPU backends gbene/HighSeas.jl#7

Open

4 tasks

christiangnrd force-pushed the feature/fft-support branch 3 times, most recently from 3c46dc2 to 60077f5 Compare February 19, 2026 16:38

simone-silvestri mentioned this pull request Feb 20, 2026

Poisson eigenvalues should use Float32 for Metal GPUs CliMA/Oceananigans.jl#5316

Open

christiangnrd force-pushed the feature/fft-support branch from 60077f5 to 25108fd Compare February 24, 2026 15:42

christiangnrd force-pushed the feature/fft-support branch from 25108fd to 57a51b7 Compare March 5, 2026 00:56

aplavin reviewed Mar 23, 2026

View reviewed changes

Comment thread src/fft.jl

christiangnrd force-pushed the feature/fft-support branch from 57a51b7 to e0db641 Compare March 24, 2026 16:52

christiangnrd force-pushed the feature/fft-support branch from e0db641 to c115f38 Compare April 8, 2026 21:14

christiangnrd added 29 commits May 18, 2026 16:46

Remove hallucinated kwargs

537b337

Add more CUDA tests

7f3a1a1

Temp disable some new tests

bb402e4

[NFC] Reorganize code

9aa1dac

[NFC] Cleanup

8de6db7

Rename to MPSGraphFFTDescriptor

6bd2d45

assert_applicable

23d9d3a

Add inplace stuff

98764ca

Merge MtfFFTPlan and MtlFFTInplacePlan

b6e17ec

Move some code around

de5f6c2

Remove unecessary helper function

c4ac247

Add output type parameter

43ad756

Big merge

18e287e

Fix show

d8a6b63

Metaprogram ALL the things

bc19414

Minor refactor

ea9d35f

Bring implementation more in line with the AbstractFFTs interface

3fb0241

Add some missing CUDA stuff

e2addbc

autoreleasepool tweaks

c32ee3e

More validation + fixup tests

7c58133

Rename stuff for consistency with CUDA fft test

bf39b14

Integer FFTs

db8abf2

Remove unused MPSGraph operations

2338fe9

They can be added in a different PR

Cleanup

9ed8793

Comment cleanup

be7ae48

Cache plans

7a87dd3

Claude was instructed to implement this like the MPSGraphs matmul caching

Copilot suggestion

7a6fa34

Fix potentially unsafe cache access

4848aff

Address feedback

d63bf16

christiangnrd force-pushed the feature/fft-support branch from 3d73d20 to d63bf16 Compare May 19, 2026 20:29

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add FFT support via AbstractFFTs interface#713

Add FFT support via AbstractFFTs interface#713
KaanKesginLW wants to merge 43 commits into
JuliaGPU:mainfrom
KaanKesginLW:feature/fft-support

KaanKesginLW commented Dec 3, 2025 •

edited by christiangnrd

Loading

Uh oh!

codecov Bot commented Dec 3, 2025 •

edited

Loading

Uh oh!

github-actions Bot left a comment •

edited

Loading

Uh oh!

This comment was marked as outdated.

This comment was marked as outdated.

This comment was marked as spam.

Uh oh!

Uh oh!

This comment was marked as resolved.

This comment was marked as outdated.

Uh oh!

This comment was marked as off-topic.

This comment was marked as resolved.

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

9 participants

Conversation

KaanKesginLW commented Dec 3, 2025 • edited by christiangnrd Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Performance

Example Usage

Uh oh!

codecov Bot commented Dec 3, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

github-actions Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Metal Benchmarks

Uh oh!

This comment was marked as outdated.

This comment was marked as outdated.

This comment was marked as spam.

Uh oh!

Uh oh!

This comment was marked as resolved.

This comment was marked as outdated.

Uh oh!

This comment was marked as off-topic.

This comment was marked as resolved.

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

9 participants

KaanKesginLW commented Dec 3, 2025 •

edited by christiangnrd

Loading

codecov Bot commented Dec 3, 2025 •

edited

Loading

github-actions Bot left a comment •

edited

Loading