WIP: enable reusing fft plans #271

IanButterworth · 2024-01-13T05:34:13Z

Requires

v1.0.0: Add RCpair copy! for complex. Rename region->dims. Remove deprecation. HolyLab/RFFT.jl#17

TODO:

Merge v1.0.0: Add RCpair copy! for complex. Rename region->dims. Remove deprecation. HolyLab/RFFT.jl#17
Register RFFT.jl in General
Tidy (remove manual RFFT add, remove demo)
Settle on naming/API i.e. should this be setup via Algorithm.FFT(img, kernel, border)?
Docstrings

Because they are not threadsafe FFTW FFT plans are behind a lock. They should be reusable for same-sized images and kernels, so this is WIP towards exposing a way to provide plans, so that imfilter! can be run concurrently on julia threads without lock conflict.

Note that this is using JuliaLang/julia#52883 to count conflicts.

There's a demo temporarily in this PR.

Before it hits 34255 lock conflicts

% FFTW_NUM_THREADS=1 BLAS_NUM_THREADS=3 ../julia/julia --project -t6,1 -q
julia> include("demo.jl")
warm run of benchmark(mats): 5.483911 seconds (2.14 M allocations: 10.671 GiB, 16.62% gc time, 34255 lock conflicts)

This PR removes all but the lock conflicts from the singlular plans, which could be moved out of the for loop

% FFTW_NUM_THREADS=1 BLAS_NUM_THREADS=3 ../julia/julia --project -t6,1 -q
julia> include("demo.jl")
warm run of benchmark(mats): 2.440659 seconds (478.12 k allocations: 7.888 GiB, 21.53% gc time, 8 lock conflicts)

src/imfilter.jl

IanButterworth · 2024-01-13T18:06:28Z

Ok, I think this seems reasonable now. @timholy would you mind reviewing for general design and I can add tests & docs if you agree

timholy · 2024-01-13T23:00:08Z

I confess I haven't thought about these plans in ages (the code I linked earlier was probably written in the Julia 0.1 days), but nothing bad jumps out about this. Thanks for doing it!!

src/ImageFiltering.jl

IanButterworth · 2024-01-25T14:32:12Z

Comparing buffered vs standard I get different values here

ImageFiltering.jl/src/imfilter.jl

Lines 875 to 877 in 78044e7

    
           B = complex(planned_rfft1(A)) * FFTW.AbstractFFTs.to1(A) 
        
           _B = rfft(A) 
        
           @show B[1:4] _B[1:4]

B[1:4] =  ComplexF32[3065.7727f0 + 113.279686f0im, -254.48924f0 - 85.513504f0im, -369.77383f0 - 146.0826f0im, -267.98865f0 - 201.43063f0im]
_B[1:4] = ComplexF32[4595.6045f0 + 0.0f0im,         3.6047888f0 + 15.441383f0im,  14.378552f0 - 19.552792f0im, -18.684448f0 - 19.284311f0im]

My trial and error phase of debugging isn't being very fruitful, so I think I'll wait for review. @timholy if you happen to have time

mkitti · 2024-01-31T21:08:24Z

Commeting to remind myself to take a look

timholy

Nothing jumps out as a serious problem just from a read through the code, but it's probably been a decade since I put the serious work into writing RFFT.jl (it used to be part of a more monolithic repository so the git history may not reflect that). Do the disagreements only arise when you use multiple threads?

demo.jl

src/RFFT.jl

codecov · 2024-03-11T19:04:31Z

Codecov Report

Attention: Patch coverage is 15.00000% with 34 lines in your changes are missing coverage. Please review.

Project coverage is 89.78%. Comparing base (66cf9d9) to head (6ef477e).

❗ Current head 6ef477e differs from pull request most recent head 9db8caf. Consider uploading reports for the commit 9db8caf to get more accurate results

Files	Patch %	Lines
src/imfilter.jl	10.52%	34 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #271      +/-   ##
==========================================
- Coverage   94.25%   89.78%   -4.48%     
==========================================
  Files          13       13              
  Lines        1706     1742      +36     
==========================================
- Hits         1608     1564      -44     
- Misses         98      178      +80

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

IanButterworth · 2024-03-11T19:24:30Z

Ok. Tests are passing here.

Note that planned_fft doesn't support non-float eltypes currently because RFFT.RCpair only supports floats (and other reasons, I guess).

However, demo.jl still fails

f1[1:4] = [0.11624783622854669, 0.15113825707056688, 0.16887924282352731, 0.20249257393260642]
f2[1:4] = [0.17744583830964283, -0.0655415564571991, -0.10802786527654895, 0.11185075216343188]
ERROR: LoadError: f1 !≈ f2

timholy

Thanks for this! Not only is the new functionality appreciated, but the various cleanups are too.

I haven't played with this myself, but does demo pass now? If so, it seems like you're almost over the hump. I don't see anything objectionable here (assuming you'll get to a point of full test coverage for the new functionality), so feel free to run with this.

Project.toml

demo.jl

timholy · 2024-04-10T10:13:43Z

src/imfilter.jl

+    return function (arr::AbstractArray{T}) where {T}
+        copy!(buf, OffsetArrays.no_offset_view(arr))
+        return plan(buf)
+    end


No test coverage here and below

Thanks. I hadn't noticed this. Sadly it turns out the new functionality isn't passing yet..

test/gabor.jl

test/nd.jl

timholy · 2024-04-10T10:29:20Z

src/imfilter.jl

+function filtfft(A::AbstractArray{C}, krn, planned_rfft1::Function, planned_rfft2::Function, planned_irfft::Function) where {C<:Colorant}
+    Av, dims = channelview_dims(A)
+    kernrs = kreshape(C, krn)
+    B = complex(planned_rfft1(Av, dims)) # TODO: dims is not supported by planned_rfft1


I don't think dims can be a point of flexibility: these plans are specific to the memory layout of the array. (The planning explores various implementations and picks the fastest discovered; performance is strongly dependent on memory layout, so the choice for one layout may not be the same as another.) You'd have to create a plan specifically to the colorant array-type.

I've added this comment to the TODO. I think I'd prefer to fix this in a follow on PR, if that sounds ok

IanButterworth marked this pull request as draft January 13, 2024 05:34

This comment was marked as outdated.

Sign in to view

IanButterworth mentioned this pull request Jan 13, 2024

imfilter "plan" ? #219

Open

This comment was marked as outdated.

Sign in to view

IanButterworth commented Jan 13, 2024

View reviewed changes

src/imfilter.jl Outdated Show resolved Hide resolved

IanButterworth changed the title ~~WIP: allow providing fft plans~~ WIP: enable reusing fft plans Jan 13, 2024

IanButterworth mentioned this pull request Jan 14, 2024

Add count of lock conflicts to @time etc. Add @lock_conflicts JuliaLang/julia#52883

Merged

JeffFessler reviewed Jan 22, 2024

View reviewed changes

src/ImageFiltering.jl Outdated Show resolved Hide resolved

IanButterworth force-pushed the ib/threading branch from 9183fb7 to 17e9b0a Compare January 23, 2024 18:57

This comment was marked as resolved.

Sign in to view

timholy reviewed Mar 1, 2024

View reviewed changes

demo.jl Outdated Show resolved Hide resolved

src/RFFT.jl Outdated Show resolved Hide resolved

src/RFFT.jl Outdated Show resolved Hide resolved

This comment was marked as outdated.

Sign in to view

IanButterworth mentioned this pull request Mar 11, 2024

v1.0.0: Add RCpair copy! for complex. Rename region->dims. Remove deprecation. HolyLab/RFFT.jl#17

Merged

IanButterworth force-pushed the ib/threading branch from a79e807 to faa2fb6 Compare March 11, 2024 16:50

allow providing fft plans

25dbbbe

IanButterworth force-pushed the ib/threading branch from 5087d8f to 25dbbbe Compare March 11, 2024 18:54

simplify demo

a948fae

IanButterworth force-pushed the ib/threading branch 2 times, most recently from e32b32f to a948fae Compare March 11, 2024 20:13

timholy reviewed Apr 10, 2024

View reviewed changes

IanButterworth added 4 commits May 7, 2024 09:15

RFFT -> RealFFTs & tidyup

6ef477e

use test_logs for overflow warnings

2988c21

bump julia compat to 1.10

270e59f

expand TODO based on comment

9a290c9

IanButterworth added 2 commits May 7, 2024 10:05

add to docs

7ff1a55

fix test coverage

9db8caf

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

WIP: enable reusing fft plans #271

WIP: enable reusing fft plans #271

IanButterworth commented Jan 13, 2024 •

edited

This comment was marked as outdated.

This comment was marked as outdated.

IanButterworth commented Jan 13, 2024

timholy commented Jan 13, 2024

This comment was marked as resolved.

IanButterworth commented Jan 25, 2024

mkitti commented Jan 31, 2024

timholy left a comment

This comment was marked as outdated.

codecov bot commented Mar 11, 2024 •

edited

IanButterworth commented Mar 11, 2024 •

edited

timholy left a comment

timholy Apr 10, 2024

IanButterworth May 7, 2024

timholy Apr 10, 2024

IanButterworth May 7, 2024

WIP: enable reusing fft plans #271

Are you sure you want to change the base?

WIP: enable reusing fft plans #271

Conversation

IanButterworth commented Jan 13, 2024 • edited

This comment was marked as outdated.

This comment was marked as outdated.

IanButterworth commented Jan 13, 2024

timholy commented Jan 13, 2024

This comment was marked as resolved.

IanButterworth commented Jan 25, 2024

mkitti commented Jan 31, 2024

timholy left a comment

Choose a reason for hiding this comment

This comment was marked as outdated.

codecov bot commented Mar 11, 2024 • edited

Codecov Report

IanButterworth commented Mar 11, 2024 • edited

timholy left a comment

Choose a reason for hiding this comment

timholy Apr 10, 2024

Choose a reason for hiding this comment

IanButterworth May 7, 2024

Choose a reason for hiding this comment

timholy Apr 10, 2024

Choose a reason for hiding this comment

IanButterworth May 7, 2024

Choose a reason for hiding this comment

IanButterworth commented Jan 13, 2024 •

edited

codecov bot commented Mar 11, 2024 •

edited

IanButterworth commented Mar 11, 2024 •

edited