WIP: enable reusing fft plans

.github/workflows/UnitTest.yml

@@ -7,8 +7,6 @@ on:
     branches:
       - master
   pull_request:
-  schedule:
-    - cron: '20 00 1 * *'
 
 jobs:
   test:
@@ -20,23 +18,18 @@ jobs:
         os: [ubuntu-latest, windows-latest, macOS-latest]
 
     steps:
-      - uses: actions/checkout@v1.0.0
+      - uses: actions/checkout@v4
       - name: "Set up Julia"
         uses: julia-actions/setup-julia@v1
         with:
           version: ${{ matrix.julia-version }}
 
-      - name: Cache artifacts
-        uses: actions/cache@v1
-        env:
-          cache-name: cache-artifacts
-        with:
-          path: ~/.julia/artifacts
-          key: ${{ runner.os }}-test-${{ env.cache-name }}-${{ hashFiles('**/Project.toml') }}
-          restore-keys: |
-            ${{ runner.os }}-test-${{ env.cache-name }}-
-            ${{ runner.os }}-test-
-            ${{ runner.os }}-
+      - uses: julia-actions/cache@v1
+
+      - run: |
+          import Pkg
+          Pkg.add(Pkg.PackageSpec(url="https://github.com/HolyLab/RFFT.jl", rev="ib/add_copy"))
+        shell: julia --project --color=yes {0}
 
       - name: "Unit Test"
         uses: julia-actions/julia-runtest@v1

Project.toml

@@ -4,6 +4,7 @@ author = ["Tim Holy <[email protected]>", "Jan Weidner <[email protected]>"]
 version = "0.7.8"
 
 [deps]
+AbstractFFTs = "621f4979-c628-5d54-868e-fcf4e3e8185c"
 CatIndices = "aafaddc9-749c-510e-ac4f-586e18779b91"
 ComputationalResources = "ed09eef8-17a6-5b46-8889-db040fac31e3"
 DataStructures = "864edb3b-99cc-5e75-8d2d-829cb0a9cfe8"
@@ -15,6 +16,7 @@ LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 OffsetArrays = "6fe1bfb0-de20-5000-8ca7-80f57d26f881"
 PrecompileTools = "aea7be01-6a6a-4083-8856-8a6e6704d82a"
 Reexport = "189a3867-3050-52da-a836-e630ba90ab69"
+RFFT = "3bd9afcd-55df-531a-9b34-dc642dce7b95"
 SparseArrays = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
 StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
 Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
@@ -31,6 +33,7 @@ ImageCore = "0.10"
 OffsetArrays = "1.9"
 PrecompileTools = "1"
 Reexport = "1.1"
+RFFT = "0.1.1"
 StaticArrays = "0.10, 0.11, 0.12, 1.0"
 Statistics = "1"
 TiledIteration = "0.2, 0.3, 0.4, 0.5"

demo.jl

@@ -0,0 +1,67 @@
+using ImageFiltering, FFTW, LinearAlgebra, Profile, Random
+using ComputationalResources
+
+FFTW.set_num_threads(parse(Int, get(ENV, "FFTW_NUM_THREADS", "1")))
+BLAS.set_num_threads(parse(Int, get(ENV, "BLAS_NUM_THREADS", string(Threads.nthreads() ÷ 2))))
+
+function benchmark_new(mats)
+    kernel = ImageFiltering.factorkernel(Kernel.LoG(1))
+    Threads.@threads for mat in mats
+        frame_filtered = deepcopy(mat[:, :, 1])
+        r_cached = CPU1(ImageFiltering.planned_fft(frame_filtered, kernel))
+        for i in axes(mat, 3)
+            frame = @view mat[:, :, i]
+            imfilter!(r_cached, frame_filtered, frame, kernel)
+        end
+        return
+    end
+end
+function benchmark_old(mats)
+    kernel = ImageFiltering.factorkernel(Kernel.LoG(1))
+    Threads.@threads for mat in mats
+        frame_filtered = deepcopy(mat[:, :, 1])
+        r_noncached = CPU1(Algorithm.FFT())
+        for i in axes(mat, 3)
+            frame = @view mat[:, :, i]
+            imfilter!(r_noncached, frame_filtered, frame, kernel)
+        end
+        return
+    end
+end
+
+function test(mats)
+    kernel = ImageFiltering.factorkernel(Kernel.LoG(1))
+    for mat in mats
+        f1 = deepcopy(mat[:, :, 1])
+        r_cached = CPU1(ImageFiltering.planned_fft(f1, kernel))
+        f2 = deepcopy(mat[:, :, 1])
+        r_noncached = CPU1(Algorithm.FFT())
+        for i in axes(mat, 3)
+            imfilter!(r_noncached, f2, deepcopy(mat[:, :, i]), kernel)
+            imfilter!(r_cached, f1, deepcopy(mat[:, :, i]), kernel)
+            @show f1[1:4] f2[1:4]
+            f1 ≈ f2 || error("f1 !≈ f2")
+        end
+        return
+    end
+end
+
+function run()
+    Random.seed!(1)
+    nmats = 10
+    mats = [rand(Float64, rand(80:100), rand(80:100), rand(2000:3000)) for _ in 1:nmats]
+
+    benchmark_new(mats)
+    for _ in 1:3
+        @time "warm run of benchmark_new(mats)" benchmark_new(mats)
+    end
+
+    benchmark_old(mats)
+    for _ in 1:3
+        @time "warm run of benchmark_old(mats)" benchmark_old(mats)
+    end
+
+    test(mats)
+end
+
+run()

src/ImageFiltering.jl

@@ -1,12 +1,14 @@
 module ImageFiltering
 
 using FFTW
+using RFFT
 using ImageCore, FFTViews, OffsetArrays, StaticArrays, ComputationalResources, TiledIteration
 # Where possible we avoid a direct dependency to reduce the number of [compat] bounds
 # using FixedPointNumbers: Normed, N0f8 # reexported by ImageCore
 using ImageCore.MappedArrays
 using Statistics, LinearAlgebra
 using Base: Indices, tail, fill_to_length, @pure, depwarn, @propagate_inbounds
+import Base: copy!
 using OffsetArrays: IdentityUnitRange   # using the one in OffsetArrays makes this work with multiple Julia versions
 using SparseArrays   # only needed to fix an ambiguity in borderarray
 using Reexport
@@ -30,7 +32,8 @@ export Kernel, KernelFactors,
     imgradients, padarray, centered, kernelfactors, reflect,
     freqkernel, spacekernel,
     findlocalminima, findlocalmaxima,
-    blob_LoG, BlobLoG
+    blob_LoG, BlobLoG,
+    planned_fft
 
 FixedColorant{T<:Normed} = Colorant{T}
 StaticOffsetArray{T,N,A<:StaticArray} = OffsetArray{T,N,A}
@@ -50,10 +53,16 @@ function Base.transpose(A::StaticOffsetArray{T,2}) where T
 end
 
 module Algorithm
+    import FFTW
     # deliberately don't export these, but it's expected that they
     # will be used as Algorithm.FFT(), etc.
     abstract type Alg end
-    "Filter using the Fast Fourier Transform" struct FFT <: Alg end
+    "Filter using the Fast Fourier Transform" struct FFT <: Alg
+        plan1::Union{Function,Nothing}
+        plan2::Union{Function,Nothing}
+        plan3::Union{Function,Nothing}
+    end
+    FFT() = FFT(nothing, nothing, nothing)
     "Filter using a direct algorithm" struct FIR <: Alg end
     "Cache-efficient filtering using tiles" struct FIRTiled{N} <: Alg
         tilesize::Dims{N}

src/imfilter.jl

@@ -826,7 +826,7 @@
     for I in CartesianIndices(axes(kern))
         krn[I] = kern[I]
     end
-    Af = filtfft(A, krn)
+    Af = filtfft(A, krn, r.settings.plan1, r.settings.plan2, r.settings.plan3)
     if map(first, axes(out)) == map(first, axes(Af))
         R = CartesianIndices(axes(out))
         copyto!(out, R, Af, R)
@@ -837,13 +837,61 @@
         src = view(FFTView(Af), axes(dest)...)
         copyto!(dest, src)
     end
-    out
+    return out
+end
+
+function buffered_planned_rfft(a::AbstractArray{T}) where {T}
+    buf = RFFT.RCpair{T}(undef, size(a))
+    plan = RFFT.plan_rfft!(buf; flags=FFTW.MEASURE)
+    return function (arr::AbstractArray{T}) where {T}
+        copy!(buf, OffsetArrays.no_offset_view(arr))
+        return plan(buf)
+    end
+end
+function buffered_planned_irfft(a::AbstractArray{T}) where {T}
+    buf = RFFT.RCpair{T}(undef, size(a))
+    plan = RFFT.plan_irfft!(buf; flags=FFTW.MEASURE)
+    return function (arr::AbstractArray{T}) where {T}
+        copy!(buf, OffsetArrays.no_offset_view(arr))
+        return plan(buf)
+    end
 end
 
+function planned_fft(A::AbstractArray{T,N},
+            kernel::ProcessedKernel,
+            border::BorderSpecAny=Pad(:replicate)
+            ) where {T<:AbstractFloat,N}
+    bord = border(kernel, A, Algorithm.FFT())
+    _A = padarray(T, A, bord)
+    bfp1 = buffered_planned_rfft(_A)
+    kern = samedims(_A, kernelconv(kernel...))
+    krn = FFTView(zeros(eltype(kern), map(length, axes(_A))))
+    bfp2 = buffered_planned_rfft(krn)
+    bfp3 = buffered_planned_irfft(_A)
+    return Algorithm.FFT(bfp1, bfp2, bfp3)
+end
+planned_fft(A::AbstractArray, kernel, border::AbstractString) = planned_fft(A, kernel, borderinstance(border))
+planned_fft(A::AbstractArray, kernel::Union{ArrayLike,Laplacian}, border::BorderSpecAny) = planned_fft(A, factorkernel(kernel), border)
+
+function filtfft(A, krn, planned_rfft1::Function, planned_rfft2::Function, planned_irfft::Function)
+    B = complex(planned_rfft1(A))
+    B .*= conj!(complex(planned_rfft2(krn)))
+    return real(planned_irfft(complex(B)))
+end
+# TODO: this does not work. See TODO below
+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
+    B .*= conj!(complex(planned_rfft2(kernrs)))
+    Avf = real(planned_irfft(complex(B)))
+    return colorview(base_colorant_type(C){eltype(Avf)}, Avf)
+end
+filtfft(A, krn, ::Nothing, ::Nothing, ::Nothing) = filtfft(A, krn)
 function filtfft(A, krn)
     B = rfft(A)
     B .*= conj!(rfft(krn))
-    irfft(B, length(axes(A, 1)))
+    return irfft(B, length(axes(A, 1)))
 end
 function filtfft(A::AbstractArray{C}, krn) where {C<:Colorant}
     Av, dims = channelview_dims(A)

test/2d.jl

@@ -36,6 +36,15 @@ using ImageFiltering: borderinstance
     end
 end
 
+function supported_algs(img, kernel, border)
+    if eltype(img) isa AbstractFloat
+        (Algorithm.FIR(), Algorithm.FIRTiled(), Algorithm.FFT(), planned_fft(img, kernel, border))
+    else
+        # TODO: extend planned_fft to support other types
+        (Algorithm.FIR(), Algorithm.FIRTiled(), Algorithm.FFT())
+    end
+end
+
 @testset "FIR/FFT" begin
     f32type(img) = f32type(eltype(img))
     f32type(::Type{C}) where {C<:Colorant} = base_colorant_type(C){Float32}
@@ -50,6 +59,7 @@ end
     # Dense inseparable kernel
     kern = [0.1 0.2; 0.4 0.5]
     kernel = OffsetArray(kern, -1:0, 1:2)
+    border = Inner()
     for img in (imgf, imgi, imgg, imgc)
         targetimg = zeros(typeof(img[1]*kern[1]), size(img))
         targetimg[3:4,2:3] = rot180(kern) .* img[3,4]
@@ -66,7 +76,7 @@ end
             @test @inferred(imfilter(f32type(img), img, kernel, border)) ≈ float32.(targetimg)
             fill!(ret, zero(eltype(ret)))
             @test @inferred(imfilter!(ret, img, kernel, border)) ≈ targetimg
-            for alg in (Algorithm.FIR(), Algorithm.FIRTiled(), Algorithm.FFT())
+            for alg in supported_algs(img, kernel, border)
                 @test @inferred(imfilter(img, kernel, border, alg)) ≈ targetimg
                 @test @inferred(imfilter(img, (kernel,), border, alg)) ≈ targetimg
                 @test @inferred(imfilter(f32type(img), img, kernel, border, alg)) ≈ float32.(targetimg)
@@ -76,12 +86,12 @@ end
             @test_throws MethodError imfilter!(CPU1(Algorithm.FIR()), ret, img, kernel, border, Algorithm.FFT())
         end
         targetimg_inner = OffsetArray(targetimg[2:end, 1:end-2], 2:5, 1:5)
-        @test @inferred(imfilter(img, kernel, Inner())) ≈ targetimg_inner
-        @test @inferred(imfilter(f32type(img), img, kernel, Inner())) ≈ float32.(targetimg_inner)
-        for alg in (Algorithm.FIR(), Algorithm.FIRTiled(), Algorithm.FFT())
-            @test @inferred(imfilter(img, kernel, Inner(), alg)) ≈ targetimg_inner
-            @test @inferred(imfilter(f32type(img), img, kernel, Inner(), alg)) ≈ float32.(targetimg_inner)
-            @test @inferred(imfilter(CPU1(alg), img, kernel, Inner())) ≈ targetimg_inner
+        @test @inferred(imfilter(img, kernel, border)) ≈ targetimg_inner
+        @test @inferred(imfilter(f32type(img), img, kernel, border)) ≈ float32.(targetimg_inner)
+        for alg in supported_algs(img, kernel, border)
+            @test @inferred(imfilter(img, kernel, border, alg)) ≈ targetimg_inner
+            @test @inferred(imfilter(f32type(img), img, kernel, border, alg)) ≈ float32.(targetimg_inner)
+            @test @inferred(imfilter(CPU1(alg), img, kernel, border)) ≈ targetimg_inner
         end
     end
     # Factored kernel
@@ -96,7 +106,7 @@ end
         for border in ("replicate", "circular", "symmetric", "reflect", Fill(zero(eltype(img))))
             @test @inferred(imfilter(img, kernel, border)) ≈ targetimg
             @test @inferred(imfilter(f32type(img), img, kernel, border)) ≈ float32.(targetimg)
-            for alg in (Algorithm.FIR(), Algorithm.FIRTiled(), Algorithm.FFT())
+            for alg in supported_algs(img, kernel, border)
                 @test @inferred(imfilter(img, kernel, border, alg)) ≈ targetimg
                 @test @inferred(imfilter(f32type(img), img, kernel, border, alg)) ≈ float32.(targetimg)
             end
@@ -106,7 +116,7 @@ end
         targetimg_inner = OffsetArray(targetimg[2:end, 1:end-2], 2:5, 1:5)
         @test @inferred(imfilter(img, kernel, Inner())) ≈ targetimg_inner
         @test @inferred(imfilter(f32type(img), img, kernel, Inner())) ≈ float32.(targetimg_inner)
-        for alg in (Algorithm.FIR(), Algorithm.FIRTiled(), Algorithm.FFT())
+        for alg in supported_algs(img, kernel, border)
             @test @inferred(imfilter(img, kernel, Inner(), alg)) ≈ targetimg_inner
             @test @inferred(imfilter(f32type(img), img, kernel, Inner(), alg)) ≈ float32.(targetimg_inner)
         end
@@ -122,7 +132,7 @@ end
         for border in ("replicate", "circular", "symmetric", "reflect", Fill(zero(eltype(img))))
             @test @inferred(imfilter(img, kernel, border)) ≈ targetimg
             @test @inferred(imfilter(f32type(img), img, kernel, border)) ≈ float32.(targetimg)
-            for alg in (Algorithm.FIR(), Algorithm.FIRTiled(), Algorithm.FFT())
+            for alg in supported_algs(img, kernel, border)
                 if alg == Algorithm.FFT() && eltype(img) == Int
                     @test @inferred(imfilter(Float64, img, kernel, border, alg)) ≈ targetimg
                 else
@@ -134,7 +144,7 @@ end
         targetimg_inner = OffsetArray(targetimg[2:end-1, 2:end-1], 2:4, 2:6)
         @test @inferred(imfilter(img, kernel, Inner())) ≈ targetimg_inner
         @test @inferred(imfilter(f32type(img), img, kernel, Inner())) ≈ float32.(targetimg_inner)
-        for alg in (Algorithm.FIR(), Algorithm.FIRTiled(), Algorithm.FFT())
+        for alg in supported_algs(img, kernel, border)
             if alg == Algorithm.FFT() && eltype(img) == Int
                 @test @inferred(imfilter(Float64, img, kernel, Inner(), alg)) ≈ targetimg_inner
             else
@@ -184,7 +194,7 @@ end
             targetimg = target1(img, kern, border)
             @test @inferred(imfilter(img, kernel, border)) ≈ targetimg
             @test @inferred(imfilter(f32type(img), img, kernel, border)) ≈ float32.(targetimg)
-            for alg in (Algorithm.FIR(), Algorithm.FIRTiled(), Algorithm.FFT())
+            for alg in supported_algs(img, kernel, border)
                 @test @inferred(imfilter(img, kernel, border, alg)) ≈ targetimg
                 @test @inferred(imfilter(f32type(img), img, kernel, border, alg)) ≈ float32.(targetimg)
             end
@@ -195,7 +205,7 @@ end
         targetimg = zerona!(copy(targetimg0))
         @test @inferred(zerona!(imfilter(img, kernel, border))) ≈ targetimg
         @test @inferred(zerona!(imfilter(f32type(img), img, kernel, border))) ≈ float32.(targetimg)
-        for alg in (Algorithm.FIR(), Algorithm.FIRTiled(), Algorithm.FFT())
+        for alg in supported_algs(img, kernel, border)
             @test @inferred(zerona!(imfilter(img, kernel, border, alg), nanflag)) ≈ targetimg
             @test @inferred(zerona!(imfilter(f32type(img), img, kernel, border, alg), nanflag)) ≈ float32.(targetimg)
         end
@@ -208,7 +218,7 @@ end
             targetimg = target1(img, kern, border)
             @test @inferred(imfilter(img, kernel, border)) ≈ targetimg
             @test @inferred(imfilter(f32type(img), img, kernel, border)) ≈ float32.(targetimg)
-            for alg in (Algorithm.FIR(), Algorithm.FIRTiled(), Algorithm.FFT())
+            for alg in supported_algs(img, kernel, border)
                 @test @inferred(imfilter(img, kernel, border, alg)) ≈ targetimg
                 @test @inferred(imfilter(f32type(img), img, kernel, border, alg)) ≈ float32.(targetimg)
             end
@@ -219,7 +229,7 @@ end
         targetimg = zerona!(copy(targetimg0))
         @test @inferred(zerona!(imfilter(img, kernel, border))) ≈ targetimg
         @test @inferred(zerona!(imfilter(f32type(img), img, kernel, border))) ≈ float32.(targetimg)
-        for alg in (Algorithm.FIR(), Algorithm.FIRTiled(), Algorithm.FFT())
+        for alg in supported_algs(img, kernel, border)
             @test @inferred(zerona!(imfilter(img, kernel, border, alg), nanflag)) ≈ targetimg
             @test @inferred(zerona!(imfilter(f32type(img), img, kernel, border, alg), nanflag)) ≈ float32.(targetimg)
         end

test/gabor.jl

@@ -7,11 +7,15 @@ using ImageFiltering, Test, Statistics
     size_y = 6*σy+1
     γ = σx/σy
     # zero size forces default kernel width, with warnings
-    @info "Four warnings are expected"
-    kernel = Kernel.gabor(0,0,σx,0,5,γ,0)
-    @test isequal(size(kernel[1]),(size_x,size_y))
-    kernel = Kernel.gabor(0,0,σx,π,5,γ,0)
-    @test isequal(size(kernel[1]),(size_x,size_y))
+
+    @test_logs (:warn, r"The input parameter size_") match_mode=:any begin
+        kernel = Kernel.gabor(0,0,σx,0,5,γ,0)
+        @test isequal(size(kernel[1]),(size_x,size_y))
+    end
+    @test_logs (:warn, r"The input parameter size_") match_mode=:any begin
+        kernel = Kernel.gabor(0,0,σx,π,5,γ,0)
+        @test isequal(size(kernel[1]),(size_x,size_y))
+    end
 
     for x in 0:4, y in 0:4, z in 0:4, t in 0:4
         σx1 = 2*x+1