Merge pull request #53 from mcabbott/upci

Add tests for LoopVectorization

Author: mcabbott

.github/workflows/ci.yml

@@ -16,7 +16,7 @@ jobs:
       fail-fast: false
       matrix:
         version:
-          - '1.3' # Replace this with the minimum Julia version that your package supports. E.g. if your package requires Julia 1.5 or higher, change this to '1.5'.
+          - '1.4' # Replace this with the minimum Julia version that your package supports. E.g. if your package requires Julia 1.5 or higher, change this to '1.5'.
           - '1' # Leave this line unchanged. '1' will automatically expand to the latest stable 1.x release of Julia.
           - 'nightly'
         os:

.travis.yml

@@ -3,8 +3,8 @@ language: julia
 os:
   - linux
 julia:
-  - 1.3
-  - 1.5
+  - 1.4
+  - 1
   - nightly
 env:
   - JULIA_NUM_THREADS=1

Project.toml

@@ -1,18 +1,16 @@
 name = "Tullio"
 uuid = "bc48ee85-29a4-5162-ae0b-a64e1601d4bc"
 authors = ["Michael Abbott"]
-version = "0.2.10"
+version = "0.2.11"
 
 [deps]
 DiffRules = "b552c78f-8df3-52c6-915a-8e097449b14b"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 Requires = "ae029012-a4dd-5104-9daa-d747884805df"
-SpecialFunctions = "276daf66-3868-5448-9aa4-cd146d93841b"
 
 [compat]
 DiffRules = "1"
 Requires = "1"
-SpecialFunctions = "0.10" # just for this bound
 julia = "1.3"
 
 [extras]
@@ -21,6 +19,7 @@ FillArrays = "1a297f60-69ca-5386-bcde-b61e274b549b"
 ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210"
 KernelAbstractions = "63c18a36-062a-441e-b654-da1e3ab1ce7c"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
+LoopVectorization = "bdcacae8-1622-11e9-2a5c-532679323890"
 NamedDims = "356022a1-0364-5f58-8944-0da4b18d706f"
 OffsetArrays = "6fe1bfb0-de20-5000-8ca7-80f57d26f881"
 Printf = "de0858da-6303-5e67-8744-51eddeeeb8d7"
@@ -31,4 +30,4 @@ Tracker = "9f7883ad-71c0-57eb-9f7f-b5c9e6d3789c"
 Zygote = "e88e6eb3-aa80-5325-afca-941959d7151f"
 
 [targets]
-test = ["Test", "CUDA", "FillArrays", "ForwardDiff", "KernelAbstractions", "LinearAlgebra", "NamedDims", "OffsetArrays", "Printf", "Random", "TensorOperations", "Tracker", "Zygote"]
+test = ["Test", "CUDA", "FillArrays", "ForwardDiff", "KernelAbstractions", "LinearAlgebra", "LoopVectorization", "NamedDims", "OffsetArrays", "Printf", "Random", "TensorOperations", "Tracker", "Zygote"]

src/eval.jl

@@ -53,12 +53,11 @@ using Requires
 
 @init @require LoopVectorization = "bdcacae8-1622-11e9-2a5c-532679323890" begin
     using .LoopVectorization
-    SVec = if isdefined(LoopVectorization, :SVec) # version 0.8, for Julia ⩽1.5
+    if isdefined(LoopVectorization, :SVec) # version 0.8, for Julia ⩽1.5
         using .LoopVectorization.VectorizationBase: SVec, Mask, prevpow2
-        Svec
     else # version 0.9, supports Julia 1.6
         using .LoopVectorization.VectorizationBase: Vec, Mask, prevpow2
-        Vec
+        SVec{N,T} = Vec{N,T}
     end
 
     # Functions needed for safe vectorised max gradient
@@ -71,7 +70,8 @@ using Requires
     @inline allzero(seen::Int) = iszero(seen)
     @inline allzero(seen::SVec{N,Int}) where {N} = iszero((!iszero(seen)).u)
 
-    @inline Tullio.anyone(cond::Mask) = cond != zero(cond)
+    # @inline Tullio.anyone(cond::Mask) = cond != zero(cond)
+    @inline Tullio.anyone(cond::Mask) = cond.u != zero(cond).u # for v0.9
 
     @require ForwardDiff = "f6369f11-7733-5829-9624-2563aa707210" begin
         # Dual numbers + svec, should live in PaddedMatricesForwardDiff?

src/macro.jl

@@ -1070,7 +1070,9 @@ function make_many_actors(act!, args, ex1, outer::Vector, ex3, inner::Vector, ex
     safe = if act! == ACT!
         isempty(store.unsafeleft)
     else # working on ∇act!
-        isempty(store.unsaferight)
+        isempty(store.unsaferight) &&
+            store.redfun == :+ && # Disable @avx for min/max grad, #53
+            store.grad != :Dual   # and for use with ForwardDiff
     end
 
     if safe && store.avx != false && isdefined(store.mod, :LoopVectorization)

test/gradients.jl

@@ -259,24 +259,24 @@ if Tullio._GRAD[] != :Dual
         dv = ForwardDiff.gradient(v -> sum(f6(m2,v)), v2)
         @test dv ≈ _gradient(sum∘f6, m2, v2)[2]
 
-        f7(x,y) = @tullio (max) z[i] := x[i,j]^2 / sqrt(y[i]) + exp(y[j])
+        f7(x,y) = @tullio (max) z[i] := x[i,j]^2 / sqrt(y[i]) + exp(y[j]) avx=false
 
         dm = ForwardDiff.gradient(m -> sum(f7(m,v2)), m2)
-        @test dm ≈_gradient(sum∘f7, m2, v2)[1]
+        @test dm ≈_gradient(sum∘f7, m2, v2)[1]  # gives wrong answers with avx, 1.4 in tests
         dv = ForwardDiff.gradient(v -> sum(f7(m2,v)), v2)
         @test dv ≈ _gradient(sum∘f7, m2, v2)[2]
 
-        f8(x,y) = @tullio (max) z[i,l] := log(x[i,j,k,l]) / y[j]^1/3
-        f9(x,y) = @tullio (min) z[i,j] := log(x[i,j,k,l]) / y[j]^1/3
+        f8(x,y) = @tullio (max) z[i,l] := log(x[i,j,k,l]) / y[j]^1/3 avx=false
+        f9(x,y) = @tullio (min) z[i,j] := log(x[i,j,k,l]) / y[j]^1/3 avx=false
 
         dm = ForwardDiff.gradient(m -> sum(f8(m,v2)), m4)
-        @test dm ≈_gradient(sum∘f8, m4, v2)[1]
+        @test dm ≈ _gradient(sum∘f8, m4, v2)[1]  # gives wrong answers with avx, 1.5 in tests
         dv = ForwardDiff.gradient(v -> sum(f8(m4,v)), v2)
         @test dv ≈ _gradient(sum∘f8, m4, v2)[2]
         dm = ForwardDiff.gradient(m -> sum(f9(m,v2)), m4)
-        @test dm ≈_gradient(sum∘f9, m4, v2)[1]
+        @test dm ≈_gradient(sum∘f9, m4, v2)[1]  # gives wrong answers with avx, repl
         dv = ForwardDiff.gradient(v -> sum(f9(m4,v)), v2)
-        @test dv ≈ _gradient(sum∘f9, m4, v2)[2]
+        @test dv ≈ _gradient(sum∘f9, m4, v2)[2]  # gives wrong answers with avx
 
     end
 

test/parsing.jl

@@ -11,7 +11,7 @@ using Tullio, Test, LinearAlgebra
     @test A == [i^2 for i in 1:10]
 
     # diagonals
-    @tullio D[i,i] := trunc(Int, sqrt(A[i]))
+    @tullio D[i,i] := trunc(Int, sqrt(A[i])) avx=false # MethodError: no method matching trunc(::Type{Int64}, ::VectorizationBase.Vec{4,Float64})
     @test D == Diagonal(sqrt.(A))
 
     # arrays of arrays
@@ -124,7 +124,7 @@ using Tullio, Test, LinearAlgebra
     @test H[1,:] == M[2,:] # but H[3,:] gets written into twice.
 
     J′ = [1,2,10]
-    @tullio H′[J′[i'],k] := A[k]
+    @tullio H′[J′[i'],k] := A[k] avx=false # StackOverflowError
     @test size(H′) == (10, length(A))
     @test H′[2,:] == A
     @test H′[3,4] == 0 # zeroed before being written into
@@ -219,7 +219,7 @@ end
     # scatter operation
     D = similar(A, 10, 10) .= 999
     inds = [2,3,5,2]
-    @tullio D[inds[i],j] = A[j]
+    @tullio D[inds[i],j] = A[j] avx=false # StackOverflowError
     @test D[2,:] == A
     @test D[4,4] != 0 # not zeroed before writing.
 
@@ -444,7 +444,7 @@ end
     # basics
     @test [prod(A)] == @tullio (*) P[_] := float(A[i])
     @test maximum(A) == @tullio (max) m := float(A[i])
-    @test minimum(A) == @tullio (min) m := float(A[i]) # fails with @avx
+    @test minimum(A) == @tullio (min) m := float(A[i])
 
     @test true == @tullio (&) p := A[i] > 0
     @test true === @tullio (&) p := A[i] > 0
@@ -505,11 +505,11 @@ end
 
     # promotion of init & += cases:
     B = rand(10)
-    @test sum(B.^2)+2 ≈ @tullio s2 := B[i]^2 init=2 threads=false
+    @test sum(B.^2)+2 ≈ @tullio s2 := B[i]^2 init=2 threads=false avx=false # InexactError: Int64 on LV 0.8
     s3 = 3
     @test sum(B.^2)+3 ≈ @tullio s3 += B[i]^2
     s4 = 4im
-    @test sum(B.^2)+4im ≈ @tullio s4 += B[i]^2
+    @test sum(B.^2)+4im ≈ @tullio s4 += B[i]^2 avx=false # TypeError: in AbstractSIMD, in T, expected T<:(Union{Bool, Float32
 
     # no reduction means no redfun, and no init:
     @test_throws LoadError @eval @tullio (max) A2[i] := A[i]^2
@@ -566,18 +566,18 @@ end
     # reading
     N = NamedDimsArray(rand(Int8,3,10), (:r, :c))
 
-    @tullio A[i,j] := N[i, j] + 100 * (1:10)[j]
+    @tullio A[i,j] := N[i, j] + 100 * (1:10)[j] avx=false # conversion to pointer not defined for NamedDimsArray
     @test A == N .+ 100 .* (1:10)'
 
-    @tullio B[i] := N[r=i, c=1]
+    @tullio B[i] := N[r=i, c=1] avx=false
     @test B == N[:,1]
 
-    @tullio C[j,i] := N[c=j, r=i] + 100 * (1:10)[j]
+    @tullio C[j,i] := N[c=j, r=i] + 100 * (1:10)[j] avx=false
     @test A == C'
     @test dimnames(C) == (:_, :_) # similar(parent(A)) avoids a bug
 
     # writing
-    @tullio M[row=i, col=j, i=1] := (1:3)[i] // (1:7)[j]
+    @tullio M[row=i, col=j, i=1] := (1:3)[i] // (1:7)[j] avx=false
     @test dimnames(M) == (:row, :col, :i)
 
 end
@@ -593,7 +593,7 @@ end
     @test_throws LoadError @macroexpand1 @tullio A[i] := (1:10)[i]^2  threads=:maybe
 
     # keyword verbose accepts values [true, false, 2, 3]
-    @tullio A[i] := (1:10)[i]^2  verbose=1
+    @tullio A[i] := (1:10)[i]^2  verbose=1 avx=false # @error: rejected by LoopVectorization's check_args
     @tullio A[i] := (1:10)[i]^2  verbose=false
     @test_throws LoadError @macroexpand1 @tullio A[i] := (1:10)[i]^2  verbose=4
 
@@ -632,7 +632,7 @@ end
     # https://github.com/mcabbott/Tullio.jl/issues/36
     # final type real, intermediate complex... not fixed yet!
     xs = randn(1000)
-    @test_throws InexactError @tullio z[i] := exp(im * xs[i] - xs[j]) |> abs2
+    @test_throws InexactError @tullio z[i] := exp(im * xs[i] - xs[j]) |> abs2  avx=false # TypeError with LV
 
     # https://github.com/mcabbott/Tullio.jl/issues/43
     P = rand(2,2,3); Diff = rand(3,3); n=4
@@ -664,7 +664,7 @@ end
        @tullio x[i] := s[i]
     end
     @test_broken [3,7] == let s = [1 2;3 4], zero=one # left because of #50
-       @tullio x[i] := s[i,j]
+       @tullio x[i] := s[i,j]  avx=false # Unexpected Pass with LV
     end
 
 end

test/runtests.jl

@@ -96,6 +96,8 @@ end
 
 #===== Zygote =====#
 
+if VERSION < v"1.6-" # Zygote isn't working on 1.6
+
 t5 = time()
 using Zygote
 
@@ -109,7 +111,6 @@ _gradient(x...) = Zygote.gradient(x...)
 @testset "gradients: Zygote + ForwardDiff" begin include("gradients.jl") end
 
 @tullio grad=Base
-if VERSION >= v"1.4" # mysterious failures on 1.3
 @testset "complex gradients with Zygote" begin
 
     x0 = [1,2,3] .+ [5im, 0, -11im]
@@ -161,10 +162,11 @@ if VERSION >= v"1.4" # mysterious failures on 1.3
 
     end
 end
-end # VERSION
 
 @info @sprintf("Zygote tests took %.1f seconds", time()-t5)
 
+end # VERSION
+
 #===== ReverseDiff =====#
 #=
 t6 = time()
@@ -200,17 +202,43 @@ _gradient(x...) = Yota.grad(x...)[2]
 =#
 
 #===== LoopVectorization =====#
-#=
+
 t8 = time()
 using LoopVectorization
 
-using LoopVectorization.VectorizationBase: SVec, Mask, prevpow2
-sv = SVec{4,Int}(1,2,3,4) # SVec{4,Int64}<1, 2, 3, 4>
-ms = Mask(0x03) # Mask{8,Bool}<1, 1, 0, 0, 0, 0, 0, 0>
-@test Tullio.onlyone(ms, 0) == Mask(0x02)
-@test Tullio.onlyone(ms, sv) == Mask(0x00)
-@test Tullio.onlyone(ms, zero(sv)) == Mask(0x02)
+if isdefined(LoopVectorization, :SVec) # version 0.8, for Julia ⩽1.5
+    using LoopVectorization.VectorizationBase: SVec, Mask
+else # version 0.9, supports Julia 1.6
+    using LoopVectorization.VectorizationBase: Vec, Mask
+    SVec{N,T} = Vec{N,T}
+end
+
+@testset "LoopVectorization onlyone" begin
+    ms = Mask{UInt8}(0x03); # Mask{8,Bool}<1, 1, 0, 0, 0, 0, 0, 0>
+    sv = SVec{4,Int}(1,2,3,4) # SVec{4,Int64}<1, 2, 3, 4>
+
+    # preliminaries:
+    @test Tullio.allzero(sv) === false
+    @test Tullio.allzero(zero(sv)) === true
+
+    @test Tullio.anyone(ms) === true
+
+    # the main function:
+    @test Tullio.onlyone(false, 0) === false
+    @test Tullio.onlyone(true, 0) === true
+    @test Tullio.onlyone(true, 1) === false
 
+    # @test Tullio.onlyone(ms, 0) === Mask{UInt8}(0x02)
+    @test Tullio.onlyone(ms, 0).u == 0x02
+    # @test Tullio.onlyone(ms, sv) === Mask{UInt8}(0x00)
+    @test Tullio.onlyone(ms, sv).u == 0x00
+    # @test Tullio.onlyone(ms, zero(sv)) === Mask{UInt8}(0x02)
+    @test Tullio.onlyone(ms, zero(sv)).u == 0x02
+end
+
+@testset "parsing + LoopVectorization" begin include("parsing.jl") end
+
+using Tracker
 GRAD = :Tracker
 _gradient(x...) = Tracker.gradient(x...)
 
@@ -220,16 +248,10 @@ _gradient(x...) = Tracker.gradient(x...)
 @tullio grad=Dual
 @testset "gradients: Tracker + ForwardDiff + LoopVectorization" begin include("gradients.jl") end
 
-GRAD = :Zygote
-_gradient(x...) = Zygote.gradient(x...)
+@info @sprintf("LoopVectorization tests took %.1f seconds", time()-t8)
 
-@tullio grad=Base
-@testset "gradients: Zygote + LoopVectorization" begin include("gradients.jl") end
+@tullio avx=false
 
-@testset "parsing + LoopVectorization" begin include("parsing.jl") end
-
-@info @sprintf("LoopVectorization tests took %.1f seconds", time()-t8)
-=#
 #===== TensorOperations =====#
 
 t9 = time()
@@ -242,6 +264,8 @@ _gradient(x...) = Tracker.gradient(x...)
 @tullio grad=Base
 @testset "gradients: Tracker + TensorOperations" begin include("gradients.jl") end
 
+if VERSION < v"1.6-" # Zygote isn't working on 1.6
+
 using Zygote
 GRAD = :Zygote
 _gradient(x...) = Zygote.gradient(x...)
@@ -271,6 +295,8 @@ _gradient(x...) = Zygote.gradient(x...)
     end
 end
 
+end # VERSION
+
 @testset "parsing + TensorOperations" begin include("parsing.jl") end # testing correct fallback
 
 @info @sprintf("TensorOperations tests took %.1f seconds", time()-t9)