Better handling of function interpolation, lower bound ArrayInterface to a version testing that size isn't broken on vectors.

Author: chriselrod

Project.toml

@@ -14,7 +14,7 @@ UnPack = "3a884ed6-31ef-47d7-9d2a-63182c4928ed"
 VectorizationBase = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
 
 [compat]
-ArrayInterface = "2.14.2"
+ArrayInterface = "2.14.9"
 DocStringExtensions = "0.8"
 IfElse = "0.1"
 OffsetArrays = "1.4.1"

src/costs.jl

@@ -463,11 +463,11 @@ const FUNCTIONSYMBOLS = IdDict{Type{<:Function},Instruction}(
     typeof(Base.ifelse) => :ifelse,
     typeof(ifelse) => :ifelse,
     typeof(identity) => :identity,
-    typeof(conj) => :conj,
-    typeof(zero) => :zero,
-    typeof(one) => :one,
-    typeof(axes) => :axes,
-    typeof(eltype) => :eltype
+    typeof(conj) => :conj
+    # typeof(zero) => :zero,
+    # typeof(one) => :one,
+    # typeof(axes) => :axes,
+    # typeof(eltype) => :eltype
 )
 
 # implement whitelist for avx_support that package authors may use to conservatively guard `@avx` application

src/graphs.jl

@@ -573,13 +573,18 @@ function instruction!(ls::LoopSet, x::Expr)
     if instr ∉ keys(COST)
         instr = gensym(:f)
         pushpreamble!(ls, Expr(:(=), instr, x))
+        Instruction(Symbol(""), instr)
+    else
+        Instruction(:LoopVectorization, instr)
     end
-    Instruction(Symbol(""), instr)
 end
 instruction!(ls::LoopSet, x::Symbol) = instruction(x)
-function instruction!(ls::LoopSet, ::F) where {F <: Function}
-    FUNCTIONSYMBOLS[F]
-    # get(FUNCTIONSYMBOLS, F, 
+function instruction!(ls::LoopSet, f::F) where {F <: Function}
+    get(FUNCTIONSYMBOLS, F) do
+        instr = gensym(:f)
+        pushpreamble!(ls, Expr(:(=), instr, f))
+        Instruction(Symbol(""), instr)
+    end
 end
 
 

src/lowering.jl

@@ -652,13 +652,13 @@ end
 
 function determine_eltype(ls::LoopSet)
     if length(ls.includedactualarrays) == 0
-        return Expr(:call, :typeof, 0)
+        return Expr(:call, lv(:typeof), 0)
     elseif length(ls.includedactualarrays) == 1
-        return Expr(:call, :eltype, first(ls.includedactualarrays))
+        return Expr(:call, lv(:eltype), first(ls.includedactualarrays))
     end
-    promote_q = Expr(:call, :promote_type)
+    promote_q = Expr(:call, lv(:promote_type))
     for array ∈ ls.includedactualarrays
-        push!(promote_q.args, Expr(:call, :eltype, array))
+        push!(promote_q.args, Expr(:call, lv(:eltype), array))
     end
     promote_q
 end

src/reconstruct_loopset.jl

@@ -5,7 +5,7 @@ function Loop(ls::LoopSet, ex::Expr, sym::Symbol, ::Type{<:AbstractUnitRange})
     start = gensym(ssym*"_loopstart"); stop = gensym(ssym*"_loopstop"); loopsym = gensym(ssym * "_loop")
     pushpreamble!(ls, Expr(:(=), loopsym, ex))
     pushpreamble!(ls, Expr(:(=), start, Expr(:call, :first, loopsym)))
-    pushpreamble!(ls, Expr(:(=), stop, Expr(:call, :last, loopsym)))
+    pushpreamble!(ls, Expr(:(=), stop, Expr(:call, lv(:last), loopsym)))
     loop = Loop(sym, 1, 1024, start, stop, false, false)::Loop
     pushpreamble!(ls, loopiteratesatleastonce(loop))
     loop
@@ -14,14 +14,14 @@ end
 
 function Loop(ls::LoopSet, ex::Expr, sym::Symbol, ::Type{OptionallyStaticUnitRange{I, Static{U}}}) where {I<:Integer, U}
     start = gensym(String(sym)*"_loopstart")
-    pushpreamble!(ls, Expr(:(=), start, Expr(:call, :first, ex)))
+    pushpreamble!(ls, Expr(:(=), start, Expr(:call, lv(:first), ex)))
     loop = Loop(sym, U - 1024, U, start, Symbol(""), false, true)::Loop
     pushpreamble!(ls, loopiteratesatleastonce(loop))
     loop
 end
 function Loop(ls::LoopSet, ex::Expr, sym::Symbol, ::Type{OptionallyStaticUnitRange{Static{L}, I}}) where {I <: Integer, L}
     stop = gensym(String(sym)*"_loopstop")
-    pushpreamble!(ls, Expr(:(=), stop, Expr(:call, :last, ex)))
+    pushpreamble!(ls, Expr(:(=), stop, Expr(:call, lv(:last), ex)))
     loop = Loop(sym, L, L + 1024, Symbol(""), stop, true, false)::Loop
     pushpreamble!(ls, loopiteratesatleastonce(loop))
     loop

test/miscellaneous.jl

@@ -787,6 +787,31 @@ function maybe_const_issue144_avx!(𝛥mat, 𝛥ℛ, mat, ℛ)
     end
     𝛥mat
 end
+    function grad!(𝛥x, 𝛥ℛ, x, 𝒶𝓍i=eachindex(x))
+        for i = 𝒶𝓍i
+            (i >= first(axes(𝛥x, 1))) & (i <= last(axes(𝛥x, 1))) && (𝛥x[i] = 𝛥x[i] + 𝛥ℛ[i])
+        end
+        𝛥x
+    end
+    function grad_avx!(𝛥x, 𝛥ℛ, x, 𝒶𝓍i=eachindex(x))
+        @avx for i = 𝒶𝓍i
+            (i >= first(axes(𝛥x, 1))) & (i <= last(axes(𝛥x, 1))) && (𝛥x[i] = 𝛥x[i] + 𝛥ℛ[i])
+        end
+        𝛥x
+    end
+    function grad_avx_base!(𝛥x, 𝛥ℛ, x, 𝒶𝓍i=eachindex(x))
+        @avx for i = 𝒶𝓍i
+            (i >= first(axes(𝛥x, 1))) & (i <= Base.last(axes(𝛥x, 1))) && (𝛥x[i] = 𝛥x[i] + 𝛥ℛ[i])
+        end
+        𝛥x
+    end
+    @eval function grad_avx_eval!(𝛥x, 𝛥ℛ, x, 𝒶𝓍i=eachindex(x))
+        @avx for i = 𝒶𝓍i
+            (i >= $first($axes(𝛥x, 1))) & (i <= $last($axes(𝛥x, 1))) && (𝛥x[i] = 𝛥x[i] + 𝛥ℛ[i])
+        end
+        𝛥x
+    end # LoadError: KeyError: key typeof(first) not found
+
 
     for T ∈ (Float32, Float64)
         @show T, @__LINE__
@@ -1015,6 +1040,7 @@ end
             rtol = ∛(eps(T))
         )
 
+        @test grad!(zeros(5), ones(5), ones(3)) ≈ grad_avx!(zeros(5), ones(5), ones(3)) ≈ grad_avx_base!(zeros(5), ones(5), ones(3)) ≈ grad_avx_eval!(zeros(5), ones(5), ones(3))
     end
     for T ∈ [Int16, Int32, Int64]
         n = 8sizeof(T) - 1