Added tests, removed some dead code.

Author: chriselrod

.appveyor.yml

@@ -4,7 +4,7 @@ environment:
   - julia_version: 1.4
   - julia_version: nightly
 platform:
-#  - x86
+  - x86
   - x64
 matrix:
   allow_failures:

Project.toml

@@ -15,10 +15,10 @@ VectorizationBase = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
 [compat]
 DocStringExtensions = "0.8"
 OffsetArrays = "1"
-SIMDPirates = "0.7.20"
-SLEEFPirates = "0.4.4"
+SIMDPirates = "0.7.23"
+SLEEFPirates = "0.4.8"
 UnPack = "0"
-VectorizationBase = "0.11.2"
+VectorizationBase = "0.11.3"
 julia = "1.1"
 
 [extras]

benchmark/driver.jl

@@ -21,12 +21,12 @@ end
 
 # sizes = 23:23
 sizes = 256:-1:2
-selfdot_bench = benchmark_selfdot(sizes)
 
 AmulB_bench = benchmark_AmulB(sizes)
 AmulBt_bench = benchmark_AmulBt(sizes)
-AtmulB_bench = benchmark_AtmulB(sizes)
 AtmulBt_bench = benchmark_AtmulBt(sizes)
+AtmulB_bench = benchmark_AtmulB(sizes)
+
 Amulvb_bench = benchmark_Amulvb(sizes)
 Atmulvb_bench = benchmark_Atmulvb(sizes)
 
@@ -36,14 +36,15 @@ filter2d_unrolled_bench = benchmark_filter2dunrolled(sizes)#512:-1:2)
 
 dot3_bench = benchmark_dot3(sizes)
 dot_bench = benchmark_dot(sizes)
+selfdot_bench = benchmark_selfdot(sizes)
 sse_bench = benchmark_sse(sizes)
 aplusBc_bench = benchmark_aplusBc(sizes)
 AplusAt_bench = benchmark_AplusAt(sizes)
 vexp_bench = benchmark_exp(sizes)
 randomaccess_bench = benchmark_random_access(sizes)
 logdettriangle_bench = benchmark_logdettriangle(sizes)
 
-v = 1
+const v = 1
 using Cairo, Fontconfig
 const PICTURES = joinpath(pkgdir(LoopVectorization), "docs", "src", "assets")
 function saveplot(f, br)

src/condense_loopset.jl

@@ -195,13 +195,6 @@ function add_external_functions!(q::Expr, ls::LoopSet)
     end
 end
 
-@inline unwrap_array(A) = A
-@inline unwrap_array(A::Union{SubArray,Transpose,Adjoint}) = parent(A)
-@inline array_wrapper(A) = nothing
-@inline array_wrapper(A::Transpose) = Transpose
-@inline array_wrapper(A::Adjoint) = Adjoint
-@inline array_wrapper(A::SubArray) = A.indices
-
 # Try to condense in type stable manner
 function generate_call(ls::LoopSet, inline_unroll::NTuple{3,Int8}, debug::Bool = false)
     operation_descriptions = Expr(:curly, :Tuple)
@@ -235,16 +228,6 @@ function generate_call(ls::LoopSet, inline_unroll::NTuple{3,Int8}, debug::Bool =
     add_external_functions!(q, ls)
     q
 end
-concat_vals() = Val{()}()
-# @generated concat_vals(::Val{N}) where {N} = Val{(N,)}()
-# @generated concat_vals(::Val{M}, ::Val{N}) where {M, N} = Val{(M,N)}()
-@generated function concat_vals(args...)
-    tup = Expr(:tuple)
-    for n in eachindex(args)
-        push!(tup.args, args[n].parameters[1])
-    end
-    Expr(:call, Expr(:curly, :Val, tup))
-end
 
 
 """

src/lowering.jl

@@ -146,45 +146,42 @@ function lower_block(
     blockq
 end
 
-function lower_llvm_unroll(ls::LoopSet, us::UnrollSpecification, n::Int, loop::Loop)
-    loopsym = names(ls)[n]
-    loop = getloop(ls, loopsym)
-    # loopsym = mangletiledsym(loopsym, us, n)
-    nisvectorized = false#isvectorized(us, n)
-    sl = startloop(loop, nisvectorized, loopsym)
-    # tc = terminatecondition(loop, us, n, loopsym, inclmask, 1)
-    looprange = if loop.startexact
-        if loop.stopexact
-            Expr(:(=), loopsym, Expr(:call, :(:), loop.starthint, loop.stophint))
-        else
-            Expr(:(=), loopsym, Expr(:call, :(:), loop.starthint, loop.stopsym))
-        end
-    elseif loop.stopexact
-        Expr(:(=), loopsym, Expr(:call, :(:), loop.startsym, loop.stophint))
-    else
-        Expr(:(=), loopsym, Expr(:call, :(:), loop.startsym, loop.stopsym))
-    end
-    body = lower_block(ls, us, n, false, 1)
-    push!(body.args, Expr(:loopinfo, (Symbol("llvm.loop.unroll.count"), 4)))
-    # q = Expr( :block, sl, Expr(:while, tc, body))
-    q = Expr(:for, looprange, body)
-    # if nisvectorized
-    #     tc = terminatecondition(loop, us, n, loopsym, true, 1)
-    #     body = lower_block(ls, us, n, true, 1)
-    #     push!(q.args, Expr(:if, tc, body))
-    # end
-    q
-end
-# tiledsym(s::Symbol) = Symbol("##outer##", s, "##outer##")
-# mangletiledsym(s::Symbol, us::UnrollSpecification, n::Int) = isunrolled2(us, n) ? tiledsym(s) : s
+# function lower_llvm_unroll(ls::LoopSet, us::UnrollSpecification, n::Int, loop::Loop)
+#     loopsym = names(ls)[n]
+#     loop = getloop(ls, loopsym)
+#     # loopsym = mangletiledsym(loopsym, us, n)
+#     nisvectorized = false#isvectorized(us, n)
+#     sl = startloop(loop, nisvectorized, loopsym)
+#     # tc = terminatecondition(loop, us, n, loopsym, inclmask, 1)
+#     looprange = if loop.startexact
+#         if loop.stopexact
+#             Expr(:(=), loopsym, Expr(:call, :(:), loop.starthint, loop.stophint))
+#         else
+#             Expr(:(=), loopsym, Expr(:call, :(:), loop.starthint, loop.stopsym))
+#         end
+#     elseif loop.stopexact
+#         Expr(:(=), loopsym, Expr(:call, :(:), loop.startsym, loop.stophint))
+#     else
+#         Expr(:(=), loopsym, Expr(:call, :(:), loop.startsym, loop.stopsym))
+#     end
+#     body = lower_block(ls, us, n, false, 1)
+#     push!(body.args, Expr(:loopinfo, (Symbol("llvm.loop.unroll.count"), 4)))
+#     q = Expr(:for, looprange, body)
+#     # if nisvectorized
+#     #     tc = terminatecondition(loop, us, n, loopsym, true, 1)
+#     #     body = lower_block(ls, us, n, true, 1)
+#     #     push!(q.args, Expr(:if, tc, body))
+#     # end
+#     q
+# end
 function lower_no_unroll(ls::LoopSet, us::UnrollSpecification, n::Int, inclmask::Bool)
     usorig = ls.unrollspecification[]
     nisvectorized = isvectorized(us, n)
     loopsym = names(ls)[n]
     loop = getloop(ls, loopsym)
-    if VERSION ≥ v"1.4" && !nisvectorized && !inclmask && isone(n) && !ls.loadelimination[] && (us.u₁ > 1) && (usorig.u₁ == us.u₁) && (usorig.u₂ == us.u₂) && length(loop) > 7
-        return lower_llvm_unroll(ls, us, n, loop)
-    end
+    # if VERSION ≥ v"1.4" && !nisvectorized && !inclmask && isone(n) && !ls.loadelimination[] && (us.u₁ > 1) && (usorig.u₁ == us.u₁) && (usorig.u₂ == us.u₂) && length(loop) > 7
+    #     return lower_llvm_unroll(ls, us, n, loop)
+    # end
     sl = startloop(loop, nisvectorized, loopsym)
     tc = terminatecondition(loop, us, n, loopsym, inclmask, 1)
     body = lower_block(ls, us, n, inclmask, 1)

src/operations.jl

@@ -25,7 +25,6 @@ function sameref(x::ArrayReference, y::ArrayReference)
     yinds = y.indices
     nrefs = length(xinds)
     nrefs == length(yinds) || return false
-    xoffs = x.offsets; yoffs = y.offsets
     for n ∈ 1:nrefs
         xinds[n] === yinds[n] || return false
     end

src/reconstruct_loopset.jl

@@ -461,7 +461,7 @@ Execute an `@avx` block. The block's code is represented via the arguments:
 - `vargs...` holds the encoded pointers of all the arrays (see `VectorizationBase`'s various pointer types).
 """
 @generated function _avx_!(::Val{UNROLL}, ::Type{OPS}, ::Type{ARF}, ::Type{AM}, ::Type{LPSYM}, lb::LB, vargs...) where {UNROLL, OPS, ARF, AM, LPSYM, LB}
-    1 + 1 # Irrelevant line you can comment out/in to force recompilation...
+    # 1 + 1 # Irrelevant line you can comment out/in to force recompilation...
     ls = _avx_loopset(OPS.parameters, ARF.parameters, AM.parameters, LPSYM.parameters, LB.parameters, vargs)
     # @show avx_body(ls, UNROLL)
     avx_body(ls, UNROLL)

test/broadcast.jl

@@ -66,6 +66,10 @@
         @test D1 ≈ D2
         fill!(D2, -999999); D2 = @avx C .+ At' *ˡ B;
         @test D1 ≈ D2
+        D1 .= view(C, 1, :) .+ A * B;
+        fill!(D2, -999999);
+        @avx D2 .= view(C1, 1, :) .+ A *ˡ B;
+        @test D1 ≈ D2
 
         if VERSION > v"1.2"
             b = rand(T,K); x = rand(R,N);

test/dot.jl

@@ -38,7 +38,7 @@ using Test
     function myselfdot(a)
         s = zero(eltype(a))
         @inbounds @simd for i ∈ eachindex(a)
-            s += a[i]*a[i]
+            s += getindex(a, i) * a[i]
         end
         s
     end
@@ -142,7 +142,7 @@ using Test
     function complex_mul_with_index_offset_avx!(c_re, c_im, a_re, a_im, b_re, b_im)
         @_avx for i = 1:length(a_re) - 1
             c_re[i] = b_re[i] * a_re[i + 1] - b_im[i] * a_im[i + 1]
-            c_im[i] = b_re[i] * a_im[i + 1] + b_im[i] * a_re[i + 1]
+            setindex!(c_im, b_re[i] * a_im[i + 1] + b_im[i] * a_re[i + 1], i)
         end
     end