Misc updates, bumping minor version due to new assumption that loop iterables are not isempty.

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <[email protected]>"]
-version = "0.7.8"
+version = "0.8.0"
 
 [deps]
 DocStringExtensions = "ffbed154-4ef7-542d-bbb7-c09d3a79fcae"
@@ -15,7 +15,7 @@ VectorizationBase = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
 [compat]
 DocStringExtensions = "0.8"
 OffsetArrays = "1"
-SIMDPirates = "0.7.23"
+SIMDPirates = "0.7.24"
 SLEEFPirates = "0.4.8"
 UnPack = "0"
 VectorizationBase = "0.11.3"

README.md

@@ -14,6 +14,14 @@ Pkg.add("LoopVectorization")
 ```
 LoopVectorization is supported on Julia 1.1 and later. It is tested on Julia 1.1, 1.3, and nightly.
 
+## Warning
+
+Misusing LoopVectorization can have [serious consequences](http://catb.org/jargon/html/N/nasal-demons.html). Like `@inbounds`, misusing it can lead to segfaults and memory corruption.
+We expect that any time you use the `@avx` macro with a given block of code that you:
+1. Are not indexing an array out of bounds. `@avx` does not perform any bounds checking.
+2. Are not iterationg over an empty collection. Iterating over an empty loop such as `for i ∈ eachindex(Float64[])` is undefined behavior, and will likely result in the out of bounds memory accesses. Ensure that loops behave correctly.
+3. Are not relying on a specific execution order. `@avx` can and will re-order operations and loops inside its scope, so the correctness cannot depend on a particular order. You cannot implement `cumsum` with `@avx`.
+
 ## Usage
 
 This library provides the `@avx` macro, which may be used to prefix a `for` loop or broadcast statement.

benchmark/benchmarkflops.jl

@@ -61,9 +61,10 @@ function matmul_bench!(br, C, A, B, i)
     @assert C ≈ Cblas "Fort builtin gemm wrong?"; fill!(C, NaN)
     br[10,i] = n_gflop / @belapsed ifgemm_builtin!($C, $A, $B)
     @assert C ≈ Cblas "ifort builtin gemm wrong?"; fill!(C, NaN)
-    br[11,i] = n_gflop / @belapsed mul!($C, $A, $B)
+    br[11,i] = n_gflop / @belapsed mul!($C, $A, $B);
+    fill!(C, NaN)
     br[12,i] = n_gflop / @belapsed dgemmmkl!($C, $A, $B)
-    @assert C ≈ Cblas "MKL JIT gemm wrong?"; fill!(C, NaN)
+    @assert C ≈ Cblas "MKL JIT gemm wrong?"
     # br[12,i] = n_gflop / @belapsed gemmavx!($C, $A, $B)
 end
 function A_mul_B_bench!(br, s, i)
@@ -222,7 +223,7 @@ function gemv_bench!(br, x, A, y, i)
     br[10,i] = n_gflop / @belapsed ifgemv_builtin!($x, $A, $y)
     @assert x ≈ xblas "ifort wrong?"; fill!(x, NaN);
     br[11,i] = n_gflop / @belapsed mul!($x, $A, $y)
-    br[11,i] = n_gflop / @belapsed dgemvmkl!($x, $A, $y)
+    br[12,i] = n_gflop / @belapsed dgemvmkl!($x, $A, $y)
     @assert x ≈ xblas "gemvmkl wrong?"; fill!(x, NaN);
 end
 function A_mul_vb_bench!(br, s, i)
@@ -455,7 +456,7 @@ function logdettriangle_bench!(br, s, i)
     br[7,i] = n_gflop / @belapsed logdet($U)
 end
 function benchmark_logdettriangle(sizes)
-    tests = ["LoopVectorization", "Julia", "Clang", "GFortran", "icc", "ifort", "Julia-builtin"]
+    tests = ["LoopVectorization", "Julia", "Clang", "GFortran", "icc", "ifort", "LinearAlgebra"]
     br = BenchmarkResult(tests, sizes)
     sm = br.sizedresults.results
     pmap(is -> logdettriangle_bench!(sm, is[2], is[1]), enumerate(sizes))

benchmark/looptests.jl

@@ -61,9 +61,9 @@ function jgemm!(𝐂, 𝐀ᵀ::Adjoint, 𝐁ᵀ::Adjoint)
     end
 end
 function gemmavx!(𝐂, 𝐀, 𝐁)
-    @avx for m ∈ 1:size(𝐀,1), n ∈ 1:size(𝐁,2)
+    @avx for m ∈ axes(𝐀,1), n ∈ axes(𝐁,2)
         𝐂ₘₙ = zero(eltype(𝐂))
-        for k ∈ 1:size(𝐀,2)
+        for k ∈ axes(𝐀,2)
             𝐂ₘₙ += 𝐀[m,k] * 𝐁[k,n]
         end
         𝐂[m,n] = 𝐂ₘₙ

benchmark/plotbenchmarks.jl

@@ -11,36 +11,43 @@ function Base.show(io::IO, br::BenchmarkResult)
 end
 
 
-using Colors, Gadfly
-const COLORS = distinguishable_colors(21, [RGB(1,1,1), RGB(0,0,0)])
-const COLOR_MAP = Dict{String,RGB{Colors.N0f8}}()
+using Colors, ColorSchemes, Gadfly
+const COLORS = [RGB(0.0,0.0,0.0),RGB(1.0,0.0,0.0)]
+# const COLORS = [RGB(0.0,0.0,0.0),RGB(0.0,1.0,0.0)]
+# const COLORS = distinguishable_colors(14, pushfirst!(get(ColorSchemes.Paired_12, (0.5:11.5) ./ 12), RGB(0.0,0.0,0.0)))
+for i ∈ 1:12 # 11 is number of tested libs - 2
+    push!(COLORS, get(ColorSchemes.cyclic_mygbm_30_95_c78_n256_s25, i/12))
+    # push!(COLORS, get(ColorSchemes.vikO, (i-0.5)/12))
+end
+# const COLOR_MAP = Dict{String,RGB{Float64}}()
+# const COLOR_MAP = Dict{String,RGB{Colors.N0f8}}()
+const COLOR_MAP64 = Dict{String,RGB{Float64}}()
 function getcolor(s::String)
-    get!(COLOR_MAP, s) do
-        COLORS[length(COLOR_MAP) + 2]
+    get!(COLOR_MAP64, s) do
+        COLORS[length(COLOR_MAP64) + 1]
     end
 end
-
+replace_and(str) = replace(str, '&' => "with")
 
 function Gadfly.plot(br::BenchmarkResult)
     res = br.sizedresults.results
     sizes = br.sizedresults.sizes
     # sizes = Vector{eltype(brsizes)}(undef, length(res))
-    tests = @view(br.tests[2:end])
-    ntests = length(tests)
+    tests = replace_and.(@view(br.tests[2:end]))
     colors = getcolor.(tests)
 
     xt = 0:20:260
     maxres = maximum(res)
     maxtick = 10round(Int, 0.1maxres)
     yt = if iszero(maxtick)
         maxtick = 10round(0.1maxres)
-        range(0, maxtick, length = 20)
-    elseif maxtick < 50
-        0:5:maxtick
-    elseif maxtick < 20
-        0:2:maxtick
+        range(0, maxres, length = 20)
     elseif maxtick < 10
         0:1:maxtick
+    elseif maxtick < 20
+        0:2:maxtick
+    elseif maxtick < 50
+        0:5:maxtick
     else
         0:10:maxtick
     end