Added explicit dependency on ArrayInterface

Author: chriselrod

Project.toml

@@ -4,11 +4,12 @@ authors = ["Mason Protter", "Chris Elrod", "Dilum Aluthge", "contributors"]
 version = "0.1.0-DEV"
 
 [deps]
+ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"
 LoopVectorization = "bdcacae8-1622-11e9-2a5c-532679323890"
 VectorizationBase = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
 
 [compat]
-BenchmarkTools = "0.5"
+ArrayInterface = "2.14"
 LoopVectorization = "0.9.14"
 VectorizationBase = "0.14.9"
 julia = "1.5"
@@ -19,8 +20,8 @@ InteractiveUtils = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 LoopVectorization = "bdcacae8-1622-11e9-2a5c-532679323890"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
-VectorizationBase = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
+VectorizationBase = "3d5dd08c-fd9d-11e8-17fa-ed2836048c2f"
 
 [targets]
 test = ["BenchmarkTools", "InteractiveUtils", "LinearAlgebra", "LoopVectorization", "Random", "VectorizationBase", "Test"]

src/Octavian.jl

@@ -1,5 +1,6 @@
 module Octavian
 
+import ArrayInterface
 import LoopVectorization
 import VectorizationBase
 

src/matmul.jl

@@ -1,6 +1,12 @@
 evenly_divide(x, y) = cld(x, cld(x, y))
 evenly_divide(x, y, z) = cld(evenly_divide(x, y), z) * z
 
+_dim1contig(::ArrayInterface.Contiguous) = false
+_dim1contig(::ArrayInterface.Contiguous{1}) = true
+dim1contig(::Type{A}) where {A <: StridedArray} = _dim1contig(ArrayInterface.contiguous_axis(A))
+dim1contig(::Type) = false
+dim1contig(A) = dim1contig(typeof(A))
+
 """
     matmul!(C::AbstractMatrix, A::AbstractMatrix, B::AbstractMatrix, _α = 1, _β = 0)
 """
@@ -10,7 +16,7 @@ function matmul!(C::AbstractMatrix{T}, A::AbstractMatrix{T}, B::AbstractMatrix{T
     M, K, N = matmul_sizes(C, A, B)
 
     # Check if maybe it's better not to pack at all.
-    if M * K ≤ _Mc * _Kc && A isa DenseArray && C isa StridedArray && B isa StridedArray && #
+    if M * K ≤ _Mc * _Kc && dim1contig(A) && LoopVectorization.check_args(C, A, B) &&
         (stride(A,2) ≤ 72 || (iszero(stride(A,2) & (VectorizationBase.pick_vector_width(eltype(A))-1)) && iszero(reinterpret(Int,pointer(A)) & 63)))
         macrokernel!(C, A, B, _α, _β)
         return C

test/matmul_coverage.jl

@@ -57,3 +57,19 @@ end
     @test @time(Octavian.matmul(A, B′)) == A * B′
     @test @time(Octavian.matmul(A′, B′)) == A′ * B′
 end
+
+@time @testset "A not-a-StrideArray" begin
+    m = 20
+    n = 20
+    k = 20
+    A = view(rand(Float64, 2m, 2k), 1:2:2m, 2:2:2k)
+    B = rand(Float64, k, n)
+    A′ = view(permutedims(parent(A))', 1:2:2m, 2:2:2k)
+    B′ = permutedims(B)'
+    @show m, k, n
+    @test @time(Octavian.matmul(A, B)) ≈ A * B
+    @test @time(Octavian.matmul(A′, B)) ≈ A′ * B
+    @test @time(Octavian.matmul(A, B′)) ≈ A * B′
+    @test @time(Octavian.matmul(A′, B′)) ≈ A′ * B′
+end
+