Update chooseorder tests for non-AVX512F

Author: chriselrod

src/determinestrategy.jl

@@ -384,8 +384,10 @@ end
 
 function solve_unroll(X, R, u₁L, u₂L, u₁step, u₂step)
     X₁, X₂, X₃, X₄ = X[1], X[2], X[3], X[4]
-    # If we don't have AVX512, masks occupy a vector register
-    VectorizationBase.AVX512F || (R[3] += 1)
+    # If we don't have AVX512, masks occupy a vector register;
+    # AVX512F is currently defined as `false` for non-x86 CPUs, but
+    # should instead define generic constant `HAS_OPMASK_REGISTERS` in VectorizationBase.jl to use here instead.
+    AVX512F || (R[3] += 1)
     R₁, R₂, R₃, R₄, R₅ = R[1], R[2], R[3], R[4], R[5]
     iszero(R₅) || return solve_unroll_iter(X, R, u₁L, u₂L, u₁step:u₁step:10, u₂step:u₂step:10)
     RR = REGISTER_COUNT - R₃ - R₄

test/gemm.jl

@@ -1,7 +1,7 @@
 @testset "GEMM" begin
     # using LoopVectorization, LinearAlgebra, Test; T = Float64
     Unum, Tnum = LoopVectorization.REGISTER_COUNT == 16 ? (2, 6) : (3, 9)
-    Unumt, Tnumt = LoopVectorization.REGISTER_COUNT == 16 ? (3, 4) : (5, 5)
+    Unumt, Tnumt = LoopVectorization.REGISTER_COUNT == 16 ? (2, 6) : (5, 5)
     if LoopVectorization.REGISTER_COUNT != 8
         @test LoopVectorization.mᵣ == Unum
         @test LoopVectorization.nᵣ == Tnum
@@ -353,7 +353,7 @@
     if LoopVectorization.REGISTER_COUNT == 32
         @test LoopVectorization.choose_order(lsr2amb) == ([:n, :m, :k], :m, :n, :m, 3, 7)
     elseif LoopVectorization.REGISTER_COUNT == 16
-        @test LoopVectorization.choose_order(lsr2amb) == ([:m, :n, :k], :n, :m, :m, 4, 2)
+        @test LoopVectorization.choose_order(lsr2amb) == ([:m, :n, :k], :m, :n, :m, 1, 6)
     end
     function rank2AmulBavx!(C, Aₘ, Aₖ, B)
         @avx for m ∈ axes(C,1), n ∈ axes(C,2)