Limit <32 register architectures to unroll/tile by up to 4.

Author: chriselrod

src/determinestrategy.jl

@@ -352,7 +352,7 @@ function solve_tilesize(
     reg_pressure::AbstractVector{Float64},
     W::Int, vectorized::Symbol
 )
-    maxTbase = maxUbase = 6#8
+    maxTbase = maxUbase = VectorizationBase.REGISTER_COUNT == 32 ? 6 : 4#8
     maxT = maxTbase#8
     maxU = maxUbase#8
     tiledloop = getloop(ls, tiled)

test/gemm.jl

@@ -294,7 +294,7 @@
     if LoopVectorization.VectorizationBase.REGISTER_COUNT == 32
         @test LoopVectorization.choose_order(lsr2amb) == ([:n, :m, :k], :k, :n, :m, 3, 6)
     else
-        @test LoopVectorization.choose_order(lsr2amb) == ([:n, :m, :k], :k, :n, :m, 1, 6)
+        @test LoopVectorization.choose_order(lsr2amb) == ([:n, :m, :k], :k, :n, :m, 2, 4)
     end
     function rank2AmulBavx!(C, Aₘ, Aₖ, B)
         @avx for m ∈ 1:size(C,1), n ∈ 1:size(C,2)

test/gemv.jl

@@ -2,7 +2,7 @@ using LoopVectorization
 using Test
 
 @testset "GEMV" begin
-    Unum, Tnum = LoopVectorization.VectorizationBase.REGISTER_COUNT == 16 ? (2, 6) : (4, 6)
+    Unum, Tnum = LoopVectorization.VectorizationBase.REGISTER_COUNT == 16 ? (3, 4) : (4, 6)
     gemvq = :(for i ∈ eachindex(y)
               yᵢ = 0.0
               for j ∈ eachindex(x)

test/miscellaneous.jl

@@ -1,6 +1,6 @@
 
 @testset "Miscellaneous" begin
-    Unum, Tnum = LoopVectorization.VectorizationBase.REGISTER_COUNT == 16 ? (2, 6) : (4, 6)
+    Unum, Tnum = LoopVectorization.VectorizationBase.REGISTER_COUNT == 16 ? (3, 4) : (4, 6)
     dot3q = :(for m ∈ 1:M, n ∈ 1:N
               s += x[m] * A[m,n] * y[n]
               end);