Test cleanup

define NNlib.leakyrelu(::AbstractSIMD) for Julia 1.6 test, and don't test `check_order` for westmere like architectures (i.e., simd width of 16 bytes, 16 registers) because we don't have a set of validated reasonable results to compare with.

test/forwarddiffext.jl

@@ -15,6 +15,11 @@ function tovec(x::ForwardDiff.Dual{T,V,N}) where {T,V,N}
   return ret
 end
 
+if LoopVectorization.ifelse !== Base.ifelse
+  @inline function NNlib.leakyrelu(x::LoopVectorization.AbstractSIMD, a=NNlib.oftf(x, NNlib.leakyrelu_a)) 
+     LoopVectorization.ifelse(x>0, float(x), NNlib.oftf(x, a*x))  # max(a*x, x) is 3x slower
+  end
+end
 
 vx0 = randnvec()
 vx1 = randnvec()

test/gemm.jl

@@ -8,7 +8,7 @@
     Unum, Tnum = LoopVectorization.register_count() == 16 ? (2, 6) : (4, 6)
   end
   Unumt, Tnumt = LoopVectorization.register_count() == 16 ? (2, 6) : (5, 5)
-  if LoopVectorization.register_count() != 8
+  if (LoopVectorization.register_count() != 8) &&  ((LoopVectorization.pick_vector_width(Float64) != 2) ||(LoopVectorization.register_count() != 16))
     @test @inferred(LoopVectorization.matmul_params()) == (Unum, Tnum)
   end
 
@@ -30,7 +30,7 @@
     end
   )
   lsAmulBt1 = LoopVectorization.loopset(AmulBtq1)
-  if LoopVectorization.register_count() != 8
+  if (LoopVectorization.register_count() != 8) &&  ((LoopVectorization.pick_vector_width(Float64) != 2) ||(LoopVectorization.register_count() != 16))
     @test LoopVectorization.choose_order(lsAmulBt1) ==
           (Symbol[:n, :m, :k], :m, :n, :m, Unum, Tnum)
   end
@@ -43,7 +43,7 @@
     end
   )
   lsAmulB1 = LoopVectorization.loopset(AmulBq1)
-  if LoopVectorization.register_count() != 8
+  if (LoopVectorization.register_count() != 8) &&  ((LoopVectorization.pick_vector_width(Float64) != 2) ||(LoopVectorization.register_count() != 16))
     @test LoopVectorization.choose_order(lsAmulB1) ==
           (Symbol[:n, :m, :k], :m, :n, :m, Unum, Tnum)
   end
@@ -56,7 +56,7 @@
     end
   )
   lsAmulB2 = LoopVectorization.loopset(AmulBq2)
-  if LoopVectorization.register_count() != 8
+  if (LoopVectorization.register_count() != 8) &&  ((LoopVectorization.pick_vector_width(Float64) != 2) ||(LoopVectorization.register_count() != 16))
     @test LoopVectorization.choose_order(lsAmulB2) ==
           (Symbol[:n, :m, :k], :m, :n, :m, Unum, Tnum)
   end
@@ -70,11 +70,9 @@
     end
   )
   lsAmulB3 = LoopVectorization.loopset(AmulBq3)
-  if LoopVectorization.register_count() != 8
+  if (LoopVectorization.register_count() != 8) &&  ((LoopVectorization.pick_vector_width(Float64) != 2) ||(LoopVectorization.register_count() != 16))
     @test LoopVectorization.choose_order(lsAmulB3) ==
           (Symbol[:n, :m, :k], :m, :n, :m, Unum, Tnum)
-  end
-  if LoopVectorization.register_count() != 8
     for (fA, fB, v, Un, Tn) ∈ [
       (identity, identity, :m, Unum, Tnum),
       (adjoint, identity, :k, Unumt, Tnumt),
@@ -178,8 +176,13 @@
   )
   lsAmuladd = LoopVectorization.loopset(Amuladdq)
   if LoopVectorization.register_count() != 8
+    if LoopVectorization.pick_vector_width(Float64) != 2
     @test LoopVectorization.choose_order(lsAmuladd) ==
           (Symbol[:n, :m, :k], :m, :n, :m, Unum, Tnum)
+    else
+    @test LoopVectorization.choose_order(lsAmuladd) ==
+          (Symbol[:m, :n, :k], :n, :m, :m, Unum, Tnum)
+     end
   end
   Atmuladdq = :(
     for m ∈ axes(A, 2), n ∈ axes(B, 2)
@@ -410,9 +413,13 @@
       @test LoopVectorization.choose_order(lsr2amb) == ([:m, :n, :k], :m, :n, :m, 3, 7)
     end
   elseif LoopVectorization.register_count() == 16
+    if LoopVectorization.pick_vector_width(Float64) == 4
     # @test LoopVectorization.choose_order(lsr2amb) == ([:m, :n, :k], :m, :n, :m, 1, 6)
     # @test LoopVectorization.choose_order(lsr2amb) == ([:m, :n, :k], :m, :n, :m, 2, 4)
-    @test LoopVectorization.choose_order(lsr2amb) == ([:n, :m, :k], :n, :m, :m, 3, 3)
+      @test LoopVectorization.choose_order(lsr2amb) == ([:n, :m, :k], :n, :m, :m, 3, 3)
+    elseif LoopVectorization.pick_vector_width(Float64) == 2
+      @test LoopVectorization.choose_order(lsr2amb) ==  ([:m, :n, :k], :n, :m, :m, 3, 3)
+    end
   end
   function rank2AmulBavx!(C, Aₘ, Aₖ, B)
     @turbo for m ∈ axes(C, 1), n ∈ axes(C, 2)