Ever so slight nudge toward more favorable patterns.

Author: chriselrod

src/determinestrategy.jl

@@ -696,7 +696,11 @@ function evaluate_cost_tile(
     u₁, u₂, ucost = solve_unroll(ls, u₁loopsym, u₂loopsym, cost_vec, reg_pressure, W, vectorized)
     outer_reduct_penalty = length(ls.outer_reductions) * (u₁ + isodd(u₁))
     favor_bigger_u₂ = u₁ - u₂
-    u₁, u₂, costpenalty * ucost + stride_penalty(ls, order) + outer_reduct_penalty + favor_bigger_u₂, choose_to_inline[]
+    u₁v = vectorized === u₁loopsym; u₂v = vectorized === u₂loopsym
+    favor_smaller_vectorized = u₁v ? ( u₁ - u₂ )  : (u₂v ?  ( u₂ - u₁ ) : 0 )
+    favor_u₁_vectorized = -0.2u₁v
+    favoring_heuristics = favor_bigger_u₂ + 0.5favor_smaller_vectorized + favor_u₁_vectorized
+    u₁, u₂, costpenalty * ucost + stride_penalty(ls, order) + outer_reduct_penalty + favoring_heuristics, choose_to_inline[]
 end
 
 function should_inline(ls::LoopSet, u₁::Int, u₂::Int)

src/lowering.jl

@@ -186,9 +186,9 @@ end
 function assume(ex)
     Expr(:call, Expr(:(.), Expr(:(.), :LoopVectorization, QuoteNode(:SIMDPirates)), QuoteNode(:assume)), ex)
 end
-function expect(ex)
-    Expr(:call, Expr(:(.), Expr(:(.), :LoopVectorization, QuoteNode(:SIMDPirates)), QuoteNode(:expect)), ex)
-end
+# function expect(ex)
+#     Expr(:call, Expr(:(.), Expr(:(.), :LoopVectorization, QuoteNode(:SIMDPirates)), QuoteNode(:expect)), ex)
+# end
 function loopiteratesatleastonce(loop::Loop, as::Bool = true)
     comp = if loop.startexact # requires !loop.stopexact
         Expr(:call, :>, loop.stopsym, loop.starthint - 1)
@@ -197,7 +197,8 @@ function loopiteratesatleastonce(loop::Loop, as::Bool = true)
     else
         Expr(:call, :>, loop.stopsym, Expr(:call, :-, loop.stopsym, 1))
     end
-    as ? assume(comp) : expect(comp)
+    # as ? assume(comp) : expect(comp)
+    assume(comp)
 end
 function lower_no_unroll(ls::LoopSet, us::UnrollSpecification, n::Int, inclmask::Bool)
     usorig = ls.unrollspecification[]

test/gemm.jl

@@ -13,7 +13,7 @@
     # @test LoopVectorization.choose_order(lsAmulBt1) == (Symbol[:n,:m,:k], :n, :m, :m, Unum, Tnum)
     @test LoopVectorization.choose_order(lsAmulBt1) == (Symbol[:n,:m,:k], :m, :n, :m, Unum, Tnum)
 
-    AmulBq1 = :(for m ∈ 1:size(A,1), n ∈ 1:size(B,2)
+    AmulBq1 = :(for n ∈ 1:size(B,2), m ∈ 1:size(A,1)
                 C[m,n] = zeroB
                 for k ∈ 1:size(A,2)
                 C[m,n] += A[m,k] * B[k,n]