Set max tile-unroll to 4 as a heuristic.

JuliaSIMD · Jan 1, 2020 · 090c173 · 090c173 · chriselrod · Jan 1, 2020
1 parent fb7bf0e
commit 090c173
Show file tree

Hide file tree

Showing 2 changed files with 3 additions and 3 deletions.
diff --git a/src/determinestrategy.jl b/src/determinestrategy.jl
@@ -242,8 +242,8 @@ function solve_tilesize(
     cost_vec::AbstractVector{Float64} = @view(ls.cost_vec[:,1]),
     reg_pressure::AbstractVector{Int} = @view(ls.reg_pres[:,1])
 )
-    maxT = isstaticloop(ls, tiled) ? looprangehint(ls, tiled) : 8#REGISTER_COUNT
-    maxU = isstaticloop(ls, unrolled) ? looprangehint(ls, unrolled) : 4#REGISTER_COUNT
+    maxT = isstaticloop(ls, tiled) ? looprangehint(ls, tiled) : 4#REGISTER_COUNT
+    maxU = isstaticloop(ls, unrolled) ? looprangehint(ls, unrolled) : 8#REGISTER_COUNT
     solve_tilesize(cost_vec, reg_pressure, maxU, maxT)
 end
 

diff --git a/test/runtests.jl b/test/runtests.jl
@@ -43,7 +43,7 @@ gemmq = :(for i ∈ 1:size(A,1), j ∈ 1:size(B,2)
       end)
 
 lsgemm = LoopVectorization.LoopSet(gemmq);
-U, T = LoopVectorization.VectorizationBase.REGISTER_COUNT == 16 ? (3,4) : (4, 6)
+U, T = LoopVectorization.VectorizationBase.REGISTER_COUNT == 16 ? (3,4) : (6, 4)
 @test LoopVectorization.choose_order(lsgemm) == (Symbol[:j,:i,:k], U, T)
 
 function mygemm!(C, A, B)