Require at least one loop to be fully unrolled to consider fully unrolling a static loop when doing convolution-like loops.

Author: chriselrod

Project.toml

@@ -1,7 +1,7 @@
 name = "LoopVectorization"
 uuid = "bdcacae8-1622-11e9-2a5c-532679323890"
 authors = ["Chris Elrod <[email protected]>"]
-version = "0.9.7"
+version = "0.9.8"
 
 [deps]
 ArrayInterface = "4fba245c-0d91-5ea0-9b3e-6abc04ee57a9"

src/lowering.jl

@@ -263,7 +263,17 @@ end
 #     )
 # end
 
+function check_full_conv_kernel(ls, us, N)
+    loop₁ = getloop(ls, names(ls)[us.u₁loopnum])
+    (isstaticloop(loop₁) && length(loop₁) == us.u₁) && return true
+    loop₂ = getloop(ls, names(ls)[us.u₂loopnum])
+    (isstaticloop(loop₂) && length(loop₂) == us.u₂) && return true
+    false
+end
 function allinteriorunrolled(ls::LoopSet, us::UnrollSpecification, N)
+    if ls.loadelimination[]
+        check_full_conv_kernal(ls, us, N) || return false
+    end
     unroll_total = 1
     for n ∈ 1:N-1
         loop = getloop(ls, names(ls)[n])