Add a padding heuristic

python/dune/perftool/sumfact/vectorization.py

@@ -150,12 +150,13 @@ def greedy_vectorization_strategy(sumfacts, width):
     sumfacts = set(sumfacts)
     horizontal = width
     vertical = 1
+    allowed_padding = 1
 
     result = {}
 
     while horizontal > 0:
         if horizontal > 1:
-            horizontal_kernels = horizontal_vectorization_strategy(sumfacts, horizontal, allow_padding=0)
+            horizontal_kernels = horizontal_vectorization_strategy(sumfacts, horizontal, allow_padding=allowed_padding)
         else:
             horizontal_kernels = {sf: sf for sf in sumfacts}
 
@@ -170,6 +171,8 @@ def greedy_vectorization_strategy(sumfacts, width):
         horizontal = horizontal // 2
         vertical = vertical * 2
 
+        # We heuristically allow padding only on the full SIMD width
+        allowed_padding = 0
     return result