diff --git a/extra/replay_pkl.py b/extra/replay_pkl.py
index 9b213e1c40..8b0a633b22 100644
--- a/extra/replay_pkl.py
+++ b/extra/replay_pkl.py
@@ -132,6 +132,7 @@ if __name__ == "__main__":
             #k.apply_opt(Opt(OptOps.PADTO, 4, 4))
             k.apply_opt(Opt(OptOps.UNROLL, 1, 0))
             k.apply_opt(Opt(OptOps.UPCAST, 2, 128))
+            k.apply_opt(Opt(OptOps.UPCAST, 1, 2))
           elif knum == 5:
             k.apply_opt(Opt(OptOps.UNROLL, 1, 0))
             k.apply_opt(Opt(OptOps.UPCAST, 2, 0))
diff --git a/tinygrad/codegen/symbolic.py b/tinygrad/codegen/symbolic.py
index f0a0e0e259..7a0558c913 100644
--- a/tinygrad/codegen/symbolic.py
+++ b/tinygrad/codegen/symbolic.py
@@ -196,7 +196,8 @@ gep_pushing = PatternMatcher([
 
 symbolic = symbolic_simple+PatternMatcher([
   # ** COMMUTATIVE flipping (only for ints) **
-  (UPat(GroupOp.Commutative, dtype=dtypes.int, name='x'), lambda x: x.replace(src=x.src[::-1]) if x.src[1].tuplize < x.src[0].tuplize else None),
+  # NOTE: this can break merging vector math by only flipping some of them
+  #(UPat(GroupOp.Commutative, dtype=dtypes.int, name='x'), lambda x: x.replace(src=x.src[::-1]) if x.src[1].tuplize < x.src[0].tuplize else None),
   # ** boolean algebra **
   (UPat.var("x") | (UPat.var("x") & UPat.var()), lambda x: x), # x|(x&y) -> x
   # ** combine terms **
diff --git a/tinygrad/runtime/ops_dsp.py b/tinygrad/runtime/ops_dsp.py
index 04d6c6d69a..610e738c2e 100644
--- a/tinygrad/runtime/ops_dsp.py
+++ b/tinygrad/runtime/ops_dsp.py
@@ -163,7 +163,8 @@ def prefetch_l1(ld:UOp):
 def vectorize_shuffle(x:UOp):
   if not all(s.op in {Ops.GEP, Ops.CONST} for s in x.src): return None
   gepped = dedup([s.src[0] for s in x.src if s.op is Ops.GEP])
-  if len(gepped) < 2: return None
+  if len(gepped) != 3: return None
+  if not all(x.dtype.scalar() is dtypes.uchar and x.dtype.count == 128 for x in gepped): return None
   arg = []
   for s in x.src:
     if s.op is Ops.GEP: