diff --git a/extra/replay_pkl.py b/extra/replay_pkl.py
index 6697bb1647..b4a333ea0b 100644
--- a/extra/replay_pkl.py
+++ b/extra/replay_pkl.py
@@ -122,11 +122,12 @@ if __name__ == "__main__":
           elif knum == 15:
             # 28x28, 192 chan, 3x3 dwconv
             k.apply_opt(Opt(OptOps.UNROLL, 1, 0))
-            k.apply_opt(Opt(OptOps.PADTO, 2, 128))
-            k.apply_opt(Opt(OptOps.UPCAST, 2, 128))
+            #k.apply_opt(Opt(OptOps.PADTO, 2, 128))
+            k.apply_opt(Opt(OptOps.UPCAST, 2, 64))
             #k.apply_opt(Opt(OptOps.UNROLL, 0, 0))
             #k.apply_opt(Opt(OptOps.UPCAST, 2, 32))
             #k.apply_opt(Opt(OptOps.UPCAST, 1, 4))
+            pass
           elif knum == 3:
             k.apply_opt(Opt(OptOps.UNROLL, 0, 8))
             k.apply_opt(Opt(OptOps.UPCAST, 1, 16))
diff --git a/tinygrad/codegen/devectorizer.py b/tinygrad/codegen/devectorizer.py
index d9a6b9cf34..8a4cd7fefe 100644
--- a/tinygrad/codegen/devectorizer.py
+++ b/tinygrad/codegen/devectorizer.py
@@ -158,7 +158,7 @@ def split_load_store(ctx:Renderer|None, ls:UOp, idx:UOp):
   must_divide = True
   if ctx is not None and ctx.device == "DSP":
     lengths = [128,64,32,16,8,4]
-    if ls.dtype.count in [128+64, 128*2+64, 128*4+64]: return None  # leave 192 alone
+    #if ls.dtype.count in [128+64, 128*2+64, 128*4+64]: return None  # leave 192 alone
     # we really want stores to be 128 for fast casting
     #if ls.op is Ops.LOAD: lengths = [192]+lengths
     #if ls.op is Ops.LOAD: lengths = [1536,1024,512,384,256,192,96]+lengths