diff --git a/extra/replay_pkl.py b/extra/replay_pkl.py
index f277740ac5..e44bf9c316 100644
--- a/extra/replay_pkl.py
+++ b/extra/replay_pkl.py
@@ -106,7 +106,10 @@ if __name__ == "__main__":
             #k.apply_opt(Opt(OptOps.UPCAST, 0, 2))
           """
           # blocked
-          if k.full_shape[-3:] == (32,3,3):
+          if len(k.full_shape) == 6 and k.full_shape[-4:] == (32,3,3,3):
+            k.apply_opt(Opt(OptOps.UPCAST, 2, 32))
+            k.apply_opt(Opt(OptOps.UPCAST, 1, 4))
+          elif k.full_shape[-3:] == (32,3,3):
             #if k.full_shape[-4]%4 != 0: k.apply_opt(Opt(OptOps.PADTO, len(k.full_shape)-4, 4))
             # 3x3 dwconv
             k.apply_opt(Opt(OptOps.UNROLL, 0, 0))
diff --git a/tinygrad/codegen/devectorizer.py b/tinygrad/codegen/devectorizer.py
index 7c1ea36dd1..2497286b16 100644
--- a/tinygrad/codegen/devectorizer.py
+++ b/tinygrad/codegen/devectorizer.py
@@ -159,7 +159,7 @@ def split_load_store(ctx:Renderer|None, ls:UOp, idx:UOp):
   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 [192, 288]: return None  # leave 192 alone
+    if ls.dtype.count in [192, 288, 160, 96]: 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