diff --git a/tinygrad/codegen/kernel.py b/tinygrad/codegen/kernel.py
index f34e426edf..c0dfe37008 100644
--- a/tinygrad/codegen/kernel.py
+++ b/tinygrad/codegen/kernel.py
@@ -454,7 +454,7 @@ class Kernel:
         # if this is small, swap it
         # NOTE: this is breaking something (should be fixed w/o padto)
         # kernel 23 is broken with this
-        #if k.full_shape[0] <= 6: k.apply_opt(Opt(OptOps.SWAP, 0, 1))
+        if k.full_shape[0] <= 6: k.apply_opt(Opt(OptOps.SWAP, 0, 1))
       elif k.full_shape[-4:] == (32,3,3,3):
         # 3x3 normal conv
         k.apply_opt(Opt(OptOps.UNROLL, 2, 0))
@@ -478,7 +478,7 @@ class Kernel:
         k.apply_opt(Opt(OptOps.UPCAST, 2, 32))
         if k.full_shape[1]%4 == 0: k.apply_opt(Opt(OptOps.UPCAST, 1, 4))
         # if the more is small, upcast it (kernel 50 is broken with this)
-        #if k.full_shape[0] <= 6: k.apply_opt(Opt(OptOps.UPCAST, 0, 0))
+        if k.full_shape[0] <= 6: k.apply_opt(Opt(OptOps.UPCAST, 0, 0))
       elif len(k.full_shape) == 2 and k.first_reduce == 1:
         # unroll to 4 if we can
         if k.full_shape[k.first_reduce]%4 == 0: k.apply_opt(Opt(OptOps.UNROLL, 0, 4))
diff --git a/tinygrad/runtime/ops_dsp.py b/tinygrad/runtime/ops_dsp.py
index 1c26778160..e9aacfd2d3 100644
--- a/tinygrad/runtime/ops_dsp.py
+++ b/tinygrad/runtime/ops_dsp.py
@@ -373,8 +373,8 @@ dsp_pm_late = PatternMatcher([
 
   #(UPat(Ops.BITCAST, src=(UPat(Ops.LOAD, name="ld"),), name="bc"),
   # lambda ld, bc: ld.src[0].src[0].cast(bc.dtype.ptr(ld.src[0].dtype.size)).load(dtype=bc.dtype)),
-  #(UPat(Ops.GEP, name="x"), lambda x: UOp(Ops.CUSTOM, x.dtype, x.src,
-  #  "__builtin_shufflevector({0}, {0}, "+','.join([f'{y:4d}' for y in x.arg])+")") if len(x.arg) > 1 and x.src[0].dtype.count > 1 else None),
+  (UPat(Ops.GEP, name="x"), lambda x: UOp(Ops.CUSTOM, x.dtype, x.src,
+    "__builtin_shufflevector({0}, {0}, "+','.join([f'{y:4d}' for y in x.arg])+")") if len(x.arg) > 1 and x.src[0].dtype.count > 4 else None),
   (UPat.var("x")+UPat(Ops.VECTORIZE,src=UPat.var("y")),
    lambda x,y: x+UOp(Ops.CUSTOMI,x.dtype,(y,),arg="{0}") if x.op is not Ops.CUSTOMI or x.arg != "{0}" else None),
   (UPat.var("x")*UPat(Ops.VECTORIZE,src=UPat.var("y")),