diff --git a/tinygrad/codegen/devectorizer.py b/tinygrad/codegen/devectorizer.py
index f2c655f248..4d4aca56ae 100644
--- a/tinygrad/codegen/devectorizer.py
+++ b/tinygrad/codegen/devectorizer.py
@@ -150,6 +150,7 @@ 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]
     # 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
     must_divide = False
   elif buf.dtype.base != dtypes.float and buf.dtype.base != dtypes.half and not isinstance(buf.dtype, ImageDType):
diff --git a/tinygrad/runtime/ops_dsp.py b/tinygrad/runtime/ops_dsp.py
index 9b71636062..f45e2a9b15 100644
--- a/tinygrad/runtime/ops_dsp.py
+++ b/tinygrad/runtime/ops_dsp.py
@@ -179,7 +179,8 @@ class DSPRenderer(ClangRenderer):
   device = "DSP"
   supports_float4 = True
   buffer_suffix = " restrict __attribute__((align_value(128)))"
-  kernel_prefix = vmemu_support + "__attribute__((noinline)) "
+  #kernel_prefix = vmemu_support + "__attribute__((noinline)) "
+  kernel_prefix = "__attribute__((noinline)) "
   pre_matcher = dsp_pm
   extra_matcher = dsp_pm_late+ClangRenderer.extra_matcher+pretty_render
   type_map = { **ClangRenderer.type_map, dtypes.uint64: "unsigned long long", dtypes.int64: "long long" }