diff --git a/tinygrad/codegen/devectorizer.py b/tinygrad/codegen/devectorizer.py index 539aa7c322..0fa56c015c 100644 --- a/tinygrad/codegen/devectorizer.py +++ b/tinygrad/codegen/devectorizer.py @@ -158,6 +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 [192]: 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