Not sure whether this is a bug or feature. I was trying figure out why certain apps have unaligned loads/stores when tile size is a multiple of burst size. And it turns out to be the type conversion on loop indices mess up constant propagation in spatial.
is a aligned load. because requestLength.s in DRAMTransfersInternal is not statically known. I would assume loop bound is already of type Index so .to[Index] effectively should be a no op?
Not sure whether this is a bug or feature. I was trying figure out why certain apps have unaligned loads/stores when tile size is a multiple of burst size. And it turns out to be the type conversion on loop indices mess up constant propagation in spatial.
For example in GEMM_Blocked,
is a unaligned load while
is a aligned load. because
requestLength.s
inDRAMTransfersInternal
is not statically known. I would assume loop bound is already of type Index so.to[Index]
effectively should be a no op?