MatchToAlloc generates an AllocSparseBuffer instead of AllocBuffer for sparse tensors, actually to make CompactBufferRegion work, we need to allocate original multi-dimensional buffers.
Reverse Cache Read/Write might generate redundant loops when generated block was placed under existing loops.
This PR fix the two issues:
AllocSparseBuffer
instead ofAllocBuffer
for sparse tensors, actually to makeCompactBufferRegion
work, we need to allocate original multi-dimensional buffers.