map-overlap-length-without-materializing-the-task-graph

map-overlap-length-without-materializing.ipynb

Rules in 1D

• 1 task per chunk from da.ones (this shouldn't be counted in the layer number of tasks, I don't think?
• If the depth is larger than the smallest chunksize, that screws up these calculations, but we won't worry about that for now
• first getitem call has 1 task per input chunk
• concatenate call - there are three concatenate calls, but if chunks are neigbouring they can share the edge ones
• overlap - there are three overlap calls, but if chunks are neigbouring they can share the edge ones
• There is one getitem call per chunk after the overlap part
• actual function passed to overlap - generates 1 task per input array chunk
• 2 tasks per chunk if trim=True (these are the block-info_trim and _trim tasks)
• next getitem calls - the

re are four getitem calls, but if chunks are neigbouring they can share the edge ones

Rules in 2D

• 1 task per chunk from da.ones (this shouldn't be counted in the layer number of tasks, I don't think?
• If the depth is larger than the smallest chunksize, that screws up these calculations, but we won't worry about that for now
• first getitem call has 1 task per input chunk
• ???concatenate call - there are three concatenate calls, but if chunks are neigbouring they can share the edge ones
• ???more getitems ??
• ???more concatenate ??
• ???overlap - there are three overlap calls, but if chunks are neigbouring they can share the edge ones
• There is one getitem call per chunk after the overlap part
• actual function passed to overlap - generates 1 task per input array chunk
• 2 tasks per chunk if trim=True (these

are the block-info_trim and _trim tasks)