Controlled Bloq learns add_my_tensors
#815
Conversation
qualtran/_infra/controlled.py
Outdated
| for cv, ctrl_reg in zip(self.ctrl_spec.cvs, self.ctrl_regs): | ||
| for idx in ctrl_reg.all_idxs(): | ||
| active_idx[ctrl_idx] = int(cv[idx]) | ||
| active_idx[ctrl_idx + len(out_ind)] = int(cv[idx]) |
There was a problem hiding this comment.
what is going on here? what's the + len(out_ind)? should active_idx be two dimensional?
There was a problem hiding this comment.
We typically pass out_idx + inp_idx to the tn.add() call. The active_idx is a tuple of length len(out_idx.shape) + len(inp_idx.shape); which is equal to len(data.shape) and is used to find the subspace of data (an n-d array) that should be actived.
Thinking of CSWAP; if you want to express data[1, :, :, 1, :, :] = _swap_matrix(); active_idx will be a tuple corresponding to (1, :, :, 1, :, :) so you can do data[active_idx] = _swap_matrix()
qualtran/_infra/controlled.py
Outdated
| active_idx: List[Union[int, slice]] = [slice(x) for x in data.shape] | ||
| ctrl_idx = 0 | ||
| for cv, ctrl_reg in zip(self.ctrl_spec.cvs, self.ctrl_regs): |
There was a problem hiding this comment.
if I understand correctly, this is highly specific to the current concrete implementation of CtrlSpec. You can imagine a world where there is an abstract CtrlSpec base class and multiple implementations that can support things other than conjunction of a bunch of equals-a-value. In my head, cvs would be an implementation detail and CtrlSpec.is_active would be the interface.
There was a problem hiding this comment.
At that point, we could iterate over all 2 ** n possible bitstrings that cvs can take and call is_active to check if that subspace should be "Activated"
There was a problem hiding this comment.
Note SumOfProducts is a generalized representation that can be used to represent any arbitrary boolean function.
| def _tensor_shape_from_signature(signature: Signature) -> Tuple[int, ...]: | ||
| return tuple( | ||
| [ | ||
| *itertools.chain.from_iterable( | ||
| (2**reg.bitsize,) * int(np.prod(reg.shape)) | ||
| for reg in [*signature.rights(), *signature.lefts()] | ||
| ) | ||
| ] | ||
| ) | ||
|
|
||
|
|
||
| def _tensor_data_from_unitary_and_signature(inp_unitary: np.ndarray, signature: Signature): |
There was a problem hiding this comment.
move to a module in qualtran.simulation.tensor?
qualtran/_infra/controlled.py
Outdated
| ctrl_idx += 1 | ||
| # Put the subbloq tensor at indices where ctrl is active. | ||
| subbloq_shape = _tensor_shape_from_signature(self.subbloq.signature) | ||
| data[tuple(active_idx)] = self.subbloq.tensor_contract().reshape(subbloq_shape) |
There was a problem hiding this comment.
consider factoring out a lot of this into a function in qualtran.simulation.tensor. Perhaps: creating the data array; and then a function that uses it to mangle the indices
|
@mpharrigan I've addressed all comments, PTAL |
| """Returns the "active" subspace corresponding to `signature` and `ctrl_spec`. | ||
|
|
||
| Assumes first n-registers for `signature` are control registers corresponding to `ctrl_spec`. |
There was a problem hiding this comment.
say somewhere that this is a sequence of indices/slices that can address into a ndarray
Fixes #650
Currently based on top of #809