qiskit 2.1.0rc1__cp39-abi3-manylinux2014_aarch64.manylinux_2_17_aarch64.whl

qiskit/synthesis/unitary/aqc/fast_gradient/fast_grad_utils.py

@@ -0,0 +1,237 @@
+# This code is part of Qiskit.
+#
+# (C) Copyright IBM 2022.
+#
+# This code is licensed under the Apache License, Version 2.0. You may
+# obtain a copy of this license in the LICENSE.txt file in the root directory
+# of this source tree or at http://www.apache.org/licenses/LICENSE-2.0.
+#
+# Any modifications or derivative works of this code must retain this
+# copyright notice, and modified files need to carry a notice indicating
+# that they have been altered from the originals.
+
+"""
+Utility functions in the fast gradient implementation.
+"""
+from __future__ import annotations
+from typing import Union
+import numpy as np
+
+
+def is_permutation(x: np.ndarray) -> bool:
+    """
+    Checks if array is really an index permutation.
+
+    Args:
+        1D-array of integers that supposedly represents a permutation.
+
+    Returns:
+        True, if array is really a permutation of indices.
+    """
+    return (
+        isinstance(x, np.ndarray)
+        and x.ndim == 1
+        and x.dtype == np.int64
+        and np.all(np.sort(x) == np.arange(x.size, dtype=np.int64))
+    )
+
+
+def reverse_bits(x: Union[int, np.ndarray], nbits: int, enable: bool) -> Union[int, np.ndarray]:
+    """
+    Reverses the bit order in a number of ``nbits`` length.
+    If ``x`` is an array, then operation is applied to every entry.
+
+    Args:
+        x: either a single integer or an array of integers.
+        nbits: number of meaningful bits in the number x.
+        enable: apply reverse operation, if enabled, otherwise leave unchanged.
+
+    Returns:
+        a number or array of numbers with reversed bits.
+    """
+
+    if not enable:
+        if isinstance(x, int):
+            pass
+        else:
+            x = x.copy()
+        return x
+
+    if isinstance(x, int):
+        res: int | np.ndarray = int(0)
+    else:
+        x = x.copy()
+        res = np.full_like(x, fill_value=0)
+
+    for _ in range(nbits):
+        res <<= 1
+        res |= x & 1
+        x >>= 1
+    return res
+
+
+def swap_bits(num: int, a: int, b: int) -> int:
+    """
+    Swaps the bits at positions 'a' and 'b' in the number 'num'.
+
+    Args:
+        num: an integer number where bits should be swapped.
+        a: index of the first bit to be swapped.
+        b: index of the second bit to be swapped.
+
+    Returns:
+        the number with swapped bits.
+    """
+    x = ((num >> a) ^ (num >> b)) & 1
+    return num ^ ((x << a) | (x << b))
+
+
+def bit_permutation_1q(n: int, k: int) -> np.ndarray:
+    """
+    Constructs index permutation that brings a circuit consisting of a single
+    1-qubit gate to "standard form": ``kron(I(2^n/2), G)``, as we call it. Here n
+    is the number of qubits, ``G`` is a 2x2 gate matrix, ``I(2^n/2)`` is the identity
+    matrix of size ``(2^n/2)x(2^n/2)``, and the full size of the circuit matrix is
+    ``(2^n)x(2^n)``. Circuit matrix in standard form becomes block-diagonal (with
+    sub-matrices ``G`` on the main diagonal). Multiplication of such a matrix and
+    a dense one is much faster than generic dense-dense product. Moreover,
+    we do not need to keep the entire circuit matrix in memory but just 2x2 ``G``
+    one. This saves a lot of memory when the number of qubits is large.
+
+    Args:
+        n: number of qubits.
+        k: index of qubit where single 1-qubit gate is applied.
+
+    Returns:
+        permutation that brings the whole layer to the standard form.
+    """
+    perm = np.arange(2**n, dtype=np.int64)
+    if k != n - 1:
+        for v in range(2**n):
+            perm[v] = swap_bits(v, k, n - 1)
+    return perm
+
+
+def bit_permutation_2q(n: int, j: int, k: int) -> np.ndarray:
+    """
+    Constructs index permutation that brings a circuit consisting of a single
+    2-qubit gate to "standard form": ``kron(I(2^n/4), G)``, as we call it. Here ``n``
+    is the number of qubits, ``G`` is a 4x4 gate matrix, ``I(2^n/4)`` is the identity
+    matrix of size ``(2^n/4)x(2^n/4)``, and the full size of the circuit matrix is
+    ``(2^n)x(2^n)``. Circuit matrix in standard form becomes block-diagonal (with
+    sub-matrices ``G`` on the main diagonal). Multiplication of such a matrix and
+    a dense one is much faster than generic dense-dense product. Moreover,
+    we do not need to keep the entire circuit matrix in memory but just 4x4 ``G``
+    one. This saves a lot of memory when the number of qubits is large.
+
+    Args:
+        n: number of qubits.
+        j: index of control qubit where single 2-qubit gate is applied.
+        k: index of target qubit where single 2-qubit gate is applied.
+
+    Returns:
+        permutation that brings the whole layer to the standard form.
+    """
+    dim = 2**n
+    perm = np.arange(dim, dtype=np.int64)
+    if j < n - 2:
+        if k < n - 2:
+            for v in range(dim):
+                perm[v] = swap_bits(swap_bits(v, j, n - 2), k, n - 1)
+        elif k == n - 2:
+            for v in range(dim):
+                perm[v] = swap_bits(swap_bits(v, n - 2, n - 1), j, n - 2)
+        else:
+            for v in range(dim):
+                perm[v] = swap_bits(v, j, n - 2)
+    elif j == n - 2:
+        if k < n - 2:
+            for v in range(dim):
+                perm[v] = swap_bits(v, k, n - 1)
+        else:
+            pass
+    else:
+        if k < n - 2:
+            for v in range(dim):
+                perm[v] = swap_bits(swap_bits(v, n - 2, n - 1), k, n - 1)
+        else:
+            for v in range(dim):
+                perm[v] = swap_bits(v, n - 2, n - 1)
+    return perm
+
+
+def inverse_permutation(perm: np.ndarray) -> np.ndarray:
+    """
+    Returns inverse permutation.
+
+    Args:
+        perm: permutation to be reversed.
+
+    Returns:
+        inverse permutation.
+    """
+    inv = np.zeros_like(perm)
+    inv[perm] = np.arange(perm.size, dtype=np.int64)
+    return inv
+
+
+def make_rx(phi: float, out: np.ndarray) -> np.ndarray:
+    """
+    Makes a 2x2 matrix that corresponds to X-rotation gate.
+    This is a fast implementation that does not allocate the output matrix.
+
+    Args:
+        phi: rotation angle.
+        out: placeholder for the result (2x2, complex-valued matrix).
+
+    Returns:
+        rotation gate, same object as referenced by "out".
+    """
+    a = 0.5 * phi
+    cs, sn = np.cos(a).item(), -1j * np.sin(a).item()
+    out[0, 0] = cs
+    out[0, 1] = sn
+    out[1, 0] = sn
+    out[1, 1] = cs
+    return out
+
+
+def make_ry(phi: float, out: np.ndarray) -> np.ndarray:
+    """
+    Makes a 2x2 matrix that corresponds to Y-rotation gate.
+    This is a fast implementation that does not allocate the output matrix.
+
+    Args:
+        phi: rotation angle.
+        out: placeholder for the result (2x2, complex-valued matrix).
+
+    Returns:
+        rotation gate, same object as referenced by "out".
+    """
+    a = 0.5 * phi
+    cs, sn = np.cos(a).item(), np.sin(a).item()
+    out[0, 0] = cs
+    out[0, 1] = -sn
+    out[1, 0] = sn
+    out[1, 1] = cs
+    return out
+
+
+def make_rz(phi: float, out: np.ndarray) -> np.ndarray:
+    """
+    Makes a 2x2 matrix that corresponds to Z-rotation gate.
+    This is a fast implementation that does not allocate the output matrix.
+
+    Args:
+        phi: rotation angle.
+        out: placeholder for the result (2x2, complex-valued matrix).
+
+    Returns:
+        rotation gate, same object as referenced by "out".
+    """
+    exp = np.exp(0.5j * phi).item()
+    out[0, 0] = 1.0 / exp
+    out[0, 1] = 0
+    out[1, 0] = 0
+    out[1, 1] = exp
+    return out

qiskit/synthesis/unitary/aqc/fast_gradient/fast_gradient.py

@@ -0,0 +1,226 @@
+# This code is part of Qiskit.
+#
+# (C) Copyright IBM 2022.
+#
+# This code is licensed under the Apache License, Version 2.0. You may
+# obtain a copy of this license in the LICENSE.txt file in the root directory
+# of this source tree or at http://www.apache.org/licenses/LICENSE-2.0.
+#
+# Any modifications or derivative works of this code must retain this
+# copyright notice, and modified files need to carry a notice indicating
+# that they have been altered from the originals.
+
+"""
+Implementation of the fast objective function class.
+"""
+
+import math
+import warnings
+import numpy as np
+
+from .layer import (
+    LayerBase,
+    Layer2Q,
+    Layer1Q,
+    init_layer2q_matrices,
+    init_layer2q_deriv_matrices,
+    init_layer1q_matrices,
+    init_layer1q_deriv_matrices,
+)
+from .pmatrix import PMatrix
+from ..cnot_unit_objective import CNOTUnitObjective
+
+
+class FastCNOTUnitObjective(CNOTUnitObjective):
+    """
+    Implementation of objective function and gradient calculator, which is
+    similar to
+    :class:`~qiskit.transpiler.aqc.DefaultCNOTUnitObjective`
+    but several times faster.
+    """
+
+    def __init__(self, num_qubits: int, cnots: np.ndarray):
+        super().__init__(num_qubits, cnots)
+
+        if not 2 <= num_qubits <= 16:
+            raise ValueError("expects number of qubits in the range [2..16]")
+
+        dim = 2**num_qubits
+        self._ucf_mat = PMatrix(num_qubits)  # U^dagger @ C @ F
+        self._fuc_mat = PMatrix(num_qubits)  # F @ U^dagger @ C
+        self._circ_thetas = np.zeros((self.num_thetas,))  # last thetas used
+
+        # array of C-layers:
+        self._c_layers = np.asarray([object()] * self.num_cnots, dtype=LayerBase)
+        # array of F-layers:
+        self._f_layers = np.asarray([object()] * num_qubits, dtype=LayerBase)
+        # 4x4 C-gate matrices:
+        self._c_gates = np.full((self.num_cnots, 4, 4), fill_value=0, dtype=np.complex128)
+        # derivatives of 4x4 C-gate matrices:
+        self._c_dervs = np.full((self.num_cnots, 4, 4, 4), fill_value=0, dtype=np.complex128)
+        # 4x4 F-gate matrices:
+        self._f_gates = np.full((num_qubits, 2, 2), fill_value=0, dtype=np.complex128)
+        # derivatives of 4x4 F-gate matrices:
+        self._f_dervs = np.full((num_qubits, 3, 2, 2), fill_value=0, dtype=np.complex128)
+        # temporary NxN matrices:
+        self._tmp1 = np.full((dim, dim), fill_value=0, dtype=np.complex128)
+        self._tmp2 = np.full((dim, dim), fill_value=0, dtype=np.complex128)
+
+        # Create layers of 2-qubit gates.
+        for q in range(self.num_cnots):
+            j = int(cnots[0, q])
+            k = int(cnots[1, q])
+            self._c_layers[q] = Layer2Q(num_qubits=num_qubits, j=j, k=k)
+
+        # Create layers of 1-qubit gates.
+        for k in range(num_qubits):
+            self._f_layers[k] = Layer1Q(num_qubits=num_qubits, k=k)
+
+    def objective(self, param_values: np.ndarray) -> float:
+        """
+        Computes the objective function and some intermediate data for
+        the subsequent gradient computation.
+        See description of the base class method.
+        """
+        depth, n = self.num_cnots, self._num_qubits
+
+        # Memorize the last angular parameters used to compute the objective.
+        if self._circ_thetas.size == 0:
+            self._circ_thetas = np.zeros((self.num_thetas,))
+        np.copyto(self._circ_thetas, param_values)
+
+        thetas4d = param_values[: 4 * depth].reshape(depth, 4)
+        thetas3n = param_values[4 * depth :].reshape(n, 3)
+
+        init_layer2q_matrices(thetas=thetas4d, dst=self._c_gates)
+        init_layer2q_deriv_matrices(thetas=thetas4d, dst=self._c_dervs)
+        init_layer1q_matrices(thetas=thetas3n, dst=self._f_gates)
+        init_layer1q_deriv_matrices(thetas=thetas3n, dst=self._f_dervs)
+
+        self._init_layers()
+        self._calc_ucf_fuc()
+        objective_value = self._calc_objective_function()
+        return objective_value
+
+    def gradient(self, param_values: np.ndarray) -> np.ndarray:
+        """
+        Computes the gradient of objective function.
+        See description of the base class method.
+        """
+
+        # If thetas are the same as used for objective value calculation
+        # before calling this function, then we re-use the computations,
+        # otherwise we have to re-compute the objective.
+        tol = math.sqrt(np.finfo(float).eps)
+        if not np.allclose(param_values, self._circ_thetas, atol=tol, rtol=tol):
+            self.objective(param_values)
+            warnings.warn("gradient is computed before the objective")
+
+        grad = np.full((param_values.size,), fill_value=0, dtype=np.float64)
+        grad4d = grad[: 4 * self.num_cnots].reshape(self.num_cnots, 4)
+        grad3n = grad[4 * self.num_cnots :].reshape(self._num_qubits, 3)
+        self._calc_gradient4d(grad4d)
+        self._calc_gradient3n(grad3n)
+        return grad
+
+    def _init_layers(self):
+        """
+        Initializes C-layers and F-layers by corresponding gate matrices.
+        """
+        c_gates = self._c_gates
+        c_layers = self._c_layers
+        for q in range(self.num_cnots):
+            c_layers[q].set_from_matrix(mat=c_gates[q])
+
+        f_gates = self._f_gates
+        f_layers = self._f_layers
+        for q in range(self._num_qubits):
+            f_layers[q].set_from_matrix(mat=f_gates[q])
+
+    def _calc_ucf_fuc(self):
+        """
+        Computes matrices ``ucf_mat`` and ``fuc_mat``. Both remain non-finalized.
+        """
+        ucf_mat = self._ucf_mat
+        fuc_mat = self._fuc_mat
+        tmp1 = self._tmp1
+        c_layers = self._c_layers
+        f_layers = self._f_layers
+        depth, n = self.num_cnots, self._num_qubits
+
+        # tmp1 = U^dagger.
+        np.conj(self.target_matrix.T, out=tmp1)
+
+        # ucf_mat = fuc_mat = U^dagger @ C = U^dagger @ C_{depth-1} @ ... @ C_{0}.
+        self._ucf_mat.set_matrix(tmp1)
+        for q in range(depth - 1, -1, -1):
+            ucf_mat.mul_right_q2(c_layers[q], temp_mat=tmp1, dagger=False)
+        fuc_mat.set_matrix(ucf_mat.finalize(temp_mat=tmp1))
+
+        # fuc_mat = F @ U^dagger @ C = F_{n-1} @ ... @ F_{0} @ U^dagger @ C.
+        for q in range(n):
+            fuc_mat.mul_left_q1(f_layers[q], temp_mat=tmp1)
+
+        # ucf_mat = U^dagger @ C @ F = U^dagger @ C @ F_{n-1} @ ... @ F_{0}.
+        for q in range(n - 1, -1, -1):
+            ucf_mat.mul_right_q1(f_layers[q], temp_mat=tmp1, dagger=False)
+
+    def _calc_objective_function(self) -> float:
+        """
+        Computes the value of objective function.
+        """
+        ucf = self._ucf_mat.finalize(temp_mat=self._tmp1)
+        trace_ucf = np.trace(ucf)
+        fobj = abs((2**self._num_qubits) - float(np.real(trace_ucf)))
+
+        return fobj
+
+    def _calc_gradient4d(self, grad4d: np.ndarray):
+        """
+        Calculates a part gradient contributed by 2-qubit gates.
+        """
+        fuc = self._fuc_mat
+        tmp1, tmp2 = self._tmp1, self._tmp2
+        c_gates = self._c_gates
+        c_dervs = self._c_dervs
+        c_layers = self._c_layers
+        for q in range(self.num_cnots):
+            # fuc[q] <-- C[q-1] @ fuc[q-1] @ C[q].conj.T. Note, c_layers[q] has
+            # been initialized in _init_layers(), however, c_layers[q-1] was
+            # reused on the previous step, see below, so we need to restore it.
+            if q > 0:
+                c_layers[q - 1].set_from_matrix(mat=c_gates[q - 1])
+                fuc.mul_left_q2(c_layers[q - 1], temp_mat=tmp1)
+            fuc.mul_right_q2(c_layers[q], temp_mat=tmp1, dagger=True)
+            fuc.finalize(temp_mat=tmp1)
+            # Compute gradient components. We reuse c_layers[q] several times.
+            for i in range(4):
+                c_layers[q].set_from_matrix(mat=c_dervs[q, i])
+                grad4d[q, i] = (-1) * np.real(
+                    fuc.product_q2(layer=c_layers[q], tmp1=tmp1, tmp2=tmp2)
+                )
+
+    def _calc_gradient3n(self, grad3n: np.ndarray):
+        """
+        Calculates a part gradient contributed by 1-qubit gates.
+        """
+        ucf = self._ucf_mat
+        tmp1, tmp2 = self._tmp1, self._tmp2
+        f_gates = self._f_gates
+        f_dervs = self._f_dervs
+        f_layers = self._f_layers
+        for q in range(self._num_qubits):
+            # ucf[q] <-- F[q-1] @ ucf[q-1] @ F[q].conj.T. Note, f_layers[q] has
+            # been initialized in _init_layers(), however, f_layers[q-1] was
+            # reused on the previous step, see below, so we need to restore it.
+            if q > 0:
+                f_layers[q - 1].set_from_matrix(mat=f_gates[q - 1])
+                ucf.mul_left_q1(f_layers[q - 1], temp_mat=tmp1)
+            ucf.mul_right_q1(f_layers[q], temp_mat=tmp1, dagger=True)
+            ucf.finalize(temp_mat=tmp1)
+            # Compute gradient components. We reuse f_layers[q] several times.
+            for i in range(3):
+                f_layers[q].set_from_matrix(mat=f_dervs[q, i])
+                grad3n[q, i] = (-1) * np.real(
+                    ucf.product_q1(layer=f_layers[q], tmp1=tmp1, tmp2=tmp2)
+                )