Add native PauliRot implementation in LightningKokkos [sc-71642]

.github/CHANGELOG.md

@@ -33,6 +33,9 @@
 * Unify Lightning Kokkos device and Lightning Qubit device under a Lightning Base device
   [(#876)](https://github.com/PennyLaneAI/pennylane-lightning/pull/876)
 
+* LightningKokkos gains native support for the `PauliRot` gate.
+  [(#855)](https://github.com/PennyLaneAI/pennylane-lightning/pull/855)
+
 ### Documentation
 
 ### Bug fixes

pennylane_lightning/core/src/simulators/lightning_kokkos/StateVectorKokkos.hpp

@@ -359,6 +359,25 @@ class StateVectorKokkos final
         }
     }
 
+    /**
+     * @brief Apply a PauliRot gate to the state-vector.
+     *
+     * @param wires Wires to apply gate to.
+     * @param inverse Indicates whether to use inverse of gate.
+     * @param params Rotation angle.
+     * @param word A Pauli word (e.g. "XYYX").
+     */
+    void applyPauliRot(const std::vector<std::size_t> &wires,
+                       const bool inverse,
+                       const std::vector<PrecisionT> &params,
+                       const std::string &word) {
+        PL_ABORT_IF_NOT(wires.size() == word.size(),
+                        "wires and word have incompatible dimensions.");
+        Pennylane::LightningKokkos::Functors::applyPauliRot<KokkosExecSpace,
+                                                            PrecisionT>(
+            getView(), this->getNumQubits(), wires, inverse, params[0], word);
+    }
+
     template <bool inverse = false>
     void applyControlledGlobalPhase(const std::vector<ComplexT> &diagonal) {
         auto diagonal_ = vector2view(diagonal);

pennylane_lightning/core/src/simulators/lightning_kokkos/bindings/LKokkosBindings.hpp

@@ -62,6 +62,14 @@ void registerBackendClassSpecificBindings(PyClass &pyclass) {
 
     registerGatesForStateVector<StateVectorT>(pyclass);
 
+    pyclass.def(
+        "applyPauliRot",
+        [](StateVectorT &sv, const std::vector<std::size_t> &wires,
+           const bool inverse, const std::vector<ParamT> &params,
+           const std::string &word) {
+            sv.applyPauliRot(wires, inverse, params, word);
+        },
+        "Apply a Pauli rotation.");
     pyclass
         .def(py::init([](std::size_t num_qubits) {
             return new StateVectorT(num_qubits);

pennylane_lightning/core/src/simulators/lightning_kokkos/gates/BasicGateFunctors.hpp

@@ -19,12 +19,15 @@
 #include "BitUtil.hpp"
 #include "GateOperation.hpp"
 #include "Gates.hpp"
+#include "Util.hpp" // exp2, INVSQRT2
+#include "UtilKokkos.hpp"
 
 /// @cond DEV
 namespace {
 using namespace Pennylane::Util;
 using Kokkos::kokkos_swap;
 using Pennylane::Gates::GateOperation;
+using Pennylane::LightningKokkos::Util::vector2view;
 } // namespace
 /// @endcond
 
@@ -1092,6 +1095,59 @@ void applyMultiRZ(Kokkos::View<Kokkos::complex<PrecisionT> *> arr_,
         });
 }
 
+template <class ExecutionSpace, class PrecisionT>
+void applyPauliRot(Kokkos::View<Kokkos::complex<PrecisionT> *> arr_,
+                   const std::size_t num_qubits,
+                   const std::vector<std::size_t> &wires, const bool inverse,
+                   const PrecisionT angle, const std::string &word) {
+    using ComplexT = Kokkos::complex<PrecisionT>;
+    constexpr auto IMAG = Pennylane::Util::IMAG<PrecisionT>();
+    PL_ABORT_IF_NOT(wires.size() == word.size(),
+                    "wires and word have incompatible dimensions.")
+    if (std::find_if_not(word.begin(), word.end(),
+                         [](const int w) { return w == 'Z'; }) == word.end()) {
+        applyMultiRZ<ExecutionSpace>(arr_, num_qubits, wires, inverse,
+                                     std::vector<PrecisionT>{angle});
+        return;
+    }
+    const PrecisionT c = std::cos(angle / 2);
+    const ComplexT s = ((inverse) ? IMAG : -IMAG) * std::sin(angle / 2);
+    const std::vector<ComplexT> sines = {s, IMAG * s, -s, -IMAG * s};
+    auto d_sines = vector2view(sines);
+    auto get_mask =
+        [num_qubits, &wires](
+            [[maybe_unused]] const std::function<bool(const int)> &condition) {
+            std::size_t mask{0U};
+            for (std::size_t iw = 0; iw < wires.size(); iw++) {
+                const auto bit = static_cast<std::size_t>(condition(iw));
+                mask |= bit << (num_qubits - 1 - wires[iw]);
+            }
+            return mask;
+        };
+    const std::size_t mask_xy =
+        get_mask([&word](const int a) { return word[a] != 'Z'; });
+    const std::size_t mask_y =
+        get_mask([&word](const int a) { return word[a] == 'Y'; });
+    const std::size_t mask_z =
+        get_mask([&word](const int a) { return word[a] == 'Z'; });
+    const auto count_mask_y = std::popcount(mask_y);
+    Kokkos::parallel_for(
+        Kokkos::RangePolicy<ExecutionSpace>(0, exp2(num_qubits)),
+        KOKKOS_LAMBDA(const std::size_t i0) {
+            const std::size_t i1 = i0 ^ mask_xy;
+            if (i0 <= i1) {
+                const auto count_y = Kokkos::Impl::bit_count(i0 & mask_y) * 2;
+                const auto count_z = Kokkos::Impl::bit_count(i0 & mask_z) * 2;
+                const auto sign_i0 = count_z + count_mask_y * 3 - count_y;
+                const auto sign_i1 = count_z + count_mask_y + count_y;
+                const ComplexT v0 = arr_(i0);
+                const ComplexT v1 = arr_(i1);
+                arr_(i0) = c * v0 + d_sines(sign_i0 % 4) * v1;
+                arr_(i1) = c * v1 + d_sines(sign_i1 % 4) * v0;
+            }
+        });
+}
+
 template <class ExecutionSpace, class PrecisionT>
 void applyGlobalPhase(Kokkos::View<Kokkos::complex<PrecisionT> *> arr_,
                       const std::size_t num_qubits,

pennylane_lightning/core/src/simulators/lightning_kokkos/gates/tests/Test_StateVectorKokkos_Param.cpp

@@ -579,6 +579,22 @@ TEMPLATE_TEST_CASE("StateVectorKokkosManaged::applyIsingXX",
                 }
             }
         }
+        SECTION("PauliRotXX 0,1") {
+            for (std::size_t index = 0; index < angles.size(); index++) {
+                StateVectorKokkos<TestType> kokkos_sv{num_qubits};
+                kokkos_sv.applyPauliRot({0, 1}, adjoint, {angles[index]}, "XX");
+                std::vector<ComplexT> result_sv(kokkos_sv.getLength(), {0, 0});
+                kokkos_sv.DeviceToHost(result_sv.data(), kokkos_sv.getLength());
+
+                for (std::size_t j = 0; j < exp2(num_qubits); j++) {
+                    CHECK((real(result_sv[j])) ==
+                          Approx(real(expected_results[index][j])));
+                    CHECK((imag(result_sv[j])) ==
+                          Approx(((adjoint) ? -1.0 : 1.0) *
+                                 imag(expected_results[index][j])));
+                }
+            }
+        }
         SECTION("IsingXX 0,2") {
             for (std::size_t index = 0; index < angles.size(); index++) {
                 StateVectorKokkos<TestType> kokkos_sv{num_qubits};
@@ -645,6 +661,22 @@ TEMPLATE_TEST_CASE("StateVectorKokkosManaged::applyIsingYY",
                 }
             }
         }
+        SECTION("PauliRotYY 0,1") {
+            for (std::size_t index = 0; index < angles.size(); index++) {
+                StateVectorKokkos<TestType> kokkos_sv{num_qubits};
+                kokkos_sv.applyPauliRot({0, 1}, adjoint, {angles[index]}, "YY");
+                std::vector<ComplexT> result_sv(kokkos_sv.getLength(), {0, 0});
+                kokkos_sv.DeviceToHost(result_sv.data(), kokkos_sv.getLength());
+
+                for (std::size_t j = 0; j < exp2(num_qubits); j++) {
+                    CHECK((real(result_sv[j])) ==
+                          Approx(real(expected_results[index][j])));
+                    CHECK((imag(result_sv[j])) ==
+                          Approx(((adjoint) ? -1.0 : 1.0) *
+                                 imag(expected_results[index][j])));
+                }
+            }
+        }
         SECTION("IsingYY 0,2") {
             for (std::size_t index = 0; index < angles.size(); index++) {
                 StateVectorKokkos<TestType> kokkos_sv{num_qubits};
@@ -700,6 +732,22 @@ TEMPLATE_TEST_CASE("StateVectorKokkosManaged::applyIsingZZ",
                 }
             }
         }
+        SECTION("PauliRotZZ 0,1") {
+            for (std::size_t index = 0; index < angles.size(); index++) {
+                StateVectorKokkos<TestType> kokkos_sv{num_qubits};
+                kokkos_sv.applyPauliRot({0, 1}, adjoint, {angles[index]}, "ZZ");
+                std::vector<ComplexT> result_sv(kokkos_sv.getLength(), {0, 0});
+                kokkos_sv.DeviceToHost(result_sv.data(), kokkos_sv.getLength());
+
+                for (std::size_t j = 0; j < exp2(num_qubits); j++) {
+                    CHECK((real(result_sv[j])) ==
+                          Approx(real(expected_results[index][j])));
+                    CHECK((imag(result_sv[j])) ==
+                          Approx(((adjoint) ? -1.0 : 1.0) *
+                                 imag(expected_results[index][j])));
+                }
+            }
+        }
         SECTION("IsingZZ 0,2") {
             for (std::size_t index = 0; index < angles.size(); index++) {
                 StateVectorKokkos<TestType> kokkos_sv{num_qubits};

pennylane_lightning/core/src/simulators/lightning_qubit/gates/cpu_kernels/GateImplementationsLM.hpp

@@ -582,6 +582,12 @@ class GateImplementationsLM : public PauliGenerator<GateImplementationsLM> {
         constexpr auto IMAG = Pennylane::Util::IMAG<PrecisionT>();
         PL_ABORT_IF_NOT(wires.size() == word.size(),
                         "wires and word have incompatible dimensions.")
+        if (std::find_if_not(word.begin(), word.end(), [](const int w) {
+                return w == 'Z';
+            }) == word.end()) {
+            applyMultiRZ(arr, num_qubits, wires, inverse, angle);
+            return;
+        }
         const PrecisionT c = std::cos(angle / 2);
         const ComplexT s = ((inverse) ? IMAG : -IMAG) * std::sin(angle / 2);
         const std::array<ComplexT, 4> sines{s, IMAG * s, -s, -IMAG * s};
@@ -615,15 +621,11 @@ class GateImplementationsLM : public PauliGenerator<GateImplementationsLM> {
             const auto count_y = std::popcount(i0 & mask_y) * 2;
             const auto count_z = std::popcount(i0 & mask_z) * 2;
             const auto sign_i0 = count_z + count_mask_y * 3 - count_y;
-            if (mask_xy) [[likely]] {
-                const auto sign_i1 = count_z + count_mask_y + count_y;
-                const ComplexT v0 = arr[i0];
-                const ComplexT v1 = arr[i1];
-                arr[i0] = c * v0 + sines[sign_i0 % 4] * v1;
-                arr[i1] = c * v1 + sines[sign_i1 % 4] * v0;
-            } else [[unlikely]] {
-                arr[i0] *= c + sines[sign_i0 % 4];
-            }
+            const auto sign_i1 = count_z + count_mask_y + count_y;
+            const ComplexT v0 = arr[i0];
+            const ComplexT v1 = arr[i1];
+            arr[i0] = c * v0 + sines[sign_i0 % 4] * v1;
+            arr[i1] = c * v1 + sines[sign_i1 % 4] * v0;
         }
     }
 

pennylane_lightning/lightning_kokkos/_state_vector.py

@@ -268,6 +268,12 @@
                 self._apply_lightning_midmeasure(
                     operation, mid_measurements, postselect_mode=postselect_mode
                 )
+            elif isinstance(operation, qml.PauliRot):
+                method = getattr(state, "applyPauliRot")
+                paulis = operation._hyperparameters["pauli_word"]
+                wires = [i for i, w in zip(wires, paulis) if w != "I"]
+                word = "".join(p for p in paulis if p != "I")  # pylint: disable=protected-access
+                method(wires, invert_param, operation.parameters, word)
             elif method is not None:  # apply specialized gate
                 param = operation.parameters
                 method(wires, invert_param, param)

pennylane_lightning/lightning_kokkos/lightning_kokkos.py

@@ -18,6 +18,7 @@
 import os
 import sys
 from dataclasses import replace
+from functools import reduce
 from pathlib import Path
 from typing import Optional
 
@@ -154,7 +155,10 @@
         return len(op.wires) < 10
     if isinstance(op, qml.GroverOperator):
         return len(op.wires) < 13
-
+    if isinstance(op, qml.PauliRot):
+        word = op._hyperparameters["pauli_word"]  # pylint: disable=protected-access
+        # decomposes to IsingXX, etc. for n <= 2
+        return reduce(lambda x, y: x + (y != "I"), word, 0) > 2
     return op.name in _operations
 
 
@@ -210,7 +214,7 @@
 
 def _adjoint_ops(op: qml.operation.Operator) -> bool:
     """Specify whether or not an Operator is supported by adjoint differentiation."""
-    return adjoint_ops(op)
+    return not isinstance(op, qml.PauliRot) and adjoint_ops(op)
 
 
 def _add_adjoint_transforms(program: TransformProgram) -> None:

tests/test_gates.py

@@ -523,8 +523,8 @@ def circuit(x):
 
 
 @pytest.mark.skipif(
-    device_name != "lightning.qubit",
-    reason="PauliRot operations only implemented in lightning.qubit.",
+    device_name not in ("lightning.qubit", "lightning.kokkos"),
+    reason="PauliRot operations only implemented in lightning.qubit and lightning.kokkos.",
 )
 @pytest.mark.parametrize("n_wires", [1, 2, 3, 4, 5, 10, 15])
 @pytest.mark.parametrize("n_targets", [1, 2, 3, 4, 5, 10, 15])
@@ -540,8 +540,12 @@ def test_paulirot(n_wires, n_targets, tol):
     init_state /= np.linalg.norm(init_state)
     theta = 0.3
 
-    for _ in range(10):
-        word = "".join(pws[w] for w in np.random.randint(0, 3, n_targets))
+    for i in range(10):
+        word = (
+            "Z" * n_targets
+            if i == 0
+            else "".join(pws[w] for w in np.random.randint(0, 3, n_targets))
+        )
         wires = np.random.permutation(n_wires)[0:n_targets]
         stateprep = qml.StatePrep(init_state, wires=range(n_wires))
         op = qml.PauliRot(theta, word, wires=wires)