Merge remote-tracking branch 'origin/main' into 29-add-swap

.github/workflows/ci_ubuntu.yml

@@ -50,7 +50,7 @@ jobs:
           verbose: 2
 
       - name: Install qulacs for Ubuntu
-        run: ./script/build_gcc.sh
+        run: CMAKE_BUILD_TYPE=Debug ./script/build_gcc.sh
 
       - name: Test in Ubuntu
         run: |
@@ -106,7 +106,5 @@ jobs:
 
       - name: Install qulacs for Ubuntu
         run: ./script/build_gcc.sh
-
-      - name: Test in Ubuntu
-        run: |
-          ninja test -C build -j $(nproc)
+
+      # GitHub Actions cannot run CUDA program

CMakeLists.txt

@@ -24,39 +24,6 @@ endif(NOT DEFINED QULACS_USE_EXE)
 message(STATUS "QULACS_USE_PYTHON = ${QULACS_USE_PYTHON}")
 message(STATUS "QULACS_USE_TEST = ${QULACS_USE_TEST}")
 
-### Compile Warnings ###
-if (${CMAKE_CXX_COMPILER_ID} STREQUAL "GNU")
-    set(WARNING_CPP "-Wall -Wdate-time -Wendif-labels -Werror=format=2 \
-        -Werror=missing-declarations -Werror=return-type -Wextra \
-        -Wfloat-equal -Wimplicit-fallthrough=5 -Wlogical-op \
-        -Wmissing-include-dirs -Wpointer-arith -Wredundant-decls \
-        -Wshadow -Wstrict-aliasing=2 -Wsuggest-attribute=noreturn -Wwrite-strings \
-        -fdiagnostics-color=auto -fstrict-aliasing")
-    # -Werror=undef is eliminated due to conflict with boost
-elseif ((${CMAKE_CXX_COMPILER_ID} STREQUAL "Clang") OR (${CMAKE_CXX_COMPILER_ID} STREQUAL "AppleClang"))
-    set(WARNING_CPP "-Wall -Wdate-time -Wendif-labels -Werror=format=2 \
-        -Werror=missing-declarations -Werror=return-type -Wextra \
-        -Wfloat-equal -Wimplicit-fallthrough \
-        -Wmissing-include-dirs -Wpointer-arith -Wredundant-decls \
-        -Wshadow -Wstrict-aliasing=2 -Wwrite-strings \
-        -fdiagnostics-color=auto -fstrict-aliasing")
-endif()
-
-### Compiler options ###
-if ((${CMAKE_CXX_COMPILER_ID} STREQUAL "GNU") OR (${CMAKE_CXX_COMPILER_ID} STREQUAL "Clang") OR (${CMAKE_CXX_COMPILER_ID} STREQUAL "AppleClang"))
-    set(CMAKE_CXX_STANDARD 20)
-    set(CMAKE_POSITION_INDEPENDENT_CODE ON)
-
-    # Enable pthread
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread")
-
-    # Enable openmp
-    if(QULACS_USE_OMP)
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fopenmp")
-        add_compile_definitions(OPENMP)
-    endif()
-endif()
-
 ### Kokkos options ###
 set(Kokkos_ENABLE_SERIAL ON CACHE BOOL "Enable Kokkos Serial backend")
 if(QULACS_USE_OMP)
@@ -133,6 +100,32 @@ else()
     message(STATUS "Skip downloding googletest")
 endif(QULACS_USE_TEST)
 
+add_library(qulacs)
+
+### Compiler options ###
+if ((${CMAKE_CXX_COMPILER_ID} STREQUAL "GNU") OR (${CMAKE_CXX_COMPILER_ID} STREQUAL "Clang") OR (${CMAKE_CXX_COMPILER_ID} STREQUAL "AppleClang"))
+    # Standard
+    target_compile_features(qulacs PUBLIC cxx_std_20)
+
+    # Warning
+    target_compile_options(qulacs PUBLIC
+        -Wall -Wextra -Wunused-parameter -Wshadow -pedantic
+        -Wsign-compare -Wtype-limits -Wuninitialized)
+
+    # Enable pthread
+    target_compile_options(qulacs PUBLIC -pthread)
+
+    # Enable openmp
+    if(QULACS_USE_OMP)
+        target_compile_options(qulacs PUBLIC -fopenmp)
+        target_compile_definitions(qulacs PUBLIC OPENMP)
+    endif()
+
+    # Debug options
+    target_compile_options(qulacs PUBLIC $<$<CONFIG:Debug>:-O0 -g -fsanitize=address,undefined>)
+    target_link_options(qulacs PUBLIC $<$<CONFIG:Debug>:-fsanitize=address,undefined>)
+endif()
+
 ### Add subdirectories ###
 add_subdirectory(qulacs)
 if(QULACS_USE_PYTHON)

exe/main.cpp

@@ -6,7 +6,7 @@ using namespace qulacs;
 
 void run() {
     auto state = StateVector::Haar_random_state(3);
-    for (int i = 0; i < state.dim(); i++) {
+    for (UINT i = 0; i < state.dim(); i++) {
         std::cout << state[i] << std::endl;
     }
 }

qulacs/CMakeLists.txt

@@ -1,7 +1,5 @@
 cmake_minimum_required(VERSION 3.21)
 
-add_library(qulacs)
-
 target_sources(qulacs PRIVATE
     gate/gate_npair_qubit.cpp
     gate/gate_one_control_one_target.cpp

qulacs/gate/constant.hpp

@@ -7,57 +7,67 @@
 
 namespace qulacs {
 //! PI value
-constexpr double PI = std::numbers::pi;
+KOKKOS_INLINE_FUNCTION
+double PI() { return 3.141592653589793; }
 
 //! square root of 2
-constexpr double SQRT2 = std::numbers::sqrt2;
+KOKKOS_INLINE_FUNCTION
+double SQRT2() { return 1.4142135623730950; }
 
 //! inverse square root of 2
-constexpr double INVERSE_SQRT2 = 0.707106781186547;
+KOKKOS_INLINE_FUNCTION
+double INVERSE_SQRT2() { return 0.707106781186547; }
 
 //! cosine pi/8
-constexpr double COSPI8 = 0.923879532511287;
+KOKKOS_INLINE_FUNCTION
+double COSPI8() { return 0.923879532511287; }
 
 //! sine pi/8
-constexpr double SINPI8 = 0.382683432365090;
+KOKKOS_INLINE_FUNCTION
+double SINPI8() { return 0.382683432365090; }
 
 //! identity matrix
-std::array<Complex, 4> I_GATE = {1, 0, 0, 1};
+matrix_2_2 I_GATE = {1, 0, 0, 1};
 //! Pauli matrix X
-std::array<Complex, 4> X_GATE = {0, 1, 1, 0};
+matrix_2_2 X_GATE = {0, 1, 1, 0};
 //! Pauli matrix Y
-std::array<Complex, 4> Y_GATE = {0, Complex(0, -1), Complex(0, 1), 0};
+matrix_2_2 Y_GATE = {0, Complex(0, -1), Complex(0, 1), 0};
 //! Pauli matrix Z
-std::array<Complex, 4> Z_GATE = {1, 0, 0, -1};
+matrix_2_2 Z_GATE = {1, 0, 0, -1};
+
 //! list of Pauli matrix I,X,Y,Z
-std::array<std::array<Complex, 4>, 4> PAULI_MATRIX = {I_GATE, X_GATE, Y_GATE, Z_GATE};
+// std::array<matrix_2_2, 4> PAULI_MATRIX = {I_GATE, X_GATE, Y_GATE, Z_GATE};
+
 //! S-gate
-std::array<Complex, 4> S_GATE_MATRIX = {1, 0, 0, Complex(0, 1)};
+matrix_2_2 S_GATE_MATRIX = {1, 0, 0, Complex(0, 1)};
 //! Sdag-gate
-std::array<Complex, 4> S_DAG_GATE_MATRIX = {1, 0, 0, Complex(0, -1)};
+matrix_2_2 S_DAG_GATE_MATRIX = {1, 0, 0, Complex(0, -1)};
 //! T-gate
-std::array<Complex, 4> T_GATE_MATRIX = {
-    COSPI8 - Complex(0, 1) * SINPI8, 0., 0., COSPI8 + Complex(0, 1) * SINPI8};
+matrix_2_2 T_GATE_MATRIX = {
+    COSPI8() - Complex(0, SINPI8()), 0., 0., COSPI8() + Complex(0, SINPI8()) * SINPI8()};
 //! Tdag-gate
-std::array<Complex, 4> T_DAG_GATE_MATRIX = {
-    COSPI8 + Complex(0, 1) * SINPI8, 0., 0., COSPI8 - Complex(0, 1) * SINPI8};
+matrix_2_2 T_DAG_GATE_MATRIX = {
+    COSPI8() + Complex(0, SINPI8()), 0., 0., COSPI8() - Complex(0, SINPI8())};
 //! Hadamard gate
-std::array<Complex, 4> HADAMARD_MATRIX = {1. / SQRT2, 1. / SQRT2, 1. / SQRT2, -1. / SQRT2};
+matrix_2_2 HADAMARD_MATRIX = {INVERSE_SQRT2(), INVERSE_SQRT2(), INVERSE_SQRT2(), -INVERSE_SQRT2()};
 //! square root of X gate
-std::array<Complex, 4> SQRT_X_GATE_MATRIX = {
+matrix_2_2 SQRT_X_GATE_MATRIX = {
     Complex(0.5, 0.5), Complex(0.5, -0.5), Complex(0.5, -0.5), Complex(0.5, 0.5)};
 //! square root of Y gate
-std::array<Complex, 4> SQRT_Y_GATE_MATRIX = {0.5 + 0.5i, -0.5 - 0.5i, 0.5 + 0.5i, 0.5 + 0.5i};
+matrix_2_2 SQRT_Y_GATE_MATRIX = {
+    Complex(0.5, 0.5), Complex(-0.5, -0.5), Complex(0.5, 0.5), Complex(0.5, 0.5)};
 //! square root dagger of X gate
-std::array<Complex, 4> SQRT_X_DAG_GATE_MATRIX = {0.5 - 0.5i, 0.5 + 0.5i, 0.5 + 0.5i, 0.5 - 0.5i};
+matrix_2_2 SQRT_X_DAG_GATE_MATRIX = {
+    Complex(0.5, -0.5), Complex(0.5, 0.5), Complex(0.5, 0.5), Complex(0.5, -0.5)};
 //! square root dagger of Y gate
-std::array<Complex, 4> SQRT_Y_DAG_GATE_MATRIX = {0.5 - 0.5i, 0.5 - 0.5i, -0.5 + 0.5i, 0.5 - 0.5i};
+matrix_2_2 SQRT_Y_DAG_GATE_MATRIX = {
+    Complex(0.5, -0.5), Complex(0.5, -0.5), Complex(-0.5, 0.5), Complex(0.5, -0.5)};
 //! Projection to 0
-std::array<Complex, 4> PROJ_0_MATRIX = {1, 0, 0, 0};
+matrix_2_2 PROJ_0_MATRIX = {1, 0, 0, 0};
 //! Projection to 1
-std::array<Complex, 4> PROJ_1_MATRIX = {0, 0, 0, 1};
+matrix_2_2 PROJ_1_MATRIX = {0, 0, 0, 1};
 //! complex values for exp(j * i*pi/4 )
-std::array<Complex, 4> PHASE_90ROT = {1., 1.i, -1, -1.i};
+matrix_2_2 PHASE_90ROT = {1., Complex(0, 1), -1, Complex(0, -1)};
 //! complex values for exp(-j * i*pi/4 )
-std::array<Complex, 4> PHASE_M90ROT = {1., -1.i, -1, 1.i};
+matrix_2_2 PHASE_M90ROT = {1., Complex(0, -1), -1, Complex(0, 1)};
 }  // namespace qulacs

qulacs/gate/gate_quantum_matrix.cpp

@@ -1,7 +1,5 @@
 #include "gate_quantum_matrix.hpp"
 
-#include "update_ops.hpp"
-
 namespace qulacs {
 void U1::update_quantum_state(StateVector& state_vector) const {
     u_gate(this->_target, this->_matrix, state_vector);
@@ -14,4 +12,4 @@ void U2::update_quantum_state(StateVector& state_vector) const {
 void U3::update_quantum_state(StateVector& state_vector) const {
     u_gate(this->_target, this->_matrix, state_vector);
 }
-}  // namespace qulacs
+}  // namespace qulacs

qulacs/gate/gate_quantum_matrix.hpp

@@ -10,7 +10,7 @@ namespace qulacs {
 class U1 : public QuantumGate {
     UINT _target;
     double _lambda;
-    std::array<Complex, 4> _matrix;
+    matrix_2_2 _matrix;
 
 public:
     U1(UINT target, double lambda) : _target(target), _lambda(lambda) {
@@ -20,20 +20,20 @@ class U1 : public QuantumGate {
 };
 class U2 : public QuantumGate {
     UINT _target;
-    double _lambda, _phi;
-    std::array<Complex, 4> _matrix;
+    double _phi, _lambda;
+    matrix_2_2 _matrix;
 
 public:
     U2(UINT target, double phi, double lambda) : _target(target), _phi(phi), _lambda(lambda) {
-        _matrix = get_IBMQ_matrix(PI / 2.0, phi, lambda);
+        _matrix = get_IBMQ_matrix(PI() / 2.0, phi, lambda);
     };
     void update_quantum_state(StateVector& state_vector) const override;
 };
 
 class U3 : public QuantumGate {
     UINT _target;
-    double _theta, _lambda, _phi;
-    std::array<Complex, 4> _matrix;
+    double _theta, _phi, _lambda;
+    matrix_2_2 _matrix;
 
 public:
     U3(UINT target, double theta, double phi, double lambda)

qulacs/gate/update_ops.hpp

@@ -45,13 +45,13 @@ void cnot_gate(UINT control_qubit_index, UINT target_qubit_index, StateVector& s
 
 void cz_gate(UINT control_qubit_index, UINT target_qubit_index, StateVector& state);
 
-std::array<Complex, 4> get_IBMQ_matrix(double _theta, double _phi, double _lambda);
+matrix_2_2 get_IBMQ_matrix(double _theta, double _phi, double _lambda);
 
 void single_qubit_dense_matrix_gate(UINT target_qubit_index,
-                                    std::array<Complex, 4> matrix,
+                                    const matrix_2_2& matrix,
                                     StateVector& state);
 
-void u_gate(UINT target_qubit_index, std::array<Complex, 4> matrix, StateVector& state);
+void u_gate(UINT target_qubit_index, const matrix_2_2& matrix, StateVector& state);
 
 void swap_gate(UINT target1, UINT target2, StateVector& state);
 

qulacs/gate/update_ops_dense_matrix.cpp

@@ -6,20 +6,20 @@
 
 namespace qulacs {
 void single_qubit_dense_matrix_gate(UINT target_qubit_index,
-                                    std::array<Complex, 4> matrix,
+                                    const matrix_2_2& matrix,
                                     StateVector& state) {
     const UINT mask = 1ULL << target_qubit_index;
     const UINT mask_low = mask - 1;
     const UINT mask_high = ~mask_low;
     auto amplitudes = state.amplitudes_raw();
     Kokkos::parallel_for(
-        state.dim() / 2, KOKKOS_LAMBDA(const UINT it) {
-            UINT basis_0 = (it & mask_low) + ((it & mask_high) << 1);
-            UINT basis_1 = basis_0 + mask;
+        state.dim() >> 1, KOKKOS_LAMBDA(const UINT it) {
+            UINT basis_0 = ((it & mask_high) << 1) | (it & mask_low);
+            UINT basis_1 = basis_0 | mask;
             Complex val0 = amplitudes[basis_0];
             Complex val1 = amplitudes[basis_1];
-            Complex res0 = matrix[0] * val0 + matrix[1] * val1;
-            Complex res1 = matrix[2] * val0 + matrix[3] * val1;
+            Complex res0 = matrix.val[0][0] * val0 + matrix.val[0][1] * val1;
+            Complex res1 = matrix.val[1][0] * val0 + matrix.val[1][1] * val1;
             amplitudes[basis_0] = res0;
             amplitudes[basis_1] = res1;
         });

qulacs/gate/update_ops_one_control_one_target.cpp

@@ -21,8 +21,8 @@ void cnot_gate(UINT control_qubit_index, UINT target_qubit_index, StateVector& s
     auto amplitudes = state.amplitudes_raw();
     Kokkos::parallel_for(
         1ULL << (n_qubits - 2), KOKKOS_LAMBDA(const UINT& it) {
-            UINT i = (it & low_mask) + ((it & mid_mask) << 1) +
-            ((it & high_mask) << 2) + control_mask;
+            UINT i =
+                (it & low_mask) | ((it & mid_mask) << 1) | ((it & high_mask) << 2) | control_mask;
             UINT j = i | target_mask;
             Kokkos::Experimental::swap(amplitudes[i], amplitudes[j]);
         });
@@ -39,14 +39,12 @@ void cz_gate(UINT control_qubit_index, UINT target_qubit_index, StateVector& sta
     const UINT low_mask = min_qubit_mask - 1;
     const UINT mid_mask = (max_qubit_mask - 1) ^ low_mask;
     const UINT high_mask = ~(max_qubit_mask - 1);
-    const UINT mask = control_mask + target_mask;
 
     auto amplitudes = state.amplitudes_raw();
     Kokkos::parallel_for(
         1ULL << (n_qubits - 2), KOKKOS_LAMBDA(const UINT& it) {
-            UINT i = (it & low_mask) + 
-                     ((it & mid_mask) << 1) +
-                     ((it & high_mask) << 2) + mask;
+            UINT i = (it & low_mask) | ((it & mid_mask) << 1) | ((it & high_mask) << 2) |
+                     control_mask | target_mask;
             amplitudes[i] *= -1;
         });
 }