merge dev into master

.github/workflows/wheels.yml

@@ -10,7 +10,7 @@ on:
   pull_request_review:
     types: [submitted, edited]
   workflow_dispatch:
-  release:
+
 
 
 jobs:

.gitignore

@@ -13,4 +13,6 @@ cmake
 *.egg-info
 test
 .vscode
-thirdparty
+thirdparty
+.pyd
+MANIFEST.in

CMakeLists.txt

@@ -2,10 +2,11 @@
 
 cmake_minimum_required(VERSION 3.14...3.22)
 
-project(qfvm LANGUAGES CXX C)
+project(qfvm LANGUAGES CXX C) 
 
 set (CMAKE_BUILD_TYPE Release)
-
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CUDA_ARCHITECTURES 70;75;80;90)
 if(SKBUILD)
 
 execute_process(
@@ -42,9 +43,6 @@ ExternalProject_Add(Eigen3
 		PREFIX             ${EIGEN3_ROOT}
 		GIT_REPOSITORY     https://gitlab.com/libeigen/eigen.git
 		GIT_TAG            3.3.9
-		# CONFIGURE_COMMAND  cd ${EIGEN3_ROOT}/src/Eigen3 && cmake -B build -DCMAKE_INSTALL_PREFIX=${EIGEN3_ROOT}
-		# BUILD_COMMAND ""
-		# INSTALL_COMMAND    cd ${EIGEN3_ROOT}/src/Eigen3 && cmake --build build --target install
 
 		CONFIGURE_COMMAND ""
 		BUILD_COMMAND ""
@@ -72,6 +70,7 @@ if(CMAKE_HOST_SYSTEM_PROCESSOR STREQUAL "x86_64" OR CMAKE_HOST_SYSTEM_PROCESSOR
 	endif()
 endif()
 
+list (APPEND PRJ_INCLUDE_DIRS src/qfvm)
 pybind11_add_module(${PROJECT_NAME} MODULE src/${PROJECT_NAME}/${PROJECT_NAME}.cpp)
 add_dependencies(${PROJECT_NAME} Eigen3) #must add dependence for ninja
 target_compile_options(${PROJECT_NAME} PUBLIC ${PRJ_COMPILE_OPTIONS})
@@ -80,4 +79,49 @@ target_link_libraries(${PROJECT_NAME} PUBLIC ${PRJ_LIBRARIES})
 set_target_properties(${PROJECT_NAME} PROPERTIES SUFFIX ${PYTHON_MODULE_EXTENSION})
 target_compile_definitions(${PROJECT_NAME} PRIVATE VERSION_INFO=${PROJECT_VERSION})
 
-install(TARGETS ${PROJECT_NAME} DESTINATION .)
+#GPU version
+if (USE_GPU)
+	add_compile_definitions(_USE_GPU)
+	enable_language(CUDA)	
+	set_source_files_properties(src/${PROJECT_NAME}/${PROJECT_NAME}.cpp PROPERTIES LANGUAGE CUDA)
+	target_link_libraries(${PROJECT_NAME} PUBLIC cudart)
+	target_compile_options(${PROJECT_NAME} PUBLIC $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda> )
+	target_include_directories(${PROJECT_NAME} PUBLIC src/qfvm_gpu)
+	target_include_directories(${PROJECT_NAME} PUBLIC src/qfvm_gpu/cuda_utils)
+	target_include_directories(${PROJECT_NAME} PUBLIC ${CUDA_INCLUDE_DIRS})
+	message("cuda_include" ${CUDA_INCLUDE_DIRS})
+	if (USE_CUQUANTUM)
+		add_compile_definitions(_USE_CUQUANTUM)
+		function(set_with_fallback VARIABLE FALLBACK)
+			if (NOT DEFINED ${VARIABLE} OR ${VARIABLE} STREQUAL "")
+			set(${VARIABLE} $ENV{${VARIABLE}} CACHE INTERNAL ${VARIABLE})
+			if (${VARIABLE} STREQUAL "")
+				if (NOT ${FALLBACK} STREQUAL "")
+				set(${VARIABLE} $ENV{${FALLBACK}} CACHE INTERNAL ${VARIABLE})
+				endif ()
+			endif ()
+			endif ()
+		endfunction()
+
+		set_with_fallback(CUSTATEVEC_ROOT CUQUANTUM_ROOT)
+
+		if (CUSTATEVEC_ROOT STREQUAL "")
+			message(FATAL_ERROR "Please set the environment variables CUSTATEVEC_ROOT or CUQUANTUM_ROOT to the path of the cuQuantum installation.")
+		endif ()
+
+		message(STATUS "Using CUSTATEVEC_ROOT = ${CUSTATEVEC_ROOT}")
+
+		set(CMAKE_CUDA_FLAGS_ARCH_SM70 "-gencode arch=compute_70,code=sm_70")
+		set(CMAKE_CUDA_FLAGS_ARCH_SM75 "-gencode arch=compute_75,code=sm_75")
+		set(CMAKE_CUDA_FLAGS_ARCH_SM80 "-gencode arch=compute_80,code=sm_80 -gencode arch=compute_80,code=compute_80")
+		set(CMAKE_CUDA_FLAGS_ARCH_SM90 "-gencode arch=compute_90,code=sm_90 -gencode arch=compute_90,code=compute_90")
+		set(CMAKE_CUDA_FLAGS_ARCH "${CMAKE_CUDA_FLAGS_ARCH_SM70} ${CMAKE_CUDA_FLAGS_ARCH_SM75} ${CMAKE_CUDA_FLAGS_ARCH_SM80} ${CMAKE_CUDA_FLAGS_ARCH_SM90}")
+		set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} ${CMAKE_CUDA_FLAGS_ARCH}")
+
+		target_include_directories(${PROJECT_NAME} PUBLIC ${CUDA_INCLUDE_DIRS} ${CUSTATEVEC_ROOT}/include)
+		target_link_directories(${PROJECT_NAME} PUBLIC ${CUSTATEVEC_ROOT}/lib ${CUSTATEVEC_ROOT}/lib64)
+		target_link_libraries(${PROJECT_NAME} PUBLIC -lcustatevec_static -lcublas )
+	endif()
+endif()
+
+install(TARGETS ${PROJECT_NAME} DESTINATION .)

README.md

@@ -22,6 +22,18 @@ pip install -r requirements.txt
 python setup.py install
 ```
 
+## GPU support
+To install PyQuafu with GPU-based circuit simulator, you need build from the source and make sure that [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads) is installed. You can run
+
+```
+python setup.py install -DUSE_GPU=ON
+```
+to install the GPU version. If you further have [cuQuantum](https://developer.nvidia.com/cuquantum-sdk) installed, you can install PyQuafu with cuQuantum support.
+```
+python setup.py install -DUSE_GPU=ON -DUSE_CUQUANTUM=ON
+```
+
+
 ## Document
 Please see the website [docs](https://scq-cloud.github.io/).
 

requirements.txt

@@ -6,4 +6,6 @@ scipy>=1.8.1
 setuptools>=58.0.4
 sparse>=0.13.0
 scikit-build>=0.16.1
-pybind11>=2.10.3
+pybind11>=2.10.3
+ply~=3.11
+Pillow~=10.0.0

setup.py

@@ -28,7 +28,7 @@
 
 setup(
     name="pyquafu",
-    version="0.2.11",
+    version="0.3.0",
     author="ssli",
     author_email="[email protected]",
     url="https://github.com/ScQ-Cloud/pyquafu",
@@ -45,4 +45,4 @@
     zip_safe=False,
     setup_cfg=True,
     license="Apache-2.0 License"
-)
+)

src/qfvm/circuit.hpp

@@ -89,7 +89,7 @@ class Circuit{
     private:
         uint qubit_num_;    
         vector<QuantumOperator> gates_{0};
-
+        uint max_targe_num_;
     public:
     Circuit();
     explicit Circuit(uint qubit_num);
@@ -99,6 +99,7 @@ class Circuit{
     void add_gate(QuantumOperator &gate);
     void compress_gates();
     uint qubit_num() const { return qubit_num_; }
+    uint max_targe_num() const {return max_targe_num_;}
     vector<QuantumOperator>gates() const { return gates_; }
 
 };
@@ -121,16 +122,21 @@ void Circuit::add_gate(QuantumOperator &gate){
 
  Circuit::Circuit(vector<QuantumOperator> &gates)
  :
- gates_(gates){
+ gates_(gates),
+ max_targe_num_(0){
     qubit_num_ = 0;
     for (auto gate : gates){
         for (pos_t pos : gate.positions()){
+            if (gate.targe_num() > max_targe_num_)
+                max_targe_num_ = gate.targe_num();
             if (pos+1 > qubit_num_){ qubit_num_ = pos+1; }
         }
     }
 }
 
 Circuit::Circuit(py::object const&pycircuit)
+:
+max_targe_num_(0)
 {
     auto pygates = pycircuit.attr("gates");
     auto used_qubits = pycircuit.attr("used_qubits").cast<vector<pos_t>>();
@@ -139,6 +145,8 @@ Circuit::Circuit(py::object const&pycircuit)
         py::object pygate = py::reinterpret_borrow<py::object>(pygate_h);
         QuantumOperator gate = from_pyops(pygate);
         if (gate){
+            if (gate.targe_num() > max_targe_num_)
+                max_targe_num_ = gate.targe_num();
             gates_.push_back(std::move(gate));
         }        
     }

src/qfvm/qfvm.cpp

@@ -2,42 +2,111 @@
 #include <pybind11/numpy.h>
 #include "simulator.hpp"
 
+#ifdef _USE_GPU
+#include <cuda_simulator.cuh>
+#endif
+
+#ifdef _USE_CUQUANTUM
+#include <custate_simu.cuh>
+#endif
+
 namespace py = pybind11;
 
 template <typename T>
-py::array_t<T> to_numpy(std::vector<T> &&src) {
-  vector<T>* src_ptr = new std::vector<T>(std::move(src));
-  auto capsule = py::capsule(src_ptr, [](void* p) { delete reinterpret_cast<std::vector<T>*>(p); });
-  return py::array_t<T>(
-    src_ptr->size(),  // shape of array
-    src_ptr->data(),  // c-style contiguous strides for vector
-    capsule           // numpy array references this parent
-  );
-}
+py::array_t<T> to_numpy(const std::tuple<T*, size_t> &src) {
+    auto src_ptr = std::get<0>(src);
+    auto src_size = std::get<1>(src);
 
+    auto capsule = py::capsule(src_ptr, [](void* p) {
+        delete [] reinterpret_cast<T*>(p);
+    });
+    return py::array_t<T>(
+        src_size,
+        src_ptr,
+        capsule
+    );
+}
 
 py::object execute(string qasm){
-    return to_numpy(simulate(qasm).move_data());
+    return to_numpy(simulate(qasm).move_data_to_python());
 }
 
-py::object simulate_circuit(py::object const&pycircuit, vector<complex<double>> const&inputstate){
-  auto circuit = Circuit(pycircuit);
-    if (inputstate.size() == 0){
+py::object simulate_circuit(py::object const&pycircuit, py::array_t<complex<double>> &np_inputstate){
+    auto circuit = Circuit(pycircuit);
+    py::buffer_info buf = np_inputstate.request();
+    auto* data_ptr = reinterpret_cast<std::complex<double>*>(buf.ptr);
+    size_t data_size = buf.size;
+
+    if (data_size == 0){
         StateVector<double> state;
         simulate(circuit, state);
-        return to_numpy(state.move_data());
+        return to_numpy(state.move_data_to_python());
     }
     else{
-      StateVector<double> state{inputstate};
+      StateVector<double> state(data_ptr, buf.size);
       simulate(circuit, state);
-      return to_numpy(state.move_data());
+      state.move_data_to_python();
+      return np_inputstate;
+    }
+}
+
+#ifdef _USE_GPU
+py::object simulate_circuit_gpu(py::object const&pycircuit, py::array_t<complex<double>> &np_inputstate){
+    auto circuit = Circuit(pycircuit);
+    py::buffer_info buf = np_inputstate.request();
+    auto* data_ptr = reinterpret_cast<std::complex<double>*>(buf.ptr);
+    size_t data_size = buf.size;
+
+
+    if (data_size == 0){
+        StateVector<double> state;
+        simulate_gpu(circuit, state);
+        return to_numpy(state.move_data_to_python());
+    }
+    else{
+      StateVector<double> state(data_ptr, buf.size);
+      simulate_gpu(circuit, state);
+      state.move_data_to_python();
+      return np_inputstate;
+    }
+}
+#endif
+
+#ifdef _USE_CUQUANTUM
+py::object simulate_circuit_custate(py::object const&pycircuit, py::array_t<complex<double>> &np_inputstate){
+    auto circuit = Circuit(pycircuit);
+    py::buffer_info buf = np_inputstate.request();
+    auto* data_ptr = reinterpret_cast<std::complex<double>*>(buf.ptr);
+    size_t data_size = buf.size;
+
+
+    if (data_size == 0){
+        StateVector<double> state;
+        simulate_custate(circuit, state);
+        return to_numpy(state.move_data_to_python());
+    }
+    else{
+      StateVector<double> state(data_ptr, buf.size);
+      simulate_custate(circuit, state);
+      state.move_data_to_python();
+      return np_inputstate;
     }
 }
+#endif
+
 
 
 PYBIND11_MODULE(qfvm, m) {
     m.doc() = "Qfvm simulator";
     m.def("execute", &execute, "Simulate with qasm");
     m.def("simulate_circuit", &simulate_circuit, "Simulate with circuit", py::arg("circuit"), py::arg("inputstate")= py::array_t<complex<double>>(0));
+
+    #ifdef _USE_GPU
+     m.def("simulate_circuit_gpu", &simulate_circuit_gpu, "Simulate with circuit", py::arg("circuit"), py::arg("inputstate")= py::array_t<complex<double>>(0));
+    #endif
+
+    #ifdef _USE_CUQUANTUM
+    m.def("simulate_circuit_custate", &simulate_circuit_custate, "Simulate with circuit", py::arg("circuit"), py::arg("inputstate")= py::array_t<complex<double>>(0));
+    #endif
 }
 

src/qfvm/simulator.hpp

@@ -66,11 +66,11 @@ void simulate(Circuit const& circuit, StateVector<data_t> & state){
                 state.apply_rz(op.positions()[1], op.paras()[0]);
                 state.apply_cnot(op.positions()[0], op.positions()[1]);
                 break;
-    
+
 
             //Other general gate
-            default: 
-            {   
+            default:
+            {
                 if (op.targe_num() == 1){
                     auto mat_temp = op.mat();
                     complex<double> *mat = mat_temp.data();
@@ -176,7 +176,7 @@ void simulate(string qasm, StateVector<double> & state){
 StateVector<double> simulate(string qasm){
     StateVector<double>state;
     simulate(qasm, state);
-    return std::move(state); 
+    return std::move(state);
 }