# pybind11

pybind11 is a lightweight, header-only C++ library that enables seamless interoperability between C++ and Python. Its primary purpose is to create Python bindings for existing C++ code with minimal boilerplate, using compile-time type introspection via C++11 features such as variadic templates, lambda functions, and tuples. Unlike Boost.Python, which it was inspired by, pybind11 carries no external dependencies beyond the C++ standard library and Python itself (CPython 3.8+, PyPy3 7.3.17+, or GraalPy 24.1+), resulting in binaries that are typically 2–5x smaller and faster to compile.

The library supports a comprehensive range of C++ constructs: free functions and class methods with overloading, instance and static attributes, properties, enumerations, inheritance with automatic downcasting, custom exceptions, iterators, smart pointers (`std::unique_ptr`, `std::shared_ptr`, and `py::smart_holder`), STL container conversions, NumPy buffer protocol integration, vectorized array operations, and embedding of the Python interpreter inside C++ applications. All standard headers are prefixed `pybind11/` and the namespace alias `namespace py = pybind11` is used throughout this documentation.

---

## PYBIND11_MODULE — Define an extension module

The `PYBIND11_MODULE` macro is the entry point for every pybind11 extension. It defines the function that Python calls when importing the module. The optional `py::mod_gil_not_used()` tag signals that the module is safe to use without the GIL (for free-threaded CPython builds).

```cpp
// example.cpp
#include <pybind11/pybind11.h>
namespace py = pybind11;

int add(int i, int j) { return i + j; }

double multiply(double x, double y = 1.0) { return x * y; }

PYBIND11_MODULE(example, m, py::mod_gil_not_used()) {
    m.doc() = "A minimal pybind11 demonstration module";

    // Bind a free function with keyword args and default values
    m.def("add", &add, "Add two integers",
          py::arg("i"), py::arg("j"));

    m.def("multiply", &multiply, "Multiply two numbers",
          py::arg("x"), py::arg("y") = 1.0);

    // Export a constant attribute
    m.attr("PI") = 3.14159265358979;
}
```

```bash
# Compile (Linux/macOS)
c++ -O3 -Wall -shared -std=c++11 -fPIC \
    $(python3 -m pybind11 --includes) \
    example.cpp -o example$(python3 -m pybind11 --extension-suffix)
```

```python
import example

print(example.add(3, 4))          # 7
print(example.multiply(5.0))      # 5.0  (default y=1.0)
print(example.multiply(x=3, y=2)) # 6.0
print(example.PI)                  # 3.14159265358979
```

---

## py::class_ — Bind C++ classes and structs

`py::class_<T>` exposes a C++ type to Python, including constructors, methods, and attributes. Methods are registered with `.def()`, read-write fields with `.def_readwrite()`, and computed properties with `.def_property()`. In pybind11 v3 it is recommended to add `py::smart_holder` as a second template argument for safe smart-pointer interop.

```cpp
#include <pybind11/pybind11.h>
namespace py = pybind11;

class Pet {
public:
    Pet(const std::string &name, int age)
        : name(name), age(age) {}
    std::string repr() const {
        return "<Pet '" + name + "' age=" + std::to_string(age) + ">";
    }
    std::string name;
    int age;
private:
    std::string secret = "treats";
};

PYBIND11_MODULE(pets, m, py::mod_gil_not_used()) {
    py::class_<Pet, py::smart_holder>(m, "Pet")
        .def(py::init<const std::string &, int>(),
             py::arg("name"), py::arg("age"))
        // Direct field access
        .def_readwrite("name", &Pet::name)
        // Read-only computed property via getter/setter
        .def_property("age",
            [](const Pet &p) { return p.age; },
            [](Pet &p, int v) {
                if (v < 0) throw std::invalid_argument("age must be >= 0");
                p.age = v;
            })
        // Lambda bound as __repr__
        .def("__repr__", &Pet::repr)
        // Dynamic attribute support (adds __dict__)
        // .def(...) with py::dynamic_attr() in constructor instead:
        ;
}
```

```python
import pets

p = pets.Pet("Molly", 3)
print(repr(p))      # <Pet 'Molly' age=3>
p.name = "Charlie"
p.age = 5
print(p.age)        # 5
try:
    p.age = -1      # raises ValueError: age must be >= 0
except Exception as e:
    print(e)
```

---

## py::class_ with dynamic attributes

Adding `py::dynamic_attr()` to the `py::class_` constructor allows Python-side code to attach arbitrary attributes at runtime, exactly like a native Python class.

```cpp
#include <pybind11/pybind11.h>
namespace py = pybind11;

struct Widget {
    int value = 0;
};

PYBIND11_MODULE(widgets, m, py::mod_gil_not_used()) {
    py::class_<Widget>(m, "Widget", py::dynamic_attr())
        .def(py::init<>())
        .def_readwrite("value", &Widget::value);
}
```

```python
import widgets

w = widgets.Widget()
w.value = 42          # C++ field
w.label = "button"    # dynamic Python attribute
w.enabled = True
print(w.__dict__)     # {'label': 'button', 'enabled': True}
```

---

## py::init — Constructors and multiple overloads

`py::init<ArgTypes...>()` wraps a constructor. Multiple constructors are registered by chaining `.def(py::init<...>())` calls. Use `py::overload_cast` to disambiguate overloaded methods.

```cpp
#include <pybind11/pybind11.h>
#include <pybind11/stl.h>
namespace py = pybind11;

struct Point {
    double x, y;
    Point() : x(0), y(0) {}
    Point(double x, double y) : x(x), y(y) {}
    void set(double x_) { x = x_; }
    void set(double x_, double y_) { x = x_; y = y_; }
    double length() const { return std::sqrt(x*x + y*y); }
};

PYBIND11_MODULE(geometry, m, py::mod_gil_not_used()) {
    py::class_<Point>(m, "Point")
        .def(py::init<>())                                  // default ctor
        .def(py::init<double, double>(),                    // (x, y) ctor
             py::arg("x"), py::arg("y"))
        // Disambiguate overloaded set()
        .def("set", py::overload_cast<double>(&Point::set),
             "Set x only", py::arg("x"))
        .def("set", py::overload_cast<double, double>(&Point::set),
             "Set x and y", py::arg("x"), py::arg("y"))
        .def("length", &Point::length)
        .def("__repr__", [](const Point &p) {
            return "Point(" + std::to_string(p.x) + ", "
                            + std::to_string(p.y) + ")";
        });
}
```

```python
import geometry

p1 = geometry.Point()          # default: (0, 0)
p2 = geometry.Point(3.0, 4.0)
print(p2.length())             # 5.0
p2.set(0.0)                    # set x only
p2.set(6.0, 8.0)               # set x and y
print(p2.length())             # 10.0
```

---

## Inheritance and automatic downcasting

C++ inheritance hierarchies are mirrored in Python by passing the C++ base type as a template parameter to `py::class_`. For polymorphic types (those with at least one virtual function), pybind11 automatically downcasts base-class pointers to the actual derived type at runtime.

```cpp
#include <pybind11/pybind11.h>
namespace py = pybind11;

struct Animal {
    virtual ~Animal() = default;
    virtual std::string sound() const = 0;
    std::string name;
    Animal(const std::string &n) : name(n) {}
};

struct Dog : Animal {
    using Animal::Animal;
    std::string sound() const override { return "woof!"; }
    std::string fetch() const { return name + " fetches the ball!"; }
};

struct Cat : Animal {
    using Animal::Animal;
    std::string sound() const override { return "meow!"; }
};

std::unique_ptr<Animal> make_animal(bool dog) {
    if (dog) return std::make_unique<Dog>("Rex");
    return std::make_unique<Cat>("Whiskers");
}

PYBIND11_MODULE(animals, m, py::mod_gil_not_used()) {
    py::class_<Animal>(m, "Animal")
        .def("sound", &Animal::sound)
        .def_readwrite("name", &Animal::name);

    py::class_<Dog, Animal>(m, "Dog")
        .def(py::init<const std::string &>())
        .def("fetch", &Dog::fetch);

    py::class_<Cat, Animal>(m, "Cat")
        .def(py::init<const std::string &>());

    m.def("make_animal", &make_animal, py::arg("dog"));
}
```

```python
import animals

a = animals.make_animal(True)
print(type(a).__name__)   # Dog  (auto-downcast from Animal*)
print(a.sound())          # woof!
print(a.fetch())          # Rex fetches the ball!

b = animals.make_animal(False)
print(type(b).__name__)   # Cat
print(b.sound())          # meow!
```

---

## Trampoline classes — Override virtual functions in Python

To allow Python subclasses to override C++ virtual methods, a *trampoline* helper class is required. It uses `PYBIND11_OVERRIDE` / `PYBIND11_OVERRIDE_PURE` macros to forward virtual calls to the Python override. With `py::smart_holder`, trampolines must also inherit `py::trampoline_self_life_support`.

```cpp
#include <pybind11/pybind11.h>
namespace py = pybind11;

class Processor {
public:
    virtual ~Processor() = default;
    virtual std::string process(const std::string &input) = 0;
    virtual std::string name() const { return "Processor"; }
};

// Trampoline
class PyProcessor : public Processor,
                    public py::trampoline_self_life_support {
public:
    using Processor::Processor;

    std::string process(const std::string &input) override {
        PYBIND11_OVERRIDE_PURE(std::string, Processor, process, input);
    }
    std::string name() const override {
        PYBIND11_OVERRIDE(std::string, Processor, name);
    }
};

std::string run(Processor &p, const std::string &data) {
    return "[" + p.name() + "] " + p.process(data);
}

PYBIND11_MODULE(processors, m, py::mod_gil_not_used()) {
    py::class_<Processor, PyProcessor, py::smart_holder>(m, "Processor")
        .def(py::init<>())
        .def("process", &Processor::process)
        .def("name", &Processor::name);

    m.def("run", &run);
}
```

```python
import processors

class UpperProcessor(processors.Processor):
    def process(self, input: str) -> str:
        return input.upper()
    def name(self) -> str:
        return "UpperProcessor"

p = UpperProcessor()
print(processors.run(p, "hello world"))
# [UpperProcessor] HELLO WORLD
```

---

## py::native_enum — Bind C++ enums to Python native enums

`py::native_enum` (introduced in pybind11 v3) binds C++ enum types to Python's standard `enum.Enum` (or `enum.IntEnum`, etc.), producing PEP 435-compatible enumerations. Always call `.finalize()` after listing all values.

```cpp
#include <pybind11/pybind11.h>
#include <pybind11/native_enum.h>
namespace py = pybind11;

enum class Color { Red = 0, Green, Blue };
enum class Direction { North = 1, South, East, West };

struct Canvas {
    Color bg = Color::White;
    Canvas(Color c) : bg(c) {}
};

PYBIND11_MODULE(graphics, m, py::mod_gil_not_used()) {
    py::native_enum<Color>(m, "Color", "enum.IntEnum", "RGB color values")
        .value("Red",   Color::Red)
        .value("Green", Color::Green)
        .value("Blue",  Color::Blue)
        .export_values()
        .finalize();

    py::native_enum<Direction>(m, "Direction", "enum.Enum")
        .value("North", Direction::North)
        .value("South", Direction::South)
        .value("East",  Direction::East)
        .value("West",  Direction::West)
        .finalize();
}
```

```python
import graphics
from graphics import Color, Direction

print(Color.Red)           # Color.Red
print(Color.Red.value)     # 0
print(isinstance(Color.Green, int))  # True (IntEnum)

d = Direction.North
print(d.name)  # North

# Iterating
for c in Color:
    print(c.name, c.value)
# Red 0 / Green 1 / Blue 2
```

---

## Return value policies

Return value policies control Python object lifetime for values returned from C++ functions, especially raw pointers. The correct policy must be chosen explicitly to avoid crashes or memory leaks.

```cpp
#include <pybind11/pybind11.h>
namespace py = pybind11;

struct Config {
    std::string host = "localhost";
    int port = 8080;
};

static Config global_config;

struct Server {
    Config config;

    Server() : config{} {}

    // Returns pointer to internal config — use reference_internal
    Config *get_config() { return &config; }

    // Returns pointer to global — use reference (C++ manages lifetime)
    static Config *global() { return &global_config; }

    // Returns a new copy — default (move/copy) policy is fine
    Config make_config() const { return config; }
};

PYBIND11_MODULE(server, m, py::mod_gil_not_used()) {
    py::class_<Config>(m, "Config")
        .def(py::init<>())
        .def_readwrite("host", &Config::host)
        .def_readwrite("port", &Config::port);

    py::class_<Server>(m, "Server")
        .def(py::init<>())
        // reference_internal: keeps Server alive while Config ref is used
        .def("get_config", &Server::get_config,
             py::return_value_policy::reference_internal)
        // reference: global object, C++ owns it
        .def_static("global_config", &Server::global,
                    py::return_value_policy::reference)
        // copy: always safe, creates independent Python-owned copy
        .def("make_config", &Server::make_config,
             py::return_value_policy::copy);
}
```

```python
import server

s = server.Server()
cfg = s.get_config()      # reference_internal: s kept alive
cfg.host = "example.com"
print(s.get_config().host) # example.com

g = server.Server.global_config()
print(g.port)              # 8080

copy_cfg = s.make_config()
copy_cfg.port = 9090
print(s.get_config().port) # 8080 (original unchanged)
```

---

## py::keep_alive and py::call_guard — Call policies

`py::keep_alive<Nurse, Patient>` prevents the *nurse* object from being garbage-collected while the *patient* object is alive. `py::call_guard<T>` wraps a function call with an RAII scope guard, commonly used to release the GIL.

```cpp
#include <pybind11/pybind11.h>
namespace py = pybind11;

struct Item {
    std::string value;
    Item(std::string v) : value(std::move(v)) {}
};

struct Container {
    std::vector<Item *> items;
    void add(Item *item) { items.push_back(item); }
    size_t size() const { return items.size(); }
};

PYBIND11_MODULE(storage, m, py::mod_gil_not_used()) {
    py::class_<Item>(m, "Item")
        .def(py::init<std::string>());

    py::class_<Container>(m, "Container")
        .def(py::init<>())
        // keep_alive<1,2>: keep the Item (arg 2) alive as long as
        // the Container (arg 1 / `this`) is alive
        .def("add", &Container::add, py::keep_alive<1, 2>())
        .def("size", &Container::size);
}
```

```python
import storage

c = storage.Container()
c.add(storage.Item("apple"))
c.add(storage.Item("banana"))
print(c.size())  # 2
# Items won't be collected while c is alive
```

---

## *args and **kwargs — Variadic Python arguments

`py::args` (a subclass of `py::tuple`) and `py::kwargs` (a subclass of `py::dict`) allow C++ functions to accept arbitrary positional and keyword arguments mirroring Python's `*args` and `**kwargs`.

```cpp
#include <pybind11/pybind11.h>
namespace py = pybind11;

py::object variadic(py::args args, const py::kwargs &kwargs) {
    py::list result;
    for (auto item : args)
        result.append(py::str("arg: ") + py::str(item));
    for (auto kv : kwargs)
        result.append(py::str(kv.first) + py::str("=") + py::str(kv.second));
    return result;
}

// Keyword-only after *args
void with_kw_only(int a, py::args rest, int b) {
    py::print("a =", a, ", b =", b, ", rest =", rest);
}

PYBIND11_MODULE(variadics, m, py::mod_gil_not_used()) {
    m.def("variadic", &variadic);
    m.def("with_kw_only", &with_kw_only,
          py::arg("a"), py::arg("b"));
}
```

```python
import variadics

print(variadics.variadic(1, 2, "three", x=10, y=20))
# ['arg: 1', 'arg: 2', 'arg: three', 'x=10', 'y=20']

variadics.with_kw_only(1, 2, 3, b=99)
# a = 1 , b = 99 , rest = (2, 3)
```

---

## py::arg options — noconvert, none, kw_only, pos_only

`py::arg` objects carry metadata that refines how function arguments are processed: `.noconvert()` disables implicit conversion, `.none(bool)` controls `None`-to-nullptr mapping, `py::kw_only()` enforces keyword-only arguments, and `py::pos_only()` enforces positional-only arguments.

```cpp
#include <pybind11/pybind11.h>
namespace py = pybind11;

struct Dog { std::string name; };

std::string bark(Dog *dog) {
    if (!dog) return "(no dog)";
    return dog->name + " says woof!";
}

double strict_double(double v) { return v * 2; }

void mixed(int pos_only, int normal, int kw_only) {}

PYBIND11_MODULE(args_demo, m, py::mod_gil_not_used()) {
    py::class_<Dog>(m, "Dog")
        .def(py::init<>())
        .def_readwrite("name", &Dog::name);

    // .none(true) — allow Python None → nullptr
    m.def("bark", &bark, py::arg("dog").none(true));

    // .noconvert() — refuse implicit __float__ conversion
    m.def("strict_double", &strict_double,
          py::arg("v").noconvert());

    // pos_only / kw_only markers
    m.def("mixed", &mixed,
          py::arg("pos_only"), py::pos_only(),
          py::arg("normal"),
          py::kw_only(), py::arg("kw_only"));
}
```

```python
import args_demo

d = args_demo.Dog()
d.name = "Rex"
print(args_demo.bark(d))     # Rex says woof!
print(args_demo.bark(None))  # (no dog)

print(args_demo.strict_double(3.0))  # 6.0
# args_demo.strict_double(MyFloat(3))  # TypeError

args_demo.mixed(1, 2, kw_only=3)    # OK
# args_demo.mixed(pos_only=1, ...)   # TypeError
```

---

## Exception handling — C++ ↔ Python translation

pybind11 automatically translates standard C++ exceptions to Python exceptions. Custom C++ exception classes can be registered with `py::register_exception`. Python exceptions raised inside C++ code arrive as `py::error_already_set`.

```cpp
#include <pybind11/pybind11.h>
namespace py = pybind11;

// Custom C++ exception
class DatabaseError : public std::runtime_error {
public:
    using std::runtime_error::runtime_error;
};

std::string query(const std::string &sql) {
    if (sql.empty())
        throw std::invalid_argument("SQL query must not be empty");
    if (sql.find("DROP") != std::string::npos)
        throw DatabaseError("Destructive queries are forbidden");
    return "result_of: " + sql;
}

void call_python_callback(py::object cb) {
    try {
        cb(42);
    } catch (py::error_already_set &e) {
        // Inspect and re-raise or handle
        if (e.matches(PyExc_ValueError)) {
            py::print("Caught Python ValueError:", e.what());
        } else {
            throw;  // re-raise other Python exceptions
        }
    }
}

PYBIND11_MODULE(db, m, py::mod_gil_not_used()) {
    // Register custom C++ exception → Python class
    py::register_exception<DatabaseError>(m, "DatabaseError",
                                          PyExc_RuntimeError);

    m.def("query", &query);
    m.def("call_python_callback", &call_python_callback);
}
```

```python
import db

try:
    db.query("")
except ValueError as e:
    print("ValueError:", e)  # SQL query must not be empty

try:
    db.query("DROP TABLE users")
except db.DatabaseError as e:
    print("DatabaseError:", e)  # Destructive queries are forbidden

def bad_cb(n):
    raise ValueError("bad value!")

db.call_python_callback(bad_cb)  # Caught Python ValueError: bad value!
```

---

## STL containers — pybind11/stl.h

Including `pybind11/stl.h` enables automatic copying conversion between C++ `std::vector`, `std::list`, `std::set`, `std::map`, `std::optional`, `std::variant` and their Python equivalents. For pass-by-reference semantics, use `PYBIND11_MAKE_OPAQUE` or `py::bind_vector` / `py::bind_map` from `pybind11/stl_bind.h`.

```cpp
#include <pybind11/pybind11.h>
#include <pybind11/stl.h>
#include <pybind11/stl_bind.h>
namespace py = pybind11;

// Make this type opaque so Python sees it as a native-like object
PYBIND11_MAKE_OPAQUE(std::vector<int>)

std::vector<std::string> filter_long(const std::vector<std::string> &words,
                                     size_t min_len) {
    std::vector<std::string> out;
    for (auto &w : words)
        if (w.size() >= min_len) out.push_back(w);
    return out;
}

std::map<std::string, int> word_lengths(const std::vector<std::string> &words) {
    std::map<std::string, int> m;
    for (auto &w : words) m[w] = (int)w.size();
    return m;
}

PYBIND11_MODULE(stl_demo, m, py::mod_gil_not_used()) {
    // Auto-converted functions
    m.def("filter_long", &filter_long, py::arg("words"), py::arg("min_len"));
    m.def("word_lengths", &word_lengths, py::arg("words"));

    // Opaque std::vector<int> exposed as a Python object with full interface
    py::bind_vector<std::vector<int>>(m, "IntVector");
}
```

```python
import stl_demo

words = ["apple", "fig", "banana", "kiwi"]
print(stl_demo.filter_long(words, 5))
# ['apple', 'banana']

print(stl_demo.word_lengths(words))
# {'apple': 5, 'banana': 6, 'fig': 3, 'kiwi': 4}

# Opaque vector — mutations are visible in C++
v = stl_demo.IntVector()
v.append(10)
v.append(20)
print(len(v), list(v))  # 2 [10, 20]
```

---

## Smart pointers — py::smart_holder and std::shared_ptr

`py::smart_holder` (added in pybind11 v3) is the recommended holder type for `py::class_`. It supports bidirectional `std::unique_ptr` and `std::shared_ptr` transfers and is required when using trampoline classes. Use `py::classh<T>` as a shorthand for `py::class_<T, py::smart_holder>`.

```cpp
#include <pybind11/pybind11.h>
namespace py = pybind11;

class Resource {
public:
    Resource(std::string name) : name_(std::move(name)) {}
    const std::string &name() const { return name_; }
    ~Resource() { /* cleanup */ }
private:
    std::string name_;
};

std::unique_ptr<Resource> make_resource(const std::string &n) {
    return std::make_unique<Resource>(n);
}

void consume(std::unique_ptr<Resource> r) {
    // Takes ownership; Python object is disowned after this call
    (void)r;
}

void share(std::shared_ptr<Resource> r) {
    // Increments refcount; both Python and C++ share ownership
    (void)r;
}

PYBIND11_MODULE(resources, m, py::mod_gil_not_used()) {
    // py::classh is shorthand for py::class_<T, py::smart_holder>
    py::classh<Resource>(m, "Resource")
        .def(py::init<std::string>())
        .def("name", &Resource::name);

    m.def("make_resource", &make_resource);
    m.def("consume", &consume);   // disowns the Python object
    m.def("share", &share);       // shared ownership
}
```

```python
import resources

r = resources.make_resource("db_conn")
print(r.name())  # db_conn

resources.share(r)   # still accessible in Python after this
print(r.name())  # db_conn

resources.consume(r) # Python object is now disowned; don't use r again
```

---

## NumPy buffer protocol and py::array_t

Include `pybind11/numpy.h` to enable zero-copy exchange between C++ data structures and NumPy arrays. Expose a buffer view with `py::buffer_protocol()` and `def_buffer()`, or accept typed arrays with `py::array_t<T>`.

```cpp
#include <pybind11/pybind11.h>
#include <pybind11/numpy.h>
namespace py = pybind11;

class Matrix {
public:
    Matrix(size_t rows, size_t cols)
        : rows_(rows), cols_(cols), data_(rows * cols, 0.0f) {}
    float &operator()(size_t r, size_t c) { return data_[r * cols_ + c]; }
    float *data() { return data_.data(); }
    size_t rows() const { return rows_; }
    size_t cols() const { return cols_; }
private:
    size_t rows_, cols_;
    std::vector<float> data_;
};

// Compute sum of all elements in a typed NumPy array
double array_sum(py::array_t<double, py::array::c_style | py::array::forcecast> arr) {
    auto buf = arr.request();
    auto *ptr = static_cast<double *>(buf.ptr);
    double total = 0.0;
    for (py::ssize_t i = 0; i < buf.size; ++i) total += ptr[i];
    return total;
}

PYBIND11_MODULE(numpy_demo, m, py::mod_gil_not_used()) {
    py::class_<Matrix>(m, "Matrix", py::buffer_protocol())
        .def(py::init<size_t, size_t>())
        .def("__setitem__", [](Matrix &m, std::pair<size_t,size_t> idx, float v) {
            m(idx.first, idx.second) = v;
        })
        .def_buffer([](Matrix &m) -> py::buffer_info {
            return py::buffer_info(
                m.data(), sizeof(float),
                py::format_descriptor<float>::format(),
                2,
                { m.rows(), m.cols() },
                { sizeof(float) * m.cols(), sizeof(float) }
            );
        });

    m.def("array_sum", &array_sum, py::arg("array"));
}
```

```python
import numpy as np
import numpy_demo

mat = numpy_demo.Matrix(3, 3)
mat[0, 0] = 1.0
mat[1, 1] = 5.0

# Zero-copy conversion to NumPy
arr = np.array(mat, copy=False)
print(arr.shape, arr.dtype)   # (3, 3) float32
print(arr[0, 0], arr[1, 1])   # 1.0  5.0

data = np.array([[1.0, 2.0], [3.0, 4.0]])
print(numpy_demo.array_sum(data))  # 10.0
```

---

## Embedding the Python interpreter

pybind11 can embed a Python interpreter into a C++ application using `py::scoped_interpreter`. Link against `pybind11::embed` in CMake. Use `py::exec`, `py::eval`, and `py::module_::import` to interact with Python.

```cpp
// main.cpp
#include <pybind11/embed.h>
#include <iostream>
namespace py = pybind11;
using namespace py::literals;

int main() {
    py::scoped_interpreter guard{};  // start interpreter (RAII)

    // Execute a Python string
    py::exec(R"(
        import sys
        print("Python", sys.version.split()[0], "embedded!")
    )");

    // Call Python from C++ and retrieve results
    auto math = py::module_::import("math");
    double result = math.attr("sqrt")(py::float_(2.0)).cast<double>();
    std::cout << "sqrt(2) = " << result << "\n";  // 1.41421...

    // Build Python objects using C++ API
    auto kwargs = py::dict("name"_a = "World", "count"_a = 3);
    py::exec(R"(
        msg = "Hello, {name}! " * count
        print(msg.format(name=name))
    )", py::globals(), kwargs);

    // Import and run a local Python module
    py::module_::import("sys").attr("path").attr("insert")(0, ".");
    // If "mymodule.py" is present, it can be imported here:
    // auto mod = py::module_::import("mymodule");

    return 0;
}
```

```cmake
# CMakeLists.txt
cmake_minimum_required(VERSION 3.15)
project(embed_example)
find_package(pybind11 REQUIRED)
add_executable(embed_example main.cpp)
target_link_libraries(embed_example PRIVATE pybind11::embed)
```

```
$ ./embed_example
Python 3.11.x embedded!
sqrt(2) = 1.41421
Hello, World! Hello, World! Hello, World!
```

---

## Build systems — CMake, scikit-build-core, setuptools

pybind11 supports multiple build systems. The recommended approach for distributing packages on PyPI is `scikit-build-core`, which uses CMake under the hood. For standalone C++ projects, `find_package(pybind11)` + `pybind11_add_module` is the cleanest path.

```cmake
# CMakeLists.txt — standalone CMake build
cmake_minimum_required(VERSION 3.15...4.2)
project(myext LANGUAGES CXX)

set(PYBIND11_FINDPYTHON ON)
find_package(pybind11 CONFIG REQUIRED)

pybind11_add_module(myext myext.cpp)
install(TARGETS myext DESTINATION .)
```

```toml
# pyproject.toml — scikit-build-core (recommended for PyPI packages)
[build-system]
requires = ["scikit-build-core", "pybind11"]
build-backend = "scikit_build_core.build"

[project]
name = "myext"
version = "0.1.0"
requires-python = ">=3.8"
```

```meson
# meson.build — Meson alternative
project('myext', 'cpp', version: '0.1.0',
        default_options: ['cpp_std=c++11'])
py = import('python').find_installation(pure: false)
pybind11_dep = dependency('pybind11')
py.extension_module('myext', 'myext.cpp',
                    install: true, dependencies: [pybind11_dep])
```

```bash
# Build and install with pip (scikit-build-core)
pip install .

# Or compile manually (Linux/macOS)
c++ -O3 -Wall -shared -std=c++11 -fPIC \
    $(python3 -m pybind11 --includes) \
    myext.cpp -o myext$(python3 -m pybind11 --extension-suffix)
```

---

## Summary

pybind11 is the de-facto standard for creating Python bindings for C++ libraries. Its primary use cases span scientific computing (exposing numerical solvers, simulators, and matrix libraries like Eigen with zero-copy NumPy integration), systems programming (wrapping OS-level C++ APIs for Python tooling), and performance-critical Python extensions (replacing pure-Python hot paths with compiled C++ that Python code calls transparently). The library's header-only design means it integrates into any existing C++ build system without linking additional shared libraries, and its `py::smart_holder` mechanism ensures safe memory ownership semantics even for complex class hierarchies with virtual functions.

Integration patterns follow a consistent structure: define C++ classes and functions, create a `PYBIND11_MODULE` block, and register everything using the fluent `py::class_<T>` and `m.def()` APIs. For packaging, `scikit-build-core` + `CMakeLists.txt` + `pyproject.toml` is the modern recommended workflow for distributing wheels on PyPI. When embedding Python inside C++ applications, `py::scoped_interpreter` provides automatic lifecycle management, and the full pybind11 binding API remains available for bidirectional C++/Python data exchange. Exception handling, GIL management (`py::gil_scoped_release` / `py::call_guard`), and STL container bridging all follow the same declarative, compile-time-verified pattern that makes pybind11 both safe and ergonomic.