aldanor / ipybind Goto Github PK

• %%pybind11
  • Enabling the extension
  • Basic usage example
  • Caching and recompilation
  • Error reporting and verbosity
  • Setting C++ standard
  • Compiler and linker flags
  • Include and library directories
  • Build output verbosity
• Notebook integration
• Compatibility

ipybind is an IPython extension that allows building and importing pybind11 modules in IPython environment, such as IPython console or Jupyter notebooks running a Python kernel.

To enable ipybind extension in the current kernel:

%load_ext ipybind

In all examples that follow we assume that the extension has been previously loaded.

%%pybind11

PYBIND11_PLUGIN(example) {
    py::module m("example");
    m.def("add", [](int x, int y) { return x + y; });
    return m.ptr();
}

This will build the extension module and import all symbols from it into the current namespace:

>>> add(1, 2)
3

Each compiled module is assigned a hash based on the code contents, arguments to %%pybind magic and Python interpreter version. Sources for compiled modules and the binaries are stored under $IPYTHONDIR/pybind11 (on Linux / macOS it's ~/.ipython/pybind11 by default). If a compiled module binary with matching hash is found, it is not rebuilt and is instead imported directly.

It is also possible to force recompilation by assigning a new unique hash (this is useful, for instance, in cases when module's code depends on 3rd-party code that may change) – this can be done by passing -f flag:

%pybind11 -f

All compiler output is captured and shown in the IPython environment (as opposed to the standard error of the shell where the kernel was launched). By default, compiler output is shown only if compilation fails; to always show all compiler output (like warnings), use -v flag:

%pybind11 -v

The following example should compile silently if built via %%pybind11 with no arguments:

%%pybind11 -vf;

PYBIND11_PLUGIN(example) {
    int x;                    // cell line 4
    py::module m("example");
    return m.ptr();
}

Building it with -v flag will display compiler warning:

--------------------------------------------------------------------------------
<source>:4:9: warning: unused variable 'x' [-Wunused-variable]
    int x;
        ^
1 warning generated.
--------------------------------------------------------------------------------

As can be seen in the above example, line numbers in reported error and warning messages should match line numbers in the input cell, including the cell magic line itself. (Line numbers can be shown in Jupyter notebooks by pressing L in command mode).

If C++ standard is not specified, it defaults to C++14. If it's not supported by the compiler, it falls back to C++11. It is also possible to specify the standard manually by passing -std option; for example:

%%pybind11 -std=c++17

Additional compiler and linker flags can be passed via -c and -Wl options respectively. Flags containing spaces or starting with a dash should be passed using -c="..." syntax; double quotes can be escaped via \". Both of these options can be specified multiple times.

%%pybind11 -c="-Wextra -fno-inline" -c="-Os"

On Linux and macOS, extensions are compiled with -flto and -fvisibility=hidden provided those flags are supported by the compiler. On Windows, extensions are built with /MP /bigobj /EHsc. The rest of the flags are provided by distutils.

Include and library directories can be specified via -I and -L options. Both of these options can be passed multiple times:

%%pybind11 -v -I /foo/bar -I="/foo bar/baz" -L/baz

Note: in conda environments, $PREFIX/include is always added to include paths and $PREFIX/lib is added to library paths (on Windows, it's $PREFIX/Library/include and $PREFIX/Library/lib).

Starting a cell with %%pybind11 changes its syntax highlighting to C++ and enables additional visual styles (so that C/C++ keywords are highlighted).

Since the whole cell starting with %%pybind11 is considered as C++ by syntax highlighter, this also includes the cell magic itself (the very first line). If it's not terminated with a semicolon, the indentation on the following lines is likely to be off. For this reason, a trailing semilocon is always ignored, such as in this example:

%%pybind11 -f -v;

Python 2 support was dropped in IPython 6.x release entirely and the clock is ticking so it makes little sense for us to support it here either.

Starting from Python 3.5, the official Windows distribution switched from MSVC 10.0 to MSVC 14.0 as a default build toolchain. The former is too ancient to work with pybind11, so we require Python 3.5 or newer.