nextgenintelligence / pynvvm Goto Github PK

pynvvm - Python Bindings to libNVVM


Introduction
============

The pynvvm package is a Python binding for the libNVVM compiler library from
NVIDIA.  This library takes LLVM (NVVM) IR input and produces NVIDIA PTX
output suitable for execution on NVIDIA GPUs on any platform.  Please see the
CUDA 5.5 documentation for a complete description of libNVVM.


Installation
============

The pynvvm package does not have any external dependencies and can be
installed with ``pip`` or ``easy_install``.

    $ pip install pynvvm

Note, however, that the package does require the libNVVM binary to be present
at runtime. See below for instructions on how to set the search path.


Using pynvvm
============

There are two primary interfaces with pynvvm; a low-level interface which
provides users with direct access to the libNVVM API, and a high-level
interface which provides a Pythonic API for the compiler routines in libNVVM.


Low-Level Interface
-------------------

The low-level interface can be found in the ``pynvvm.interface`` module. An
instance of the interface can be obtained by calling the ``NVVMInterface``
constructor:

    from pynvvm.interface import NVVMInterface

    inter = NVVMInterface()

By default, the ``NVVMInterface`` object will attempt to load the libNVVM
shared library from ``LD_LIBRARY_PATH`` on Linux, ``DYLD_LIBRARY_PATH`` on
Mac, or ``PATH`` on Windows.  An optional parameter to the ``NVVMInterface``
constructor provides the absolute path to the libNVVM shared library and
overwrites the system search path.  For example, on Linux:

    from pynvvm.interface import NVVMInterface

    inter = NVVMInterface('/usr/local/cuda-5.5/nvvm/lib64/libnvvm.so')

**NOTE**: It is important that the specified binary match the architecture of
the Python interpreter under which your program is running.

Once an interface object is created, it provides access to all of the libNVVM
API functions as regular Python functions. However, instead of returning a
libNVVM status code, each function returns either a string (for output
functions) or None.  If an error occurs within libNVVM, an ``NVVMException``
exception is raised with the corresponding status code.

Note that the ``nvvmGetProgramLogSize`` and ``nvvmGetCompiledResultSize``
functions are *not* exposed.  Instead, the ``nvvmGetProgramLog`` and
``nvvmGetCompiledResult`` functions automatically determine the correct size
and return a UTF-8 encoded Python string.

Full Example:

    from pynvvm.interface import NVVMInterface, NVVMException

    module = ... ## Populate NVVM IR or bitcode

    inter = NVVMInterface()
    prog = inter.nvvmCreateProgram()
    try:
        inter.nvvmAddModuleToProgram(prog, module, 'mymodule')
        inter.nvvmCompileProgram(prog, ['-ftz=1'])
        ptx = inter.nvvmGetCompiledResult(prog)
    except NVVMException as e:
        print('Error: %s' % repr(e))



High-Level Interface
--------------------

For clients wanting a higher-level interface to libNVVM, the ``Program`` class
in ``pynvvm.compiler`` provides such an interface. The usage is similar to
that of the ``NVVMInterface`` class, but the API is more Pythonic and you do
not need to worry about maintaining NVVM objects.


    from pynvvm.compiler import Program, ProgramException

    module = ... ## Populate NVVM IR or bitcode

    try:
        prog = Program()
        prog.add_module(module, 'mymodule')
        ptx = prog.compile(['-ftz=1'])
    except ProgramException as e:
        print('Error: %s' % repr(e))

As with ``NVVMInterface``, the ``Program`` constructor accepts an optional
path to the libNVVM binary.

Please see ``samples/ptxgen.py`` for a complete example of an NVVM IR -> PTX
compiler using the higher-level interface.