cuda4py

Python cffi CUDA bindings and helper classes.

Tested with Python 2.7, Python 3.6 and PyPy on Linux with CUDA 8.0.

To compile kernel code written in C++, nvcc should be in PATH and exported functions should be marked as extern "C" (for Windows, cl.exe should be in PATH also). Functions in plain PTX can be used without nvcc.

To use CUBLAS, libcublas.so (cublas64_65.dll) should be present. To use CUDNN, libcudnn.so (cudnn64_65.dll) should be present. To use CUFFT, libcufft.so (cufft64_65.dll) should be present. To use CURAND, libcurand.so (curand64_65.dll) should be present.

Not all CUDA api is currently covered.

To install the module run:

python setup.py install

or just copy src/cuda4py to any place where python interpreter will be able to find it.

To run the tests, execute:

for Python 2.7:

PYTHONPATH=src nosetests -w tests

for Python 3.4:

PYTHONPATH=src nosetests3 -w tests

for PyPy:

PYTHONPATH=src pypy -m nose -w tests

Example usage:

import cuda4py as cu
import logging
import numpy


if __name__ == "__main__":
    logging.basicConfig(level=logging.DEBUG)
    ctx = cu.Devices().create_some_context()
    module = cu.Module(
        ctx, source=
        """
        extern "C"
        __global__ void test(const float *a, const float *b,
                             float *c, const float k) {
          size_t i = blockDim.x * blockIdx.x + threadIdx.x;
          c[i] = (a[i] + b[i]) * k;
        }
        """)
    test = cu.Function(module, "test")
    a = numpy.arange(1000000, dtype=numpy.float32)
    b = numpy.arange(1000000, dtype=numpy.float32)
    c = numpy.empty(1000000, dtype=numpy.float32)
    k = numpy.array([0.5], dtype=numpy.float32)
    a_buf = cu.MemAlloc(ctx, a)
    b_buf = cu.MemAlloc(ctx, b)
    c_buf = cu.MemAlloc(ctx, c)
    test.set_args(a_buf, b_buf, c_buf, k)
    test((a.size, 1, 1))
    c_buf.to_host(c)
    max_diff = numpy.fabs(c - (a + b) * k[0]).max()
    logging.info("max_diff = %.6f", max_diff)

Released under Simplified BSD License. Copyright (c) 2014, Samsung Electronics Co.,Ltd.