Skip to content

Latest commit

 

History

History
94 lines (70 loc) · 2.91 KB

README.md

File metadata and controls

94 lines (70 loc) · 2.91 KB

pyopencl_blas

PyOpenCL wrappers for AMD clMathLibraries.clBLAS.

Currently only supports a small subset of the BLAS library (specifically the SWAP, SCAL, COPY, AXPY, GEMV, and GEMM families of functions), for real numbers (i.e., 32-bit or 64-bit floats) and complex numbers (i.e., complex64 and complex128).

Installation

First, install PyOpenCL as per the instructions. You will also need to install Cython. Additionally, the OpenCL headers are required for building:

sudo apt-get install opencl-headers

Download the latest clBLAS release (2.4.0 at the time of writing), and unpack it somewhere (I suggest unpacking to /opt/). Inside, you will find an include directory and a lib directory (called lib64 on 64-bit machines). Your machine will need to know where to find the libraries when running the program. On Linux, you can do this by putting a file in /etc/ld.so.conf.d/ that contains the full path to the lib directory (the file name must end in .conf). Then call sudo ldconfig (you can do sudo ldconfig -v | grep libclBLAS to make sure that the library has been detected). I am not sure how to add the libraries on other systems.

Open up setup.py and change the include dirs in the extension to target your OpenCL include directory and your clBLAS include directory (which you just installed), respectively. Also change the library dir to target your clBLAS library directory (again, which you just installed).

Then, build the project:

python setup.py build_ext --inplace

It should compile without errors, and create pyopencl_blas/blas.so (as well as a corresponding blas.cpp file).

You can now install the project:

python setup.py install --user

or do a "developer" install:

python setup.py develop --user

The latter will mean that changes to this source directory will show up when you import the package (making it easy to develop). You do not need the --user flag if installing to a virtualenv (they're great; check them out!).

Usage

The basic usage is to start up PyOpenCL as usual, create some PyOpenCL Arrays, and pass them to the BLAS functions.

import numpy as np
import pyopencl
import pyopencl.array
import pyopencl_blas
pyopencl_blas.setup()  # initialize the library

ctx = pyopencl.create_some_context()
queue = pyopencl.CommandQueue(ctx)

dtype = 'float32'  # also supports 'float64', 'complex64' and 'complex128'
x = np.array([1, 2, 3, 4], dtype=dtype)
y = np.array([4, 3, 2, 1], dtype=dtype)

clx = pyopencl.array.to_device(queue, x)
cly = pyopencl.array.to_device(queue, y)

# call a BLAS function on the arrays
pyopencl_blas.axpy(queue, clx, cly, alpha=0.8)
print("Expected: %s" % (0.8 * x + y))
print("Actual:   %s" % (cly.get()))

See the examples folder for more examples.