pyopencl benchmark

clbench.py

import pyopencl as cl
from pyopencl import array as clarray
from pyopencl import clmath
from pyopencl import clrandom
import numpy
from time import time

def timeit(f, args):
    n = 100
    t = 0
    for _ in range(n):
        start = time()
        f(*args)
        stop = time()
        t += stop - start
    print t / n

def runnumpy(n):
    a = numpy.random.rand(n).astype(numpy.float32)
    b = numpy.random.rand(n).astype(numpy.float32)

    timeit(lambda a, b: a + b, (a, b))
    timeit(lambda a: a + 1.0, (a,))
    timeit(lambda a, b: a * b, (a, b))
    timeit(lambda a: a * 2.0, (a,))
    timeit(lambda a: numpy.sin(a), (a,))
    timeit(lambda a: numpy.sum(a), (a,))
    timeit(lambda a, b: numpy.dot(a, b), (a, b))
    timeit(lambda : numpy.random.rand(n), ())

def runcl(device, n):
    ctx = cl.Context([device])
    queue = cl.CommandQueue(ctx)

    a = clrandom.rand(queue, n, numpy.float32)
    b = clrandom.rand(queue, n, numpy.float32)
    cl.enqueue_barrier(queue)

    def add(a, b):
        a + b
        cl.enqueue_barrier(queue)
    timeit(add, (a, b))

    def add_s(a):
        a + 1.0
        cl.enqueue_barrier(queue)
    timeit(add_s, (a,))

    def mul(a, b):
        a * b
        cl.enqueue_barrier(queue)
    timeit(mul, (a, b))

    def mul_s(a):
        a * 2.0
        cl.enqueue_barrier(queue)
    timeit(mul_s, (a,))

    def sin(a):
        clmath.sin(a)
        cl.enqueue_barrier(queue)
    timeit(sin, (a,))

    def sum(a):
        clarray.sum(a)
        cl.enqueue_barrier(queue)
    timeit(sum, (a,))

    def dot(a, b):
        clarray.dot(a, b)
        cl.enqueue_barrier(queue)
    timeit(dot, (a,b))

    def rand():
        clrandom.rand(queue, n, numpy.float32)
        cl.enqueue_barrier(queue)
    timeit(rand, ())

ns = [1000, 10000, 100000, 1000000]

for platform in cl.get_platforms():
    for device in platform.get_devices():
        print device.name
        for n in ns:
            print n
            runcl(device, n)

print 'numpy'
for n in ns:
    print n
    runnumpy(n)