PyOpenCL Parallel Patterns: Map/Elementwise

Setup code

In [1]:
import pyopencl as cl
import pyopencl.array
import pyopencl.clrandom
import numpy as np
In [2]:
ctx = cl.create_some_context()
queue = cl.CommandQueue(ctx)
In [3]:
n = 10**7
a = cl.clrandom.rand(queue, n, np.float32)
b = cl.clrandom.rand(queue, n, np.float32)

A simple 'target application'

We would like to evaluate this linear combination:

In [4]:
c1 = 5*a + 6*b

A problem with this is that every single operator (all three of them--and easily more for complicated expressions) corresponds to a kernel call, which can lead to high overhead. Let's try and avoid that by stuffing the entire operation into one kernel, in turn saving lots of memory traffic:

In [5]:
from pyopencl.elementwise import ElementwiseKernel
In [6]:
lin_comb = ElementwiseKernel(ctx,

        "float a, float *x, float b, float *y, float *c",

        "c[i] = a*x[i] + b*y[i]")
In [7]:
c2 = cl.array.empty_like(a)
lin_comb(5, a, 6, b, c2)
Out[7]:
<pyopencl.cffi_cl.Event at 0x7f859c0d0fd0>
In [8]:
import numpy.linalg as la
print(la.norm(c1.get() - c2.get()))
0.0

Timing ElementwiseKernel

Did this optimization pay off?

In [11]:
from time import time
queue.finish()
start_time = time()

for i in range(10):
    c1 = 5*a + 6*b
    
queue.finish()
print("elapsed: {0} s".format(time()-start_time))
elapsed: 5.775973320007324 s
In [12]:
from time import time
queue.finish()
start_time = time()

for i in range(10):
    lin_comb(5, a, 6, b, c2)
    
queue.finish()
print("elapsed: {0} s".format(time()-start_time))
elapsed: 2.476457357406616 s
In [ ]: