Opencl Vs Numpy Simple

March 1st, 2020

Simplest example to showcase the performance difference between OpenCL (using GPUs) and Numpy (with CPU) for vector processing.

import pyopencl as cl
import numpy as np
import pyopencl.array

interactive=True allows us to choose a configuration from various configurations present in a system.

ctx = cl.create_some_context(interactive=True)

Choose platform:
[0] <pyopencl.Platform 'NVIDIA CUDA' at 0x2a8ca50>
[1] <pyopencl.Platform 'Intel Gen OCL Driver' at 0x7fa9c756e4e0>
Choice [0]:0
Set the environment variable PYOPENCL_CTX='0' to avoid being asked again.

pyopencl provides an ipython extension which reduces a lot of boilerplate code

%load_ext pyopencl.ipython_ext

Write a kernel function which will add two vectors and store the result to C.

\[C(i) = A(i) + B(i), \text{for i} \in [0, n)\]

%%cl_kernel -o "-cl-fast-relaxed-math"
__kernel void sum_vector(__global const float *A, __global const float *B, __global float *C) {
    int gid = get_global_id(0);
    C[gid] = A[gid] + B[gid];
}

Here, I am creating three buffers $A$ (a), $B$ (b) and $C$ (c). a and b are allocated memory on both Host (CPU) and device (GPU). c is allocated memory on device.

n = 100000000
queue = cl.CommandQueue(ctx)
a_host = np.random.randn(n).astype(np.float32)
a = cl.array.to_device(queue, a_host)

b_host = np.random.randn(n).astype(np.float32)
b = cl.array.to_device(queue, b_host)
c = cl.array.empty_like(a)

Simple benchmark code for measuring performance of OpenCL (backed by GPU)

def _run_sum():
    event = sum_vector(queue, (n,), None, a.data, b.data, c.data)
    cl.wait_for_events([event])
    
%timeit _run_sum()

8.08 ms ± 94.4 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

%timeit a_host + b_host

89.7 ms ± 308 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

So, OpenCL code runs 10X faster than normal Numpy code.

a_host + b_host

array([ 0.4168285 , -0.5317373 , -0.55536956, ..., -1.4577644 ,
        2.5533843 ,  1.0661736 ], dtype=float32)

array([ 0.4168285 , -0.5317373 , -0.55536956, ..., -1.4577644 ,
        2.5533843 ,  1.0661736 ], dtype=float32)