Writing Your First GPU Kernel in Python with Numba and CUDA -

Picture by Writer | Ideogram

GPUs are nice for duties the place it is advisable do the identical operation throughout completely different items of knowledge. This is named the Single Instruction, A number of Information (SIMD) method. In contrast to CPUs, which solely have just a few highly effective cores, GPUs have 1000’s of smaller ones that may run these repetitive operations abruptly. You will notice this sample rather a lot in machine studying, for instance when including or multiplying giant vectors, as a result of every calculation is unbiased. That is the perfect state of affairs for utilizing GPUs to hurry up duties with parallelism.

NVIDIA created CUDA as a method for builders to jot down applications that run on the GPU as a substitute of the CPU. It’s primarily based on C and allows you to write particular capabilities referred to as kernels that may run many operations on the identical time. The issue is that writing CUDA in C or C++ isn’t precisely beginner-friendly. You must cope with issues like handbook reminiscence allocation, thread coordination, and understanding how the GPU works at a low stage. This may be overwhelming particularly in case you’re used to writing code in Python.

That is the place Numba will help you. It permits writing CUDA kernels with Python utilizing the LLVM (Low Degree Digital Machine) compiler infrastructure to immediately compile your Python code to CUDA-compatible kernels. With just-in-time (JIT) compilation, you may annotate your capabilities with a decorator, and Numba handles every little thing else for you.

On this article, we’ll use a typical instance of vector addition, and convert easy CPU code to a CUDA kernel with Numba. Vector addition is a perfect instance of parallelism, as addition throughout a single index is unbiased of different indices. That is the proper SIMD state of affairs so all indices might be added concurrently to finish vector addition in a single operation.

Word that you’ll require a CUDA GPU to comply with this text. You need to use Colab’s free T4 GPU or an area GPU with NVIDIA toolkit and NVCC put in.

# Setting Up the Surroundings and Putting in Numba

Numba is obtainable as a Python package deal, and you’ll set up it with pip. Furthermore, we’ll use numpy for vector operations. Arrange the Python surroundings utilizing the next instructions:

python3 -m venv venv
supply venv/bin/activate
pip set up numba-cuda numpy

# Vector Addition on the CPU

Let’s take a easy instance of vector addition. For 2 given vectors, we add the corresponding values from every index to get the ultimate worth. We are going to use numpy to generate random float32 vectors and generate the ultimate output utilizing a for loop.

import numpy as np 

N = 10_000_000 # 10 million components 
a = np.random.rand(N).astype(np.float32) 
b = np.random.rand(N).astype(np.float32) 
c = np.zeros_like(a) # Output array 

def vector_add_cpu(a, b, c): 
    """Add two vectors on CPU""" 
    for i in vary(len(a)): 
        c[i] = a[i] + b[i]

Here’s a breakdown of the code:

Initialize two vectors every with 10 million random floating-point numbers
We additionally create an empty vector c to retailer the consequence
The vector_add_cpu perform merely loops by every index and provides the weather from a and b, storing the lead to c

It is a serial operation; every addition occurs one after one other. Whereas this works advantageous, it is not probably the most environment friendly method, particularly for big datasets. Since every addition is unbiased of the others, this can be a excellent candidate for parallel execution on a GPU.

Within the subsequent part, you will notice the way to convert this identical operation to run on the GPU utilizing Numba. By distributing every element-wise addition throughout 1000’s of GPU threads, we will full the duty considerably sooner.

# Vector Addition on the GPU with Numba

You’ll now use Numba to outline a Python perform that may run on CUDA, and execute it inside Python. We’re doing the identical vector addition operation however now it may possibly run in parallel for every index of the numpy array, resulting in sooner execution.

Right here is the code for writing the kernel:

from numba import config

# Required for newer CUDA variations to allow linking instruments. 
# Prevents CUDA toolkit and NVCC model mismatches.
config.CUDA_ENABLE_PYNVJITLINK = 1

from numba import cuda, float32

@cuda.jit
def vector_add_gpu(a, b, c):
	"""Add two vectors utilizing CUDA kernel"""
	# Thread ID within the present block
	tx = cuda.threadIdx.x
	# Block ID within the grid
	bx = cuda.blockIdx.x
	# Block width (variety of threads per block)
	bw = cuda.blockDim.x

	# Calculate the distinctive thread place
	place = tx + bx * bw

	# Make sure that we do not exit of bounds
	if place

Let’s break down what is occurring above.

// Understanding the GPU Operate

The @cuda.jit decorator tells Numba to deal with the next perform as a CUDA kernel; a particular perform that can run in parallel throughout many threads on the GPU. At runtime, Numba will compile this perform to CUDA-compatible code and deal with the C-API transpilation for you.

@cuda.jit
def vector_add_gpu(a, b, c):
	...

This perform will run on 1000’s of threads on the identical time. However we’d like a method to determine which a part of the info every thread ought to work on. That’s what the following few strains do:

tx is the thread’s ID inside its block
bx is the block’s ID inside the grid
bw is what number of threads there are in a block

We mix these to calculate a novel place, which tells every thread which component of the arrays it ought to add. Word that the threads and blocks may not at all times present a legitimate index, as they function in powers of two. This will result in invalid indices when the vector size just isn’t conforming to the underlying structure. Due to this fact, we add a guard situation to validate the index, earlier than we carry out the vector addition. This prevents any out-of-bound runtime error when accessing the array.

As soon as we all know the distinctive place, we will now add the values identical to we did for the CPU implementation. The next line will match the CPU implementation:

c[position] = a[position] + b[position]

// Launching the Kernel

The gpu_add perform units issues up:

It defines what number of threads and blocks to make use of. You’ll be able to experiment with completely different values of block and thread sizes, and print the corresponding values within the GPU kernel. This will help you perceive how underlying GPU indexing works.
It copies the enter arrays (a, b, and c) from the CPU reminiscence to the GPU reminiscence, so the vectors are accessible within the GPU RAM.
It runs the GPU kernel with vector_add_gpu[blocks_per_grid, threads_per_block].
Lastly, it copies the consequence again from the GPU into the c array, so we will entry the values on the CPU.

# Evaluating the Implementations and Potential Speedup

Now that now we have each the CPU and GPU variations of vector addition, it’s time to see how they examine. It is very important confirm the outcomes and the execution enhance we will get with CUDA parallelism.

import timeit

c_cpu = time_cpu()
c_gpu = time_gpu()

print("Outcomes match:", np.allclose(c_cpu, c_gpu))

cpu_time = timeit.timeit("time_cpu()", globals=globals(), quantity=3) / 3
print(f"CPU implementation: {cpu_time:.6f} seconds")

gpu_time = timeit.timeit("time_gpu()", globals=globals(), quantity=3) / 3
print(f"GPU implementation: {gpu_time:.6f} seconds")

speedup = cpu_time / gpu_time
print(f"GPU speedup: {speedup:.2f}x")

First, we run each implementations and test if their outcomes match. That is essential to ensure our GPU code is working accurately and the output needs to be the identical because the CPU’s.

Subsequent, we use Python’s built-in timeit module to measure how lengthy every model takes. We run every perform just a few occasions and take the typical to get a dependable timing. Lastly, we calculate what number of occasions sooner the GPU model is in comparison with the CPU. It is best to see an enormous distinction as a result of the GPU can do many operations directly, whereas the CPU handles them one after the other in a loop.

Right here is the anticipated output on NVIDIA’s T4 GPU on Colab. Word that the precise speedup can differ primarily based on CUDA variations and the underlying {hardware}.

Outcomes match: True
CPU implementation: 4.033822 seconds
GPU implementation: 0.047736 seconds
GPU speedup: 84.50x

This easy take a look at helps reveal the ability of GPU acceleration and why it’s so helpful for duties involving giant quantities of knowledge and parallel work.

# Wrapping Up

And that’s it. You have got now written your first CUDA kernel with Numba, with out truly writing any C or CUDA code. Numba permits a easy interface for utilizing the GPU by Python, and it makes it a lot easier for Python engineers to get began with CUDA programming.

Now you can use the identical template to jot down superior CUDA algorithms, that are prevalent in machine studying and deep studying. Should you discover an issue following the SIMD paradigm, it’s at all times a good suggestion to make use of GPU to enhance execution.

The entire code is obtainable on Colab pocket book that you could entry right here. Be at liberty to try it out and make easy modifications to get a greater understanding of how CUDA indexing and execution works internally.

Kanwal Mehreen is a machine studying engineer and a technical author with a profound ardour for information science and the intersection of AI with medication. She co-authored the e book “Maximizing Productiveness with ChatGPT”. As a Google Technology Scholar 2022 for APAC, she champions range and tutorial excellence. She’s additionally acknowledged as a Teradata Variety in Tech Scholar, Mitacs Globalink Analysis Scholar, and Harvard WeCode Scholar. Kanwal is an ardent advocate for change, having based FEMCodes to empower ladies in STEM fields.

Writing Your First GPU Kernel in Python with Numba and CUDA