Writing Your First GPU Kernel in Python with Numba and CUDA

Picture by Writer | Ideogram

GPUs are nice for duties the place it’s worthwhile to do the identical operation throughout completely different items of information. This is called the Single Instruction, A number of Knowledge (SIMD) strategy. In contrast to CPUs, which solely have just a few highly effective cores, GPUs have hundreds of smaller ones that may run these repetitive operations suddenly. You will notice this sample quite a bit in machine studying, for instance when including or multiplying giant vectors, as a result of every calculation is unbiased. That is the best situation for utilizing GPUs to hurry up duties with parallelism.

NVIDIA created CUDA as a approach for builders to jot down packages that run on the GPU as an alternative 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 similar time. The issue is that writing CUDA in C or C++ isn’t precisely beginner-friendly. It’s important to cope with issues like handbook reminiscence allocation, thread coordination, and understanding how the GPU works at a low degree. This may be overwhelming particularly when you’re used to writing code in Python.

That is the place Numba may also help you. It permits writing CUDA kernels with Python utilizing the LLVM (Low Stage Digital Machine) compiler infrastructure to instantly compile your Python code to CUDA-compatible kernels. With just-in-time (JIT) compilation, you’ll be able to annotate your capabilities with a decorator, and Numba handles the whole lot else for you.

On this article, we are going to use a standard 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 situation so all indices will be added concurrently to finish vector addition in a single operation.

Be aware that you’ll require a CUDA GPU to observe this text. You should use Colab’s free T4 GPU or a neighborhood GPU with NVIDIA toolkit and NVCC put in.

# Setting Up the Surroundings and Putting in Numba

 
Numba is offered as a Python package deal, and you may set up it with pip. Furthermore, we are going to 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 end result
• The vector_add_cpu perform merely loops by means of 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 fantastic, it isn’t probably the most environment friendly strategy, particularly for big datasets. Since every addition is unbiased of the others, it is a good candidate for parallel execution on a GPU.

Within the subsequent part, you will notice the best way to convert this similar operation to run on the GPU utilizing Numba. By distributing every element-wise addition throughout hundreds 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 could 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

	# Be certain that we do not exit of bounds
	if place 

Let’s break down what is going on 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 may 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 hundreds of threads on the similar time. However we want a approach 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 singular place, which tells every thread which component of the arrays it ought to add. Be aware that the threads and blocks may not all the time present a legitimate index, as they function in powers of two. This may increasingly result in invalid indices when the vector size will not be 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 similar 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 may experiment with completely different values of block and thread sizes, and print the corresponding values within the GPU kernel. This may also 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 end result 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 we’ve each the CPU and GPU variations of vector addition, it’s time to see how they evaluate. You will need to confirm the outcomes and the execution increase 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 necessary to ensure our GPU code is working accurately and the output must 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 instances and take the typical to get a dependable timing. Lastly, we calculate what number of instances sooner the GPU model is in comparison with the CPU. It’s best to see an enormous distinction as a result of the GPU can do many operations directly, whereas the CPU handles them one by one in a loop.

Right here is the anticipated output on NVIDIA’s T4 GPU on Colab. Be aware 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 check helps display the facility of GPU acceleration and why it’s so helpful for duties involving giant quantities of information and parallel work.

# Wrapping Up

 
And that’s it. You could have now written your first CUDA kernel with Numba, with out really writing any C or CUDA code. Numba permits a easy interface for utilizing the GPU by means of 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. In case you discover an issue following the SIMD paradigm, it’s all the time a good suggestion to make use of GPU to enhance execution.

The entire code is offered on Colab pocket book that you could entry right here. Be happy to check it out and make easy adjustments 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 drugs. She co-authored the e-book “Maximizing Productiveness with ChatGPT”. As a Google Technology Scholar 2022 for APAC, she champions variety and educational excellence. She’s additionally acknowledged as a Teradata Range 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.