Hello world with CUDA

CUDA is becoming one of the essential tools to learn. It is going to help you even you are not in Deep Learning world. Gaming industry, Finance Industry and even there are databases which runs on GPU.

In this blog post we will try to get started quickly with CUDA. At the end of the blog there are some resources which can help you learn CUDA in detail.

Launch GPU on aws.

If you already know or you have a GPU instance, then you can skip this section.

First we will see that how do we spin up a gpu instance on aws to write our first CUDA code. We need a CUDA supported GPU. I know I know aws is a bit costly and there are many alternatives like paperspace, FloydHub etc but we will stick to aws because let’s say I ain’t spending my money 😛

When you login to aws console, you will be asked to select an AMI(Amazon Machine Image). I prefer “Deep Learning AMI(ubuntu)” because it comes with all the pre installed packages I need and it is free!

The second step here is to select the instance type. I have selected “p2.xlarge”.

after this, you can directly do “review and finish”. Don’t forget to download “.pem” file.

Then give your .pem right permissions.

chmod 400 name-of-your-key.pem

Now we are set to login to our gpu instance –

ssh -i name-of-your-key.pem ubuntu@your-instance-ip

Yay! we are in.

Writing your first CUDA code.

Congratulation! you have come this far. In this section we will write our first hello world CUDA program in cpp(You can also write the code in C).

Now, Let’s try to understand the above code. The __global__ specifier indicates a function that runs on GPU. These function can be called through host code, e.g. the main() function in the example, and is also known as “kernels“. When a kernel is called, its execution configuration is provided through <<<...>>> syntax, e.g. cuda_hello<<<1,1>>>(). This is called “kernel launch“.

Keep in mind that if you compile the above file, say “hello-world.cu”. It will give you printf not found error. To resolve this problem, include #<stdio.h> rather than #include<iostream>.

The extension for CUDA files is “.cu” .

Compiling CUDA code.

nvcc hello-world.cu -o hello

Further Reading –

I hope you enjoyed reading this blog post and If there are any doubts then feel free to reach out. This blog was meant to be a kickstarter we will try to dive deep into CUDA in coming blogs.

Originally published at vipulvaibhaw.com on March 17, 2019.