SSE & AVX Vectorization

Marchete
164.1K views

Open Source Your Knowledge, Become a Contributor

Technology knowledge has to be shared and made accessible for free. Join the movement.

Create Content

First AVX Code: SQRT calculation

Now that we have reviewed all the requirements, the autovectorization, and AVX intrinsics, we can create our first manually vectorized program. In this exercise, you need to vectorize a sqrt calculation of float numbers. We will explicitly use the __m256 datatype to store our floats, reducing the overhead in data loading.

Vectorized SQRT

You will probably see a 600% performance improvement or more. That is, once you have the data loaded, AVX will perform up to 7 times faster than normal sqrtf. The theoretical limit is 800%, but it's rarely achieved. You can expect between a 300% and 600% average increase.

Open Source Your Knowledge: become a Contributor and help others learn. Create New Content