— Dense Linear Algebra Algorithms on AMD GPUs

Blog | About | Timeline | Links | Art | Archive


image thumbnail

Dense Linear Algebra Algorithms on AMD GPUs

I’m excited to have recently published my first journal paper with ICL! You can check it out on HGPU. TL;DR: we port a math library to AMD GPUs and get much better performance than AMD’s existing vendor libraries

You can check out the paper at the bottom of this page, or at this URL

I wrote this paper with Ahmad Abdelfattah, Stan Tomov, and Jack Dongarra at the Innovative Computing Laboratory, while a research assistant. We used auto-translation tools to convert the source code (in addition to some manual changes) to port from the existing CUDA platform to the new HIP/ROCm platform.

While doing this, we also did performance tuning on algorithms, specifically aimed at improving performance on the new AMD hardware. There are a number of differences (such as warp size, memory hierarchy, and compiler optimizations) that allowed us to improve performance, bringing it a lot closer to the peak performance on these GPUs.


My paper was added to the OS hackathon resources page