Another great thing about this work is showing how flexible the generalizations of "sum" can be. A lot of problems can be restated in terms of an associative operator, and doing so unlocks all that parallel goodness.
I was just curious to ask, as opposed to what all models?
A few years ago I worked on a deep learning project using parallel prefix sum as a new way to accelerate recurrent neural nets on GPUs. The paper in this post was the most important reference and source of inspiration. I'm happy to see this paper shared on HN in hopes that it also sparks ideas in others.
The marching cubes algorithm I was implementing generated a lot of sparse datasets, and I wanted to send only the vertices that were going to be rendered. Got a lot of speedup when all was said and done — and learning this algorithm, and the parallel way of thinking in general, was eye opening.
Thanks to the timeless nature of the PDF format (and the supporting applications), I'm able to read a digital content created 3 decades ago. What also blows my mind is that virtually everything about computers has changed over these decades and yet, here I'm reading this brilliant paper.
There is something about the PDF format that gives me real joy just to look at the beautifully rendered content -- letters and pictures. For me, no other format comes close to PDF. In the physical world Springer books are a good match but that's probably because they are typeset in PDF too.
Prefix sums and their applications [pdf] - https://news.ycombinator.com/item?id=17621998 - July 2018 (1 comment)
Prefix Sums and Their Applications (1993) [pdf] - https://news.ycombinator.com/item?id=7800594 - May 2014 (12 comments)