Previous | Next --- Slide 43 of 58
Back to Lecture Thumbnails
jinghuan

I saw three questions from a paper and was confused about why do following questions matter to performance. 1. When and where should the value at each coordinate in each function be computed? 2. Where should they be stored? 3. How long are values cached and communicated across multiple consumers, and when are they independently recomputed by each?

devs

I actually read the same paper - link for those curious (I think Kayvon may have gotten some of the figures from there e.g. Figure 4) : http://people.csail.mit.edu/jrk/halide-pldi13.pdf.

Something that baffles me is that I can't really seem to find any examples/explanations of the 5th primitive reuse. In this paper, they give a nice visualizations over the four scheduling examples listed here, but no sign of "reuse". Is there a reason why reuse is used less/not spoken about as much?

Also, as for your question, I believe they have a really nice example a little further down (the one Kayvon used in class) with blurx and out where they talk about tradeoffs between all the different scheduling techniques.