Nice! Actually, we don't have any superoptimization papers on our reading list yet, and I've been meaning to ask if you have a favorite.
-
-
Replying to @lindsey
there are a lot of good choices depending on what you want to emphasize. I like the Bansal/Aiken paper from ASPLOS '06 a lot, and also "Synthesis of Loop Free Programs" from like PLDI '14
2 replies 0 retweets 6 likes -
-
-
Replying to @johnregehr @lindsey
I liked this class when I took it in 2012: http://web.stanford.edu/class/cs357/ Alex and David gave the class, with Clark as guest. I'm thinking of this paper: https://raw.githubusercontent.com/StanfordPL/stoke/develop/docs/papers/asplos13.pdf … I still have some notes and the readings. Looks like I did this for the class:https://github.com/jfbastien/BitBlast …
4 replies 1 retweet 7 likes -
I used STOKE once to generate some SIMD code! It was fun
1 reply 0 retweets 4 likes -
I think it's so cool that it actually works
1 reply 0 retweets 1 like -
I found STOKE useful as sort of a guide to give me a rough idea of the best way to do something, which I then implemented manually. There’s more work to be done before I’d be comfortable actually shipping the generated code verbatim.
1 reply 0 retweets 1 like -
basically all superoptimizers to date have this issue, that a human should be in the loop somewhere. it's something I'd like to fix.
2 replies 0 retweets 1 like -
I’m actually OK with this, though. I think I want an “algorithm idea generator” more than a superoptimizer.
3 replies 0 retweets 1 like
In my case (PNG prediction during decoding), I knew what I wanted to do, and I could write some basic code, but I wasn’t sure if there was some clever trick I was missing. A superoptimizer would be useful to point out tricks that I missed.
-
-
The model for STOKE is too naive though. It doesn’t know about critical path, superscalar, loop stream, etc. It only counts per-instruction cost. It’s good at high level things, but many of the details it could learn. It’s also missing concurrency correctness and debugging
1 reply 0 retweets 0 likes -
it doesn't use something like llvm-mca??
1 reply 0 retweets 0 likes - 2 more replies
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.