| programmerjake | roasting arm's reduce pseudo-code: https://zulip-archive.rust-lang.org/stream/257879-project-portable-simd/topic/pairwise.20ops.20would.20be.20nice.html#270330368 | 04:24 |
|---|---|---|
| programmerjake | https://rust-lang.zulipchat.com/#narrow/stream/257879-project-portable-simd/topic/pairwise.20ops.20would.20be.20nice/near/270330368 | 04:25 |
| programmerjake | scroll up a little from where i linked | 04:26 |
| lkcl | programmerjake, the second link is requesting a login (which i'm not going to do) | 10:41 |
| lkcl | the first i can see no problem | 10:42 |
| lkcl | the original algorithm you wrote jacob carries-over state, i believe (which we can't do). | 10:44 |
| lkcl | it's ok to do tricks such as write to the result vector | 10:45 |
| lkcl | (even though the result will be in element 0) | 10:45 |
| lkcl | that way a context save/restore can snap in/out in the middle of the reduce | 10:46 |
| lkcl | and it's ok to store "REMAP" state in SVREMAP SPRs (the indices / for-loops) | 10:47 |
| lkcl | also another trick is Rc=1, if there's any single-bit boolean state needed as part of the reduce, that can go into a Vector of CRs (vector-version of CR0) | 10:48 |
| lkcl | the other trick - which i am really reluctant to have, it would be last resort - is that if there is an interrupt in the middle you chuck the entire lot and start again on restore | 10:50 |
| lkcl | i think the algorithm you came up with would fit into the SVREMAP hardware | 12:04 |
| lkcl | i leave it to you to work that out | 12:05 |
| *** iridium.libera.chat sets mode: +o lkcl | 21:12 | |
Generated by irclog2html.py 2.17.1 by Marius Gedminas - find it at https://mg.pov.lt/irclog2html/!