Saturday, 2023-04-15

*** gnucode <gnucode!~gnucode@user/jab> has quit IRC03:20
markoslkcl, did you read the comments about the software ecosystem?10:01
markosthey basically shot themselves in the foot, it's MIPS but not entirely MIPS so everything is totally incompatible10:01
markosso you can't reuse debian mips ports on this cpu10:02
markosour approach is much more sane10:02
markosit will save us *years* of porting development10:02
lkcltoshywoshy, this one's really short and easy to do
lkclthis one's longer and will need slightly more investment of time in the unit tests
lkclbut what jacob wrote for crternlogi can be used as a template10:08
lkclalthough bmask might be easier as a next candidate after cprop?10:09
*** ghostmansd[m] <ghostmansd[m]!> has quit IRC10:58
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has joined #libre-soc10:58
lkclls008 v1 is in, after i added appendices for the element stepping functions11:00
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has quit IRC11:00
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has joined #libre-soc11:01
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has quit IRC11:07
*** ghostmansd[m] <ghostmansd[m]!> has joined #libre-soc11:17
*** ghostmansd[m] <ghostmansd[m]!> has quit IRC11:39
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has joined #libre-soc11:40
lkclmarkos, you have some warm words about hyperscan?11:59
lkclemail me them, or add them directly to the latex i'm happy with either12:00
lkcltexstudio does a great job on that btw12:01
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has quit IRC12:20
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has joined #libre-soc12:21
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has quit IRC12:32
*** ghostmansd[m] <ghostmansd[m]!> has joined #libre-soc12:32
*** ghostmansd[m] <ghostmansd[m]!> has quit IRC12:47
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has joined #libre-soc12:47
markoslkcl, I will prepare something and send it to you in the next minutes14:49
markosagain, we cannot do anything on *hyperscan*14:49
markosit's always vectorscan14:50
markoshyperscan is Intel's and I have zero control over it, I tried14:50
markosit's not exactly nitpicking but I'm careful not to promise something I cannot deliver14:50
markosit's API and ABI compatible so for end users it should not make a difference14:51
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has quit IRC16:12
*** ghostmansd[m] <ghostmansd[m]!~ghostmans@> has joined #libre-soc16:14
lkclthx - mainly, what do we plan to do, what's the goal?17:07
lkcli.e. what's the goal of the project? in one line: "the plan is to port VectorScan to SVP64"17:08
lkcland to see if estimates of power reduction and performance can be achieved17:09
lkclmarkos, ok let me know if you're happy with this;a=blob;f=conferences/ngisearch2023/ngisearch2023.tex;hb=HEAD17:21
lkclvery very brief, it's literally "what is it, what does it do, and what are we going to attempt"17:21
lkclit'll be like under a minute available per slide, maybe only 40 seconds.17:24
lkclthere's 10 projects, total time is 60 minutes, that's 6 minutes per project *including Q&A*17:24
markosyes, that's correct17:47
markosbasically, it's as you said, port it to SVP6417:47
markosideally we are aiming for full porting of every algorithm, but if that's not possible then at least it should be usable even in scalar form in some algorithms17:48
markosremember it doesn't even compile currently if no SIMD is available17:48
markosso one of the intermediate goals is to provide normal scalar implementations of algorithms and port those to SVP6417:48
markoswhich is a good thing in the long run, it will run on all arches with working C compiler17:49
lkclbleugh! :)17:51
lkcli remember you mentioned about going back to an older version17:51
markosto look for the scalar versions yes, but I wouldn't actually "revert" the whole project to an older version18:01
markosso this would be the first step in my approach, check the original scalar functions, make sure it runs with NO SIMD unit available18:02
markoseg on SFFS :)18:02
markosand then start porting each function to SVP64 through the emulator18:02
markosthe nice thing is that we have a huge unit test suite so it really means we can get a full 100% usable port18:03
markosbtw, what's the video call medium? jitsi/zoom/other?18:05
markosI don't remember receiving any such link18:05
markosah it's jitsi18:05
lkclthe one on hivane.net21:05
*** DevHONKIntel486S <DevHONKIntel486S!~allonsenf@2001:470:69fc:105::2:4883> has joined #libre-soc21:05

Generated by 2.17.1 by Marius Gedminas - find it at!