| *** choozy <choozy!~choozy@75-63-174-82.ftth.glasoperator.nl> has quit IRC | 01:18 | |
| lkcl | ghostmansd[m], no *no* aliases are to be added, that is *not* on the list. | 02:21 |
|---|---|---|
| openpowerbot | [irc] <programmerjake> the part of fminmax that needs redoing instead of just adding is changing the old fmin*/fmax* ops to be aliases of fminmax, since that was still labeled redo and you only seemed to object to c[ft]fpr* aliases, i assumed fminmax's aliases are included | 02:25 |
| openpowerbot | [irc] <programmerjake> though, i'll note that fmin*./fmax*. with Rc=1 were deleted from binutils but are aliases of fminmax. | 02:27 |
| openpowerbot | [irc] <programmerjake> so, if we have fmin*/fmax* aliases at all, those should be added for consistency | 02:28 |
| lkcl | programmerjake: please *do NOT* keep adding to the scope of the task for ghostmansd[m] | 02:46 |
| lkcl | aliases ARE NOT, WERE NOT and WILL NEVER BE part of the scope of work until such time as a FULL budget and task analysis has been carried out | 02:47 |
| lkcl | my estimates put the task of adding aliases at a MASSIVE 18k EUR budget | 02:47 |
| lkcl | you HAVE to stop adding things to people's plates by changing the scope of their task when you have repeatedly demonstrated a lack of skill in this area. | 02:48 |
| lkcl | your skill is "deep algorithm and research" and i am astounded by how much expertise and knowledge you have in this area | 02:49 |
| lkcl | but putting it bluntly: you suck at time management :) | 02:49 |
| lkcl | help other people to do research: yes. (the list-analysis of instructions was extremely useful) | 02:50 |
| *** gnucode <gnucode!~gnucode@user/jab> has quit IRC | 02:50 | |
| lkcl | define the scope of what they need to do and what they *do not* need to do: no. please *don't* load work onto other people's plates, it's simply not your responsibility: it's theirs to negotiate, ok? | 02:51 |
| lkcl | (and for me to keep an eye on, as the project lead, keeping the entire frickin project roadmap - past present and future - in my head) | 02:52 |
| openpowerbot | [irc] <programmerjake> you can lecture me all you want but it doesn't change that i thought you intended him to do the tasks that i explained to ghostmansd, as i explained ^ | 02:52 |
| lkcl | jacob: you *should not* have added aliases *at all* to ghostmansd[m]'s list of tasks... period. | 02:54 |
| lkcl | this was discussed many months ago, i said NO to aliases and it means NO. | 02:54 |
| lkcl | NO means NO. | 02:54 |
| lkcl | we have had this discussion and you goddamn need to listen | 02:54 |
| lkcl | i am dangerously ill due to overwork and verbal abuse from several different directions. | 02:55 |
| lkcl | i cannot be put under stress under *any* circumstances | 02:55 |
| openpowerbot | [irc] <programmerjake> i understood that to mean no aliases in the simulator, not saying anything about binutils | 02:56 |
| lkcl | this is real serious, i was at "hypertension level 2" blood pressure levels a couple of weeks ago, which is the point where you get strokes and heart attacks. | 02:56 |
| lkcl | if we can't *test* the aliases what's the f*****g point of adding them to binutils?? | 02:57 |
| lkcl | you have *no idea* how to do linking of tasks together in a chain, it's just not in your skillset | 02:57 |
| lkcl | but i can - i've been doing it for decades. | 02:57 |
| openpowerbot | [irc] <programmerjake> because i recalled there being some tests while those were still real insns and not aliases | 02:58 |
| lkcl | and the list of tasks that are inter-dependent - i counted SEVEN of them - came out at a whopping EIGHTEEN THOUSAND EUROs. | 02:58 |
| lkcl | none of which is accounted for in ANY of the budgets because i had ALREADY EXCLUDED ALIASES ENTIRELY | 02:58 |
| lkcl | because without openpower-isa unit tests WE CANNOT TEST BINUTILS. | 02:58 |
| lkcl | okay?? | 02:58 |
| lkcl | you made an assumption. | 02:58 |
| lkcl | you didn't *ASK* | 02:59 |
| openpowerbot | [irc] <programmerjake> binutils has its own tests too | 02:59 |
| lkcl | and you added to ghostmansd[m]'s tasklist FOR WHICH HE IS GETTING PAID LESS MONEY by saying "yes go ahead and do the work" | 02:59 |
| lkcl | now i f*****g well have to write to him to say "terribly sorry ghostmansd[m] but you now have to do EVEN MORE WORK for the same amount of money by taking those out" | 03:00 |
| lkcl | which is why | 03:00 |
| lkcl | i said | 03:00 |
| lkcl | and repat | 03:00 |
| lkcl | STOP interfering in other people's tasks when you DON'T have time and budget management skills | 03:00 |
| openpowerbot | [irc] <programmerjake> if i asked about everything we'd never get any work done, this is why #1125 took >1 week instead of a day, because I was busy asking you what you wanted done instead of you leaving it up to me to pick | 03:00 |
| lkcl | help them with their technical tasks but DO NOT interfere with what they do | 03:01 |
| lkcl | i can't cope with this | 03:01 |
| openpowerbot | [irc] <programmerjake> i assumed ghostmansd knew that and so would choose which to do as part of that task from the list i gave, i probably should have said that. sorry | 03:02 |
| openpowerbot | [irc] <programmerjake> sorry luke & dimitry for the trouble i've inadvertently caused | 03:24 |
| openpowerbot | [irc] <programmerjake> lkcl: sorry for being rude to you, i'll try to follow what you asked me to do. | 06:29 |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@broadband-90-154-80-239.ip.moscow.rt.ru> has quit IRC | 07:46 | |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@176.59.40.164> has joined #libre-soc | 07:47 | |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@176.59.40.164> has quit IRC | 07:53 | |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@46.39.248.179> has joined #libre-soc | 07:54 | |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@46.39.248.179> has quit IRC | 08:09 | |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@176.59.40.59> has joined #libre-soc | 08:09 | |
| ghostmansd[m] | Hey guys, keep calm :-) | 08:13 |
| ghostmansd[m] | I don't feel stressed being hit by bunch of aliases | 08:14 |
| ghostmansd[m] | I added them because on TODO list some insns had "no aliases" but fminmax didn't | 08:14 |
| ghostmansd[m] | So, since we already had added minmax aliases, I felt this implied I should support those for fminmax as well | 08:15 |
| ghostmansd[m] | Dropping them is easy, I'll do it, and will also drop minmax aliases until we have a clear and obvious format which allows to automate it at least somewhat | 08:16 |
| openpowerbot | [irc] <programmerjake> ok, sounds good to me! | 08:17 |
| ghostmansd[m] | So, no hard feelings, please, we need to be cooperative and constructive :-) | 08:17 |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@176.59.40.59> has quit IRC | 08:25 | |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@broadband-90-154-80-239.ip.moscow.rt.ru> has joined #libre-soc | 08:25 | |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@broadband-90-154-80-239.ip.moscow.rt.ru> has quit IRC | 10:49 | |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@176.59.57.3> has joined #libre-soc | 10:50 | |
| *** hl <hl!~hl@user/hl> has quit IRC | 10:59 | |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@176.59.57.3> has quit IRC | 11:05 | |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@85.30.228.54> has joined #libre-soc | 11:05 | |
| *** hl <hl!~hl@user/hl> has joined #libre-soc | 11:07 | |
| *** hl <hl!~hl@user/hl> has quit IRC | 12:07 | |
| *** hl <hl!~hl@user/hl> has joined #libre-soc | 12:11 | |
| openpowerbot | [irc] <sadoon[m]1> 87/161 success, 3 failures, still ongoing for debian | 12:13 |
| openpowerbot | [irc] <sadoon[m]1> God bless bash scripts | 12:13 |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@85.30.228.54> has quit IRC | 12:42 | |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@176.59.160.56> has joined #libre-soc | 12:42 | |
| ghostmansd[m] | sadoon[m], amen! | 12:49 |
| *** lxo <lxo!~lxo@gateway/tor-sasl/lxo> has quit IRC | 12:51 | |
| lkcl | sadoon: not baaad | 12:56 |
| *** lkcl <lkcl!lkcl@freebnc.bnc4you.xyz> has quit IRC | 13:05 | |
| *** awilfox <awilfox!~awilfox@kelsey.foxkit.us> has quit IRC | 13:05 | |
| *** lxo <lxo!~lxo@gateway/tor-sasl/lxo> has joined #libre-soc | 13:05 | |
| *** lkcl <lkcl!lkcl@freebnc.bnc4you.xyz> has joined #libre-soc | 13:05 | |
| *** awilfox <awilfox!~awilfox@kelsey.foxkit.us> has joined #libre-soc | 13:05 | |
| *** hl <hl!~hl@user/hl> has quit IRC | 13:16 | |
| *** hl <hl!~hl@user/hl> has joined #libre-soc | 13:18 | |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@176.59.160.56> has quit IRC | 13:18 | |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@broadband-90-154-80-239.ip.moscow.rt.ru> has joined #libre-soc | 13:18 | |
| openpowerbot | [irc] <sadoon[m]1> :D | 14:45 |
| openpowerbot | [irc] <sadoon[m]1> Failures are gpgv, libc-bin, and libc6 (which I assume is the same source package as libc-bin) | 14:46 |
| openpowerbot | [irc] <sadoon[m]1> I'm chilling and waiting for it to finish and then will start manually trying the failed ones, gpgv should be easy | 14:46 |
| lkcl | libc6 i'd kinda expect to fail, without some persuasion: there *should* be patches upstream for removing "#ifdef POWER9" and replacing with "#ifdef HAVE_VSX" and so on | 15:04 |
| lkcl | but it will depend on the version | 15:04 |
| openpowerbot | [irc] <sadoon[m]1> It built fine on gentoo so I don't expect it to be something major | 15:18 |
| openpowerbot | [irc] <sadoon[m]1> One day I'll have to get myself a proper IDE or configure my vimrc to make it act like one for such cases | 15:20 |
| openpowerbot | [irc] <sadoon[m]1> My vimrc already has a bunch of useful stuff but searching in multiple files and directories, that I'm missing | 15:20 |
| lkcl | i create shell-scripts for that, i have one called "rgfind.sh" specially for python files - find . -name "*.py" | xargs grep "$@" | 15:40 |
| lkcl | then call them out with vim macros on either "," or "-" or "=" which are free keys | 15:40 |
| lkcl | map = :'a,!fmt^M | 15:41 |
| lkcl | is a good one | 15:41 |
| lkcl | use marker-a ('a), move the cursor, then press "=" and the text between the two is reformatted with /bin/fmt | 15:42 |
| lkcl | map , ^[:w^M:!python %:t^M^M | 15:42 |
| lkcl | saves then runs the current file as a python script | 15:42 |
| lkcl | = , ` - and + are i believe all "free for macro use" but you can likely find them online | 15:43 |
| lkcl | no reason why you shouldn't put the current line contents into a yank-buffer which is then extracted and used as the term you want to recursively search for with /usr/bin/find | 15:44 |
| lkcl | (or grep -r) | 15:44 |
| openpowerbot | [irc] <sadoon[m]1> It's a bit too tedious :P | 16:03 |
| *** octavius <octavius!~octavius@92.40.168.138.threembb.co.uk> has joined #libre-soc | 17:26 | |
| openpowerbot | [irc] <programmerjake> luke, i think i'd like to next work on implementing fminmax and c[ft]fpr* in ieee754fpu, what do you think? | 17:59 |
| openpowerbot | [irc] <programmerjake> part of that will be adding minimal fpscr support to ieee754fpu | 18:00 |
| openpowerbot | [irc] <programmerjake> and rounding modes (as in the classes for all that, existing pipelines won't be modified yet) | 18:01 |
| openpowerbot | [irc] <programmerjake> they'd likely be sub-tasks of https://bugs.libre-soc.org/show_bug.cgi?id=1026 | 18:03 |
| openpowerbot | [irc] <programmerjake> c[ft]fpr* is a lot simpler in hw than the pseudocode makes it look | 18:04 |
| ghostmansd[m] | No idea WTF happened to binutils between revisions, but the fact is that the last sync with master eats the whole RAM when I build it with multiple jobs. | 18:05 |
| ghostmansd[m] | The prick literally eats all the memory the VM has. | 18:06 |
| openpowerbot | [irc] <programmerjake> basically cffpr* is extract fields, add constant to exponent, shift & mask, and then round | 18:06 |
| ghostmansd[m] | Probably it's related to clangd. I"ll check. | 18:06 |
| openpowerbot | [irc] <programmerjake> clangd?! that shouldn't be part of binutils builds...it's for IDEs | 18:07 |
| openpowerbot | [irc] <sadoon[m]1> That's weird yeah | 18:08 |
| openpowerbot | [irc] <sadoon[m]1> Unless there's another clangd we're not aware of | 18:09 |
| ghostmansd[m] | Yep, it's part of IDE | 18:09 |
| openpowerbot | [irc] <sadoon[m]1> welp, libsystemd also failed, gonna have to look into that also | 18:09 |
| openpowerbot | [irc] <programmerjake> or are you running builds from an ide? | 18:09 |
| ghostmansd[m] | programmerjake, yep, correct | 18:09 |
| openpowerbot | [irc] <sadoon[m]1> Could have some crypto stuff that depends on 128 float | 18:10 |
| ghostmansd[m] | And until today this has never been a problem :-) | 18:10 |
| openpowerbot | [irc] <programmerjake> hmm, set clangd to only use a few threads? may reduce memory usage | 18:10 |
| ghostmansd[m] | In fact, I never ever thought this can be an issue :-) | 18:10 |
| ghostmansd[m] | I killed the IDE process and now build w/o it | 18:11 |
| ghostmansd[m] | Things seem to work, the build is in progress | 18:11 |
| openpowerbot | [irc] <programmerjake> how much ram does your vm have? | 18:11 |
| ghostmansd[m] | 8 GiB | 18:11 |
| ghostmansd[m] | I really thought that's sufficient :-) | 18:11 |
| ghostmansd[m] | It's not like that I have plenty of RAM | 18:12 |
| openpowerbot | [irc] <programmerjake> oh, if you can set it to at least 16GiB | 18:12 |
| openpowerbot | [irc] <programmerjake> if you have 16GiB host ram, try setting it to 14GiB | 18:12 |
| ghostmansd[m] | Then this will be the whole RAM I have :-D | 18:12 |
| openpowerbot | [irc] <sadoon[m]1> Less jobs perhaps? | 18:13 |
| ghostmansd[m] | Ok, might give it a shot... But wouldn't it be way too much? | 18:13 |
| ghostmansd[m] | I used `make -j`, I assumed it spawns nproc processes... | 18:13 |
| ghostmansd[m] | But I really assume it's clangd | 18:13 |
| openpowerbot | [irc] <programmerjake> make -j spawns infinite processes, you need make -j$(nproc) | 18:14 |
| ghostmansd[m] | Because this one is the first to be killed by OOM killer | 18:14 |
| ghostmansd[m] | Really? Ok, this might be the reason... I think I used nproc before... | 18:14 |
| openpowerbot | [irc] <programmerjake> or make -j2 or -j4 | 18:14 |
| ghostmansd[m] | That said, -j runs OK without IDE | 18:15 |
| openpowerbot | [irc] <programmerjake> yeah, make -j has a bad default | 18:15 |
| openpowerbot | [irc] <sadoon[m]1> Can't come to the meeting tonight, apologies | 20:05 |
| *** lkcl <lkcl!lkcl@freebnc.bnc4you.xyz> has quit IRC | 20:25 | |
| *** lkcl <lkcl!lkcl@freebnc.bnc4you.xyz> has joined #libre-soc | 20:25 | |
| openpowerbot | [irc] <lkcl__> sadoon no problem | 20:34 |
| openpowerbot | [irc] <lkcl__> programmerjake, great idea (fpscr support etc.) | 20:35 |
| openpowerbot | [irc] <lkcl__> remember that exceptions are nothing particularly special, but that if they *could* occur, the "Shadow Hold" flag has to be held until the point where it is 100% known that under no circumstances could an exception be raised from that point onwards in relation to that result | 20:36 |
| openpowerbot | [irc] <lkcl__> (this *could* be half-way down a pipeline!) | 20:36 |
| openpowerbot | [irc] <lkcl__> therefore one of the slightly-intrusive outputs from the FPU when a pipeline is created will have to be an *array* of output bits indicating: | 20:38 |
| openpowerbot | [irc] <lkcl__> pipeline stage 1: exception could still occur for this particular result | 20:38 |
| openpowerbot | [irc] <lkcl__> pipeline stage 2: exception could still occur for this result | 20:38 |
| openpowerbot | [irc] <lkcl__> .... | 20:38 |
| openpowerbot | [irc] <lkcl__> pipeline stage N: exception could NOT still occur for this result because it's now "safe" | 20:39 |
| openpowerbot | [irc] <lkcl__> if done instead as a FSM it's much easier, the FSM would have one single bit (stored in a DFF) | 20:40 |
| openpowerbot | [irc] <lkcl__> saying "yep i don't know if it's safe yet that an exception could occur, so for safety's sake i'm leaving the flag high" | 20:40 |
| openpowerbot | [irc] <lkcl__> several clock cycles later... | 20:41 |
| openpowerbot | [irc] <lkcl__> "ok i know now that the safe-point is passed, we're now on to 'normalisation' (or whatever), i can drop the "exception-might-occur" flag" | 20:41 |
| openpowerbot | [irc] <lkcl__> at which point once that flag is dropped, it is an ABSOLUTE 100% ABSOLUTE 100% without fail ABSOLUTE and ABSOLUTE 100% guaranteed indication that the result *CAN AND WILL* be written to the FPR | 20:42 |
| openpowerbot | [irc] <lkcl__> because from that point onwards once that flag is dropped there is ZERO possibility of an exception being raised for that result | 20:43 |
| openpowerbot | [irc] <lkcl__> that's the "contract" that an OoO system has to have, built-in, right deep into the system. | 20:44 |
| openpowerbot | [irc] <lkcl__> we've not had to deal with this before because this is the first time that Exceptions like this have to be raised (from a pipelined resource) | 20:44 |
| openpowerbot | [irc] <lkcl__> at present i "avoid" this problem in TestIssuer by *NOT PERMITTING* anything to occur whilst a LD/ST is in progress... LD/ST being the only place where an exception could occur | 20:45 |
| openpowerbot | [irc] <programmerjake> ok, i'll create some tasks later | 20:45 |
| openpowerbot | [irc] <lkcl__> awesome. | 20:45 |
| openpowerbot | [irc] <lkcl__> this is all in the notes on "Shadow Hold" in the 3d_gpu architecture somewhere | 20:46 |
| openpowerbot | [irc] <lkcl__> Shadow-Hold and Shadow-Cancel | 20:46 |
| openpowerbot | [irc] <lkcl__> it *may* be easier and simpler to do a FSM-only FPU | 20:47 |
| openpowerbot | [irc] <lkcl__> (just like in microwatt) | 20:48 |
| openpowerbot | [irc] <lkcl__> then to get parallelism just... cut/paste multiple FSMs and wire them all up to a multi-input multi-output ReservationStation. | 20:48 |
| openpowerbot | [irc] <lkcl__> there's a class already written that does that, which wires into the ALUCompUnits. | 20:49 |
| openpowerbot | [irc] <programmerjake> i think we should have pipelined FPUs, so we have much less work later when wiring them up to the full OoO cpu | 20:49 |
| openpowerbot | [irc] <lkcl__> ver vague recollection of them being called ReservationStations2? | 20:49 |
| openpowerbot | [irc] <lkcl__> no jacob: please *re-read* what i just wrote above | 20:50 |
| openpowerbot | [irc] <lkcl__> pipelines are *not* the sole exclusive means and method of parallelism. | 20:50 |
| openpowerbot | [irc] <programmerjake> for the reason that 1 pipeline is much smaller than 3-4 fsms for many fp ops due to sharing shifters/mul units etc. | 20:50 |
| openpowerbot | [irc] <lkcl__> it is possible to lay down *MULTIPLE* FSMs and present *all* of them via fan-out FSM1-FSM2-FSM3-FSM4-FSM5 ... fan-in *as if* they were a 5-stage "pipeline" | 20:51 |
| openpowerbot | [irc] <lkcl__> too complex at this point | 20:51 |
| openpowerbot | [irc] <lkcl__> we can get to that later - and there's techniques for dealing with that, by micro-coding | 20:51 |
| openpowerbot | [irc] <lkcl__> you have the *FSM* connected to those pipelines performing shift/multiply (!!!) | 20:52 |
| openpowerbot | [irc] <programmerjake> well, fminmax is like a 1 stage pipeline and c[ft]fpr is like 2 stage | 20:52 |
| openpowerbot | [irc] <lkcl__> as an "extra user" | 20:52 |
| openpowerbot | [irc] <lkcl__> if you can please do things as "split stages" (one class per stage) just like we did with Div Units then a "reorg" can be carried out later | 20:53 |
| openpowerbot | [irc] <programmerjake> all the fsm stuff is too complex due to having to keep multiple fsms out of sync so they don't conflict with shared stuff | 20:54 |
| openpowerbot | [irc] <lkcl__> exaclty: then don't share *anything* at this point | 20:54 |
| openpowerbot | [irc] <programmerjake> yeah, like div is the plan (though much less stages) | 20:54 |
| openpowerbot | [irc] <lkcl__> it's too big a task, but i did actually prepare the ReservationStations(2?) classes for this | 20:55 |
| openpowerbot | [irc] <programmerjake> for 1-2 stages, pipelines are simpler for me to think about than fsms | 20:55 |
| openpowerbot | [irc] <lkcl__> so please don't attempt to "share" anything | 20:55 |
| openpowerbot | [irc] <lkcl__> please split *every* class into "stages" - just like is done in the Div unit work, and in Mul. | 20:56 |
| openpowerbot | [irc] <programmerjake> sharing is part of my reason for not doing fsms, so if i'm doing pipelines no sharing need occur | 20:56 |
| openpowerbot | [irc] <lkcl__> actually, they could and most likely will | 20:56 |
| openpowerbot | [irc] <lkcl__> by - again - splitting out into "micro-coding" jobs | 20:57 |
| openpowerbot | [irc] <lkcl__> so there would be "partial results" called OP_FIRST_PHASE_OF_SOME_FPTASK and OP_FINAL_PHASE_OF_SOME_FPTASK | 20:57 |
| openpowerbot | [irc] <programmerjake> well, ttyl, meeting time | 20:57 |
| openpowerbot | [irc] <lkcl__> where *in between* that, the output from one pipeline gets passed to the Shift pipeline, with a notification "please drop this into OP_FINAL_PHASE_OF_SOME_FPTASK pipeline when you're done" | 20:58 |
| openpowerbot | [irc] <lkcl__> (!!!!!) | 20:58 |
| openpowerbot | [irc] <lkcl__> that's "micro-coding" | 20:58 |
| lkcl | octavius, ghostmansd[m] bi-weekly meeting | 21:01 |
| ghostmansd[m] | 1 sec, have to change the place to connect :-) | 21:16 |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@broadband-90-154-80-239.ip.moscow.rt.ru> has quit IRC | 21:23 | |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@176.59.55.211> has joined #libre-soc | 21:24 | |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@176.59.55.211> has quit IRC | 21:32 | |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@broadband-90-154-80-239.ip.moscow.rt.ru> has joined #libre-soc | 21:33 | |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@broadband-90-154-80-239.ip.moscow.rt.ru> has quit IRC | 22:06 | |
| *** ghostmansd[m] <ghostmansd[m]!~ghostmans@176.59.55.33> has joined #libre-soc | 22:06 | |
| *** octavius <octavius!~octavius@92.40.168.138.threembb.co.uk> has quit IRC | 22:12 | |
| lkcl | programmerjake: this is what the "difference" is between pipelines and FSMs (effectively nothing as far as the MultiCompUnits are concerned) https://git.libre-soc.org/?p=libreriscv.git;a=blob;f=3d_gpu/pipeline_vs_fsms.jpg;hb=HEAD | 23:34 |
| lkcl | the caveat is: if you want to avoid starvation of the ReservationStations you *must* have equal-or-greater RS entries (top and bottom) than you have EITHER pipeline-stages *OR* number-of-FSMs | 23:36 |
| lkcl | a Micro-Coded design would DEFINITELY have more RS#s than either-pipelinestages-or-num-FSMs | 23:37 |
| lkcl | and would "mark" the partial-result as "to be fed BACK into ANOTHER MultiCompUnit for additional processing" | 23:37 |
| openpowerbot | [irc] <programmerjake> yeah, i'm familiar with how pipelines vs fsms work in our core, iirc i've written several of them both | 23:37 |
| lkcl | but that is *not* anything to do with whether FSMs are used or pipelines are used | 23:38 |
| lkcl | my point is that it is a completely separate concept called "micro-coding" which i tried to implement in RSes about... oo... 3 years ago, and it went very badly :) | 23:38 |
| lkcl | but it turned out i had bugs in the ReservationStations class, hence i rewrote it and it became ReservationStations2 | 23:39 |
| lkcl | but for now, duplication (no sharing of *anything*) is perfectly fine as it's just too complex to design Micro-Coding within the available budget | 23:40 |
| openpowerbot | [irc] <programmerjake> in the case of fminmax, the pipeline is small enough that there's really no point trying to share hardware very much, it's basically a 64-bit integer compare, 2x nan-detection and some muxes | 23:40 |
| lkcl | yehyeh, the irony is that the "routing" ends up dwarfing the size of the stages | 23:41 |
| lkcl | especially if you get to say 18-long "processing" and have 18-long RSes, that's QTY 18x something mad like 192 wires | 23:41 |
| lkcl | which is just... mental. | 23:42 |
| openpowerbot | [irc] <programmerjake> well, i'll go fill out tasks now... | 23:42 |
| lkcl | :) | 23:42 |
| *** gnucode <gnucode!~gnucode@user/jab> has joined #libre-soc | 23:42 | |
| openpowerbot | [irc] <programmerjake> what fun...my computer wouldn't turn on until I unplugged it for 30s...must have tripped the over-power protection or something ... time to get a better power supply | 23:50 |
Generated by irclog2html.py 2.17.1 by Marius Gedminas - find it at https://mg.pov.lt/irclog2html/!
