IRC logs for #openrisc Tuesday, 2014-11-04

--- Log opened Tue Nov 04 00:00:13 2014
poke53282stekern: It is indeed the snoop hit in
poke53282At least for two cores and switching the core after every instruction.01:33
poke53282You know, when I first saw this snoop hit in the doc, I thought that I don't need it.01:34
poke53282And I wondered how this hit should behave if you access the 32 Bit word with an non 32 bit aligned 16 Bit write.01:35
poke53282I thought this would be undefined and no one would ever implement it this way. Well, I was wrong :)01:36
poke53282The first guy who comes implemented the locks using this behavior.01:38
poke53282Do you use any other snoop hits like this?01:39
mithroanyone here going to
stekernpoke53282: to be fair, I stole that particular piece of code from:
stekernmithro: I'm tempted, mostly considering how the weather start to look like here in Finland at the moment04:31
poke53282stekern: A solution which is based solely on the atomic instructions would be preferable. So, increasing the size of the spinlock to 2 32 Bit words would do it for example.04:39
poke53282Or implementing the restore with lwa and swa04:40
poke53282If the arm implementation really works the same way then it is a shame, that this is not documented properly.04:42
stekernI don't understand the problem, I think it's just using a natural property of how things are supposed to work? (well, I understand that it's problematic for you when you write the emulator, but apart from that ;))04:44
poke53282If this behavior would be solely part of the l.swa instruction I would be fine. But at the moment it looks like that every write to the memory has to be checked against n variables, with n the number of cores. This is a huge waste of energy just to win one or two cycles in the atomicity operations.04:49
poke53282By the way. 16 cores still don't work. So I still haven't figured out the main bug :)04:52
stekernwell, to be honest, I don't know if other code make use of this property as well. But that particular piece of code or not, you'll have to check all writes from other codes, that's kind of in the definition of an atomic rmw.04:54
poke53282I will think about it and check the speed penalty. Maybe I can get rid of this check with a few tricks. But first I need a stable emulation.05:04
poke53282I don't like the definition, but it might be necessary for security. Who knows. What happens for a DMA access?05:06
stekernit's not necessarily that dma accesses cause snoop hits, but implementations might05:27
stekernbut that's more related to cache coherency, not atomicity05:28
-!- Netsplit *.net <-> *.split quits: zama06:19
-!- Netsplit over, joins: zama06:20
mor1kx[mor1kx] bandvig pushed 1 new commit to withfpu:
mor1kxmor1kx/withfpu 602c45f Andrey Bacherov: Continue re-factoring for pipelined FPU: multiplier. Plus cosmetics for add/sub.07:55
wallento1stekern: can I find libpthread or1k-sysdeps for uClibc anywhere?09:58
stekernwallento1: there's only support for the old Linux threads in the openrisc port of uClibc10:03
wallento1okay, thanks. what is the problem with the new?10:04
stekernno other problem than that no-one done the work with NPTL in uClibc10:05
wallentookay, I will ask my student then10:05
stekernpersonally, I shifted interest towards musl (which have a good pthreads implementation)10:05
wallentoyes, we might do the same, but the fiasco OS somehow depends on uClibc10:07
wallentoneed to have a closer look10:07
stekernblueCmd_ has done the necessary NPTL work in glibc, so if you have someone doing the work, peeking at glibc first might be a good idea (and ask advice from blueCmd_ might be another good idea ;))10:07
wallentothanks! I found the nptl there, we will try to integrate this one10:12
olofkI also want students to do stuff for me11:22
olofkbtw stekern, I'm almost done with wb_sdram_ctrl now, and it's currently 8-14% faster now11:22
olofk(with my benchmark)11:22
olofkAre there any real benchmarks that are affected by memory speeds? Would be interesting to see that11:23
stekernolofk: you can try the membench I wrote11:40
stekernit's very naive, but gives some indication11:40
olofkbare-metal or linux?11:41
olofkstekern: It's too bloated11:50
stekernI agree, the memory usage is way over the top11:55
olofkWorks, but need some cleanup12:43
olofkHow do people name git tags to avoid getting them confused with branches? Just v1.0 ?12:47
stekernolofk: why didn't you build upon the latest version of wb_sdram_ctrl?13:18
stekern is already there13:19
stekern(but by franck)13:19
stekernI'll need to get my de0 nano back and give it a try, I've lent it to a guy at work13:31
olofkstekern: What the hell... I wonder why I missed that15:45
olofkah... now I remember. I wanted to merge the fixes that _franck_ did, so I checked out the version just before he forked it15:46
olofkThinking of how to merge this makes me want to cry15:46
olofkgit, I hate you so much15:46
olofkor actually... I'm not at all sure what I did15:52
stekernme neither, since you seem to have some of the new stuff manually merged in ;)15:54
stekernjust do a 'git pull -r' and fix the conflicts15:54
olofkOn which repo? :)15:54
stekernmy repo15:54
olofkAnd then force push it to my github repo?15:55
stekernyes, that you can do when you verified that you didn't screw up your changes when you fixed the conflicts ;)15:56
juliusbI presume a few of you here recieved that email from Jerry Harthcock about the GPU thing16:03
juliusbit's worth Googling his name, too, if you have a moment16:03
juliusban even more internet history :)16:03
juliusberr, even more *interesting* internet history16:03
juliusbThis is the best one, though:
juliusbalthough his GPU thing looks real, at least16:11
poke53282why he is talking about Android applications?16:20
poke53282Well, I can try to compile Dalvik :)16:21
poke53282But his steps make sense.16:31
wallento"Harthcock also wants to profit from importing alien technology by assembling a crack team of 10 professional psychics."16:32
wallento"Harthcock also wants to profit from importing alien technology by assembling a crack team of 10 professional OpenRISC developers."16:32
wallentoAll in16:32
wallentoIt actually sounds interesting16:32
wallentoalthough I am sceptical about OpenRISC in there16:33
poke53282The article is fun16:33
wallentoI think the large cores in the GPGPU need to be more powerful16:33
poke53282"Beginning in 1998, I have been formally trained to the advanced level (Stage-6) in Controlled Remote Viewing by ex-Defense Intelligence Agency military remote viewer"16:34
wallentonevertheless, it might be a chance to bring in stuff like out-of-order into mor1kx16:34
wallentoor multi-issue16:34
wallentomaybe the name is just some stealth mode name?16:35
wallentolike herbert schlangemann16:35
poke53282Offtopic: Take a look at the video.
poke53282I have a Oculus DK1 and know the potential. My next apartment must have an additional room for this ;)17:17
stekernjuliusb: how do you know that your google name searches hit the right guy?19:33
stekernI mean, I don't mind if you believe that I'm a 65 year old general in the Swedish military ;)19:35
olofkI'm a really old guitarist on google19:35
stekernolofk: but we look much younger in person!19:39
olofkAnd thinner :)19:39
rschmidlinHi there19:39
stekernhi rschmidlin19:39
olofkCan I just change remote "origin" to a different repo in .git/config to push it somewhere else? Or will that make the git gods angry?19:40
stekernolofk: at least we're still alive, in contrast to juliusb:
olofkYeah, I also use a lot19:41
olofkI can never remember where I put all my graves, so it's really helpful19:42
olofkstekern: I think I managed to sort out the repo situation now and force-pushed to olofk/wb_sdram_ctrl19:43
stekernI think you can, but you can also do 'git remote remove origin; git remote add origin <other origin>'19:43
olofkrschmidlin: Did you get my shared document invitation?19:44
rschmidlinyes, I did. I’m trying to put down some ideas.19:45
olofkPerfect. I'm out of ideas myself19:45
rschmidlinanyone want to join to input ideas?19:54
rschmidlinI have come up with 4 main goals for now19:55
olofk1. Start foundation19:56
olofk2. Collect money19:56
olofk3. ???19:56
olofk4. Profit19:56
rschmidlinI’m trying more to put the goals together first19:56
rschmidlinbut straight to point 4 works perfectly well too19:56
rschmidlinok, I think that’s what I have without further brainstorming.20:20
olofkI'm back now. Will read what you have written down20:34
olofkrschmidlin: Did you see my messages on google docs? It looks like you are just disconnecting and reconnecting all the time20:45
rschmidlinlet me see20:45
olofkstekern: Do I need IMMU and DMMU to run Linux?21:01
olofkrschmidlin: Yep. We're on a roll now :)21:02
stekernolofk: not sure if I should take that question seriously? =P21:10
rschmidlinit’s out21:10
olofkstekern: ah right. That was the whole point of uClinux21:11
stekernthe answer is yes, unless you add no-mmu support to openrisc Linux21:12
stekern...but I see that you didn't actually state that this was about openrisc in your question, so the answer is no.21:13
olofkYes. It's for one of the many other processors that I'm heavily involved with :)21:14
rschmidlinSee you guys, take care!21:16
--- Log closed Wed Nov 05 00:00:14 2014

Generated by 2.15.2 by Marius Gedminas - find it at!