#qi-hardware IRC log for Wednesday, 2012-06-20

mthNO_HZ only applies when the system is idle, not with one task running00:00
DocScrutinizer06U might e right00:00
DocScrutinizer06o/00:01
mthAyla: can the USB controller be suspended while the freq change happens?00:02
mthlike we do for the MSC00:02
DocScrutinizer06if you wanna drive this ahead, w/o me playing advocado diabolo(?), provide CPU cycles breakdowns, for the IRQ handler invoked every other time, vs the one task running. Give a ratio00:03
AylaI don't see a reason why it couldn't00:03
DocScrutinizer06Ayla: suspending USB is a friggin complicated and annoying ting to do00:06
DocScrutinizer06and takes tiiiiime00:06
Aylahow much time?00:07
DocScrutinizer06but depending on controller you don't need to suspend00:07
DocScrutinizer06(howm much) milliseconds00:07
DocScrutinizer06several00:07
DocScrutinizer06you wanna switch clock so fast it "doesn't even matter"00:08
DocScrutinizer06ideally clock of usb uart is independent00:08
Aylaideally.00:09
mthAyla: I don't think we should be doing PLL changes on demand anyway00:09
mthonly divider changes00:09
DocScrutinizer06everything after uart is basically not related to clock00:09
mthand then only the CPU divider has to be changed00:09
DocScrutinizer06mth: +++00:09
Aylayes00:09
DocScrutinizer06thats the way00:10
DocScrutinizer06you need to consider other timeframe based hw IF too00:10
DocScrutinizer06like rs23200:11
Aylabut that differs from how the "ondemand" governor works00:11
DocScrutinizer06basically all async hw if00:11
mthwe can write our own governor00:11
mthI'm not even sure how much is gained by running at a lower freq00:11
mthvs running fast and then being idle for a while00:11
DocScrutinizer06the trick is to provide an API that changes CPU clock *only*00:12
DocScrutinizer06mth: nuttin00:12
DocScrutinizer06speed to idle00:12
Aylawe'd gain a couple of milliwatts00:12
DocScrutinizer06of course depending on overhead00:12
DocScrutinizer06for switching00:12
mthAyla: how did you measure that?00:13
Aylawith a multimeter :p00:13
DocScrutinizer06~listvalues to idle00:13
mthI mean, how did you generate the load?00:13
infobotFactoid search of 'to idle' by value returned no results.00:13
DocScrutinizer06~listvalues idle00:13
infobotFactoid search of 'idle' by value (18 of 52): /proc/uptime ;; 50 fps ;; atrap a285 ;; autolog ;; cmd: grope (.*?) ;; condor ;; disk striping with parity ;; efnet server ;; gpppkill ;; grope ;; hlt ;; idle ;; jamuraa ;; jargon finger ;; killer ;; loll ;; note: the point of the game ;; now kirma's task-bootstrap upload.00:13
Aylausing "yes" IIRC00:14
mthbut that generated an unlimited load00:14
DocScrutinizer06meh00:14
mthin which case an ondemand-style governor won't scale down00:14
DocScrutinizer06shouldn't00:15
AylaI didn't measure with the ondemand governor00:15
mthyou'd need a load of about 30% or so to get a useful measurement00:15
DocScrutinizer06calculate prime factors of 8744 every hsync00:15
mthlet's say you compute primes, then computing all primes will simply run faster on a higher clock, so you do more work while consuming more energy00:15
DocScrutinizer06err vsync00:15
DocScrutinizer06;-)00:15
mthcomputing a fixed number of primes per second would be a fair load to compare on00:16
DocScrutinizer06yep00:16
AylaI'd say the latter is better00:16
mthso you'd get to see the effect of a slow clock being busy 60% of the time vs a fast clock busy 30% of the time00:16
DocScrutinizer06rush to idle 00:16
DocScrutinizer06is it00:16
DocScrutinizer06usually00:16
Aylajudging from the general performance gaps between frequencies, and http://wiki.dingoonity.org/index.php?title=OpenDingux:Development#Power_consumption_measurements00:17
DocScrutinizer06unless idle is <10%00:17
mthin PC overclocking, lower freqs save power because the high freqs need a higher voltage supplied to the CPU00:17
mthbut we're not changing any voltages00:17
DocScrutinizer06where "10" is a guestimatedepending on overhead o switch clock speeds00:17
mththat is with the PLL at 34 vs 408 MHz though00:18
DocScrutinizer06idiots at #maemo do OC with UNDERvolting ;-D00:18
mthcould you also measure the PLL at 408 MHz and a high CPU divider?00:18
mthDocScrutinizer06: well, if overheating is the enemy and not instability, it might work00:19
Aylahow could I force a high divider?00:19
whitequarkDocScrutinizer06: on jz4740 everything is bound to master clock00:19
mthhardcode it in the cpufreq driver is the easiest way, I guess00:19
whitequarkhence you cannot easily scale cpufreq if you want your peripherals to work00:19
whitequarklcd, usb, ram... everything00:19
DocScrutinizer06overheating is 3rd or 5th concern though00:19
whitequarkexceptionally stupid design imo00:20
DocScrutinizer06whitequark: we had same on S3C2410 iirc00:20
mthyes, you can only scale cpufreq drastically, like 1/3 or 1/2 of the PLL freq,  but not 80%00:20
DocScrutinizer06on s2c2442 iirc it been decoupled00:20
whitequarks3c? sounds samsung, hmm00:20
DocScrutinizer06samsung S3C, Neo00:21
DocScrutinizer06O00:21
mthI even had to patch the cpufreq infrastructure to get this PLL-changing approach to work at all00:21
DocScrutinizer06M00:21
whitequarkyeah, already found00:21
DocScrutinizer06OM even00:21
whitequarkmth: wasn't there also some other issue, like cpu clock being bound to ram/bus/dma clock?00:22
whitequarkwhere you could theoretically scale it 1/3 or 1/2, but it won't work in practice due to bugs or some other issues00:22
DocScrutinizer06RAM bus, sure00:22
mththere are various restrictions about the dividers00:22
whitequarkI recall reading the datasheet and thinking that one cannot really do *any* cpufreq scaling on that chip00:22
mthalso, there is the case where raising the PLL freq means you have to increase the memory divider and you end up with a slower system overall00:22
DocScrutinizer06that's why many devices don't run RAM at max speed00:23
mthsince slower SDRAM has more impact than a slightly faster CPU00:23
DocScrutinizer06incl gta02 iirc00:23
whitequarkyeah00:23
whitequarkjz4740 divider system isn't suited to scaling at all00:24
mthwe should be testing CL=2 vs CL=3 for the SDRAM, it's possible CL=3 aligns better with the high PLL values we'd like for the CPU00:24
whitequarkit's a set-and-forget system00:24
whitequarkI'd say that proper wait states would give more in that case00:24
mthexcept that our users like to overclock a bit if it means getting 60 fps instead of 45 fps from an emulator or game00:24
whitequarkwell, it00:25
whitequark's pretty much the same set-and-forget00:25
mthCL is a kind of wait state, isn't it?00:25
whitequarkmhm... CL?00:25
mthcas latency00:25
whitequarkah, yes00:25
whitequarkwell, I'm not sure how exactly it affects power consumption00:26
Aylait does align better with the CPU, that's for sure00:26
whitequarkwith static circuits it's the count of cycles which matters, and (I hope) the CPU is pipelined anyway00:26
mthprobably CL=2 is better for that; fewer clock pulses can only be good00:26
mthbut I don't know if the difference is significant or not00:26
Aylacan't we reconfigure the RAM on demand?00:27
whitequarkpretty perversive thing to do00:27
mthwe can write to its config regs at any time afaik, but I don't know if it will have side effects00:27
whitequarkI'd say that it will introduce general system instability through transition effects00:28
whitequarki.e. if the cpu will decide it's time to fill its cache exactly at the time another cycle would be inserted00:28
whitequarkand we don't know much about 4740 internals00:29
mthAyla: see it as a selling point for ubiboot ;)00:29
Aylait wouldn't be hard to patch u-boot00:29
whitequarku-boot entirely consists of patches, no? :)00:29
whitequarkof the vendor quality, mostly00:30
mthwith jzboot you can simply set the CL value in the config file00:30
Aylaand currently ubiboot sets CL=2 too00:30
mthI don't know what u-boot consists of exactly, I got completely lost in its sources00:30
whitequarkpure shit meters deep00:31
whitequarkjust like C is a pdp-11 assembler dreaming it's a programming language, u-boot is a yet another secondary bootloader which thinks it's an OS00:32
whitequarkshut up and load the damned sectors!00:34
whitequarkor pages00:34
whitequarkor whatever unit your non-volatile storage has00:34
whitequarkfucking insomnia.00:35
Aylawithout insomnias, I wouldn't go very fast on my projects ;)00:36
mthnot for me: I still sleep 7 hours, just weird hours00:37
mthso there is no time gained00:37
mthand quite a lot of time lost if I go to bed and don't fall asleep00:37
whitequarkmth: yeah, I understand you well00:37
whitequarkif I'm let to myself I'd go to bed at 8AM and wake up after 7-12 hours00:38
whitequarkbut unfortunately every other person in this world lives in exactly invert intervals00:38
Aylahow can you sleep 7 hours when you go to sleep at 5am? You don't go to work? :)00:38
mthif you work at home, you can pick any hours of the day you like00:39
whitequarkyou can have different kinds of work :) but yes, I don't work in plain old office00:39
Aylaah, ok00:39
whitequarkwell, sometimes I do, when all else fails to give profit :)00:39
whitequarknot quite often through00:39
Aylado you enjoy working at home?00:39
mthyes and no00:40
AylaI don't think I would like it, I'd prefer to meet some other people from time to time00:40
mthwhat I miss most is a "sparring partner" to discuss design ideas00:40
whitequarkI generally live with my own projects, whatever they are00:40
whitequarkthrough working at home means more procrastination00:40
mthnot necessarily in the same room, but more than one person on a project00:40
whitequarkmth: +100000:41
whitequarkpublic irc channels often help, through00:41
mthI do meet other people almost every day, but they're not colleagues00:41
whitequarkalso, when I do work at the daytime, I often do that in a cafe00:42
whitequarkprovides a nice background noise00:42
DocScrutinizer06duh, shutup and don't awake me :-P00:42
Action: DocScrutinizer06 scrolls back to see how loing he slept00:42
whitequarkDocScrutinizer06: when I can't sleep, no one can :D00:42
Aylawhitequark: hey, nice00:42
DocScrutinizer06wow, 15min00:42
mthI use a different computer with no mail or IRC or anything else distracting running00:43
whitequarkDocScrutinizer06: do you sleep with your n900? :)00:43
mthand put on some music that I've listened to before, so it doesn't draw much attention00:43
DocScrutinizer06nah, we got separate beds00:43
DocScrutinizer06this is lenovo t500 though00:43
mthI have done meetings in a cafe though00:44
whitequarkmth: I just run irc in irssi. it doesn't capture my attention by its own then, same for email. And when I need to get distracted, I just switch desktops00:44
whitequarkalso, yes, music00:44
mthin theory my mail doesn't have to distract me, but if I see unread mail I want to check it, even though I know there is no reason to00:44
mthbut without a mail counter I actually check it more often00:44
whitequarkI had some of my most productive hours when I were sleep deprived and listening to the same track for 200th or so time00:44
whitequarker, I was00:45
Action: DocScrutinizer06 killed mailcounter systray icon, whe the 5 digits didn't fit anymore00:45
whitequarke.g., yesterday was the first time I truly lead one of my major projects to 1.0.0 point00:46
whitequarkat the last 8h I did more than in previous four days00:46
mthI excluded most mailinglists from the counter, it's more like an inbox counter now00:46
DocScrutinizer06I googled my name yesterday, amazing experience00:46
DocScrutinizer06esp since there are still others like me who do firebrigade, gardening...00:47
DocScrutinizer06and one even does linux kernel (NIC et al)00:47
DocScrutinizer06dunno how to exclude dirs from kmail counter00:48
Action: whitequark never had mailcounter. too lazy to set up00:48
whitequarkthrough I have kind of a mailcounter on my phone, when it's on wifi00:48
mthDocScrutinizer06: I'm using kmail as well, but I can't find the option anymore00:50
mthI think it used to be in the context menu for a mail folder00:50
whitequarkDocScrutinizer06: http://talk.maemo.org/member.php?s=78b7634e2cefca7edf93bb87f77ef9c7&u=35572 is that your photo on the right? :)00:51
mthby the way, I've had major problems after switching to akonadi storage, so if you haven't done so yet, I'd recommend not doing that for a long time00:51
mthI once programmed a KDE2 panel applet to show new mail status (no count, just a flag)00:58
mthbefore that became standard functionality00:58
AylaKDE 2? :)01:03
Aylasounds like prehistoric01:03
mthI started with KDE 1.1, I think01:25
DocScrutinizer06whitequark: guess ;-)01:35
DocScrutinizer06http://talk.maemo.org/member.php?u=3557201:38
Action: DocScrutinizer06 wonders what that "s=<randome large numer>" came from01:39
DocScrutinizer06yeah and kde3 was so good kde4 is still struggling to catch up01:43
DocScrutinizer06wtf is akonadi?01:44
mthit's some kind of storage backend separate from the mail reader itself01:52
mthnot a bad idea in itself, imo, but the implementation is very immature01:52
qi-bot[commit] Xiangfu: milkymist: files: automatic mount exroot at /dev/sda1 (master) http://qi-hw.com/p/openwrt-packages/2fbc9e803:32
qi-bot[commit] Xiangfu: urjtag: minor fix, create usr/lib folder (master) http://qi-hw.com/p/openwrt-packages/935266e03:40
qi-botThe build was successful: http://fidelio.qi-hardware.com/~xiangfu/build-nanonote/openwrt-xburst.minimal-20120619-1445 17:00
whitequarkLOL17:29
whitequarkSGS2 RIL has a SecSetIMEI call.17:30
whitequarkdoes what it says on the box.17:30
whitequarkDocScrutinizer06: shouldn't radios be certifier or something like that? ;)17:31
whitequark*certified17:31
wpwrakmaybe it doesn't work ? :)18:17
wpwrake.g., on an RSX-11 system (RSX-11 was relatively VMS-like a PDP-11 operating system) i was once shown not only SET PASSWORD, but also a command SHOW PASSWORD. not sure if it came like that from DEC or is someone had added it a s a joke. the output suggested the latter :)18:30
larscwere you able to run it without having to enter the password first?18:41
wpwrakoh yes. it output several diagnostics. accessing passowrd database, then some problem, ... then it suggested to try to decode it for you, worked on it for a while, gave up,18:51
wpwrakthen it showed a world map, "war games" like. wrote some more stuff. until you eventually got a final fatal error.18:52
larschehe :)19:08
viricmh these loongson fpu things run super bad.19:19
viricffmpeg fails, lame fails, libvorbis fails...19:20
whitequarkwpwrak: it has been verified to work correctly19:59
whitequarkthat's the worst part19:59
wpwrakperhaps they'll try to make it less obvious the next time :)20:01
kristianpaulcool cisco routers have altera, lattice fpga's and cpld20:18
kristianpaultheir main chip seems to be self brandend..20:18
kristianpaulwonder what hapennded with that network procesor core announced on open cores time ago..20:20
wpwrakanyone knows a nice palindrome made of up 6-10 left-right symmetrical upper-case characters ? (AHIMOTUVWXY)20:29
whitequarkwpwrak: grep thinks there are no such words in /usr/share/dict/american-english20:34
whitequarkthe best match was20:34
whitequark1.9.3-p194 :012 > words.grep /([AHIMOTUVWXY])([AHIMOTUVWXY])([AHIMOTUVWXY])\3\2\1/i => ["Chattahoochee\n", "Chattahoochee's\n"] 20:34
wpwraknaw, thre must be some. e.g., "aha", "tit", ...20:35
whitequarkyou said 6-1020:35
wpwrakah, you already took that into account20:35
whitequarkyup20:36
wpwrakdidn't find any in the usual palindrome lists either. one with six characters was listed ("MAXXAM") but that seems to be a name20:38
wpwrakwell. it'll have to do for testing. thanks !20:42
--- Thu Jun 21 201200:00

Generated by irclog2html.py 2.9.2 by Marius Gedminas - find it at mg.pov.lt!