23:14 < matches> And now I am basically adding random lines of code and running the tester and seeing if it still works :S
--- Day changed Fri Sep 12 2014
00:15 < matches> This reminds me of the James Micken comment about addition and multiplication basically being the same thing...
+01:19 < matches> Oh my goodness
+01:19 < matches> I think ParanoidNumber just worked for one very specific case!
+01:19 < matches> ParanoidNumber: 0.3 + 0.3 + 0.3 = 9/10, ToDouble() -> 0.9000000000000000222044604925031308084726
+01:20 < matches> double: 0.3 + 0.3 + 0.3 = 0.9000000000000001332267629550187848508358
+01:20 < matches> And strtod("0.9") = 0.9000000000000000222044604925031308084726
+01:23 < matches> For 0.2+0.2+0.3 = 7/10 it has an error of 1e-16 compared to doubles which apparently have no error
+01:24 < matches> But as floats it has an error of 1e-8 compared to 4e-8
+01:24 < matches> Hmm, why does 7.0 / 10.0 give a bigger error than 0.2 + 0.2 + 0.3
+01:24 < matches> Oh
+01:24 < matches> It'll be denormals probably
+01:25 < matches> That is... quiet irritating
+01:27 < matches> So... it is sometimes better than a float but only if there is a full moon and you sacrifice thrice the number of one tenth of unity goats three times
+01:28 < matches> The other times it breaks because subtraction is totally FITH
+01:29 < matches> Oh, it should be able to do 1/3 + 1/3 + 1/3 except my calculator tester doesn't allow me to enter non decimal numbers
+01:32 < matches> I think 0.3+0.3+0.3 is the only thing it can do better than regular floats
+01:32 < matches> Woo
+01:40 < matches> Ah it is better at beating doubles when it is using doubles as the digit type
+01:40 < matches> But it still is sometimes worse if you have a small number of operations
+01:40 < matches> Probably because one division is less precise than several additions
+01:42 < matches> I'm going to keep telling myself that this is useful
+01:42 < matches> After I sleep
+01:51 < Pommers> It isn't currently a full moon...
+01:51 < Pommers> Also
+01:51 < Pommers> Go to bed
+11:27 < matches> It was a full moon after I sacrificed the goats
+11:40 < Pommers> Stop toying with space matches
+--- Day changed Sat Sep 13 2014
+09:25 < unmercifulfish> I feel that the logs for this chat should just be added to qdb
+09:25 < unmercifulfish> < matches> It was a full moon after I sacrificed the goats
+12:36 < Pommers> Should make a QDB bot for submitting them
+16:20 < matches> When sulix is awarded the nobel prize for his work on quad trees, I will be proud to be a part of this channel
+16:21 < Pommers> Until then you won't be?
+16:23 < matches> I've been thinking of adding a "grep -v matches" before the automatic log commit
+17:45 < matches> I am a horrible person
+17:45 < matches> I am exploiting that "delete" acting on a NULL pointer has no effect
+17:46 < matches> "delete Operation(new ParanoidNumber(a), ADD)"
+17:47 < matches> Where Operation will either return the first argument or NULL
+17:50 < matches> At the moment I'm very pleased with how cleverly I am about to shoot myself in the foot
+17:50 < matches> We'll see how I feel after I pull the trigger
+17:53 < matches> I have either written an extremely elegant or an extremely segfaulting function
+17:57 < matches> The latter is looking more and more likely
+17:57 < matches> But I'm sure there is beauty underneath
+17:57 < matches> Like the ugly duckling!
+17:59 < matches> There's still hope, it turns out I'm segfaulting way before I even get to that
+18:01 < matches> Apparently 2+1 is 6, but doubles think 2+1 is 5
+18:02 < matches> "Freedom is the freedom to say 2+2=4, if that is granted, all else follows"
+18:06 < matches> Ok, 1 + 1 is 2
+18:06 < matches> I think we can call that a result
+18:06 < matches> Write the thesis and we're done here
+22:18 < matches> 1 + 0.3 = 13/10 and 13/10 + 0.2 = Segfault
+22:18 < matches> Getting close
+22:19 < matches> Now I'm doomed to not sleep until it works because I'm never going to remember how this works
+22:20 < Pommers> Got a whiteboard?
+22:20 < matches> I do actually but it's in the other room
+22:20 < matches> And it does not have space for this
+22:22 < matches> I don't think I've ever used so many delete operators...
+22:23 < matches> Also constructing a 4 diretional list on the stack then stealing one of the branches
+22:23 < matches> It's a game of "dodge the double delete!"
+22:24 < matches> Which reminds me, we have copious memory leaks in an external library
+22:24 < matches> 'ld' ?
+22:25 < matches> /lib/x86_64-linux-gnu/ld-2.19.so
+22:25 < matches> Not sure if that's something I'm doing wrong
+22:26 < matches> Turns out you can't pipe valgrind's output to less
+--- Day changed Sun Sep 14 2014
+18:23 < matches> http://szmoore.net/ipdf/sam/abstract.txt
+18:24 < matches> You know
+18:24 < matches> It sounds a lot more boring than it has been
+18:24 < matches> I also need a title
+18:24 < matches> "Number Representations and Precision in Vector Graphics" ?
+18:24 < matches> I'm too scared to pick something punny...
+18:26 < matches> Libreoffice thinks that is 147 words and pluma thinks it is 150...
+18:26 < matches> Oh
+18:26 < matches> hyphenated IEEE-754
+18:28 < matches> I think the first sentence could probably be less mind numbingly boring
+18:28 < matches> My last thesis started with "In this project" which seems terribly not academic
+18:29 < matches> (Every time I read that thesis I get a little more convinced that the markers didn't actually read any of it)
+18:29 < matches> ParanoidNumbers are not going well
+18:30 < matches> Like, it's hard to debug where things go wrong when you have a list like this:
+18:30 < matches> debug: main (tests/paranoidcalculator.cpp:75) - a is: {3*(23*(398123+(1/10+(2/100+(3/1000+(5/10000+(1/100000+(2/1e+06+(2/1e+07+(3/1e+08)))))))))+(3*(398123+(1/10+(2/100+(3/1000+(5/10000+(1/100000+(2/1e+06+(2/1e+07+(3/1e+08)))))))))/10+(4*(398123+(1/10+(2/100+(3/1000+(5/10000+(1/100000+(2/1e+06+(2/1e+07+(3/1e+08)))))))))/100+(1*(398123+(1/10+(2/100+(3/1000+(5/10000+(1/100000+(2/1e+06+(2/1e+07+(3/1e+08)))))))))/1000+(2*(398123+(1/10+(2/100+(3/1000+(5/10000+(1/100000+(2/1e+06+(2/1e+07+(3/1e+08)))))))))/10000+(3*(398123+(1/10+(2/100+(3/1000+(5/10000+(1/100000+(2/1e+06+(2/1e+07+(3/1e+08)))))))))/100000))))))/10+(1*(23*(398123+(1/10+(2/100+(3/1000+(5/10000+(1/100000+(2/1e+06+(2/1e+07+(3/1e+08)))))))))+(3*(398123+(1/10+(2/100+(3/1000+(5/10000+(1/1 ...
+18:31 < matches> That's without simplifying it
+18:31 < matches> But when it gives the wrong results before you even simplify it, that's bad
+18:32 < matches> I'm preeety sure it's not just giving the wrong results because it has a stupid number of terms/factors
+18:32 < matches> It's wrong enough that there's an extra factor or a factor missing or I don't even
+18:34 < matches> Hmm
+18:34 < matches> I want to add "Unlike paper, ..." before the third sentence
+18:34 < matches> I'm going to Captain Obvious
+18:38 < matches> :(
+18:38 < matches> Panic mode is well and truly activated now
+--- Day changed Mon Sep 15 2014
+14:00 < sulix> matches: You need to put the title in: "Keepin' it real: Staying afloat on the sea of document precision"
+--- Day changed Tue Sep 16 2014
+12:25 < matches> I think I'm sinking
+12:27 * sulix has failed to come up with a suitable sinking/syncing pun.
+12:27 < sulix> Or, for that matter, a reason why the quadtree clipping sometimes breaks.
+12:33 < matches> Hey, at least multiplication doesn't sometimes break
+12:33 < matches> That's fairly major
+12:34 < matches> My graph traversal for the paranoid number is somewhat flawed
+12:34 < matches> Yeah it's a giant graph
+12:34 < matches> Smell the memory cooking
+12:34 < matches> I should do some performance tests
+12:34 < matches> Even if I don't have anything to performance test...
+12:34 < matches> We did promise graphs
+12:35 < matches> (The other kind of graph)
+12:35 < sulix> How are you traversing the graph?
+12:35 < sulix> (The first kind)
+12:36 < matches> Originally I was alternating between multipying and adding, but that doesn't work if the first term in a factor has factors of its own
+12:36 < matches> I could probably still use that but I'd have to add zeroes or something
+12:37 < matches> Thinking of it as a linked list is wrong anyway, as factors can have terms and terms can have factors and I'm not making any sense am I
+12:37 < matches> Maths can be made up of maths
+12:37 < matches> I kind of regret basically promising to have this done in my abstract...
+12:37 < sulix> Hmmm...
+12:40 < matches> Anyway, either I am creating the graph wrong and traversing it correctly, or creating it right and traversing it incorrectly, or they are just both totally wrong
+12:41 < matches> Leaning towards the latter
+12:43 < sulix> You need to traverse it from the bottom-up, I think.
+12:50 < matches> Yeah, I need to change both the creation and traversal
+12:50 < matches> Repeated divisions need to change into a division of a number multiplied by a number
+12:51 < matches> Then the traversal is relatively trivial
+12:55 < matches> Division is annoying
+12:55 < matches> Why does maths need division
+12:55 < matches> All the other operations make sense
+12:56 < matches> Someone should reinvent mathematics without a concept of division
+12:56 < matches> We'd all be happier
+12:56 < sulix> I'm pretty certain there's a formal proof of that.
+12:57 < sulix> Most number systems don't have the concept of division.
+12:57 < matches> Haha
+12:57 < sulix> (The integers, for example)
+12:57 < matches> They have division with remainder though?
+12:58 < matches> Ah
+12:58 < matches> But you don't get a single number out
+12:58 < matches> You get two
+12:58 < matches> So it's not a proper operation?
+13:00 < sulix> It's not a binary operation, and it's not the inverse of division.
+13:00 < sulix> *multiplication
+13:07 < Pommers> Guild debate. Pretty sure I have the best camera angle here
+13:43 < matches> Guild debates are beyond the scope of this project
+13:44 < matches> Urgh I'm regretting distinguishing between ADD and SUBTRACT now
+13:44 < matches> Just ADD and MULTIPLY should be all the universe needs
+13:51 < matches> You have to be doing something wrong when you are using a triple pointer
+13:57 < matches> Delicious triple pointers
+13:57 < matches> Now I feel like a triple chocolate milkshake
+13:58 < matches> Ok, I have integers 0->9 adding and subtracting preeettty well now
+14:13 < sulix> Ah, the power of modern technology!
+14:29 < sulix> When in doubt, more debug output!
+14:55 < sulix> Ah ha! I have solve (read: hacked horribly around) the quadtree bug.
+14:56 < Pommers> http://xkcd.com/859/
+14:56 < sulix> Basically, adding a bunch of fake roots in fixes it.
+14:56 < sulix> Details at 11.
+14:56 < sulix> Pommers: You are a terrible person.
+14:57 < Pommers> Oh wait. It's there. I should sit closer to my screen
+15:32 < matches> Ok, I'm pretty sure Paranoid Numbers are the worst idea ever
+15:32 < matches> Possibly worse than iRRAM
+15:32 < matches> I mean, in theory you can simplify things
+15:32 < matches> But when you can't, you just use ridiculous amounts of memory
+15:33 < matches> I don't think you can simplify that many things either
+15:33 < matches> If I get it working enough I can demonstrate that it is a terrible idea at least...
+16:46 < matches> Oh my goodness it is kind of working
+16:46 < matches> I've said this before...
+16:47 < matches> But it's at least at the point where it gives suspiciously similar answers to the right answer
+16:48 < matches> As opposed to suspiciously off by factors of a random number
+17:31 < matches> So I think all I have to do now is simplify the damn things
+17:31 < matches> There are all sorts of amazing optimisations like not using a std::vector for the terms and not using seperate std::vector for + and -
+17:31 < matches> And I think you can probably optimise it even more if you just use doubles
+17:31 < matches> But
+17:31 < matches> It is better at repeated division/multiplications
+17:32 < matches> See TestMulDivIntegers
+17:32 < matches> By better I mean slower
+18:50 < matches> I have pushed a thing
+18:50 < matches> Try not to judge me too harshly
+18:51 < matches> Also there's no way it will compile on Cabellera anymore
+23:06 < matches> So
+23:06 < matches> I just made SVGs load into the centre of the view...
+23:06 < matches> And then I realised that is exactly the sort of thing that will break sulix's quad tree
+23:06 < matches> Sorry sulix
+23:07 < matches> My role in this project is basically to make sulix have to do more work...
+23:21 < Pommers> Why'd you break cabellera?
+23:30 < matches> I didn't break cabellera
+23:30 < matches> I used C++11 auto range iterators
+23:30 < matches> for (auto add : m_next[add])
+23:31 < matches> As opposed to: for (vector<ParanoidNumber*>::iterator i = m_next[ADD].begin(); i != m_next[ADD].end(); ++i) ParanoidNumber * add = *i; ...
+23:31 < matches> Cabellera's g++ is too old for such amazing things
+23:32 < matches> But I was at a point where I needed to change like 50 things from iterating over linked lists to iterating over some kind of hideous graph of std::vectors
+23:32 < matches> Stuff typing all that
+23:32 < matches> I broke the quad tree because I thought it would be nice to have SVGs automagically appear in the middle of the screen
+23:33 < matches> As opposed to the top left underneath all the debug output
+23:49 < sulix> matches: You are too kind.
+23:49 < sulix> The quadtrees are actually working fine at the moment, by virtue of me modifying SolveCubic to no longer solve cubics.
+23:51 < Pommers> return null;?
+--- Day changed Wed Sep 17 2014
+00:03 < matches> Wait what
+00:04 < matches> Did you remove Cubic Beziers
+00:04 < matches> Because I liked having Cubic Beziers
+00:04 < sulix> I'm still trying to work out why it works myself.
+00:04 < sulix> Nope, still cubic béziers, solvecubic now just returns some things which I'm pretty certain actually aren't roots as well.
+00:05 < Pommers> Do either of you want to trade? I'll work on iPDF and you can fix my housemates computer
+00:05 < matches> Hah
+00:05 < matches> Set up that Jenkins server and we'll talk
+00:06 < Pommers> The Container exists for Jenkins
+00:06 < Pommers> So it's been started
+00:06 < matches> I wish CS would stop thinking up names for things that are also names for other things
+00:06 < matches> Container
+00:06 < matches> Puppet
+00:06 < matches> Orchestration
+00:06 < matches> For god's sake
+00:06 < matches> Paranoid Number is totally a legit name
+00:07 < Pommers> Stop being paranoid about things matches
+00:07 < matches> "Tree" is acceptable because it at least looks like a tree
+00:07 < Pommers> Also, coming to dinner?
+00:07 < matches> There is dinner?
+00:07 < matches> When?
+00:07 < matches> It's midnight...
+00:07 < Pommers> Thursday
+00:07 < matches> Oh
+00:07 < matches> I suppose
+00:08 < matches> Thursday is typically the post meeting freak the hell out day
+00:08 < matches> Monday Tuesday and Wednesday being the pre meeting freak the hell out days
+00:08 < matches> Friday is the I should really prepare for my Physics tute day
+00:09 < Pommers> Saturday is "o god coderdojo day"
+00:09 < matches> Yeah
+00:09 < matches> Sunday is I should really not be playing computer games but I will anyway day
+00:09 < matches> No wait last sunday was a freak the hell out day as well, until they made me play monopoly
+00:09 < matches> Sorry, monotony
+00:10 < matches> Anyway how do you get valgrind to print number of flops
+00:10 < sulix> I'm just s/.*day/PANIC day/g at the moment.
+00:10 < matches> Your thing works!
+00:10 < matches> Sort of
+00:11 < sulix> thing?
+00:11 < matches> Quadtrees
+00:11 < Pommers> "thing"
+00:11 < sulix> Technically it works*
+00:11 < Pommers> "it"
+00:11 < matches> I have about 4 things I still haven't even tried yet
+00:11 < matches> And a couple of things that will work if you wait the lifetime of the universe
+00:12 < Pommers> Hmm... I should make my bed while this backup happens
+00:12 < matches> Bed is tempting
+00:12 < sulix> *The roots I'm using might not actually be roots, it doesn't work if you either pan, zoom out or anything other than zoom in on a single point forever.
+00:12 < Pommers> But having too much fun coding?
+00:13 < matches> I'm working towards actually making graphs of something
+00:13 < matches> I'm still working out exactly what I will graph
+00:13 < matches> flops seems important
+00:13 < matches> I'm basically replacing a single inexact flop with a lot of usually but not always exact flops
+00:14 < matches> And several hundred MB of memory...
+00:18 < matches> Ok, this can't be right
+00:18 < matches> To render one frame of an empty document takes 137,708 F64 operations
+00:18 < matches> I assume that means double
+00:19 < Pommers> That could use some optimizing...
+00:19 < Pommers> Make all the functions return null and see how that speeds it up
+00:19 < matches> -_-
+00:19 < matches> This is using regular doubles
+00:19 < matches> Not Paranoid Numbers
+00:20 < matches> ParanoidNumbers takes about 5 minutes to do 20000 operations at the moment...
+00:20 < sulix> I could believe it.
+00:22 < Pommers> I'd be paranoid if it took that long
+00:22 < matches> Don't make fun of the Paranoid Numbers
+00:22 < matches> They are a work of art
+00:23 < sulix> Part of me really wants to try compiling the quadtree with them, and part of me is terrified at the prospect.
+00:23 < matches> There isn't a REALTYPE for them yet
+00:25 < matches> They might work better if you had a giant Trie of them but I have no idea how to structure that
+00:25 < matches> At the moment there is one tree per number and the tree can have a variable number of children for each node... :S
+00:31 < matches> Anyway what performance metric do we want
+00:31 * matches consults the proposal
+00:32 < Pommers> What will be easier to do? Time or space?
+00:32 < matches> Time
+00:32 < matches> But Rowan was kind of keen on Space
+00:32 < matches> Number of flops is probably the best
+00:32 < Pommers> Just put some ascii art daleks in
+00:34 < matches> Ok, so 1 frame = 30,897 F32, 129,042 F64 and 1000 frames = 4,364,459 F32, 6,933,526 F64
+00:35 < matches> 6800 doubles to render a blank document
+00:35 < matches> Actually
+00:35 < matches> !
+00:35 < matches> There is a rectangle in it!
+00:35 < matches> Of width and height 0
+00:40 < Pommers> I almost have the container for jenkins fully set up now
+00:43 < Pommers> I just need to give it an IP and then just set up jenkins
+00:45 < matches> I was joking about Jenkins...
+00:46 < Pommers> I actually have a use for it...
+01:16 < Pommers> Housemates computer is fixed
+01:16 * Pommers sleeps
+01:29 < matches> Dammit
+01:29 < matches> I entered a timestamp of 10:56 instead of 22:56 and now my commit is at 13:27
+01:30 < matches> I tried git commit --amend --date= but it didn't quite work
+01:31 < sulix> Would it count as cheating on the performance metrics if time travel was involved?
+01:31 < matches> Nooo
+01:32 < sulix> As you can see, ParanoidNumber actually takes only -12 hours to render a frame by being so cautious it violates causality.
+01:32 < matches> So, the "spatial" metric can be: Number of pixels not the same, and mean minimum distance horizontally/vertically to the nearest set pixel in the original
+01:32 < matches> Because I said so
+01:32 < matches> A possible other metric would be the Fourier Transform
+01:33 < matches> But that's stupid
+01:33 < matches> -12 hours is pretty good
+14:43 < Pommers> Jenkins is set up and I can let you create an account if you want to play with it
+17:01 < matches> Best idea ever
+17:02 < matches> A performance test that saves the graph as an svg
+17:02 < matches> Then runs itself on the graph...
+17:05 < Pommers> recursive testing?
+17:05 < matches> Yes!
+17:07 < Pommers> Set your testing up in jenkins
+17:08 < matches> This is performance tests not sanity tests
+17:08 < matches> Or whatever they're called
+17:08 < matches> We don't need to make sure things work...
+17:08 < matches> We just need them to not work fastly
+18:35 < matches> SDL_WINDOW_HIDDEN is useful
+18:36 < matches> Actually I'm surprised it works
+18:37 < matches> As much as having ipdf constantly flashing up and stealing my window focus is exciting, I am relieved to be able to stop that
+18:37 < matches> Now I just need to stop the monitor from randomly flashing its menu and/or turning off...
+18:42 < matches> So I can show some graphs tomorrow, although they may not be graphs of actually useful things
+19:44 < matches> I hope you like ipython because we're using ipython
+19:50 < matches> git.ucc seems sick...
git.ucc.asn.au / ipdf / documents.git / blobdiff
UCC git Repository :: git.ucc.asn.au