Programming puzzles: processing lists!
|
06-12-2017, 07:46 PM
(This post was last modified: 06-13-2017 10:24 AM by pier4r.)
Post: #141
|
|||
|
|||
RE: Programming puzzles: processing lists!
Neat addition. And yes the revlist works.
So I got some surprising experiences. First the FOR is not as slow as I though compared to DOSUBS / DOLIST (I though was like 10 times slower). 10 times iterating over a list of 100 elements. dosusbs: avg 0.48 s dolist: avg 0.50 s for get: avg 0.93 s (nice pun) start geti: avg 1.19 s start headList: avg 0.42 s (surprise! same speed category of dosusbs/dolist) . wrong, the code had a bug. start headList: avg 1.66 s start DUP TAIL SWAP HEAD: avg 0.96 s start lrot: avg 3.50 s (I had hopes on this, but likely rotating a list is not that easy) So if exploding a list with userRPL gets a speed (maybe only for 100 elements, I need to test with 1000) comparable to dolist/dosubs and way better than the compact TAIL + HEAD (those may be few commands but I beleive are intensive), I was thinking that exploding the list and handling it with sysRPL would be super fast, like 3 times faster. Also I was thinking that LROT would have been faster, maybe is the DUP + HEAD operations that makes it slow. if the "headTail" procedure holds its speed, it is a great replacement for dosubs/dolists for iterations, at least in userRPL. Also using a for is not as bad as I thought. Going for 1000 elements now. Of course, if you see other ways (debug friendly) to iterate over a list, please tell me! Code:
Wikis are great, Contribute :) |
|||
06-12-2017, 09:41 PM
Post: #142
|
|||
|
|||
RE: Programming puzzles: processing lists!
(06-12-2017 07:46 PM)pier4r Wrote: So I got some surprising experiences. First the FOR is not as slow as I though compared to DOSUBS / DOLIST (I though was like 10 times slower). I'm a little surprised that the GET/GETI commands did that well, but I'd guess you'd start to see them slow down a bit as the list sizes grow. LROT is actually fairly simple, but it does take some time. It basically determines the two sublists that need to be swapped based on the given parameter, makes the sublists, then reassembles them. I might be able to speed it up some with a slightly different approach, but I'm guessing it won't be a significant difference in the final timings. I should add this to my list of things to check, though. The bulk of the time in the LROT loop is probably from LROT itself. DUP is very fast (it simply copies a pointer, not the actual data), and HEAD is pretty fast as well. My testing on lists of 100 items showed an average timing of .0336 seconds, so a loop of 100 iterations taking 3.36 seconds just for LROT alone would be expected. The remainder of the time for the other commands (and loop overhead) seems reasonable to me. The internal representation of a list isn't very sophisticated. It's basically a prologue token, each object's data one after the other, and an epilogue token. So any commands that access individual objects within a list have to (literally) traverse/skip each and every object before the one in question in order to access it. In the case of embedded sublists, each element in them also has to be traversed. There's no way to skip over entire embedded sublists, as lists don't contain size information. Even getting the length of a list requires traversing all contained elements. So you can see where the longer a list is, the more time is required to deal with the list's elements. HEAD is relatively fast because only the first element has to be accessed. TAIL can quickly skip the first element, but it still has to make a copy of the remainder of the list. So it's more sensitive to list size. I haven't yet traced through DOSUBS and DOLIST to see how they work, but I suspect they do something similar to STREAM (which I have looked at) that allows relatively quick access to list contents by isolating one element at a time in the runstream. That method isn't available to UserRPL programs, as it requires manipulation of the return stack -- a technique that has to be executed very carefully to avoid blowing everything up. |
|||
06-12-2017, 09:51 PM
(This post was last modified: 06-12-2017 09:52 PM by pier4r.)
Post: #143
|
|||
|
|||
RE: Programming puzzles: processing lists!
For lrot I thought you exploded the list and rolled the elements, as I do in user rpl.
This should be faster I think. And thanks for sharing insights! Wikis are great, Contribute :) |
|||
06-12-2017, 10:19 PM
Post: #144
|
|||
|
|||
RE: Programming puzzles: processing lists!
(06-12-2017 09:51 PM)pier4r Wrote: For lrot I thought you exploded the list and rolled the elements, as I do in user rpl. Explode/roll/implode would work ok for an argument of 1/-1, but you'd have to loop the rolls for other numbers. Once list sizes get into the hundreds, rolls start to slow down. Try the "roll" method on a list of 1000 items rotated by 500 (or -500). How does it do? LROT handles that in about .19 seconds. The method I want to test is similar to what I'm currently doing, but would involve loading the elements onto the stack from two different streams and then imploding the whole list. It may not actually be any faster than splitting a list into two pieces and reassembling them, but it's worth a try. |
|||
06-12-2017, 10:26 PM
(This post was last modified: 06-13-2017 08:16 AM by pier4r.)
Post: #145
|
|||
|
|||
RE: Programming puzzles: processing lists!
(06-12-2017 10:19 PM)DavidM Wrote: Explode/roll/implode would work ok for an argument of 1/-1, but you'd have to loop the rolls for other numbers. Once list sizes get into the hundreds, rolls start to slow down. Try the "roll" method on a list of 1000 items rotated by 500 (or -500). How does it do? LROT handles that in about .19 seconds. Understood, I thought it was always quite quick. Then Lrot seems to be effective on a big chunk of rotations but not many little ones. Good to know. The test with lists of 1000 elements is still running, over an hour now. I'll be interested in the results. Wikis are great, Contribute :) |
|||
06-13-2017, 12:43 AM
Post: #146
|
|||
|
|||
RE: Programming puzzles: processing lists!
(06-12-2017 10:26 PM)pier4r Wrote: The test with lists of 1000 elements is still running, over an hour now. I ran a single cycle with 1000 list elements, and got these results: "dosubs" { 7.8397 } "dolist" { 7.7578 } "for get" { 34.002 } "start geti" { 54.728 } "start headList" { 535.6086 } "start tail head" { 33.1366 } "start lrot" { 4244.4562 } The LROT loop is most likely a victim of multiple garbage collection events, as the expected time based on previous timings would have been closer to 385 seconds. Garbage Collections can be made worse by using certain techniques, though I must confess the reasons aren't always obvious. It could be that simply splitting a large list into two sublists is one of those techniques that causes problems. It doesn't make sense to me that it would be any worse than exploding everything then imploding it back, but the evidence speaks for itself! I'll continue to test some other methods for LROT. I may also "special case" it to use a roll/unroll if the argument is 1 or -1, since that would almost certainly be better than creating two lists to join when one of them has only one element. |
|||
06-13-2017, 08:23 AM
Post: #147
|
|||
|
|||
RE: Programming puzzles: processing lists!
(06-13-2017 12:43 AM)DavidM Wrote: I ran a single cycle with 1000 list elements, and got these results: You saved me. The night is passed but the 50g is still working. Based on your measurements I would expect 13.5 hours for 10 cycles. Impressive how headList dropped his efficiency and started to be not linear (so exploding a list does not help too) and impressive for + get. I remember now I tested it with 1000 elements and I was getting like 35 seconds. Even for it is not linear (one would expect 9 seconds, 10 times 0.9, not 35) but still only "5" times slower than the DOx commands. Also tail+head very quick. Quote:The LROT loop is most likely a victim of multiple garbage collection events, as the expected time based on previous timings would have been closer to 385 seconds. Yes I guess special cases of an "handful" of elements rotated (I do not know a meaningful threshold, but maybe up to +10 , -10 ?) would make LROT really effective even for multiple calls. Wikis are great, Contribute :) |
|||
06-13-2017, 10:06 AM
(This post was last modified: 06-13-2017 11:49 AM by pier4r.)
Post: #148
|
|||
|
|||
RE: Programming puzzles: processing lists!
In the meanwhile my 50g finished. Those are the results for 10 iterations over lists of 1000 elements (the same as the one from David, just to confirm the results. Wait not, there is one that does not fit)
"dosubs" avg { 7.85 } "dolist" avg { 7.76 } "for get" avg { 34.08 } "start geti" avg { 54.86 } wrong: "start headList" avg { 4.12 } (instead of 535 s. What's going on here? Surely replicating the test in different setups helps) "start headList" avg { likely 535 as David tested } I found a bug and I checked the code for a list of 100 elements, it returns 1.66s on average (I corrected the older post). So likely it uses 530 seconds or more on 1000 elements. I do not have the time, at the moment, to let it run. "start tail head" avg { 33.38 } "start lrot" avg { 4867.04 } I'll have to double check the headList result because it seems to good to be true. Moreover david reports 535 seconds, mine 4 seconds. Edit: checked, indeed there was a bug. Anyway, all in all, the for + get is quite feasible as subsitute for nested iterations of dolist/dosubs Wikis are great, Contribute :) |
|||
06-13-2017, 06:23 PM
Post: #149
|
|||
|
|||
RE: Programming puzzles: processing lists!
(06-13-2017 10:06 AM)pier4r Wrote: ...Anyway, all in all, the for + get is quite feasible as subsitute for nested iterations of dolist/dosubs After spending more time than I care to admit trying alternative methods for LROT, I've come to the conclusion that any method that involves breaking the list apart into its constituent elements will eventually lead to a heart-stopping garbage collection if the function is called repeatedly in succession with a large-ish list as its argument. I was able to get it to the point where they were a bit less often, but it only delayed the inevitable. So I've decided to try going the route I originally wanted to go in the first place, but thought might not be worth the hassle: I'll recreate the list by copying the two blocks of memory containing the reversed parts into a new list object directly. It will require a bit more setup (and a lot more code), but it should be a relatively fast option that is nicer to the overall memory footprint of the machine. It will be a good exercise that may have other benefits as well for the other functions I'm contemplating. Regarding the GET performance, it occurs to me that I should probably temper my criticism of its performance. GET is better than PUT, owing to the fact that it's relatively quick to skip a series of objects when you don't actually have to explode them. But I still maintain that if you know you need to manipulate more than a couple of items in a list (or array, though they are a bit more efficient), it's better to break them apart and then reassemble them later than to loop a bunch of GET/PUT pairs. |
|||
06-13-2017, 10:00 PM
Post: #150
|
|||
|
|||
RE: Programming puzzles: processing lists!
Well from a completely ignorant point of view I somehow believe that consuming a list (for iterations) is somehow quick (indeed your tail head method is as fast as for + get). Just one has to extract elements properly. For this I asked you a sysrpl approach, because maybe there is a way.
Wikis are great, Contribute :) |
|||
06-14-2017, 05:40 AM
Post: #151
|
|||
|
|||
RE: Programming puzzles: processing lists!
(06-13-2017 06:23 PM)DavidM Wrote: I'll recreate the list by copying the two blocks of memory containing the reversed parts into a new list object directly. It will require a bit more setup (and a lot more code), but it should be a relatively fast option that is nicer to the overall memory footprint of the machine. First stab at the new version of LROT (LROT7 below) is promising: \begin{array}{|lrrrrc|} \hline \textbf{Command} & \textbf{50} & \textbf{100} & \textbf{500} & \textbf{1000} & \textbf{Cycles} \\ \hline \textbf{LROT} & 0.0259 & 0.0339 & 0.1006 & 0.1820 & 50 \\ \textbf{LROT7} & 0.0262 & 0.0299 & 0.0592 & 0.0945 & 50 \\ \hline \end{array} Timings are comparable to the original for smaller lists, but it scales nicely as the list sizes grow. And it's much nicer with regards to the memory footprint, in that it doesn't explode the list and rebuild it. Completed the 1000-element "stress test" (1000 loops with DUP/HEAD/etc.) in about 125 seconds. |
|||
06-14-2017, 08:04 AM
(This post was last modified: 06-14-2017 08:06 AM by pier4r.)
Post: #152
|
|||
|
|||
RE: Programming puzzles: processing lists!
Nice. and 125 seconds vs 4800 is quite an improvement. Because, at least from my pov, I do believe that the average use case of rotating a list is not rotating it once, but multiple times.
Wikis are great, Contribute :) |
|||
06-14-2017, 08:57 AM
(This post was last modified: 06-14-2017 12:08 PM by pier4r.)
Post: #153
|
|||
|
|||
RE: Programming puzzles: processing lists!
@DavidM: another request if I may ask.
POS , as far as I know, returns the first matching position in a list. What if I want to know all the positions that matches the element? I did, for one of the challenges, an helper routine. Do you find reasonable to develop one for your library? Small argument by me: whether a routine it is good already in userRPL, it does not mean that it ready available for a user (which has to find it or develop it), that is the reason to have a "useful" library. request: Code:
edit. another request, list contains Code:
request: does your library complement goferlist or not? (I would say: until you find faster and/more flexibile code, yes) Because otherwise diff/intersect may be useful. edit. Also #36 Code:
edit: check goferlist from time to time, I have to say that the documentation of the single commands is, well, too "twitter like". Sometimes really cryptic. Whatever the product, a proper documentation is part of its value. Wikis are great, Contribute :) |
|||
06-14-2017, 11:49 AM
Post: #154
|
|||
|
|||
RE: Programming puzzles: processing lists!
ALLPOS
Code: \<< or, when no match returns an empty resultlist: Code: \<< Cheers, Werner 41CV†,42S,48GX,49G,DM42,DM41X,17BII,15CE,DM15L,12C,16CE |
|||
06-14-2017, 02:34 PM
Post: #155
|
|||
|
|||
RE: Programming puzzles: processing lists!
(06-14-2017 08:04 AM)pier4r Wrote: Nice. and 125 seconds vs 4800 is quite an improvement. Because, at least from my pov, I do believe that the average use case of rotating a list is not rotating it once, but multiple times. I agree with you about the multiple rotations, which is part of the reason I went ahead and did this. To be fair, this is (for a 50g) a fairly extreme situation due to the list size. Exploding a list and then rebuilding it leaves behind unreferenced objects; the number of discards will be (at a minimum) equal to the number of objects in the list + the original list. When you do that rapidly in succession, the "garbage" builds up quickly. It's inevitable that a GC will take place, and when it does, it has to locate/flag-objects-for-disposal/re-compress all of the space they were inhabiting. The changes I made still leave a large block of memory disposed, but it's contiguous and easily "skippable". When watching a loop make 1000 calls to LROT (v7) operating on a list of 1000 elements, you can barely discern when the GCs occur. The previous version left 1000 disposable objects behind every iteration, so the GCs would lock up the system for a painful period every time they happened. I should note that this phenomenon is not unique to my library; any activity (whether built-in or user-programmed) that explodes/implodes large lists will be vulnerable to this issue. |
|||
06-14-2017, 03:27 PM
Post: #156
|
|||
|
|||
RE: Programming puzzles: processing lists!
(06-14-2017 08:57 AM)pier4r Wrote: @DavidM: another request if I may ask. I think of this "working project" as a collaborative effort, and I'm happy to consider all suggestions for additional functions/commands. It continues to be a good exercise for me to consider the application of list processing in solving problems as opposed to my more "iterative" inclinations. I've been somewhat concerned that showing my earlier timing comparisons with GoferList functions would be perceived as my wanting to replace it (or even "compete" with it). That definitely wasn't/isn't my intent. It seemed like a good benchmark for comparison with similar functions that I'd already created, so I thought it would make sense to show the data that way. GoferList appears to have been originally created with the idea that it would replicate known list functions from the Haskell language, so I'm guessing that the original developers were less concerned about documenting the commands and only felt the need to provide the minimum needed to use the commands in an RPL context. They probably figured that it's main appeal would be to people who already knew what the functions did, or would seek the meaning in other ways. I definitely see my library as a complement to GoferList, even though there is a small amount of overlap. If I'm to implement more functions, I'd prefer to focus on those things that aren't already available in GoferList (or trivial to implement). As an example, many of the GoferList commands are designed to apply a user-defined function to a process. I have no desire to replicate that sort of functionality, as I don't think I could really add anything useful to those commands. Gilles' suggestion of a DOPERM command would represent new functionality that definitely would add value (and isn't trivial), so it's high on my list to investigate. So that's more the type of thing that interests me at present. I'll add your recent suggestions to my list of things to ponder, as I can see where those functions would be useful in a variety of contexts. |
|||
06-14-2017, 03:38 PM
Post: #157
|
|||
|
|||
RE: Programming puzzles: processing lists! | |||
06-14-2017, 06:02 PM
(This post was last modified: 06-14-2017 06:04 PM by pier4r.)
Post: #158
|
|||
|
|||
RE: Programming puzzles: processing lists!
Werner nice input, although your very usage of the stack sometimes is not so easy to follow if one loses track of what is on the stack (like the fact that you collect only one input in a variable, the other is on the stack). I prefer something that may be bigger but easier to read if I did not touch it for long.
Anyway, the #35 , that with "for + get" is feasible (still, I think that even with for + get the #20 would be quite more demanding), seems working. Actually I am not sure I tested all the corner cases so if you want to play along, do it. It uses, I think, the library of david and also goferlist in case some commands were missing (but I avoided to replicate them) The #35 requires the #36 and an helper. If you have tips were I can improve (keeping things readable), let me know. For the timings, well, I'll do if there is a bit of competition. Code:
Wikis are great, Contribute :) |
|||
06-15-2017, 06:21 PM
Post: #159
|
|||
|
|||
RE: Programming puzzles: processing lists!
I thought that it might be worth a few more bytes to have two more special-purpose commands that could be wrapped around the same core that's at the heart of LROT: LRLL (List Roll) and LRLLD (List Rolldown). These are functionally equivalent to 1 LROT and -1 LROT, but faster since the amount of rotation doesn't have to be computed from an argument. It also seemed to me that a rotation of 1 or -1 would be common enough to warrant a separate command for those purposes.
The names were chosen to coincide with similar RPL stack-based functionality. LRLL does the same thing as Code: \<< LRLLD is the same, just replace ROLL with ROLLD. |
|||
06-15-2017, 06:30 PM
Post: #160
|
|||
|
|||
RE: Programming puzzles: processing lists!
Reading the replies of all experts is just fascinating.
|
|||
« Next Oldest | Next Newest »
|
User(s) browsing this thread: 6 Guest(s)