April 2008 - Boost-users - lists.preview.boost.org

Wrapping a nested c++ container to python
by RR 29 Apr '08

29 Apr '08

Hello I have some trouble with handling c++-container. I have these code: typedef struct foo { typedef std::vector<double> pt_coords; typedef std::vector<pt_coords> pt_list; }; foo return_foo(); Now I want to wrap these, but I don't know how! I have test this, but I think it's wrong: BOOST_PYTHON_MODULE(...) { implicitly_convertible<foo::pt_coords,tuple>(); implicitly_convertible<foo::pt_list,list>(); } By using this in Python I get: TypeError: No Python class registered for C++ class std::vector<std::vector<double, std::allocator<double> >, std::allocator<std::vector<double, std::allocator<double> > > > How to do this right? Greetings

2 3

Re: [Boost-users] math
by Hicham Mouline 28 Apr '08

28 Apr '08

>Hicham Mouline wrote: >>> Are there libraries with class templates for the following concepts: >>> >>> . notion of mathematical function f(x1,x2,.,xn), and perhaps >>> derivation/integration >>> >>> . any-order multidim partial differential equations? >>> >>> . multidim stochastic differential equations? >Not at present, sorry, >Regards, John. If I were to attempt to implement such things, I would need to familiarize myself with boost::function maybe, boost::mpl, boost::math for special functions? I just finished reading the interesting "c++ template metaprogramming" which highlights such a large number of ideas that I will read once more. Assuming the general case of a set of stochastic processes x1....xm (X is a vector of xi): dx1 = a1(t, X) dt + b11(t,X) dW1 + b12(t,X)dW2 + .... + b1n(t,X) dWn ... dxm = am(t, X) dt + bm1(t,X) dW1 + bm2(t,X)dW2 + .... + bmn(t,X) dWn where ai: functionals of time and the state vector X bij: functionals of time and the state vector X Wj: non deterministic processes, governed by various statistical distributions (e.g. Normal) (perhaps ones listed in boost::math), with correlation matrix M n and m would have to be runtime variables in my case as the end-user might have thousands of processes. How would one model such a thing? class StochasticProcessesSet { template<class A, class B, class W> explicit StochasticProcessesSet(const std::vector<const A*>, const std::vector<const B*>, const std::vector<const W*>) { // 1. for correctness, runtime assertion that the sizes of the 3 vectors // are consistent with the system of equations above // 2. how to enforce that A and B are functionals that take the appropriate // number of arguments. // It seems impossible to mix a compile-time check on the template arguments // with the runtime variables n and m // 3. How to enforce W is one of the stat distributions } }; I would want to have special cases (perhaps for efficiency) for the above such as: 1 stochastic process dependent on n Wi with functionals dependent just on itself and time class StochasticProcess { .... }; What relationship should StochasticProcess have with StochasticProcessesSet? A final special case would be 1 process driven only by source of randomness, not n. Clearly, these first thoughts are not applicable, given than n and m need to be runtime. Regards,

1 0

[general] General how-to-use-the-library documentation?
by Jason Sachs 28 Apr '08

28 Apr '08

Is there any other general (e.g. not specific to any of the particular libraries) about compiling/linking application programs that use Boost libraries? The only one I found was this: http://www.boost.org/doc/libs/1_35_0/more/getting_started/index.html I don't see the BOOST_ALL_DYN_LINK #define mentioned anywhere (e.g. to use shared libraries rather than static libraries, define this symbol in your makefile / IDE) other than the source files themselves, and am wondering what other usage parameters I need to know about. (p.s. the links to alternative format documentation on this page http://www.boost.org/doc/libs/1_35_0/doc/html/about.html are broken and yield 404 Not Found)

1 0

Integer bounds checking with boost?
by Graham Reitz 28 Apr '08

28 Apr '08

I could have sworn I saw an example using boost to do simple numeric bounds checking. For example (silly): enum number_range {A=0,B=1,C=2,D=3,E=4,F=5,G=6,H=7}; class check_number_range { public: check_number_range(number_range a_number) { // Can this check be boostified? <--- Question located here. if ((a_number) > -1) && (a_number) < 8)) { // do something } else { throw out_of_range; } } }; thanks much, graham

1 0

[numeric_conversion] rounding
by Peter 28 Apr '08

28 Apr '08

I'm wondering how to use the numeric_conversion library to perform rounding based on types. Assume I have a double value of 3.7. If I convert it to integer, the result should be 4. If I convert it to float, the result should be 3.7. The types I want to convert from and to are both specified in a template, i.e. they form an open set of types. >From reading the numeric_conversion docs I tend to think that the library can do what I want, but I'm unable to find out how. Can anybody help? Peter

2 3

Re: [Boost-users] [MultiIndex] re-indexing on load
by Mark Westcott 28 Apr '08

28 Apr '08

Hi Joaquin. > As you have found out, loading involves some amount of reindexing, > though the algorithm is designed to minimize this, let me summarize > how serialization of multi_index_containers is performed: > > Saving a multi_index_container internally consists in: > > 1. Saving each of the elements, in the order they are traversed > with index #0. > 2. For each index other than #0, some "readjusting" info is saved if > necessary. More on this later. > > Loading on the other hand reduces to: > > 1. Loading each element and inserting them with hinted insertion > at the end of the container: for index #0, hinting the end() position > reduces the nomber of key comparisons performed: for the rest > of key based indices, hinting end() usually does not have any impact > and insertion takes as many comparison ops as a regular insertion. > 2. After step 1 the elements in index #0 are listed in the exact > order they were originally saved, as it should be. As for the > rest of the indices, they can be differences in traversal order > for non key-based indices (sequenced and random_access) > and for ordered non-unique indices (equivalent elements will be > adjacent as ordered non-unique indices guarantee, but they might > be shuffled with respect to the original situation). The "adjusting" > info stored in saving step 2 is used here to restore the exact > original traversing order for every index of the container. > > So, to try to minimize your loading time you'd need to first make > a guess about where the most overhead lies: is it in step 1 > or step 2? Step 2 can only be significant if any of your indices > *other than #0* is of type ordered_non_unique, sequenced > or random_access. Assuming you indeed have a situation where > step 2 causes some impact, you can: > > a) if affordable, move the problematic index to position #0, so that it > won't get any adjustment later. > b) replace Boost.MultiIndex built-in serialization with the a handmade > method so as to avoid step 2 altogether. Please find attached a sketchy > implementation example of this idea. Note that this method does > not guarantee exact traversal order replication for indices other > than #0. > > In any case, do not expect spectacular improvements after applying > any of the approaches above, as more likely most of the overhead > comes from step 1 anyway. A more radical solution, but also one > with a very high impact in your application code, is to put the > multi_index_container in a memory-mapped file, so that serialization > reduces to dumping the memory segment. You can find an example > of this technique at: > > http://www.boost.org/doc/libs/1_35_0/libs/multi_index/doc/examples.html#exa… Thanks for the explanation and ideas. Based on the information you provided, I guessed that moving from hashed indices (where the hash function is relatively expensive) to sorted_non_unique (with a cheap comparison function) would improve things; and indeed that change has got the load hugely quicker. I then tried the 'poor man serialization' example that you attached. As you suspected, the performance effect of this was pretty small and so I guess it is indeed step 1 that is causing the issue (I have several ordered_non_uniques). I do like the memory mapped file idea, but as you say, the impact on the code is large; I think too large for the time being, though it may be sometime I come back to if needs be in the future. I think, for the time being at least, the performance is ok now (using ordered rather than hashed indices). Is there any chance that future versions of MI might be able to store the hashes/complete ordering whilst serializing, or is that Just Not Possible (TM)? Thanks again, Mark

2 1

portable binary archive major bugfix
by Pfligersdorffer, Christian 28 Apr '08

28 Apr '08

Hello everybody, if you're using or evaluating my portable binary archives be sure to get version 2.0 from the boost vault (http://www.boost-consulting.com/vault/) (Remeber it only works with boost 1.33 and 1.34.) Version 1.x contained a serious bug that effectively transformed most of negative integral values into positive values! For example the two numbers -12 and 234 were stored in the same 8-bit pattern and later always restored to 234. This was fixed in this version in a way that does not change the interpretation of existing archives that happend to work because there were no negative numbers - as is unfortunately the case in our application or the bug would have surfaced long ago. Archives created by version 2.0 and containing negative numbers will raise an integer type size exception when reading it with version 1.x. Thanks to Markus Frohnmaier for testing the archives and finding the bug. PS @ Robert Ramey: Your portable binary example also suffers from this bug. In fact I felt free to copy it from you ;-) Ah, I see now that you fixed it in version 1.35. Didn't think you'd bother to since you wanted to remove it completely... Regards, -- Christian Pfligersdorffer Software Engineering http://www.eos.info

1 0

Re: [Boost-users] a problem about boost::regex
by Guo Xu 28 Apr '08

28 Apr '08

Hi, John I have a string splited by "\0" ( it means {'\\','0'}, the bellow is the same ), for example: "\0xxxx\0yyyy\0zzzz\0". And there is some special situation like "\0xxx\\0\0yyy\\\0zzz\0" , you can think it is similarly to c style string, the two '\\' means a single '\' and the '\\0' is not a spliter. i want to split them in to 0) xxx\\0 1) yyy\\ c) zzz how can i do it in boost::regex? thx,all > Message: 7 > Date: Sat, 26 Apr 2008 12:52:43 +0100 > From: "John Maddock" <john(a)johnmaddock.co.uk> > Subject: Re: [Boost-users] a problem about boost::regex > To: <boost-users(a)lists.boost.org> > Message-ID: <002d01c8a794$0d593e50$eeea0352@fuji> > Content-Type: text/plain; format=flowed; charset="UTF-8"; > reply-type=original > > Guo Xu wrote: > >> a regex expression: > >> (?<=[^\\](?:\\\\)*)\\0 > >> can work well in c#, but error in boost::regex > >> why? > > Variable length lookbehind isn't supported in Boost.Regex (or in Perl for > that matter). > > Regards, John. > >

1 0

re; Boost CGI
by Brad 28 Apr '08

28 Apr '08

Forget this post, as one of the guys on IRC just pointed out, Darren has recently submitted changes via SVN so is obviously still active. Sorry for the noise guys. Cheers, Brad

1 0

[serialization] Stateful serialization
by Scott Howlett 27 Apr '08

27 Apr '08

Hello, In a recent post, I asked whether it was possible to swap deserialized pointer values for already-existing ones so I could merge the incoming data with objects already in memory. I guess the answer is no... An alternative is to serialize some other identifier instead of writing pointers, and then on de-serialization I find my existing objects via those identifiers. This works fine, but it leads to another question: How do I make my internal objects available to the deserialization routines so that they have an opportunity to look up the necessary information? Or more generally, How do I do stateful serialization / deserialization? For example, maybe I want to supply options that fine-tune how my data is serialized or deserialized. The only parameter that is passed around to all the serialization functions is the Archive, so if I want to add options or other context information, it seems like I should add this to the Archive. That's simple enough to do; I just made a wrapper Archive that can sit atop any other Archive and contain any other options / context data I need. On the one hand it seems sneaky to hijack the Archive for this purpose, but on the other hand iostreams do basically the same thing via the various ios_base calls (setf, width, precision, etc.) so maybe it's not so bad. Am I perhaps overlooking some facility already present in the serialization library? Thanks, Scott

3 3