- Boost-users - lists.boost.org

boost::signals build and C++Builder 6 Pro in 1.29.0
by Russell Hind 07 Mar '03

07 Mar '03

I have built the boost::signals library with C++Builder 6 Pro Update 4 and am running into problems: a) We tend to use double word alignment, not quad word (which is the default in C++Builder). It appears that the libraries are built with quad word, but the headers don't enforce this, so when I change the options to double word (in our project) and try to construct a signal with boost::signal0<void> s; I get an access violation when constructing a signal (from line 266 of shared_count.hpp) The call stack is 00417F3F __internal_malloc 00417DD8 __internal_malloc 00417539 ___org_malloc 00416EA2 ___org_new 004076E1 boost::detail::shared_count::shared_count(this=:0012FBB4, p=:007F3928, d={ }, =:007F3928) 004063E3 boost::shared_ptr<boost::signals::detail::signal_base_impl>::shared_ptr<boost::signals::detail::signal_base_impl>(this=:0012FBB0, p=:007F3928) 0040546D boost::shared_ptr<boost::signals::detail::signal_base_impl>::reset<boost::signals::detail::signal_base_impl>(this=:0012FF70, p=:007F3928) 00403F51 boost::signals::detail::signal_base::signal_base(this=:0012FF70, comp=:0012FC24) 0040232E boost::signal0<void, boost::last_value<void>, int, _STL::less<int>, boost::function0<void, boost::empty_function_policy, boost::empty_function_mixin, _STL::allocator<boost::function_base> > >::signal0<void, boost::last_value<void>, int, _STL::less<int>, boost::function0<void, boost::empty_function_policy, boost::empty_function_mixin, _STL::allocator<boost::function_base> > >(this=:0012FF70, c=:0012FC7C, comp=:0012FC74) 00401BF3 main(argc=1, argv=:007F21E4) 00420862 __startup b) If I do use quad word alignment, then the object constructs correctly, but when I connect a method to it with void VoidFunc(void) { } int main(int argc, char* argv[]) { boost::signal0<void> s; s.connect(&VoidFunc); } the program hangs. Turning on CodeGuard shows a similar problem to that above. The object was allocated with a size of 12 bytes, but code is trying to access byte 16 in the object (I guess this is a similar alignment problem). The callstack is 0041E44E boost::counted_base::dispose(this=:00C04360) 00427203 boost::signals::detail::signal_base_impl::connect_slot 00404741 boost::signal0<void, boost::last_value<void>, int, _STL::less<int>, boost::function0<void, boost::empty_function_policy, boost::empty_function_mixin, _STL::allocator<boost::function_base> > >::connect(this=:0012FF60, in_slot=:0012FDA8) 004026C5 main(argc=1, argv=:00C021E4) 00433802 __startup It says the this pointer block is only 12 bytes long, but the code is trying to access byte 16. Do (or should) all the bost headers force alignment before all structures/classes and then restore the alignment afterwards? Currently, I have to link directly with the signals source code to get use the library. Has anything to do with this been changed in the builds of 1.30.0? Thanks Russell

8 27

Using Unit Tests w/Codewarrior Pro 8.3
by paul_snively <paul_snively＠yahoo.com> 07 Mar '03

07 Mar '03

Folks, I'd like to be able to run a boost unit test, e.g. is_convertible_test.cpp, as kind of a one-shot apart from having to deal with the rest of boost. So I read the documentation on the unit test framework, and it seems like this should be relatively straightforward. I've created a Carbon Console project in Codewarrior, which contains the following files: is_convertible_test.cpp init.cpp (from type_traits/test) unit_test_main.cpp unit_test_suite.cpp unit_test_result.cpp unit_test_log.cpp unit_test_monitor.cpp execution_monitor.cpp unit_test_parameters.cpp test_tools.cpp This all compiles and links file, which is a relief in and of itself. :-) However, when I simply run the app, I get "*** No errors detected". If I remove is_convertible_test.cpp and rebuild and run again, I get the same result. :-) So my suspicion is that I need to do more to let the framework know that it needs to run is_convertible_test.cpp. Can someone please clue me in as to what that is, in minimal terms? Thanks!

4 4

shared_ptr 1.27 w/gcc 2.96 (VxWorks)
by Brian 06 Mar '03

06 Mar '03

We are using boost 1.27 (I know, I know) with the real-time operating system VxWorks, which ships with its own "special" version of gcc 2.96 (not that same as Red Hats - I know, I know). Today I finally got around to providing a version of atomic_count.hpp for VxWorks. I had to play some games with #defines to get BOOST_HAS_THREADS and all the other things set up right before my implementation would get chosen. Once that was squared away, I put a cout in the constructor of my atomic_count class just be sure it was all working. I then stamped down a shared_ptr<int> and much to my surprise: no text output. I traced it back to the constructor for counted_base_impl (a templated class). That constructor was being called, but it inherits from counted_base, and that constructor was just not getting called! I searched your bug list on source forge but found no mention of this. But then again, its not really a boost bug! I then took a look at boost 1.29 and noticed that lots of things had been reworked in this area, so now looks like the reason to start using 1.29. I'm afraid I am stuck with VxWorks and that compiler though... So I'm just throwing this out in case anyone is remotely in the same situation as we are. I will start using 1.29 and see what happens. Hopefully its not too hard to customize the OS specific part in 1.29. I also suggest a note get added to the shared_ptr documentation about this OS specific mutex stuff thats going on. I got curious about how that was done and decided to look. Had I not done this, my shared_ptr's reference count would not have been thread safe and I wouldn't have known this. Also, is there a "what has changed in this new version of boost" file somewhere? Thanks again everyone. BN

1 1

Re: [Boost-Users] Unable to BOOST_INSTALL_PROPERTY
by Administrator 06 Mar '03

06 Mar '03

Martin, I did a simple search through the Boost website, and on http://www.boost.org/libs/graph/doc/using_adjacency_list.htm, I found this: You can also use enum's instead of struct's to create tag types. Create an enum type for each property. The first part of the name of the enum type must be edge, vertex, or graph followed by an underscore, with a _t at the end. Inside the enum, define a value with the same name minus the _t. Then invoke the BOOST_INSTALL_PROPERTY macro. enum edge_myflow_t { edge_myflow }; enum edge_mycapacity_t { edge_mycapacity }; namespace boost { BOOST_INSTALL_PROPERTY(edge, myflow); BOOST_INSTALL_PROPERTY(edge, mycapacity); } compared to: > namespace boost{ > enum vertex_xx_t { vertex_xx = 1111}; > BOOST_INSTALL_PROPERTY(vertex,vertex_xx); > } which means it should be: > enum vertex_xx_t { vertex_xx = 1111}; > namespace boost{ > BOOST_INSTALL_PROPERTY(vertex,xx); > } Dale

1 0

[BGL] Preconditioned color maps?
by Tarjei Knapstad 05 Mar '03

05 Mar '03

In an algorithm I'm working on I need to do an undirected_dfs using a visitor for analysis. I know my starting vertex, and I also want the DFS to skip one of the starting vertex's adjacent vertices (i.e. don't proceed in the "direction" from the starting vertex). Example: d-e-f / a-b-c \ g-h-i Starting vertex is 'c' and I want to eliminate the "d-e-f" branch, so the DFS finds "c-b-a" and "c-g-h-i". My first idea was to set up a vertex color map, and set the color to black for vertex 'd' in the example above, but that doesn't do much good as the undirected_dfs sets all colors to white initially. My second thought was to feed the vertex I want to eliminate and the vertex color map to my dfs_visitor implementation, something like: discover_vertex(u, g) { if (u == blocked) { vertex_colormap[u] = black; } } Will the latter approach work satisfactory? If not, is there some other way which I can use to precondition the color maps used in the BGL algorithms? Thanks, -- Tarjei

4 10

Re: [Boost-Users] NDEBUG, affects floating point math?
by Stephen Crowley 05 Mar '03

05 Mar '03

On Fri, Feb 28, 2003 at 09:53:56AM +0100, Joerg Walter wrote: > > I wrote a small piece of C code to calculate a symmetric-matrix by > > row-matrix product by directly accessing the memory (which btw is about > > 1000% faster than ublas's generated code with gcc 3.2.3 and -O3, why is > > ublas so slow?) > > If you have *not* defined NDEBUG, ublas uses many internal consistency > checks and usually no expression templates. I was referring to the case when NDEBUG is defined and comparing symmetric_matrix by general_matrix products. uBLAS is quite slow with this operation. However, I compared my symmetric_matrix by general_matrix_transpose product and ublas was only 3-4 times slower than the optimized C, not too bad. > > And when I compile with -DNDEBUG the results differ > > by around +/- 1e-16, but when its not defined the results match almost > > perfectly. > > You've probably hit a bug, may be an old one (please consider to check > against CVS, too). Could you please send or post a small sample, if the > problem persists? I think I know the source of the problem now, and I'm not sure if it's one that boost should handle or not. I attached my test code at the end of this email, only 2.7k, I hope it isn't inappropriate for the list. 387 floating point registers are stored as 80 bits internally and rounded to 64 bits when assigned to memory. SSE2 floating-point registers are stored in the native 64-bit format. The following table lists the compiler flags used and the accumulated absolute difference between the two products over 100 randomly generated matrices. gcc version 3.2.3 20030221 (Debian prerelease) no compiler flags diff sum [5,3]((0,0,0),(0,0,0),(0,0,0),(0,0,0),(0,0,0)) -DNDEBUG diff sum [5,3]((0,0,0),(0,0,0),(0,0,0),(0,0,0),(0,0,0)) -O1 diff sum [5,3]((2.91694e-15,2.65348e-15,2.7027e-15),(4.4049e-15,3.29251e-15,3.44863e-15),(4.10436e-15,4.14816e-15,3.24393e-15),(3.57873e-15,3.75741e-15,2.64068e-15),(5.03091e-15,2.79637e-15,4.06099e-15)) -O1 -DNDEBUG diff sum [5,3]((0,0,0),(0,0,0),(0,0,0),(0,0,0),(0,0,0)) -O2 diff sum [5,3]((4.37584e-15,4.8659e-15,3.61516e-15),(4.61783e-15,3.77736e-15,4.04169e-15),(4.46691e-15,5.10876e-15,5.00576e-15),(3.64118e-15,4.91968e-15,5.17641e-15),(3.97273e-15,4.25528e-15,3.75394e-15)) -O2 -DNDEBUG diff sum [5,3]((0,0,0),(0,0,0),(0,0,0),(0,0,0),(0,0,0)) -O3 diff sum [5,3]((4.96131e-15,3.88968e-15,4.34245e-15),(5.07754e-15,5.91974e-15,4.19109e-15),(4.89539e-15,4.29279e-15,5.06344e-15),(4.79781e-15,4.66033e-15,3.77996e-15),(4.75227e-15,4.72539e-15,3.33414e-15)) -O3 -DNDEBUG diff sum [5,3]((4.60569e-15,2.67928e-15,2.7079e-15),(3.21054e-15,2.98112e-15,2.61249e-15),(3.91701e-15,3.60822e-15,2.93515e-15),(3.23006e-15,3.42434e-15,3.53363e-15),(3.18105e-15,2.77035e-15,3.67198e-15)) -O1 -mfpmath=sse,387 -march=pentium4 -msse2 diff sum [5,3]((0,0,0),(0,0,0),(0,0,0),(0,0,0),(0,0,0)) -O1 -mfpmath=sse,387 -march=pentium4 -msse2 -DNDEBUG diff sum [5,3]((0,0,0),(0,0,0),(0,0,0),(0,0,0),(0,0,0)) -O2 -mfpmath=sse,387 -march=pentium4 -msse2 diff sum [5,3]((0,0,0),(0,0,0),(0,0,0),(0,0,0),(0,0,0)) -O2 -mfpmath=sse,387 -march=pentium4 -msse2 -DNDEBUG diff sum [5,3]((0,0,0),(0,0,0),(0,0,0),(0,0,0),(0,0,0)) -O3 -mfpmath=sse,387 -march=pentium4 -msse2 diff sum [5,3]((0,0,0),(0,0,0),(0,0,0),(0,0,0),(0,0,0)) -O3 -mfpmath=sse,387 -march=pentium4 -msse2 -DNDEBUG diff sum [5,3]((3.62991e-15,3.74006e-15,4.62824e-15),(4.00548e-15,4.5311e-15,3.60129e-15),(3.48332e-15,5.11743e-15,3.08174e-15),(3.49373e-15,4.05752e-15,4.2865e-15),(4.81169e-15,4.51549e-15,4.18589e-15)) Is this perhaps a compiler bug? Part of the problem actually comes from whether the products are accumulated into a register by the compiler, or whether they are accumulated into memory. If accumulated into a 387 register it has higher-precision before it is rounded back down to 64-bits to be stored into memory. If the compiler generates code to accumulate the products directly into memory then the results of each individual product are rounded to 64-bits at every iteration and the result will be slightly less accurate. Now, when the -msse2 flag is used, the registers are 64-bit only and thus no accuracy is gained by using them, only speed. -- Stephen [Non-text portions of this message have been removed]

3 4

GCC 3.2 compiler error using Boost::Threads::Mutex::scoped_lock
by Michael Hunley 05 Mar '03

05 Mar '03

Disclaimer: I am relatively new to gcc, linux and Boost. I'm hoping someone out in the group who is much more knowledgable about all of them can help me get past a hurdle in code I am porting from Windows. I am getting a vague error message from g++ (gcc 3.2-7 under Redhat Linux 8.0) in my code using the Boost scoped_lock. I have wrapped the Boost versions of mutex and mutex::scoped_lock as: class PMutex { public: PMutex() {} virtual ~PMutex() {} protected: boost::mutex m_Mutex; friend class PScopedLock; }; class PScopedLock { public: explicit PScopedLock( PMutex& M ) : m_ScopedLock(M.m_Mutex) {} virtual ~PScopedLock() {} static PScopedLock* Create( PMutex& M ) { return (PScopedLock*) ::new PScopedLock(M); } void Destroy( void ) { delete this; } protected: void* operator new(size_t) {} // disallowed, must declare on the stack boost::mutex::scoped_lock m_ScopedLock; friend class PCondition; }; The error is "ISO C++ forbids declarations of 'type name' with no type". I get that error on several lines, but most notably on a simple declaration like: PScopedLock Lock(MyMutex); From what little I can find on the newsgroups, this sort of error has been noted when use of template template parameters are not declared quite correctly (without the trailing ::template from what I can gather). But as far as I can tell, I should not be having that issue. I am not knowledgeable enough about templates, Boost or gcc/g++ to sort out what I am running into. Boost builds fine under gcc, the code I am porting compiles fine under MSVC6 and Kylix3, but not gcc. Can anyone help me out? thanks. michael

2 2

Has anyone compiled Boost successfully on OS390?
by kant_rakesh <kant_rakesh＠yahoo.com> 05 Mar '03

05 Mar '03

Hi, I am trying to get boost compiled on OS390. Has anyone done it successfully so far? The compiler on OS390(version2 release10) is complaining of empty template parameter. It does not like 'template <>', which is being used frequently in the library. Any ideas about how to successfully build this on OS390? Thanks and Regards Rakesh Kant

2 1

boost-users@lists.boost.org ?
by speedsnaii <speedsnaii＠yahoo.de> 04 Mar '03

04 Mar '03

I tried to subscribe to the list boost-users(a)lists.boost.org I was thinking that this is just an alias list of this one. But it seems not to be true, is it? I would prefer to use the mail interface. Can anyone enligthen me if this is possible, please? Thank you, roland

1 0

uBLAS: reference to container data & iterator declaration
by Roland Schwarz 04 Mar '03

04 Mar '03

Hi, *) First question: I am trying to get a reference to data of a container in the following way: matrix<double> m(3,3); vector<double>& v(m.column(2)); But this seems not to be possible. Is there a way to achieve this behaviour? What I basically want to achive is having functor object that holds a reference to a row/column of my container matrix. >From my questions about argument types I know that m.column(2) is a kind of an expression. How can I get the underlying reference to data if there even is one? *) Second question: matrix<double> m(10,3); ... fill the matrix ... matrix<double>::iterator it; // wrong this does not work!! for (it = m.column(2).begin(); it != m.colum(2).end(); ++it) { cout << *it << " " << foo(*it) << endl; } How would I declare my iterator? I understand when I use std::library-algorithms the automatic template parameter defer will give me the correct type from begin(). But in my case how should I declare the iterator? Thank you for help. Roland

1 0