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

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

multivoque application algorithms
by Alexandre Carsac 03 Mar '03

03 Mar '03

I applologize, perhaps this is the wrong place, Just wondering if the boost Graph library can be used for multivoque application. ------------------------------------------------------------------------------------------- let be : E = {x0, x1, ...,xp} F = {y0, y1, ...,yq} G': y -> {xa.....xm} and card G'(y) = constant = m G : x -> {ya.....yn} and card G(x) < M for all x belongs to E G' is the reverse application of G. G is irregular. h : x -> h(x) whose values are in [0,1] The problem is to find the y0 so that G'(y0) maximise h(xa), ...,h(xm). (assume that h(xi) near of 1 for all xi in G'(y0) ) One simple approch would be to calculate h(x) for all x associated to all y and keep the maximum, but this lead to too much calculations. ------------------------------------------------------------------------------------------------- I try to organized E into a binary tree. The idee is to shrink E so that each schink step truncate the search space. lets take x0 so that card G(x0) be maximum. We have the composed application of G with its reverse one G': G o G' (x0) = {xaa....xpa, xab,...xpb, xpa,....xpn} means that x0 is tied to some other x's in E by G. if h(x0) is the not near 1, then the y we are looking for will have fewer chance to belongs to G'(x0), so it is not worth to look at all the h(xaa), ...h(xpn). This would create a bi-partion on E and so on. --------------------------------- Do You Yahoo!? -- Une adresse @yahoo.fr gratuite et en français ! Testez le nouveau Yahoo! Mail [Non-text portions of this message have been removed]

1 0

Re: [Boost-Users] stop a thread
by William E. Kempf 03 Mar '03

03 Mar '03

> From: "hlh771 <hlh771(a)yahoo.com.cn>" <hlh771(a)yahoo.com.cn> > > How to stop(or cancel) a thread? Suppose I have a main thread and a > worker thread, the worker thread does something looply,and when the > main thread get a message indicating the process should exit, I want > stop the worker thread,how to do that?I have read other thread > library,their thread class supply a method called stop or cancel > which stops the thread. Their Implementation is:the thread class has > a data member indicating whether the thread should stop,and the > thread function looply check the data member whether the thread > should stop.Generally the data member's type is a bool or Event. > Is it the best way to stop the thread? How the boost to deal with > this? > Thanks in advance. Today you should use a flag and a condition/mutex combination. Direct support for cancellation is being provided in a future release. William E. Kempf wekempf(a)cox.net

3 4

Exporting references into boost::python
by Nicholas Francis 02 Mar '03

02 Mar '03

A newbie question regarding boost::python. I need to make a templatized function that returns a boost::python::object containing a reference to an object I already have. Basically: template<class T> object convertToPython (T &obj) { return object (obj) } All classes I pass through T are already wrapped (they are also not copy-constructible). I tried with the code found at python.orgs wiki: template<class T> T& identity(T& x) { return x; } template<class T> object get_object_reference (T& x) { object f = make_function (&identity<T>, return_value_policy<reference_existing_object()); return f(x); } this code gave me a compile error (on CW 8.3, Mac OS X). Apparently CW is convinced that the &identity<T> maps to void. I then tried to hardcode a type just to see if it worked at all: InputManager &identity (InputManager &x) { return x; } object get_object_reference (InputManager &x) { object f = make_function ( &identity, return_value_policy<reference_existing_object()); return f(x); } This compiled, but gave me a typeerror at runtime: (No to_python (by-value) converter found for c++ type: InputManager) I've seccesfully wrapped functions like this: InputManager &GetInputManager(); BOOST_PYTHON_MODULE (Engine) { def ( "GetInputManager", &GetInputManager, return_value_policy<reference_existing_object() ); } So it should be able to work. Any help would be greatly appreciated.... TIA, Nicholas

2 1

VC7 says 'fatal C1067' compiling multi_array?
by Clift Norris 01 Mar '03

01 Mar '03

Greetings: I'm using VC7 and boost_1_29_0 to compile this code that is using multi_array.hpp. typedef boost::multi_array<double, 4> theta4_t; typedef theta4_t::index index; theta4_t Theta_Trun( boost::extents[10000][2][10][256] ); I'm receiving this error: "fatal error C1067: compiler limit: debug information module size exceeded" Of course, the code compiles fine using the "Release" configuration, but I need the "Debug" one also. Is there a workaround for this, other than the 'make your variables shorter' suggested in the VC7 documentation? Thanks, -Clift

2 1

are namespaces of boost libs required to be lowercase?
by Markus Werle 28 Feb '03

28 Feb '03

Hi! If someone wants to contribute code to boost: Is she required to provide the names of all namespaces lowercase? E.g. someone provides Elefant, the big library. Is she allowed to have namespace Elefant and perhaps namespace Elefant::Private, etc. I could not find any hint except this here, which did not really help me, since I do not know the standard library's conventions. Any link/ hint appreciated: -- citation from the boost docs -- "Naming conventions rationale The C++ standard committee's Library Working Group discussed this issue in detail, and over a long period of time. The discussion was repeated again in early boost postings. A short summary: Naming conventions are contentious, and although several are widely used, no one style predominates. Given the intent to propose portions of boost for the next revision of the C++ standard library, boost decided to follow the standard library's conventions." best regards, Markus

4 6