Re: [boost] [random] lagged_fibonacci607 with Intel compiler for Linux v10 produces negative output values

From: Mikko Vainio (mikko.vainio_at_[hidden])
Date: 2007-07-19 01:08:17

Matthias Troyer wrote:
> On 17 Jul 2007, at 14:58, Mikko Vainio wrote:
> 
>> Dear boosters,
>>
>> Has anyone else noticed that lagged_fibonacci607 produces negative
>> random numbers when compiled with the Intel C++ compiler for Linux  
>> v10.0
>> using Pentium4 processor specific optimizations? As far as I  
>> understand,
>> the random numbers generated by lagged_fibonacci607 should be  
>> within the
>> interval [0,1). GCC and the Intel compiler with less aggressive
>> optimizations generate values within that range.
>>
>> This is probably a bug in the Intel compiler, but if anyone has a
>> workaround please let me know.
> 
> 
> Can you see it just with the default seed?
> Which compiler options do you use?
> 
> Matthias
> 

Yes. Based on little experimenting I'd say it occurs with any seed.

The error occurs when any SSE optimizations are turned on: icpc -O2 
-msse ....

Below is a short program used for the testing on a 2GHz Core Duo T2500 
running Fedora 7.

/Mikko

--------[scissors]--------------------------

#include <iostream>
#include <vector>
#include <boost/random/lagged_fibonacci.hpp>

using namespace std;

int main( int argc, char* argv[] )
{
        boost::lagged_fibonacci607 fibo;
        fibo.seed( 34597245 );
        unsigned int nErrors = 0;
        double rndom;
        for( size_t j = 0; j < 10000; ++j )
        {
                rndom = fibo();
                if( 1. < rndom || 0. > rndom )
                        ++nErrors;
        }
        cout << "errors = " << nErrors << endl;
        return 0;
}