Re: [Boost-bugs] [Boost C++ Libraries] #7951: ranlux24_base(0) should be equal to ranlux24_base()

Date view Thread view Subject view Author view

Subject: Re: [Boost-bugs] [Boost C++ Libraries] #7951: ranlux24_base(0) should be equal to ranlux24_base()
From: Boost C++ Libraries (noreply_at_[hidden])
Date: 2013-01-30 03:12:07

#7951: ranlux24_base(0) should be equal to ranlux24_base()
-----------------------------------------------------+----------------------
  Reporter: Stephan T. Lavavej <stl@…> | Owner: no-maintainer
      Type: Bugs | Status: new
 Milestone: To Be Determined | Component: random
   Version: Boost 1.52.0 | Severity: Problem
Resolution: | Keywords:
-----------------------------------------------------+----------------------

Old description:

> After fixing a bug in VC11's <random> (see [1]), I compared VC's fixed
> output to Boost 1.52.0's output and I believe I found a bug in Boost.
> Here's a self-contained test case:
>
> C:\Temp>type meow.cpp
> #include <ios>
> #include <iostream>
>
> #ifdef USE_BOOST
> #include <boost/random.hpp>
> using std::boolalpha;
> using std::cout;
> using std::endl;
> using boost::random::minstd_rand0;
> using boost::random::mt19937;
> using boost::random::ranlux24_base;
> #else
> #include <random>
> using namespace std;
> #endif
>
> template <typename Engine> typename Engine::result_type run_10k(Engine
> engine) {
> for (int i = 0; i < 9999; ++i) {
> engine();
> }
>
> return engine();
> }
>
> int main() {
> cout << boolalpha;
>
> #ifdef USE_BOOST
> cout << "Using boost." << endl;
> #else
> cout << "Using std." << endl;
> #endif
>
> cout << "N3485 26.5.5 [rand.predef]/1 satisfied: " << (run_10k(
> minstd_rand0()) == 1043618065) << endl;
> cout << "N3485 26.5.5 [rand.predef]/3 satisfied: " << (run_10k(
> mt19937()) == 4123659995) << endl;
> cout << "N3485 26.5.5 [rand.predef]/5 satisfied: " <<
> (run_10k(ranlux24_base()) == 7937952 ) << endl;
>
> cout << "I believe minstd_rand0(0) == minstd_rand0() should be
> true: "
> << (minstd_rand0(0) == minstd_rand0()) << endl;
> cout << "I believe mt19937(0) == mt19937() should be
> false: "
> << (mt19937(0) == mt19937()) << endl;
> cout << "I believe ranlux24_base(0) == ranlux24_base() should be
> true: "
> << (ranlux24_base(0) == ranlux24_base()) << endl;
>
> cout << "run_10k( minstd_rand0(0)): " << run_10k( minstd_rand0(0)) <<
> endl;
> cout << "run_10k( mt19937(0)): " << run_10k( mt19937(0)) <<
> endl;
> cout << "run_10k(ranlux24_base(0)): " << run_10k(ranlux24_base(0)) <<
> endl;
> }
>
> With my fixed build of Milan/VC12, I'm getting:
>
> C:\Temp>cl /EHsc /nologo /W4 /MTd meow.cpp
> meow.cpp
>
> C:\Temp>meow
> Using std.
> N3485 26.5.5 [rand.predef]/1 satisfied: true
> N3485 26.5.5 [rand.predef]/3 satisfied: true
> N3485 26.5.5 [rand.predef]/5 satisfied: true
> I believe minstd_rand0(0) == minstd_rand0() should be true: true
> I believe mt19937(0) == mt19937() should be false: false
> I believe ranlux24_base(0) == ranlux24_base() should be true: true
> run_10k( minstd_rand0(0)): 1043618065
> run_10k( mt19937(0)): 1543171712
> run_10k(ranlux24_base(0)): 7937952
>
> (Note that VC11 RTM/Update 1 will suffer the bug described by [1].)
>
> If I use Boost's implementation with VC11 Update 1, I get:
>
> C:\Temp>cl /EHsc /nologo /W4 /MTd /DUSE_BOOST meow.cpp /I
> boost-1.52.0\include
> meow.cpp
>
> C:\Temp>meow
> Using boost.
> N3485 26.5.5 [rand.predef]/1 satisfied: true
> N3485 26.5.5 [rand.predef]/3 satisfied: true
> N3485 26.5.5 [rand.predef]/5 satisfied: true
> I believe minstd_rand0(0) == minstd_rand0() should be true: true
> I believe mt19937(0) == mt19937() should be false: false
> I believe ranlux24_base(0) == ranlux24_base() should be true: false
> run_10k( minstd_rand0(0)): 1043618065
> run_10k( mt19937(0)): 1543171712
> run_10k(ranlux24_base(0)): 14007167
>
> Here's why I believe that ranlux24_base(0) should be equal to
> ranlux24_base(). ranlux24_base is a subtract_with_carry_engine. 26.5.3.3
> [rand.eng.sub]/7 says:
>
> "explicit subtract_with_carry_engine(result_type value = default_seed);
> [...]
> linear_congruential_engine<result_type,
> 40014u,0u,2147483563u> e(value == 0u ? default_seed : value);"
>
> /7 doesn't refer to "value" anywhere else. Therefore, 0 should behave
> identically to the default_seed 19780503u.
>

> (For comparison, VC and Boost agree that minstd_rand0(0) ==
> minstd_rand0(). 26.5.3.1 [rand.eng.lcong]/5 says:
>
> "explicit linear_congruential_engine(result_type s = default_seed);
> [...]
> If c mod m is 0 and s mod m is 0, sets the engine's state to 1, otherwise
> sets the engine's state to s mod m."
>
> The default_seed is 1u and minstd_rand0 sets c to 0, so that's why 0
> behaves identically to the default_seed.)
>
> [1] http://connect.microsoft.com/VisualStudio/feedback/details/776456

New description:

 After fixing a bug in VC11's <random> (see [1]), I compared VC's fixed
 output to Boost 1.52.0's output and I believe I found a bug in Boost.
 Here's a self-contained test case:
 {{{
 C:\Temp>type meow.cpp
 #include <ios>
 #include <iostream>

 #ifdef USE_BOOST
     #include <boost/random.hpp>
     using std::boolalpha;
     using std::cout;
     using std::endl;
     using boost::random::minstd_rand0;
     using boost::random::mt19937;
     using boost::random::ranlux24_base;
 #else
     #include <random>
     using namespace std;
 #endif

 template <typename Engine> typename Engine::result_type run_10k(Engine
 engine) {
     for (int i = 0; i < 9999; ++i) {
         engine();
     }

     return engine();
 }

 int main() {
     cout << boolalpha;

     #ifdef USE_BOOST
         cout << "Using boost." << endl;
     #else
         cout << "Using std." << endl;
     #endif

     cout << "N3485 26.5.5 [rand.predef]/1 satisfied: " << (run_10k(
 minstd_rand0()) == 1043618065) << endl;
     cout << "N3485 26.5.5 [rand.predef]/3 satisfied: " << (run_10k(
 mt19937()) == 4123659995) << endl;
     cout << "N3485 26.5.5 [rand.predef]/5 satisfied: " <<
 (run_10k(ranlux24_base()) == 7937952 ) << endl;

     cout << "I believe minstd_rand0(0) == minstd_rand0() should be
 true: "
                     << (minstd_rand0(0) == minstd_rand0()) << endl;
     cout << "I believe mt19937(0) == mt19937() should be
 false: "
                          << (mt19937(0) == mt19937()) << endl;
     cout << "I believe ranlux24_base(0) == ranlux24_base() should be
 true: "
                    << (ranlux24_base(0) == ranlux24_base()) << endl;

     cout << "run_10k( minstd_rand0(0)): " << run_10k( minstd_rand0(0)) <<
 endl;
     cout << "run_10k( mt19937(0)): " << run_10k( mt19937(0)) <<
 endl;
     cout << "run_10k(ranlux24_base(0)): " << run_10k(ranlux24_base(0)) <<
 endl;
 }

 With my fixed build of Milan/VC12, I'm getting:

 C:\Temp>cl /EHsc /nologo /W4 /MTd meow.cpp
 meow.cpp

 C:\Temp>meow
 Using std.
 N3485 26.5.5 [rand.predef]/1 satisfied: true
 N3485 26.5.5 [rand.predef]/3 satisfied: true
 N3485 26.5.5 [rand.predef]/5 satisfied: true
 I believe minstd_rand0(0) == minstd_rand0() should be true: true
 I believe mt19937(0) == mt19937() should be false: false
 I believe ranlux24_base(0) == ranlux24_base() should be true: true
 run_10k( minstd_rand0(0)): 1043618065
 run_10k( mt19937(0)): 1543171712
 run_10k(ranlux24_base(0)): 7937952

 (Note that VC11 RTM/Update 1 will suffer the bug described by [1].)

 If I use Boost's implementation with VC11 Update 1, I get:

 C:\Temp>cl /EHsc /nologo /W4 /MTd /DUSE_BOOST meow.cpp /I
 boost-1.52.0\include
 meow.cpp

 C:\Temp>meow
 Using boost.
 N3485 26.5.5 [rand.predef]/1 satisfied: true
 N3485 26.5.5 [rand.predef]/3 satisfied: true
 N3485 26.5.5 [rand.predef]/5 satisfied: true
 I believe minstd_rand0(0) == minstd_rand0() should be true: true
 I believe mt19937(0) == mt19937() should be false: false
 I believe ranlux24_base(0) == ranlux24_base() should be true: false
 run_10k( minstd_rand0(0)): 1043618065
 run_10k( mt19937(0)): 1543171712
 run_10k(ranlux24_base(0)): 14007167
 }}}

 Here's why I believe that ranlux24_base(0) should be equal to
 ranlux24_base(). ranlux24_base is a subtract_with_carry_engine. 26.5.3.3
 [rand.eng.sub]/7 says:

 "explicit subtract_with_carry_engine(result_type value = default_seed);
 [...]
 linear_congruential_engine<result_type,
 40014u,0u,2147483563u> e(value == 0u ? default_seed : value);"

 /7 doesn't refer to "value" anywhere else. Therefore, 0 should behave
 identically to the default_seed 19780503u.


 (For comparison, VC and Boost agree that minstd_rand0(0) ==
 minstd_rand0(). 26.5.3.1 [rand.eng.lcong]/5 says:

 "explicit linear_congruential_engine(result_type s = default_seed);
 [...]
 If c mod m is 0 and s mod m is 0, sets the engine's state to 1, otherwise
 sets the engine's state to s mod m."

 The default_seed is 1u and minstd_rand0 sets c to 0, so that's why 0
 behaves identically to the default_seed.)

 [1] http://connect.microsoft.com/VisualStudio/feedback/details/776456 

--
Comment (by marshall):
 Fixed the formatting in the description.

 Also, I verified the same behavior occurs with clang/libc++
-- 
Ticket URL: <https://svn.boost.org/trac/boost/ticket/7951#comment:1>
Boost C++ Libraries <http://www.boost.org/>
Boost provides free peer-reviewed portable C++ source libraries.
Date view Thread view Subject view Author view

This archive was generated by hypermail 2.1.7 : 2017-02-16 18:50:11 UTC