IEEE exception

[dpo]

Is there a way to compile the ASL so IEEE exceptions return Inf or NaN instead of producing an error? On some Linux machines, IEEE exceptions results in an error message such as "Error while evaluating
objective.", while on OSX, they seem to simply return Inf or NaN.

Thanks.

[vitaut]

ASL disables IEEE exceptions unless ASL_FPINIT_KEEP_TRAPBITS is defined:

mp/src/asl/solvers/fpinit.c

Lines 130 to 137 in 446d112

	#if defined(_FPU_IEEE) && defined(_FPU_EXTENDED) && defined(_FPU_DOUBLE)
	__fpu_control = _FPU_IEEE - _FPU_EXTENDED + _FPU_DOUBLE;
	_FPU_SETCW(__fpu_control);
	#elif defined(__i386__) || defined(__x86_64__)
	__fpu_control = 0x27f;
	_FPU_SETCW(__fpu_control);
	#elif defined(FE_ALL_EXCEPT)
	fedisableexcept(FE_ALL_EXCEPT);

However, the "Error while evaluating objective" is usually reported when a math function returns an error via errno and not because of IEEE exceptions:

mp/src/asl/solvers/rops2.c

Line 534 in 446d112

if (errchk(rv))

I'm not sure if you can disable this, but you can handle errors yourself via asl->i.err_jmp_.

[dpo]

Sorry for the slow reply. I'm observing an erroneous function and gradient values on Linux, but they're ok on OSX. Technically, the evaluation should produce an IEEE error and return +inf. On Linux, I get zero instead.

I'm trying to determine whether the issue comes from ASL or the fact that on Linux, ASL is compiled with gcc (on OSX I use clang by default). When I try to build with gcc 4.9 on OSX, I get the error message

mp-2.0.1/src/os.cc:76:12: error: 'strlen' is not a member of 'std'
     size = std::strlen(&buffer[0]);
            ^

I added #include <cstring> in os.cc and everything compiles, but I get a test failure:

The following tests FAILED:
     14 - os-test (Failed)
Errors while running CTest

I invoke cmake with -DBUILD_SHARED_LIBS=True -DCMAKE_CXX_COMPILER=g++-4.9 -DCMAKE_C_COMPILER=gcc-4.9.

Any ideas?

[vitaut]

What version do you use? The strlen error should be fixed already :

mp/src/os.cc

Line 29 in 51ae35d

#include <cstring> // for std::strlen

.

Could you post the test log Testing/Temporary/LastTest.log that can be found in the build directory?

[dpo]

I use version 2.0.1. Somehow the #include <cstring> isn't there. I'll try a fresh download.

Here's the test log: https://gist.github.com/f8b120d6bc81cb273f5a

EDIT: I confirm that os.cc doesn't have a #include <cstring> in release 2.0.1.

[vitaut]

I meant it is fixed in the master branch. I can release a minor release with a fix, but let's try to address other issues first.

Strangely I couldn't reproduce the os-test failure on my OS X systems. Could you by any chance post the output of CMake run on a clean mp repo? This could give more info about your setup for me to reproduce the issue.

[mlubin]

I just tested with ampl/mp master and the erroneous function value issue isn't resolved on linux.

[mlubin]

No apparent difference when compiling with -DASL_FPINIT_KEEP_TRAPBITS.

[vitaut]

To clean things up a bit here I've moved the os-test failure to a separate issue (#61) and prepared a corrective release 2.0.2 that fixes the compilation error on OS X.

Regarding the main issue, as I mentioned in #59 (comment), ASL has macro WANT_INFNAN which enables Infinity & NaN propagation. However it only works with some operations at the moment.

[vitaut]

@dpo I'm closing this issue since it is possible to enable Infinity propagation at least for some operations with WANT_INFNAN, but feel free to reopen if you want some specific operation, which currently doesn't handle WANT_INFNAN, to handle it.