Re: GLPK compiled with -march=native produces different results

[Domingo Alvarez Duarte]

Hello Peter !

Every day is a day to learn something new !

Thank you for reply !

Indeed using this option I'm getting identical outputs.

With the modifications I did to GLPK 4.65 (one of then to allow eliminate "xassert" defining "-DNDEBUG") https://github.com/mingodad/GLPK

====

CFLAGS="-g -O3 -march=native -ffp-contract=off -DWITH_SPLAYTREE -DNDEBUG -flto" ./configure

====

Cheers !

On 17/7/20 18:04, Peter Cawley wrote:

Perhaps you are looking for -ffp-contract=off. With the default -ffp-contract=fast, GCC will make use of FMA instructions if the -march architecture supports them. FMA instructions are usually more accurate than separate multiply then add, but do not give bitwise identical results to a multiply followed by an add (because FMA has no rounding step between the multiply and the add).

On Fri, Jul 17, 2020 at 4:34 PM Domingo Alvarez Duarte <mingodad@gmail.com> wrote:

Hello !

Doing tests with GLPK I noticed that if I compile it with "-march=native
-g -O2" on a computer with an "i7" processor the solution for several
files change, the objective function result remain the same but other
values varies from model to model.

Example dist.mod:

====

      6 rlim[w96]    NU             3             0 3      -11283.2 #
only -O2
====

      6 rlim[w96]    NU             3             0 3      -11283.1 #
-O2 -march=native

====

Example mfasp.mod:

====

      1 r[1,2]                      9             1   # only -O2
      2 r[2,3]                      2             1
====

      1 r[1,2]                      1             1   # -O2 -march=native
      2 r[2,3]                     10             1
====


Cheers !