Re: [Patch, Fortran] PR 82143: add a -fdefault-real-16 flag

["Dominique d'Humières" <dominiq@lps.ens.fr>]

> Le 18 sept. 2017 à 20:13, Janus Weil <janus@gcc.gnu.org> a écrit :
> 
> 2017-09-18 11:31 GMT+02:00 Dominique d'Humières <dominiq@lps.ens.fr>:
>> (1) real(16) is an order of magnitude slower than real(8) for the codes I have tested (a long time ago). So its real utility is quite low.
> 
> I am fully aware that performance with quad-precision is lower than
> with double precision. How much will certainly depend on the specifics
> of the code in question.

Indeed! The polyhedron test protein.f90 goes from 19s to 32s.

> The flag I'm proposing would help in evaluating this performance hit.

My memory did not serve me well, for floating point intensive calculations it is not one order of magnitude but two (the one order of magnitude was probably for IBM REAL(16), i.e., two RAL(8))

Results for the original test_fpu.f90

Test1 - Gauss 2000 (101x101) inverts  1.1 sec  Err= 0.000000000000104
Test2 - Crout 2000 (101x101) inverts  1.2 sec  Err= 0.000000000000289
Test3 - Crout  2 (1001x1001) inverts  1.0 sec  Err= 0.000000000000082
Test4 - Lapack 2 (1001x1001) inverts  0.8 sec  Err= 0.000000000001333
                             total =  4.1 sec

Results for the REAL(10) variant

Test1 - Gauss 2000 (101x101) inverts  5.7 sec  Err= 0.000000000000000
Test2 - Crout 2000 (101x101) inverts  2.9 sec  Err= 0.000000000000000
Test3 - Crout  2 (1001x1001) inverts  2.8 sec  Err= 0.000000000000000
Test4 - Lapack 2 (1001x1001) inverts  5.4 sec  Err= 0.000000000000000
                             total = 16.8 sec

Results for the REAL(16) variant

Test1 - Gauss 2000 (101x101) inverts 119.8 sec  Err=      0.0000000000000000000000000000001018749071
Test2 - Crout 2000 (101x101) inverts 122.9 sec  Err=      0.0000000000000000000000000000003968958171
Test3 - Crout  2 (1001x1001) inverts 116.6 sec  Err=      0.0000000000000000000000000000002046368750
Test4 - Lapack 2 (1001x1001) inverts 102.2 sec  Err=      0.0000000000000000000000000000033745673898
                             total = 461.6 sec

channel.f90 went from 1s to 97s, induct.f90 from 7s to 1135s, and tfft.f90 from 1.4s to 58s.
> 
>> (2) I think your time would be better used by dealing with your assigned PRs.
> 
> I think I can very well decide for myself where to waste my spare time.

Indeed! However, as soon as you post something to a list, you are interacting (at least you are expecting it) with other people: exchanging opinion, reviewing submitted patches, bike-shedding, … .

Any way, in my mind my comment was an appreciation of your past, recent, and hopefully future contribution to gfortran. If I have upset you, it was not intended and please accept my deepest apologies.

> There were actually times when I enjoyed contributing to
> gfortran and reading this list very much, but recently it's really
> becoming a PITA and I feel like I could spend my time on much nicer
> things ...
> 
> Over and out,
> Janus

Concerning bike-shedding, please don’t add a new syntax for -fdefault-*.

Cheers,

Dominique