This is the mail archive of the
fortran@gcc.gnu.org
mailing list for the GNU Fortran project.
Polyhedron Benchmark Tests - TEST_FPU.F90
- From: THOMAS Paul Richard <prthomas at drfccad dot cea dot fr>
- To: "'fortran at gcc dot gnu dot org'" <fortran at gcc dot gnu dot org>
- Cc: "'paulthomas2 at wanadoo dot fr'" <paulthomas2 at wanadoo dot fr>
- Date: Fri, 4 Mar 2005 14:55:36 +0100
- Subject: Polyhedron Benchmark Tests - TEST_FPU.F90
I just compiled a version of the above (http://www.polyhedron.co.uk/), after
eliminating code that exposed two bugs:
(A) This causes a segmentation fault before execution starts
REAL(RK8) :: pool(101,101,1000), pool3(1001,1001) ! random numbers to invert
EQUIVALENCE (pool,pool3) ! use same pool numbers for test 3,4
(B) Constructs such as
index((/j,n+1-m/)) = index((/n+1-m,j/))
cause:
pb04/win/f90/test_fpu.f90:106: internal compiler error: in
gfc_conv_ss_descriptor, at fortran/trans-array.c:1261
whilst:
imax = MAXLOC(ABS( (/ (b(j+i-1,i),i=1,n-j+1) /) ))
produces
TEST_FPU.f90:165: fatal error: gfc_todo: Not Implemented: Unable to find
scalarization loop specifier
The above will be the subject of PRs to be posted some time this weekend,
after checking with CVS version and the existing PRs.
The workarounds are quite easy andthe following comparison is quite
interesting:
All run from the Cygwin window/Windows2000/PentiumIV 2.4GHz
************************CYGWIN/gfortran 20050131****************************
$ gfc -O2 TEST_FPU.f90
$ ./a
Benchmark running, hopefully as only ACTIVE task
Test1 - Gauss 2000 (101x101) inverts 38.8 sec Err= 0.000000000000002
Test2 - Crout 2000 (101x101) inverts 9.4 sec Err= 0.000000000000000
Test3 - Crout 2 (1001x1001) inverts 11.5 sec Err= 0.000000000000002
Test4 - Lapack 2 (1001x1001) inverts 8.4 sec Err= 0.000000000000358
total = 68.1 sec
************************CYGWIN/g95
20050208**********************************
$ g95 -O2 TEST_FPU.f90
$ ./a
Benchmark running, hopefully as only ACTIVE task
Test1 - Gauss 2000 (101x101) inverts 49.3 sec Err= 0.000000000000003
Test2 - Crout 2000 (101x101) inverts 24.5 sec Err= 0.000000000000001
Test3 - Crout 2 (1001x1001) inverts 14.4 sec Err= 0.000000000000002
Test4 - Lapack 2 (1001x1001) inverts 16.9 sec Err= 0.000000000000245
total =105.1 sec
************************Windows/gfc
20050208*********************************
$ d:/gfc/bin/gfortran -O2 TEST_FPU.f90
$ ./a
Benchmark running, hopefully as only ACTIVE task
Test1 - Gauss 2000 (101x101) inverts 20.0 sec Err= 0.000000000000002
Test2 - Crout 2000 (101x101) inverts 8.0 sec Err= 0.000000000000000
Test3 - Crout 2 (1001x1001) inverts 12.0 sec Err= 0.000000000000002
Test4 - Lapack 2 (1001x1001) inverts 8.0 sec Err= 0.000000000000358
total = 48.0 sec
************************Windows/DF5.0
rev-D**********************************
> C:\df>f90 /FAST test_fpu.f90 /LINK /STACK:8000000
$ c:/df/test_fpu
Benchmark running, hopefully as only ACTIVE task
Test1 - Gauss 2000 (101x101) inverts 5.6 sec Err= 0.000000000000003
Test2 - Crout 2000 (101x101) inverts 6.0 sec Err= 0.000000000000003
Test3 - Crout 2 (1001x1001) inverts 11.1 sec Err= 0.000000000000042
Test4 - Lapack 2 (1001x1001) inverts 8.0 sec Err= 0.000000000000583
total = 30.6 sec
Thus, the Windows version of gfortran is competitive with Digital, except
for test1. I will attempt to identify the source of the factor 3 in time
for Test1. Typical times of ~19.5-23.8 sec are reported for WIndows on a
3.06GHz PIV by Polyhedron for Intel/VisualDF/LF95. FTN95 is 39.3s
The Windows version of g95 now comes bundled with its own cygwin1.dll, so I
did not dare screw up my existing Cygwin installation. Although a lot
slower on this version of TEST_FPU.F90, it should be said that g95 compiles
it out of the box. The execution time is about the same for either version
of TEST_FPU.