This is the mail archive of the
fortran@gcc.gnu.org
mailing list for the GNU Fortran project.
RE: cygwin/gfortran 4.3 install error?
Thanks to both of you for your suggestions. I managed to compile and
install gfortran. After this, the cygwin version runs the first test
case reasonably well (correct results, respectable times). Got a bit
better throughput on the make on a dual CPU using make -j 2.
Art
Here's my matmul example. The hand written examples are triple loops
doing a matrix multiply. The intrinsic matmul is the third case.
Arithmetic is 64-bit reals. Problem size is (1000,2000) times
(2000,1000).
art@brahms ~/Fortran/MATMUL
$ gfortran -O3 -msse3 -mtune=i686 -o matmul_g matmul.f90
art@brahms ~/Fortran/MATMUL
$ ./matmul_g
kind real, integer = 8 4
lenx, leny = 1000 2000
loc(a) = 1AFF610
loc(b) = BBD210
loc(c) = 2A41A10
loc(c1) = 41C010
loc(c2) = 31E2C10
hand written time (sec) = 3.125000
hand written time (sec), i/j loops inverted = 13.07800
matmul time (sec) = 3.266001
result difference at: 1 1 407F5874120437D9 407F5874120437D4
STOP mismatch
The above mismatch isn't significant.
Here's an Intel v10 run using the same source file:
art@brahms ~/Fortran/MATMUL
$ ./matmul.10
kind real, integer = 8 4
lenx, leny = 1000 2000
loc(a) = 14009B140
loc(b) = 140FDD540
loc(c) = 141F1F940
loc(c1) = 1426C0B40
loc(c2) = 142E61D40
hand written time (sec) = 2.765625
hand written time (sec), i/j loops inverted = 0.7500000
matmul time (sec) = 0.7968750
So, gfortran's timing results are respectable.
For reference:
art@brahms ~/Fortran/MATMUL
$ uname -a
CYGWIN_NT-5.2-WOW64 brahms 1.5.24(0.156/4/2) 2007-01-31 10:57 i686
Cygwin
art@brahms ~/Fortran/MATMUL
$ which gfortran
/usr/local/bin/gfortran
art@brahms ~/Fortran/MATMUL
$ ls -lt /usr/local/bin/gfortran
-rwxr-xr-x 2 art None 372230 Apr 25 22:46 /usr/local/bin/gfortran
art@brahms ~/Fortran/MATMUL
$ gfortran -v
Using built-in specs.
Target: i686-pc-cygwin
Configured with: ./configure
Thread model: single
gcc version 4.1.2