• downs <default_357-line yahoo.de> Feb 24 2007
• Manfred Nowak <svv1999 hotmail.com> Feb 24 2007
• Dave <Dave_member pathlink.com> Feb 24 2007
• downs <default_357-line yahoo.de> Feb 25 2007
• Dave <Dave_member pathlink.com> Feb 25 2007
• Dave <Dave_member pathlink.com> Mar 08 2007

I just ran the samples/dhry.d compiled with DMD 1.007 vs GDC .23 svn on gcc
4.1.1 using MinGW, on a Pentium M with 1.6g.
The used flags were -O -release -inline for DMD and -ffast-math -O3 -frelease
-finline-functions for GDC.
The observed VAX MIPS ratings were 3280 for GDC, but only 2570 for DMD.
(Note that the GDC rating could be increased to 3480 by adding -msse
-mfpmath=sse,387, but that is somewhat risky).
All tests were run with realtime priority.
Do you think this difference in a synthetic benchmark translates into an actual
speed difference for real-world applications?
Did I miss any DMD compiler flags?
Greetings   --downs

downs wrote

 I just ran the samples/dhry.d


Warning! That is a sample only!

I remember tests of my own with similar results. But on more careful 
analysis it turned out, that dhry.d seems to be unusable as a 
benchmark.

AFAIR this was due to some function f() which converges to the value 
1.0. GCC's math library allowed that function to reach 1.0, whereas DMD 
let it stay at 1.0-epsilon.

I did not analyse that any further because I am at no means an expert 
in benchmarking and in addition do not want to explore that field.

My impression was though, that because f() was used in some 
multiplication GCC may have made itself an unfair advantage: 
multiplication with 1.0 seems to be much easier (and therefore faster!)
than multiplication with 1.0-epsilon.

-manfred

downs wrote:
 I just ran the samples/dhry.d compiled with DMD 1.007 vs GDC .23 svn on gcc
4.1.1 using MinGW, on a Pentium M with 1.6g.
 The used flags were -O -release -inline for DMD and -ffast-math -O3 -frelease
-finline-functions for GDC.
 The observed VAX MIPS ratings were 3280 for GDC, but only 2570 for DMD.
 (Note that the GDC rating could be increased to 3480 by adding -msse
-mfpmath=sse,387, but that is somewhat risky).
 All tests were run with realtime priority.
 Do you think this difference in a synthetic benchmark translates into an
actual speed difference for real-world applications?
 Did I miss any DMD compiler flags?
 Greetings   --downs


The DMD version maybe using a higher resolution timer -- GetTickCount() vs
time() for the GDC 
version. If so, that may have a lot to do with it.

On a P4 linux box, if I use an external timer ('time'), DMD consistently
outperforms GDC with the 
switches above by about 15% or so.

Dave Wrote:

 downs wrote:
 I just ran the samples/dhry.d compiled with DMD 1.007 vs GDC .23 svn on gcc
4.1.1 using MinGW, on a Pentium M with 1.6g.
 The used flags were -O -release -inline for DMD and -ffast-math -O3 -frelease
-finline-functions for GDC.
 The observed VAX MIPS ratings were 3280 for GDC, but only 2570 for DMD.
 (Note that the GDC rating could be increased to 3480 by adding -msse
-mfpmath=sse,387, but that is somewhat risky).
 All tests were run with realtime priority.
 Do you think this difference in a synthetic benchmark translates into an
actual speed difference for real-world applications?
 Did I miss any DMD compiler flags?
 Greetings   --downs


 The DMD version maybe using a higher resolution timer -- GetTickCount() vs
time() for the GDC 
 version. If so, that may have a lot to do with it.
 
 On a P4 linux box, if I use an external timer ('time'), DMD consistently
outperforms GDC with the 
 switches above by about 15% or so.


18,219s for GDC vs 22,422s for DMD. So the difference seems less, but it's
still there.

downs wrote:
 Dave Wrote:
 
 downs wrote:
 I just ran the samples/dhry.d compiled with DMD 1.007 vs GDC .23 svn on gcc
4.1.1 using MinGW, on a Pentium M with 1.6g.
 The used flags were -O -release -inline for DMD and -ffast-math -O3 -frelease
-finline-functions for GDC.
 The observed VAX MIPS ratings were 3280 for GDC, but only 2570 for DMD.
 (Note that the GDC rating could be increased to 3480 by adding -msse
-mfpmath=sse,387, but that is somewhat risky).
 All tests were run with realtime priority.
 Do you think this difference in a synthetic benchmark translates into an
actual speed difference for real-world applications?
 Did I miss any DMD compiler flags?
 Greetings   --downs


 version. If so, that may have a lot to do with it.

 On a P4 linux box, if I use an external timer ('time'), DMD consistently
outperforms GDC with the 
 switches above by about 15% or so.


 18,219s for GDC vs 22,422s for DMD. So the difference seems less, but it's
still there.


Hmmm - maybe something to do with MinGW / Pentium M?

Also I have GCC v4.0.3, gdc 0.21 installed but I don't think it would make that
big of diff (not a 
complete reversal of the results anyway)??

Here's what I get:

# dmd -O -inline -release dhry -ofdhry_dmd; for x in 1 2 3; do time dhry_dmd
>/dev/null; done
gcc dhry.o -o dhry_dmd -m32 -lphobos -lpthread -lm

real    0m2.221s
user    0m2.200s
sys     0m0.008s

real    0m2.421s
user    0m2.408s
sys     0m0.004s

real    0m2.299s
user    0m2.288s
sys     0m0.000s

# gdc -ffast-math -O3 -frelease -finline-functions -msse -mfpmath=sse dhry.d -o
dhry_gdc; for x in 1 
2 3; do time dhry_gdc >/dev/null; done

real    0m2.827s
user    0m2.816s
sys     0m0.004s

real    0m2.827s
user    0m2.812s
sys     0m0.004s

real    0m2.802s
user    0m2.788s
sys     0m0.000s

I'll see about upgrading gdc (something I've been meaning to do anyhow) and
check it out...

Dave wrote:
 downs wrote:
 Dave Wrote:

 downs wrote:
 I just ran the samples/dhry.d compiled with DMD 1.007 vs GDC .23 svn 
 on gcc 4.1.1 using MinGW, on a Pentium M with 1.6g.
 The used flags were -O -release -inline for DMD and -ffast-math -O3 
 -frelease -finline-functions for GDC.
 The observed VAX MIPS ratings were 3280 for GDC, but only 2570 for DMD.
 (Note that the GDC rating could be increased to 3480 by adding -msse 
 -mfpmath=sse,387, but that is somewhat risky).
 All tests were run with realtime priority.
 Do you think this difference in a synthetic benchmark translates 
 into an actual speed difference for real-world applications?
 Did I miss any DMD compiler flags?
 Greetings   --downs


 GetTickCount() vs time() for the GDC version. If so, that may have a 
 lot to do with it.

 On a P4 linux box, if I use an external timer ('time'), DMD 
 consistently outperforms GDC with the switches above by about 15% or so.


 18,219s for GDC vs 22,422s for DMD. So the difference seems less, but 
 it's still there.


 Hmmm - maybe something to do with MinGW / Pentium M?
 
 Also I have GCC v4.0.3, gdc 0.21 installed but I don't think it would 
 make that big of diff (not a complete reversal of the results anyway)??
 
 Here's what I get:
 
 # dmd -O -inline -release dhry -ofdhry_dmd; for x in 1 2 3; do time 
 dhry_dmd >/dev/null; done
 gcc dhry.o -o dhry_dmd -m32 -lphobos -lpthread -lm
 
 real    0m2.221s
 user    0m2.200s
 sys     0m0.008s
 
 real    0m2.421s
 user    0m2.408s
 sys     0m0.004s
 
 real    0m2.299s
 user    0m2.288s
 sys     0m0.000s
 
 # gdc -ffast-math -O3 -frelease -finline-functions -msse -mfpmath=sse 
 dhry.d -o dhry_gdc; for x in 1 2 3; do time dhry_gdc >/dev/null; done
 
 real    0m2.827s
 user    0m2.816s
 sys     0m0.004s
 
 real    0m2.827s
 user    0m2.812s
 sys     0m0.004s
 
 real    0m2.802s
 user    0m2.788s
 sys     0m0.000s
 
 I'll see about upgrading gdc (something I've been meaning to do anyhow) 
 and check it out...


Using gdc 0.23 & gcc 4.1.1, and DMD v1.007 (slight improvement for gdc, but DMD
is still a tad 
faster on my P4 2.2 Ghz, FC5 machine):

# dmd -O -inline -release dhry -ofdhry_dmd; for x in 1 2 3; do time dhry_dmd
>/dev/null; done
gcc dhry.o -o dhry_dmd -m32 -lphobos -lpthread -lm

real    0m2.297s
user    0m2.288s
sys     0m0.000s

real    0m2.223s
user    0m2.212s
sys     0m0.004s

real    0m2.310s
user    0m2.300s
sys     0m0.000s

# gdc -ffast-math -O3 -frelease -finline-functions -msse -mfpmath=sse dhry.d -o
dhry_gdc; for x in 1 
2 3; do time dhry_gdc >/dev/null; done

real    0m2.598s
user    0m2.584s
sys     0m0.000s

real    0m2.595s
user    0m2.580s
sys     0m0.004s

real    0m2.640s
user    0m2.620s
sys     0m0.004s