Thread: double precision error on Mac OSX

Hi users,

I am having problems when I want to code in double precision and compile on a Max OSX, 10.5, darwin. I compile with gcc-4.0.1

Would that be a compiler fault on the machine? Is it machine dependent? How can I solve this issue?
What I obtain is actually a wrong result; I proved it by coding the same software suing floats instead.

Is anyone a Mac user here and who normally uses doubles instead of floats?

thank you in advance,
All the best,
cfd

Yes, I use a Mac and use doubles and have not had any problems (not directly attributable to my failure to write working code). Do you have a problem? If so, what is it, actually?

hi tabstop,

This is the problem: when I code my program using floats the software performs correctly, but when I change "float" with "double" everywhere, the code still compiles, but the executable outputs wrong numbers. Very different results from the float case

thanks for helping

"Very different results from the float case" is kind of what you would expect, n'est-ce pas? I presume you're doing a "toy" problem where you know the answers to make sure everything's good. Are you calling other (people's) code? If so, did that change to double too?

hi, actually, I would expect different results within certain limits (after a certain floating point); what I am getting is NAN where I am supposed to obtain real value.

I am not calling functions written elsewhere in this code actually. I need to use doubles for the problem at hand (grid generation).

You should show some code. NaN often derives from dividing by zero, but it doesn't have to.

Hi again tabstop,

I would like to show the code, but it's an actual software that I am developing and it has about 15 functions. What I did, I entered eachone of the them and changed the floats into double where needed (including the *.h). In compilation I get no warning or errors and it executes fine. However, when I go and plot my output, that is when I get the unexpected error.

If it is difficult to show the code and that is the only way to try to solve the problem, I will try again to substitute the floats as little at the time as possible, and compile step by step.

thanks again for helping. It helped knowing that you usually code in double on your os x and it works fine.

Best
cfd

If you get totally different results, like NaN maybe you have some issue with declarations. Without some code we can't tell if your functions just return doubles or you pass pointers or maybe you have doubles inside some structs so just make sure all declarations are proper, you have included all required headers and such. Do you compile with default settings (which are useless / dangerous) or do you use something like -Wall as compiler flags?

Hi idelovski, thanks for writing,

SInce it is kind of difficult to send the code or part of it, I can surely send the makefile:

Code:

CC = gcc     #Compiler is gcc
LD = gcc -lm #Linker is gcc also

CFALGS = -Wall -shared
CDEBUG = -g

#CBLAS = -lgsl -lgslcblas

# Linker flags go here. Currently there aren't any, but if we'll switch to
# code optimization, we might add "-s" here to strip debug info and symbols.
LDFLAGS =

# List of source codes and generated object files:
SRCS = main.c nrutil.c visit_writer.c load_input.c PRINT.c linspace.c nrspline.c define_boundary.c parabola.c gaussj.c interpol_spline.c arco.c TFI.c GRID2CONN.c wrt2plotfile.c external_geometries.c
OBJS = main.o nrutil.o visit_writer.o load_input.o PRINT.o linspace.o nrspline.o define_boundary.o parabola.o gaussj.o interpol_spline.o arco.o TFI.o GRID2CONN.o wrt2plotfile.o external_geometries.o

# ProgRam executable file name:
EXE = mymeshGen2D_struct.a

# Top-level rule, to compile everything
all: $(EXE)

# Rule to Link the program
$(EXE): $(OBJS)
	$(LD) -L/usr/local/lib $(OBJS) $(GSLLIB) $(CBLAS) -o $(EXE)
	
# Rule for all files:
$(OBJS): $(SRCS)
	$(CC) $(CFLAGS) $(CDEBUG) -I/usr/include -c $(SRCS)

I will be re-checking the arguments and declarations again.

Thank you

cfdprogrammer, unless you are explicitly prohibited from posting the code, there's not really a reason not to. It's not like someone is going to steal your CFD code and go and write their own from yours...there are already plenty of mature free/open-source CFD programs out there (code_saturne, Elmer, Gerris Flow Solver) that are far more appealing than unfinished code.

Sometimes, when you do a operation with left result as double, and expression
having a "long", some compilers might require a explicit cast:
long a;
double b = a / a;

Such things may not happen using float as left since it fits the legacies address sizes
and usually the compiler knows how to handle it.
float c = a / a; // may not have any problems with that

so

double d = (double) a / a; would fix your problem

Regards,
yezaim

Originally Posted by yezaim

Sometimes, when you do a operation with left result as double, and expression
having a "long", some compilers might require a explicit cast:
long a;
double b = a / a;

Such things may not happen using float as left since it fits the legacies address sizes
and usually the compiler knows how to handle it.
float c = a / a; // may not have any problems with that

so

double d = (double) a / a; would fix your problem

Regards,
yezaim

Hi yezaim,

thanks for helping. This is the hint I needed. It is a problem in the use of doubles on a machine I barely started to use.

Thank you again, I appreciate it
Best