Intel CPUs need to be configured to correctly execute my code for arbitrary floating-point precision arithmetic and robust geometric predicates. The problem with these CPUs in their default state is that they use extended precision internal floating-point registers, which defeat Dekker's standard techniques for computing the roundoff error of a floating-point operation. These processors can be configured to round internally stored floating-point values to single or double precision, but the commands for doing so are compiler-dependent.
Depending on what C compiler and operating system you use with your PC, you might need to figure out your compiler's procedures for adjusting the floating-point control register, and add them to the file predicates.c or triangle.c. The correct place to add them is at the beginning of the procedure exactinit(). You will probably also need to add an #include statement somewhere early in the file.
If you are compiling with gcc under LINUX, be sure the LINUX symbol is defined during the compilation. If you are compiling with Microsoft C, be sure the CPU86 symbol is defined during the compilation. These symbols cause the inclusion of the following code for configuring the floating-point registers, which I suspect works, but I'm not really sure.
For gcc running under Linux, the statements are as follows. The following line appears with the other #include statements.
#include <fpu_control.h>The following lines appear at the beginning of the procedure exactinit().
int cword; #ifdef SINGLE cword = 4210; /* set FPU control word for single precision */ #else /* not SINGLE */ cword = 4722; /* set FPU control word for double precision */ #endif /* not SINGLE */ _FPU_SETCW(cword);For Microsoft C, the incantations are as follows.
#include <float.h> #ifdef SINGLE _control87(_PC_24, MCW_PC); /* set FPU control word for single precision */ #else /* not SINGLE */ _control87(_PC_53, MCW_PC); /* set FPU control word for double precision */ #endif /* not SINGLE */If you have any corrections to this information, or know how to configure Intel or other FPUs under other compilers or operating systems, please send me email at
.Alternatively, the following might also work with gcc, even if you're not using Linux. (The definition of set_ctrlword() must occur before the definition of exactinit().)
void set_ctrlword(v)
int v;
{
asm("fldcw %0" :: "m" (v));
}
#ifdef SINGLE
set_ctrlword(4210); /* set FPU control word for single precision */
#else /* not SINGLE */
set_ctrlword(4722); /* set FPU control word for double precision */
#endif /* not SINGLE */