A gentle introduction to floating point arithmetic.pptVIP

A gentle introduction to floating point arithmetic Ho Chun Hok (cho@doc.ic.ac.uk) Custom Computing Group Seminar25 Nov 2005 IEEE 754: floating point standard Normal numbers (when exponent 0 and max exponent) v = (-1)^s x 2^exponent x (1.fraction) Subnormal numbers (when exponent = 0) v = (-1)^s x 2^exponent x (0.fraction) Special numbers (when exponent = max exponent) Infinity, Nan (not a number) precisions Single: esize = 8, fsize = 23, vsize = 32 Double: esize = 11, fsize = 52, vsize = 64 Double extended, vsize 64 Operations: +, -, x, /, sqrt, f2i, i2f, compare, f2d, d2f Rounding Nearest even, +inf, -inf, towards 0 What IEEE 754 standard supposes to be Approximation to real number with expected error the epsilon can of any real number can be determined when mapping to floating point number Results of all operations can be correctly rounded in case of inexact result Ensure some math properties hold (in general), x+y==y+x, -(-a) == a, a=b c=0 ? a*c = b*c, x+0 == x, y*y = 0 All exception can be detected Using exception flags Same results across different machines How to ensure the standard? Processor? Rounding numbers in different mode Gradual underflow Raise exceptions Operating System? Handle exception Handle function which may not be supported in hardware (what if a processor cannot handle subnormal number) Keep track of the floating point unit state, (precision, rounding mode) Programming Language? Well-defined semantic for floating point (yes, we have infamous JAVA language) Compiler? Preserve the semantic defined in that language Programmer? read: What Every Computer Scientist Should Know About Floating-Point Arithmetic Case study 1 int main (void) { double ref,index; double tmp; int i; ref = (double) 169.0/ (double) 170.0; for(i=0;i250;i++){ index=i; if(ref == (double) (index/(index+1.0)) ) break; } printf(i=%d\n, i); return 0; } Visual C compiler, running on P-M GCC, running on Pentium 4 (skokie) g