Re: Floating point constant question

In my opinion, it depends on the programmer what accuracy is acceptable
(and hence what result are valid). Luckily, SPEC agrees. SPEC actually
allows deviation in results for SPEC89/92. The below was gleaned from
SPECfp92 benchmark Makefiles:


Benchmark Allowed difference in results
========= =============================


013.spice2g6: Relative tolerance 0.01
015.doduc: Absolute tolerance 0.009 or Relative tolerance 0.05
for result.float
Absolute tolerance 0.009 or Relative tolerance 0.02
for result.iter
034.mdljdp2: Relative tolerance 0.025
039.wave5: Relative tolerance 0.03
047.tomcatv: Relative tolerance 0.001
052.alvinn: Relative tolerance 0.001
056.ear: Absolute tolerance 1e-7
077.mdljsp2: Relative tolerance 0.01
078.swm256: Relative tolerance 0.01 -a0.001
089.su2cor: Relative tolerance 0.01
090.hydro2d: Relative tolerance 0.01
094.fpppp: Relative tolerance 0.000000001 for result.float
Relative tolerance 0.00025 for result.iter


The definitions of these terms (paraphrased from the spiff man page):


Relative tolerance: A fraction of the larger (in absolute terms) of the
two floating point numbers being compared. Thus, a relative tolerance of
0.1 will cause the two floating point numbers 1.0 and 0.9 to be deemed
within tolerance. The numbers 1.0 and 0.89 will be outside the tolerance.


Absolute tolerance: A fraction representing the maximum allowed absolute
difference between numbers.
--
Tom Conte Department of Electrical and Computer Engineering
conte@ece.scarolina.edu University of South Carolina, Columbia, SC 29208
--

Return to the comp.compilers page.
Search the comp.compilers archives again.