Comparing Results between APE100 and APEmille

To avoid abundant rounding differences arising from the MAF (Multiply Add Fused) precision enhancement of the APEmille floating point unit (see intranet/rounding.html), it is necessary that the APEmille codes do not use non-trivial normal operations (but only as separate MUL's and ADD's). For ktc this is achieved by using option "-n", while xtc and rtc do not generate non-trivial normal operations in general. In addition, when using psk one must use the option "-n" to avoid normalization in the optimization step.
Besides rounding differences from the use of normal operations, there may still remain differences between APEmille and APE100 execution due to a minor IEEE- incompatibility of the APE100 MAD (see intranet/rounding.html).

Bit-level Verification between SF and HW

1. SF has to be run with the environment variable

   	SF_USE_FILU
   

   set!!
2. When using psk, its normalization step may change the rounding of the FP arithmetics. Therfore, when using psk either switched off normalization in psk (by using option -n)
   or
   execute SF on the pop'ed basm-file (generated from the jasm file after the pop step of psk: use psk option -k to preserve the jasm-file *-psk.jasm after the pop step and uncomment the line "\include instr.exp")
3. Use binary or hex output formats (not usual ascii format) for floating point data, otherwise least significant bits might be lost!