From: krishna.myneni@ccreweb.org   
      
   On 5/28/24 06:11, Krishna Myneni wrote:   
   ...   
   > Updated benchmark code also allows for validating the arithmetic of F*/   
   > for all of the 1 million test cases. At the moment, all arithmetic cases   
   > use only "normal" fp numbers i.e. they do not include zero, infinity, or   
   > subnormal numbers -- special tests are needed for such cases.   
   > ...   
      
   I found the reason for hanging in my implementation of F*/ -- the   
   problem was in the word DP_NORMAL_FRACTION (the only word which has an   
   indefinite loop in the code). The version which doesn't go into an   
   infinite loop is given below:   
      
   : dp-normal-fraction ( u -- expinc u')   
       dup DP_FRACTION_BITS 1+ rshift   
       dup IF   
         tuck rshift   
       ELSE   
         drop -1 swap \ -1 u   
         BEGIN   
           dup DP_IMPLIED_BIT and 0=   
         WHILE   
           1 lshift   
           swap 1- swap   
         REPEAT   
       THEN   
      
      
   Now the benchmark code runs. It shows two things (see below):   
      
   1. The execution time is poor compared to Anton's implementation, about   
   8x slower.   
      
   2. Half of the arithmetic results are wrong, with the max relative error   
   being as high as 2/3.   
      
   So, I clearly have some work to do to fix the arithmetic. It was   
   fortuitous that it passed my test cases, but clearly the test cases in   
   my original implementation of F*/ were insufficient.   
      
   --   
   Krishna   
      
      
   bench-f*/   
   603 ms   
     ok   
   check-results   
      
   1000000 samples   
      
   # tolerance errors: 501374   
   # sign errors:           0   
   Max relative error: 0.666667   
     ok   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)