|
楼主 |
发表于 2006-3-20 13:40:06
|
显示全部楼层
又出了个s38版本,试图解决之前针对FPU的版本所带来的计算精度方面的问题。
并不是说之前的版本计算精度差,其实精度过高也可能导致计算结果验证失败,因为绝大多数用户用的还是官方的windows版本,结果验证的时候一般也就是参考这个版本的计算结果。
作者原话:
I tried to find the reason of the validation problem.
I know that the precision of C37 has to be very near to the original.
I made a code that gave poor results, but the validator accepted them!
So, I examined the codes again and I saw there is a critical point.
Sometimes the exponent of the number format is not enough.
Probably this is the reason of the validation problem of SSE based linux application too.
exponent size of SSE -> 8 bit
exponent size of SSE2 -> 11 bit
exponent size of FPU -> 15 bit
C37 uses FPU (maximum precision) for the calculation, but it needs about 4 bit longer exponent than original.
I checked C37 on my computers with 70-80 units without any error.
A longer test period would be able to determine the frequency of the faults.
I hope that S38 will solve this exponent problem.
It consists a very aggressive precision combination method to extend the precision.
[ Last edited by Youth on 2006-3-20 at 13:41 ] |
|