Re: Fastcode UpperCase B&V 3.0

Hi

A quick summary

> For this to be possible we must decide upon the validity of the new
> functions by Aleksandr. Do they allocate to much memory for the result
> string? If yes - can we accept that?

Accepted

> Then we must select a solution to the problem that we measure a slowdown
if
> validation has been run before the benchmark. Possible solutions:
>
> 1) The cheap hack - do not allow benchmarking after validation.
> 2) Decrease loop counts in Validate 6 and 7 - validation becomes less
> throrough.
> 3) Find the exact problem and the exact cure.

3) Use BucketMem_Asm

> Then we must be sure that there are no problems regarding the benchmarking
> order of functions.

No problem found

> We must also make sure that there is no problem regarding precision in
> "Bechmark" versus "Benchmark All".

Probably no problem. It is not confirmed 100% by the test made so far. A
general improvement in accuracy will be enough to remove any doubts? One
more test will be made. If we are under 1% in error it is OK.

> Then we must recheck the general precision and rework the benchmark if
> needed.

No of reruns increased.

> Then we must find new subbenchmark weigths.

I find the weigths that will make the fastest function score the same in
sub1 and sub2 on the P4 1600. This weigth will be used in LowerCase too.

Best regards
Dennis


.