Re: [ANN] LuaProfiler 2.0

[Rici Lake <lua@...>]

On 14-Jun-05, at 7:19 PM, Chris Marrin wrote:

> I have gotten recent experience with QueryPerformanceCounter(). I 
> naively tried to get its value and divide it by the value in 
> QueryPerformanceFrequency(), first converting both to doubles. I found 
> that this conversion reduced the precision down to about 100ms. This 
> is because on my shiny new 3GHz P4 machine, the frequency is 3GHz! 
> That means the counter increments by 3 billion every second. Since 
> it's a 64 bit counter it will still not rollover for about 200 years. 
> But if your machine has been running for a month or so like mine was, 
> the number gets too big to be represented very precisely by a double!
I find this a bit puzzling. A double has 53 bits of precision. A day 
has 86400 seconds; log2 of 86400*3e9 is 47.88; in other words, a double 
can accurately represent a number of 3GHz ticks up to a bit over 34 
days. After that, it should lose one bit of precision, which should be 
barely noticeable. In a year, you would lose less than four bits of 
precision. If you are reduced to a precision of 100ms, you are losing 
something like 28 bits of precision. This suggests that you were doing 
arithmetic with floats rather than doubles; floats have 24 bits of 
precision, or 29 less than doubles, which would work out to about the 
loss you are reporting.