| Re: Sky is falling, maybe? [message #68226] |
Thu, 08 October 2009 13:21 |
Jean H.
Messages: 472
Registered: July 2006
|
Senior Member |
|
|
Jean H. wrote:
> Lasse Clausen wrote:
>
>> I see a distinct difference in the SMOOTH output after the very uppy-
>> downy bit of the data. It seems the documentation should be changed
>> from
>
> Hi,
>
> I don't see any difference here...
>
> IDL> print, !version
> { x86 Win32 Windows Microsoft Windows 7.0.8 Feb 9 2009 32 64}
>
> Jean
IDL> a=smooth(power, 12)
IDL> b=smooth(power, 12, /nan)
IDL> print, where(a-b ne 0)
-1
...and power,a and b are still floats!
Jean
|
|
|
|
| Re: Sky is falling, maybe? [message #68228 is a reply to message #68226] |
Thu, 08 October 2009 13:01  |
Foldy Lajos
Messages: 268
Registered: October 2001
|
Senior Member |
|
|
On Thu, 8 Oct 2009, JohnSmith wrote:
>
> "Jean H." <jghasban@DELTHIS.ucalgary.ANDTHIS.ca> wrote in message
> news:haldn9$lr4$1@news.ucalgary.ca...
>> Lasse Clausen wrote:
>>
>>> I see a distinct difference in the SMOOTH output after the very uppy-
>>> downy bit of the data. It seems the documentation should be changed
>>> from
>>
>> Hi,
>>
>> I don't see any difference here...
>>
>> IDL> print, !version
>> { x86 Win32 Windows Microsoft Windows 7.0.8 Feb 9 2009 32 64}
>>
>> Jean
>
> I did, which seems odd:
>
> IDL> print,!version
> { x86_64 Win32 Windows Microsoft Windows 7.1 Apr 21 2009 64 64}
>
The x86 version uses the x87 FPU (80 bit internally), while the x86_64
version probably uses the SSE/SSE2 FPU unit (strictly 64 bit). This may
create different result (the sky is falling, again :-)
regards,
lajos
|
|
|
|
| Re: Sky is falling, maybe? [message #68229 is a reply to message #68228] |
Thu, 08 October 2009 12:50 |
JohnSmith
Messages: 4
Registered: September 2009
|
Junior Member |
|
|
"Jean H." <jghasban@DELTHIS.ucalgary.ANDTHIS.ca> wrote in message
news:haldn9$lr4$1@news.ucalgary.ca...
> Lasse Clausen wrote:
>
>> I see a distinct difference in the SMOOTH output after the very uppy-
>> downy bit of the data. It seems the documentation should be changed
>> from
>
> Hi,
>
> I don't see any difference here...
>
> IDL> print, !version
> { x86 Win32 Windows Microsoft Windows 7.0.8 Feb 9 2009 32 64}
>
> Jean
I did, which seems odd:
IDL> print,!version
{ x86_64 Win32 Windows Microsoft Windows 7.1 Apr 21 2009 64 64}
|
|
|
|
| Re: Sky is falling, maybe? [message #68230 is a reply to message #68229] |
Thu, 08 October 2009 12:14 |
Jean H.
Messages: 472
Registered: July 2006
|
Senior Member |
|
|
Lasse Clausen wrote:
> I see a distinct difference in the SMOOTH output after the very uppy-
> downy bit of the data. It seems the documentation should be changed
> from
Hi,
I don't see any difference here...
IDL> print, !version
{ x86 Win32 Windows Microsoft Windows 7.0.8 Feb 9 2009 32 64}
Jean
|
|
|
|
| Re: Sky is falling, maybe? [message #68231 is a reply to message #68230] |
Thu, 08 October 2009 12:11 |
Foldy Lajos
Messages: 268
Registered: October 2001
|
Senior Member |
|
|
On Thu, 8 Oct 2009, Lasse Clausen wrote:
> I find the following odd but maybe the sky is just falling and one of
> you guys can explain why this happens. Try running
>
> power = randomu(1001, 150)
> power[77+lindgen(10)*3] = 1e+7
> help, where(~finite(power))
> plot, power, yrange=[.1, 10]
> loadct, 12
> oplot, smooth(power, 12, /nan), thick=3, color=20
> oplot, smooth(power, 12), thick=3, color=120
> end
>
> On my machine
>
> IDL> print, !version
> { x86_64 linux unix linux 7.0 Oct 25 2007 64 64}
>
> I see a distinct difference in the SMOOTH output after the very uppy-
> downy bit of the data. It seems the documentation should be changed
> from
>
> SMOOTH should never be called without the NAN keyword if the input
> array may possibly contain NaN values.
>
> to
>
> SMOOTH should never be called without the NAN. Period.
>
> Again, maybe I'm missing something but the SMOOTH function seems like
> a pretty straight forward piece of code - without ever having seen it,
> of course - that leaves very little room for error. But by the same
> token we all know that "Every program has at least one bug and can be
> shortened by at least one instruction - from which, by induction - it
> can be shown that every program can be reduced to one instruction that
> doesn't work".
>
> So long
> Lasse
>
Yes, the sky is falling, again :-) Try with 'power=double(power)' and the
difference will disappear (randomu and 1e7 are too far apart for float).
smooth(...) and smooth(..., /nan) use different algorithms. The previous
one uses a sliding window, while the other does not (the sliding window
would give all NaNs after the first NaN).
regards,
lajos
|
|
|
|
| Re: Sky is falling, maybe? [message #68232 is a reply to message #68231] |
Thu, 08 October 2009 12:04 |
JohnSmith
Messages: 4
Registered: September 2009
|
Junior Member |
|
|
"Lasse Clausen" <lbnc@lbnc.de> wrote in message
news:771d3ad1-2210-4252-87ad-10af20c3f397@m1g2000vbi.googleg roups.com...
> I find the following odd but maybe the sky is just falling and one of
> you guys can explain why this happens. Try running
>
> power = randomu(1001, 150)
> power[77+lindgen(10)*3] = 1e+7
> help, where(~finite(power))
> plot, power, yrange=[.1, 10]
> loadct, 12
> oplot, smooth(power, 12, /nan), thick=3, color=20
> oplot, smooth(power, 12), thick=3, color=120
> end
>
> On my machine
>
> IDL> print, !version
> { x86_64 linux unix linux 7.0 Oct 25 2007 64 64}
>
> I see a distinct difference in the SMOOTH output after the very uppy-
> downy bit of the data. It seems the documentation should be changed
> from
>
> SMOOTH should never be called without the NAN keyword if the input
> array may possibly contain NaN values.
>
> to
>
> SMOOTH should never be called without the NAN. Period.
my guess, and I have not thought hard about it, is that
the NAN keyword forces smooth to work in double precision internally.
Note: the difference goes away if you put a
power = double(power)
right after the randomu() call.
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
