| Re: Turning off math error checking for a code block [message #28871] |
Fri, 18 January 2002 07:47  |
Paul van Delst
Messages: 364
Registered: March 1997
|
Senior Member |
|
|
Kenneth Bowman wrote:
>
> In article <888zawn5bq.fsf@catspaw.jpl.nasa.gov>, Vapuser <vapuser@catspaw.jpl.nasa.gov> wrote:
>
>> The help says that for !except=1 (the default) it only reports the
>> exception upon arriving at back at an interactive prompt. On my SGI, it
>> still does the `where' and returns the correct answer, it just
>> complains.
>
> The problem is that changing !EXCEPT does not reset the accumulated error status, so the errors that occur with !EXCEPT = 0 are still present. Hence the need to call CHECK_MATH (twice).
>
> I want to be able to run my whole code with !EXCEPT = 2 (or 1) but not generate errors for the blocks that I *know* will have floating point exceptions in them.
>
> I think my solution is probably the simplest way to do what I want to do.
Given the (valid) reason the NaN's are in your array (flagging missing values) and that you
want to _not_ accumulate certain well-identified errors, I think you're right.
paulv
--
Paul van Delst Religious and cultural
CIMSS @ NOAA/NCEP purity is a fundamentalist
Ph: (301)763-8000 x7274 fantasy
Fax:(301)763-8545 V.S.Naipaul
|
|
|
|
| Re: Turning off math error checking for a code block [message #28876 is a reply to message #28871] |
Thu, 17 January 2002 16:06  |
Craig Markwardt
Messages: 1869
Registered: November 1996
|
Senior Member |
|
|
k-bowman@null.com (Kenneth Bowman) writes:
> In article <3C47094C.1F1879D2@ssec.wisc.edu>, "Liam E. Gumley" <Liam.Gumley@ssec.wisc.edu> wrote:
>
>> The FINITE function returns 1 where the argument is finite, and 0 where
>> the argument is infinite *or* NaN (see p. 134 of my book). Try the
>> following:
>>
>> x_min = 2.0
>> index = where(finite(x) eq 1, count)
>> if (count gt 0) then print, where(x[index] lt x_min)
>
> I am aware of that. These are relatively large vectors (10^5 to
> 10^6 elements), however, and this operation is repeated many times,
> so I am trying to avoid extracting the finite values (or creating an
> array index to them). This is my "innermost loop", and efficiency
> is important. I know there are NaN's. I prefer to simply turn off
> the error messages.
I have found that an operation on an array which contains NANs is
slowed down considerably. I think it is because each operation causes
a floating point exception which is handled in the OS. I use WHERE
most of the time when this comes up. Occassionally I get "floating
exception" messages, but big whoop.
Craig
--
------------------------------------------------------------ --------------
Craig B. Markwardt, Ph.D. EMAIL: craigmnet@cow.physics.wisc.edu
Astrophysics, IDL, Finance, Derivatives | Remove "net" for better response
------------------------------------------------------------ --------------
|
|
|
|
|
|
|
|
| Re: Turning off math error checking for a code block [message #28881 is a reply to message #28879] |
Thu, 17 January 2002 13:55 |
Vapuser
Messages: 63
Registered: November 1998
|
Member |
|
|
Paul van Delst <paul.vandelst@noaa.gov> writes:
> Vapuser wrote:
>>
>> k-bowman@null.com (Kenneth Bowman) writes:
>>>
>>> I have an array x that is likely to have missing values in it,
>>> indicated by NaN's. I would like to search the array for values
>>> less than x_min. Because of the NaN's, WHERE generates a
>>> floating point error, e.g.,
>>>
>>> IDL> print, x
>>> 0.00000 NaN 2.00000 3.00000
>>> IDL> print, where(x lt 2.0)
>>> 0
>>> % Program caused arithmetic error: Floating illegal operand
>>>
>>
>> Hmmmm..... I don't get this result.
>
> Sure you do - you just have a 0 instead of a 3 as the 4th element of y:
>
Oooopppsss. My error! I guess I just can't read today. For some
reason I thought that Ken's x[3] == 0 too!
mea culpa.
>> IDL> y=[0,!values.f_nan,2,0.]
>
> The where result of the original is one index - the 0th one:
>
Yes.
> IDL> x=[0,!values.f_nan,2,3.]
> IDL> print,where( x LT 2,nx),nx
> 0
> 1
> % Program caused arithmetic error: Floating illegal operand
>
> paulv
>
> p.s. After trying the simple example above, it sure is annoying to
> get that error message. :o\
>
True, but setting !except=0 makes that go away.
whd
--
William Daffer: 818-354-0161: William.Daffer@jpl.nasa.gov
|
|
|
|
|
|
| Re: Turning off math error checking for a code block [message #28888 is a reply to message #28883] |
Thu, 17 January 2002 12:54 |
Paul van Delst
Messages: 364
Registered: March 1997
|
Senior Member |
|
|
Vapuser wrote:
>
> k-bowman@null.com (Kenneth Bowman) writes:
>
>> I have an array x that is likely to have missing values in it, indicated by NaN's. I would like to search the array for values less than x_min. Because of the NaN's, WHERE generates a floating point error, e.g.,
>>
>> IDL> print, x
>> 0.00000 NaN 2.00000 3.00000
>> IDL> print, where(x lt 2.0)
>> 0
>> % Program caused arithmetic error: Floating illegal operand
>>
>
> Hmmmm..... I don't get this result.
Sure you do - you just have a 0 instead of a 3 as the 4th element of y:
> IDL> y=[0,!values.f_nan,2,0.]
The where result of the original is one index - the 0th one:
IDL> x=[0,!values.f_nan,2,3.]
IDL> print,where( x LT 2,nx),nx
0
1
% Program caused arithmetic error: Floating illegal operand
paulv
p.s. After trying the simple example above, it sure is annoying to get that error message. :o\
--
Paul van Delst Religious and cultural
CIMSS @ NOAA/NCEP purity is a fundamentalist
Ph: (301)763-8000 x7274 fantasy
Fax:(301)763-8545 V.S.Naipaul
|
|
|
|
| Re: Turning off math error checking for a code block [message #28890 is a reply to message #28888] |
Thu, 17 January 2002 11:30 |
Vapuser
Messages: 63
Registered: November 1998
|
Member |
|
|
k-bowman@null.com (Kenneth Bowman) writes:
> I have an array x that is likely to have missing values in it, indicated by NaN's. I would like to search the array for values less than x_min. Because of the NaN's, WHERE generates a floating point error, e.g.,
>
> IDL> print, x
> 0.00000 NaN 2.00000 3.00000
> IDL> print, where(x lt 2.0)
> 0
> % Program caused arithmetic error: Floating illegal operand
>
Hmmmm..... I don't get this result.
IDL Version 5.3 (IRIX mipseb). (c) 1999, Research Systems, Inc.
Installation number: 12619.
Licensed for use by: Jet Propulsion Lab
IDL> y=[0,!values.f_nan,2,0.]
IDL> print,where( y LT 2,nx),nx
0 3
2
% Program caused arithmetic error: Floating illegal operand
IDL> !except=0
IDL> print,where( y LT 2,nx),nx
0 3
2
IDL> exit
The help says that for !except=1 (the default) it only reports the
exception upon arriving at back at an interactive prompt. On my SGI, it
still does the `where' and returns the correct answer, it just
complains.
>
> As best I understand the interaction between !EXCEPT and CHECK_MATH,
> in order to suppress this error message, while still checking errors
> elsewhere in the code, I must do the following:
>
> error = CHECK_MATH(/PRINT) ;If any errors have occurred, print
> save_except = !EXCEPT ;Save current exception flag
> !EXCEPT = 0 ;Set exception flag to 0
> i = WHERE(x LT x_min, ni) ;Find all x < x_min
> error = CHECK_MATH() ;Clear accumulated error status
> !EXCEPT = save_except ;Restore exception flag
>
> Am I making this harder than it needs to be?
>
You could use FINITE. But why bother? Your method saves you one
iteration over the array.
whd
--
William Daffer: 818-354-0161: William.Daffer@jpl.nasa.gov
|
|
|
|
| Re: Turning off math error checking for a code block [message #28891 is a reply to message #28890] |
Thu, 17 January 2002 12:30 |
John-David T. Smith
Messages: 384
Registered: January 2000
|
Senior Member |
|
|
Paul van Delst wrote:
>
> Kenneth Bowman wrote:
>>
>> In article <3C47094C.1F1879D2@ssec.wisc.edu>, "Liam E. Gumley" <Liam.Gumley@ssec.wisc.edu> wrote:
>>
>>> The FINITE function returns 1 where the argument is finite, and 0 where
>>> the argument is infinite *or* NaN (see p. 134 of my book). Try the
>>> following:
>>>
>>> x_min = 2.0
>>> index = where(finite(x) eq 1, count)
>>> if (count gt 0) then print, where(x[index] lt x_min)
>>
>> I am aware of that. These are relatively large vectors (10^5 to 10^6 elements),
>> however, and this operation is repeated many times, so I am trying to avoid
>> extracting the finite values (or creating an array index to them). This is my
>> "innermost loop", and efficiency is important. I know there are NaN's. I prefer
>> to simply turn off the error messages.
>
> Hmm. This is straying way off topic...and don't take it the wrong way or anything, but how come
> you don't prefer to simply prevent the NaNs from occurring in the first place?
>
> (To the NG) Does IDL stop processing compound logical tests before they're completed? What
> about:
>
> i = WHERE((FINITE(x) EQ 1) AND (x LT x_min), ni)
>
> Will the second test for any particular index value still get performed if the first one fails?
> I should look this up in me IDL book, I know......
No, IDL booleans are not short-circuiting, which is a real pain for
cases like:
if n_elements(a) ne 0 AND a eq 4 then ....
which requires more deeply nested if's to pull off. That said, they
*are* overloaded to perform array based boolean computations, so it's
not clear if a short-circuiting version would even have been possible.
The only solution might have been to keep the array and scalar boolean
operators separate.
JD
|
|
|
|
| Re: Turning off math error checking for a code block [message #28892 is a reply to message #28890] |
Thu, 17 January 2002 12:23 |
Paul van Delst
Messages: 364
Registered: March 1997
|
Senior Member |
|
|
Kenneth Bowman wrote:
>
> In article <3C47094C.1F1879D2@ssec.wisc.edu>, "Liam E. Gumley" <Liam.Gumley@ssec.wisc.edu> wrote:
>
>> The FINITE function returns 1 where the argument is finite, and 0 where
>> the argument is infinite *or* NaN (see p. 134 of my book). Try the
>> following:
>>
>> x_min = 2.0
>> index = where(finite(x) eq 1, count)
>> if (count gt 0) then print, where(x[index] lt x_min)
>
> I am aware of that. These are relatively large vectors (10^5 to 10^6 elements),
> however, and this operation is repeated many times, so I am trying to avoid
> extracting the finite values (or creating an array index to them). This is my
> "innermost loop", and efficiency is important. I know there are NaN's. I prefer
> to simply turn off the error messages.
Hmm. This is straying way off topic...and don't take it the wrong way or anything, but how come
you don't prefer to simply prevent the NaNs from occurring in the first place?
(To the NG) Does IDL stop processing compound logical tests before they're completed? What
about:
i = WHERE((FINITE(x) EQ 1) AND (x LT x_min), ni)
Will the second test for any particular index value still get performed if the first one fails?
I should look this up in me IDL book, I know......
paulv
>
> P.S. For symmetry, wouldn't it be nice to have an INFINITE function?
And don't forget the complementary NOT_A_NAN() function. :o)
--
Paul van Delst Religious and cultural
CIMSS @ NOAA/NCEP purity is a fundamentalist
Ph: (301)763-8000 x7274 fantasy
Fax:(301)763-8545 V.S.Naipaul
|
|
|
|
| Re: Turning off math error checking for a code block [message #28894 is a reply to message #28890] |
Thu, 17 January 2002 11:48 |
k-bowman
Messages: 12
Registered: December 2001
|
Junior Member |
|
|
In article <3C47094C.1F1879D2@ssec.wisc.edu>, "Liam E. Gumley" <Liam.Gumley@ssec.wisc.edu> wrote:
> The FINITE function returns 1 where the argument is finite, and 0 where
> the argument is infinite *or* NaN (see p. 134 of my book). Try the
> following:
>
> x_min = 2.0
> index = where(finite(x) eq 1, count)
> if (count gt 0) then print, where(x[index] lt x_min)
I am aware of that. These are relatively large vectors (10^5 to 10^6 elements), however, and this operation is repeated many times, so I am trying to avoid extracting the finite values (or creating an array index to them). This is my "innermost loop", and efficiency is important. I know there are NaN's. I prefer to simply turn off the error messages.
Perhaps a /NAN keyword on the WHERE function would be useful? (as in, for example, TOTAL)
Ken
P.S. For symmetry, wouldn't it be nice to have an INFINITE function?
|
|
|
|
|
|
| Re: Turning off math error checking for a code block [message #28897 is a reply to message #28896] |
Thu, 17 January 2002 09:26 |
Liam E. Gumley
Messages: 378
Registered: January 2000
|
Senior Member |
|
|
Kenneth Bowman wrote:
>
> I have an array x that is likely to have missing values in it, indicated by NaN's.
> I would like to search the array for values less than x_min. Because of the NaN's,
> WHERE generates a floating point error, e.g.,
>
> IDL> print, x
> 0.00000 NaN 2.00000 3.00000
> IDL> print, where(x lt 2.0)
> 0
> % Program caused arithmetic error: Floating illegal operand
>
> As best I understand the interaction between !EXCEPT and CHECK_MATH,
> in order to suppress this error message, while still checking errors elsewhere
> in the code, I must do the following:
>
> error = CHECK_MATH(/PRINT) ;If any errors have occurred, print
> save_except = !EXCEPT ;Save current exception flag
> !EXCEPT = 0 ;Set exception flag to 0
> i = WHERE(x LT x_min, ni) ;Find all x < x_min
> error = CHECK_MATH() ;Clear accumulated error status
> !EXCEPT = save_except ;Restore exception flag
>
> Am I making this harder than it needs to be?
The FINITE function returns 1 where the argument is finite, and 0 where
the argument is infinite *or* NaN (see p. 134 of my book). Try the
following:
x_min = 2.0
index = where(finite(x) eq 1, count)
if (count gt 0) then print, where(x[index] lt x_min)
Cheers,
Liam.
Practical IDL Programming
http://www.gumley.com/
|
|
|
|
| Re: Turning off math error checking for a code block [message #28966 is a reply to message #28876] |
Fri, 18 January 2002 09:22 |
Martin Downing
Messages: 136
Registered: September 1998
|
Senior Member |
|
|
craig wrote:
> I have found that an operation on an array which contains NANs is
> slowed down considerably. I think it is because each operation causes
> a floating point exception which is handled in the OS. I use WHERE
> most of the time when this comes up. Occassionally I get "floating
> exception" messages, but big whoop.
To illustrate craigs point:
IDL> a = replicate(!values.f_nan,1024,1024)
IDL> b = replicate(2.0,1024,1024)
IDL> help, a,b
A FLOAT = Array[1024, 1024]
B FLOAT = Array[1024, 1024]
IDL> t = systime(1) & for i =0,9 do c = total(a * 2) & print, systime(1) - t
3.1250000
IDL> t = systime(1) & for i =0,9 do c = total(a * 2) & print, systime(1) - t
3.1240001
IDL> a = replicate(2.0,1024,1024)
IDL> help, a
A FLOAT = Array[1024, 1024]
IDL> t = systime(1) & for i =0,9 do c = total(a * 2) & print, systime(1) - t
0.71099997
|
|
|
|
| Re: Turning off math error checking for a code block [message #28968 is a reply to message #28879] |
Fri, 18 January 2002 09:45 |
Vapuser
Messages: 63
Registered: November 1998
|
Member |
|
|
k-bowman@null.com (Kenneth Bowman) writes:
> In article <888zawn5bq.fsf@catspaw.jpl.nasa.gov>, Vapuser <vapuser@catspaw.jpl.nasa.gov> wrote:
>
>> The help says that for !except=1 (the default) it only reports the
>> exception upon arriving at back at an interactive prompt. On my SGI, it
>> still does the `where' and returns the correct answer, it just
>> complains.
>
> The problem is that changing !EXCEPT does not reset the accumulated
> error status, so the errors that occur with !EXCEPT = 0 are still
> present. Hence the need to call CHECK_MATH (twice).
>
Oh, absolutely. I was only addressing the issue of the annoying
error message itself. !except=0 doesn't turn off the exception, just
the message. It's somewhat misnamed, I would have called it
'math_error_verbosity' or some such.
> I want to be able to run my whole code with !EXCEPT = 2 (or 1) but
> not generate errors for the blocks that I *know* will have floating
> point exceptions in them.
>
> I think my solution is probably the simplest way to do what I want to do.
>
I agree completely.
> Thanks everybody, Ken
whd
--
William Daffer: 818-354-0161: William.Daffer@jpl.nasa.gov
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
