floating point exception from visit 1.12.0

[markcmiller86]

Date: Fri, 14 Aug 2009 02:47:01 -0700
From: andrew at fluidgravity.co.uk (Andrew Wood)
This post was imported from the [email protected] email archive

---

Hi,

I've just been testing visit 1.12.0 before upgrading from version 1.11.2.

One of my tests involves drawing a 3D material plot and saving it to a
png file. Visit displayed the plot correctly in the window, but failed
to save it to the png file, reporting a floating point exception in the
bash console. The test works fine with version 1.11.2.

A 2D plot did get saved to the png file. However, the png image is
corrupted with vertical lines covering the image. I'd mentioned this
issue before in a previous email, and also that I'd managed to get round
it by using a copy of the Mesa library that had been compiled on a
different machine. I've not tried that solution with the new version yet.

I rebuilt visit so I could run it through a debugger. Here's what I found:

1. The floating point exception occurs in the function
'LabelExponent(double min, double max)'.
The arguments min and max are both 0.0.
So the line 'double pow10 = log10(range)' sets pow10 = -Inf.

2. LabelExponent is called from the method VisWinAxes::AdjustValues,
which is called from the method VisWinAxes::UpdateView

3. The offending arguments to LabelExponent are initialised to 0.0 in
UpdateView and set in the method VisWinAxes::GetRange.

4. In GetRange, there is a statement 'switch (vw->GetWindowMode())'

5. vw->GetWindowMode() returns WINMODE_3D, which is not in the list of
cases in the switch statement, so the arguments that are eventually
passed to LabelExponent remain unchanged.

I diffed the source file between versions 1.12.0 and 1.11.2 to see if
there was a simple and obvious error that I could quickly fix myself,
but didn't see anything.

I hope that useful.

Regards,
Andy.

---

Dr Andrew Wood
Tel: +44 (0)1334 460805
Fax: +44 (0)1334 460813

Fluid Gravity Engineering Ltd is registered in the UK with registration
number 1674369. The registered address is Fluid Gravity Engineering Ltd,
Unit 1, The Old Coach House, 1 West Street, Emsworth, Hampshire, PO10 7DX.

---

[markcmiller86]

Date: Fri, 14 Aug 2009 06:28:20 -0700
From: miller86 at llnl.gov (Mark Miller)
This post was imported from the [email protected] email archive

---

Super bug report! Sorry you ran into this, again!

I tried to duplicate here and was unable to.

Can you provide information on platform/os? Also, is it practical to
provide the data files you ran these tests on? Or, if they are one of
our files, the names?

Thanks for any add'l info.

Mark

On Fri, 2009-08-14 at 02:47, Andrew Wood wrote:
> Hi,
>
> I've just been testing visit 1.12.0 before upgrading from version 1.11.2.
>
> One of my tests involves drawing a 3D material plot and saving it to a
> png file. Visit displayed the plot correctly in the window, but failed
> to save it to the png file, reporting a floating point exception in the
> bash console. The test works fine with version 1.11.2.
>
> A 2D plot did get saved to the png file. However, the png image is
> corrupted with vertical lines covering the image. I'd mentioned this
> issue before in a previous email, and also that I'd managed to get round
> it by using a copy of the Mesa library that had been compiled on a
> different machine. I've not tried that solution with the new version yet.
>
> I rebuilt visit so I could run it through a debugger. Here's what I found:
>
> 1. The floating point exception occurs in the function
> 'LabelExponent(double min, double max)'.
> The arguments min and max are both 0.0.
> So the line 'double pow10 = log10(range)' sets pow10 = -Inf.
>
> 2. LabelExponent is called from the method VisWinAxes::AdjustValues,
> which is called from the method VisWinAxes::UpdateView
>
> 3. The offending arguments to LabelExponent are initialised to 0.0 in
> UpdateView and set in the method VisWinAxes::GetRange.
>
> 4. In GetRange, there is a statement 'switch (vw->GetWindowMode())'
>
> 5. vw->GetWindowMode() returns WINMODE_3D, which is not in the list of
> cases in the switch statement, so the arguments that are eventually
> passed to LabelExponent remain unchanged.
>
> I diffed the source file between versions 1.12.0 and 1.11.2 to see if
> there was a simple and obvious error that I could quickly fix myself,
> but didn't see anything.
>
> I hope that useful.
>
> Regards,
> Andy.

---

Mark C. Miller, Lawrence Livermore National Laboratory
email: mailto:miller86 at llnl.gov
(M/T/W) XXXXXXXXXXXXXX (!!LLNL BUSINESS ONLY!!)
(Th/F)  XXXXXXXXXXXXXX (!!LLNL BUSINESS ONLY!!)

---

[markcmiller86]

Date: Fri, 14 Aug 2009 06:42:29 -0700
From: jsmeredith at ornl.gov (Meredith, Jeremy S.)
This post was imported from the [email protected] email archive

---

What platform is this on?  I'm surprised this is causing such bad behavior -- maybe there's a way to
keep this floating point behavior from causing an exception?

Nevertheless, it sounds like you're building from source?  It's an easy fix, I think -- in
src/avt/VisWindows/Colleagues/VisWinAxes.C, line 980, I think you can simply replace this line:
double pow10 = log10(range);

With this:
double pow10 = -5.0;
if (range > 0)
pow10 = log10(range);

I made a similar change in the past to a different file, but it was to address a legitimate case
where the range truly was zero.  I'm a little surprised this crops up in the case you mentioned.
Still, the change I suggest is the right thing to do, so I'll make sure it gets committed for our
next release, but let me know if that doesn't fix the problem you're seeing so we can investigate
further.

---

Jeremy Meredith
Oak Ridge National Laboratory

> -----Original Message-----
> From: Andrew Wood [mailto:andrew at fluidgravity.co.uk]
> Sent: Friday, August 14, 2009 5:47 AM
> To: visit-users at email.ornl.gov
> Subject: [visit-users] Floating point exception from visit 1.12.0
>
> Hi,
>
> I've just been testing visit 1.12.0 before upgrading from version
> 1.11.2.
>
> One of my tests involves drawing a 3D material plot and saving it to a
> png file. Visit displayed the plot correctly in the window, but failed
> to save it to the png file, reporting a floating point exception in the
> bash console. The test works fine with version 1.11.2.
>
> A 2D plot did get saved to the png file. However, the png image is
> corrupted with vertical lines covering the image. I'd mentioned this
> issue before in a previous email, and also that I'd managed to get
> round
> it by using a copy of the Mesa library that had been compiled on a
> different machine. I've not tried that solution with the new version
> yet.
>
> I rebuilt visit so I could run it through a debugger. Here's what I
> found:
>
> 1. The floating point exception occurs in the function
> 'LabelExponent(double min, double max)'.
> The arguments min and max are both 0.0.
> So the line 'double pow10 = log10(range)' sets pow10 = -Inf.
>
> 2. LabelExponent is called from the method VisWinAxes::AdjustValues,
> which is called from the method VisWinAxes::UpdateView
>
> 3. The offending arguments to LabelExponent are initialised to 0.0 in
> UpdateView and set in the method VisWinAxes::GetRange.
>
> 4. In GetRange, there is a statement 'switch (vw->GetWindowMode())'
>
> 5. vw->GetWindowMode() returns WINMODE_3D, which is not in the list of
> cases in the switch statement, so the arguments that are eventually
> passed to LabelExponent remain unchanged.
>
> I diffed the source file between versions 1.12.0 and 1.11.2 to see if
> there was a simple and obvious error that I could quickly fix myself,
> but didn't see anything.
>
> I hope that useful.
>
> Regards,
> Andy.
>
> --
> Dr Andrew Wood
> Tel: +44 (0)1334 460805
> Fax: +44 (0)1334 460813
>
> Fluid Gravity Engineering Ltd is registered in the UK with registration
> number 1674369. The registered address is Fluid Gravity Engineering
> Ltd,
> Unit 1, The Old Coach House, 1 West Street, Emsworth, Hampshire, PO10
> 7DX.
> --
> List subscription information:
> https://email.ornl.gov/mailman/listinfo/visit-users

---

[markcmiller86]

Date: Mon, 17 Aug 2009 04:37:20 -0700
From: andrew at fluidgravity.co.uk (Andrew Wood)
This post was imported from the [email protected] email archive

---

Thanks for the response.

The os is Linux, openSUSE 11.0 (x86_64).

Unfortunately, I don't think it would be practical to provide the data
files - the readers we use include a lot of our own proprietary code for
post-processing.

I tried another test problem that was produced by a different hydrocode
and reader. This test wrote out the png file (albeit producing a
corrupted image). Stepping through with the debugger, I found that the
function LabelExponent is still called with min = max = 0.0, and pow10
is set to -Inf, in exactly the same way as in the case that failed.
However, the function still returns without throwing a floating point
exception. In the previous case the floating point exception would
actually occur in the line 'return (int)ipow10;'.

The fix Jeremy suggested in his reply appears to resolve the problem.

Andy.

Mark Miller wrote:
> Super bug report! Sorry you ran into this, again!
>
> I tried to duplicate here and was unable to.
>
> Can you provide information on platform/os? Also, is it practical to
> provide the data files you ran these tests on? Or, if they are one of
> our files, the names?
>
> Thanks for any add'l info.
>
> Mark
>
> On Fri, 2009-08-14 at 02:47, Andrew Wood wrote:
>> Hi,
>>
>> I've just been testing visit 1.12.0 before upgrading from version 1.11.2.
>>
>> One of my tests involves drawing a 3D material plot and saving it to a
>> png file. Visit displayed the plot correctly in the window, but failed
>> to save it to the png file, reporting a floating point exception in the
>> bash console. The test works fine with version 1.11.2.
>>
>> A 2D plot did get saved to the png file. However, the png image is
>> corrupted with vertical lines covering the image. I'd mentioned this
>> issue before in a previous email, and also that I'd managed to get round
>> it by using a copy of the Mesa library that had been compiled on a
>> different machine. I've not tried that solution with the new version yet.
>>
>> I rebuilt visit so I could run it through a debugger. Here's what I found:
>>
>> 1. The floating point exception occurs in the function
>> 'LabelExponent(double min, double max)'.
>> The arguments min and max are both 0.0.
>> So the line 'double pow10 = log10(range)' sets pow10 = -Inf.
>>
>> 2. LabelExponent is called from the method VisWinAxes::AdjustValues,
>> which is called from the method VisWinAxes::UpdateView
>>
>> 3. The offending arguments to LabelExponent are initialised to 0.0 in
>> UpdateView and set in the method VisWinAxes::GetRange.
>>
>> 4. In GetRange, there is a statement 'switch (vw->GetWindowMode())'
>>
>> 5. vw->GetWindowMode() returns WINMODE_3D, which is not in the list of
>> cases in the switch statement, so the arguments that are eventually
>> passed to LabelExponent remain unchanged.
>>
>> I diffed the source file between versions 1.12.0 and 1.11.2 to see if
>> there was a simple and obvious error that I could quickly fix myself,
>> but didn't see anything.
>>
>> I hope that useful.
>>
>> Regards,
>> Andy.

---

Dr Andrew Wood
Tel: +44 (0)1334 460805
Fax: +44 (0)1334 460813

Fluid Gravity Engineering Ltd is registered in the UK with registration
number 1674369. The registered address is Fluid Gravity Engineering Ltd,
Unit 1, The Old Coach House, 1 West Street, Emsworth, Hampshire, PO10 7DX.

---