I was stumped by a segmentation fault in MATLAB.
It seems like it was caused by an anonymous function that was loaded from a mat file.
The original anonymous function handle was:
#(x)scaledNlfun(x,#logexp1,1e3)
But when it is loaded, it becomes:
#sf%1#(x)scaledNlfun(x,#logexp1,1e3)
It seems to be okay, when I call it in command line, but it creates a segmentation fault (or Segmentation violation) within a function. Not the function call itself, but a few lines after that. In debugging mode, if I step through the statement, it is fine as well.
The stack trace shows bunch of
[ 0] 0x00002b20b97baba4 /usr/local/MATLAB/R2013a/bin/glnxa64/libmwm_interpreter.so+04127652
and it happens on both MATLAB 2012a and 2013a on a Linux 2.6.18-371.3.1.el5 SMP.
This function handle was saved within a parfor loop using '-v7.3' option because the struct that contains the handle was too big. If I replace the anonymous function after loading the mat file, everything works fine, so I'm thinking the matlab load function has a bug.
Unfortunately, I cannot create a minimal example to reproduce the error. I tried saving anonymous function handles within parfor with '-v7.3', but without the other complex data structures, it seems to work fine. But I have 80 mat files that would reliably crash matlab (many of them more than 1GB).
In any case, does anybody know what that "#sf%" mean? (it's not the stateflow toolbox)
The core of the problem seems to be that you have #sf%1# where you would expect # just looking at this, I can think of a few possibilities:
Somehow sf%1# was inserted after the original #
Somehow #sf%1# was substituted in place of the original #
Somehow #sf%1 was attached before the original #
I would actually bet on the third one, but here are the most logical scenarios I can think of that could cause this problem:
Perhaps there was an invisible char?
Perhaps some kind of strange character conversion?
Probably a situation where two things are stored in a variable instead of one. Perhaps something like #s or #sf and some separating characters.
All in all, this does not explain why it would go well if you run the entire program in the console, but perhaps you just ran part of it. In that case these could be some things to look out for.
Related
I'm extremely new to both working on Linux and Fortran, so apologies if this is a basic question.
I am trying to firstly use fwrite to save a 60x150 array that I've produced in MATLAB as a binary file, which I am then attempting to load and read in Fortran as a 60x150 array again.
In Matlab, I have used the following code to save the array. In this case the name of the array in the workspace is VP, and I'm saving it to a file also called 'VP':
>> fileID = fopen('VP','w');
>> fwrite(fileID,VP,'real*8');
>> fclose(fileID)
Next, I am copying the file over from Windows to a linux ssh server (I'm not sure if this is relevant, but thought it's worth including anything that might help).
Now, in my Fortran code, I've got:
REAL(KIND=kind(1.0D0)), DIMENSION(60,150) :: VP
...
open(unit, file="LOCATION/VP", access = "stream", form = "unformatted", iostat = stat)
if(stat /= 0) labort("Failed to open input file")
print *, wl
DO inx2=1,60
DO inx=1,150
print *,inx
READ(unit,*) VP(inx2,inx)
ENDDO
ENDDO
print *,VP(1:10,1)
Now, when I compile this there are no errors. However, when I run it, it gets to exactly the first "READ(unit,*) VP(inx2,inx)" before failing (you can tell from the print just before it).
I get the error:
forrtl: severe (257): formatted I/O to unit open for unformatted transfers, unit 111, file LOCATION/VP
Obviously I would like my actual result to be the function running and ending up with the same values in the array.
Now I've seen the question before, specifically for this error message, but that was answered by including access="stream" which I have already. Basically I am not sure at what point I am getting something wrong in this process, any help would be appreciated.
Note some things I have tried is changing the precision in fwrite, and swapping the inx2 and inx values around (but it fails on the first one so I don't believe that's the error).
Again this might be just a fundamental issue with my understanding of Fortran because I've been thrown in the deep end a bit with a project I'm working on (most of the code I'm running was produced by someone else, I'm just trying to edit a small part of it).
Edit:
Okay, thank you so much francescalus! He found the solution was to edit the line to READ(unit) VP(inx2,inx) in order to get it to run. However, the values I get by running the next line of code:
print *,VP(1:10,1)
Only the first value matches the first value in my original matlab array. Displaying VP(1:5,1:5) of the matlab array there aren't any other matching values. I might be able to figure this out on my own but as I'm already here I might as well ask as I haven't fully completed the original question (although got over a big hurdle!).
Edit 2:
Okay the next bit I've managed to figure out for myself. If anyone is searching though it was simply a case of swapping the DO loops. ie
DO inx=1,150
DO inx2=1,60
print *,inx
READ(unit,*) VP(inx2,inx)
ENDDO
ENDDO
Thanks for the help.
My problem is the following:
I have very many (~1000) mutually calling Matlab scripts, which are very poorly written, regularly damage each other's environments and generally became unmanageable.
One of the reasons I even got this problem is that I need to write a testsuite covering a big part of them. Luckily, for most of them the main criterion of 'correctness' is 'they don't crash'.
Just running them one by one in a loop is generally not an option, because they regularly call clear classes, close all, clc, shadow built-in functions and operators, et cetera.
So my original aim was to find a way to run a matlab script in sort of an 'isolated environment', but I didn't find a good way to do it. (Suggestions welcome, but it is not the main question.)
Since I will need to convert them all to functions anyway, I am looking for some way to do it auto-magically, or at least semi-automatically.
What I can mean semi-automatically:
Just add a line function varargout = $filename( varargin ) as the first line of the file, and end as the last one. This will at least make them runnable as functions with feval and all such functions and (more importantly) prevent them from damaging the test-runner.
Do point 1 and scan the file for referencing undeclared variables and add them as function arguments. This should be also doable, since the names of the variables are known. This will not help identifying output variables, but will still be a lot of help. For example, we could pack the whole workspace into one big output structure.
Do a runtime version of point 2. This way the 'magical converter' can actually track execution environments (workspaces) and identify which variables are implicitly used as 'input arguments' of a script, and which would be later used 'output arguments'. This option looks EXPHARD, but for a small number of calls should be not too bad in practice.
Point 1 I can implement myself using sed, as I also can get rid of all clear classes and clc, but the options 2 and 3 seem much harder. Is there anything at least remotely resembling options 2 or 3?
I recently started using Live editor in MATLAB and I inserted a function inside it. But, apparently, I cannot execute that particular section of code where I type function. Even the section break disappears.
Is it that using function is not suitable for live editor?
Apparently the MATLAB parser did not join the 20th century until partway through 2016, and could not interpret function definitions in scripts (live or otherwise) until R2016b. In the web-based docs, there is a notice at Add Functions to Scripts, but it took me a while to find this out because the builtin docs in R2016a or earlier do not explicitly contain this information. It is implied by the tutorials that tell you to create a new file for each function (which to me, a python programmer, sounds more like strange advice than a restriction).
Trying to define a function in a live script gives confusing errors. For example, if you create a cell with this content:
function y = myfunc(x)
y = 2*x;
end
It will underline the keyword function with a popup error that reads:
Parse error at FUNCTION: usage might be invalid MATLAB syntax.
Might be? Whom shall I ask? Upon running the cell, it prints an error after the first line:
All functions in a script must be closed with an end.
I eventually made this discovery myself thanks to a helpful message if the first thing you happen to do in a new empty live script is to start typing function on the first line; as soon as you hit the spacebar a message pops up at the top saying:
Functions and classes are not supported in the Live Editor. To continue, save the file as a plain text code file (.m).
It should work as when you add a function inside a script. For example, like this:
What function are you exactly trying to code?
In MATLAB, how do you tell where in the code a variable is getting output?
I have about 10K lines of MATLAB code with about 4 people working on it. Somewhere, someone has dumped a variable in a MATLAB script in the typical way:
foo
Unfortunately, I do not know what variable is getting output. And the output is cluttering out other more important outputs.
Any ideas?
p.s. Anyone ever try overwriting Standard.out? Since MATLAB and Java integration is so tight, would that work? A trick I've used in Java when faced with this problem is to replace Standard.out with my own version.
Ooh, I hate this too. I wish Matlab had a "dbstop if display" to stop on exactly this.
The mlint traversal from weiyin is a good idea. Mlint can't see dynamic code, though, such as arguments to eval() or string-valued figure handle callbacks. I've run in to output like this in callbacks like this, where update_table() returns something in some conditions.
uicontrol('Style','pushbutton', 'Callback','update_table')
You can "duck-punch" a method in to built-in types to give you a hook for dbstop. In a directory on your Matlab path, create a new directory named "#double", and make a #double/display.m file like this.
function display(varargin)
builtin('display', varargin{:});
Then you can do
dbstop in double/display at 2
and run your code. Now you'll be dropped in to the debugger whenever display is implicitly called by the omitted semicolon, including from dynamic code. Doing it for #double seems to cover char and cells as well. If it's a different type being displayed, you may have to experiment.
You could probably override the built-in disp() the same way. I think this would be analagous to a custom replacement for Java's System.out stream.
Needless to say, adding methods to built-in types is nonstandard, unsupported, very error-prone, and something to be very wary of outside a debugging session.
This is a typical pattern that mLint will help you find:
So, look on the right hand side of the editor for the orange lines. This will help you find not only this optimization, but many, many more. Notice also that your variable name is highlighted.
If you have a line such as:
foo = 2
and there is no ";" on the end, then the output will be dumped to the screen with the variable name appearing first:
foo =
2
In this case, you should search the file for the string "foo =" and find the line missing a ";".
If you are seeing output with no variable name appearing, then the output is probably being dumped to the screen using either the DISP or FPRINTF function. Searching the file for "disp" or "fprintf" should help you find where the data is being displayed.
If you are seeing output with the variable name "ans" appearing, this is a case when a computation is being done, not being put in a variable, and is missing a ';' at the end of the line, such as:
size(foo)
In general, this is a bad practice for displaying what's going on in the code, since (as you have found out) it can be hard to find where these have been placed in a large piece of code. In this case, the easiest way to find the offending line is to use MLINT, as other answers have suggested.
I like the idea of "dbstop if display", however this is not a dbstop option that i know of.
If all else fails, there is still hope. Mlint is a good idea, but if there are many thousands of lines and many functions, then you may never find the offender. Worse, if this code has been sloppily written, there will be zillions of mlint flags that appear. How will you narrow it down?
A solution is to display your way there. I would overload the display function. Only temporarily, but this will work. If the output is being dumped to the command line as
ans =
stuff
or as
foo =
stuff
Then it has been written out with display. If it is coming out as just
stuff
then disp is the culprit. Why does it matter? Overload the offender. Create a new directory in some directory that is on top of your MATLAB search path, called #double (assuming that the output is a double variable. If it is character, then you will need an #char directory.) Do NOT put the #double directory itself on the MATLAB search path, just put it in some directory that is on your path.
Inside this directory, put a new m-file called disp.m or display.m, depending upon your determination of what has done the command line output. The contents of the m-file will be a call to the function builtin, which will allow you to then call the builtin version of disp or display on the input.
Now, set a debugging point inside the new function. Every time output is generated to the screen, this function will be called. If there are multiple events, you may need to use the debugger to allow processing to proceed until the offender has been trapped. Eventually, this process will trap the offensive line. Remember, you are in the debugger! Use the debugger to determine which function called disp, and where. You can step out of disp or display, or just look at the contents of dbstack to see what has happened.
When all is done and the problem repaired, delete this extra directory, and the disp/display function you put in it.
You could run mlint as a function and interpret the results.
>> I = mlint('filename','-struct');
>> isErrorMessage = arrayfun(#(S)strcmp(S.message,...
'Terminate statement with semicolon to suppress output (in functions).'),I);
>>I(isErrorMessage ).line
This will only find missing semicolons in that single file. So this would have to be run on a list of files (functions) that are called from some main function.
If you wanted to find calls to disp() or fprintf() you would need to read in the text of the file and use regular expresions to find the calls.
Note: If you are using a script instead of a function you will need to change the above message to read: 'Terminate statement with semicolon to suppress output (in scripts).'
Andrew Janke's overloading is a very useful tip
the only other thing is instead of using dbstop I find the following works better, for the simple reason that putting a stop in display.m will cause execution to pause, every time display.m is called, even if nothing is written.
This way, the stop will only be triggered when display is called to write a non null string, and you won't have to step through a potentially very large number of useless display calls
function display(varargin)
builtin('display', varargin{:});
if isempty(varargin{1})==0
keyboard
end
A foolproof way of locating such things is to iteratively step through the code in the debugger observing the output. This would proceed as follows:
Add a break point at the first line of the highest level script/function which produces the undesired output. Run the function/script.
step over the lines (not stepping in) until you see the undesired output.
When you find the line/function which produces the output, either fix it, if it's in this file, or open the subfunction/script which is producing the output. Remove the break point from the higher level function, and put a break point in the first line of the lower-level function. Repeat from step 1 until the line producing the output is located.
Although a pain, you will find the line relatively quickly this way unless you have huge functions/scripts, which is bad practice anyway. If the scripts are like this you could use a sort of partitioning approach to locate the line in the function in a similar manner. This would involve putting a break point at the start, then one half way though and noting which half of the function produces the output, then halving again and so on until the line is located.
I had this problem with much smaller code and it's a bugger, so even though the OP found their solution, I'll post a small cheat I learned.
1) In the Matlab command prompt, turn on 'more'.
more on
2) Resize the prompt-y/terminal-y part of the window to a mere line of text in height.
3) Run the code. It will stop wherever it needed to print, as there isn't the space to print it ( more is blocking on a [space] or [down] press ).
4) Press [ctrl]-[C] to kill your program at the spot where it couldn't print.
5) Return your prompt-y area to normal size. Starting at the top of trace, click on the clickable bits in the red text. These are your potential culprits. (Of course, you may need to have pressed [down], etc, to pass parts where the code was actually intended to print things.)
You'll need to traverse all your m-files (probably using a recursive function, or unix('find -type f -iname *.m') ). Call mlint on each filename:
r = mlint(filename);
r will be a (possibly empty) structure with a message field. Look for the message that starts with "Terminate statement with semicolon to suppress output".
Suppose I'm using the MATLAB IDE and happen to have some very large objects in my workspace (e.g. arrays of 500k+ elements). Now, suppose that I stupidly and accidentally double click on one of these very large variables, which triggers a load to the array editor. Unfortunately, with arrays this big, MATLAB just hangs.
I've tried CTRL+C, CTRL+BREAK, CTRL+D, but none seem able to interrupt the behavior of the IDE. I know I can force matlab to quit, but reading all of those variables into the workspace in the first place takes a lot of time, and I may have unsaved changes in an editor window, etc.
The variable editor is launched using the command openvar. To solve your problem you can take advantage of a Matlab quirk that causes functions to be masked by variables with the same name. For example if you create a variable named plot the plot() function stops working.
The solution, although hackish, is to simply create an empty variable named openvar. Then anytime attempt to open the variable editor will fail because the function openvar is being hidden by the variable.
If you want to use the variable editor again simple call clear openvar to delete the variable and the function will be unmasked.
I found a way, but it's not the best, it requires a change of path and back once to get a handle to the original openvar
function openvar(name,array)
persistent org_openvar
if isempty(org_openvar)
curdir=pwd;
cd(fullfile(matlabroot,'toolbox/matlab/codetools'));
org_openvar = #openvar;
cd(curdir);
end
if numel(array)>1e5
if strcmp(questdlg(sprintf('Opening ''%s'' which has %d elements.\n\nAre you sure? This is gonna take a while!',name,numel(array)), ...
'Variable editor','Yes','Cancel','Cancel') , 'Yes')
org_openvar(name,array)
end
else
org_openvar(name,array)
end
end
getting that handle is the biggest problem, calling it is just fine. If openvar would be built in, you could use the function builtin:
builtin('openvar',name,array)
but this is unfortunately not the case :(
str2func in combination with the complete path also doesn't work, at least I don't get it to work...