Imshow on Parallel Computing Toolbox

I am converting my code to utilise the Parallel Computing Toolbox and I am getting a peculiar error in an imshow instruction:

Error using imshow (line 209) IMSHOW unable to display image.

Error in GPUDemo (line 9) parfor (HSize = 1:n);

Caused by: Error using imshow (line 209) IMSHOW unable to display image.

line 209 of Imshow is the error message:

if ~images.internal.isFigureAvailable() error(message('images:imshow:needJavaFigure')); end

The error appears only when I use the matlabpool.

Obviously I am using many pools in order to have each pool applying a different value for a parameter (eg. HSize or Sigma in an imfilter) to the image

So I cant quite get the Aprils fool joke here: http://blogs.mathworks.com/steve/?s_cid=global_nav :(

Thank you.

Answer 1

[Walter Roberson]

It is not possible to display graphics from inside a worker. The joke was in saying that it was possible.

Note the reference to the non-existent R2011c, and to running on TRS-80's.

[Dimitris Vayenas]

Indeed I understood the joke. I just didn't find it very funny when I discovered that it is not possible, as the "parallelism" in my case is limited on applying a different set of values for my filters. So I was thinking that one worker will be dealing with, lets say, HSize of 1, while the other with HSize of 2 etc.

So, I was using this:

figure, imshow(Final,'InitialMagnification','fit'), title(['Magnification: ' num2str(Mag) ' Sigma: ',num2str(Sigma), ' HSize: ', num2str(HSize), '  Time:   ', num2str(gpuTime)] );

In order to print in my images the processing time that it takes, depending on various parameters and given the transfers between GPU and CPU as I have to "gather" in the CPU in order to perform _ rgb2gray, imresize, edge,imclearborder, bwconncomp,labelmatrix, label2rgb and regionprops_ that are currently unsupported in the toolkit.

Any advice will be greatly appreciated.

Thank you

[Walter Roberson]

I would have thought you could use all of those routines other than displaying the image, just as you can use them if you launch with -nodisplay

[Dimitris Vayenas]

Correct but I was doing it for my own convenience. To have the images showing up the figures as the loop runs... :P

Do you know why this is not possible? Is it for "memory management" reasons (kind of buffering the results from each thread prior to display etc)?

Also do we have any news about the progress on the making edge, imresize, rgb2gray and imclearborder CUDA\GPU compatible? I noticed that there are CUDA libraries (such as CUVI) that wrap most of this functionality and was surprised to discover that the Parallel Processing Toolbox does not support them, as for example it is common practice to turn a colour image (which you can read on the GPU) into black and white ( which you cannot do in the GPU) prior (or after) applying a filter (which you can do on the GPU), but that you may need to resize or find edges (that you cannot do on the GPU), prior to dilating\eroding it(that you can do on the GPU) etc. you get my point. It doesn't "compute"... :(

[Walter Roberson]

As far as I have determined, the protocol used between the parent and the workers is a message-passing protocol that (for whatever reasons) has not been designed to handle graphics messages. It appears that except for spmd, the protocols might possibly have been designed to send in the data and only listen for the result message -- though perhaps that approach would be too simplistic to account for distributed arrays.

Each worker is a separate process that does not have access to the graphics (which is in the memory space of the parent process... which might not even be on the same computer.)

I do agree that it would seem plausible to be able to handle graphics through a message passing protocol (e.g., X Windows is designed that way.) I do not know why that was not used.

rgb2gray() is, at heart, a fairly simply matrix multiplication; it should not be difficult to write a replacement for it that would execute on GPU, except possibly over the issue of single precision vs double precision.

[Dimitris Vayenas]

Thank you for your response Walter. Regretfully, it does confirm my fears. I guess it is somehow related to the current implementation of asynchronous processes, usually based on the futures abstraction (which is a mistake if you ask me ;)).

I get your point about the rgb2gray, it "crashes" on the input/datatype validation as I am sure you know. The trouble is that if I have such an overhead between transferring the images during their trips from GPU to CPU why not have this functionality built in at Matlab. Same as with imresize and with edge.

But please let me return back to the more practical part of the my original enquiry: Is there a way to have a version of the image even saved (with no figure, imshow), where I can write additional information to it, such as the values of various parameters, or the times of various computations, under the multi-worker scenario?

Answer 2

[Daniel Golden]

I'm running R2014a and the image and imagesc functions do work within parfor loops, although the imshow function does not. If you substitute image for imshow, follow it with this command to get similar results:

axis image off

However, the pixels shown on the screen won't be 1:1 like they are by default with the imshow command.

Answer 3

[Dimitris Vayenas]

Thank you very much Daniel!

I will try it and let you know. Glad to hear that Mathworks are giving so Serious consideration to our recommendations; in such challenging cases in particular.

Here is a benchmark btw: http://www.slideshare.net/vdimitris/hybrid-cpu-gpu-matlab-image-processing-benchmarking

(PS:I do not mind how it appears on the screen as I am using for metavisualisation applications)

Answer 4

[kinjal shah]

what is about the imshow error? what is solution?

[shital shinde]

anybody please tell me how to convert matlab code into many core and multi core. I attached the file for image watermarking. but i want to perform parallelization on it. please help me for code.

[Walter Roberson]

You already asked two questions on this topic.