k-Wave

1. KHO
   

   Member
   Joined: Nov '12
   Posts: 11

   Hi,

   I tried using the k-Wave Toolbox with GPUmat. Part of it already runs without any problems but now I came across a missing GPUmat function that seems to break the code. Matlab complains about knowing no function max for data type GPUsingle. When I had a closer look at GPUmat it seemed that there really were a few essential functions missing, i.e. max, min, isnan, isinf, ...

   So I tried to find a module for GPUmat that would include these, but couldn't really come up with one. I found the DeconvNetToolbox2, which includes GPUmat modules (which also include the missing functions), but even though I could get it to compile the module, it causes Matlab to crash with a Segmentation Fault, when I try using it.

   How did you solve this problem? One could probably convert to double, do the max(), and then convert back, but that is a rather bad and inefficient solution.

   This was using Matlab R2011a on Win64 (but I don't think that is the actual problem) and the code was linked to CUDA Toolkit Version 4.2.

   Posted 12 years ago #

2. Bradley Treeby
   

   Developer
   Joined: Mar '10
   Posts: 1,643

   Hi KHO,

   Thanks for your feedback and sharing your efforts to get it working. Unfortunately we don't have a simple solution to the missing functions. From my testing notes, these are the current restrictions with using k-Wave and GPUmat:

   - the option 'PlotLayout' must be set to false (no max/min functions)
   - the 'p_max' and 'u_max' options for sensor.record don't work (no max/min functions)
   - if using a Cartesian sensor mask with nearest neighbour interpolation, 'DataRecast' must be set to true (re-ordering enlarges the output matrix on the fly which is not supported)

   Perhaps you have also noticed others?

   The GPUmat toolbox is no longer being actively developed, but the codes have been released open-source (GPL). If you're familiar with CUDA, one option would be to try and add the missing functions to the toolbox.

   Brad.

   Posted 12 years ago #

3. KHO
   

   Member
   Joined: Nov '12
   Posts: 11

   Hi Brad,

   actually using the DeconvNetToolbox2 Modules, I was at least able to get it to run on a linux machine, yesterday. Within this box, max, min and some other useful functions are defined that can be integrated into GPUmat. I am not familiar enough with CUDA to figure out why this did not work on windows because in my opinion this should make no difference for the program code.

   The toolbox (i didn't use anything except for the GPUmat module "other"):
   http://www.matthewzeiler.com/software/DeconvNetToolbox2/

   This pretty much only has to be copied to the modules directory of GPUmat on a 64 bit machine though it's possible that I had to modify a few include paths for the compilation. They include a matlab make script to use (I believe I had to modify that a little for the CUDA Paths).

   Posted 12 years ago #

4. Bradley Treeby
   

   Developer
   Joined: Mar '10
   Posts: 1,643

   Hi KHO,

   Thanks for the update, and for managing to find the missing functions! I'm sure others will find this useful. It seems the DeconvNetToolbox2 also includes precompiled mex files for 64-bit linux in the "/modules/mdz/other/@GPUtype folder", although I haven't tried using these yet.

   Brad.

   Posted 12 years ago #

RSS feed for this topic

Reply

You must log in to post.