k-Wave

1. Neo
   

   Member
   Joined: May '21
   Posts: 5

   Hi,
   I have a quick question about the new kspaceFirstOrder2DG Function which was published with the 1.3 release of kwave. Until now I used kspaceFirstOrder2D with DataCast set to gpuArray-single. To speed up simulations I want to switch to the optimized simulation function and thus tried the kspaceFirstOrder2DG function. I downloaded the compiled binary fdiles and put them into the binary folder of the kwave toolbox.
   I wanted to be sure, that both functions return the same simulation result but actually this is not the case. When I compare the simulation results, I get and RMSE of 0.0379 and a global maximum absolut difference of 0.3797.
   The deviation is small and the reconstructed b-Mode images look the same, but I wonder if this difference between the simulation results is normal?

   The manual says that "not all simulation options are currently supported by the CPU/GPU code. The sensor mask must be given as either a binary matrix or (...). All display parameters are ignored". Are there other restrictions which one should know, when using the optimized code?

   (I get the same difference in data when I use kspaceFirstOrder2D with DataCast set to Single and kspaceFirstOrder2DC)

   Posted 3 years ago #

2. Neo
   

   Member
   Joined: May '21
   Posts: 5

   I have uploaded some Matlab Files so that you can easily check these differences yourself.
   You can download these files here:
   https://ruhr-uni-bochum.sciebo.de/s/LxtleUpNbREHSWi
   The password is: kwave13

   Just download the folder "kWave_OptimizedSimFct_Debugging", add all subfolders to the active Matlab path and run the script "MainScript_SimFctDebugging.m". The Script automatically loads the already simulated data that is stored in the folder "SimulationData" and visualizes the RF data.

   I suspect that the difference is caused by numerical inaccuracies/noise, that is different for the optimized and non-optimized functions. Or is the simulation in the optimized simulation functions implemented differently than in the Matlab version? I would really like to know if these differences are normal or not and what actually causes them.

   The manual says on page 58:
   "To test whether the CPU/GPU codes are working correctly, open the “Diagnostic Ultrasound Simultion, Simulating Ultrasound Beam Patterns Example”, replace kspaceFirstOrder3D
   with kspaceFirstOrder3DC or kspaceFirstOrder3DG, then run the simulation and verify
   the outputs are the same."
   If the difference in simulation results is normal, this section of text causes some confusion

   Posted 3 years ago #

3. Bradley Treeby
   

   Developer
   Joined: Mar '10
   Posts: 1,643

   Hi Neo,

   If you normalise to the maximum value in the grid, the differences between the C++ and MATLAB codes should be on the order of single precision (e.g., 1e-5 to 1e-6). We have a test code that runs a large number of comparisons under different scenarios.

   There are a number of options that aren't supported in the C++/CUDA codes, and unfortunately there is no currently no warning that these are being ignored. I will address this with the next release.

   In the interim, some of the non-supported options are:
   - Anything related to display or movies (there is no display)
   - Cartesian sensor mask
   - medium.alpha_mode
   - medium.alpha_filter
   - medium.alpha_sign
   - sensor.directivity_angle
   - sensor.directivity_size
   - sensor.directivity_pattern

   Hope that helps,

   Brad

   Posted 3 years ago #

RSS feed for this topic

Reply

You must log in to post.