k-Wave

1. vich
   

   Member
   Joined: Jul '19
   Posts: 19

   Hi there,

   Just passing by to report what I think may be a bug: when I compile kspaceFirstOrder-CUDA (__KWAVE_GIT_HASH__="468dc31c2842a7df5f2a07c3a13c16c9b0b2b770") on Linux with gcc version 13.2.1, the compilation succeeds and the binary runs successfully. But in fact it only appears to run successfully. In some simulations of mine, after running kspaceFirstOrder-CUDA, MATLAB's h5read() reads in the /p_final and /p fields of the output HDF5 file as all-NaN arrays (of the expected size), the three /u[xyz]_non_staggered as all-zeros arrays, and /p_min as an array with all entries set to 3.4028e+38 (which equals realmax('single')), for example. The compiled binary produces no unusual messages at runtime.

   However, when I revert to gcc version 12.3.1 and compile kspaceFirstOrder-CUDA anew, keeping all else unchanged, the resulting binary works properly; the output HDF5 file contains meaningful information.

   I tried fiddling with the NVIDIA drivers and the CUDA Toolkit, but I don't think they are to blame. The main reason being (as I found out later) that the kspaceFirstOrder-OMP (__KWAVE_GIT_HASH__="0ba023063e3f29685e1e346f56883378d961f9f1") binary also suffers from the same symptoms (except /p_min becomes an all-NaN array) when compiled with the more recent gcc 13.2.1; but not with the older gcc 12.3.1.

   I am not sure if that is a problem with my particular system or if the bug can be reproduced elsewhere. Maybe this info can be useful to someone. In any case, though, I think maybe the external binaries could do with some sort of sanity check on the output; something inexpensive, just to catch egregious conditions?

   Thanks for your great software!

   Posted 11 months ago #

2. aschrempf
   

   Member
   Joined: Jul '23
   Posts: 1

   Hi,
   I can confirm that. Tried to build on macOS (according to https://github.com/lennartverhagen/k-Wave-CPUGPU-macOS) -- so completely different platform -- with gcc 14 and got always NaN after simulation. Also no error indication during simulation. Simulation on Matlab-Version of kspaceFirstOrder ist pretty stable.
   Thanks for the hint, "rich". After compiling with gcc 12 everything works as expected!

   Thanks also for the great k-wave SW!

   Kind regards
   Andreas

   Posted 8 months ago #

3. Jiri Jaros
   

   Developer
   Joined: Feb '12
   Posts: 118

   Hi, we now about and working on releasing the fix. The GNU compiler changed the way it treats infinity.

   The following lines now cause the simulation to end with NaNs.

   if (absorbNabla1[i] == std::numeric_limits<float>::infinity()) absorbNabla1[i] = 0.0f;
   if (absorbNabla2[i] == std::numeric_limits<float>::infinity()) absorbNabla2[i] = 0.0f;

   a simple, however, not the best fix is:

   // Comparison with infinity is not supported with --ffast-math enabled
   if (absorbNabla1[i] >= std::numeric_limits<float>::max()) absorbNabla1[i] = 0.0f;
   if (absorbNabla2[i] >= std::numeric_limits<float>::max()) absorbNabla2[i] = 0.0f;

   Posted 4 months ago #

RSS feed for this topic

Reply

You must log in to post.