Introduction

I will write the introduction to this project later when I am free.

Prerequirement

• Scons: We use scons to compile the projects.
• Boost library: We use some functions in boost library.
• Math Library: Please compile blas, lapack, scalapack library in your own machine and put them into the scalapack directory.
• Madagascar: We use madagascar to visualize the results.
• Mpich: We use mpich to make enfwi parallel.

Build

Change compiler_set and addition_includes in SConstruct according to your actual compiler.

run scons to build.

	vim Sconstruct	# change the compiler_set and addition_includes
	scons

Run

The recommended way to run the programs is to create separate folders for each cases. The following is a general work flow in normal cases:

  cp -r data/ ~/
  mkdir -p job/01
  cd job
  vim README.md   # write some comments about case 01
  cd 01
  cp ../../model/marm/* .
  vim SConstruct  # update the parameters is necessary
  vim run.sh      # choose the task you want to start
  ./run.sh				

If you want to run enfwi or enfwi-sw, you will need perturbation files as input. You can generate this file by compiling and runing the ./run.sh noise to produce shotsnoise.rsf, and then modify the relative file name in the job/01/Sconstruct.

Show result

In the same directory where you run the case, you can check the result including objective function values, L1 L2 normalized model fit and inverted images.

• you can display the numeric values of absolute objective function values by sfdisfil < absobjs.rsf. The sfdisfil command also works for norobjs.rsf, l1norm.rsf and l2norm.rsf.

• you can display the plots by running sfpen < absobjs.vpl, sfpen vel300.vpl ...

• of course, you can also checkout the log manually.

SWC++ compiler issues

• we can only pass the -O0 optimized flags to the mpicxx -ver 5.421-sw-437 -host when compiling rsf library. Compile it with -O2 will creash the program.

• the follow codes in src/common/parabola-vertex.cpp

  if (std::abs(k2 - k1) < 0.001 * (std::max(std::abs(k2), std::abs(k1))) ||
      (xv == -std::numeric_limits<double>::quiet_NaN())) {
    WARNING() << "THE SET OF POINTS DON'T FIT PARABOLIC WELL, SET y TO y3 ON PURPOSE
    xv = x3;
    yv = y3;
  }

it cannot support std::numeric_limits<double>::quiet_NaN()), so I use std::isnan

• posix_memalign doesn't work, use malloc instead.

Remaining issues (TODO)

• the perturbation file for ENFWI is generated from matlab code in matlab directory. Not all machine has matlab installed. So it is a good idea to convert the function into C++ code or use madagascar to generate the file.

• in ENFWI, the velocity samples in different processes is gathered/scattered via MPI_Send and MPI_Recv because different velocity resides in different part of memory, which is not in a contineous way. If the total numer of samples becomes large (tens of thousands), the communication performance may decrease dramatically. One possible solution is to refactor the Velocity class as a wrapper, making a contains a pointer which points to the actual velocity data. (Solved)

• the memory is not enough on SW when the number of samples is 100. 50 is ok. (Solved)

Contact us

If you have any questions or any suggestions, you can send email to [email protected] or [email protected].