AOB

School project for optimizing calculations, the example used is a fluid simulator. By DRISSI Reda and HASAN JUSUF Yusnanda - IATIC4 ISTY

PREREQUESITES

You need

• openjdk-x-jdk switch the x with your version
• swing
• gcc with openmp support

MAKING AND BUILDING

• If your libgomp.so is elsewhere you can find it with :

  find / \( -path /mnt -o -path /media -o -path /home -o -path /tmp -o -path /cache \) -prune -o -name "libgomp.so" 2>/dev/null
  

No need to search for it in tmp mnt media home root unless you have a reason to think it's there (please don't)

• You can build the project with make
• then execute it with make run
• At runtime we added the export LD_LIBRARY_PATH to avoid having to do it each time you open a terminal

TRYING OUT DIFFERENT OPTIS

Currently our targets are

• nopti for the basic code with no optimizations
• fast for compilation time optimizations
• pinc for the pre increment optimization
• muldiv for mul/div switching
• inl inlining build_index and setBoundry
• omp openMP integration
• fin all of the above You can execute this command in case of repetitive testing for fastest results

   make clean && make <target> && make run

Using make with no parameters is going to launch the highest level of optimization

CURRENT IMPROVEMENTS

• Using openmp which is self explanatory
• Using inline functions because calling functions is more time consuming than copying them
• Using Ofast mostly for -ftree-vectorize and -frename-registers and -ffast-math
• Using -mavx2 then changing the c constant to 1/c so that we can change division into multiplication
  because division does not benefit from AVX

USELESS IMPROVEMENTS

• using a variable to store the build_index(i,j,grid_size) then just adding subtracting the needed value in order to avoid computing the core value each time had no real performance improvement.
• In setBoundry getting the if statement that is independent of the iterations out of the core function had no real added value either.
• Dividing the long calculations to avoid mixing add and mul hassles.
• using pre increment instead of post increment (in case gcc doesn't handle that which it most probably does)
• unrolling loops even manually is useless because iterations are dependent

IMPROVEMENT IDEAS

• Using intel intrinsic fonctions in order to use sse registers (better result than -mavx2 in theory)
• Verifying whether build_index and setBoundry functions have room for improvements.