Datastructure

Exercise0 : Particle datastructure

In hands-on/memory/datastructure Modify Particle.h Modify the class GoodParticle with the following members:

• 1 const std::string to hold the particle’s name;
• 3 doubles for the x, y, z velocities
• 3 bools to mark if there has been a collision along the x, y z directions
• 1 float for the mass
• 1 float for the energy
• 3 doubles for the px, py, pz coordinates
• 1 const int for the particle’s id

What is the best order for your members?

compile and run good_particle.cc to check the size of your GoodParticle class

Exercise1: Access Pattern

Now let’s use our GoodParticle (nicely organized struct/class) and let’s do some computations

• Write a function to initialize a collectin of N GoodParticles
  • The function should also initialize all the memebers of each object
• Write another function that takes as input the collection of GoodParticles and x_max value
  • Iterate over all the elements and for each element:
    • Generate a random number representing a time interval t
    • Update the position x-->x = x + px/m * t
    • If the updated position x is less than 0 or greater than x_max
      • set hit_x to true
    • Else:
      • Set hit_x to false and change the sign of px
• How are you accessing the members in the datastructure?

Exercise2: AoS vs SoA

In hands-on/memory/datastructure Modify Particle.h Modify the class ParticleSoA with the following members:

• Implement the functions described in Exercise1 for the SoA object
• Work with fixed size objects
• Measure the time for AoS implementation and SoA implementation and compare them
• What happens if you compile with some optimization? `-O1’, ‘-O2’, ‘-O3’?
• How does SoA interface change with respect to AoS?

Bonus: Matrix

In hands-on/memory/datastructure

Compile matrix.cc and check the timing of doing matrix multiplication in “naive” way against the “transposing” method. Then:

• Comment out the execution of the “transpose” method and compile again with -O3 and -g.
  • Run perf stat -B -e L1-dcache-loads,L1-dcache-load-misses,LLC-loads,LLC-load-misses,cycles,instructions ./a.out and have a look at the cache missing and CPU cycles
• Do the same commenting out the execution of the naive method and run perf again