Overall build
1. Type make
2. Produces 3 executables: fjlt, kdtree, main
3. fjlt - runs a FJLT transform on randomly
generated nxd to nxk
4. kdtree - builds a kd-tree from nxd dataset
5. main - Generates random nxd data, and transforms it to nxk
using FJLT and then builds the tree
Parallel sort:
To compile on Hogwarts:
1. run this line in the shell
source /opt/intel/Compiler/11.0/084/bin/iccvars.sh intel64
source /opt/intel/Compiler/11.1/064/bin/ifortvars.sh intel64
2. use ICC to compile it:
icc -openmp proj.cpp myRNG.cpp -o parallelSort.exe
3. executable takes 3 inputs:
i. log(# elements in array)
ii. print input array? (1 = Yes)
iii. print output array? (1 = Yes)
ex. OMP_NUM_THREADS=4 ./parallelSort.exe 21 0 0
This will generate a random array of floats with 2^21 elements
and sort it in parallel using 4 threads. The input nor the
output will be printed to the screen.
Fast Johnson-Lindenstrauss Transform:
FJLT implementation complete
To compile:
1. g++ main.cc fjlt.cc io.cc -o fjlt -fopenmp -lfftw3 -lblas -lm
2. As of now executable doesnt take any inputs.
Just creates some data randomly and does the projections on those
Building kd-tree
Compiling:
icc -openmp -openmp-task intel proj.cpp
Executing:
OMP_NUM_THREADS=<N> ./a.out n d
where, N = number of openmp threads
n = log of dataset size
d = log of dimension size
Example:
OMP_NUM_THREADS=4 ./a.out 8 7
The above command does:
1. Generate a random data points of size 2^8. Each point will belong to a 2^7 dimension space.
2. Run the program with 4 openmp threads
IMPORTANT:
It is important to compile the program as instructed above, because the program uses Intel task-queuing pragma
to parallelize recursive functions. This is not a standard OpenMP feature and other compiles cannot compile the pragmas.
If a different compiler is used, the pragmas will be ignored and the performance of the program will be badly affected.