fangde / gpu_badmm_mt Goto Github PK

Mass transportation problem: C, a, b.  The default a = 1, b = 1;
1. The file C,a,b for mass transportation problem are generated using matlab storing in a binary file in row-major order. For example,
m = 1024;
n = m;
C = rand(m,n);
fid = fopen([int2str(n) 'C.dat'],'w');
fwrite(fid,[m,n],'int');
fwrite(fid,C','float');
fclose(fid);

2. compile cuda code and tune parameters
nvcc -o badmm_mt badmm_mt.cu badmm_kernel.cu -lcublas -arch sm_13
nvprof ./badmm_mt dim_file_name rho max_iteration tolerance_stop_badmm num_steps_print save_output
a. dim_file_name: the name of cost matrix is stored in binary with name dim*1024C.dat. The name of a,b files are dim*1024a.dat and dim*1024b.dat
b. rho (float): the parameter of badmm
c. max_iteration (int): the maximum number of iterations of badmm
d. tol (float): the stopping condition of primal and dual residuals of badmm. If less than tol, badmm will stop.
e. num_steps_print (int): print intermediate results every num_steps_print step. If num_steps_print = 0, no print. 
f. save_output: write the final result of mass transportation into a binary file named X_out.dat
The default values will be chosen if setting them to zero or no specification


3. read X_out to matlab
fid = fopen('X_out.dat','r');
X = fread(fid,[n,m],'float');
fclose(fid);
X = X';