How codeml detects positive selection

Hi,

I am using the codeml package of the PAML software for positive selection analysis. I have tried to read the manual as well as the other literature of the software but still I have a question about how it actually detects the positive selection. I got the fact that if we take dN/dS as the average over all the codon sites in the gene, then it will rarely be >1 because not all codon sites are under positive selection. Thus the software implements different site model with different omega ratios for different codon sites, but still I did not understand how it calculates dN/dS for individual sites. I also read that it first calculates the dN/dS for individual site and then sum it over all the sites to get the final value of dN/dS over whole gene, but again how is the dN/dS calculated for individual codon sites?