Computing Item-to-Item Similarity Using Cosine

I have a "User x Item" matrix as below:

user item1 item2 item3
u1   2     0     3
u2   1     2     0
u3   4     3     1
u4   0     2     2


I want to computer the similarity between items based on users. For example, to calculate similarity between items i1 and i2, I only choose users who have assigned values to both of these items.

The cosine similarity between two items (i1 and i2) is as follows:

cos(v1,v2) = (1*2 + 4*3)/sqrt[(1 + 4)*(16+9)]


My question is for item-to-item similarity, should I consider all users, or just common users who assigned values to the items?