attention calculation at matrix level

Question

I am confused with the calculation in self attention or attention only.

Lets talk about self-attention first and I have:

x -> [batch_size, query_len, em