In the spec, the pseudocode defines them as

   FMADD:   addend + ( op1 * op2)
   FMSUB:   addend + (-op1 * op2)
   FNMADD: -addend + (-op1 * op2)
   FNMSUB: -addend + ( op1 * op2)

which is to say that in all three cases the result is the output of a fused multiply-add; the difference is just whether (a) the addend and (b) op1 are negated before feeding them in.

IEEE 754-2008 only defines that you have to provide a fusedMultiplyAdd() operation, so it's up to the CPU architecture whether folding in the free negations is something they want to provide, and if so, where exactly the negation is done.