First, you should look into whether the auto-vectorization can help your application. That's parallelism within a single thread.
For threaded parallelism, if your program is easily parallelizable, the option /Qparallel may work.
If you use MATMUL on large enough data sets, /Qopt-matmul implicitly invokes the MKL library. Consideration of direct MKL library function calls may give better performance as well as a much wider variety of optimized parallel functions.
