librsb
A shared memory parallel sparse matrix library including Sparse BLAS.
... is especially well suited for symmetric and transposed multiplication variants. On these variants, librsb has been found to be faster than Intel MKL's implementation for CSR. Most numerical kernels code is auto generated, and the supported numerical types can be chosen by the user at build time.
librsb implements the Sparse BLAS standard, as specified in the BLAS Forum documents.