Try redsvd (BSD license). It implements clean and very efficient, modern algorithms for SVD, including partial (truncated) SVD.
Redsvd is built on top of the beautiful C++ templating library, eigen3. Since you mention installing is an issue, you'll like to hear eigen3 requires no installation. It's just templates (C++ header files).
Also, there don't exist "more than one decomposition of a single matrix". The SVD decomposition always exists and is unique, up to flipping signs of the corresponding U/V vectors.
