Input matrix data must be written in a column-major fashion, and the matrix is assumed to be Hermitian positive-definite. The output matrix will be written in column-major fashion. The Cholesky operation is susceptible to catastrophic cancellation; therefore, it is recommended to ensure your matrix is well-conditioned.
Data is operated on in vecSampleNum * vecSampleNum chunks, and thus only chunks along and below the diagonal are operated on. Upper matrix data is assumed to be zero, thus chunks in the upper-triangular output are undefined.
In a single-tile implementation, TP_DIM must be a multiple of vecSampleNum. In multi-tile implementations, TP_DIM / TP_GRID_DIM must be a multiple of vecSampleNum, with a maximum TP_DIM of 1024. Resulting sub-matrices must fit in local storage.
TP_CASC_LEN must be between 1 and TP_DIM, with a max value of 32.