Design Notes - 2023.2 English

Vitis Libraries

Release Date
2023.2 English
  • Target: \(A=UMV*\), \(A[M*N]\) is input matrix, \(U[M*M]\), \(V*[N*N]\) and \(R[M*N]\) are the output matrix via singular value decomposition.
  • DataType supported: cfloat.
  • DataSize supported: input matrix size \(M\) is the times of 4 and no bigger than 1024, and \(N\) shoulb be no bigger than 256.
  • Description:
    Singular value decomposition took one-sided Jacobi algorithm to solve. It’s an iterative approximation algorithm. Each time the AIE graph perform “column-sweep” on the input matrixs and make columns of its inputs more “orthogonal”. After number of iterations, the final result is considered to be good enough as approximation of result of singular value decomposition. The number of iteration needed depends on size of matrix and precision requirements. Thus our design does not put any assumption on it and leave it to users to determine how many iterations they need.