Abstract:

We propose a vectorization method of QR decomposition with  Givens rotation on Matrix processors. According to the systematic characteristics of Matrix architecture, the computation tasks are evenly distributed to all vector processing elements by optimizing the memory access to vector data and calculation. We also design a double DMA buffering scheme to smooth the data transfers, which can fully overlap the kernel computation time and the DMA data transfer time so that the kernel computation is always at its peak speed and the best computation efficiency is achieved.  Experimental results show that the proposal can achieve higher computation efficiency and performance speedup.

Key words: QR decomposition;vector processor;Givens rotation;software pipeline