Multiloop Parallelisation Using Unrolling and Fission

Figure 7

Mapping the speech recognition system to the heterogeneous computing system. “vq3” are the unrolled 3 iterations of vector quantisation. “autocc12” are the unrolled 12 iterations for autocorrelation. “hmmdec12” are the unrolled 12 iterations for HMM decoding. Two “hmmdec12” mean the outer loop of HMM decoding is unrolled for 2 iterations, that is, decoding of two words is executed in parallel.