A media-oriented vector architectural extension with a high bandwidth cache system
Ye Gao, Naoki Shoji, Ryusuke Egawa, Hiroyuki Takizawa, Hiroaki Kobayashi · 2012
Although modern SIMD extensions enable to improve the performance of MMAs, the limited parallel-processing capability and low memory bandwidth are their drawbacks to execute next generation MMAs. This paper proposes a novel media-oriented vector architectural extension (MVPX) to aim at accelerating a wide range of next generation MMAs. MVPX consists of an out-of-order vector processing mechanism (OVPM) and a high bandwidth cache system (MVP-cache). OVPM enables MVPX to take advantage of high data level parallelism for various MMAs. Meanwhile, MVP-cache can reduce the data transfer from/to off-chip memory accesses, and thus increase the sustained memory bandwidth by exploiting reusable vector data. It has a low power/area overhead approach to increasing memory bandwidth, compared with approaches to increasing the off-chip memory bandwidth. As a result, MVPX expands the potential of GPPs for next generation media processing in a power efficient way.