Initial Evaluation of Multimedia Extensions on VLIW Architectures

Media processing has motivated strong changes in the focus and design of processors. The inclusion of μSIMD multimedia extensions such as MMX is a cost effective option to improve the performance of those regions of the program with large amounts of DLP. This paper provides an initial evaluation of μSIMD and vector-SIMD enhanced VLIW architectures. We show that these two architectures execute respectively an average of 40% and 57% fewer operations than the reference VLIW architecture. However, when most of the available DLP parallelism has been exploited via multimedia extensions or wide-issue static scheduling, the remaining of the program exhibits only modest amounts of ILP (1.40 operations per cycle for a 8-issue width architecture). We claim that, in general, vector-SIMD extensions achieve the highest speed-ups while still reducing the fetch pressure, although for wide-issue μSIMD architectures reach a similar performance at a lower cost.