Intra-vector SIMD instructions for core specialization

Current research is mainly focussing on exploiting TLP to increase performance. Another avenue, however, for achieving performance scalability is specialization. In this paper we propose application specific intra-vector instructions for two dimensional signal processing kernels. In such kernels usually significant data rearrangement overhead is required in order to use the SIMD capabilities. When using the intra-vector instructions the overhead can be avoided. We have implemented intra-vector instructions in the Cell SPU core and measured speedups of up to 2.06, with an average of 1.45.