Scalability and Complexity of 2-Dimensional SIMD Extensions
Esther Salamí, Mateo Valero Cortés, Alex Ramírez, Friman Sánchez, Mauricio A. Álvarez · Dialnet (Universidad de la Rioja) · 2004
Current SIMD extensions have probed to be effective for incrementing the performance of general purpose and embedded microprocessors for multimedia applications, but they seem to be not capable of scale their performance as required by new multimedia standards and applications. In this paper we describe and evaluate the scaling of a SIMD extension based on matrix registers and present a performance comparison with the scaling of SIMD extensions used in current microprocessors. This matrix architecture adapts well to the data structures used in image, video and audio applications and can be incorporated without bigger modifications into a high performance superscalar processor. Evaluations have demonstrated a significant performance improvement over conventional SIMD extensions. Speed-up over a 2-way superscalar processors with MMX-like extension goes up to 3.3X for complete applications. The reduced instruction count, that reduces the pressure in the front end, and the distributed nature of the register file make the matrix extension a complexityeffective solution suitable for the next generation of high performance processors. Key words— Computer architecture, SIMD, multimedia extensions, vector processors.