VLIW DSP vs. superscalar implementation of a baseline 11.263 video encoder

A Very Long Instruction Word (VLIW) processor and a superscalar processor can execute multiple instructions simultaneously. A VLIW processor depends on the compiler and programmer to find the parallelism in the instructions, whereas a superscaler processor determines the parallelism at runtime. This paper compares TI TMS320C6700 VLIW digital signal processor (DSP) and SimpleScalar superscalar implementations of a baseline 11.263 video encoder in C. With level two C compiler optimization, a one-way issue superscalar processor is 7.5 times faster than the VLIW DSP for the same processor clock speed. The superscalar speedup from one-way to four-way issue is 2.88:1, and from four-way to 256-way issue is 2.43:1. To reduce the execution time on the C6700, we write assembly routines for sum-of-absolute-difference, interpolation, and reconstruction, and place frequently used code and data into on-chip memory. We use TI's discrete cosine transform assembly routines. The hand optimized VLIW DSP implementation is 61/spl times/ faster than the C version compiled with level two optimization. Most of the improvement was due to the efficient placement of data and programs in memory. The hand optimized VLIW implementation is 14% faster than a 256-way superscalar implementation without hand optimizations.