Abstract:

Compilers or programmers have to use a permutation instruction to reorganize the element of vectors to get the correct operands for the SIMD instructions in the design of a SIMD program recently. But these redundant permutation instructions result in performance loss. This paper proposes a new intermediate representation, which contains enough address messages for the operands of scalar and vector. It makes the generation of data permutations be postponed and the redundant permutation instruction be reduced. We utilize this new intermediate representation to implement an optimization algorithm on data permutation, which can effectively reduce the performance loss from permutation instruction. The test result to a group of typical multimedia program shows that the algorithm can achieve 7% performance acceleration on average.

Key words: intermediate representation;SIMD;data permutation