IDISA+: A portable model for high performance SIMD programming

Most of today’s commodity processors have single-instruction multiple-data (SIMD) instructions built in and provide SIMD within a register. However, different processor vendors tend to have different SIMD instruction sets which poses significant challenges to cross-platform SIMD programming. This thesis proposes a model called IDISA+ to overcome the compatibility issues and enable portable SIMD programming. There are more than 60 well-selected SIMD operations defined in the model, which are believed to support a broad range of applications. We have implemented the model as a toolkit with two components, a code generator for producing portable libraries and a test suite for both correctness and performance analysis on the libraries. For performance concerns, our model uses a least instruction count mechanism to select the best among implementation alternatives of library routines. The experimental results demonstrate the effectiveness of the generator and show that generated libraries in our model perform better than hand-tuned libraries.