Facilitate SIMD-Code-Generation in the Polyhedral Model by Hardware-aware Automatic Code-Transformation

Although Single Instruction Multiple Data (SIMD) units are available in general purpose processors already since the 1990s, state-of-the-art compilers are often still not capable to fully exploit them, i.e., they may miss to achieve the best possible performance. We present a new hardware-aware and adaptive loop tiling approach that is based on polyhedral transformations and explicitly dedicated to improve on auto-vectorization. It is an extension to the tiling algorithm implemented within the PluTo framework [4, 5]. In its default setting, PluTo uses static tile sizes and is already capable to enable the use of SIMD units but not primarily targeted to optimize it. We experimented with different tile sizes and found a strong relationship between their choice, cache size parameters and performance. Based on this, we designed an adaptive procedure that specifically tiles vectorizable loops with dynamically calculated sizes. The blocking is automatically fitted to the amount of data read in loop iterations, the available SIMD units and the cache sizes. The adaptive parts are built upon straightforward calculations that are experimentally verified and evaluated. Our results show significant improvements in the number of instructions vectorized, cache miss rates and, finally, running times.