在MLIR上透過隨機優化演算法自動化Affine排程並加速Halide

Abstract

電腦視覺算法在影像中心的設備上的使用越來越多，這使得我們需要優化這些演算法以提高其在資源受限的平台上的性能。Halide是一種在C++上的特定領域語言用來處理高性能圖像處理演算法，可以將演算法與指令排程分離。本論文提出了一種使用模擬退火的隨機演算法優化排程的方法，以提高Halide電腦視覺演算法的性能，從而實現在有限時間內探索仿射轉換的全局最優解。為了將Halide程式碼轉換為MLIR，我們使用了新的編譯流程，即Halide到MLIR（HTM）編譯器。此方法的效能將在不同平台上進行評估，例如x86、Arm和RISC-V。研究展示了MLIR在仿射方言(Affine Dialect)上的轉換和優化能力，並顯示了使用隨機演算法優化基礎建設以充分最佳化MLIR不同優化能力的必要性。

The increasing usage of computer vision algorithms in camera-centric devices has led to a growing need for optimizing these algorithms to improve their performance on resource-constrained platforms. Halide is a language specific to image processing algorithms that separates the algorithm's scheduling from its implementation, resulting in high performance. This thesis proposes an approach to improve the performance of Halide computer vision algorithms using stochastic algorithms such as simulated annealing to optimize scheduling, which enables exploring the global optimum of affine scheduling in constrained time. To convert the Halide program to MLIR, we use novel compile flows, namely the Halide to MLIR (HTM) converter. The efficacy of the approach will be evaluated on different platforms, such as x86, ARM, and RISC-V. The study demonstrates the potential of MLIR's transformation and optimization capabilities on Affine dialects and highlights the need for tuning infrastructure to fully leverage MLIR's optimization capabilities.