Halide

Abstract

Image processing pipelines combine the challenges of stencil computations and stream programs. They are composed of large graphs of different stencil stages, as well as complex reductions, and stages with global or data-dependent access patterns. Because of their complex structure, the performance difference between a naive implementation of a pipeline and an optimized one is often an order of magnitude. Efficient implementations require optimization of both parallelism and locality, but due to the nature of stencils, there is a fundamental tension between parallelism, locality, and introducing redundant recomputation of shared values.We present a systematic model of the tradeoff space fundamental to stencil pipelines, a schedule representation which describes concrete points in this space for each stage in an image processing pipeline, and an optimizing compiler for the Halide image processing language that synthesizes high performance implementations from a Halide algorithm and a schedule. Combining this compiler with stochastic search over the space of schedules enables terse, composable programs to achieve state-of-the-art performance on a wide range of real image processing pipelines, and across different hardware architectures, including multicores with SIMD, and heterogeneous CPU+GPU execution. From simple Halide programs written in a few hours, we demonstrate performance up to 5× faster than hand-tuned C, intrinsics, and CUDA implementations optimized by experts over weeks or months, for image processing applications beyond the reach of past automatic compilers.

Automatically harnessing sparse acceleration

“…Many of these contain some degree of autotuning functionality to achieve good performance across different platforms [50]. Halide [39] was designed for image processing. [52].…”

On the Autotuning Potential of Time-stepping methods from Scientific Computing

“…From these DSL descriptions, a compiler can generate different code variants that can be tested in an online phase. An important application area for these approaches are stencil computations, see PATUS [24], Pochoir [25] and Halide [26] for approaches in this direction. All approaches mentioned above are application-specific, i.e., they have been developed for a specific application area and exploit specific properties of this application area.…”

“…The generation of different code variants is an important part of many autotuning approaches. The polytope model [30] and compiler-based approaches [26] are often used in this context. The resulting number of code variants can be large and efficient heuristic search strategies are important, including Simulated Annealing or Nelder-Mead [31].…”

Learning to optimize halide with tree search and random programs

“…They are computationally intense, and implementations often have to be highly optimized by experts, at great cost, to be usable in practice. The Halide programming language has proven to be a powerful tool for this task because it separates the algorithm -what you want to compute -from the schedule -how you want to compute it, including choices about memory locality, redundant computation, and parallelism [Ragan-Kelley et al 2012, 2013. While Halide makes it easy to try different schedules, writing schedules that achieve high performance is hard: it requires expertise in hardware architecture and optimization, and even then, the space of possible schedules is enormous and their performance can be difficult to predict.…”