LMS Halide

This repository contains the source code for a replica of Halide, a DSL for image processing, using the lightweight modular staging library.

The DSL uses multi-stage programming to generate efficent, low-level C code from high-level Scala. Example DSL programs can be found in /src/main/test/scala/TestPrograms.scala, and the corresponding C code in /testOutput. In order to actually manipulate images, the output C code should be compiled with /builder.c - this requires that the libpng library is installed for PNG encoding / decoding.

Example

The following is a DSL implementation of a simple box-blur:

trait TwoStageBoxBlurFast extends TestPipeline {
	override def prog(in: Input, w: Rep[Int], h: Rep[Int]): Rep[Unit] = {
		val f = func[Short] {
			(x: Rep[Int], y: Rep[Int]) => (in(x, y) + in(x+1, y) + in(x-1, y)) / 3.toShort
		}
		val g = final_func[Short] {
			(x: Rep[Int], y: Rep[Int]) => (f(x, y) + f(x, y+1) + f(x, y-1)) / 3.toShort
		}

    f.storeRoot()
    f.computeAt(g, "y")
    f.vectorize("x", 16)
    g.vectorize("x", 16)
  }
}

This example produces a 'schedule' where the f and g x-loops are vectorised, and f is stored globally but computed per y-coordinate of g.