---

Writing Kernels

The heart of any image processing filter is the kernel file. A kernel file contains one or more kernel routines and any required subroutines. A kernel routine gets called once for each pixel for the destination image. The routine must return a vec4 data type. Although this four-element vector typically contains pixel data, the vector is not required to represent a pixel. However, the kernel routines in this chapter produce only pixel data because that’s the most common data returned by a kernel routine.

A kernel routine can:

• Fabricate the data. For example, a routine can produce random pixel values, generate a solid color, or produce a gradient. It can generate patterns, such as stripes, a checkerboard, a starburst, or color bars.

• Modify a single pixel from a source image. A kernel routine can adjust hue, exposure, white point values, replace colors, and so on.

• Sample several pixels from a source image to produce the output pixel. Stylize filters such as edge detection, pixellate, pointillize, gloom, and bloom use this technique.

• Use location information from one or more pixels in a source image to produce the output pixel. Distortion effects, such as bump, pinch, and hole distortions are created this way.

• Produce the output pixel by using data from a mask, texture, or other source to modify one or more pixels in a source image. The Core Image stylize filters—height field from mask, shaded material, and the disintegrate with mask transition—are examples of filters that use this technique.

• Combine the pixels from two images to produce the output pixel. Blend mode, compositing, and transition filters work this way.

This chapter shows how to write a variety of kernel routines. First you’ll see what the programming constraints, or rules, are. Then you’ll learn how to write a simple filter that operates on one input pixel to produce one output pixel. As the chapter progresses, you’ll learn how to write more complex kernel routines, including those used for a multipass filter.

Although the kernel routine is where the per-pixel processing occurs, it is only one part of an image unit. You also need to write code that provides input data to the kernel routine and performs a number of other tasks as described in “Writing the Objective-C Portion.” Then you’ll need to bundle all the code by following the instructions in “Preparing an Image Unit for Distribution.”

Before continuing in this chapter, see Core Image Kernel Language Reference for a description of the language you use to write kernel routines. Make sure you are familiar with the constraints discussed in “Kernel Routine Rules.”

Contents:

Writing Simple Kernel Routines
Testing Kernel Routines in Quartz Composer
Writing Advanced Kernel Routines
Writing Kernel Routines for a Detective Lens
Solution to the Kernel Challenge
Next Steps

---