Imaging applications can take advantage of Hyper-Threading technology, reaping substantial performance benefits.

---

from Intel Corp.

Desktop developers, it's time to thread your apps. Today's operating systems are engineered to make best use of computer resources, and Hyper-Threading Technology (HT technology) better utilizes those resources. This article illustrates how imaging applications can take advantage of this groundbreaking technology, reaping substantial performance benefits. The code samples discussed in this paper show an18–20% performance gain.

Threading a serial application
To reap the benefits of HT technology, you need to thread your applications correctly. Choose the most effective threading method to get off to a fast start. To do this, you need to describe your application in terms of two threading or parallel programming models: data parallelism (data decomposition) and task-level parallelism (functional decomposition).

Data decomposition
Data decomposition or data parallelism is a programming model that implies creation of multiple threads to independently process smaller pieces of data. This implies that the same independent operation is applied repeatedly to different data. Compute-intensive tasks with a large degree of independence like compute-intensive loops in applications are good candidates for data decomposition.

A good decomposition requires that the threads not conflict with each other and operate with completely unrelated data. Work to avoid multiple threads accessing the same data at the same time and computations that depend on each other. Good decomposition may require a code rewrite to avoid thread conflicts.

Functional decomposition
Functional decomposition or task-level parallelism is a programming model that implies creation of multiple threads, mapped to process independent functional tasks that execute asynchronously. Again when running multiple threads it's important to minimize thread conflicts.

Parallelizing imaging applications
In the case of imaging applications, it's pretty easy to find compute-intensive tasks to parallelize. Image processing algorithms consist of pixel operations that are often independent. Hence their computations can be done in parallel.