Intel® 64 and IA-32 Architectures Optimization Reference Manual

MULTICORE AND HYPER-THREADING TECHNOLOGY8.2.1 Parallel Programming ModelsTwo common programming models for transforming independent task requirementsinto application threads are:• Domain decomposition• Functional decomposition8.2.1.1 Domain DecompositionUsually large compute-intensive tasks use data sets that can be divided into anumber of small subsets, each having a large degree of computational independence.Examples include:• Computation of a discrete cosine transformation (DCT) on two-dimensional databy dividing the two-dimensional data into several subsets and creating threads tocompute the transform on each subset• Matrix multiplication; here, threads can be created to handle the multiplication ofhalf of matrix with the multiplier matrixDomain Decomposition is a programming model based on creating identical orsimilar threads to process smaller pieces of data independently. This model can takeadvantage of duplicated execution resources present in a traditional multiprocessorsystem. It can also take advantage of shared execution resources between twological processors in HT Technology. This is because a data domain thread typicallyconsumes only a fraction of the available on-chip execution resources.Section 8.3.5, “Key Practices of Execution Resource Optimization,” discusses additionalguidelines that can help data domain threads use shared execution resourcescooperatively and avoid the pitfalls creating contentions of hardware resourcesbetween two threads.8.2.2 Functional DecompositionApplications usually process a wide variety of tasks with diverse functions and manyunrelated data sets. For example, a video codec needs several different processingfunctions. These include DCT, motion estimation and color conversion. Using a functionalthreading model, applications can program separate threads to do motion estimation,color conversion, and other functional tasks.Functional decomposition will achieve more flexible thread-level parallelism if it isless dependent on the duplication of hardware resources. For example, a threadexecuting a sorting algorithm and a thread executing a matrix multiplication routineare not likely to require the same execution unit at the same time. A design recognizingthis could advantage of traditional multiprocessor systems as well as multiprocessorsystems using processors supporting HT Technology.8-5