Applications Tuning for Streaming SIMD Extensions

In early 1997, Intel formed an engineering lab whose charter was to apply a new set of instructions to the optimization of software applications. This lab worked with commercial software companies to increase the performance of their applications by using these new instructions. Two years later, this new instruction set has been made public as a principal new feature of the Pentium® III processor, the Streaming SIMD Extensions. Many of the commercial software companies’ applications on which the lab consulted have been brought to market, demonstrating significant performance improvements by using the Streaming SIMD Extensions. This paper describes many of the principles and concepts developed as a result of that activity. The Streaming SIMD Extensions expand the Single Instruction/Multiple Data (SIMD) capabilities of the Intel® Architecture. Previously, Intel® MMXTM instructions allowed SIMD integer operations. These new instructions implement floating-point operations on a set of eight new SIMD registers. Additionally, the Streaming SIMD Extensions provide new integer instructions operating on the MMX registers as well as cache control instructions to optimize memory access. Applications using 3D graphics, digital imaging, and motion video are generally well suited for optimization with these new instructions. Data organization plays a pivotal role in the performance of applications in the above areas. This paper explores three data organizations (Array of Structure, Structure of Array, and Hybrid data orders) and their impact on SIMD processing performance. The impact of cache control instructions, such as the prefetch instructions, is also examined. Examples of applying the Streaming SIMD Extensions to 3D transform and lighting, bilinear interpolation, video block matching, and motion compensation are considered. The principles applied in these examples can be extended to many other algorithms and applications.