Intel® 64 and IA-32 Architectures Optimization Reference Manual

OPTIMIZING FOR SIMD FLOATING-POINT APPLICATIONSNow consider the case when the data is organized as SoA. Example 6-2 demonstrateshow four results are computed for five instructions.Example 6-2. Pseudocode for Vertical (xxxx, yyyy, zzzz, SoA) Computationmulpsmulpsmulpsaddpsaddps; x*x' for all 4 x-components of 4 vertices; y*y' for all 4 y-components of 4 vertices; z*z' for all 4 z-components of 4 vertices; x*x' + y*y'; x*x'+y*y'+z*z'For the most efficient use of the four component-wide registers, reorganizing thedata into the SoA format yields increased throughput and hence much better performancefor the instructions used.As seen from this simple example, vertical computation yielded 100% use of theavailable SIMD registers and produced four results. (The results may vary based onthe application.) If the data structures must be in a format that is not “friendly” tovertical computation, it can be rearranged “on the fly” to achieve full utilization of theSIMD registers. This operation is referred to as “swizzling” operation and the reverseoperation is referred to as “deswizzling.”6.5.1.2 Data SwizzlingSwizzling data from one format to another may be required in many algorithms whenthe available instruction set extension is limited (for example: only SSE is available).An example of this is AoS format, where the vertices come as XYZ adjacent coordinates.Rearranging them into SoA format (XXXX, YYYY, ZZZZ) allows more efficientSIMD computations.For efficient data shuffling and swizzling use the following instructions:• MOVLPS, MOVHPS load/store and move data on half sections of the registers.• SHUFPS, UNPACKHPS, and UNPACKLPS unpack data.To gather data from four different memory locations on the fly, follow these steps:1. Identify the first half of the 128-bit memory location.2. Group the different halves together using MOVLPS and MOVHPS to form an XYXYlayout in two registers.3. From the 4 attached halves, get XXXX by using one shuffle, YYYY by usinganother shuffle.ZZZZ is derived the same way but only requires one shuffle. Example 6-3 illustratesthe swizzle function.6-7