A single/double precision floating-point reciprocal unit design for multimedia applications

Modern graphic processors, multimedia processors, and general-purpose processors with multimedia extensions provide SIMD floating-point instructions. SIMD floating-point reciprocal operation is commonly used in 2D and 3D applications, which mostly use single precision floating-point operands. Consequently, efficient single precision units are crucial for high performance systems. This paper introduces a packed floating point reciprocal unit that can perform reciprocal of either a double precision or two parallel single-precision floating-point numbers using fast multipliers.