Abstract:

Based on the analysis on the existing problems in the single precision inverse algorithm, we design and implement a single precision inverse based on the first order Taylor series. Compared with the traditional algorithm, the resource consumption, operation cycle and efficiency are improved. The main logic module of this floating point algorithm is composed of a 24 bit integer adder, a ROM and a 24 bit multiplier. The mantissa in range of ［1, 2) is divided to 4096 intervals on average, and the reciprocal square of the starting point of each interval is stored in a lookup table. Then the first order Taylor series is applied to compute the inverse value of each interval. Simulation results are consistent with the theoretical results, which meets the accuracy requirement of the single precision. This algorithm has been successfully applied to the third generation of GPU JM7200.

Key words: GPU, first order Taylor series, look-up table, single precision inverse