发明名称 |
Block floating-point representation of data samples |
摘要 |
A method for representing a plurality of data samples is disclosed. The method generally includes steps (A) and (B). Step (A) may generate an exponent common to all of the data samples using a circuit. Step (B) may generate a plurality of mantissae based on the exponent. Each of the mantissae generally corresponds to a different one of the data samples. |
申请公布号 |
US9423999(B1) |
申请公布日期 |
2016.08.23 |
申请号 |
US200912582831 |
申请日期 |
2009.10.21 |
申请人 |
Ambarella, Inc. |
发明人 |
Linzer Elliot N. |
分类号 |
G06F7/38;G06F7/483;G06F9/30 |
主分类号 |
G06F7/38 |
代理机构 |
Christopher P. Maiorana, PC |
代理人 |
Christopher P. Maiorana, PC |
主权项 |
1. A method for representing a plurality of data samples, comprising the steps of:
(A) calculating a maximum value of said data samples in a camera; (B) generating an exponent common to all of said data samples based on said maximum value and a threshold value in an iterative sequence; (C) generating a plurality of mantissae by quantizing said data samples based on a plurality of randomized quantization bins and said exponent, wherein each of said mantissae corresponds to a different one of said data samples; (D) buffering a block containing said mantissae and said exponent in a memory; (E) converting said block as buffered in said memory into a fixed-point representation; and (F) processing said data samples as a picture after conversion to said fixed-point representation. |
地址 |
Santa Clara CA US |