Computing method of floating-point represented data
AssigneeFujitsu Ten Limited
InventorMasatsugu Kamimura
Life of this patent
Estimating the term…
Abstract
A computing method of floating-point represented data including dividing data x with a n (n>M) bit length into high rank data X' and low rank data .alpha. with a mutual m bit length, a mantissa part of the high rank data X' leaving the same bits as those of the data x by k bits (k.gtoreq.1) from its initial bit to replace the rest of the bits of X' by zero, the low rank data .alpha. being formed as X--X', and computing other data together with the high rank data X' and the low rank data .alpha..
Claims4
A computing method of a floating-point represented bit data X having an exponential part and a mantissa part, the mantissa part including first and second mantissa parts, comprising the steps of:
- (a) storing the floating-point represented bit data X in an n bit length register;
- (b) storing a floating-point represented high rank bit data X' having first and second high rank data parts, with the exponential part and the first mantissa part of the floating-point represented bit data X in the n bit length register in the first high rank data part and with zeros in the second high rank data part, in a first m bit length register wherein n is greater than m;
- (c) storing a floating-point represented low rank bit data .alpha. obtained by subtracting said floating-point represented high rank bit data X' from said floating-point represented bit data X in the n bit length register, in a second m bit length register;
- (d) computing other floating-point bit data stored in a third m bit length register, using said floating-point represented high rank bit data X' and said floating-point represented low rank bit data .alpha. registered in the first and second m bit length registers to obtain accurate floating-point data;
- and
- (e) storing said accurate floating-point data in a fourth m bit length register.
Dependencies are inferred from the claim text.