meissner@dg_rtp.UUCP (Michael Meissner) (12/24/85)
The following describes the Data General Eclipse floating point format. Note, that some of the submissions were wrong, in specifying that the DG format is exactly the same as the IBM format. It is true we use the same number of bits, but the IBM format expresses the exponent as a power of 2, whereas in the DG format, the exponent is a power of 16. Single Precision +-+-------+------------------ ........ --------+ |S| Exp | Mantissa | +-+-------+------------------ ........ --------+ 1 7 bits 24 bits bit Double Precision +-+-------+------------------ ........ --------+ |S| Exp | Mantissa | +-+-------+------------------ ........ --------+ 1 7 bits 56 bits The exponent is in excess-64 notation, representing the appropriate power of 16. The mantissa is grouped into 6 (single) or 14 (double) hexidecimal digits. For the 16-bit Eclipses, the mantissa does not have to be normalized. For the 32-bit MV line, the POOPS manual specifies that the number must always be normalized, and implementation dependent results occur if the number is not normalized. In practice, the MV/8000 will normalize the number before doing any calculations, and the MV/10000 will generate a floating point fault if given an unnormalized number. One special guarantee is given in that loads and stores affecting floating point registers do not change the format, so that a floating load, followed by a floating store can be used to move 64 bits of information. Floating Point Hexidecimal 0: 00000000 1: 41100000 2: 41200000 3: 41300000 4: 41400000 5: 41500000 6: 41600000 7: 41700000 8: 41800000 9: 41900000 10: 41A00000 11: 41B00000 12: 41C00000 13: 41D00000 14: 41E00000 15: 41F00000 16: 42100000 17: 42110000 32: 42200000 64: 42400000 128: 42800000 256: 43100000 512: 43200000 0.5: 40800000 0.25: 40400000 0.125: 40200000 6.25e-02: 40100000 -1: C1100000 Michael Meissner Data General Corporation ...{ ihnp4, decvax }!mcnc!rti-sel!dg_rtp!meissner
hes@ecsvax.UUCP (Henry Schaffer) (12/25/85)
> > The following describes the Data General Eclipse floating point format. > Note, that some of the submissions were wrong, in specifying that the DG > format > is exactly the same as the IBM format. It is true we use the same number of > bits, but the IBM format expresses the exponent as a power of 2, whereas in > the > DG format, the exponent is a power of 16. In the IBM format, the exponent (also called characteristic) is the power of 16 (not 2) by which the fraction (also called mantissa) must be multiplied. With a base of 2, the range of floating point numbers would be *much* too small unless the exponent were much longer. > ... > > The exponent is in excess-64 notation, representing the appropriate power > of 16. The mantissa is grouped into 6 (single) or 14 (double) hexidecimal > digits. > This also describes the IBM floating point format. > > Floating Point Hexidecimal > > 0: 00000000 [*] > 1: 41100000 [*] > 2: 41200000 [*] > 3: 41300000 [*] > ... > 9: 41900000 [*] > 10: 41A00000 [*] > 11: 41B00000 [*] > ... > 15: 41F00000 [*] > 16: 42100000 [*] > ... > 32: 42200000 [*] > ... > 256: 43100000 [*] > ... > 0.5: 40800000 [*] > 0.25: 40400000 [*] > 0.125: 40200000 > 6.25e-02: 40100000 [*] > -1: C1100000 [*] > * These values were in the IBM manual I looked in and agree exactly with the DG values shown. I don't see any difference between the IBM and DG formats from what is given here. > Michael Meissner --henry schaffer
ark@alice.UucP (Andrew Koenig) (12/25/85)
> The following describes the Data General Eclipse floating point format. > Note, that some of the submissions were wrong, in specifying that the DG format > is exactly the same as the IBM format. It is true we use the same number of > bits, but the IBM format expresses the exponent as a power of 2, whereas in the > DG format, the exponent is a power of 16. In IBM format, the exponent is a power of 16 also. Numbers may be unnormalized, but the only floating-point instructions that produce unnormalized results are "add unnormalized" and "subtract unnormalized."