Nicolas Cellier uploaded a new version of Kernel to project The Trunk:
http://source.squeak.org/trunk/Kernel-nice.892.mcz ==================== Summary ==================== Name: Kernel-nice.892 Author: nice Time: 23 December 2014, 11:50:48.613 pm UUID: 85017561-ba41-4e1f-9c6b-37f97730b5ea Ancestors: Kernel-eem.891 Fix Float class>>fromIEEE32Bit: in the case of gradual underflow again... Details: A single precision (32bits) float gradual underflow is of the form biased exponent = 0 (-127 unbiased, but true value to take into account is -126) significand of the form 0.bbb....b (23 bit after the floating point, no implied 1) Once converted to double precision, this will not underflow... So we must shift the significand left until the highest bit pass left of the floating point : for example 0.001b...b ==> 1.b...b000 adjust the exponent (which is -126) by same number of shift (-129 in this example since there are 3 shifts) then remove the leading bit left of floating point - it is already taken into account as the implied 1 now that we do not underflow. =============== Diff against Kernel-eem.891 =============== Item was changed: ----- Method: Float class>>fromIEEE32Bit: (in category 'instance creation') ----- fromIEEE32Bit: word "Convert the given 32 bit word (which is supposed to be a positive 32bit value) from a 32bit IEEE floating point representation into an actual Squeak float object (being 64bit wide). Should only be used for conversion in FloatArrays or likewise objects." | sign mantissa exponent newFloat delta | word negative ifTrue: [^ self error:'Cannot deal with negative numbers']. word = 0 ifTrue: [^ Float zero]. sign := word bitAnd: 16r80000000. word = sign ifTrue: [^self negativeZero]. exponent := ((word bitShift: -23) bitAnd: 16rFF) - 127. mantissa := word bitAnd: 16r7FFFFF. exponent = 128 ifTrue:["Either NAN or INF" mantissa = 0 ifFalse:[^ Float nan]. sign = 0 ifTrue:[^ Float infinity] ifFalse:[^ Float negativeInfinity]]. exponent = -127 ifTrue: [ "gradual underflow (denormalized number) Remove first bit of mantissa and adjust exponent" delta := mantissa highBit. + mantissa := (mantissa bitAnd: (1 bitShift: delta - 1) - 1) bitShift: 24 - delta. - mantissa := (mantissa bitShift: 1) bitAnd: (1 bitShift: delta) - 1. exponent := exponent + delta - 23]. "Create new float" newFloat := self new: 2. newFloat basicAt: 1 put: ((sign bitOr: (1023 + exponent bitShift: 20)) bitOr: (mantissa bitShift: -3)). newFloat basicAt: 2 put: ((mantissa bitAnd: 7) bitShift: 29). ^newFloat! |
This code is acrobatically converting It might even be faster than these LargeInteger arithmetics (at least until 64bits spur)single precision float bits -> double precision float bits Frankly, it would be much easier conceptually to convert via single precision float bits -> abstract float representation (sign, exponent , significand) -> double precision bits Indeed, the second conversion is more than easy via (significandAsInteger asFloat timesTwoPower: exponent - bias). 2014-12-23 23:51 GMT+01:00 <[hidden email]>: Nicolas Cellier uploaded a new version of Kernel to project The Trunk: |
Free forum by Nabble | Edit this page |