| Anonymous | Login | Signup for a new account | 02-09-2010 14:33 UTC |
| Main | My View | View Issues | Change Log | Docs |
| Viewing Issue Simple Details [ Jump to Notes ] | [ View Advanced ] [ Issue History ] [ Print ] | ||||||||
| ID | Category | Severity | Reproducibility | Date Submitted | Last Update | ||||
| 0003504 | [Squeak] Kernel | minor | always | 04-26-06 04:16 | 04-29-06 22:54 | ||||
| Reporter | nicolas cellier | View Status | public | ||||||
| Assigned To | |||||||||
| Priority | normal | Resolution | fixed | ||||||
| Status | closed | Product Version | 3.9 | ||||||
| Summary | 0003504: Float asInteger conversion is inexact... | ||||||||
| Description |
| x | x := (10 raisedTo: 16) asFloat. { "This one give correctly 0" (x asTrueFraction - (10 raisedTo: 16)) isZero. "This one give incorrectly -1" (x asInteger - (10 raisedTo: 16)) isZero. "Guilty code lie in Float>>truncated: this one give 999999999999999" (x quo: 1073741823.0) * 1073741823 + (x rem: 1073741823.0) truncated. }. |
||||||||
| Additional Information |
Integer numbers are represented exactly in IEEE 754 floating point as long as they do not exceed the mantissa limit (53 bits in double precision). (10 raisedTo: 16) > (2 raisedTo: 53), however, 16 trailing bits are zero... So there is no round of error and (10 raisedTo: 16) is represented exactly. In fact, it is ((5 raisedTo: 16) asFloat timesTwoPower: 16), and (5 raisedTo: 16) < (2 raisedTo: 53). Problem is located in Float>>truncated asTrueFraction does the right job, but other branch with quo: and rem: might imply inexact arithmetic and round off error... |
||||||||
| Attached Files |
 Kernel-Float-truncated-Test.1.cs [^] (819 bytes) 04-26-06 05:29 Kernel-Float-truncated-Patch.1.cs [^] (1,031 bytes) 04-26-06 05:30 |
||||||||
|
|
|||||||||
 Relationships |
||||||
|
||||||
|
Notes |
|
|
(0004813 - 975 - 1074 - 1074 - 1074 - 1074 - 1074) nicolas cellier 04-26-06 05:29 |
Instead of using SmallInteger maxVal (1073741823), i suggest using (SmallInteger maxVal bitShift: -1)+1 (that is 2 raisedTo: 29, or 536870912) Since it is a power of two, float division should be exact. I provide a TestCase and a patch. Tested only on 1000 random numbers between 1.0e12 and 2.0e16... No performance loss. Note that mean penalty of using self asTrueFraction truncated would be less than two on my AMD ATHLON linux, i wonder if it would not be better to use this simpler code. { Time millisecondsToRun: [| r x | r := Random new. 10000 timesRepeat: [x := r next * 1.0e16 + 1.0e12. x abs < 2.0e16 ifTrue: [(x quo: 536870912.0) * 536870912 + (x rem: 536870912.0) truncated] ifFalse: [x asTrueFraction]]] . Time millisecondsToRun: [| r | r := Random new. 10000 timesRepeat: [(r next * 1.0e16 + 1.0e12) truncated ]] . Time millisecondsToRun: [| r | r := Random new. 10000 timesRepeat: [(r next * 1.0e16 + 1.0e12) asTrueFraction truncated]] . } |
|
(0004827 - 10 - 10 - 10 - 10 - 10 - 10) ducasse 04-29-06 22:54 |
in 39-7026 |
|
Issue History |
|||
| Date Modified | Username | Field | Change |
| 04-26-06 04:16 | nicolas cellier | New Issue | |
| 04-26-06 05:29 | nicolas cellier | Note Added: 0004813 | |
| 04-26-06 05:29 | nicolas cellier | File Added: Kernel-Float-truncated-Test.1.cs | |
| 04-26-06 05:30 | nicolas cellier | File Added: Kernel-Float-truncated-Patch.1.cs | |
| 04-29-06 22:54 | ducasse | Status | new => closed |
| 04-29-06 22:54 | ducasse | Note Added: 0004827 | |
| 04-29-06 22:54 | ducasse | Resolution | open => fixed |
| 04-30-06 21:17 | MarcusDenker | Relationship added | related to 0003510 |
| Mantis 1.0.8[^]
Copyright © 2000 - 2007 Mantis Group
48 total queries executed. 34 unique queries executed. |  |
