One of the common problems faced by many programmers is the ensuring the accuracy of floating point arithmetic, mainly due to the rounding errors that can occur when calculations are performed on numbers that cannot be fully represented in a binary floating-point format (The size of this format determines the precision, and therefore the accuracy of calculations).
OpenInsight is built with the Microsoft C++ compiler, which limits the floating-point format to 64-bits (known as the “Double” type), and Basic+ variables that represent non-integer numbers use this type internally. The use of this 64-bit format is one of the problems commonly noticed by developers who have moved their systems from the older Advanced Revelation platform to OpenInsight, because the internal floating-point format for R/Basic variables was the 80-bit “Long Double” type instead. This means that calculations from the same RBasic/Basic+ code running on ARev and OI may produce different results due to this reduction in precision.
(The Long Double is unfortunately not supported by the MS C++ compiler as, according to the VC++ compiler team back in 2006: “The major reason is that FP code generation has been switching to the use of SSE/SSE2/SSE3 instruction sets instead of the x87 FP stack since that is what both the AMD and Intel recent and future chip generations are focusing their performance efforts on. These instruction sets only support 32 and 64 bit FP formats”).
The “integer workaround”
A common workaround for precision problems is to control the calculations at each step so you are effectively dealing with integer operations – this can be done by using the “MD” IConv/OConv functions to control the precision, or by simply multiplying values by a known factor and dividing the result again afterwards.
Both of these methods can make the code messy and obscure the intent, and, depending on the values used, may result in integer overflow if they are too large (though this is less likely to happen on a 64-bit system like OpenInsight 10 however).
The “Extended Precision Operators”
Another option to help mitigate these calculation problems are the Extended Precision Operators that were introduced in OpenInsight 9.3:
These do allow you to specify the precision to use at each step, but suffer from the need to rewrite existing code to use them. It is also easy to lose precision if you inadvertently use a “normal” operator in between them as well, as the following example demonstrates:
a = _divx( c, b, 24 ) ; // "a" is full precision (24) if ( a > 1 ) then ; // "a" converted to "double" by ">" operator // do stuff end b = _addx( a, z ) ; // "a" is no longer full precision passed to _addx
The new “Extended Precision Mode”
With the upcoming release of version 10.0.7, OpenInsight now supports a new feature for dealing with high precision calculations called Extended Precision Mode (EP Mode). When this mode is enabled all the normal maths operators switch into an “extended mode” and the results of calculations are stored in Basic+ variables using a new internal type introduced specifically for maintaining the precision.
This means that existing code can be reused by simply adding statements to activate and deactivate the mode as needed via the new SetEPMode() function , e.g:
Compile Function Mickey_Mouse_EPM_Test( void ) Declare Function GetEPMode $Insert Logical epMode = GetEPMode() Call SetEPMode( FALSE$ ) GoSub runTest Call SetEPMode( TRUE$ ) GoSub runTest Call SetEPMode( epMode ) Return runTest: a = 10.12346 * (22/7); b = 100000; For x = 1 To 1000 b -= a Next Call Send_Dyn( b ) Return
(Normal) 68183.4114285739 (EPMode) 68183.41142857142857142857142857143725
The following operators are affected by the EP Mode:
+ += - -= * / == or = != or <> > < >= <= mod() int() abs() atan() cos() exp() ln() pwr() sin() sqrt() tan()
By default EP Mode is not enabled because calculations are slower due to the extra processing needed to maintain precision, and this would have a detrimental effect on the performance of your applications if it were permanently enabled, so you should only use it when absolutely required.
You should also note that the EP Mode and Precision are set to their default values for each new request made to the engine (i.e each event or web-request). This is to protect against situations where an error condition could force the engine to abort processing before the EP Mode settings could be reset, thereby leaving it in an undesired state (This is similar to the way UTF-8 mode works so that data integrity is preserved).
Controlling the precision level
The actual level of precision (i.e. the decimal places) is controlled by another setting, which is updated by the new SetEPModePrecision() function (The greater the precision the longer calculations will take to perform). By default the precision is set to 32.
decPlaces = GetEPModePrecision() Call SetEPModePrecision( 24 )
Default EP settings
The default settings can be changed in the Application Properties dialog launched from the IDE Settings menu.
Further reading on floating point arithmetic
More information on floating point arithmetic can be found by following the links below:
Did you mean “Iconv” rather than “Oconv” in the the “integer workaround” paragraph?
Indeed – updated thanks.
Pingback: String comparison in OpenInsight – Part 3 – Linguistic Mode | Building OpenInsight 10