Annotating Code
This guide gives an overview over the way functions and variables can be annotated in C-code. Consider this guide as a cheat sheet. For annotating functions click here and for variables, here.
This guide however does not provide all the details regarding annotations, for which please read the the Design Document
1. Function Approximation
Key/value pairs
| key | value range | short description | notes |
|---|---|---|---|
| name | Any string | this uniquely identifies this function | Will be set by clang, should not be set manually |
| numSegments | 0 to 2^segmentsBits-1 | This defines the maximum number of segments (primary + secondary) in which the function ought to be divided | - |
| numPrimarySegments | (0 to 2^segmentsBits-1) | This defines the maximum number of primary segments in which the function ought to be divided | - |
| segments | “uniform”, “log-left”, “log-right”, or “min-error” | This defines how the segments should be distributed in the function to be approximated | See table below for an overview over the segmentation strategies. |
| segments2 | “uniform”, “log-left”, “log-right”, or “min-error” | This defines how the subsegments (secondary segments) should be distributed in the function to be approximated | Experimental |
| explicitSegments | [min,max],(min,max),… | If the programmer knows the segments beforehand, he can explicitly state them as comma seperated value ranges | - |
| approximation | “linear”, or “step” | Describes which strategy should be used to approximate the function within each segment | - |
| bounds | (min, max) | If you know your the range of the input values to the function beforehand you can specify it here | - |
Segmentation Strategies:
| value | short description | notes |
|---|---|---|
| uniform | all segments are uniformly distributed | - |
| log-left | the size of segments increases by the power of 2 starting from the right | - |
| log-right | the size of segments increases by the power of 2 starting from the left | - |
| min-error | the segments are chosen by the minimizing the approximation error depending on the approximation strategy | You need to supply the key/value pair approximation |
Example Functions
2. Variable Approximation
Key/Value Pairs
| key | valid values | short description | notes |
|---|---|---|---|
| neglect_amount | 2,4,7,10,15,20,27 | Describes how many least significant bits should be neglected on operations with this variable | mutually exclusive with mask |
| mask | 0b1111110, 0b1111100, …, 0b0000000 | Similar to neglect_amount, but the neglected bits are specified as a mask | mutually exclusive with neglect_amount |
| inject | 0, 1 | A variable, that has inject=1 is marked as approximate for operations that use this variable as an input. | - |
| relax | 0,1 | A variable, that has relax=1 can be approximated on operation even when it’s inputs are seen as precise | Default=1 |