This folder contains scripts and files to assess the quality of aspects of the AVR-LibC implementation.
The accuracy of the floating-point implementation can be assessed with
the math-prec.sh bash script in this folder.
-
An avr-gcc installation.
-
The avrtest AVR core simulator. avrtest uses the host's native floating-point implementation to provide syscalls for IEEE-754 single and double emulation on the target. This implies the requirement that the host floating-point format matches the target format: litte endian, 8-bit exponent for 32-bit values, and 11-bit exponent for the 64-bit values. Otherwise, avrtest will terminate with an error message when a floating-point syscall is invoked.
-
The Gnuplot graphing utility. When Gnuplot is not available, the
math-prec.shscript can still be used with option-P(don't generate plots).
A simple way to run the script is to
$ cd <sources>/tests/simulate/quality
$ ./math-prec.sh -i
which uses the AVR-LibC from an [i]nstalled avr-gcc toolchain.
It reads function specifications from funcs.txt, and the
console output will be something like:
== quoted lines are: <x> <abs-err> <rel-err> <rel-err-bits> # <x-hex> <cycles>
== writing out-prec/math-prec.html...
== sinf [-0.5, 6.5] == single ========================================
== calc sinf data to: out-prec/e-sinf.data
== sinf min rel error: 2.12777781 -1.1921e-07 -1.4044e-07 -21.76 # 0x1.105b06p1 1858
== sinf max rel error: 6.28333330 2.3283e-10 1.5732e-06 -18.28 # 0x1.922222p2 2039
== sinf min abs error: 1.57222223 -1.1921e-07 -1.1921e-07 -22.00 # 0x1.927d28p0 1880
== sinf max abs error: 4.60555553 1.1921e-07 1.1989e-07 -21.99 # 0x1.26c16cp2 1739
== sinf max cycles: 6.28333330 2.3283e-10 1.5732e-06 -18.28 # 0x1.922222p2 2039
== sinf mean cycles: 1757.7
== plot sinf abs error to: out-prec/e-sinf-abs.png
== plot sinf rel error to: out-prec/e-sinf-rel.png
== plot sinf rel error in bits to: out-prec/e-sinf-bit.png
...
