-
[FEATURE] opt = auto: automatically select bb or bb-eval.
-
[FEATURE] --print-second-order-errors: an option for printing second order errors.
-
The performance of Taylor forms generation for large expressions is improved.
-
Removed proof recording. Use FPTaylor v0.9.3 for recording proofs.
-
Minimal version of OCaml is increased to 4.03.
-
[FEATURE] --print-hex-floats: Printing important floating-point values as exact hexadecimal numbers.
-
[FEATURE] --export-options: Exporting all options into a given file.
-
[FEATURE] --print-precision: Error bounds are printed as correctly rounded decimal numbers with the given number of digits.
-
Moved
tools/plot.py
toscripts/plot/plot.py
. -
Added regression tests.
-
[FEATURE] bb-eval optimization: the same as bb but does not require an OCaml compiler to be installed.
-
ExprOut printer for JavaScript.
-
find-bounds = false by default.
-
bb and bb-eval: evaluate common subexpressions only once.
-
JavaScript version of FPTaylor and FPTaylor's website.
-
Maxima: the simplificaiton function is
ratexpand
now. -
FPCore export: context annotations for mixed precision benchmarks.
-
export: A separate tool for exporting FPCore benchmarks.
-
[FEATURE] ULP error (experimental; almost always suboptimal near powers of 2).
-
[FEATURE, can break existing FPTaylor input files] The rounding operation
rnd
without parameters can be used in input files. The rounding type is defined by thedefault-rnd
option (the default value isrnd64
). Parameters of rounding operations must be given in square brackets. (Parameters were in parentheses before.) -
[FEATURE, can break existing FPTaylor input files] Variables without explicit types are assigned the default type defined by the
default-var-type
option (the default value isfloat64
). (Variables without explicit types were assumed to be of thereal
type before.)
-
[FEATURE]
tools/plot.py
: a script for plotting 1d error models. -
[FEATURE] Exporting ErrorBounds input files.
-
[FEATURE] Exporting to the FPCore format.
-
[FEATURE] Several tasks in one files (inside curly braces).
-
[FIX] Correct rounding of constants which are representable as
float64
. It was possible to get incorrect results for rounding modes different fromrnd64
. For instance, the result ofa = rnd32(1p+50 + 1)
waserr = 0
.