FPBench is a standard benchmark suite for the floating point community. The benchmark suite contains a common format for floatingpoint computation and metadata and a common set of accuracy measures:
 The FPCore input format
 FPCore example inputs
 Metadata for FPCore benchmarks
 Standard measures of error
FPCore metadata properties
The metadata properties allow describing additional information about each benchmark, such as their name, description, precision, citations, or preconditions.
FPBench 1.2 defines the meaning of the following properties:
 :name
 A string name for the benchmark. If a name is needed and not provided, we recommend using the body expression.
 :description
 A string description for the benchmark and its inputs. We recommend describing the physical meaning and units of inputs when applicable.
 :cite
 A list of symbols that describe the sources of the benchmark. We recommend making available a BibTeX file which uses the same symbols as keys.
 :precision

Describes the floatingpoint precision used
for rounding, using a list
naming a precision family and its parameters, or using a
shorthand symbol
synonymous with such a list.
 (float e nbits)
 An IEEE754 binary floatingpoint value with e
exponent bits and nbits total bits.
binary32
andbinary64
are common shorthands, for(float 8 32)
and(float 11 64)
respectively.  (posit es nbits)
 Posits with es exponent bits and nbits total bits.
Common shorthands include
posit8
for(posit 0 8)
,posit16
for(posit 1 16)
,posit32
for(posit 2 32)
, andposit64
for(posit 3 64)
.  (fixed scale nbits)
 Fixed point, with nbits total bits and a scale factor of 2 ^ scale. One bit will store the sign, using two's complement. For example, (fixed 0 8) can represent the 8bit integers from 128 to 127, while (fixed 8 8) can represent magnitudes from 128/128 to 127/128 in steps of 1/128. Overflow behavior is determined by the overflow property.
 real
 A special precision for arbitrary real numbers, where no rounding takes place.
 integer
 A special precision for mathematical integers, where no rounding takes place. Implementations should assume that all values with this precision are indeed integers: if a noninteger value is computed, implementations can signal an error.
:precision
property is specified, implementations should assume a sane default, such asbinary64
or possiblyreal
depending on the application domain.  :round

The floatingpoint rounding mode used for rounding. Expected to be one of the IEEE 754 modes
nearestEven
,nearestAway
,toPositive
,toNegative
, ortoZero
.  :overflow

The rounding behavior to use when a fixedpoint value overflows.
Possible values include:
 infinity
 Produce a floatingpointlike infinity on overflow.
 clamp
 Clamp the value to the largest representable magnitude.
 wrap
 Wrap the significand using two's complement arithmetic.
 :pre

A precondition on inputs to the benchmark. Tools should not use
points that fail the precondition to determine the accuracy of the
benchmark. It is expected to be
an expression that
returns a boolean value. A rounding context can be given explicitly
for the expression, but if it is not, the default rounding context is assumed
to have
real
precision rather than the overall properties specified for the FPCore.  :spec

A realvalued function approximated by the benchmark. If not provided,
then it is assumed to be the body of the benchmark. As with preconditions
specified with
:pre
, a rounding context can be provided explicitly, but it is assumed to havereal
precision rather than the overall properties specified for the FPCore.  :mathlibrary

The library used in the source to implement floatingpoint
operations. It is expected to be a symbol identifying a library
implementation and version, such as
gnulibm2.34
.
Additionally, properties prefixed by the name of a tool (such
as :xxxfoobar
) are reserved for definition by that
tool. Only that tool may define the meaning of those properties;
other tools should not depend on the meaning of those
properties, nor must the defining tool keep the meaning constant.