The examples in this directory show how BSV is used to implement an
IEEE-754 floating point operations.   Use of parameterized
functions and type inference is used extensively to maximize code
reuse.  The examples are an adder, multiplier, and mac (multiply and
accumulate) unit.  These examples are available in both IEEE 32 and
IEEE 64 bit formats, where the addition of the 64-bit format involved
the addition of a 10 line function.



- FPLibrary.bsv:  Contains common functions which can become the
building block of a floating point library.  Examples of functions
included here conversion to and from IEEE format and an internal
format, normalization, rounding, finding leading bit.

- FPAdd.bsv:  Contains the add specific functions, and bsv code which
creates the FPAdd and FPAdd64 modules.

- FPMult.bsv:   Contains the multiply specific functions, and bsv code
which creates the FPMult and FPMult64 modules.

- FPMac.bsv:   Contains the multiple and accumulate code for 32 and 64
bit versions.  Each of these function is 15 lines of code.


- TesterLib.bsv:  Contains interface definitions and common module
wrappers for the binary (add and multiple) and ternary (mac) functions.

- Tester.bsv:  Contains testbench code which 

- dumper.C:   Is a c file which is used to generate test data for the
add, multiple and mac modules.


Extensions:
There are many directions which to take this code, specifically, the
addition of other functions, such as abs, comparisons, conversions,
alternate rounding modes, optimizing by combining normalizing and
rounding, etc. 

It may be interesting to build, from these primitives, a module which
implements (A*B + C*D) where the addition occurs in extended
precision, before the final rounding and normalization.

There may be other optimizations and supporting primitives which can
be implemented based on the higher-level function.  For example, the
normalize, round and normalize1 functions can be combined during
additions, since at most one normalization will be required.

The current code has not be extensively tested and may not correctly
flag overflow and underflow conditions.  The same holds for
denormalized IEEE representations.