Release 0.6.0 "Christmas": Introducing variable precision

[c3d]

This release was a bit longer in coming than earlier ones, because we are about
to reach the limits of what can fit on a DM42. This release uses 711228 bytes
out of the 716800 (99.2%).

Without the Intel Decimal Library code, we use only 282980 bytes. This means
that the Intel Decimal Library code uses 60.2% of the total code space. Being
able to move further requires a rather radical rethinking of the project, where
we replace the Intel Decimal Library with size-optimized decimal code.

As a result, release 0.6.0 introduces a new table-free and variable-precision
implementation of decimal computations. In this release, most operations are
implemented, but some features are still missing (e.g. Gamma function). This
release will be simultaneous with 0.5.2, which is functionally equivalent but
still uses the Intel Decimal library. The new implementation is much more
compact, allowing us to return to normal optimizations for the DM42 and regain
some of the lost performance. On the other hand, having to switch to a table
free implementation means that it's significantly slower than the Intel Decimal
Library. The upside of course is that you can compute with decimal numbers that
have up to 9999 digits, and a decimal exponent that can be up to 2^60
(1 152 921 504 606 846 976).

New features

Variable precision decimal floating point implementation for arithmetic,
trigonometrics, logs, exponential and integer factorial. Other functions may
still return "unimplemented error".

Bug fixes

None. If anything, this release introduces bugs in computations and performance
regressions. However, it frees a lot of space for further DM42 development.

Improvements

The Precision setting now sets the number of digits with a granularity of one,
between 3 and 9999. Ideal use of memory is with multiples of 12 digits, e.g. 12,
24 or 36 digits, where decimal packing does not cause lost bits.

Performance on the DM42 is somewhat improved, since it is now possible to return
to a higher level of optimization.

Regressions

In addition to lower performance and unimplemented functions, this version no
longer builds a compatible QSPI. This means that returning to the DM42 requires
flashing both the QSPI and the PGM file.

Full Changelog: v0.5.2...v0.6.0

---

This discussion was created from the release Release 0.6.0 "Christmas": Introducing variable precision.