Hi,
This is to announce that I've been working on a rewrite of Guile's array
system, entirely in Scheme.
https://notabug.org/lloda/guile-newra
https://github.com/lloda/guile-newra
Newra works on Guile 2.2, but it's quite a bit faster on Guile 2.9.
There are several reasons for this rewrite. A major one is that the current
array system forces constant switches between C and Scheme and is opaque to the
Scheme compiler. As the Scheme compiler improves this becomes more noticeable.
The work is at a stage where all of the functionality of the old system is
available in the new one, except for a few items that require patches to Guile
itself (e.g. compiler support for literals, the C interface). These are listed
at the end of the README. On the other hand the new system has a few extra
whistles, such as basic rank extension / broadcasting and lazy index vectors.
This is still not a full featured array system, but it's a superset of what
Guile offers now and hopefully it's a better base to build on.
There is a test suite and a benchmark. The documentation consists of the
README, plus docstrings. The output of the benchmark on two different machines:
debian i7-4930K - https://paste.debian.net/1094626/
mac i5-8500B - https://paste.debian.net/1094627/
On Guile 2.9 some things are appreciably faster already, others are about the
same. The only places where the C is still out of reach is where there is a
fast path, such as array-copy! between #t types, which ends in a memcpy() in C.
--
The reason I announce this now is that I'm at a point where some design
decisions are needed — the old API is inconvenient and we should have something
that looks closer to they typical array API. I also could use some help with
optimization. Of course a review or any other kind of feedback would be welcome
as well.
To be clear, it is my intention, in due course, to propose Newra for inclusion
in Guile.
Regards
- Daniel
