SIMD not present in Libre/Open hardware OpenPOWER implementations [was Re: enabling/disabling AltiVec in Firefox and derived browsers (ArcticFox)]
changing subject, for reference / background:
* Paul Mackerras is working on an experimental branch to add VSX
he was experimenting to see what was needed to get Fedora booting. the internal design is a Finite State Machine. multiple clocks per instruction (due to internal 64 bit pathways)
* neither A2I nor A2O have VSX and an estimate for adding it to these high-performance gate-level designs would be around 2 years
* LibreSOC we are just not going to add VSX. the development cost is far too high, the performance nowhere near that of Vectors, software complexity far too high and L1 cache usage is compromised.
all of these designs - all four - have internal 64 bit pathways. OpenPOWER instruction decoding is complex even without SIMD (4,000 gates) and adding VSX multiplies that by three or four. that's enough gates to do a decent embedded RISC core in any other RISC ISA.
IBM had years in which to incrementally extend SIMD operations. Jeffrey in another post kindly outlines the progression.
now, at POWER10 with 18 billion transistors, the barrier to entry is so high that if someone doesn't put their foot down and say "no" to SIMD there isn't going to *be* any new OpenPOWER hardware other than that from IBM [not, at least, capable of running standard ppc64le distros that is]
we seriously considered doing an entirely new ppc64le-eabi-1.5/1.9 Debian Distro port at one point, going *backwards* to the time when SIMD was not mandatory but doing LE rather than BE, but the risk of it being viewed in the same way as "rasbian" is too great.
On Tuesday, March 2, 2021, Riccardo Mottola <email@example.com> wrote:
> Emulation at kernel level is painfully slow,
seriously, i kid you not, it is infinitely better than trying to implement VSX in hardware. we would spend so long implementing it that it would delay LibreSOC *beyond* the point where money from NLnet was available, jeapordising the entire project in the process.
given a choice between "painfully slow right now but fixable in software later" and "completely destroying any chance of completing and delivering even any hardware at all" it's hardly a choice :)
the Cray-style Vectorisation being added will smoke SIMD in the long run, once the ABIs and compilers are sorted.
> yes enabling runtime libraries could be done, requires extensive work in
> upstream code.
this is a better situation than an entire new distro port. we may have to have one anyway: all timescale estimates which start from defining a new triplet and going from there are around 3-5 years.
if a new EABI has to be defined and spec'd as well it's even longer.
> An easier version is the path that TenFourFox and other follow: just
> provide two binaries, which is what I intend to do with ArcticFox.
> However if Debian wants to come up with the pain of two (or more?) FF
deep breath, this may be a sane medium term solution. long term the separation of SIMD is needed behind dynamic loadable libraries (and HWCAPS in glibc6) rather than assuming it is 100% guaranteed available.
LibreSOC in particular needs to appear to go *backwards* in terms of performance before it can go forwards, once the Cray-style Vectorisation hits gcc properly.
then other hardware can also do variants of the same libraries (including POWER9/10).
On Tuesday, March 2, 2021, Jeffrey Walton <firstname.lastname@example.org> wrote:
> Based on my experience with Botan and Crypto++... VSX is available
> with POWER7 and -mvsx compiler option. VSX is part of POWER8 core and
> does not need a compiler option.
as demonstrated by A2O/I in particular there is unfortunately a problem with referring to IBM proprietary processor names as the canonical definition of available features: A2I/O are Power v2.06/7 compliant but *still do not have VSX*.
this is because the feature is optional except by the time you get to the AIX Compliancy Subset. see v3.0C or v3.1 first few pages, copy easily available at http://ftp.libre-soc.org
note that it really does say "SIMD is optional" for Linux/UNIX subset, Floating Point Embedded subset and Fixed Point subset.
many people misinterpret / misread that document including myself for several months.
this conflation is caused by the fact that only IBM processors, which happen to go by proprietary names POWER7-10, are commonly available. NXP Quorl, not so well known, which is v2.08B compliant, used in the PowerPC Notebook, is going EOL.
> VSX is a lot like Intel tic/toc features. VSX allows a 64-bit vector
> loads and stores, but it does not provide operations on 64-bit
> vectors. You have to use POWER8 to get the 64-bit add (addudm),
> subtract (subudm), etc.
this illustrates very nicely the progression over time (many years) as the team inside IBM ramped up the capabilities.
we can see very unfortunately that they too were seduced by what SIMD says it can do. as they progressed far beyond what other OpenPOWER Foundation members were able to handle (NXP Quorl for example: NXP has made it clear they have no intention of implementing v3.0) IBM became the only implementor.
this is a very delicate situation with not very good options. ideas appreciated.
crowd-funded eco-conscious hardware: https://www.crowdsupply.com/eoma68