Google Groups no longer supports new Usenet posts or subscriptions. Historical content remains viewable.
Dismiss

Libre-SOC OpenPOWER Project

77 views
Skip to first unread message

lkcl

unread,
Apr 14, 2021, 6:10:02 PM4/14/21
to
hi folks i thought it might be time to let everyone know of the plans for Libre-SOC.

it is a new SoC based on OpenPOWER, with a roadmap to ramp up from single core 300mhz (as an entirely Libre ASIC, using opencores PHYs, coriolis2 for P&R and Chips4Makers Cell Libraries), to quad core within 2 years, then 8 core and on to a 64 monster in 6 years time.

we aim to extend and adapt the OpenPOWER ISA, given that there now exists a way to propose extensions.  this will include full Cray-style Variable-length Vectorisation (VSX is 15 year old SIMD) and adding 3D and Video opcodes directly to OpenPOWER (*NOT* as a separate core), such that Vulkan Shader binaries may be JIT translated then directly executed natively.

the kicker: after reading the v3.0B ISA manual and seeing that of its 1300 pages, fully 50% was dedicated to VSX, we simply flatly refused point blank to consider implementing it.  this decision was reinforced after seeing that in binutils ppc port, a staggering 4500 assembly mnemonics are listed, 90% of which are VSX.

purely from a logistics perspective it would be insane to attempt to implement them (minimum 3 man-years).  there are so many opcodes that the IBM POWER9 implementation actually has to have a 2 stage instruction decoder.  just the integer opcodes (only 70 of them) requires 4,000 gates: imagine how many would be needed for VSX.

what we intend to do instead is to add emulation in the linux kernel (extending lib/ppc/sstep.c) as a stopgap measure.  however we will not be implementing the VSX instructions by hand: we will be auto-extracting the pseudocode from the v3.0B specification, and compiling it to c.

this is already how our python-based OpenPOWER ISA emulator works: the pseudocode is converted to python.  outputting c instead of python is a trivial next evolutionary step. 

that leaves a few years breathing space in which to define a new EABI which does not have VSX as mandatory.  the mistake made a few years ago of adding VSX/VMX as mandatory, on the basis that it improved performance [for IBM POWER9] is a costly one.  Microwatt, A2O, A2I and LibreSOC - none of these open cores have VSX or VMX, and consequently cannot run debian, fedora, arch, or in fact any modern mainline GNU/Linux distro at all.

this is a massive project, for which we have grants from NLnet to help pay for it, if anyone would like to help.

main site link: http://libre-soc.org

l.

0 new messages