Taken for Granted

Today (Tuesday February 24, 2009) I attended the North American SystemC Users Group meeting held in conjunction with DVCon 2009.   The agenda contained the normal OSCI update (from Ken Tallo of Intel, OSCI Chairman), and several presentations from vendors and users.   Many of the slides are available for download from the agenda site.

The first sign that attendance was going to be down on the 2008 DVCon meeting was the room chosen – less than 1/2 the size of the room last year.  I counted between 35 and 40 people in the room including organisers and presenters – again less than 1/2 the numbers that my (admittedly aged) memory recalled from 2008.   And there were perhaps more vendors/tool suppliers in the room, and more marketing people, than hard core users (although that was an impression only).

Some of the highlights from OSCI in Ken Tallo’s talk:

• TLM-2 LRM public review to start Q2 of 2009
• AMS draft 1 kit is in public review through March of 2009
• Synthesis working group should start public review of their v1.2 in Q2 of 2009
• A new working group, CCI:  Configuration, Control and Inspection, was started.  See the announcement.   It was stated that 38 individuals representing 24 organisations were involved.

The most interesting talk to me was by Herve Alexanian of Sonics on “OCP Sockets for TLM-2.0“.  The talk was created by Herve, Mark Burton and Robert Gunzel of Greensocs, and James Aldis of TI.  OCPIP has been one of the more advanced groups doing SystemC modelling, going back many years.  They are an interesting example of a complex interconnect protocol, modelled at various levesl, and migrating to use TLM-2.0 where possible as a base.   They have defined 5 abstraction levels from TL 0 (cycle accurate signals/pins = RTL level) through to TL 4 (Programmers View kind of level).  Some of these have OSCI near-direct equivalents, some do not, but they try to use whatever they can from TLM-2 as a base, through extending the generic payload and base protocol, or creating specific alternatives, for example.   TL 1 is a cycle accurate level but not modelling wires and signals (corresponding to my own definition of a CA transaction type model).    TL1 and TL4 look to me to be the two most useful modelling levels (at least from the perspective of an IP provider who currently provides equivalents to both of these).  The OCP release of their models is planned Q2 of 2009.

The OCP models add necessary capabilities (eg. to model bus locking, semaphores, etc., and the phases requied for cycle accurate modelling in TL1) and uses a socket model derived from the Greensocs Greensocket.

One conclusion that can be drawn from the OCP work is that TLM-2 is missing key items needed to model many real bus protocols (something we’ve known for a while).  And although it can be extended or used, such extensions, if not co-ordinated well by a consortium such as OCP-IP, will cause interoperability problems.   I think the OSCI TLM working group could find considerable value in using what OCPIP has done in developing a “TLM 3″, assuming that is on their roadmap, and including support for cycle-accurate TLM – but not being privy to what goes on in OSCI, I can only hope that the overlap in membership between OSCI TLM and OCP-IP is having such an influence.

On the other hand, as Herve mentioned, if you are not trying to model complex bus protocols at a cycle accurate level, but perhaps are only interested in virtual platform creation at the PV (TL4) kind of level or slightly below, TLM-2.0 may be enough.   Recent announcements from Open Virtual Platforms (OVP) indicates that they moved from their proprietary modelling protocol to one based on SystemC TLM-2.0 with only a small drop in performance (about 3%).

I will look forward to the release of the OCP work when it comes out.