Re: [oc] Beyond Transmeta...

["Jim Dempsey" <tapedisk@ameritech.net>]

 

----- Original Message -----

From: "Holger Baxmann" <holger@bitwind.org>

To: <cores@opencores.org>

Sent: Wednesday, February 12, 2003 4:23 PM

Subject: Re: [oc] Beyond Transmeta...

>
> But, in difference to 'sims': it is not necessary to keep the single
> identity for every bit while it is traveling the system, is it ?
>
> bax
>

The techniques for optimizing the performance of the system would

require the bits (or bit streams) to exist in multiple places at once

and in different time states. A simple illustration would be

 

    aaaa = bbbb + cccc

    bbbb = aaaa + cccc

 

Becomes: (set to display in courier)

 

   bbbb+

   cccc+

  aaaa+/

  cccc+

 bbbb/

 

Note that you cannot visualize bbbb as a whole entity at all points

in time. In time state 3 the lsb of bbbb transitions to new state

while the more significant bits of bbbb (of the first use) have

yet to be used. There undoubtedly will be cases where the bit

"consumption" rate of a bit stream longer than 1 clock tick. In this

case if the input variable is recomputed elsewhere then the

first use stream is conceptualized as producing a loop. This

is too hard to diagram here in text glyphs.

 

Conceptually, if a "function" block were to use a variable at a particular

time state it would always clock in one bit per clock. Internally it can

consume (use) the bits at a longer time interval (e.g. one bit every

other clock). You can view this as a serial FIFO where input and

output are at different clock frequencies.

 

The implementation would not use a FIFO inside each function block.

Instead. The bit stream variables, at any one time, can live at different

time states in multiple locations of the device. Example:

(set to display in courier)

 

  bbbbbbb

    b

    bbb

 

In the above if the bit stream of bbbbbbb were forked and if the leading

bit of the upper and lower streams were used at the same time in

the device then the lower use is as the state of bbbbbbb was two time

states earlier. If the loop is longer than the width of bbbbbbb then it

becomes possible to use an earlier state of bbbbbbb even after a

later state of bbbbbbb was completely modified. The delay loops

and/or ersatz FIFOs are an integral part of the dynamic wiring of the

device.

 

This is quite a departure from programming languages such as C++,

FORTRAN, etc. Initial uses of the device would make use of cross

compilers where you start your programming experience using something

familiar. Later on a new language would evolve beginning with some

obtuse precompiler statements/directives.

 

I envision that the ideal programming tool would be a visual tool that

resemble something like needlepoint that is "performed" on a time domain

scroll (player piano like). The width of the roll is respective of the size

of the device BSPLD or whatever acronym suits you best. If you use a

smaller device than optimally (performance wise) the length of the roll

extends (i.e. the computation takes longer). You likely would perform

the needlepointing on an abstraction pane and a separate pane would

be drawn to illustrate the utilization of the target device.

 

Jim Dempsey