Skipping the preflight check

A project log for Libre Gates

A Libre VHDL framework for the static and dynamic analysis of mapped, gate-level circuits for FPGA and ASIC. Design For Test or nothing !

Yann Guidon / YGDESYann Guidon / YGDES 12/07/2020 at 23:170 Comments

The last log Plot twist explains the "shadowing flaw" in the new net probing algorithm. A solution was proposed and this log tries to get the details straight.

Inspired from the CORS nomenclature, the previous solution runs a first round of checks before starting the main loop, so it's called the "preflight check".

Instead of serialising the driver's identifier, the signal sends a value that encodes the type of the driver, so 'U' would naturally mean "unconnected". This then helps the next phase to send data when partial inputs are available.

A new code path must be added, by using a special value of 0 for the shared variable multiply_pass.

Fine so far but...

Why did I go from "OR" to "AND" logic to propagate the update event ? It simplifies the logic because I don't have to check if the output value has already been sent.

This is because now, setting the output also modifies the "decumulator" and it would be out of sync if more than one event triggers it.

In fact the "preflight check" can be skipped if each input and output has a flag to indicate that the corresponding signal has been processed (accumulated or decumulated). I know it would take a bit more room but... It's only 5 bits for the LUT4 and the previously proposed solution would use more. The output port is scanned manually at the end of each cycle so it's not affected, unlike when a preflight check is used.

There are two new challenges to solve : how/where to solve the flags, and when/how to toggle them.

That should work...




AAaaaannnd... It works ! Amazingly with pretty few hicups and few compile errors on the first try. Go get it : LibreGates_20201209.tbz !

The probe is enabled for 4 out of the 6 tests and even spotted a flaw in the BigLFSR one (which I will leave for the lulz because it doesn't really matter). I should benchmark its performance now. The next step will be to reimplement the depthlist system.