In Katabatic & Kite, remove all hard-coded values related to track pitches.
* New: In <Session>, add more convenience function to access RoutingGauge
characteristics.
* New: In <AutoSegment>, <AutoContact>, add support for the "depth spin",
that is, if the source/target contacts are going "top" or "down".
Used to compute the perpandicular pitch. Need a small modification
of the revalidation mechanism. The observers of <AutoSegment> are
notified when the spin changes.
* New: In <AutoSegment>, the getPPitch() method allow to compute the
"perpandicular pitch". For now it is simply the greatest from the
source perpandicular pitch and the target perpandicular pitch.
Make uses of the "depth spin".
* New: In <TrackElement>, <TrackSegment>, cache the perpandicular pitch.
Updated through the notification from the observable.
- New: In NegociateWindow/RoutingEvent, adds a more comprehensive stage
"Repair". Perform in three stage: first try to place with a relaxed
constraint (one GCell on each side). Second try to minimize the faulty
segment. Third perform another "repack perpandicular" but this time
the faulty segment is re-inserted *before* any of it's perpandiculars.
- New: In RoutingEvent::cacheAxisHint(), when a segment has a parent, that
is comes for a "moveUp()", uses the parent axis hint as it's own.
- New: In State::slackenTopology(), in the global FSM, adds a special
operation when reaching MaximumSlack: forceOverLocals(), try to insert
the global on track containing only local segments. Should tend to
concentrate locals on a small set of shared tracks. Most useful on the
highest layers.
- New: In State::slackenTopology(), in the "MoveUp" state, try to find the
more appropriate segment to move up (Manipulator::desaturate()).
Effectively move up the longest segment fully enclosing the one we are
processing.
- New: In State::slackenTopology(), add a check for fully blocked segments
in the local segment FSM. Calls State::solveFullBlocked().
- New: In KiteEngine::createGlobalGraph(), decrease the vertical capacity
of one track inside the core. Helps smooth the vertical density.
- Change: In Manipulator::insertInTrack(), when a track is freed for a
to be inserted changes the priorities so that the segment is immediatly
inserted. Parallels ripeds and theirs perpandiculars are replaced
only *after*. This is the opposite of the previous behavior.
- Change: In NegociateWindow::NegociateOverlapCost(), account the costs
of terminals only for deep depth layers (M1, M2 & M3).
- Change: In RoutingEvent::insertInTrack(), expand the excluded interval
by a half-pitch (2.5l) instead of one lambda.
- Change: In State::State(), do not uses DiscardGlobal if the ripup count
exceed 5. Case of the "Strap" segments that can be ripped a lot
before changing state.
- Change: In State::_processNegociate(), no longer lock into position
(fixed) the local terminal segments as a last resort.
- Change: In RoutingEvent::_processNegociate(), no longer ripup perpandiculars
when a segment is inserted in a free space. Reduce the number of
events whithout degrading the routing quality.
- Change: In State::conflictSolve1_v1b(), if getLongestConflict() is nul,
ignore the track, the conflict must occurs on another track.
- Change: In TrackCost, add a flag support. First uses, a flags to prevent
a local of the topmost layer to ripup a global which is in moveUp
state.
- Bug: In State::solveFullBlockage(), after have been freed, reset the
segment state to "moveUp".
- Bug: In manipulator::minimize(), the axisHint was miscalculated if the
punctual span was empty.
- New: In RoutingEvent, while routing a full chip, transient full blockages
may happens due to the initial position of some perpandiculars on theirs
optimal positions. Check for it and perform a "ripupPerpandicular" with
the perpandiculars re-routed first (on normal operation, this is the
reverse).
- Change: In NegociateWindow::NegociateOverlapCost(), do not account terminals
if the layer is *above* METAL3 as it is unlikely that it's truly directly
connected to a terminal (true at least for our designs).
- Change: In RoutingEvent::State, add a "fullBlocked" flag to perform the
full blockage direction while building the state object.
- Change: In RoutingEvent::cacheAxisHint(), when the TrackSegment has parent,
uses the parent current axis position and not it's axis hint.
- Change: In Manipulator::insertInTrack(), when the overlapped segment is a
Local and is completly enclosed (shrinkLeft & shrinkRight), no longer
rip it up but force a shrink left/right instead.
- Bug: In RoutingEvent, the event sorting function was sorting in the *wrong*
order! The less prioritary first! With the correct sort, we won an
additionnal 30% in speed (total: 69%). From the "reference" time we have
a 3.2 speed-up. And we can successfully process denser designs...
- New: In BuildPowerRails, special processing for the power ring segments.
The "diagonal" of vias at each corner is causing a misbehavior of the
routing algorithm (due to fully saturated GCells in one direction).
As a temporary fix, extend the segments so they form a "square corner".
(problem arise on "d_in_i(22)").
- New: In RoutingEvent::_processNegociate, disable the "isForcedToHint()"
feature. No noticeable loss of quality or speed.
- New: In TrackElement/TrackSegment, wraps the AutoSegment parent's mechanism.
Allows to gets the DataNegociate of either the segment or it's parent.
- New: State::solveFullBlockages(), dedicated method to solves the case when
all the allowed tracks of a segment are blocked, tries to moves up
local segments and to break-up global ones.
- New: RoutingEventLoop, a more sophisticated way to detect looping.
Maintain a dynamic histogram of the last N (default 10) segments routeds,
with the count of how many times they have occurred. If that count
exeed 40, we *may* be facing a loop.
- Change: In State::conflictSolve1, implement new policy. The global segments
no more can be broken by local ones. The idea behind is that breaking
a global on the request of a local will only produce more cluttering
in the GCell. Globals must be keep straigth pass through, especially
inside near-saturated GCells. Globals breaking however can occurs at
another global's request.
- Change: In TrackCost, implement the new policy about locals segments that
cannot break globals segments. The sorting class now accept flags to
modulate the sorting function. Two options avalaibles: IgnoreAxisWeigth
(to uses for strap segments) and DiscardGlobals (to uses with locals).
- Change: In TrackCost, the "distance to fixed" have now an upper bound of
50 lambdas (no need to be greater because it means it's outside the
begin & en GCells). Take account not only of fixed segment, but also
of placed segments which makes bound.
- Bug: In Track::_check(), while calling each individual TrackSegment check,
uses it as the *first* argument of the "or", otherwise it may not be
called.
- Bug: In ProtectRoutingPad, loop over segment Collections while modificating
it was producing non-deterministic results. The fact that a collection
must be not modificated while beeing iterated is becoming a more and more
painful problem.
- Change: In TrackCost, the ripup cost is now computed from a mix between
the number of ripup *and* the data state. Restore a correct balance
of the ripping up between segments.
- New: In RoutingEvent::State::conflictSolve1(), when a very long Segment
enters in conflictSolve1(), still prefers to move it up in one piece,
but if it's not possible, it is unlikely that it will overlap with
only one or two other Segment. In that cases, blindy break it up in
the middle.
- New: In RoutingEvent::Manipulator, another overload of makeDogLeg to
allow break up at one precise point.
- Change/Bug: In RoutingEvent::Manipulator::Relax(), reset the state of
the non-overlapping Segment fragments instead of inheriting blindly
of the state of the breaked up one. Gives some supplemental slack
that allows to converge.
- Note: Now the strategy is to disable the uses of GCellRoutingSet by
setting the expand step to 0.99. This induces only a slight slow
dow and memory increase and passes all the tests...
- Mark: This revision passes all the benches except for idct, vld &
ieee_division.
./coriolis/src/crlcore,
./coriolis/src/knik,
./coriolis/src/katabatic,
./coriolis/src/kite,
./coriolis/src/equinox,
./coriolis/src/solstice,
./coriolis/src/ispd:
- SVN MOVE: Source tree simplification & uniformisation. Now all tools
are at the same level, directly under the root of the repository.
No more "coriolis/src".
Jean-Paul Chaput