This commit degrades the run success rate of ARMv2a to 87% (40 iters).
* New: In CRLcore/etc/.../kite.conf, add configuration parameters:
katana.termSatReservedlocal
katana.termSatthreshold
for the new edge capacity computation system.
* New: In CRLcore/etc/symbolic/phenitec06/, add support for N. Shimizu
small I/O pads (supplied in phlib80). Tune various parameters of
Anabatic/Katana to increase routing success.
* Change: In CRLcore/alliance/ap/ApParser, make Pin external components,
so RoutingPad will be build upon in global routing.
Do not complain when a I/O pad has a physical instance that did
not exists in the netlist. Just create it (appeared in phlib80).
When no netlist instance exists in a pad, the pad Cell is still
considered as terminal.
* New: In Etesian::BloatCells, new profile named "3metals" better suited
for two routing metals technologies (i.e. Phenitec).
* New: In Anabatic::RawGCellsUnder, new CTOR which take only source &
target points instead of a segment. Needed to manage wide segment for
which the axis to consider is not that of the segment (one axis for
each track it intersect).
* New: In Anabatic::GCell, add a RoutingPad count attribute, for Edge
reservation computation.
* New: In AnabaticEngine::computeEdgeCapacities(), instead of decreasing
all edges of a fixed amount (hTrackReservedLocal), guess the GCell
cluttering from the number of RoutingPads that it contains.
For non-saturated GCells, the four edges are decreased by the number
of RoutingPads. We use the maximum from the two neigboring GCells.
The hTrackReservedLocal parameter is now used only as a *maximum*
that the edge reservation can reach.
If a GCell is saturated (more than 8 RoutingPads, the saturation is
propagated horizontally to 2 neigboring GCells).
* Change: In AutoContactTerminal::getNativeConstraintBox(), use a more
flexible gauge name matching for terminal vertical extensions correction.
Namely, match all "msxlib*" kind of gauges.
* Change: In AutoSegment::setAxis(), add the ability to force the axis
position, even if it is a non-canonical segment. Maybe needed in the
initialisation steo, before the first canonisation is performed.
* New: In NetBuilder, added new methods _do_1G_1PinM1() and _do_2G_1PinM1(),
to manage coronas for Phenitec designs.
To avoid various side effects from segments being too close from
the north / east side of the routing area, make those segments fixeds.
* Change: In KatanaEngine::annotateGlobalGraph(), the management of wide
wires was wrong. The axis to use to find the underlying GCells is the
one of the track, not of the segment. This was creating bad edge
capacity computation under the power ring of a block and subsequently
routing failures.
* New: In Kanata::Manipulator, added method reprocessParallels(), not used
though, but keep it anyway, might be of use later...
* New: In Kanata::Manipulator, added method avoidBlockage() for terminal
METAL2 in non-preferred direction, restrict the terminal and turn
constraint box at the current position of the perpandicular, so it
doesn't create a deadlock in METAL2.
* Change: In SegmentFsm::conflictSolveByPlaceds(), if we cannot break
using the whole overlap, try the first atomic overlap.
* New: In SegmentFsm::_slackenStrap(), manage conflict between a non-prefered
segment and a blockage, this when to call avoidBlockage()...
* New: In Katana::Configuration, management of the new edge computation
parameters:
katana.termSatReservedlocal
katana.termSatthreshold
* New: In Cumulus/plugins/Core2Chip, support for Phenitec I/O pads.
This non-deterministic behavior was showing only in the ARMv2a benchmark
around event 180k...
* Bug: In Anabatic::Session::_netInvalidateds & _netRevalidateds, the
set<> was still sorted on pointers. As contacts & segments can be
created to maintain connexity after a layer change, we got a
discrepency in objects Ids that may generate a change in ordering
later.
* Bug: In Katana::Session::_doRemovalEvents(), the set of Tracks that
got elements deleted was still using pointers. Now we use a TrackSet
sorted on (direction,depth,axis). This should not have created
a change in the results, as destructions do not change Ids, but it
genereate extra differences in traces.
* Change: In Hurricane::SharedName, replace the incremental Id by a hash key.
This is to ensure better deterministic properties. Between use cases,
additional strings may have to be allocated, shitfing the ids. Even if
hash can be duplicated, we should be able to ensure that the absolute
order in map table should be preserved. Supplemental strings are inserted
in a way that keep the previous order.
* Change: In CRL/etc/symbolic/cmos/kite.conf, add "katabatic.routingGauge"
default parameter value ("sxlib").
* Change: In CRL/etc/common/technology.conf, define minimal spacing for
symbolic layers too (added for METAL4 only for now).
* Change: In CRL::Histogram, extend support to dynamically sized histograms.
Add a text pretty print with table and pseudo-curve.
* Change: In Cumulus/plugins/ClockTreePlugin, create blockage under the
block corona corners so the global router do not draw wire under them.
This was creating deadlock for the detailed router.
When the abutment has to be computed, directly use Etesian to do it
instead of duplicating the computation in the Python plugin.
* New: In Etesian, as Coloquinte seems reluctant to evenly spread the
standard cells, we trick it by making them bigger during the placement
stage. Furthermore, we do not not uniformely increase the size of the
cells but create a "bloating profile" based on cell size, cell name
or it's density of terminals. Currently only two profiles are defined,
"disabled" which does nothing and "nsxlib" targeted on 4 metal layer
technologies (aka AMS 350nm, c35b4).
* Bug: In Knik::MatrixVertex, load the default routing gauge using the
configuration parameter "katabatic.routingGauge" as the default one
may not be the first registered one.
* New: In AnabaticEngine::setupNetDatas(), build a dynamic historgram of
the nets terminal numbers.
* Bug: In Anabatic::AutoContact::Invalidate(), always invalidate the
contact cache when topology is invalidated. In case of multiple
invalidations, if the first did not invalidate the cache, later one
that may need it where not allowed to do so. The end result was correct
nonetheless, but it did generate annoying error messages.
* Bug: In Anabatic::AutoContactTurn::updateTopology(), bad computation
of the contact's depth when delta == 2.
* Bug: In Anabatic::Gcell::getCapacity(), was always returning the west
edge capacity, even for the westermost GCell, should be the east
edge in that case.
* New: In Anabatic::AutoSegment, introduce a new measure "distance to
terminal". This is the minimal number of segments separating the
current one from the nearest RoutingPad. This replace the previous
"strong terminal" and "weak terminal" flags.
This distance is used by Katana to sort the events, we route the
segments *from* the RoutingPads *outward*. The idea being that if we
cannot event connect to the RoutingPad, there is no points continuing
as thoses segments are the more constraineds. This gives an order close
to the simple ascending metals but with better results.
* New: In Anabatic::AutoSegment, introduce a new flag "Unbreakable", disable
dogleg making on those segments. mainly intended for local segments
directly connecteds to RoutingPads (distance == 0).
* New: In Anabatic::AutoSegment, more aggressive reducing of segments.
Now the only case where a segment cannot be reduced is when it is
one horizontal branch in a HTee or a vertical on a VTee. Check if,
when not accounted the source & target VIAs are still connex, if so,
allow reducing.
* New: In Anabatic::AutoContact, new state flags CntVDogleg & CntHDogleg
mainly to prevent making doglegs twice on a turn contact. This is to
limit over-fragmentation. If one dogleg doesn't solve the problem,
making a second one will make things worse only...
* Bug: In Anabatic::Configuration::selectRpcomponent(), we were choosing
the component with the *smallest* span instead of the *bigger* one.
* New: In Anabatic::GCell, introduce a new flag "GoStraight" to tell that
no turn go be made inside those GCells. Mainly used underneath a block
corona.
* New: In AnabaticEngine::layerAssign(), new GCellRps & RpsInRow to manage
GCells with too many terminals. Slacken at least one RoutingPad access
when there is more than 8 RoutingPad in the GCell (slacken or change
a vertical METAL2 (non-preferred) into a METAL3).
* Change: In Anabatic::NetBuilderHV, allow the use of terminal connection
in non-preferred direction. That is, vertical METAL2 directly connected
to the RoutingPad (then a horizontal METAL2). This alllows for short
dogleg without clutering the METAL3 layer (critical for AMS c35b4).
Done in NetBuilderHV::doRp_Access(), with a new UseNonPref flag.
Perform some other tweaking on METAL1 access topologies, to also
minimize METAL3 use.
* New: In AnabaticEngine::computeNetConstraints(), also compute the
distance to RoutingPad for segments. Set the Unbreakable flag, based
on the distance and segment length (local, short global or long global).
New local function "propagateDistanceFromRp()".
* Change: In AnabaticEngine.h, the sorting class for NetData, SparsityOrder,
is modificated so net with a degree superior to 10 are sorted first,
whatever their sparsity. This is to work in tandem with GlobalRouting.
* New: In Katana::TrackSegmentNonPref, introduce a class to manage segment
in non-preferred routing direction. Mostly intended for small METAL2
vertical directly connected to RoutingPad. Modifications to manage
this new variant all through Katana.
* Change: In Katana::GlobalRoute, DigitalDistance honor the GoStraight flag
of the GCell. Do not make bend inside thoses GCells.
* Change: In KatanaEngine::runGlobalRouter(), high degree nets (>= 10) are
routed first and whitout the global routing estimation. There should be
few of them so they wont create saturations and we want them as straight
as possible. Detour are for long be-points.
Set the saerch halo to one GCell in the initial routing stage (before
ripup).
* Bug: In KatanaEngine & NegociateWindow, call _computeCagedconstraints()
inside NegociateWindow::run(), as segments are inserted into tracks
only at that point so we cannot make the computation earlier.
* Change: In Katana::Manipulator::repackPerpandiculars(), add a flag to
select whether to replace the perpandiculars *after* or *before* the
current segment.
* Change: In Katana::NegociateWindow::NegociateOverlapCost(), when the
segment is fully enclosed inside a global, the longest overlap cost
is set to the shortest global hoverhang (before or after).
When the cost is for a global, set an infinite cost if the overlapping
segment has a RP distance less or equal to 1 (this is an access segment).
* Bug: In Katana::PowerRailsPlane::Rail::doLayout(), correct computation of
the segments extension cap.
* New: In Katana::QueryPowerRails::addToPowerRail(), add support for Pad.
* Change: In Katana/PreProcess::protectCagedTerminals(), apply the contraints
to any turn connected to the first segment of the RoutingPad so the
perpandicular constraints got propagated to the perpandicular segment...
* Change: In RoutingEvent, cache the "distance to RP" value.
* Change: In RoutingEvent::Key::compare(), sort *first* on distance to
RoutingPad, then layer depth. If both distance to RoutingPad is null,
then sort on segment length.
* Change: In RoutingEvent::_processRepair(), try a repack perpandicular with
perpandiculars first (then with perpandicular last, then give up).
* Change: In SegmentFsm::bindToTrack() and moveToTrack(), set an axis hint
when creating the insertion event.
* Change: In SegmentFsm::_slackenStrap(), add a step through slacken between
minimize and maximum slack (wihch directly end up in unimplemented).
* Change: In Session::_addInsertEvent(), add an axis parameter needed when
the axis of the segment is not the one of the track (case of wide
segments or non-preferred direction).
* Bug: In Track::_preDestroy(), bad management of the TrackElement reference
count. Destroy the segment only when reaching zero...
* Bug: In Track::expandFreeIneterval(), forgotten to manage case when there
is a set of overlaping segments at the "end" of the track, the
EndIsTrackMax was not set.
* Change: In TrackCost::Compare, increase the cost when an overlaping
segment is at it's ripup limit. We should try *not* to rip it up if
we can. Add a dedicated flag "AtRipupLimit".
* Change: In TrackElement, add proxies for isUnbreakable(), new function
updateTrackSpan().
* New: In TrackFixedSegment CTOR, when a supply wire of METAL2 or above is
found, make the underlying GCells "GoStraight".
* New: In TrackElement::canDogleg(GCell*), check for already done perpandicular
dogleg on source/target (reject if so).
* Bug: In Katana::Track::expandFreeInterval(), the OutsideElement case was
not handled correctly, the end index must be increased in that case.
As a result, this function was returning too short an interval.
This was affecting the computation of perpandicular free interval
in DataNegociate::update() for perpandicular fixed but not in a track.
* New: In Anabatic::NetBuilder and NetBuilderVH, now manage GCells with
one Pin, and up to three globals. Used to put terminals at the
edge of a block.
* New: In EtesianEngine, now manage placed Pins.
(done through extractInstanceName())
* Change: In CRL::RoutingGauge::getHorizontalGauge() no longer skip PinOnly
gauges to find the reference one (same goes for vertical).
* New: Anabatic::AutoContact::setLayerAndWidth() to set both layer and
VIA width/side according to the RoutingGauge. If the delta in zero,
use the metal gauge, and the VIA gauge otherwise.
* Bug: In Anabatic::AutoContactTerminal, Anabatic::AutoContactTurn,
Anabatic::AutoContactHTee & Anabatic::AutoContactVTee, in the
updateTopology() method, set both the layer and the VIA with when
there is a change of layer. Note that this default size may be
overriden later by updateSeize() in the case of non-default width
segments.
* New: In Anabatic::AutoSegment, new overload setLayer(size_t) to set
both layer and segment with according to the routing gauge.
* Bug: In Anabatic::AutoHorizontal and Anabatic::AutoVertical, in method
_makeDogleg(), make use of the new setLayer() to correctly set up
the wire width.
Idem for Anabatic::AutoSegment::changeDepth() and ::makeDogleg(),
and in Anabatic::LayerAssign.
The short net mode degrade the routing in some cases. This will be
fixed in a next batch of commits.
* New: In Hurricane::NetRoutingProperty, added "ShortNet" flag for Nets
that are completly inside *one* GCell.
* Bug: In CRL::BlifParser::Model::staticInit(), when looking for the
output of zero and one cell, also skip the blockage net (as well as
automatic and supplies).
* New: In Anabatic::AutoSegment, added "ShortNet" flag to know if the
segment is part of a short net (fully included in *one* GCell).
Also add accessor/mutators for the _analogMode flag (was it ever
used before?).
* New: In Anabatic::NetBuilder::singleGCell(), if a RoutingPad is
vertically small, add a vertical segment to give it some slack.
* New: In Anabatic::Dijkstra::_materialize(), detect "short net" as
they have only one GCell in their source list...
* Bug: In AnabaticEngine::_loadGrbyNet(), reset the AutoSegment
"short net" and "analog mode" creation flags between two different
nets.
* New: In Katana::Configuration, added dedicated ripup for short net
segmnts.
* New: In Katana: partially implemented support for "short dogleg", that
is dogleg that are always kept in same metal because they connect
neighboring perpandicular tracks. Not finished neither activated
yet.
* New: In Katana::TreckElement and derived, export the the *short net*
support from AutoSegment.
* Bug: In Katana::RoutingEvent::_processRepair(), when a segment is
successfully inserted, re-process any perpandicular that is in
repair state, as it may have a new chance to be placed.
* New: In Katana::SegmentFsm::slackenTopology(), always reject short nets.
* Bug: In Katana::Track::check(), correctly handle wide segments instead
of issuing false check messages.
* New: In Anabatic::LayerAssign, a new mode "LayerAssignNoGlobalM2V" has
been added to manage the 3 routing metal technologies like AMS 350nm.
The standard cells have all their connectors punctual and aligned on
an horizontal line in the middle of the Cell. This is a design
inherited from the channel routing times that makes global routing
in metal2 *through* a standard cell almost impossible (except for
directly neighboring cells). Thus, the layer assignment must move
up all the metal2 that span more than two GCells.
* Change: In Katana::Manipulator::ripupPerpandiculars(), when a perpandicular
has a fixed axis and it's (sole) underlying track is taken, ripup
the other segment to ensure the perpandicular interval will be free.
We need to do that, because in DataNegociate::update(), if the
perpandicular is taken, the perpandicular interval will always been
empty preventing the segment to have a valid track span (so directly
failing).
* Change: In Katana::SegmentFsm::conflictSolveByPlaceds() the fixed axis
segments must also been taken into account as conflicts.
Also correct a small bug, the first conflicting segment in the
interval was not taken into account correctly.
* Change: In Anabatic::AutoSegments collections, change the type of all
the flags that where in "unsigned int" (32 bits) to Flags (uint64_t)
as there is now more than 32 flags for functions.
* New: In Ababatic::Constants, added new flag Flags::WithPerpands, which
makes the number of flags tip over 32 bits, thus making mandatory
to uses Flags and not unsigned int.
* New: In Anabatic::AutoSegments_Perpandiculars, manage a new flag
Flags::WithDoglegs to allow to propagate through global segments that
are connecteds via doglegs on local segments. Meaning that there is
a good chance that they could be aligned.
Slighly change the way we propagate on aligned segments: no longer
check for VTee or HTee, but only for same direction and layer as
master.
* New: In Anabatic & Katana, replace all the "int", "long" and their
variants by the less implementation ambiguous "int32_t", "int64_t"
(and variant). This should help to better detect bit trucation in
flags.
Use the type to give a hint about the flags kind:
- Type "Flags", for flags shared among Anabatic & Katana
functions/methods (may also appear in some objects states).
- Type "uint32_t" for flags belonging to an object internal
state of from Hurricane functions flags (those should be
grouped in a Flag subclass in a perfect world).
* Bug: In Hurricane, in StaticObservable::getObserver(), if the slot
pointer is NULL, do not try to access the owner. Returns NULL, so
the caller can be aware of the situation...
* Change: In Hurricane, in BreakpointWidget & ExceptionWidget some
cosmetic changes (fonts and window sizes).
* Bug: In Anabatic, In AutoHorizontal::getConstraints(), take into account
the constraints from the source AutoContact, as it holds the constraints
transmitted by the RoutingPads and sets up by propageConstraintsFromRp().
It is likely to be a bug affecting the original Katabatic as well.
* Change: In Anabatic, in RawGCellsUnder(), check that the segment is not
completly oustside the cell abutment box and truncate the coordinates
to the part that is inside. Use the "shrink" if we reach the east/north
border.
* Change: In Anabatic, in Configuration, no more decorator because we will
use a true derived relationship. Katana *derives* from *Anabatic* and do
not *decorate* it, so the Configuration can do the same. It also implies
that we directly create a Katana engine, not an Anabatic one.
* Change: In Anabatic, in Session, do not allow the opening of the Session
in a standalone fashion (with a static method). Instead it must be opened
using the relevant method of the Anabatic/Katana engine. This ensure we
are opening the right Session type.
* Change: In Anabatic, in AutoSegment_Aligneds() collection the seed segment
is not part of the collection by default, but will be included if the
Flags::WithSelf is set.
* Change: In Configuration, all the flags value are now defined in two steps.
Declared in the header and initialized in the module. This is to prevent
the fact that on some cases, in relation with the Python "extern C" part
modules, we need a true allocated variable. It was causing weird linking
problems.
A side effect is that they can no longer be used as entry is switches,
have to replace them by if/else.
* New: In Anabatic, new GCell::getNeighborAt() utility function.
* Bug: In Anabatic, in GCell::doGrid(), tag all the GCells of the grid with
the grid type... Back annote all the edges capacity (north & east) with
the reserved local capacity.
* New: Complete portage of Kite over Anabatic. The new engine is christened
"Katana" for Kite-Analogic. When it's capabilities and performances
will be on a part with Kite, it is to completly replace it (and take
back the "Kite" name). Preliminary tests seems to show that, contrary
to intuition (because built on a more complex/slower grid), it is even
slightly faster than Kite 8-).
Jean-Paul Chaput