Documentation restructuration:
* Bug: in <hurricane>: bad figure for transf-R1.
* Change, in <kite>: Remove deprecated configuration variables.
* New: documentation <tool> to regroup all documentation not directly
related to one tool.
* Change, in <crlcore>, move the user's guide and the top doc index
into <documentation>.
* New, in <documenation>: added Python script demo based on AM2901.
- In Session::_canonize(): if the canonical segment has never had it's axis
sets, put it into *optimal* position instead of simply constraint
compliant. Force realignment of the segment sets as rounding errors
may introduce slight misalignment on the axis (causing error laters
in the various contact topological updates).
This also seems to solve the ludicrous problem of negative ratio of
expansion wires... (segments *not* at their optimal position at the
start).
- In LayerAssign, KatabaticEngine::layerAssign(), suppress the triple loop
enclosign the desaturation. Now desaturate each layer exactly once and
from bottom to top.
- New: In LayerAssignement & GCell, perform layer assignment on whole net
trunk, instead on a segment by segment basis.
- New: In GCell, fragmentation analysis. When a layer change is requested,
it allows to check if the additionnal perpandiculars created to
maintain connexity will not over fragment the layer. This is particu-
larly critical for M3->M5 which create M4 in perpandiculars, and M4
may be subject to very dense configuration.
- New: In AutoSegment & LoadGrByNet, adds a flags telling if the net's
degree equal 2. Has not proven to be useful yet.
- New: In AutoSegment, adds a flag telling if the segment is the perpan-
dicular part of a dogleg. Has not proven useful yet.
- New: In GCell, compute a "fragmentation" indice, the average free track
length, the more fragmented, the smaller the indice.
- New: In AutoSegment::canMoveUp(), when requested, not only check for free
(parallel) track in the up layer but also check that the begin/end
local segments created to maintain layer connexity do not creates too
much fragmentation. This fragment could prove critical in higly dense
GCells.
- Change: In LayerAssign, disable the ::rpDesaturate() steps, after optimization
the router can handle 10 terminals per GCells in M2 without problem.
And ::rpDesaturate() must uses GCell::stepNetDesaturate() and *not*
directly ::stepDesaturate().
- New: In AutoSegment, adds a "_parent" attribute to keep track of the fragmentation
processus. Currently used only for strap segments, points to the original
segment in the appropriate direction (before the split).
- New: In GCell & LayerAssign, new method of layer assignment. Move up
the whole net trunk if only one of it's segment is inside an over-saturated
GCell. AutoSegment are moved up only if there is at least 2 free tracks
remaining on the upper level.
- Change: In Session::_canonize(), uses the lowest segment Id as canonical.
More reliable than geometricals criterions in the end. Assuming that the
segments are being created in deterministic order, which *should* be the
case consediring the way we are walking through the global routing.
- Change: In AutoSegment, completly suppress the CompareCanonical(), replace
it by the much simpler CompareId().
- Change: In GCell::rpDesaturate(), stops desaturation when bottom density
is under 0.5, otherwise we are causing a severe imbalance in M2/M4
densities. All wires pushed up to M4...
- Change: In ChipTools, for the Pad's RoutingPad, reslect the best component
using the one in the lowest layer. To avoid problem when splitting
AutoContact as we expect the base Contact to be on the lower layer.
- Bug: In GCellConfiguration::_GCell_xG_xL1_xL3(), add H/V alignement constraints
in fork case. This allow NE/SW contact to be splitted correctly later.
- Bug: In AutoContact::split(), the connexity on the splitted contacts was
not correctly restored, leading to canonization and parentage looping
errors. This was concealed by the Kite Track::_check() bug (incomplete
individual TrackSegment checking).
- Change: In GCell & GCellGrid, unificated way of computing all the GCell's
densities.
- Change: In GCell::stepDesaturate(), uses shearUp() instead of moveUp().
- Change: In AutoSegment, new method ::shearUp() to avoid moveUp() saturating
GCells as side-effect.
- Change: In LayerAssign, desaturate until the last two top layers are reacheds
instead of only METAL2 & METAL3.
- Change: In Configuration/GCellGrid/GCell, correct computation of the edges
capacity (based on the numbers of avalaibles layers and not hard-wired to
4).
- Change: In KatabaticEngine, the containers used to store the "nets to
route" is changed from a vector<> to a set<> sorted by the Nets names.
Uses more memory but allows much faster Net deletion.
- New: In _loadNetGlobalRouting(Net*), detect non-routed Nets. A Net is
considered as unrouted if it's 10 first RoutingPads have no slave
Segment. The Net is withdrawn from the set of Nets to route.
- Bug: In GCellConfiguration::_GCell_1G_1L1(), correctly compute the
access flag passed to _GCell_rp_Access().
- Change: In AutoContact::restrictConstraintBox() adds "warnOnError"
parameter to shut down warnings.
- New: KatabaticEngine::_computeNetConstraints() now also returns a set<>
of segments on which constraints are incompatibles. All the NetConstraints
related functions are modificateds accordingly.
- Change: In KatabaticEngine::_loadNetGlobalRouting(), create dogleg
on overconstrained Segments computed by a call to _computeNetConstraints().
Overconstraineds Segments can occurs when a global is directly
connecting two punctual RoutingPads (see "ieee_division").
As a side effect, and to avoid looping each Net is revalidated
*before* making any dogleg. So Nets are revalidated one by one instead
of alltogether.
- Change: In AutoSegment::_toConstraintAxis(), do nothing if constraint
interval is empty.
- Bug: In KatabaticEnginge::_desaturate() (part of LayerAssign), reset
the moved up Segment pointer between iterations. Suppress the warning
about the GCell's DyKeyQueue beeing not revalidated on deletion.
- Change: In LoadGrByNet, manage the 4G+3L1 configuration with the generic
_GCell_xG_xL1_xL3(). Needed for "dct_lvl2".
- Bug: In GCell, bad cast between <size_t> and <unsigned int>, showed only
under 64 bits.
- New: In KatabaticEngine::printMeasures(), adds a support for measurments,
both time and memory size in Mb are stored for each tag.
- New: Adds measurements "Gates", "GCells", "load" & "assign".
- New: In GCell & LayerAssign, adds a supplemental desaturation steps when
a GCell containts too many (more than 7) RoutingPads in M1. This is
especially useful for the eFPGA matrix with it's custom made cells.
- Change: Uses a map instead of a sorted vector in the layer assignment
steps. It was used to order GCells by density, and causing terrific
slowdown on big examples. Namely the eFPGA 16x16 took 1h20, now it's
down to 21 seconds...
Devellop a new object <DyKeyQueue> to handle map with elements
whose key *can* change. Next step is to normalize the key caching
mechanism and to templatize.
- Change: In LayerAssign, account the number of globals AutoSegments
moved up.
./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