* Bug: In Anabatic::Edge::getDistance(), remove the additionnal 0.1
added to horizontal edges. This was for testing before the hScaling
parameter was added (to the distance computation in GlobalRoute).
* New: Anabatic::Path_Edges, collectio to walkthrough all the edges
between two node. More complex than in Knik as we are no longer
using a regular grid. We may request the north bound path or south
bound path.
Collection returned by AnabaticEngine::getEdgesUnderPath().
* New: In Anabatic::NetData, add a new flag GlobalEstimated to tell if
the net RMST has been computed (using FLUTE).
* New: In Anabatic::PriorityQueue, used to sort Vertexes by increasing
distances, add a new criterion to be used in case of distance
equality. The attractor which should be the center of the search
area. In case of equality, we choose the Vertex which is closest
to the attractor. Give a small improvement, and more "dendritic"
trees.
For a more simple implementation of the comparison function it is
made as a static member (so no two Dijkstra objects at the same
time...).
* Change: In Anabatic::Edge, make the estimate occupance a floating
point number instead of an integer.
* New: In Katana::GlobalRoute, finally implement the estimated congestion
driven router. Net RMST estimated using FLUTE.
Use the historic cost from Knik implementation and not the one
given in Damien's thesis, which seems not be the same and a bit
strange.
* New: In KatanaEngine, add the ability to exclude nets from routing,
and export it to Python.
* Change: In Hurricane::CellWidget, set the minimal size to 350 pixels
to fit my normal DPI secondary screen...
* Change: In Hurricane::Error(), reactivate the backtrace generation by
default. Seriously slow down the program each time an Error is to
be constructed.
* Bug: In Analog::Device::preCreate(), check for NULL Technology before
attempting to use it.
* Change: In Hurricane/Analog, remove all '*Arguments*' classes and their
Python interface. It was an obsoleted way of passing devices parameters
to the Python layout generators (located in Oroshi). Now we just get
them straight from the Device with the getParamter() method.
* Change: In CRL::System CTOR, add Python pathes for Oroshi & Karakaze.
* Change: In Oroshi/Python/WIP_*.py layout generator scripts, remove
all uses of the "Arguments". Directly access the parameters through
the device itself. Make the checkCoherency() with identical arguments
as of layout().
* New: Bora tool that performs analog place & route. Based on a slicing
tree representation. It is the thesis work of Eric Lao.
Code beautyfication and some programming cleanup.
* New: Karakaze tool, provide the Python base class AnalogDesign used
to build an analog design. Create/configure devices and assemble
them in a slicing tree.
* Change: In Unicorn/cgt.py, display the stack trace in case of an
ImportError exception as well as for other exceptions.
Add Bora to the set for included tool engines.
* Bug: In Anabatic::Dijkstra, the degree of a vertex (the number of neighbors
belonging to the same net) was miscalculated. This was leading, in the
materialize step to some feed-through vertexes not being broken.
Leading in turn to incomplete transformation of the detailed routing.
Also in _trackback(), the degree of the first vertex we were backtracking
from was not incremented.
* Bug: In Anabatic::Dijkstra::materialize(), systematically use
GCell::breakGoThrough() on both source and target. This is needed when
we are in the ripup phase as both source and target can be go-through.
This was also leading to incomplete detailed routing transformation.
* Change: In Anabatic::Edge::ripup(), ripup one third of the segments instead
of thoses exeeding the global length threshold. This way we are sure to
desaturate an edge. Needs to be further calibrated.
* Change: In Aanabatic::GCell::breakGoThrough(), no longer return NULL.
Return existing gcontact if any. Break if it is a go-through and create
a new gcontact in last resort. Maybe rename this function.
* New: In Anabatic::Configuration, new parameters:
- anabatic.edgeHScaling, to adjust the length of the horizontal edges
relative to the vertical ones (this is a ratio).
- anabatic.globalIterations, set the maximum number of ripup passes
of the global router.
* New: In CRL/etc/*/kite.conf, added new parameters anabatic.edgeHScaling
and anabatic.globalIterations.
* New: In Katana::GlobalRoute::DigitalDistance, take into account the new
edgeHScaling factor. Must be used when the capacity of V-edges differs
greatly for H-edges (case of AMS 350nm c35b4 for instance).
* Bug: In Katana::GlobalRoute::DigitalDistance, the historic cost is
computed for an edge length of "1". Must be multiplicated by the
current edge length to have any measurable effect.
This bug is finally explaining why the ripup was producing the
same solutions over and over, the historical cost was negligible!
* New: In Hurricane::Entity, add an id counter limit and a memory size
limit. The two limits are checked only when a new Entity object is
created. This should help avoiding massive memory links.
* New: In CRL Core, add a "crlcore.groundName" and a "crlcore.powerName"
parameter to specify the name of the ground/power signals to be
created if they are missing in a Cell. For Alliance libraries it
would be "vss" & "vdd" (default values), but for real technologies,
it is often "gnd!" & "vdd!".
The Blif parser is modificated to make use of it.
* Bug: In AnabaticEngine::unify(), set the resulting unified segment in
the center of the GCells common side. Gcells under a segment are
found by using the edge that cover the segment axis. When we have
a "bend" GCell stack and the axis is wrong, they could be ommited.
This was causing deleted segments to be not removed from some
Edges, then core dump.
* Change: In Anabatic::AutoSegment::create(), smarter choosing of the
reference contact, select the fixed or terminal one instead of
always the source one.
* New: In Anabatic::Edge::isEnding(), new function to check if a
segment going through an Edge is starting/ending in either source
or target GCell of the edge (active only when running in channel
mode).
* New: In Anabatic::Edge::add(), a segment takes part in the occupancy
only if it is not ending in either source or target (channel mode
only). The occupancy due to terminal is pre-computed in Katana.
* New: In Anabatic::Edge::ripup(), in channel mode, never ripup a
segment which is ending in either source or target (we *have* to
access this edge to connect to the terminal).
* Bug: In Anabatic::GCell::hcut() and vcut(), force the update of
the Edge which is on the side that will get splitted by the cut.
It's capacity will be reduced to it must be updated.
* Change: In Anabatic::GCell::updateGContacts() add a flag to conditionnally
update horizontals or verticals only. We may require only a partial
update when resizing the GCell in only one direction.
This, again, related to the fact that we compute the GCells under
a segment thanks to it's axis position, so we need to be very careful
when modificating axis.
* Change: In Katana::Block::resizeChannels(), only update GContact vertical
position. Do not disturb X positions of segments.
* Bug: In Katana::GlobalRoute::DigitalDistance, in channel mode, some
Edges can have a zero capacity, but still be reachable if the net has
a terminal in either source or target. Look for this case and return
a distance of zero instead of "unreachable". This was causing the
global routing not to complete in channel mode.
For computing the edge distance, makes the vertical edges much more
long (10 times) than the horizontal ones as the vertical capacity is
very limited. Hard coded for now, should make it a parameter in the
future.
* Change: In KatanaEngine::annotateGlobalGraph(), decrease the capacity
of edges with reserveCapacity for each terminal inside a GCell.
Both north and south edges are decreased as we a terminal will
block both north and south edges.
As a counterpart, the Edge capacity is not decreased when the
global router connect to a terminal.
* Change: In Katana::RoutingEvent::revalidate(), when in repair stage,
do not expand the slack for horizontal segments in channel mode.
So they may not overlap the standard cell row.
* Bug: In Stratus documentation, do not use the french option in babel,
the documentation is in english!
* New: In Documentation, added Hurricane/Python tutorial, part for drawing
layout.
* New: In Anabatic::Edge::ripup(), sort edges to ripup by decreasing
length so the longest gets riped up first.
In two metals mode, ripup all segments *except* those who have
terminal contact in the GCell, because they must use it anyway.
* Bug: In Anabatic::GCell::getEdgeAt(), the size hint flag was not
correctly selected, leading to the use of an edge in the wrong
side. So deleted segments where not deleted from their edges,
leading to core dump. Don't understand how it can have not shown
until now.
* New: In KatanaEngine::annotateGlobalGraph(), for two metal gauges,
decrease the edge capacity on both north and south side for each
terminal in the GCell. As a terminal must be accessed through
north or south it blocks the whole vertical track (hence both
edges).
* New: In Anabatic::EdgeCapacity, dedicated object to manage the capacities
of an edge by layer. This needed now because with real technologies
layers capacities differs (unlike with symbolic technologies).
This object is separated to be shared between Edges with identical
characteristics (direction+interval). Deletion is automatic and
done through refcounting. All the already allocateds EdgeCapacity
are kept into a set in the AnabaticEngine (key is (direction,interval)).
* New: In Anabatic::Edge, capacities are stored in a shared EdgeCapacity
object. The total capacity can be annotated (i.e. decreased).
EdgeCapacity attribute is created during the materialize() call.
The capacities are computed at this time.
The incCapacity() function is renamed in reserveCapacity().
* New: In Anabatic::AnabaticEngine, added attribute _edgeCapacitiesLut
to store the shared EdgeCapacity.
Lookup/Creation of an EdgeCapacity is done through _createCapacity().
* Change: In Anabatic::Constants, rename Flags::IllimitedCapacity into
Flags::InfiniteCapacity. Add Flags::NullCapacity (both for Edges).
* Change: In Anabatic::GCell, implement a by depth (for layer) getCapacity().
This modification did expose a bug in the density calculation :
per depth density where divided by the complete density instead of the
density's depth. This was leading to greatly underestimated densities.
Thoses underestimations where preventing Dijkstra and layer assignement
to manage congestion correctly (in fact, it was acting as if there
never was congestion).
Also avoid a divide by zero (thus -NAN showing in densities).
* Change: In Anabatic::GCell, rename setEdgeOccupancy() into the more
accurate forceEdgesCapacities().
Note for Eric: only the first Edge on each side has it's capacity
forced. What if there's more than one Edge ?
* Change: In Anabatic::AutoContactTerminal::getNativeConstraintBox(),
when the anchor is a RoutingPad (which must be always the case),
perform the true computation of it's position based on the
segment occurrence. It is a important change, previously the
area was in fact the "center line" of the connector while now
it is really an area (mandatory for "half-offgrid" terminals of
real technologies).
The change is not complete yet, the area should be shrinked
by the half size of a VIA, because the area applies to the center
coordinate of the VIA (to be done quickly).
* Bug: In Anabatic::AutoContactTurn::updateTopology(), when a dogleg
is created (restore connexity after a layer change) the layer of
the VIA, based on the segments it connects to must be re-computed
*after* the dogleg has been made.
* Change: In all files of Anabatic, when comparing two layers, no longer
use the Layer pointer itself, but the layer mask. This allow a
transparent management of both real and symbolic layers (which
do share the same mask). Real metal layers (not VIAs) will be
BasicLayer and symbolic metal layers will be RegularLayer.
* New: Anabatic::Configuration::selectRpComponent(), select the best
RoutingPad component for metal1 terminals. Look for the metal1
component with the biggest accessibility on-grid.
RoutingPad using other metals are left untoucheds.
* New: New function Anabatic::Vertex::getNeighbor(Edge*) to get the
neighbor Vertex through an Edge*. This method allows to write
clearer code as we no longer need to access the neighbor through
the underlying GCell.
Also add proxies for GCell methods in Vertex.
* Bug: In Anabatic::Dijkstra::_toSources(), in the ripup stage, when
a component with multiples vertexes is reached *and* two of it's
vertexes are reached *at the same time* (one from which we backtrack
and one still in the queue) extraneous edges may be created by
_materialize(). Case occurs on snx/c35b4, "abc_5360_n903_1".
To solve this, Dijkstra::_toSource() is modificated, the "from"
edges of the newly reacheds vertexes are reset to NULL, *except*
for the one we will be backtracking from. That is, the one given
in the source argument.
* Change: In Anabatic::NetBuilder class, put the various Hooks and
RoutingPad sorting functions as class ones.
* Bug: In AutoSegment::setLayer(), raise the SegInvalidatedFayer flag.
This unset flag was causing AutoContactTurn::updateTopology()
to not work as expected and making gaps, this was the cause of
the last remaining warnings about layer connexity.
* New: In AnabaticEngine::invalidateRoutingPads() this method is a temporary
workaround for a Hurricane problems. When an instance is moved, the
RoutingPads that use it must be moved accordingly, but they are not
invalidated so they stay in the wrong QuadTree.
New method ::_resizeMatrix() to be called when the associated Cell
is resized.
* Bug: In AutoHorizontal::getConstraints() and AutoVertical::getConstraints(),
the *target* constraints where never merged.
* Change: In AutoHorizontal::getCells() and AutoVertical::getGCells(),
now return a boolean to tell if it was ok (must not encounter a NULL
GCell while progessing from source to target).
* New: In Anabatic::Configuration and Anabatic:Session, create new methods:
- getDHorizontalLayer()
- getDhorizontalDepth()
- getDHorizontalWidth()
- getDHorizontalPitch()
And so on for Vertical and Contact.
They supply depth-independant informations about the H/V layers to
build the initial detailed routing.
The AutoSegment::create() methods have been modificated accordingly.
* New: In Anabatic::GCell, add two new types "StdCellRow" and "ChannelRow"
for implementing 2-Metal blocks.
Rename the GCell::setXY() method in GCell::setSouthWestCorner(),
move the contents of GCell::updateContactsPosition() into it and
suppress it.
WARNING: In case of a GCell shrink this may cause problems. But for
now we only expand...
New method GCell::getNetCount() to count the number of Net going
though the GCell.
* Change: In Anabatic::Edge, add specific support for capacity of 2-Metal
routing channels.
* Change: In Anabatic::Dijsktra various methods, replace the "gcell->isMatrix()"
calls by "not gcell->isAnalog()". Add more check so that the methods
pertaining to the analog routing (GRData) are not called in digital
mode.
* New: In Anabatic::Dijkstra::materialize(), add support for 2-Metal specific
cases. That is, always break in case of vertical pass-through or
U-turn. The global routing must always be broken in H-Channel.
* New: In Anabatic::GCell & Anabatic::Edge, make use of the Session mechanism
to ensure the revalidation. The "::revalidate()" method is then moved
as "::materialize()" (overload of Go) and "::_invalidate()" becomes
"::invalidate()"
* Change: In LoadGlobalRouting, cosmetic rename of SortHkByX in SortHookByX.
* New: In GCellTopology, added support for building 2-Metal topologies.
* ForkStack is now an object attribute as many methods do need it.
* To push segments/hook on the stack, a new method "push()" is
available. Perform NULL and fromHook checking. Can also setup
_southWestContact or _northEastContact if it is the "from" edge.
* N/S/E/W edges are now vector as in digital channel mode there
can be more than one.
* Added build topological build methods:
- doRp_2m_Access() RoutingPad stem access.
- _do_2m_1G_1M1() North or south access.
- _do_2m_2G_1M1() North AND south access.
- _do_2m_xG() H-Channel routing.
* New: In Anabatic::Matrix, new ::resize() function, as Cell can be resizeds.
* New: In Anabatic::Vertex, new static method ::getValueString() for a
friendly text rendering.
* New: In Katana::DigitalDistance, support for channel routing.
* Change: In KatanaEngine::digitalSetup() and KatanaEngine::runGlobalrouter(),
for channel routing, calls to setupPowerRails() and
protectRoutingPads() must be called after the core block has
been fully dimensionned.
::runGlobalrouter() contains the code tasked with the grid creation
and channel sizing.
* New: In KatanaEngine: Added support for core block, for 2-Metal routing.
May be expanded for over-the-cell routing in the future.
Added methods :
- isDigitalMode()
- isAnalogMode()
- isMixedMode()
- isChannelMode()
- getBlock() / addBlock()
- setupChannelMode()
- createChannel()
* New: In Katana, new class Block to manage core blocks and perform
channel routing.
* New: In Katana::Session, new convenience method "isOpen()".
in Dijkstra.cpp/.h, Edge.cpp/.h, GCell.cpp/.h:
*Change: GlobalRouting wires can have different wires' width. One Net's wires has one width only.
in NetRoutingProperty.cpp/.h, PyNetRoutingProperty.cpp/.h:
*New: A wire width parameter is added. It can be set through python script.
* Change: In Hurrican::Component::_preDestroy(), the cascaded destruction
of slaves Components and subsequent ring merge and Rubber creation
was using set<> sorted on pointer values. This was making the order
of destruction of the slaves components non-deterministic. Now the
sets are sorted according to their id (see Entity).
* Change: In Hurricane::SharedProperty, the _ownerSet is now a vector<>
instead of a set<>. This will slow the search process, but hopefully
this is needed only during construction/destruction of a Property
so the slow down shouldn't be too much.
This has also an impact on Hurricane::Relation.
* Bug: In Anabatic::RawGCellsUnder, *again*, the north and east borders
of the whole area *are* includeds (shut up disgraceful warning).
* New: In Anabatic::GCell, add a display mode to select between boundary
display (for analogic) and density display (for numeric).
* Bug: In KatanaEngine::runGlobalRouter(), do not check if an Edge cannot
be desatured when the vector or overloaded Edges is empty.
(one less disgraceful warning)
* New: In GraphicKatanaEngine::drawGCell(), support for drawing the GCells
in density mode. Use the fire scale and the MaxDensity mode.
Setup the GCell drawing mode in GraphicKatanaEngine::initGCell().
* Change: In GraphicKatanaEngine::drawEdge(), adjust the various thresholds
for showing the Edge and its label.
* New: In CRL Core, adds "anabatic.gcell.displayMode" to the set of Anabatic
parameters.
Adjust the showing threshold for GCells in "display.conf" so when the
zoom level is low, we still can see the density map.
* Change: In Anabatic:
- Long wires, going straight accross multiple GCells were splitted
into a long sequence of segment+contact, one for each gone-through
GCell. It was memory-consuming and algorithmically inefficient.
Now there is only one straight wire.
As a consequence, Edges now store the list of all segments
going through them. That way we do not need to load a Net into
Dijkstra before riping up some of it's segment. The ripup mechanims
can now be implemented right at AnabaticEngine level, without
dependencies over Dijkstra. Note that long wires are automatically
broken in two if we need to attach a new branch on it (breakAt()).
And fused back into one if a branch is removed (unify()).
* Bug: In Hurricane, in tstream (Commons.h), in the *template* overload of
operator<<(), do not use the operator<<() of ostream as it will be
looked up in "stage 1" (template definition) and so will miss all the
overloads added later and built over getString<>(). Instead, make use
of getString<>(), which, as another template will be looked up in
"stage 2" (template instanciation) and at that point will have all the
needed template specialisation of getString<>().
We also need to define new stream manipulators to be able to create
a matching template overload not dependant from the implementation.
To avoid name clashes, we prefix a 't'. For now, only 'tsetw()' is
refined.
As a side effect, we cannot directly print bit-fields into the stream,
we must go through an intermediate variable (happens once in AutoContact).
* New: In Anabatic:
- In AnabaticEngine, keep track of overflowed edges.
- In AnabaticEngine, getNetsFromedge() to lookup all nets going
through an Edge.
- In Configuration, read the Kite "reserved local" parameter to
decrease the Edge capacity (it's a guessing of the cost of the
local routing).
- In Edge, add an attribute to know if there is an associated
segment of the current net (set by Dijkstra::_traceback()).
Transparently manage the overflowed edges.
- In GCell_Edges, correct a filtering bug when not all sides are
selecteds.
- New GCell::getEdgeTo() to find the edge between two adjacent
GCells.
- New GCell::unrefContact() to automatically removes global contacts
no longer used by any global segments (used during the ripup
step).
- In Dijkstra::load(), now able to "reload" and already partially
or completly routed net (look for Contact of "gcontact" layer
and their attached segments).
- In Dijkstra, keep the last net loaded until the next one is.
Put the cleanup operations in an isolated function "_cleanup()".
- In Dijkstra::_selectFirstsource() and run(), load first source
component made of multiple vertexes.
- In Dijkstra::_trackback(), link the Net segments to the Edges.
- New Dijkstra::ripup(), Dijkstra::_propagateRipup() to perform
the ripup of one edge of a Net (must be loaded in Dijkstra first).
Dijkstra::_tagConnecteds() setup the connexId of a set of Vertexes
- that are connecteds through edges *with* segments.
- In GraphicAnabaticengine & GlobalRoute.cpp, embryo of a global
routing tool with ripup & reroute.
* Bug: In Anabatic:
- In _propagate(), on reaching a target, forgot to remove it from
the queue before pushing it back with the new distance. It also
simplificate the core algorithm as target as treated normal nodes.
* New: In Anabatic:
- Update cdebug to use the fastest macro version.
- More readable drawings of GCells and Edges.
- Added timer support.
- The distance is now computed in DbU::Unit (aka long) and not in
normalized float.
- The distance function is now a callback (std::function<>) that
can be changed (a default is provided at initialization).
- New concept of branch in the agglomerated connex component.
Each trace back part create a "branch" (tagged with a "branchId").
When a node is reached with the same distance, but from two
different branches, choose the the branch that was lastly created.
This create a slightly different tree which grows outward from
the newest branches.
- Makes the horizontal edges *slightly* longer than the vertical ones
to skew the tree to use vertical edges, as it is usually less
congested than the horiontal one (due to metal1 cell terminals).
It is also my understanding that it is useful to reduce the
number of vias, whithout introducing a via cost.
* New: In Bootstrap:
- Script sprof.py to perform sprof & demangle libraries execution
profile.
* ToDo: In Anabatic:
- Corner optimization. Sometimes when two corners are possible, the
wrong one is choosen. That is, one of it's edge cannot be used for
further grow of the tree.
* Bug: In Anabatic:
- In Dijktra::load(), do not create Contact while looping over the Net's
components (for RoutingPads).
- In Dijkstra::propagate(), reset the connexId of reached Vertexes
on updating the stamp to avoid using connexId from previous runs.
Push the vertexes on the queue while backtracking (with a distance
of 0.0). Do not reset the "_from" as we need it in Dijkstra::toWires()
to know how were the routing is.
- In Edge::getDistance(), convert to float more early so the normalisation
do not always end up in zero.
- In GCell::_add(), when the axis of the new edge is equal to one already
on the given side, adds it *after* the existing edge and not before.
(to be coherent with the way GCells are split in two)
- In GCell::hcut() and GCell::vcut(), create an Edge if is equal to X/Y Max,
*but* the new chunk is flat. The new chunk is then expected to expand
"below".
- In GraphicsAnabaticEngine::drawGCell(), display the id of the GCell in
a small box at the center of it's bounding box. Nothing displayed for
flat GCells, to avoid cluttering.
* Bug: In Anabatic:
- In Edge, materialize was done before the Edge get it's final position.
Must be invalidated afterwards for the UpdateSession to put him in
the correct QuadTree.
- Functional (debugged) Dijkstra skeleton.
* Change: In Hurricane, in CellWidget, enable the drawing of ExtensionGos
according to the the threshold paramaters (as layers do).
* Change: In CRL Core, provide DrawingStyle for Anabatic::GCell &
Anabatic::Edge. Use the exact name of the ExtensionGo.
In kite.conf, add settings for the size of an edge as the optimal
size may change between analog & digital designs.
* New: In Anabatic:
- Support for fast position query using a matrix.
- Computation of the Edge capacity, based on the routing gauge parameters.
- Graphic display of the edges capacity. Added configuration
parameters to size the bounding box of an Edge.
* Bug: In Anabatic:
- GCell::_moveEdges was both too complex (some cases never arise the
way we divide GCells) and bugged.
Jean-Paul Chaput
EricLaoGitHub