Note: Keep the various hfnsX.py as toolboxes for future experiments.
* New: cumulus/plugins/block/hfns3.py, build the trunk of the
net as a RMST. First with the "Iterative One Steiner Point"
(terribly slow above 100 points) then with FLUTE.
For the global routing trunk, must be very cautious to
check that the cluster "graph point" is the one of the
it's buffer RoutingPad so both end up in the same GCell.
* New: cumulus/plugins/block/timing.py, stub for basic
timing computation and conversion between sink and
capacitance. Currently based on the fake 350nm given as
example in SxLib ("man sxlib"...).
* New: In Hurricane::NetRoutingProperty, add and change the meaning
of the following flags:
- ManualGlobalRoute : now means that a global routing *trunk*
is present, made of "gmetalh", "gmetalv" & "gcontact".
- Manualdetailroute : added, get the former meaning of
ManualGlobalRoute, that is, the detailed routing is
already present for this net, but can be changed by the
detailed router. Implies that it respect the Terminal,
HTee & VTee structuration.
* New: Add Anabatic::Diskstra::loadFixedGlobal(), to account
a manually global net into the edges capacities.
* New: In Anabatic::Edges::ripup(), exclude manually global routed
segments from the ripup. Change the segment sorting function
so that thoses segments are put in head of list (considered
as "smaller").
* Change: In AnabaticEngine::setupPreRouteds(), now detect manual
global routed and manual detail routed signals, and tag them
accordingly.
* New: In AnabaticEngine::Configuration & Session, add proxies
for the global routing layers ("gmetalh", "gmetalv", "gcontact").
* New: In Anabatic::Constants, add flags for global fixed and
detail routed nets.
* Change: In KatanaEngine::updateEstimateDensity(), now use int64_t
for flute coordinates.
* New: Add CRL::RoutingGauge::hasLayer(), to know if a layer is
managed by the gauge (comparison by mask).
* New: In flute, added a Python binding. Contains two methods: "readLUT()",
to load the POWV9.dat and POST9.dat and "flute()" to build a RSMT.
"flute()" takes a tuple of positions (themselves 2-uple) like:
( (x0,y0), (x1,y1,), ... (xN,yM) )
and returns a tuple of 3-uple of branches:
( (n0,x0,y0), (n1,x1,y1), ... )
In "flute.h", set the distance type (DTYPE) to int64_t to always
accomodate DbU::Unit.
As it now uses the Hurricane Python interface and the path utilities
from CRL Core, move it's compilation *after* them (see build.conf in
bootstrap).
To perform HFNS, recursively makes clusters of closest sinks, less than
30 sinks and with a control of the half-perimeter at all the clusters
levels. Clearly needs lots of improvements but the backbone of the
feature works. Make use of the pool buffers as do the clock tree.
Clustering is done with a degenerated Kruskal algorithm.
This is rougly based on the article:
Buffered Steiner Trees for Difficult Instances
Charles J. Alpert, Member, IEEE
IEEE TCAD, Vol. 21, No. 1, January 2002
0278–0070/02$17.00 (c) 2002 IEEE
* In NetBuilder::_do_xG_1PinM3(), in the three global case, an
incorrect connexion was made.
* In NetBuilder::_do_2G_xM1_1PinM2(), in the global E+W case,
segment between vtee1 <-> turn1 must be vertical.
* Change: In Hurricane::RoutingPad::getPlugOccurrence(), it was
forbidden to rebuild the Occurrence over anything other than
Plug. But as we now also uses Pins, it becomes useful to
rebuild the Occurrence on a Pin. The name of the function
stays, even if its less meaningful now.
* Bug: In Isobar::PyOccurrence, use "compare by value" instead of
"compare by pointer" for the Python comparison function. So we
really can compare identical Occurrences.
* New: In Isobar::PyPoint, export the "manhattanDistance()" function.
* New: In Isobal::PyCellViewer, add the following funtions to the
exported interface:
- "select()"
- "unselect()"
- "unselectAll()"
- "reframe()"
- "setShowSelection()"
So now we can easily highlight components from Python (to ease
debug).
* New: In Hurricane::CellViewer, create a proxy for the CellWidget
"reframe()" function.
When displaying big designs, the drawing was very slow because all
the graphical objects were browsed *before* deciding if they were
big enough for display. So we introduce a new mode of working of
the QuadTree to that the ones with areas *below* a certain thresold
just get skipped. Of course, the previous behavior must be preserved
(when the threhsold is zero or negative) as most of the time, we
*want* *all* the objects under an area, not only the biggest ones.
For now, in CellWidget, the visibility threshold is set to 20 pixels.
With this, we can display the Libre-SOC "test_issuer" of 100Kgates
in between 20 to 30s.
* New: In Hurricane::QuadTree_GosUnder collection & locator, add a
"threshold" argument so that QuadTree which both area sides are
below that lenght will be ignored (no walkthrough at all).
Zero (or negative) threshold means that no QuadTree will be
pruned, we get back to the previous behavior.
* New: In Hurricane::Slice::getComponentsUnder(), add the threshold
argument (with a default to 0).
* New: In Hurricane::Cell::getInstancesUnder(), add the threshold
argument (with a default to 0).
* New: In Hurricane::Query, added support for the threshold parameter.
In doQuery(), instances that have *both* side under the threshold
are pruned (slightly different from QuadTree areas).
* New: In Hurricane::CellWidget::_redraw(), set the visibility
threshold to 20 pixels (arbitrary, must be parametrized).
* New: In Hurricane::Cell, added destroyPhysical() method.
Remove *all* physical components and the DeepNets (so make
a virtual *unflatten*) in the correct dependency order.
In particular, RoutingPads, that relies on Occurrence over
physical components must be destroyeds *before* the entity
they are based on is destroyed.
* New: In CRL::PyAllianceFramework, export getCatalog(), in PyCatalog,
export the getState(name) method and add a PyCatalog_Link().
In PyCatalogState, add PyCatalogState_Link(), cannot use the
macro because of the C++ name resolution operator (Catalog::State).
* New: In Isobar, export the RoutingPads collection to the Python interface.
* New: In Hurricane::Net, export the getRoutingPads() method to the
Python interface (hence the need of the previous export).
* Bug: In cumulus/plugins/alpha/block/clocktree.ClockTree.splitClock(),
forgot to wrap inside an UpdateSession.
* Bug: In cumulus/plugins/alpha/block/clocktree.spares.raddTransNet(),
forgot to set the masterNet for recursive call when the plug already
exists.
* New: Added multiple clock support in H-Tree generation in alpha/block.
* New: In CRL/etc/<NODE>/<TECH>/plugins.py, added three new parameters
for block plugin config:
"block.spareSide" : The size of the minimum side of a buffered area.
(quad-tree leaf).
"spare.buffer" : The model of the cell buffer to be used.
"spare.maxSinks" : max number of sinks on a buffer before issuing
a warning (non-blocking).
* Bug: In Etesian::BloatCell::getAb(), never apply the bloat profile to
fixed cells. In case of buffers from the spare maxtrix it was leading
Coloquinte to detect an overlap of fixed cells, which it do not
support (rightly so).
* Change: In CRL/helpers.Trace, flush stderr before issuing the trace
message to avoid mixing up stdout & stderr (sometimes misleading).
* Change: In CRL/helpers.overlay.CfgCache, no longer display all the
applied setting after a call to apply(). Too verbose.
* Change: In CRL::Environment, regex_t are now pointers instead of values,
this way the "in initialization" flag can be removed (maybe still too
complicated).
* Bug: In Anabatic::AutoHorizontal::_slacken(), when creating the target
dogleg, if the target was a fixed pin (M2) on the east side, the
vertical was put righ on the side, which is *not* an allowed track.
The exact east side is not covered by a vertical track (this comes
from a policy adpopted for the GCell, to avoid overlapping tracks
on the border). So we got an error message when katana tries to
bind it to a track. Now shift one perpandicular pitch to the *left*
for the EAST pins.
* Bug: In Coloquinte::rough_legalizer, when running in debug mode,
some assert where popping up. But, in normal mode the placer
just went fine. So we assume they where too broad in their
scope. Commented out.
* Change: In Anabatic::NetBuilderHV::_do_1G_xM1_1PinM2(), the M2 pin
(horizonal) was directly aligned with the outgoing global through
a H-Tee. This was too rigid. Now it can be misaligned.
* New: In Katana::Manipulator::moveUp(), if the length of the wire is
greater than 60 pitchs, increase the reserve from 0.5 track to
1.0 track. This is to left more free tracks when lots of very
long segments are moved up. Make the run 17 of the ARM test bench
pass.
* Change: In AnabaticEngine::_desaturate(), now use a STL priority
queue built upon a vector with duplicates of the GCell keys.
Much much more faster than the previous set with a custom
sorting key over the GCell. Insertion & removal where getting
a lot of time. Now it's down to less than a minutes, even for
big designs.
* Change; In Anabatic::GCells, add the ability to clone the GCell
key, to use it independently in GCells sorting containers.
This way the cloned value do not change when the GCell is
updated. Keep track of the latest (most up to date) cloned
key.
* New: In CRL/helpers.overlay.CfgCache, add support for context manager,
so we can progressively remove the use of Configuration. One class
to use in all contexts (immedaite setting or storing a set of
Cfg parameters).
* Bug: In Etesian::toColoquinte(), inadvertently stopped taking into
account fixed instances... Compute correctly the number of
instances to place (all instances minus fixed ones..).
* Bug: In cumulus/plugins/block/block.py, when a block is found to have
a layout, that is, is already placed & routed, it's instances must
be flagged as FIXED instead or PLACED so Etesian will really sees
them as unmovable and don't account them to be placed.
* New: In EtesianEngine::toColoquinte(), fixed blocks must not be
accounted as instances to be placed.
When, there is no instances to place, just issue an error message
and return.
* Bug: In Anabatic::NetBuilderHV::_do_2G_xM1_1PinM2(), in the case of a
bend of global wires, the topology was faulty. Too rigid topology,
the global wire was forcibly aligned on the external Pin.
* New: In Katana::Configuration, two new parameters hTrackReservedMin
and same for V. Forced decrease of all edges capacities, to use
when the global router is too optimistic...
* New: In EtesianEngine::setDefaultAb(), add computation of AB when
using fixed width (fixed height) was already available.
Add attribute management for the fixed width along with a
Python export.
* New: In Anabatic::AutoContactTerminal, new method "isOnPin()" to
know if the terminal contact is on a Pin. Introduced in base
class.
* Bug: In Anabatic::AutoHorizontal::_canSlacken() & AutoVertical,
forbid slackening from a Pin. As Pin are all on the side of the
cell (aligneds) and the perpandicular segment will be locked on
the Pin, it generates intractable overlaps for the router.
* Bug: In Anabatic::Configuration::selectRpComponent(), always select
the Pin as the best component. It is mandatory for a block that
we pick up the terminal at the periphery and not one deep inside
the block.
* Bug: In Hurricane::RoutingPad::setOnBestcomponent(), same rule as
in Anabatic, if there is any, the Pin must always be the best
RoutingPad component.
* New: In CRL/helpers/overlay.py, CfgCache class to hold a set of
configuration parameters and apply it on demand. It has a
different behavior than Configuration.
Jean-Paul Chaput
Marie-Minerve Louërat