Repository navigation
Conversation
A cross-die net arrives from 3D RCX as three SPEF pieces: each die's wiring (read with -path <chip_inst>) and the die-to-die bond (read at top scope). read_spef left them on three different nets, so a driver found only its own die's piece. On the 3dic_cross fixture chipA's driver saw 2 fF instead of ~103 fF, and changing the bond resistance did not change timing at all. The dbChipNet above a bonded boundary is the net all three pieces belong to, and the net the bonds SPEF is written against. Three dbNetwork overrides ascend to it: - net(const Pin*) on a chiplet boundary bterm returns the chip net. SpefReader::dspfBegin climbs term -> pin -> net to pick the owning net of a -path read, so this is what merges the die pieces. - highestConnectedNet() on a bonded die net returns the chip net, so Parasitics::findParasiticNet lets a driver on either die find it. - highestNetAbove() must agree, or ensureParasiticNode marks every node of the merged network external. A die-net -> chip-net index is built with the existing unfolded maps. staToDb(const Net*) returns null for a chip net instead of casting it. Read the bonds SPEF before the die SPEFs: a top-scope read replaces a net's parasitic network, a -path read merges into it. Known limitation: Tcl net reports (report_net) walk the network through OpenSTA's generic traversal, which now climbs from a bonded die net to the chip net but does not descend into the other die, so they list the far die's bump pad but not its loads. Timing and SPEF reading use dbNetwork directly and are unaffected. Also not covered: estimate_parasitics and rsz build parasitics on the die net, which delay calculation no longer looks up for a bonded net on a 3D-IC design. This PR is for parasitics read from SPEF. New test 3dic_spef. 3dic_cross.ok updated for the report_net change. Signed-off-by: dsengupta0628 <dsengupta@precisioninno.com>
There was a problem hiding this comment.
Code Review
This pull request introduces support for cross-die parasitics in 3DIC designs by mapping die-side nets to their corresponding chip-nets above the chiplet boundary. It implements helper methods to retrieve these chip-nets and updates network queries like net(), highestConnectedNet(), and highestNetAbove() to return the chip-net when applicable. A new regression test, 3dic_spef, is also added. The review feedback highlights a potential undefined behavior in staToDb() where a sibling class pointer is downcast directly to dbNet* to check its object type; casting to the common base class odb::dbObject* first is recommended to resolve this safely.
Calling getObjectType() through a dbNet* that may point at a dbChipNet is undefined behavior (sibling classes). Cast to the common dbObject base first, as staToDbChipNet() and staToDb(const Port*) already do. Signed-off-by: dsengupta0628 <dsengupta@precisioninno.com>
Signed-off-by: dsengupta0628 <dsengupta@precisioninno.com> # Conflicts: # src/dbSta/test/CMakeLists.txt
osamahammad21
left a comment
There was a problem hiding this comment.
LGTM. The logic is accepted with the known limitations listed in the description. This still needs a review from a dbSta code owner.
Summary
On 3D-IC designs,
read_spefdoes not merge the parasitics of a cross-die net. 3D RCX writes such a net as three SPEF pieces:read_spef -path <chip_inst>;Each piece lands on a different net, so a driver sees only its own die's wire. On master the bond resistance has no effect on timing, and cross-die paths come out optimistic.
This supersedes #11189, rebased onto current master.
Fix
All three pieces belong to the
dbChipNetabove a bonded boundary, which is also the net the bonds SPEF is written against. ThreedbNetworkoverrides return that chip net:net(const Pin*)on a chiplet boundary btermnullptrdbChipNetabovehighestConnectedNet()on a bonded die netdbChipNetabovehighestNetAbove()on a bonded die netdbChipNetabovenet()is howSpefReader::dspfBeginpicks the owning net of a-pathread, so it is what merges the die pieces into one network.highestConnectedNet()is howParasitics::findParasiticNetfinds a driver's parasitics, so a driver on either die finds the merged network.highestNetAbove()has to agree with it. OtherwiseensureParasiticNodemarks every node of the merged network as external.Supporting changes:
chipNetAbove(),chipNetAboveNet()andchipNetOfPadITerm().staToDb(const Net*)now returns null for a chip net instead of casting it to the wrong type.There is no new command, no Tcl or SWIG change, and no
src/stachange.Usage
read_3dbx top.3dbx read_spef top.bonds.spef ;# top scope, read FIRST read_spef -path die_a_0 top.die_a.spef read_spef -path die_b_0 top.die_b.spefThe bonds SPEF must be read first. A top-scope read replaces a net's parasitic network, while a
-pathread merges into it, so the other order silently discards the die pieces.Changes since #11189
const Net*API.chipA/bump_q:PAD), not the oldchipA:q. Timing is the same either way.visitConnectedPinschecks that the 3D-IC maps are up to date once per net, not once per iterm.Type of Change
Impact
3dic_speffixture (two dies; the bond resistance is swept):chipA/buf/ZThis matches an independent user report on #11189: 102.88 fF and 4.63 ns at 50 kΩ with the PR applied.
Non-3D-IC designs are unaffected, because
has3DicChip()is false for them.Known limitations
report_neton a bonded die net shows only part of the cross-die net. Tcl net reports go through OpenSTA's generic net walk. That walk now climbs to the chip net but does not descend into the other die, so the report lists the far die's bump pad but not its loads.3dic_cross.okis updated for this. Timing and SPEF reading usedbNetworkdirectly and are not affected.estimate_parasiticsand rsz build parasitics on the die net, and delay calculation no longer looks there for a bonded net on a 3D-IC design. This PR covers parasitics read from SPEF only.-pathparasitics are not stitched by this PR.Verification
./etc/Build.sh). Built with Bazel (bazel build //:openroad);./etc/Build.shnot run.bazel test //src/dbSta/test:all: 2555/2555 pass.3dic_spef: two dies, a 2 kΩ bond, and checks on the merged driver load and the cross-die arrival.Related Issues
dbInsts) indbNetwork::bumpPadITerm; see the comment there.