STA API¶
The STA is built in C++ with heavy use of STL (Standard Template Libraries). It also uses the zlib library to read compressed Liberty, Verilog, SDF, SPF, and SPEF files.
The sub-directories of the STA code are:
| Directory | Description |
|---|---|
doc |
Documentation files. |
util |
Basic utilities. |
liberty |
Liberty timing library classes and file reader. |
network |
Network and library API used by all STA code. |
verilog |
Verilog netlist reader that implements the network API. |
graph |
Timing graph built from network and library cell timing arcs. |
sdc |
SDC timing constraint classes. |
sdf |
SDF reader, writer and annotator. |
dcalc |
Delay calculator API and implementations. |
search |
Search engine used to annotate the graph with arrival, required times and find timing check slacks. |
parasitics |
Parasitics API, Spef and Spf readers. |
app |
Interface between Tcl and STA (built with SWIG). Main program definition. |
tcl |
User interface code. SDC argument parsing. |
Each sub-directory builds a library that is linked to build the STA executable or linked into another application.
The file Coding guidelines defines naming conventions used in the code.
Major components of the STA such as the network, timing graph, sdc,
and search are implemented as separate classes. The Sta class
contains an instance of each of these components.
The Sta class defines the bulk of the externally visible API used by
the Tcl interface, and coordinates operations that involve multiple
components. For example, when a false path command is entered into
the Tcl command interpreter, the Sta passes the declaration on to the
Sdc component and tells the Search component to invalidate all arrival
and required times.
Applications should call functions defined by the Sta class rather
than functions defined by the components. Calling functions defined
by the components will get you in trouble unless you understand them
in detail. For example, telling the delay calculator to recompute the
delays leaves the arrival times that depend on them wrong. Always
remember that the Sta coordinates the components.
In general, objects passed as arguments to Sta functions that are
constructors become "owned" by the STA and should not be deleted by
the caller. For example, a set of pins passed into
Sta::makeExceptionFrom are used in the resulting object (rather than
copied into another set). On the other hand, strings passed as
arguments are copied by the Sta functions before they are retained in
STA data structures.
In many cases the major components contain pointers to other
components. The StaState class is a simple container for these
components that makes initialization of pointers to the components
easier.
An STA with modified behavior can be built by defining classes derived
from the component classes and overloading some of the member
functions (which may have to be modified to be virtual). Components
are created by Sta::makeComponents(). The Sta::makeComponents()
function in turn calls each of the Sta::make<Component> component
constructors. These constructors can be overloaded by redefining them
in a class derived from Sta. Because the components refer to each
other, Sta::updateComponentsState() must be called to notify the
components if any of them are changed after creation.
Units¶
Units for values in Sta and liberty data structures are always the
following:
| Quantity | Unit |
|---|---|
| time | seconds |
| length | meters |
| capacitance | farads |
| resistance | ohms |
| voltage | volts |
| current | amperes |
| power | watts |
All file readers and the user interface are responsible for converting any user input or output to these units.
Utilities¶
The STA uses the C++ standard library containers (std::vector,
std::map, std::set, and related types). Helper functions for
lookups and iteration live in ContainerHelpers.hh (sta::findKey,
sta::findKeyValue, and range helpers). Prefer range-for over
hand-written iterators.
All printing done by the STA core is done using the Report class API.
The report class supports output redirection to a file and logging to
a file. The Tcl interpreter prints to "channels" that are
encapsulated by functions in the ReportTcl class. Printing inside
the STA is directed to the Tcl channels so that it appears with the
Tcl interpreter output.
Network¶
The network API is the key to making the STA a timing engine that can be bolted onto another application. This API allows the STA to efficiently communicate with external network data structures without the overhead of making and maintaining a copy of it.
The network API encapsulates both library and netlist accessors. Libraries are composed of cells that have ports that define connections to the cell. Netlists are built out of cell instances, pins and nets.
The ConcreteLibrary and ConcreteNetwork classes are used by the STA
netlist readers (notably Verilog). These class definitions are to
support a stand alone STA that does not depend on external netlist
data structures.
External network data structures are interfaced to the STA by casting pointers to network objects across the interface. The external objects do not have to be derived from STA network base classes. The network API functions are typically very thin functions that cast the STA network types to the external class types and call the corresponding external network database accessor.
Bus ports are expanded into ports for each bit in the bus, and iterators are provided for the expanded and unexpanded set of cell ports.
Network instances are calls of cells in the design hierarchy. Both hierarchical and leaf instances are in the network. Hierarchical instances have children instances at the next lower hierarchy level. Leaf instances have liberty cells with timing model data. At the top of the hierarchy is a top level instance that has instances for the top level netlist. If a cell has multiple instances the entire sub-tree of hierarchy is repeated in the network. This "unfolded" network representation allows optimization to specialize instances of a hierarchical block. A "folded" network representation that has only one sub-tree for each hierarchical block means that all copies must have identical sub-trees, preventing optimizations that specialize the contents.
Pins are a connection between an instance and a net corresponding to a port. For bus ports each bit in the bus has a corresponding pin (library iterators can be used to find the pins that correspond to all of the bits in a bus). Ports on the top level instance also have pins in the network that are the top level inputs and outputs.
Nets connect together a group of pins. Both hierarchical and leaf pins are on a net. Nets can connect pins on multiple levels of hierarchy.
The network objects inside the STA are always pointers to instances of undefined class objects. The implementation and definition of the network objects themselves is never visible inside the STA. The network API is implemented as an adapter that performs all operations on all network objects. There is one network adapter instance used by all STA code. For example, to find the cell of an instance:
Cell *cell = network->cell(instance);
The network adapter returns iterators for looping over groups of network objects. For example, the following code iterates over the children of the top level instance.
Instance *top_instance = network->topInstance();
InstanceChildIterator *child_iter = network->childIterator(top_instance);
while (child_iter->hasNext()) {
Instance *child = child_iter->next();
...
}
delete child_iter;
An adapter to a network database is built by defining a class derived
from the base class Network, or NetworkEdit if it supports incremental
editing operations. network/ConcreteNetwork.cc is an example of
a network adapter that supports hierarchy. An example of a network adapter
for a flat DEF based netlist is in OpenROAD:
dbNetwork.hh,
dbNetwork.cc.
A network adaptor to interface to an external network database must
define the virtual functions of the Network class (about 45
functions). The external network objects do not have to use any STA
network objects as base classes or even be C++ objects. These network
adapter functions should cast the network object pointers to the
underlying network object.
Network adapters built on the Network class must define the following
functions to find corresponding liberty objects.
virtual LibertyLibrary *libertyLibrary(Library *library) const;
virtual LibertyLibrary *makeLibertyLibrary(const char *name,
LibraryAnalysisPt *ap);
virtual LibertyCell *libertyCell(Cell *cell) const;
virtual LibertyPort *libertyPort(Port *port) const;
The NetworkLiberty class provides implementations of the first two
functions for derived network classes.
If the network adapter implements the NetworkEdit API the following
Tcl commands are supported:
make_cell
replace_cell
delete_cell
make_net
delete_net
connect_pins
disconnect_pins
Each of these commands call corresponding functions in network/NetworkEdit.i
that notify the Sta before and/or after the network operation is
performed.
Liberty¶
The liberty timing library reader builds classes that are derived from the concrete library classes. In addition to the library, cell and port classes, there are classes to represent timing arcs, timing models, wireload models, operating conditions, and scale factors for derating timing data.
Timing arcs are grouped into sets of arcs between a pair of cell ports. For example, a buffer has two timing arcs between the input and output; one for a rising output and another for a falling output. The timing arcs are:
A r -> Z r
A f -> Z f
Since a buffer is non-inverting, the timing arc set is positive-unate. Similarly, an inverter has two negative-unate timing arcs.
A f -> Z r
A r -> Z f
On the other hand, a multiplexor, has a non-unate path from the select input to the output because a rise or fall change on the input can cause the output to either rise or fall. There are four timing arcs in this arc set:
S f -> Z r
S f -> Z f
S r -> Z r
S r -> Z f
The liberty file reader can be customized to read attributes that are not used by the STA.
Graph¶
The timing graph is the central data structure used by the delay calculation and search algorithms. It is annotated with timing arc delay values and slews (from SDF or a delay calculator). A forward search annotates the graph with arrival times, and a backward search annotates required times.
The graph is composed of vertices and edges. Each pin in the design has a vertex. Bidirect pins have two vertices, one for its use as an input and another for its use as an output.
The Network adapter supplies functions to find and set the index
(unsigned) of a graph vertex corresponding to a pin.
Network::vertexIndex(const Pin *pin) const;
Network::setVertexIndex(Pin *pin, VertexIndex index);
An STL map can be used for the lookup, but it is rather memory hungry compared to storing the value in the pin structure.
A pointer to the vertex used for a bidirectional pin driver is kept in
a map owned by the Graph class.
Edges in the graph connect vertices. The pins connected together by a net have wire edges between the pin vertices. Timing arc sets in the leaf instance timing models have corresponding edges in the graph between pins on the instance.
SDC¶
There is no support for updating SDC when network edits delete the instance, pin, or net objects referred to by the SDC.
Delay Calculation¶
The graph is annotated with arc delay values and slews (also known as
transition times) by the graph delay calculator or the SDF reader.
The GraphDelayCalc class seeds slews from SDC constraints and uses a
breadth first search to visit each gate output pin. The GraphDelayCalc
then calls a timing arc delay calculator for each timing arc and
annotates the graph arc delays and vertex slews.
The delay calculator is architected to support multiple delay
calculation results. Each result has an associated delay calculation
analysis point (class DcalcAnalysisPt) that specifies the operating
conditions and parasitics used to find the delays.
The ArcDelayCalc class defines the API used by the GraphDelayCalc to
calculate the gate delay, driver slew, load delays and load slews
driven by a timing arc. The following delay calculation algorithms
are defined in the dcalc directory.
UnitDelayCalc¶
All gate delays are 1. Wire delays are zero.
LumpedCapArcDelayCalc¶
Liberty table models using lumped capacitive load (RSPF pi model total capacitance). Wire delays are zero.
DmpCeffElmoreDelayCalc¶
RSPF (Driver Pi model with elmore interconnect delays) delay calculator. Liberty table models using effective capacitive model as described in "Performance Computation for Precharacterized CMOS Gates with RC Loads", Florentin Dartu, Noel Menezes and Lawrence Pileggi, IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, Vol 15, No 5, May 1996. Wire delays are computed by applying the driver waveform to the RSPF dependent source and solving the RC network.
DmpCeffTwoPoleDelayCalc¶
Driver Pi model with two pole interconnect delays and effective
capacitance as in DmpCeffElmoreDelayCalc.
Other delay calculators can be interfaced by defining a class based on
ArcDelayCalc and using the registerDelayCalc function to register it
for the set_delay_calculator Tcl command. The Sta::setArcDelayCalc
function can be used to set the delay calculator at run time.
Search¶
A breadth first forward search is used to find arrival times at graph
vertices. Vertices are annotated with instances of the Event class to
record signal arrival and required times. As each vertex is visited
in the forward search its required time is found using If the vertex
is constrained by setup or hold timing checks, min/max path delay
exceptions or gated timing checks its required time is found from the
SDC. The slack is the difference between the vertex required time and
arrival time. If the vertex is constrained it is scheduled for a
breadth first backward search to propagate required times to the fanin
vertices. Separate events (and hence arrival and required times) are
used for each clock edge and exception set that cause a vertex to
change.
Arrival, required and slack calculations are incremental using a level based "lazy evaluation" algorithm. The first time arrival/required times are found for a vertex the arrival/required times are propagated to/from the vertex's logic level. After that no search is required for any vertex with a lower/higher logic level when the arrival/required time is requested.
Clock arrival times are found before data arrival times by
Search::findClkArrivals(). Clock arrival times include insertion delay
(source latency).
When an incremental netlist change is made (for instance, changing the
drive strength of a gate with swap_cell), the STA incrementally updates
delay calculation, arrival times, required times and slacks. Because
gate delay is only weakly dependent on slew, the effect of the change
will diminish in gates downstream of the change. The STA uses a
tolerance on the gate delays to determine when to stop propagating the
change. The tolerance is set using the
Sta::setIncrementalDelayTolerance function.
void Sta::setIncrementalDelayTolerance(float tol);
The tolerance is a percentage (0.0:1.0) change in delay that causes downstream delays to be recomputed during incremental delay calculation. The default value is 0.0 for maximum accuracy and slowest incremental speed. The delay calculation will not recompute delays for downstream gates when the change in the gate delay is less than the tolerance. Required times must be recomputed backward from any gate delay changes, so increasing the tolerance can significantly reduce incremental timing run time.
Tcl Interface¶
The interface from Tcl to C++ is written in a SWIG
(www.swig.org) interface description
(app/StaApp.i). SWIG generates the interface code from the
description file.
All commands are written in Tcl. SDC argument parsing and checking is done with Tcl procedures that call a SWIG interface function.
The Tcl sta namespace is used to segregate internal STA functions
from the global Tcl namespace. All user visible STA and SDC commands
are exported to the global Tcl namespace.
A lot of the internal STA state can be accessed from Tcl to make
debugging easier. Some debugging commands are not intended for casual
users and live in the sta namespace. Others, such as report_arrival,
report_required, report_slack, and report_edges, are also exported
to the global namespace. Examples:
report_arrival
report_required
report_slack
report_edges
report_slews
sta::report_level pin
sta::report_constant pin|instance
sta::report_network
sta::network_pin_count
sta::network_net_count
sta::network_leaf_instance_count
sta::network_leaf_pin_count
Additionally, many of the STA network and graph objects themselves are
exposed to Tcl using SWIG. These Tcl objects have methods for
inspecting them. Examples of how to use these methods can be found in
the tcl/Graph.tcl and tcl/Network.tcl files.
Architecture alternatives for using the STA Engine¶
There are a number of alternatives for using the STA engine with an application.
STA with Tcl application¶
The simplest example is an application written in Tcl. The application
calls STA commands and primitives defined in the SWIG C++/Tcl
interface. A stand-alone STA executable is built and a Tcl file that
defines the application is included as part of the STA by modifying
CMakeLists.txt to add the Tcl file to STA_TCL_FILES (encoded into
StaTclInitVar.cc).
The user calls STA commands to read design files (liberty, verilog, SDF, parasitics) to define and link the design. The user defines SDC commands or sources an SDC file. The user calls the application's Tcl commands.
A simple gate sizer is an example of an application that can be built this way because it has very little computation in the sizer itself. STA Tcl commands can be used to find the worst path and upsize gates or insert buffers.
STA with C++ application¶
The application is built by adding C++ files to the app directory and
modifying CMakeLists.txt to include them in the executable. Interface
commands between C++ and Tcl are put in a SWIG .i file in the app
directory and modifying app/StaApp.i to include them. Tcl commands are
added to the STA by modifying CMakeLists.txt to add the application's
Tcl files to STA_TCL_FILES.
The user calls STA commands to read design files (liberty, verilog, SDF, parasitics) to define and link the design. The user defines SDC commands or sources an SDC file. The user calls the application's Tcl commands.
C++ application without native Network data structures linking STA libraries¶
The application builds main() and links STA libraries. On startup it
calls STA initialization functions like staMain() defined in
app/StaMain.cc.
The application must link and instantiate a Tcl interpreter to read
SDC commands like staMain(). The application can choose to expose the Tcl
interpreter to the user or not. The STA depends on the following data
that can be read by calling Tcl commands or Sta class member functions.
Liberty files that define the leaf cells used in the design.
Read using the read_liberty command or by calling Sta::readLibertyFile().
Verilog files that define the netlist. Read using the read_verilog
command or by calling readVerilogFile() (see verilog/Verilog.i
read_verilog).
Link the design using the link_design command or calling Sta::linkDesign().
SDC commands to define timing constraints.
Defined using SDC commands in the Tcl interpreter, or sourced
from a file using Tcl_Eval(sta::tclInterp()).
Parasitics used by delay calculation.
Read using the read_spef command, Sta::readSpef(), or
using the Sta::Parasitics class API.
The application calls network editing functions such as
Sta::deleteInstance() to edit the network.
C++ application with native Network data structures linking STA libraries¶
The application defines a Network adapter (described above) so that
the STA can use the native network data structures without duplicating
them in the STA. The application defines a class built on class Sta
that defines the makeNetwork() member function to build an instance of
the network adapter.
The application builds main() and links STA libraries. On startup it
calls STA initialization functions like staMain() defined in
app/StaMain.cc. The application reads the netlist and builds network
data structures that the STA accesses through the Network adapter.
The application must link and instantiate a Tcl interpreter to read
SDC commands like staMain(). The application can choose to expose the Tcl
interpreter to the user or not. The STA depends on the following data
that can be read by calling Tcl commands or Sta class member functions.
Liberty files that define the leaf cells used in the design.
Read using the read_liberty command or by calling Sta::readLibertyFile.
SDC commands to define timing constraints.
Defined using SDC commands in the Tcl interpreter, or sourced
from a file using sta::sourceTclFile.
Parasitics used by delay calculation.
Read using the read_spef command, Sta::readSpef(), or
using the Sta::Parasitics class API.
The application calls network editing before/after functions such as
Sta::deleteInstanceBefore() to notify the Sta of network edits.
A placement tool is likely to use this pattern to integrate the STA because the DEF file includes netlist connectivity.