Universal Verification Methodology (UVM) 1.2 User’s Guide ¶

March 23, 2020

Mandatory copyrights:

*Copyright © 2011 - 2015 Accellera. All rights reserved.*

Copyright © 2011 - 2015 Accellera Systems Initiative
(Accellera). All rights reserved.

Accellera Systems Initiative, 8698 Elk Grove Bldv Suite 1, #114, Elk
Grove, CA 95624, USA

Copyright © 2013 - 2015 Advanced Micro Devices, Inc. All
rights reserved.

Advanced Micro Devices (AMD), 7171 Southwest Parkway, Austin, TX
78735

Copyright © 2011 - 2015 Cadence Design Systems,
Inc. (Cadence). All rights reserved.

Cadence Design Systems, Inc., 2655 Seely Ave., San Jose, CA 95134,
USA.

Copyright © 2011 - 2015 Mentor Graphics, Corp. (Mentor). All
rights reserved.

Mentor Graphics, Corp., 8005 SW Boeckman Rd., Wilsonville, OR 97070,
USA

Copyright © 2013 - 2015 NVIDIA CORPORATION. All rights
reserved.

NVIDIA Corporation, 2701 San Tomas Expy, Santa Clara, CA 95050

Copyright © 2011 - 2015 Synopsys, Inc. (Synopsys). All rights
reserved.

Synopsys, Inc., 690 E. Middlefield Rd, Mountain View, CA 94043

Copyright © 2019 - 2023 Tuomas Poikela. All rights reserved.

This product is licensed under the Apache Software Foundation’s Apache License, Version 2.0, January 2004. The full license is available at: http://www.apache.org/licenses/.

Warning

This is semi-manual translation from the original SystemVerilog UVM 1.2 User's Guide to uvm-python. It is work-in-progress. It’s a work-in-progress and contains still a lot of errors and references to SystemVerilog facilities. This will be gradually fixed over time. Feel free to contribute to this userguide with a pull request. Even small fixes will help in making it better!

Notices

While this guide offers a set of instructions to perform one or more specific verification tasks, it should be supplemented by education, experience, and professional judgment. Not all aspects of this guide may be applicable in all circumstances. The UVM 1.2 User’s Guide does not necessarily represent the standard of care by which the adequacy of a given professional service must be judged nor should this document be applied without consideration of a project’s unique aspects. The original SystemVerilog version of this guide has been approved through the Accellera consensus process and serves to increase the awareness of information and approaches in verification methodology. This guide may have several recommendations to accomplish the same thing and may require some judgment to determine the best course of action.

The uvm-python Class Reference represents the foundation used to create the UVM 1.2 User’s Guide. This guide is a way to apply the UVM 1.2 Class Reference, but is not the only way. Standards are an important ingredient to foster innovation and continues to encourage industry innovation based on its standards.

Table of Contents

Universal Verification Methodology (UVM) 1.2 User’s Guide

0. Getting started ¶

uvm-python uses cocotb Makefiles to launch the simulations and to control the simulation arguments. It is also possible to use cocotb-test instead of Makefiles. See README.md in the github repository for a minimal example. You also need an HDL simulator. Verilator and Icarus Verilog are freely available and can be used with uvm-python.

0.1 Installation ¶

You can install uvm-python as a normal Python package. It is recommended to use [venv](https://docs.python.org/3/library/venv.html) to create a virtual environment for Python prior to installation:

git clone https://github.com/tpoikela/uvm-python.git
cd uvm-python
python -m pip install . # If venv is used
# Or without venv, and no sudo:
python -m pip install --user .  # Omit --user for global installation

See [Makefile](test/examples/simple/Makefile) for working examples. You can also use Makefiles in test/examples/simple as a template for your project.

First example ¶

uvm-python must be installed prior to running the example. Alternatively, you can create a symlink to the uvm source folder:

cd test/examples/minimal
ln -s ../../../src/uvm uvm

You can find the source code for this example in test/examples/minimal. This example creates a test component, registers it with the UVM factory, and starts the test.

You can execute the example by running SIM=icarus make. Alternatively, you can run it with SIM=verilator make:

# File: Makefile
TOPLEVEL_LANG ?= verilog
VERILOG_SOURCES ?= new_dut.sv
TOPLEVEL := new_dut
MODULE   ?= new_test
include $(shell cocotb-config --makefiles)/Makefile.sim

The top-level module must match TOPLEVEL in Makefile:

// File: new_dut.sv
module new_dut(input clk, input rst, output[7:0] byte_out);
    assign byte_out = 8'hAB;
endmodule: new_dut

The Python test file name must match MODULE in Makefile:

# File: new_test.py
import cocotb
from cocotb.triggers import Timer
from uvm import *

class NewTest(UVMTest):

    async def run_phase(self, phase):
        phase.raise_objection(self)
        await Timer(100, "NS")
        phase.drop_objection(self)

uvm_component_utils(NewTest)

@cocotb.test()
async def test_dut(dut):
    await run_test('NewTest')

To run the example, execute:

SIM=icarus make

1. Overview ¶

This chapter provides a quick overview of UVM by going through the typical testbench architecture and introducing the terminology used throughout this User’s Guide. Then, it also touches upon the UVM Base Class Library (BCL) developed by Accellera.

1.1 The Typical UVM Testbench Architecture ¶

The UVM Class Library provides generic utilities, such as component hierarchy, transaction library model (TLM), configuration database, etc., which enable the user to create virtually any structure he/she wants for the testbench. This section provides some broad guidelines for a type of recommended testbench architecture by using one viewpoint of this architecture in an effort to introduce the terminology used throughout this User’s Guide, as shown in Figure 1.

NB: This particular representation is not intended to represent all possible testbench architectures or every type use of UVM in the electronic design automation (EDA) industry.

alternate text — Typical UVM Testbench Architecture¶

1.1.1 UVM Testbench¶

The UVM Testbench is simply a Python function decorated with @cocotb.test. The Design under Test (DUT) module is passed to the testbench as an object, and no explicit signal connections have to be made by the user. This is an advantage over SystemVerilog-based testbenches where DUT-module has to be instantiated and signal connections have to be made explicitly. The UVM Test class can be declared in a separate file and then imported into the testbench. If the verification collaterals include module-based components, they should be instantiated in a wrapper HDL module outside Python code. The UVM Test is dynamically instantiated at run-time, allowing the an HDL DUT to be compiled/elaborated once, and run with many different tests.

Note that in some architectures, the Testbench term is used to refer to a special module that encapsulates verification collaterals only, which in turn are integrated up with the DUT.

1.1.2 UVM Test¶

The UVMTest is the top-level UVM Component in the UVM Testbench. The UVM Test typically performs three main functions: Instantiates the top-level environment, configures the environment (via configuration objects, factory overrides or the configuration database), and applies stimulus by invoking UVM Sequences through the environment to the DUT.

Typically, there is one base UVM Test with the UVM Environment instantiation and other common items. Then, other individual tests will extend this base test and configure the environment differently or select different sequences to run.

1.1.3 UVM Environment¶

The UVM Environment is a hierarchical component that groups together other verification components that are interrelated. Typical components that are usually instantiated inside the UVM Environment are UVM Agents, UVM Scoreboards, or even other UVM Environments. The top-level UVM Environment encapsulates all the verification components targeting the DUT.

For example: In a typical system on a chip (SoC) UVM Environment, you will find one UVM Environment per IP (e.g., PCIe Environment, USB Environment, Memory Controller Environment, etc.). Sometimes, those IP Environments are grouped together into Cluster Environments or subsystem environments (e.g., IO Environment, Processor Environment, etc.) that are grouped together eventually in the top-level SoC Environment.

1.1.4 UVM Scoreboard¶

The UVM Scoreboard’s main function is to check the behavior of a certain DUT. The UVM Scoreboard usually receives transactions carrying inputs and outputs of the DUT through UVM Agent analysis ports (connections are not depicted in Figure 1), runs the input transactions through some kind of a reference model (also known as the predictor) to produce expected transactions, and then compares the expected output versus the actual output.

There are different methodologies on how to implement the scoreboard, the nature of the reference model, and how to communicate between the scoreboard and the rest of the testbench.

1.1.5 UVM Agent¶

The UVM Agent is a hierarchical component that groups together other verification components that are dealing with a specific DUT interface (see Figure 2). A typical UVM Agent includes a UVM Sequencer to manage stimulus flow, a UVM Driver to apply stimulus on the DUT interface, and a UVM Monitor to monitor the DUT interface. UVM Agents might include other components, like coverage collectors, protocol checkers, a TLM model, etc.

The UVM Agent needs to operate both in an active mode (where it is capable of generating stimulus) and a passive mode (where it only monitors the interface without controlling it).

1.1.6 UVM Sequencer¶

The UVM Sequencer serves as an arbiter for controlling transaction flow from multiple stimulus sequences. More specifically, the UVM Sequencer controls the flow of UVM Sequence Items transactions generated by one or more UVM Sequences.

1.1.7 UVM Sequence¶

A UVM Sequence is an object that contains a behavior for generating stimulus. UVM Sequences are not part of the component hierarchy. UVM Sequences can be transient or persistent. A UVM Sequence instance can come into existence for a single transaction, it may drive stimulus for the duration of the simulation, or anywhere in-between. UVM Sequences can operate hierarchically with one sequence, called a parent sequence, invoking another sequence, called a child sequence.

To operate, each UVM Sequence is eventually bound to a UVM Sequencer. In practice this means that UVMSequence.start receives the sequencer as an argument. Multiple UVM Sequence instances can be bound to the same UVM Sequencer. Top-level sequences do not have to run on a sequencer, and can start other sequences instead.

1.1.8 UVM Driver¶

The UVM Driver receives individual UVM Sequence Item transactions from the UVM Sequencer and applies (drives) it on the DUT Interface. Thus, a UVM Driver spans abstraction levels by converting transaction-level stimulus into pin-level stimulus. It also has a TLM port to receive transactions from the Sequencer and access to the DUT interface in order to drive the signals.

1.1.9 UVM Monitor¶

The UVM Monitor samples the DUT interface and captures the information there in transactions that are sent out to the rest of the UVM Testbench for further analysis. Thus, similar to the UVM Driver, it spans abstraction levels by converting pin-level activity to transactions. In order to achieve that, the UVM Monitor typically has access to the DUT interface and also has a TLM analysis port to broadcast the created transactions through.

The UVM Monitor can perform internally some processing on the transactions produced (such as coverage collection, checking, logging, recording, etc.) or can delegate that to dedicated components connected to the monitor’s analysis port.

1.2 The UVM Class Library ¶

The UVM Class Library provides all the building blocks you need to quickly develop well-constructed, reusable, verification components and test environments. The library consists of base classes, utilities, and macros. Figure 3 shows a subset of those classes.

Components may be encapsulated and instantiated hierarchically and are controlled through an extendable set of phases to initialize, run, and complete each test. These phases are defined in the base class library, but can be extended to meet specific project needs. See the UVM 1.2 Class Reference for more information.

The advantages of using the UVM Class Library include:

a) A robust set of built-in features—The UVM Class Library provides many features that are required for verification, including complete implementation of printing, copying, test phases, factory methods, and more. b) Correctly-implemented UVM concepts—Each component in the block diagram in Figure 1 and Figure 2 can be derived from a corresponding UVM Class Library component. Using these baseclass elements increases the readability of your code since each component’s role is predetermined by its parent class.

The UVM Class Library also provides various utilities to simplify the development and use of verification environments. These utilities support configurability by providing a standard resource sharing database. They support debugging by providing a user-controllable messaging utility for failure reporting and general reporting purposes. They support testbench construction by providing a standard communication infrastructure between verification components (TLM) and flexible verification environment construction (UVM factory). Finally, they also provide mixins for allowing more compact coding styles. These mixins offer similar capabilities as macros in the SystemVerilog (SV) version.

This User’s Guide will touch on most of these utilities; for the complete list, see the UVM 1.2 Class Reference.

2. Transaction-Level Modeling (TLM)¶

2.1 Overview ¶

One of the keys to verification productivity is to think about the problem at a level of abstraction that makes sense. When verifying a DUT that handles packets flowing back and forth, or processes instructions, or performs other types of functionality, you must create a verification environment that supports the appropriate abstraction level. While the actual interface to the DUT ultimately is represented by signal-level activity, experience has shown that it is necessary to manage most of the verification tasks, such as generating stimulus and collecting coverage data, at the transaction level, which is the natural way engineers tend to think of the activity of a system.

UVM provides a set of transaction-level communication interfaces and channels that you can use to connect components at the transaction level. The use of TLM interfaces isolates each component from changes in other components throughout the environment. When coupled with the phased, flexible build infrastructure in UVM, TLM promotes reuse by allowing any component to be swapped for another, as long as they have the same interfaces. This concept also allows UVM verification environments to be assembled with a transaction-level model of the DUT, and the environment to be reused as the design is refined to RTL. All that is required is to replace the transaction-level model with a thin layer of compatible components to convert between the transaction-level activity and the pin-level activity at the DUT.

The well-defined semantics of TLM interfaces between components also provide the ideal platform for implementing mixed-language verification environments. In addition, TLM provides the basis for easily encapsulating components into reusable components, called verification components, to maximize reuse and minimize the time and effort required to build a verification environment.

This chapter discusses the essential elements of transaction-level communication in UVM, and illustrates the mechanics of how to assemble transaction-level components into a verification environment. Later in this document we will discuss additional concerns in order to address a wider set of verification issues. For now, it is important to understand these foundational concepts first.

2.2 TLM, TLM-1, and TLM-2.0 ¶

TLM, transaction-level modeling, is a modeling style for building highly abstract models of components and systems. It relies on transactions (see Section 2.3.1.1), objects that contain arbitrary, protocol-specific data to abstractly represent lower-level activity. In practice, TLM refers to a family of abstraction levels beginning with cycle-accurate modeling, the most abstract level, and extending upwards in abstraction as far as the eye can see. Common transaction-level abstractions today include: cycle-accurate, approximately-timed, loosely-timed, untimed, and token-level.

The acronym TLM also refers to a system of code elements used to create transaction-level models. TLM-1 and TLM-2.0 are two TLM modeling systems which have been developed as industry standards for building transaction-level models. Both were built in SystemC and standardized within the TLM Working Group of the Open SystemC Initiative (OSCI). TLM-1 achieved standardization in 2005 and TLM-2.0 became a standard in 2009. OSCI merged with Accellera in 2013 and the current SystemC standard used for reference is IEEE 1666-2011.

TLM-1 and TLM-2.0 share a common heritage and many of the same people who developed TLM-1 also worked on TLM-2.0. Otherwise, they are quite different things. TLM-1 is a message passing system. Interfaces are either untimed or rely on the target for timing. None of the interfaces provide for explicit timing annotations. TLM-2.0, while still enabling transfer of data and synchronization between independent processes, is mainly designed for high performance modeling of memory-mapped bus-based systems. A subset of both these facilities has been implemented in uvm-python and is available as part of UVM.

2.3 TLM-1 Implementation ¶

The following subsections specify how TLM-1 is to be implemented in uvm-python.

2.3.1 Basics
2.3.1.1 Transactions
2.3.1.2 Transaction-Level Communication
2.3.1.3 Basic TLM Communication
2.3.1.4 Communicating between Processes
2.3.1.5 Blocking versus Nonblocking
2.3.1.6 Connecting Transaction-Level Components
2.3.1.7 Peer-to-Peer connections
2.3.1.8 Port/Export Compatibility
2.3.2 Encapsulation and Hierarchy
2.3.2.1 Hierarchical Connections
2.3.2.2 Connection Types
2.3.3 Analysis Communication
2.3.3.1 Analysis Ports
2.3.3.2 Analysis Exports

2.3.1 Basics ¶

Before you can fully understand how to model verification at the transaction level, you must understand what a transaction is.

2.3.1.1 Transactions ¶

In UVM, a transaction is a class object that includes whatever information is needed to model a unit of communication between two components. In the most basic example, a simple bus protocol transaction to transfer information would be modeled as follows:

class simple_trans(UVMSequenceItem);
    def __init__(self, name):
        super().__init__(name)
        self.data = 0
        self.rand('data')
        self.addr = 0x0
        self.rand('addr')
        self.kind = WRITE
        self.rand('kind', [WRITE, READ])
        self.constraint(lambda addr: addr < 0x2000)

The transaction object includes variables, constraints, and other fields and methods necessary for generating and operating on the transaction. Obviously, there is often more than just this information that is required to fully specify a bus transaction. The amount and detail of the information encapsulated in a transaction is an indication of the abstraction level of the model. For example, the simple_trans transaction above could be extended to include more information, such as the number of wait states to inject, the size of the transfer, or any number of other properties. The transaction could also be extended to include additional constraints. It is also possible to define higher-level transactions that include some number of lower-level transactions. Transactions can thus be composed, decomposed, extended, layered, and otherwise manipulated to model whatever communication is necessary at any level of abstraction.

2.3.1.2 Transaction-Level Communication ¶

Transaction-level interfaces define a set of methods that use transaction objects as arguments. A TLM port defines the set of methods (the application programming interface (API)) to be used for a particular connection, while a TLM export supplies the implementation of those methods. Connecting a port to an export allows the implementation to be executed when the port method is called.

2.3.1.3 Basic TLM Communication ¶

The most basic transaction-level operation allows one component to put a transaction to another. Consider Figure 4.

The square box on the producer indicates a port and the circle on the consumer indicates the export. The producer generates transactions and sends them out its put_port:

class producer (UVMComponent);

    def __init__(self, name, parent):
        super().__init__(name, parent)
        self.put_port = UVMBlockingPutPort(“put_port”, self)
        ...

    async def run(self):
        for _ in range(N):
            t = simple_trans()
            # Generate t.
            await self.put_port.put(t);

NOTE—The uvm_*_port is parameterized by the transaction type that will be communicated. This may either be specified directly or it may be a parameter of the parent component.

The actual implementation of the put() call is supplied by the consumer:

class consumer(UVMComponent):

    def __init__(self, name, parent):
        super().__init__(name, parent)
        self.put_export = UVMBlockingPutImp(“put_export”, self)

    # 2 parameters ...
    async def put(self, t):
        if t.kind == READ:
            # Do read.
        elif t.kind == WRITE:
            # Do write.

NOTE—The UVM*Imp takes two parameters: the type of the transaction and the type of the object that declares the method implementation.

NOTE—The semantics of the put operation are defined by TLM. In this case, the put() call in the producer will block until the consumer’s put implementation is complete. Other than that, the operation of producer is completely independent of the put implementation (UVMPutImp). In fact, consumer could be replaced by another component that also implements put and producer will continue to work in exactly the same way. The modularity provided by TLM fosters an environment in which components may be easily reused since the interfaces are well defined.

The converse operation to put is get. Consider Figure 5.

In this case, the consumer requests transactions from the producer via its get port:

class get_consumer (UVMComponent):
    def __init__(self, name, parent):
        super().__init__(name, parent)
        self.get_port = UVMBlockingGetPort(“get_port”, self)
        ...

    async def run_phase(self, phase):
        for _ in range(N):
            # Generate t.
            t = simple_trans()
            await self.get_port.get(t);

The get() implementation is supplied by the producer:

class get_producer (UVMComponent):

    def __init__(self, name, parent):
        super().__init__(name, parent)
        self.get_export = UVMBlockingGetImp("get_export", self)

    async def get(self, t):
        tmp = simple_trans()
        // Assign values to tmp.
        t.copy(tmp)

As with put() above, the get_consumer’s get() call will block until the get_producer’s method completes. In TLM terms, put() and get() are blocking methods.

NOTE — In both these examples, there is a single process running, with control passing from the port to the export and back again. The direction of data flow (from producer to consumer) is the same in both examples.

2.3.1.4 Communicating between Processes ¶

In the basic put example above, the consumer will be active only when its put() method is called. In many cases, it may be necessary for components to operate independently, where the producer is creating transactions in one process while the consumer needs to operate on those transactions in another. UVM provides the UVMTLMFIFO channel to facilitate such communication. The UVMTLMFIFO implements all of the TLM interface methods, so the producer puts the transaction into the UVMTLMFIFO, while the consumer independently gets the transaction from the fifo, as shown in Figure 6.

When the producer puts a transaction into the fifo, it will block if the fifo is full, otherwise it will put the object into the fifo and return immediately. The get operation will return immediately if a transaction is available (and will then be removed from the fifo), otherwise it will block until a transaction is available. Thus, two consecutive get() calls will yield different transactions to the consumer. The related peek() method returns a copy of the available transaction without removing it. Two consecutive peek() calls will return copies of the same transaction.

2.3.1.5 Blocking versus Nonblocking ¶

The interfaces that we have looked at so far are blocking—the tasks block execution until they complete; they are not allowed to fail. There is no mechanism for any blocking call to terminate abnormally or otherwise alter the flow of control. They simply wait until the request is satisfied. In a timed system, this means that time may pass between the time the call was initiated and the time it returns.

In contrast, a nonblocking call returns immediately. The semantics of a nonblocking call guarantee that the call returns in the same delta cycle in which it was issued, that is, without consuming any time, not even a single delta cycle. In UVM, nonblocking calls are modeled as functions:

class consumer (UVMComponent):

    def __init__(self, name, parent):
        super().__init__(name, parent)
        self.get_port = UVMGetPort("get_port", self)

    async def run_phase(self phase):
        ...
        for _ in range(10):
            t = []
            if (self.get_port.try_get(t)) # Do something with t[0]
            ...

If a transaction exists, it will be returned in the argument and the function call itself will return TRUE. If no transaction exists, the function will return FALSE. Similarly, with try_peek(). The try_put() method returns TRUE if the transaction is sent.

2.3.1.6 Connecting Transaction-Level Components ¶

With ports and exports defined for transaction-level components, the actual connection between them is accomplished via the connect() method in the parent (component or env), with an argument that is the object (port or export) to which it will be connected. In a verification environment, the series of connect() calls between ports and exports establishes a netlist of peer-to-peer and hierarchical connections, ultimately terminating at an implementation of the agreed-upon interface. The resolution of these connections causes the collapsing of the netlist, which results in the initiator’s port being assigned to the target’s implementation. Thus, when a component calls:

put_port.put(t)

the connection means that it actually calls:

target.put_export.put(t)

where target is the connected component.

2.3.1.7 Peer-to-Peer connections ¶

When connecting components at the same level of hierarchy, ports are always connected to exports. All connect() calls between components are done in the parent’s connect() method:

class my_env(UVMEnv):

    ...
    def connect_phase(self, phase);
        # component.port.connect(target.export);
        self.producer.blocking_put_port.connect(fifo.put_export);
        self.get_consumer.get_port.connect(fifo.get_export);
        ...

2.3.1.8 Port/Export Compatibility ¶

Another advantage of TLM communication in UVM is that all TLM connections are checked for compatibility before the test runs. In order for a connection to be valid, the export must provide implementations for at least the set of methods defined by the port and the transaction type parameter for the two must be identical. For example, a blocking_put_port, which requires an implementation of put() may be connected to either a blocking_put_export or a put_export. Both exports supply an implementation of put(), although the put_export also supplies implementations of try_put() and can_put().

2.3.2 Encapsulation and Hierarchy ¶

The use of TLM interfaces isolates each component in a verification environment from the others. The environment instantiates a component and connects its ports/exports to its neighbor(s), independent of any further knowledge of the specific implementation. Smaller components may be grouped hierarchically to form larger components. Access to child components is achieved by making their interfaces visible at the parent level. At this level, the parent simply looks like a single component with a set of interfaces on it, regardless of its internal implementation.

2.3.2.1 Hierarchical Connections ¶

Making connections across hierarchical boundaries involves some additional issues, which are discussed in this section. Consider the hierarchical design shown in Figure 7.

The hierarchy of this design contains two components, producer and consumer. producer contains three components, stim, fifo, and conv. consumer contains two components, fifo and drv. Notice that, from the perspective of top, the producer and consumer appear identical to those in Figure 4, in which the producer’s put_port is connected to the consumer’s put_export. The two fifos are both unique instances of the same UVMTLMFIFO component.

In Figure 7, connections A, B, D, and F are standard peer-to-peer connections as discussed above. As an example, connection A would be coded in the producer’s connect() method as:

gen.put_port.connect(fifo.put_export);

Connections C and E are of a different sort than what have been shown. Connection C is a port-to-port connection, and connection E is an export-to-export connection. These two kinds of connections are necessary to complete hierarchical connections. Connection C imports a port from the outer component to the inner component. Connection E exports an export upwards in the hierarchy from the inner component to the outer one. Ultimately, every transaction-level connection must resolve so that a port is connected to an export. However, the port and export terminals do not need to be at the same place in the hierarchy. We use port-to-port and export-to-export connections to bring connectors to a hierarchical boundary to be accessed at the next-higher level of hierarchy.

For connection E, the implementation resides in the fifo and is exported up to the interface of consumer. All export-to-export connections in a parent component are of the form:

export.connect(subcomponent.export)

so connection E would be coded as:

class consumer(UVMComponent):
    put_export = uvm_put_export(trans)
    fifo = uvm_tlm_fifo(trans)

    def connect_phase(self, phase):
        super().connect_phase(phase)
        self.put_export.connect(self.fifo.put_export)
        self.bfm.get_port.connect(self.fifo.get_export)

Conversely, port-to-port connections are of the form:

subcomponent.port.connect(port);

so connection C would be coded as:

class producer(UVMComponent):
    put_port = uvm_put_port()
    c = conv()

def connect_phase(self, phase):
    super().connect_phase(phase)
    self.c.put_port.connect(self.put_port)
    # ...

2.3.2.2 Connection Types ¶

Table 1 summarizes connection types and elaboration functions.

Table 1—TLM Connection Types

Connection type connect form¶
port-to-export	comp1.port.connect(comp2.export);
port-to-port	subcomponent.port.connect(port);
export-to-export	export.connect(subcomponent.export);

NOTE—The argument to the port.connect() method may be either an export or a port, depending on the nature of the connection (that is, peer-to-peer or hierarchical). The argument to export.connect() is always an export of a child component.

2.3.3 Analysis Communication ¶

The put/get communication as described above allows verification components to be created that model the “operational” behavior of a system. Each component is responsible for communicating through its TLM interface(s) with other components in the system in order to stimulate activity in the DUT and/or respond its behavior. In any reasonably complex verification environment, however, particularly where randomization is applied, a collected transaction should be distributed to the rest of the environment for end-to-end checking (scoreboard), or additional coverage collection.

The key distinction between the two types of TLM communication is that the put/get ports typically require a corresponding export to supply the implementation. For analysis, however, the emphasis is on a particular component, such as a monitor, being able to produce a stream of transactions, regardless of whether there is a target actually connected to it. Modular analysis components are then connected to the analysis_port, each of which processes the transaction stream in a particular way.

2.3.3.1 Analysis Ports ¶

The UVMAnalysisPort (represented as a diamond on the monitor in Figure 8) is a specialized TLM port whose interface consists of a single function, write(). The analysis port contains a list of analysis_exports that are connected to it. When the component calls analysis_port.write(), the analysis_port cycles through the list and calls the write() method of each connected export. If nothing is connected, the write() call simply returns. Thus, an analysis port may be connected to zero, one, or many analysis exports, but the operation of the component that writes to the analysis port does not depend on the number of exports connected. Because write() is a void function, the call will always complete in the same delta cycle, regardless of how many components (for example, scoreboards, coverage collectors, and so on) are connected.

Code example:

class get_consumer_with_ap(get_consumer):

    def __init__(self, name, parent):
        super().__init__(name, parent)
        self.ap = UVMAnalysisPort(“analysis_port”, self)
        ...

    async def run_phase(self, phase):
        ...
        for i in range(10):
            t = []
            if self.get_port.get(t):
                # Do something with t.
                tx = t[0]
                self.ap.write(tx) # Write transaction.

In the parent environment, the analysis port gets connected to the analysis export of the desired components, such as coverage collectors and scoreboards.

2.3.3.2 Analysis Exports ¶

As with other TLM connections, it is up to each component connected to an analysis port to provide an implementation of write() via an analysis_export. The UVMSubscriber base component can be used to simplify this operation, so a typical analysis component would extend UVMSubscriber as:

class sub1(UVMSubscriber):

    # my_env env;

    def write(self, t);
        # Call desired functionality in parent.

As with put() and get() described above, the TLM connection between an analysis port and export, allows the export to supply the implementation of write(). If multiple exports are connected to an analysis port, the port will call the write() of each export, in order. Since all implementations of write() must be functions, the analysis port’s write() function completes immediately, regardless of how many exports are connected to it:

class my_env(UVMEnv):

    # get_component_with_ap g;
    # sub1 s1; sub2 s2;

    def __init__(self, name, parent):
        super().__init__(name,parent)
        self.s1 = sub1("s1")
        self.s1.env = self
        self.s2 = sub2("s2")
        self.s2.env = self
        self.g = get_component_with_ap()
    endfunction

    def connect_phase(self, phase):
        self.g.ap.connect(s1.analysis_export);
        # to illustrate analysis port can be connected to multiple
        # subscribers; usually the subscribers are in separate components
        self.g.ap.connect(s2.analysis_export);

When multiple subscribers are connected to an analysis_port, each is passed a pointer to the same transaction object, the argument to the write() call. Each write() implementation must make a local copy of the transaction and then operate on the copy to avoid corrupting the transaction contents for any other subscriber that may have received the same pointer.

UVM also includes an analysis_fifo, which is a uvm_tlm_fifo that also includes an analysis export, to allow blocking components access to the analysis transaction stream. The analysis_fifo is unbounded, so the monitor’s write() call is guaranteed to succeed immediately. The analysis component may then get the transactions from the analysis_fifo at its leisure.

2.4 TLM-2.0 Implementation ¶

Warning

TLM-2.0 interaction/integration with SystemC is untested at this point of development. This depends heavily on the simulation support as well.

The following subsections specify how TLM-2.0 is to be implemented in Python.

2.4.1 Generic Payload
2.4.1.1 Attributes
2.4.1.2 Accessors
2.4.1.3 Extensions
2.4.2 Core Interfaces and Ports
2.4.3 Blocking Transport
2.4.4 Nonblocking Transport
2.4.5 Sockets
2.4.6 Time
2.4.7 Use Models

2.4.1 Generic Payload ¶

TLM-2.0 defines a base object, called the generic payload, for moving data between components. In SystemC, this is the primary transaction vehicle. In uvm-python, this is the default transaction type, but it is not the only type that can be used (as will be explained more fully in Section 2.4.2).

2.4.1.1 Attributes ¶

Each attribute in the SystemC version has a corresponding member in the uvm-python generic payload:

m_address  # type: int
m_command  # type: uvm_tlm_command_e
m_data  # type: List[int]
m_length  # type: int
m_response_status  # type: uvm_tlm_response_status_e
m_dmi  # type: bool
m_byte_enable  # type: List[int]
unsigned m_byte_enable_length  # type: int
m_streaming_width  # int

The data types of most members translate directly into uvm-python. Unsigned int in SystemC become int in Python. m_data and m_byte_enable, which are defined as type char in SystemC, are defined as List of ints. uvm_tlm_command_e and uvm_tlm_response_status_e are enumerated types. They are defined as:

class uvm_tlm_command_e(Enum):
    UVM_TLM_READ_COMMAND = auto()
    UVM_TLM_WRITE_COMMAND = auto()
    UVM_TLM_IGNORE_COMMAND = auto()

class uvm_tlm_response_status_e(IntEnum):
    OK_RESPONSE = 1
    INCOMPLETE_RESPONSE = 0
    GENERIC_ERROR_RESPONSE = -1
    ADDRESS_ERROR_RESPONSE = -2
    COMMAND_ERROR_RESPONSE = -3
    BURST_ERROR_RESPONSE = -4
    BYTE_ENABLE_ERROR_RESPONSE = -5

All of the members of the generic payload can be randomized. This enables instances of the generic payload to be randomized. uvm-python allows arrays, including dynamic arrays to be randomized.

2.4.1.2 Accessors ¶

In SystemC, all of the attributes are private and are accessed through accessor methods. In Python, this restriction is removed. The following access methods are implemented:

def uvm_tlm_command_e get_command();
def void set_command(uvm_tlm_command_e command);
def bit is_read();
def void set_read();
def bit is_write();
def void set_write();
def void set_address(bit [63:0] addr);
def bit[63:0] get_address();
def void get_data (output byte p []);
def void set_data_ptr(ref byte p []);


def int unsigned get_data_length();
def void set_data_length(int unsigned length);
def int unsigned get_streaming_width();
def void set_streaming_width(int unsigned width);
def void get_byte_enable(output byte p[]);
def void set_byte_enable(ref byte p[]);
def int unsigned get_byte_enable_length();
def void set_byte_enable_length(int unsigned length);
def void set_dmi_allowed(bit dmi);
def bit is_dmi_allowed();
def uvm_tlm_response_status_e get_response_status();
def void set_response_status(uvm_tlm_response_status_e status);
def bit is_response_ok();
def bit is_response_error();
def string get_response_string();

The accessor functions let you set and get each of the members of the generic payload. This implies a slightly different use model for the generic payload than in SystemC. The way the generic payload is defined in SystemC does not encourage you to create new transaction types derived from uvm_tlm_generic_payload. Instead, you would use the extensions mechanism (see Section 2.4.1.3). Thus, in SystemC, none of the accessors are virtual.

In uvm-python, an important use model is to add randomization constraints to a transaction type. This is most often done with inheritance—take a derived object and add constraints to a base class. These constraints can further be modified or extended by deriving a new class, and so on. To support this use model, the accessor functions are virtual, and the members are protected and not local.

2.4.1.3 Extensions ¶

The generic payload extension mechanism is very similar to the one used in SystemC; minor differences exist simply due to the lack of function templates in Python. Extensions are used to attach additional application-specific or bus-specific information to the generic bus transaction described in the generic payload.

An extension is an instance of a user-defined container class based on the uvm_tlm_extension class. The set of extensions for any particular generic payload object are stored in that generic payload object instance. A generic payload object may have only one extension of a specific extension container type.

Each extension container type is derived from the uvm_tlm_extension class and contains any additional information required by the user:

class gp_Xs_ext(uvm_tlm_extension):

    def __init__(string name = ""):
        super().__init__(name)
        self.Xmask = []

uvm_object_utils_begin(gp_Xs_ext)
uvm_field_int_array('Xmask', UVM_ALL_ON)
uvm_object_utils_end(gp_Xs_ext)

To add an extension to a generic payload object, allocate an instance of the extension container class and attach it to the generic payload object using the set_extension() method:

Xs = gp_Xs_ext()
gp.set_extension(Xs)

The static function ID() in the user-defined extension container class can be used as an argument to the function get_extension method to retrieve the extension (if any) of the corresponding container type— if it is attached to the generic payload object:

Xs = gp.get_extension(gp_Xs_ext.ID)

The following methods are also available in the generic payload for managing extensions:

get_num_extensions()
clear_extension()
clear_extensions()

clear_extension() removes any extension of a specified type. clear_extensions() removes all extension containers from the generic payload.

2.4.2 Core Interfaces and Ports ¶

In the uvm-python implementation of TLM-2.0, we have provided only the basic transport interfaces. They are defined in the uvm_tlm_if#() class:

class uvm_tlm_if():

The interface class is parameterized with the type of the transaction object that will be transported across the interface and the type of the phase enum. The default transaction type is the generic payload. The default phase enum is:

typedef enum {
    UNINITIALIZED_PHASE, BEGIN_REQ, END_REQ, BEGIN_RESP, END_RESP
} uvm_tlm_phase_e;

Each of the interface methods take a handle to the transaction to be transported and a handle to a timescale- independent time value object. In addition, the nonblocking interfaces take a reference argument for the phase:

virtual function uvm_tlm_sync_e nb_transport_fw(
    T t, ref P p, input uvm_tlm_time delay
);
virtual function uvm_tlm_sync_e nb_transport_bw(
    T t, ref P p, input uvm_tlm_time delay
);
virtual task b_transport(T t, uvm_tlm_time delay);

In SystemC, the transaction argument is of type T&. Passing a handle to a class in SystemVerilog most closely represents the semantics of T& in SystemC. One implication in SystemVerilog is transaction types cannot be scalars. If the transaction argument was qualified with ref, indicating it was a reference argument, then it would be possible to use scalar types for transactions. However, that would also mean downstream components could change the handle to a transaction. This violates the required semantics in TLM-2.0 as stated in rule 4.1.2.5-b of the TLM-2.0 LRM, which is quoted here.

“If there are multiple calls to nb_transport associated with a given transaction instance, one and the same transaction object shall be passed as an argument to every such call. In other words, a given transaction instance shall be represented by a single transaction object.”

The phase and delay arguments may change value. These are also references in SystemC; e.g., P& and sc_time&. However, phase is a scalar, not a class, so the best translation is to use the ref qualifier to ensure the same object is used throughout the call sequence.

The uvm_tlm_time argument, which is present on all the interfaces, represents time. In the SystemC TLM-2.0 specification, this argument is reference to an sc_time variable, which lets the value change on either side. This was translated to a class object in SystemVerilog in order to manage timescales in different processes. Times passed through function calls are not automatically scaled. See Section 2.4.6 for more details.

An important difference between TLM-1 and TLM-2.0 is the TLM-2.0 interfaces pass transactions by reference and not by value. In SystemC, transactions in TLM-1 were passed as const references and in TLM-2.0 just as references. This allows the transaction object to be modified without copying the entire transaction. The result is much higher performance characteristics as a lot of copying is avoided. Another result is any object that has a handle to a transaction may modify it. However, to adhere to the semantics of the TLM-2.0 interfaces, these modifications must be made within certain rules and in concert with notifications made via the return enum in the ‘nb_*’ interfaces and the phase argument.

2.4.3 Blocking Transport ¶

The blocking transport is implemented as follows:

async def b_transport(t, delay):

The b_transport task transports a transaction from the initiator to the target in a blocking fashion. The call to b_transport by the initiator marks the first timing point in the execution of the transaction. That first timing point may be offset from the current simulation by the delay value specified in the delay argument. The return from b_transport by the target marks the final timing point in the execution of the transaction. That last timing point may be offset from the current simulation time by the delay value specified in the delay argument. Once the task returns, the transaction has been completed by the target. Any indication of success or failure must be annotated in the transaction object by the target.

The initiator may read or modify the transaction object before the call to b_transport and after its return, but not while the call to b_transport is still active. The target may modify the transaction object only while the b_transport call is active and must not keep a reference to it after the task return. The initiator is responsible for allocating the transaction object before the call to b_transport. The same transaction object may be reused across b_transport calls.

2.4.4 Nonblocking Transport ¶

The blocking transport is implemented using two interfaces:

function uvm_tlm_sync_e nb_transport_fw(T t, ref P p, input
    uvm_tlm_time delay);
function uvm_tlm_sync_e nb_transport_bw(T t, ref P p, input
    uvm_tlm_time delay);

nb_transport_fw transports a transaction in the forward direction, that is from the initiator to the target (the forward path). nb_transport_bw does the reverse, it transports a transaction from the target back to the initiator (the backward path). An initiator and target will use the forward and backward paths to update each other on the progress of the transaction execution. Typically, nb_transport_fw is called by the initiator whenever the protocol state machine in the initiator changes state and nb_transport_bw is called by the target whenever the protocol state machine in the target changes state.

The ‘nb_*’ interfaces each return an enum uvm_tlm_sync_e. The possible enum values and their meanings are shown in Table 2.

Table 2—uvm_tlm_sync_e enum Description

uvm_tlm_sync_e enum Description¶
Enum value	Interpretation
UVM_TLM_ACCEPTED:	Transaction has been accepted. Neither the transaction object, the phase nor the delay arguments have been modified.
UVM_TLM_UPDATED:	Transaction has been modified. The transac tion object, the phase or the delay arguments may have been modified.
UVM_TLM_COMPLETED:	Transaction execution has completed. The transaction object, the phase or the delay arguments may have been modified. There will be no further transport calls associated with this transaction.

The P argument of nb_transport_fw and nb_transport_bw represents the transaction phase. This can be a user-defined type that is specific to a particular protocol. The default type is uvm_tlm_phase_e, whose values are shown in Table 3. These can be used to implement the Base Protocol.

Table 3—uvm_tlm_phase_e Description

uvm_tlm_phase_e Description¶
Enum value	Interpretation
UNITIALIZED_PHASE	Phase has not yet begun
BEGIN_REQ	Request has begun
END_REQ	Request has completed
BEGIN_RESP	Response has begun
END_RESP	Response has terminated

The first call to nb_transport_fw by the initiator marks the first timing point in the transaction execution. Subsequent calls to nb_transport_fw and nb_transport_bw mark additional timing points in the transaction execution. The last timing point is marked by a return from nb_transport_fw or nb_transport_bw with UVM_TLM_COMPLETED. All timing points may be offset from the current simulation time by the delay value specified in the delay argument. An nb_transport_fw call on the forward path shall under no circumstances directly or indirectly make a call to nb_transport_bw on the backward path, and vice versa.

The value of the phase argument represents the current state of the protocol state machine. Any change in the value of the transaction object should be accompanied by a change in the value of phase. When using the Base Protocol, successive calls to nb_transport_fw or nb_transport_bw with the same phase value are not permitted.

The initiator may modify the transaction object, the phase and the delay arguments immediately before calls to nb_transport_fw and before it returns from nb_transport_bw only. The target may modify the transaction object, the phase and the delay arguments immediately before calls to nb_transport_bw and before it returns from nb_transport_fw only. The transaction object, phase and delay arguments may not be otherwise modified by the initiator or target.

The initiator is responsible for allocating the transaction object before the first call to nb_transport_fw. The same transaction object is used by all of the forward and backward calls during its execution. That transaction object is alive for the entire duration of the transaction until the final timing point. The same transaction object may be reused across different transaction execution that do not overlap in time.

2.4.5 Sockets ¶

In TLM-1, the primary means of making a connection between two processes is through ports and exports, whereas in TLM-2.0 this done through sockets. A socket is like a port or export; in fact, it is derived from the same base class as ports and export, namely uvm_port_base. However, unlike a port or export a socket provides both a forward and backward path. Thus, you can enable asynchronous (pipelined) bi-directional communication by connecting sockets together. To enable this, a socket contains both a port and an export.

Components that initiate transactions are called initiators and components that receive transactions sent by an initiator are called targets. Initiators have initiator sockets and targets have target sockets. Initiator sockets can only connect to target sockets; target sockets can only connect to initiator sockets.

Figure 9 shows the diagramming of socket connections. The socket symbol is a box with an isosceles triangle with its point indicating the data and control flow direction of the forward path. The backward path is indicated by an arrow connecting the target socket back to the initiator socket. Section 3.4 of the TLM-2.0 LRM fully explains sockets, initiators, targets, and interconnect components.

Sockets come in several flavors: Each socket is an initiator or a target, a passthrough, or a terminator. Furthermore, any particular socket implements either blocking interfaces or nonblocking interfaces. Terminator sockets are used on initiators and targets as well as interconnect components as shown in Figure 9. Passthrough sockets are used to enable connections to cross hierarchical boundaries.

The cross product of {initiator, target} X {terminator, passthrough} X {blocking, nonblocking} yields eight different kinds of sockets. The class definitions for these sockets are as follows:

class uvm_tlm_nb_passthrough_initiator_socket #(type T=uvm_tlm_generic_payload,
type P=uvm_tlm_phase_e) extends
uvm_tlm_nb_passthrough_initiator_socket_base #(T,P);

class uvm_tlm_nb_passthrough_target_socket#(
    type T=uvm_tlm_generic_payload,
    type P=uvm_tlm_phase_e
) extends uvm_tlm_nb_passthrough_target_socket_base #(T,P);

class uvm_tlm_b_passthrough_initiator_socket #(
    type T=uvm_tlm_generic_payload
) extends
uvm_tlm_b_passthrough_initiator_socket_base #(T);

class uvm_tlm_b_passthrough_target_socket #(
    type T=uvm_tlm_generic_payload
)
extends uvm_tlm_b_passthrough_target_socket_base #(T);

class uvm_tlm_b_target_socket #(
    type T=uvm_tlm_generic_payload,
    type IMP=int
) extends uvm_tlm_b_target_socket_base #(T);

class uvm_tlm_b_initiator_socket #(type T=uvm_tlm_generic_payload)

extends uvm_tlm_b_initiator_socket_base #(T);

class uvm_tlm_nb_target_socket #(
    type T=uvm_tlm_generic_payload,
    type P=uvm_tlm_phase_e, type IMP=int
) extends uvm_tlm_nb_target_socket_base #(T,P);

class uvm_tlm_nb_initiator_socket #(
    type T=uvm_tlm_generic_payload,
    type P=uvm_tlm_phase_e, type IMP=int
) extends uvm_tlm_nb_initiator_socket_base #(T,P);

Table 4 shows the different kinds of sockets and how they are constructed.

Table 4—Socket Construction

uvm-python Checks and Coverage Construct Usage Overview¶
Socket	Blocking	Nonblocking
initiator	IS-A forward port	IS-A forward port; HAS-A backward imp
target	IS-A forward imp	IS-A forward imp; HAS-A backward port
passthrough initiator	IS-A forward port	IS-A forward port; HAS-A backward export
passthrough target	IS-A forward export	IS-A forward port; HAS-A backward export

IS-A and HAS-A are types of object relationships. IS-A refers to the inheritance relationship and HAS-A refers to the ownership relationship. For example, if you say D is a B, it means D is derived from base B. If you say object A HAS-A B, it means B is a member of A.

Each <socket_type>::connect() calls super.connect(), which performs all the connection mechanics. For the nonblocking sockets which have a secondary port/export for the backward path, connect() is called on the secondary port/export to form a backward connection.

Each socket type provides an implementation of the connect() method. Connection is defined polymorphically using the base class type as the argument:

def connect(provider)

Each implementation of connect() in each socket type does run-time type checking to ensure it is connected to allowable socket types. For example, an initiator socket can connect to an initiator passthrough socket, a target passthrough socket, or a target socket. It cannot connect to another initiator socket. These kinds of checks are performed for each socket type.

2.4.6 Time ¶

Warning

This is section on Time must be revised for uvm-python.

Integers are not sufficient on their own to represent time without any ambiguity; you need to know the scale of that integer value, which is conveyed outside of the integer. In SystemVerilog, this is based on the timescale that was active when the code was compiled. SystemVerilog properly scales time literals, but not integer values because it does not know the difference between an integer that carries an integer value and an integer that carries a time value. time variables are simply 64-bit integers, they are not scaled back and forth to the underlying precision. Here is a short example that illustrates part of the problem:

\`timescale 1ns/1ps
module m();

    time t;

    initial begin
        #1.5;
        $write("T=%f ns (Now should be 1.5)\n", $realtime());
        t = 1.5;
        #t; // 1.5 will be rounded to 2
        $write("T=%f ns (Now should be 3.0)\n", $realtime());
        #10ps;
        $write("T=%f ns (Now should be 3.010)\n", $realtime());
        t = 10ps; // 0.010 will be converted to int (0)
        #t;
        $write("T=%f ns (Now should be 3.020)\n", $realtime());
    end

endmodule

yields:

T=1.500000 ns (Now should be 1.5)
T=3.500000 ns (Now should be 3.0)
T=3.510000 ns (Now should be 3.010)
T=3.510000 ns (Now should be 3.020)

Within SystemVerilog, we have to worry about different time scales and precision. Because each endpoint in a socket could be coded in different packages and, thus, be executing under different timescale directives, a simple integer cannot be used to exchange time information across a socket.

For example:

\`timescale 1ns/1ps

package a_pkg; class a;

function void f(inout time t);
    t += 10ns;
endfunction endclass

endpackage

\`timescale 1ps/1ps

program p;
import a_pkg::*;

time t = 0;

initial begin
    a A = new;
    A.f(t);
    #t;
    $write("T=%0d ps (Should be 10,000)\n", $time());
end

endprogram

yields:

T=10 ps (Should be 10,000)

Scaling is needed every time you make a procedural call to code that may interpret a time value in a different timescale. Using the uvm_tlm_time type:

`timescale 1ns/1ps

package a_pkg;

    import uvm_pkg::*;

    class a;

    function void f(uvm_tlm_time t);
        t.incr(10ns, 1ns);
    endfunction

    endclass

endpackage

`timescale 1ps/1ps

program p;

    import uvm_pkg::*;
    import a_pkg::*;

    uvm_tlm_time t = new;

    initial begin
        a A = new;
        A.f(t);
        #(t.get_realtime(1ns));
        $write("T=%0d ps (Should be 10,000)\n", $time());
    end
endprogram

yields:

T=10000 ps (Should be 10,000)

To solve these problems, the uvm_tlm_time class contains the scaling information so that as time information is passed between processes, which may be executing under different time scales, the time can be scaled properly in each environment.

2.4.7 Use Models ¶

Since sockets are derived from UVMPortBase, they are created and connected in the same way as port and exports. You can create them in the build phase and connect them in the connect phase by calling UVMPortBase.connect(). Initiator and target termination sockets are the end points of any connection. There can be an arbitrary number of passthrough sockets in the path between the initiator and target.

Some socket types must be bound to imps—implementations of the transport tasks and functions. Blocking terminator sockets must be bound to an implementation of b_transport(), for example. Nonblocking initiator sockets must be bound to an implementation of nb_transport_bw and nonblocking target sockets must be bound to an implementation of nb_transport_fw. Typically, the task or function is implemented in the component where the socket is instantiated and the component type and instance are provided to complete the binding.

Consider, for example, a consumer component with a blocking target socket:

class consumer (UVMComponent);

    def __init__(self, name,  parent):
        super().__init__(name, parent)
        self.target_socket = None

    def build_phase(self, phase):
        self.target_socket = uvm_tlm_b_target_socket("target_socket", self, self);

    # Note: async must be used to allow usage of await
    async def b_transport(self, t, delay):
        await Timer(5)
        uvm_info("consumer", t.convert2string(), UVM_LOW)

The interface task b_transport is implemented in the consumer component. The consumer component type is used in the declaration of the target socket, which informs the socket object of the type of the object containing the interface task, in this case b_transport(). When the socket is instantiated this is passed in twice, once as the parent, just like any other component instantiation, and again to identify the object that holds the implementation of b_transport(). Finally, in order to complete the binding, an implementation of b_transport() must be present in the consumer component.

Any component that has a blocking termination socket, nonblocking initiator socket, or nonblocking termination socket must provide implementations of the relevant components. This includes initiator and target components, as well as interconnect components that have these kinds of sockets. Components with passthrough sockets do not need to provide implementations of any sort. Of course, they must ultimately be connected to sockets that do provide the necessary implementations.

3. Developing Reusable Verification Components ¶

This chapter describes the basic concepts and components that make up a typical verification environment. It also shows how to combine these components using a proven hierarchical architecture to create reusable verification components. The sections in this chapter follow the same order you should follow when developing a verification component:

Modeling Data Items for Generation
Transaction-Level Components
Creating the Driver
Creating the Sequencer
Creating the Monitor
Instantiating Components
Creating the Agent
Creating the Environment
Enabling Scenario Creation
Managing End of Test
Implementing Checks and Coverage

NOTE—This chapter builds upon concepts described in Chapter 1 and Chapter 2.

3.1 Modeling Data Items for Generation ¶

Data items:

— Are transaction objects used as stimulus to the device under test (DUT). — Represent transactions that are processed by the verification environment. — Are instances of classes that you define (“user-defined” classes). — Capture and measure transaction-level coverage and checking.

NOTE—The UVM Class Library provides the UVMSequenceItem base class. Every user-defined data item should be derived directly or indirectly from this base class.

To create a user-defined data item:

a) Review your DUT’s transaction specification and identify the application-specific properties, constraints, tasks, and functions. b) Derive a data item class from the UVMSequenceItem base class (or a derivative of it). c) Define a constructor for the data item. d) Add control fields (“knobs”) for the items identified in Step (a) to enable easier test writing. e) Use UVM field macros to enable printing, copying, comparing, and so on. f) Define do functions for use in creation, comparison, printing, packing, and unpacking of transaction data as needed (see Section 6.7).

UVM has built-in automation for many service routines that a data item needs. For example, you can use:

- print() to print a data item.
- copy() to copy the contents of a data item.
- compare() to compare two similar objects.

UVM allows you to specify the automation needed for each field and to use a built-in, mature, and consistent implementation of these routines.

To assist in debugging and tracking transactions, the uvm_transaction base class provides access to a unique transaction number via the get_transaction_id() member function. In addition, the UVMSequenceItem base class (extended from uvm_transaction) also includes a get_transaction_id() member function, allowing sequence items to be correlated to the sequence that generated them originally.

The class simple_item in this example defines several random variables and class constraints. The UVM mixin functions implement various utilities that operate on this class, such as copy, compare, print, and so on. In particular, the uvm_object_utils function registers the class type with the common factory:

class simple_item(UVMSequenceItem)
def __init__(self name="simple_item"):
 super().__init__(name)
 self.addr = 0x0
 self.randr('addr', range(0x2000))
 self.data = 0x0
 self.randr('data', range(0x1000))
 self.delay = 0x0
 self.randr('delay', range(1 << 32 - 1))
# UVM automation macros for general objects
uvm_object_utils_begin(simple_item)
uvm_field_int('addr', UVM_ALL_ON)
uvm_field_int('data', UVM_ALL_ON)
uvm_field_int('delay', UVM_ALL_ON)
uvm_object_utils_end

Line 1 Derives data items from UVMSequenceItem so they can be generated in a procedural sequence. See Section 3.10.2 for more information.

Line 5,Line 7 and Line 9

Add constraints to a data item definition in order to:

Reflect specification rules. In this example, the address must be less than 0x2000. Specify the default distribution for generated traffic. For example, in a typical test most transactions should be legal.

Line 11-Line 15 Use the UVM functions to automatically implement functions such as copy(), compare(), print(), pack(), and so on. Refer to “Macros” in the UVM 1.2 Class Reference for information on the uvm_object_utils_begin, uvm_object_utils_end, uvm_field_*, and their associated macros.

3.1.1 Inheritance and Constraint Layering ¶

In order to meet verification goals, the verification component user might need to adjust the data-item generation by adding more constraints to a class definition. In SystemVerilog, this is done using inheritance. The following example shows a derived data item, word_aligned_item, which includes an additional constraint to select only word-aligned addresses:

class word_aligned_item extends simple_item;

    constraint word_aligned_addr { addr[1:0] == 2'b00; }
    `uvm_object_utils(word_aligned_item)

    // Constructor
    function new (string name = "word_aligned_item");
        super.new(name);
    endfunction : new

endclass : word_aligned_item

To enable this type of extensibility:

— The base class for the data item (simple_item in this chapter) should use virtual methods to allow derived classes to override functionality. — Make sure constraint blocks are organized so that they are able to override or disable constraints for a random variable without having to rewrite a large block. — Note that fields can be declared with the protected or local keyword to restrict access to properties. This, however, will limit the ability to constrain them with an inline constraint.

3.1.2 Defining Control Fields (“Knobs”)¶

The generation of all values of the input space is often impossible and usually not required. However, it is important to be able to generate a few samples from ranges or categories of values. In the simple_item example in Section 3.1, the delay property could be randomized to anything between zero and the maximum unsigned integer. It is not necessary (nor practical) to cover the entire legal space, but it is important to try back-to-back items along with short, medium, and large delays between the items, and combinations of all of these. To do this, define control fields (often called “knobs”) to enable the test writer to control these variables. These same control knobs can also be used for coverage collection. For readability, use enumerated types to represent various generated categories.

Knobs Example:

typedef enum {ZERO, SHORT, MEDIUM, LARGE, MAX} simple_item_delay_e;

class simple_item extends UVMSequenceItem;

    rand int unsigned addr;
    rand int unsigned data;
    rand int unsigned delay;
    rand simple_item_delay_e delay_kind; // Control field
    // UVM automation macros for general objects
    `uvm_object_utils_begin(simple_item)
    `uvm_field_int(addr, UVM_ALL_ON)
    `uvm_field_enum(simple_item_delay_e, delay_kind, UVM_ALL_ON)
    `uvm_object_utils_end

    constraint delay_order_c { solve delay_kind before delay; }

    constraint delay_c {(delay_kind == ZERO) -> delay == 0;
        (delay_kind == SHORT) -> delay inside { [1:10] };
        (delay_kind == MEDIUM) -> delay inside { [11:99] };
        (delay_kind == LARGE) -> delay inside { [100:999] };
        (delay_kind == MAX ) -> delay == 1000;
        delay >=0; delay <= 1000;
    }
endclass : simple_item

Using this method allows you to create more abstract tests. For example, you can specify distribution as:

constraint delay_kind_d {
    delay_kind dist {ZERO:=2, SHORT:=1, MEDIUM:=1, LONG:=1, MAX:=2};
}

When creating data items, keep in mind what range of values are often used or which categories are of interest to that data item. Then add knobs to the data items to simplify control and coverage of these data item categories.

3.2 Transaction-Level Components ¶

As discussed in Chapter 2, TLM interfaces in UVM provide a consistent set of communication methods for sending and receiving transactions between components. The components themselves are instantiated and connected in the testbench, to perform the different operations required to verify a design. A simplified testbench is shown in Figure 10 (for simplicity, typical containers like environment and agents are not shown here).

The basic components of a simple transaction-level verification environment are:

a) A stimulus generator (sequencer) to create transaction-level traffic to the DUT. b) A driver to convert these transactions to signal-level stimulus at the DUT interface. c) A monitor to recognize signal-level activity on the DUT interface and convert it into transactions. d) An analysis component, such as a coverage collector or scoreboard, to analyze transactions.

As we shall see, the consistency and modularity of the TLM interfaces in UVM allow components to be reused as other components are replaced and/or encapsulated. Every component is characterized by its interfaces, regardless of its internal implementation (see Figure 11). This chapter discusses how to encapsulate these types of components into a proven architecture, a verification component, to improve reuse even further.

Figure 11 shows the recommended grouping of individual components into a reusable interface-level verification component agent. Instead of reusing the low-level classes individually, the developer creates a component that encapsulates it’s sub-classes in a consistent way. Promoting a consistent architecture makes these components easier to learn, adopt, and configure.

3.3 Creating the Driver ¶

The driver’s role is to drive data items to the bus following the interface protocol. The driver obtains data items from the sequencer for execution. The UVM Class Library provides the uvm_driver base class, from which all driver classes should be extended, either directly or indirectly. The driver has a TLM port through which it communicates with the sequencer (see the example below). The driver may also implement one or more of the run-time phases (run and pre_reset - post_shutdown) to refine its operation.

To create a driver:

Derive from the uvm_driver base class.
If desired, add UVM
infrastructure macros for class properties to implement utilities for printing, copying, comparing, and so on.
Obtain the next data item from the
sequencer and execute it as outlined above.
Declare a virtual
interface in the driver to connect the driver to the DUT.

Refer to Section 3.10.2 for a description of how a sequencer, driver, and sequences synchronize with each other to generate constrained random data.

The class simple_driver in the example below defines a driver class. The example derives simple_driver from uvm_driver (parameterized to use the simple_item transaction type) and uses the methods in the seq_item_port object to communicate with the sequencer. As always, include a constructor and the uvm_component_utils mixin to register the driver type with the common factory.

class simple_driver(UVMDriver):
    simple_item s_item;
    virtual dut_if vif;

    # Constructor
    def __init__(self, name="simple_driver", parent):
        super().__init__(name, parent);


    def build_phase(self, phase):
        inst_name = ""
        super().build_phase(phase);
        if(!UVMConfigDb#(virtual dut_if)::get(this, "","vif",vif))
            uvm_fatal("NOVIF", {"virtual interface must be set for: ",
                get_full_name(),".vif"});


    async def run_phase(self, phase):
        while True:
            # Get the next data item from sequencer (may block).
            seq_item_port.get_next_item(s_item);
            # Execute the item.
            drive_item(s_item);
            seq_item_port.item_done(); // Consume the request.


    async def drive_item(self, item):
        ... # Add your logic here.


uvm_component_utils(simple_driver)

Line 1 Derive the driver.

Line 5 Add UVM infrastructure macro.

Line 13 Get the resource that defines the virtual interface

Line 22 Call get_next_item() to get the next data item for execution from the sequencer.

Line 25 Signal the sequencer that the execution of the current data item is done.

Line 30 Add your application-specific logic here to execute the data item.

More flexibility exists on connecting the drivers and the sequencer. See Section 3.5.

3.4 Creating the Sequencer ¶

The sequencer generates stimulus data and passes it to a driver for execution. The UVM Class Library provides the UVMSequencer base class, which is parameterized by the request and response item types. The UVMSequencer base class contains all of the base functionality required to allow a sequence to communicate with a driver. The uvm_sequencer gets instantiated directly, with appropriate parameterization as shown in Section 3.8.1, Line 3. In the class definition, by default, the response type is the same as the request type. If a different response type is desired, the optional second parameter must be specified for the uvm_sequencer base type:

uvm_sequencer #(simple_item, simple_rsp) sequencer;

Refer to Section 3.10.2 for a description of how a sequencer, driver, and sequences synchronize with each other to generate constrained-random data.

3.5 Connecting the Driver and Sequencer ¶

The driver and the sequencer are connected via TLM, with the driver’s seq_item_port connected to the sequencer’s seq_item_export (see Figure 12). The sequencer produces data items to provide via the export. The driver consumes data items through its seq_item_port and, optionally, provides responses. The component that contains the instances of the driver and sequencer makes the connection between them. See Section 3.8.

The seq_item_port in UVMDriver defines the set of methods used by the driver to obtain the next item in the sequence. An important part of this interaction is the driver’s ability to synchronize to the bus, and to interact with the sequencer to generate data items at the appropriate time. The sequencer implements the set of methods that allows flexible and modular interaction between the driver and the sequencer.

3.5.1 Basic Sequencer and Driver Interaction ¶

Basic interaction between the driver and the sequencer is done using the tasks get_next_item() and item_done(). As demonstrated in the example in Section 3.3, the driver uses get_next_item() to fetch the next randomized item to be sent. After sending it to the DUT, the driver signals the sequencer that the item was processed using item_done().Typically, the main loop within a driver resembles the following pseudo code:

while True:
    t = []
    await self.get_next_item(t) # Send item following the protocol. item_done();
    req = t[0]
end

NOTE—get_next_item() is blocking until an item is provided by the sequences running on that sequencer.

3.5.2 Querying for the Randomized Item ¶

In addition to the get_next_item() task, the UVMSeqItemPullPort class provides another task, try_next_item(). This task will return in the same simulation step if no data items are available for execution. You can use this task to have the driver execute some idle transactions, such as when the DUT has to be stimulated when there are no meaningful data to transmit. The following example shows a revised implementation of the run_phase() task in the previous example (in Section 3.3), this time using try_next_item() to drive idle transactions as long as there is no real data item to execute:

async def run_phase(self, phase):
    while True:

    # Try the next data item from sequencer (does not block).
    t = []
    await seq_item_port.try_next_item(t)
    s_item = t[0] if len(t) > 0 else None
    if s_item is None:
        # No data item to execute, send an idle transaction. ...
    end
    else:
        # Got a valid item from the sequencer, execute it. ...
        # Signal the sequencer; we are done.
        seq_item_port.item_done();

3.5.3 Fetching Consecutive Randomized Items ¶

In some protocols, such as pipelined protocols, the driver may operate on several transactions at the same time. The sequencer-driver connection, however, is a single item handshake which shall be completed before the next item is retrieved from the sequencer. In such a scenario, the driver can complete the handshake by calling item_done() without a response and provide the response by a subsequent call to put_response(r) with the real response data.

3.5.4 Sending Processed Data back to the Sequencer ¶

In some sequences, a generated value depends on the response to previously generated data. By default, the data items between the driver and the sequencer are copied by reference, which means that changes the driver makes to the data item will be visible inside the sequencer. In cases where the data item between the driver and the sequencer is copied by value, the driver needs to return the processed response back to the sequencer. Do this using the optional argument to item_done():

seq_item_port.item_done(rsp)

or using the put_response() method:

await seq_item_port.put_response(rsp)

or using the built-in analysis port in UVMDriver:

rsp_port.write(rsp)

NOTE—Before providing the response, the response’s sequence and transaction id must be set to correspond to the request transaction using rsp.set_id_info(req).

NOTE—put_response() is a blocking method, so the sequence must do a corresponding get_response(rsp).

With the basic functionality of driver-sequencer communication outlined above, the steps required to create a driver are straightforward.

3.5.5 Using TLM-Based Drivers ¶

The seq_item_port, which is built into uvm_driver, is a bidirectional port. It also includes the standard TLM methods get() and peek() for requesting an item from the sequencer, and put() to provide a response. Thus, other components, which may not necessarily be derived from uvm_driver, may still connect to and communicate with the sequencer. As with the seq_item_port, the methods to use depend on the interaction desired:

# Pause sequencer operation while the driver operates on the transaction.
await peek(req) # Process req operation.
await get(req) # Allow sequencer to proceed immediately upon driver receiving transaction.

await get(req) # Process req operation.

The following also apply.

— peek() is a blocking method, so the driver may block waiting for an item to be returned. — The get() operation notifies the sequencer to proceed to the next transaction. It returns the same transaction as the peek(), so the transaction may be ignored.

To provide a response using the blocking_slave_port, the driver would call:

await seq_item_port.put(rsp)

The response may also be sent back using an analysis_port as well.

3.6 Creating the Monitor ¶

The monitor is responsible for extracting signal information from the bus and translating it into events, data, and status information. This information is available to other components and to the test writer via standard TLM interfaces and channels. The monitor should never rely on state information collected by other components, such as a driver, but it may need to rely on request-specific id information in order to properly set the sequence and transaction id information for the response.

The monitor functionality should be limited to basic monitoring that is always required. This can include protocol checking—which should be configurable so it can be enabled or disabled—and coverage collection. Additional high-level functionality, such as scoreboards, should be implemented separately on top of the monitor.

If you want to verify an abstract model or accelerate the pin-level functionality, you should separate the signal-level extraction, coverage, checking, and the transaction-level activities. An analysis port should allow communication between the sub-monitor components (see the UVM 1.2 Class Reference).

Monitor Example

The following example shows a simple monitor which has the following functions:

— The monitor collects bus information through a virtual interface (xmi). — The collected data is used in coverage collection and checking. — The collected data is exported on an analysis port (item_collected_port).

Actual code for collection is not shown in this example. A complete example can be found in the UBus example in ubus_master_monitor.py:

class master_monitor(UVMMonitor):

    virtual bus_if xmi; // SystemVerilog virtual interface bit
    checks_enable = 1; // Control checking in monitor and interface. bit
    coverage_enable = 1; // Control coverage in monitor and interface.

    # event cov_transaction; // Events needed to trigger covergroups

    #covergroup cov_trans @cov_transaction;
    #    option.per_instance = 1;
    #    ... // Coverage bins definition
    #endgroup : cov_trans

    def __init__(self, name, parent):
        super().__init__(name, parent);
        cov_trans = new();
        cov_trans.set_inst_name({get_full_name(), ".cov_trans"});
        trans_collected = simple_item()
        self.item_collected_port = UVMAnalysisPort("item_collected_port", self)
        self.xmi = None  # Obtain the bus handle somehow
        self.cov_transaction = Event('evt_cov_transaction')

    async def run_phase(self, phase):
        await self.collect_transactions() # collector task.

    async def collect_transactions();
        while True:
            await RisingEdge(self.xmi.sig_clock)
            # Collect the data from the bus into trans_collected.
            if (self.checks_enable):
                self.perform_transfer_checks()
            if (self.coverage_enable)
                self.perform_transfer_coverage()
            self.item_collected_port.write(trans_collected)
        end

    def perform_transfer_coverage(self):
        self.cov_transaction.set()

    def perform_transfer_checks(self):
        ... # Perform data checks on trans_collected.

uvm_component_utils_begin(master_monitor)
uvm_field_int('checks_enable', UVM_ALL_ON)
uvm_field_int('coverage_enable', UVM_ALL_ON)
uvm_component_utils_end(master_monitor)

The collection is done in a task (collect_transactions) which is spawned at the beginning of the run() phase. It runs in an endless loop and collects the data as soon as the signals indicate that the data is available on the bus.

As soon as the data is available, it is sent to the analysis port (item_collected_port) for other components waiting for the information.

Coverage collection and checking are conditional because they can affect simulation run-time performance. If not needed, they can be turned off by setting coverage_enable or checks_enable to 0, using the configuration mechanism. For example:

UVMConfigDb.set(this,“*.master0.monitor”, “checks_enable”, 0)

If checking is enabled, the task calls the perform_transfer_checks function, which performs the necessary checks on the collected data (trans_collected). If coverage collection is enabled, the task emits the coverage sampling event (cov_transaction) which results in collecting the current values.

NOTE—SystemVerilog does not allow concurrent assertions in classes, so protocol checking can also be done using assertions in a SystemVerilog interface.

3.7 Instantiating Components ¶

The isolation provided by object-oriented practices and TLM interfaces between components facilitate reuse in UVM enabling a great deal of flexibility in building environments. Because each component is independent of the others, a given component can be replaced by a new component with the same interfaces without having to change the parent’s connect() method. This flexibility is accomplished through the use of the factory in UVM.

When instantiating components in UVM, rather than calling its constructor (in bold below):

class my_component (UVMComponent);

# my_driver driver;
...
def build_phase(self, phase):
    super().build_phase(phase)
    driver = my_driver(“driver”,self)

components should be instantiated using the create() method:

class my_component (UVMComponent);

    # my_driver driver;
    ...
    def build_phase(self, phase):
        super().build_phase(phase)
        driver = my_driver.type_id.create("driver", self)

The factory operation is explained in Section 6.2. The type_id::create() method is a type-specific static method that returns an instance of the desired type (in this case, my_driver) from the factory. The arguments to create() are the same as the standard constructor arguments, a string name and the parent component. The use of the factory allows the developer to derive a new class extended from my_driver and cause the factory to return the extended type in place of my_driver. Thus, the parent component can use the new type without modifying the parent class.

For example, for a specific test, an environment user may want to change the driver. To change the driver for a specific test:

a) Declare a new driver extended from the base component and add or modify functionality as desired:

class new_driver(my_driver):
... # Add more functionality here.

b) In your test, environment, or testbench, override the type to be returned by the factory:

def build_phase(self,  phase):
    self.set_type_override_by_type(my_driver.get_type(),
        new_driver.get_type())
    super().build_phase(phase)

The factory also allows a new type to be returned for the creation of a specific instance as well. In either case, because new_driver is an extension of my_driver and the TLM interfaces are the same, the connections defined in the parent remain unchanged.

3.8 Creating the Agent ¶

An agent (see Figure 13) instantiates and connects together a driver, monitor, and sequencer using TLM connections as described in the preceding sections. To provide greater flexibility, the agent also contains configuration information and other parameters. As discussed in Section 1.1.5, UVM recommends that the verification component developer create an agent that provides protocol-specific stimuli creation, checking, and coverage for a device. In a bus-based environment, an agent usually models a master, a slave, or an arbiter component.

3.8.1 Operating Modes ¶

An agent has two basic operating modes:

— Active mode, where the agent emulates a device in the system and drives DUT signals. This mode requires that the agent instantiate a driver and sequencer. A monitor also is instantiated for checking and coverage.

— Passive mode, where the agent does not instantiate a driver or sequencer and operates passively. Only the monitor is instantiated and configured. Use this mode when only checking and coverage collection is desired.

The class simple_agent in the example below instantiates a sequencer, a driver, and a monitor in the recommended way. Instead of using the constructor, the UVM build phase is used to configure and construct the subcomponents of the agent. Unlike constructors, this virtual function can be overridden without any limitations. Also, instead of hard coding, the allocation type_id::create() is used to instantiate the subcomponents. The example in “To change the driver for a specific test:” in Section 3.8 illustrates how you can override existing behavior using extends:

class simple_agent extends uvm_agent;
... // Constructor and UVM automation macros
uvm_sequencer #(simple_item) sequencer;
simple_driver driver;
simple_monitor monitor;
// Use build_phase to create agents's subcomponents.
virtual function void build_phase(uvm_phase phase);
super.build_phase(phase)
monitor = simple_monitor.type_id.create("monitor",this);
if (is_active == UVM_ACTIVE) begin
// Build the sequencer and driver.
sequencer =
uvm_sequencer#(simple_item).type_id.create("sequencer",this);
driver = simple_driver.type_id.create("driver",this);
end
endfunction : build_phase
virtual function void connect_phase(uvm_phase phase);
if(is_active == UVM_ACTIVE) begin
driver.seq_item_port.connect(sequencer.seq_item_export);
end
endfunction : connect_phase
endclass : simple_agent

NOTE—Invoking super.build_phase() (see Line 8) enables the automatic configuration for UVM fields declared via the ‘uvm_field_*’ macros during the build phase.

Line 9 The monitor is created using create().

Line 10 - Line 15 The if condition tests the is_active property to determine whether the driver and sequencer are created in this agent. If the agent is set to active (is_active = UVM_ACTIVE), the driver and sequencer are created using additional create() calls.

Both the sequencer and the driver follow the same creation pattern as the monitor.

This example shows the is_active flag as a configuration property for the agent. You can define any control flags that determine the component’s topology. At the environment level, this could be a num_masters integer, a num_slaves integer, or a has_bus_monitor flag. See Chapter 7 for a complete interface verification component example that uses all the control fields previously mentioned.

NOTE—create() should always be called from the build_phase() method to create any multi-hierarchical com- ponent.

Line 18 - Line 20 The if condition should be checked to see if the agent is active and, if so, the connection between the sequencer and driver is made using connect_phase().

3.8.2 Connecting Components ¶

The connect_phase() phase, which happens after the build phase is complete, should be used to connect the components inside the agent. See Line 18 - Line 20 in the example in Section 3.8.1.

3.9 Creating the Environment ¶

Having covered the basic operation of transaction-level verification components in a typical environment above, this section describes how to assemble these components into a reusable environment (see Figure 14). By following the guidelines here, you can ensure that your environment will be architecturally correct, consistent with other verification components, and reusable. The following sections describe how to create and connect environment sub-components.

3.9.1 The Environment Class ¶

The environment class is the top container of reusable components. It instantiates and configures all of its subcomponents. Most verification reuse occurs at the environment level where the user instantiates an environment class and configures it and its agents for specific verification tasks. For example, a user might need to change the number of masters and slaves in a new environment as shown below:

class ahb_env(UVMEnv):

    int num_masters;
    ahb_master_agent masters[];
    `uvm_component_utils_begin(ahb_env)
    `uvm_field_int(num_masters, UVM_ALL_ON)
    `uvm_component_utils_end

    virtual function void build_phase(phase);
        string inst_name;
        super.build_phase(phase);
        if(num_masters ==0))
            `uvm_fatal("NONUM",{"'num_masters' must be set";
        masters = new[num_masters];

        for(int i = 0; i < num_masters; i++) begin
            $sformat(inst_name, "masters[%0d]", i);
            masters[i] = ahb_master_agent.type_id.create(inst_name,this);
        end
        // Build slaves and other components.
        endfunction

    function new(string name, uvm_component parent);
        super.new(name, parent);
    endfunction : new

endclass

NOTE—Similarly to the agent, create is used to allocate the environment sub-components. This allows introducing derivations of the sub-components later.

3.9.2 Invoking build_phase ¶

Earlier versions of UVM supported a manual invocation of several flow tasks, such as build, connect, run and more (essentially the phase tasks without the _phase suffix). The usage of these tasks is now deprecated and the manual invocation of these tasks is considered an error. All phase tasks are now automatically invoked by the UVM Class Library. Any connections between child components should be made in the connect() function of the parent component.

3.10 Enabling Scenario Creation

The environment user will need to create many test scenarios to verify a given DUT. Since the verification component developer is usually more familiar with the DUT’s protocol, the developer should facilitate the test writing (done by the verification component’s user) by doing the following:

— Place knobs in the data item class to simplify declarative test control.

— Create a library of interesting reusable sequences.

The environment user controls the environment-generated patterns configuring its sequencers. The user can:

— Define new sequences that generate new transactions.

— Define new sequences that invoke existing sequences.

— Override default knobs on data items to modify driver and overall environment behavior.

— “Enable” any new behavior/sequences (see Section 3.10.3).

In this section we describe how to create a library of reusable sequences and review their use. For more information on how to control environments, see Section 4.6.

3.10.1 Declaring User-Defined Sequences

Sequences are made up of several data items, which together form an interesting scenario or pattern of data. Verification components can include a library of basic sequences (instead of single-data items), which test writers can invoke. This approach enhances reuse of common stimulus patterns and reduces the length of tests. In addition, a sequence can call upon other sequences, thereby creating more complex scenarios.

NOTE—The UVM Class Library provides the UVMSequence base class. You should derive all sequence classes directly or indirectly from this class.

To create a user-defined sequence:

a) Derive a sequence from the UVMSequence base class and specify the request and response item type parameters. In the example below, only the request type is specified, simple_item. This will result in the response type also being of type simple_item. b) Use the `uvm_object_utils macro to register the sequence type with the factory. c) If the sequence requires access to the derived type-specific functionality of its associated sequencer, add code or use the ’uvm_declare_p_sequencer macro to declare and set the desired sequencer pointer. d) Implement the sequence’s body task with the specific scenario you want the sequence to execute. In

the body task, you can execute data items and other sequences (see Section 3.10.2).

The class simple_seq_do in the following example defines a simple sequence. It is derived from UVMSequence and uses the `uvm_object_utils macro. The example then defines a simple_sequencer class on which the simple_seq_do sequence can run:

class simple_seq_do extends UVMSequence #(simple_item);

    rand int count;
    constraint c1 { count >0; count <50; }
    // Constructor
    function new(string name="simple_seq_do");
        super.new(name);
    endfunction
    //Register with the factory
    `uvm_object_utils(simple_seq_do)
    // The body() task is the actual logic of the sequence.
    virtual task body();
    repeat(count) // Example of using convenience macro to execute the item
        \`uvm_do(req)
    endtask : body
endclass : simple_seq_do

class simple_sequencer extends uvm_sequencer #(simple_item) ;

    // same parameter as simple_seq_do
    `uvm_component_utils(simple_sequencer)
    function new (string name=“simple_sequencer”, uvm_component parent);
        super.new(name,parent) ;
    endfunction
endclass

3.10.2 Sending Subsequences and Sequence Items

Sequences allow you to define:

— Streams of data items sent to a DUT. — Streams of actions performed on a DUT interface.

You can also use sequences to generate static lists of data items with no connection to a DUT interface.

3.10.2.1 Basic Flow for Sequences and Sequence Items

To send a sequence item, the body() of a sequence needs to create() the item (use the factory), call start_item() on the item, optionally randomize the item, and call finish_item() on the item. To send a subsequence, the body() of the parent sequence needs to create the subsequence, optionally randomize it, and call start() for the subsequence. If the subsequence item has an associated response, the parent sequence can call get_response(). See Section 6.7.

For some advanced use cases, there may be a use for some callbacks through the flow.

Figure 15 and Figure 16 show a complete flow for sequence items and sequences that has been implemented in the uvm_do macros. The entire flow includes the allocation of an object based on factory settings for the registered type, which is referred to as “creation” in this section. After creation, comes the initialization of class properties. The balance of the object processing depends on whether the object is a sequence item or a sequence: the pre_do(), mid_do(), and post_do() callbacks of the sequence and randomization of the objects are called at different points of processing for each object type as shown in the figures. Note that the pre_body() and post_body() methods are not called for subsequences.

3.10.2.2 Sequence and Sequence Item Macros

This section describes the sequence and sequence functins, uvm_do and uvm_do_with. The uvm_do function and variations provide a convenient set of calls to create, randomize, and send transaction items in a sequence. The uvm_do function delays randomization of the item until the driver has signaled that it is ready to receive it and the pre_do method has been executed. Other function variations allow constraints to be applied to the randomization (uvm_do_with) or bypass the randomization altogether.

3.10.2.2.1 uvm_do

This function takes as an argument a variable of type UVMSequence or UVMSequenceItem. An object is created using the factory settings and assigned to the specified variable. Based on the processing in Figure 15, when the driver requests an item from the sequencer, the item is randomized and provided to the driver.

The simple_seq_do sequence declaration in the example in Section 3.10.1 is repeated here. The body of the sequence invokes an item of type simple_item, using the uvm_do function:

class simple_seq_do extends UVMSequence #(simple_item);

    ...
    // Constructor and UVM automation macros
    // See Section 4.7.2

    virtual task body();
        uvm_do(req)
    endtask : body
endclass : simple_seq_do

Similarly, a sequence variable can be provided and will be processed as shown in Figure 16. The following example declares another sequence (simple_seq_sub_seqs), which uses `uvm_do to execute a sequence of type simple_seq_do, which was defined earlier:

class simple_seq_sub_seqs extends UVMSequence #(simple_item);

    ...
    // Constructor and UVM automation macros
    // See Section 4.7.2
    simple_seq_do seq_do;
    virtual task body();
        uvm_do(seq_do)
    endtask : body
endclass : simple_seq_sub_seqs

3.10.2.2.2 uvm_do_with

This function is similar to `uvm_do (Section 3.10.2.2.1). The first argument is a variable of a type derived from UVMSequenceItem, which includes items and sequences. The second argument can be any valid inline constraints that would be legal if used in arg1.randomize() with inline constraints. This enables adding different inline constraints, while still using the same item or sequence variable.

Example

This sequence produces two data items with specific constraints on the values of addr and data:

class simple_seq_do_with(UVMSequence):

    ...
    // Constructor and UVM automation macros
    // See Section 4.7.2
    virtual task body();

        `uvm_do_with(req, { req.addr == 16'h0120; req.data == 16'h0444; } )
        `uvm_do_with(req, { req.addr == 16'h0124; req.data == 16'h0666; } )
    endtask : body
endclass : simple_seq_do_with

If constraints are used simply to set parameters to specific values, as in the previous example, the macro can be replaced with a user-defined task:

class simple_seq_do_with extends UVMSequence #(simple_item);

    task do_rw(int addr, int data);
        item= simple_item.type_id.create("item",,get_full_name());
        item.addr.rand_mode(0);
        item.data.rand_mode(0);
        item.addr = addr;
        item.data = data;
        start_item(item);
        randomize(item);
        finish_item(item);
    endtask

    virtual task body();
        repeat (num_trans)
            do_rw($urandom(),$urandom());
    endtask
    ...
endclass: simple_seq_do_with

3.10.3 Starting a Sequence on a Sequencer

Sequencers do not execute any sequences by default. The start() method needs to be called for one or more sequences to source any transactions. That start() call can be provided directly in user code. Alternatively, the user can specify a sequence to be started automatically upon a certain phase via the UVMConfigDb.

3.10.3.1 Manual Starting

The user can instantiate and randomize a sequence instance and call start() for that instance at any point.

3.10.3.2 Using the Automated Phase-Based Starting

As each run-time phase starts, the sequencer will check for the existence of a resource corresponding to that phase to determine if there is a sequence to start automatically. Such a resource may be defined in user code, typically the test. For example, this resource setting causes the specified sequencer instance to be triggered by starting main_phase and creating an instance of the loop_read_modify_write_seq sequence, then randomize it and start executing it:

UVMConfigDb.set(self,
    ".ubus_example_tb0.ubus0.masters[0].sequencer.main_phase",
    "default_sequence", loop_read_modify_write_seq.type_id.get()
)

It is also possible to start a specific instance of a sequence:

lrmw_seq = loop_read_modify_write_seq.type_id.create(
    “lrmw”,, self.get_full_name()) # set parameters in lrmw_seq, if desired
UVMConfigDb.set(this, ".ubus_example_tb0.ubus0.masters[0].sequencer.main_phase",
    "default_sequence", lrmw_seq)

By creating a specific instance of the sequence, the instance may be randomized and/or specific parameters set explicitly or constrained as needed. Upon entering the specified phase, the sequence instance will be started. The sequencer will not randomize the sequence instance.

3.10.4 Overriding Sequence Items and Sequences

In a user-defined UVMTest, e.g., base_test_ubus_demo (discussed in Section 4.5.1), you can configure the simulation environment to use a modified version of an existing sequence or a sequence item by using the common factory to create instances of sequence and sequence-item classes. See Section 6.2 for more information.

To override any reference to a specific sequence or sequence-item type:

a) Declare a user-defined sequence or sequence item class which derives from an appropriate base class. The following example shows the declaration of a basic sequence item of type sim- ple_item and a derived item of type word_aligned_item. b) Invoke the appropriate uvm_factory override method, depending on whether you are doing a type-wide or instance-specific override. For example, assume the simple_seq_do sequence is executed by a sequencer of type simple_sequencer (both defined in Section 3.10.1). You can choose to replace all processing of simple_item types with word_aligned_item types. This can be selected for all requests for simple_item types from the factory or for specific instances of simple_item. From within an UVM component, the user can execute any of the following:

// Affect all factory requests for type simple_item.
set_type_override_by_type(simple_item::get_type(),
    word_aligned_item::get_type());
// Affect requests for type simple_item only on a given sequencer.
set_inst_override_by_type("env0.agent0.sequencer.*",
    simple_item::get_type(), world_aligned_item::get_type());
// Alternatively, affect requests for type simple_item for all
// sequencers of a specific env.
set_inst_override_by_type("env0.*.sequencer.*",
    simple_item::get_type(), word_aligned_item::get_type());

Allocate the item using the factory (i.e., with create(), see Section 3.10.2); any existing override requests will take effect and a word_aligned_item will be created instead of a simple_item.

3.11 Managing End of Test

UVM provides an objection mechanism to allow hierarchical status communication among components. There is a built-in objection for each phase, which provides a way for components and objects to synchronize their testing activity and indicate when it is safe to end the phase and, ultimately, the test.

In general, the process is for a component or sequence to raise a phase objection at the beginning of an activity that must be completed before the phase stops and to drop the objection at the end of that activity. Once all of the raised objections are dropped, the phase terminates.

In simulation, agents may have a meaningful agenda to be achieved before the test goals can be declared as done. For example, a master agent may need to complete all its read and write operations before the run phase should be allowed to stop. A reactive slave agent may not object to the end-of-test as it is merely serving requests as they appear without a well-defined agenda.

On the other hand, for the sequences, there are three possible ways the phase objection can be handled.

Non-Phase Aware Sequences

1) The caller will handle phase objection raise/drop around sequence invocation. 2) The sequence itself is not phase aware.

class test extends ovm_test;

task run_phase(uvm_phase phase);

phase.raise_objection(this); seq.start(seqr); phase.drop_objection(this); endtask endclass

Phase Aware Sequences(Explicit Objection)

1) The caller will pass the starting phase reference before starting the sequence. 2) The sequence will explicitly call raise/drop to control the objection. 3) Where exactly the raise/drop is called is up to the user design. It might be called in pre/ post_body (extra care is needed in this case as pre/post_body are not always called), or pre/post_start, or even the body itself. It might also be raised and dropped multiple times within the sequence execution to guard a specific critical logic class

test extends ovm_test;

task run_phase (uvm_phase phase); seq.set_starting_phase(phase); seq.start(seqr); endtask endclass

class seq extends UVMSequence #(data_item);

task body();

uvm_phase p = get_starting_phase(); if(p) p.raise_objection(this);

//some critical logic If(p) p.drop_objection(this); endtask endclass

Phase Aware Sequences (Implicit Objection)

1) The caller will pass the starting phase reference before starting the sequence. 2) Within the sequence (mostly inside seq::new), the user will call set_automatic_phase_objection(1); 3) uvm_sequence_base will handle automatic phase raise/drop before/after pre/post_start.

class test extends ovm_test;

task run_phase (uvm_phase phase); seq.set_starting_phase(phase); seq.start(seqr); endtask endclass

class seq extends UVMSequence #(data_item);

function new(string name = “seq”);

super.new(name); set_automatic_phase_objection(1); endfunction task body();

// Sequence logic with no objection // as it is already handled in the base class endtask endclass

Note that if you are using UVM task-based phases’ default_sequence mechanism, the “caller” will be the UVM sequencer, in which case you can’t really do Option a without some workarounds. So you will default to Option b or Option c.

When all objections are dropped, the currently running phase is ended. In practice, there are times in simulation when the “all objections dropped” condition is temporary. For example, concurrently running processes may need some additional cycles to convey the last transaction to a scoreboard.

To accommodate this, you may use the phase_ready_to_end() method to re-raise the phase objection if a transaction is currently in-flight.

Alternatively, you may set a drain time to inject a delay between the time a component’s total objection count reaches zero for the current phase and when the drop is passed to its parent. If any objections are re- raised during this delay, the drop is canceled and the raise is not propagated further. While a drain time can be set at each level of the component hierarchy with the adding effect, typical usage would be to set a single drain time at the env or test level. If you require control over drain times beyond a simple time value (for example, waiting for a few clock cycles or other user-defined events), you can also use the all_dropped callback to calculate drain times more precisely. For more information on the all_dropped callback, refer to uvm_objection in the UVM 1.2 Class Reference.

Vertical reuse means building larger systems out of existing ones. What was once a top-level environment becomes a sub-environment of a large environment. The objection mechanism allows sub-system environment developers to define a drain time per sub-system.

3.12 Implementing Checks and Coverage

Checks and coverage are crucial to a coverage-driven verification flow. Python/uvm-python allows the usage shown in Table 5 for assert, @Coverpoint, and @Covergroup constructs.

There are no temporal/concurrent assertions implemented via special syntax. Users must implement these assertions using normal Python procedural code and cocotb/uvm-python.

Table 5—Python/uvm-python Checks and Coverage Construct Usage Overview

uvm-python Checks and Coverage Construct Usage Overview¶
class	interface	package	module	initial	always	generate	program
assert	no	yes	no	yes	yes	yes	yes	yes
@Coverpoint	no	yes	yes	yes	yes	yes	yes	yes
@Covergroup	yes	yes	yes	yes	no	no	yes	yes

In a verification component, checks and coverage are defined in multiple locations depending on the category of functionality being analyzed. In Figure 17, checks and coverage are depicted in the UVMMonitor and interface. The following sections describe how the assert, cover, and covergroup constructs are used in the Ubus verification component example (described in Chapter 7).

3.12.1 Implementing Checks and Coverage in Classes

Class checks and coverage should be implemented in the classes derived from UVMMonitor. The derived class of UVMMonitor is always present in the agent and, thus, will always contain the necessary checks and coverage. The bus monitor is created by default in an env and if the checks and coverage collection is enabled the bus monitor will perform these functions. The remainder of this section uses the master monitor as an example of how to implement class checks and coverage, but they apply to the bus monitor as well.

You can write class checks as procedural code or SystemVerilog immediate assertions.

Tip: Use immediate assertions for simple checks that can be written in a few lines of code and use functions for complex checks that require many lines of code. The reason is that, as the check becomes more complicated, so does the debugging of that check.

NOTE—Concurrent assertions are not allowed in SystemVerilog classes per the IEEE1800 LRM.

The following is a simple example of an assertion check. This assertion verifies the size field of the transfer is 1, 2, 4, or 8. Otherwise, the assertion fails:

function void ubus_master_monitor::check_transfer_size();
    check_transfer_size : assert(trans_collected.size == 1 \|\|
        trans_collected.size == 2 \|\| trans_collected.size == 4 \|\|
        trans_collected.size == 8) else begin

    // Call DUT error: Invalid transfer size! end
endfunction : check_transfer_size

The following is a simple example of a function check. This function verifies the size field value matches the size of the data dynamic array. While this example is not complex, it illustrates a procedural-code example of a check:

function void ubus_master_monitor::check_transfer_data_size();
    if (trans_collected.size != trans_collected.data.size())
        // Call DUT error: Transfer size field / data size mismatch.
endfunction : check_transfer_data_size

The proper time to execute these checks depends on the implementation. You should determine when to make the call to the check functions shown above. For the above example, both checks should be executed after the transfer is collected by the monitor. Since these checks happen at the same instance in time, a wrapper function can be created so that only one call has to be made. This wrapper function follows.

function void ubus_master_monitor::perform_transfer_checks();

check_transfer_size(); check_transfer_data_size(); endfunction : perform_transfer_checks

The perform_transfer_checks() function is called procedurally after the item has been collected by the monitor.

Functional coverage is implemented using SystemVerilog covergroups. The details of the covergroup (that is, what to make coverpoints, when to sample coverage, and what bins to create) should be planned and decided before implementation begins.

The following is a simple example of a covergroup:

// Transfer collected beat covergroup.

covergroup cov_trans_beat @cov_transaction_beat;

    option.per_instance = 1; beat_addr :
    coverpoint addr {
        option.auto_bin_max = 16;
    }

    beat_dir : coverpoint trans_collected.read_write;

    beat_data : coverpoint data {
        option.auto_bin_max = 8;
    }

    beat_wait : coverpoint wait_state {
        bins waits[] = { [0:9] };
        bins others = { [10:$] };
    }
    beat_addrXdir : cross beat_addr, beat_dir;
    beat_addrXdata : cross beat_addr, beat_data;
endgroup : cov_trans_beat

This embedded covergroup is defined inside a class derived from UVMMonitor and is sampled explicitly. For the above covergroup, you should assign the local variables that serve as coverpoints in a function, then sample the covergroup. This is done so that each transaction data beat of the transfer can be covered. This function is shown in the following example:

def perform_transfer_coverage(self):
    self.cov_trans.sample()  # another covergroup
    for i in range(self.trans_collected.size):
        self.addr = self.trans_collected.addr + i
        self.data = self.trans_collected.data[i]
        self.wait_state = self.trans_collected.wait_state[i]
        self.cov_trans_beat.sample()

This function covers several properties of the transfer and each element of the dynamic array data. SystemVerilog does not provide the ability to cover dynamic arrays. You should access each element individually and cover that value, if necessary. The perform_transfer_coverage() function would, like perform_transfer_checks(), be called procedurally after the item has been collected by the monitor.

3.12.2 Implementing Checks and Coverage in Interfaces

Interface checks are implemented as assertions. Assertions are added to check the signal activity for a protocol. The assertions related to the physical interface are placed in the env’s interface. For example, an assertion might check that an address is never X or Y during a valid transfer. Use assert as well as assume properties to express these interface checks.

An assert directive is used when the property expresses the behavior of the device under test. An assume directive is used when the property expresses the behavior of the environment that generates the stimulus to the DUT. The mechanism to enable or disable the physical checks performed using assertions is discussed in Chapter 3.12.3.

3.12.3 Controlling Checks and Coverage

You should provide a means to control whether the checks are enforced and the coverage is collected. You can use an UVM bit field for this purpose. The field can be controlled using the UVMConfigDb interface. Refer to UVMConfigDb in the UVM 1.2 Class Reference for more information. The following is an example of using the checks_enable bit to control the checks:

if (checks_enable) perform_transfer_checks();

If checks_enable is set to 0, the function that performs the checks is not called, thus disabling the checks. The following example shows how to turn off the checks for the master0.monitor:

UVMConfigDb.set(this,"masters[0].monitor", "checks_enable", 0);

The same facilities exist for the coverage_enable field in the Ubus agent monitors and bus monitor.

4. Using Verification Components ¶

This chapter covers the steps needed to build a testbench from a set of reusable verification components. UVM accelerates the development process and facilitates reuse. UVM users will have fewer hook-up and configuration steps and can exploit a library of reusable sequences to efficiently accomplish their verification goals.

In this chapter, a distinction is made between the testbench integrator, who is responsible for the construction and configuration of the testbench, and the test writer who might have less knowledge about the underlying construction of the testbench, but wants to use it for creating tests. The test writer may skip the configuration sections and move directly into the test-creation sections.

The steps to create a testbench from verification components are:

Review the reusable verification component configuration parameters.
Instantiate and configure reusable verification components.
Create reusable sequences for interface verification components (optional).
Add a virtual sequencer (optional).
Add checking and functional coverage extensions.
Create tests to achieve coverage goals.

Before reading this chapter make sure you read Chapter 1. It is also recommended (but not required) that you read Chapter 3 to get a deeper understanding of verification components.

4.1 Creating a Top-Level Environment ¶

Instantiating individual verification components directly inside of each test has several drawbacks:

The test writer must know how to configure each component.
Changes to the topology require updating multiple test files, which can turn
into a big task.
The tests are not reusable because they rely on a specific environment structure.

The top-level environment is a container that defines the reusable component topology within the UVM tests. The top-level environment instantiates and configures the reusable verification IP and defines the default configuration of that IP as required by the application. Multiple tests can instantiate the top-level environment class and determine the nature of traffic to generate and send for the selected configuration. Additionally, the tests can override the default configuration of the top-level environment to better achieve their goals.

Figure 17 shows a typical verification environment that includes the test class containing the top-level environment class. Other verification components (or environments) are contained inside the top-level environment.

4.2 Instantiating Verification Components ¶

This section describes how the environment integrator can use verification components to create a top-level environment that can be reused for multiple tests. The following example uses the verification IP in Chapter 7. This interface verification component can be used in many environments due to its configurability, but in this scenario it will be used in a simple configuration consisting of one master and one slave. The top-level environment sets the applicable topology overrides.

The following also apply.

interface ports

— Examples for the UVMConfigDb.set calls can be found within the: build_phase() function.

class ubus_example_env(UVMEnv):


    def __init__(self, name, parent);
        super.new(name, parent);
        self.ubus0 = None  # Scoreboard to check the memory operation of the slave
        self.scoreboard0 = None

    def build_phase(self, phase):
        super().build_phase(phase)  # Configure before creating the
        # subcomponents.
        UVMConfigDb.set(self,"ubus0", "num_masters", 1);
        UVMConfigDb.set(self,".ubus0", "num_slaves", 1);
        self.ubus0 = ubus_env.type_id.create("ubus0", self)
        self.scoreboard0 = ubus_example_scoreboard.type_id.create("scoreboard0", self)


    def connect_phase(self, phase):
        # Connect slave0 monitor to scoreboard.
        self.ubus0.slaves[0].monitor.item_collected_port.connect(
            self.scoreboard0.item_collected_export)

    def end_of_elaboration_phase(self, phase);
        # Set up slave address map for ubus0 (basic default).
        ubus0.set_slave_address_map("slaves[0]", 0, 0xFFFF);

uvm_component_utils(ubus_example_env)  # UBus reusable environment

Other configuration examples include:

— Set the masters[0] agent to be active:

UVMConfigDb.set(self,"ubus0.masters[0]", "is_active", UVM_ACTIVE)

— Do not collect coverage for masters[0] agent:

UVMConfigDb.set(self, "ubus0.masters[0].monitor", "coverage_enable", 0)

— Set all slaves (using a wildcard) to be passive:

UVMConfigDb.set(self,"ubus0.slaves*", "is_active", UVM_PASSIVE)

Many test classes may instantiate the top-level environment class above, and configure it as needed. A test writer may use the top-level environment in its default configuration without having to understand all the details of how it is created and configured.

The ubus_example_env’s __init__ constructor is not used for creating the top-level environment subcomponents because there are limitations on overriding __init__ in object-oriented languages such as SystemVerilog. Instead, use a virtual build_phase() function, which is a built-in UVM phase.

The UVMConfigDb.set calls specify that the number of masters and slaves should both be 1. These configuration settings are used by the ubus0 environment during the ubus0 build_phase(). This defines the topology of the ubus0 environment, which is a child of the ubus_example_env.

In a specific test, a user might want to extend the ubus_env and derive a new class from it. create() is used to instantiate the subcomponents (instead of directly calling the new() constructor) so the ubus_env or the scoreboard classes can be replaced with derivative classes without changing the top-level environment file. See Section 6.2.3 for more information.

super.build_phase() is called as the first line of the ubus_example_env’s build() function. If the UVM field automation mixins are used, this updates the configuration fields of the ubus_example_tb.

connect_phase() is used to make the connection between the slave monitor and the scoreboard. The slave monitor contains a TLM analysis port which is connected to the TLM analysis export on the scoreboard. connect_phase() is a built-in UVM phase.

After the build_phase() and connect_phase() functions are complete, the user can make adjustments to run-time properties since the environment is completely elaborated (that is, created and connected). For example, the end_of_elaboration_phase() function makes the environment aware of the address range to which the slave agent should respond.

4.3 Creating Test Classes ¶

The UVMTest class defines the test scenario and goals. Part of that responsibility involved selecting which top level environment will be used, as well as enabling configuration of the environment and its children verification components. Although IP developers provide default values for topological and run-time configuration properties, the test writer may use the configuration override mechanism provided by the UVM Class Library when customization is required. The test writer may additionally provide user-defined sequences in an included file or package. A test provides data and sequence generation and inline constraints. Test files are typically associated with a single configuration. For usage examples of test classes, refer to Section 4.5.

Tests in UVM are classes that are derived from an UVMTest class. Using classes allows inheritance and reuse of tests. Typically, a base test class is defined that instantiates and configures the top-level environment (see Section 4.5.1), and is then extended to define scenario-specific configurations such as which sequences to run, coverage parameters, etc. The test instantiates the top-level environment just like any other verification component (see Section 4.2).

4.4 Verification Component Configuration ¶

4.4.1 Verification Component Configurable Parameters¶

Based on the protocols used in a device, the integrator instantiates the needed verification components and configures them for a desired operation mode. Some standard configuration parameters are recommended to address common verification needs. Other parameters are protocol- and implementation-specific.

Examples of standard configuration parameters:

— An agent can be configured for active or passive mode. In active mode, the agent drives traffic to the DUT. In passive mode, the agent passively checks and collects coverage for a device. A rule of thumb to follow is to use an active agent per device that needs to be emulated, and a passive agent for every RTL device that needs to be verified. — The monitor collects coverage and checks a DUT interface by default. The user may disable these activities by the standard checks_enable and coverage_enable parameters.

Examples of user-defined parameters:

— The number of master agents and slave agents in an AHB verification component. — The operation modes or speeds of a bus.

A verification component should support the standard configuration parameters and provide user-defined configuration parameters as needed. Refer to the verification component documentation for information about its user-defined parameters.

4.4.2 Verification Component Configuration Mechanism¶

UVM provides a configuration mechanism (see Figure 18) to allow integrators to configure an environment without needing to know the verification component implementation and hook-up scheme. The following are some examples:

UVMConfigDb.set(self,"*.masters[0]", "master_id", 0);
UVMConfigDb.set(self, "*.ubus0.masters[0].sequencer.main_phase", "default_sequence", read_modify_write_seq.type_id.get());
UVMConfigDb.set(this,"ubus_example_env0.*","vif",vif);
UVMResourceDb.set(“anyobject”, “shared_config”, data, this);

The UVMConfigDb is a type-specific configuration mechanism, offering a robust facility for specifying hierarchical configuration values of desired parameters. It is built on top of the more general purpose uvm_resource_db which provides side-band (non-hierarchical) data sharing. The first example above shows how to set in integral value for the master_id field of all master components whose instance name ends with masters[0]. The second example shows how to tell the masters[0].sequencer to execute a sequence of type read_modify_write_seq upon entering the main phase. The third example shows how to define the virtual interface type that all components under ubus_example_env0 should use to set their vif variable. The last example shows how to store some shared resource to a location where any object anywhere in the verification hierarchy can access it. When the uvm_resource_db::set() call is made from a class, the last parameter should be this to allow debugging messages to show where the setting originated.

4.4.3 Choosing between uvm_resource_db and UVMConfigDb¶

The UVMConfigDb and uvm_resource_db share the same underlying database. Because of this, it is possible to write to the database using UVMConfigDb.set and retrieve from the database using uvm_resource_db::read_by_name(). The primary reason for using one method over the other is whether or not a hierarchical context is important to the setting. For configuration properties that are related to hierarchical position, e.g., “set all of coverage_enable bits for all components in a specific agent”, UVMConfigDb is the correct choice. UVMConfigDb was architected to provide the required semantic for hierarchical configuration. Likewise, for cases where a configuration property is being shared without regard to hierarchical context, uvm_resource_db should be used.

4.4.4 Using a Configuration Class¶

Some verification components randomize configuration attributes inside a configuration class. Dependencies between these attributes are captured using constraints within the configuration object. In such cases, users can extend the configuration class to add new constraints or layer additional constraints on the class using inline constraints. Once configuration is randomized, the test writer can use UVMConfigDb.set to assign the configuration object to one or more environments within the top- level environment. Setting resources allows a configuration to target multiple sub-environments of the top-level environment regardless of their location, which allows for the build process of the top-level environment to be impacted without having to extend it.

4.5 Creating and Selecting a User-Defined Test ¶

In UVM, a test is a class that encapsulates test-specific instructions written by the test writer. This section describes how to create and select a test. It also describes how to create a test family base class to verify a topology configuration.

4.5.1 Creating the Base Test¶

The following example shows a base test that uses the ubus_example_env defined in Section 4.2. This base test is a starting point for all derivative tests that will use the ubus_example_env. The complete test class is shown here:

class ubus_example_base_test(UVMTest):


    def __init__(self, name="ubus_example_base_test", parent=None):
        super().__init__(name, parent)
        self.test_pass = True
        self.ubus_example_tb0 = None
        self.printer = None

    def build_phase(self, phase):
        super().build_phase(phase)
        # Enable transaction recording for everything
        UVMConfigDb.set(self, "*", "recording_detail", UVM_FULL)
        # Create the tb
        self.ubus_example_tb0 = ubus_example_tb.type_id.create("ubus_example_tb0", self)
        # Create a specific depth printer for printing the created topology
        self.printer = UVMTablePrinter()
        self.printer.knobs.depth = 3
        arr = []
        if UVMConfigDb.get(None, "*", "vif", arr) is True:
            UVMConfigDb.set(self, "*", "vif", arr[0])
        else:
            uvm_fatal("NOVIF", "Could not get vif from config DB")

The build_phase() function of the base test creates the ubus_example_env. The UVM Class Library will execute the build_phase() function of the ubus_example_base_test for the user when cycling through the simulation phases of the components. This fully creates the top-level environment as each sub-component will create their own children components in their respective build_phase() functions.

All of the definitions in the base test are inherited by any test that derives from ubus_example_base_test. This means any derivative test will not have to build the top-level environment if the test calls super.build_phase(). Likewise, the run_phase() task behavior can be inherited, as well as all other phases. If the current implementation does not meet the needs of the extended test, the build_phase() and/or run_phase(), as well as any other run-time phase methods may be refined as needed because they are all virtual.

4.5.2 Creating Tests from a Test-Family Base Class¶

The test writer can derive from the base test defined in Section 4.5.1 to create tests that reuse the same topology. Since the top-level environment is created by the base test’s build_phase() function, and the run_phase() task defines the run phase, the derivative tests can make minor adjustments. (For example, changing the default sequence executed by the agents in the environment.) The following is a simple test that inherits from ubus_example_base_test.

class test_read_modify_write(ubus_example_base_test):

    def __init__(self, name="test_read_modify_write", parent=None):
        super().__init__(name, parent)

    def build_phase(self, phase):
        UVMConfigDb.set(self, "ubus_example_tb0.ubus0.masters[0].sequencer.run_phase",
                "default_sequence", read_modify_write_seq.type_id.get())
        UVMConfigDb.set(self, "ubus_example_tb0.ubus0.slaves[0].sequencer.run_phase",
                "default_sequence", slave_memory_seq.type_id.get())
        #    // Create the tb from super class
        ubus_example_base_test.build_phase(self, phase)


    async def run_phase(self, phase):
        phase.raise_objection(self, "test_read_modify_write OBJECTED")
        master_sqr = self.ubus_example_tb0.ubus0.masters[0].sequencer
        slave_sqr = self.ubus_example_tb0.ubus0.slaves[0].sequencer

        uvm_info("TEST_TOP", "Forking master_proc now", UVM_LOW)
        master_seq = read_modify_write_seq("r_mod_w_seq")
        master_proc = cocotb.fork(master_seq.start(master_sqr))

        slave_seq = slave_memory_seq("mem_seq")
        slave_proc = cocotb.fork(slave_seq.start(slave_sqr))
        await sv.fork_join_any([slave_proc, master_proc])
        phase.drop_objection(self, "test_read_modify_write drop objection")


uvm_component_utils(test_read_modify_write)

This test changes the default sequence executed by the masters[0] agent and the slaves[0] agent. It is important to understand that super.build_phase(), through the base class, will create the top-level environment, ubus_example_env0, and all its subcomponents. Therefore, any configuration that will affect the building of these components (such as how many masters to create) must be set before calling super.build_phase(). For this example, since the sequences don’t get started until a later phase, they could be called after super.build_phase().

This test relies on the ubus_example_base_test implementation of the run_phase() phase.

4.5.3 Test Selection¶

After defining user-defined tests (described in Section 4.5.2), the uvm_pkg::run_test() task needs to be invoked to select a test to be simulated. Its prototype is:

async def run_test(test_name="")

UVM supports declaring which test is to be run via two separate mechanisms. The testname (i.e. the name which was used to register the test with the factory) can be passed directly to the run_test() task or it can be declared on the command-line via +UVM_TESTNAME. If both mechanisms are used, the command line takes priority. Once a test name has been selected, the run_test() task calls the factory to create an instance of the test with an instance name of uvm_test_top. Finally, run_test() starts the test by cycling through the simulation phases.

The following example shows how the test with the type name test_read_modify_write (defined in Section 4.5.2) can be provided to the run_test() task. A test name is provided to run_test() via a simulator command-line argument. Using the simulator command-line argument avoids having to hardcode the test name in the task which calls run_test(). In an initial block, call the run_test() as follows:

@cocotb.test
async def run_the_test(dut):
    await run_test(dut=dut)

To select a test of type test_read_modify_write (described in Section 4.5.2) using simulator command-line option, use the following command:

% SIM=sim*simulator-command other-options* +UVM_TESTNAME=test_read_modify_write
# Or Makefile based approach
% SIM=sim make UVM_TEST=test_read_modify_write

If the test name provided to run_test() does not exist, the simulation will exit immediately via a call to $fatal. If this occurs, it is likely the name was typed incorrectly or the uvm_component_utils macro was not used.

By using this method and only changing the +UVM_TESTNAME argument, multiple tests can be run without having to recompile or re-elaborate the testbench or underlying design.

4.6 Creating Meaningful Tests ¶

The previous sections show how test classes are put together. At this point, random traffic is created and sent to the DUT. The user can change the randomization seed to achieve new test patterns. To achieve verification goals in a systematic way, the user will need to control test generation to cover specific areas.

The user can control the test creation using these methods:

— Add constraints to control individual data items. This method provides basic functionality (see

Section 4.6.1). — Use UVM sequences to control the order of multiple data items. This method provides more flexibil-

ity and control (see Section 4.7.2).

4.6.1 Constraining Data Items¶

By default, sequencers repeatedly generate random data items. At this level, the test writer can control the number of generated data items and add constraints to data items to control their generated values.

To constrain data items:

Identify the data item classes and their generated fields in the verification component.
Create a derivation of the data item class that adds or overrides default constraints.
In a test, adjust the environment (or a subset of it) to use the newly-defined data items.
Run the simulation using a command-line option to specify the test name.

Data Item Example:

from enum import Enum
import uvm

class Parity(Enum):
    BAD_PARITY = 0
    GOOD_PARITY = 1

class UartFrame(uvm.UVMSequenceItem):
    def __init__(self, name=""):
        super().__init__(name)
        self.transmit_delay = 0
        self.start_bit = 0
        self.payload = 0
        self.stop_bits = 0
        self.error_bits = 0
        self.parity = 0
        self.parity_type = Parity.GOOD_PARITY

    def calc_parity(self):
        # implementation
        pass

uvm_object_utils_begin(UartFrame)
uvm_field_int('transmit_delay', UVM_ALL_ON)
uvm_field_int('start_bit', UVM_ALL_ON)
uvm_field_int('payload', UVM_ALL_ON)
uvm_field_int('stop_bits', UVM_ALL_ON)
uvm_field_int('error_bits', UVM_ALL_ON)
uvm_field_int('parity', UVM_ALL_ON)
uvm_field_enum('parity_type', Parity, UVM_ALL_ON + UVM_NOCOMPARE)
uvm_object_utils_end(UartFrame)

// different than zero.
# constraint error_bits_c {error_bits != 4'h0;}
// Default distribution constraints

# constraint default_parity_type {parity_type dist {
# GOOD_PARITY:=90, BAD_PARITY:=10};} // Utility functions

The uart_frame is created by the uart environment developer.

4.6.2 Data Item Definitions¶

A few fields in the derived class come from the device specification. For example, a frame should have a payload that is sent to the DUT. Other fields are there to assist the test writer in controlling the generation. For example, the field parity_type is not being sent to the DUT, but it allows the test writer to easily specify and control the parity distribution. Such control fields are called “knobs”. The verification component documentation should list the data item’s knobs, their roles, and legal range.

Data items have specification constraints. These constraints can come from the DUT specification to create legal data items. For example, a legal frame must have error_bits_c not equal to 0. A different type of constraint in the data items constrains the traffic generation. For example, in the constraint block default_parity_type (in the example in Section 4.6.1), the parity bit is constrained to be 90-percent legal (good parity) and 10-percent illegal (bad parity).

4.6.3 Creating a Test-Specific Frame¶

In tests, the user may wish to change the way data items are generated. For example, the test writer may wish to have short delays. This can be achieved by deriving a new data item class and adding constraints or other class members as needed:

// A derived data item example
// Test code
class short_delay_frame extends uart_frame;

    // This constraint further limits the delay values.
    constraint test1_txmit_delay {transmit_delay < 10;}
    uvm_object_utils(short_delay_frame)

    function new(input string name="short_delay_frame");
        super.new(name);
    endfunction

endclass: short_delay_frame

Deriving the new class is not enough to get the desired effect. The environment needs to be configured to use the new class (short_delay_frame) rather than the verification component frame. The UVM Class Library factory mechanism can be used to introduce the derived class to the environment:

class short_delay_test(UVMTest):

    def __init__(self, name="short_delay_test", parent=None):
        super().__init__(name, parent)
        self.uart_tb0 = None

    def build_phase(self, phase):
        super.build_phase(phase)  # Use short_delay_frame throughout the environment.
        factory.set_type_override_by_type(uart_frame.get_type(),
            short_delay_frame.get_type())
        self.uart_tb0 = uart_tb.type_id.create("uart_tb0", self)


    async def run_phase(self, phase):
        uvm_top.print_topology()

uvm_component_utils(short_delay_test)

Calling the factory function set_type_override_by_type() (in bold above) instructs the environment to use short-delay frames.

At times, a user may want to send special traffic to one interface but keep sending the regular traffic to other interfaces. This can be achieved by using set_inst_override_by_type() inside an UVM component:

set_inst_override_by_type("uart_env0.master.sequencer.*",
    uart_frame::get_type(), short_delay_frame::get_type());

Wildcards can also be used to override the instantiation of a few components, e.g.

set_inst_override_by_type("uart_env*.master.sequencer.*",
    uart_frame::get_type(), short_delay_frame::get_type())

4.7 Virtual Sequences ¶

Section 4.6 describes how to efficiently control a single-interface generation pattern. However, in a system- level environment, multiple components are generating stimuli in parallel. The user might want to coordinate timing and data between the multiple channels. Also, a user may want to define a reusable system-level scenario. Virtual sequences are associated with a virtual sequencer and are used to coordinate stimulus generation in a testbench hierarchy. In general, a virtual sequencer contains references to its subsequencers, that is, driver sequencers or other virtual sequencers in which it will invoke sequences. Virtual sequences can invoke other virtual sequences associated with its sequencer, as well as sequences in each of the subsequencers. However, virtual sequencers do not have their own data item and therefore do not execute data items on themselves. Virtual sequences can execute items on other sequencers that can execute items.

Virtual sequences enable centralized control over the activity of multiple verification components which are connected to the various interfaces of the DUT. By creating virtual sequences, existing sequence libraries of the underlying interface components and block-level environments can be used to create coordinated system-level scenarios.

In Figure 19, the virtual sequencer invokes configuration sequences on the ethernet and cpu verification components. The configuration sequences are developed during block-level testing.

There are three ways in which the virtual sequencer can interact with its subsequencers:

“Business as usual”—Virtual subsequencers and subsequencers send
transactions simultaneously.
Disable subsequencers—Virtual sequencer is the only one driving.
Using grab() and ungrab()—Virtual sequencer takes control of the
underlying driver(s) for a limited time.

When using virtual sequences, most users disable the subsequencers and invoke sequences only from the virtual sequence. For more information, see Section 4.7.3.

To invoke sequences, do one of the following: Use the appropriate uvm_do function. Use the sequence start() method.

4.7.1 Creating a Virtual Sequencer¶

For high-level control of multiple sequencers from a single sequencer, use a sequencer that is not attached to a driver and does not process items itself. A sequencer acting in this role is referred to as a virtual sequencer.

To create a virtual sequencer that controls several subsequencers:

Derive a virtual sequencer class from the uvm_sequencer class.
Add references to the sequencers where the virtual sequences will
coordinate the activity. These references will be assigned by a higher-level component (typically the top-level environment).

The following example declares a virtual sequencer with two subsequencers. Two interfaces called eth and cpu are created in the build function, which will be hooked up to the actual sub-sequencers:

class simple_virtual_sequencer extends uvm_sequencer;

    eth_sequencer eth_seqr;
    cpu_sequencer cpu_seqr; // Constructor

    function new(input string name="simple_virtual_sequencer",
            input uvm_component parent=null);
        super.new(name, parent);
    endfunction

    // UVM automation macros for sequencers
    `uvm_component_utils(simple_virtual_sequencer)

endclass: simple_virtual_sequencer

Subsequencers can be driver sequencers or other virtual sequencers. The connection of the actual subsequencer instances via reference is done later, as shown in Section 4.7.4.

4.7.2 Creating a Virtual Sequence¶

Creating a virtual sequence is similar to creating a driver sequence, with the following differences:

— A virtual sequence use `uvm_do_on or `uvm_do_on_with to execute sequences on any of the

subsequencers connected to the current virtual sequencer. — A virtual sequence uses `uvm_do or `uvm_do_with to execute other virtual sequences of this sequencer. A virtual sequence cannot use `uvm_do or `uvm_do_with to execute items. Virtual sequencers do not have items associated with them, only sequences.

To create a virtual sequence:

Declare a sequence class by deriving it from UVMSequence, just like
a driver sequence.
Define a body() method that implements the desired logic of the sequence.
Use the `uvm_do_on (or `uvm_do_on_with) macro to invoke sequences in the
underlying subsequencers.
Use the `uvm_do (or `uvm_do_with) macro to invoke other virtual
sequences in the current virtual sequencer.

The following example shows a simple virtual sequence controlling two subsequencers: a cpu sequencer and an ethernet sequencer. Assume the cpu sequencer has a cpu_config_seq sequence in its library and the ethernet sequencer provides an eth_large_payload_seq sequence in its library. The following sequence example invokes these two sequencers, one after the other.

class simple_virt_seq extends UVMSequence;

… // Constructor and UVM automation macros // A sequence from the cpu sequencer library cpu_config_seq conf_seq; // A sequence from the ethernet subsequencer library eth_large_payload_seq frame_seq; // A virtual sequence from this sequencer’s library random_traffic_virt_seq rand_virt_seq;

virtual task body();

// Invoke a sequence in the cpu subsequencer. `uvm_do_on(conf_seq, p_sequencer.cpu_seqr) // Invoke a sequence in the ethernet subsequencer. `uvm_do_on(frame_seq, p_sequencer.eth_seqr) // Invoke another virtual sequence in this sequencer. `uvm_do(rand_virt_seq) endtask : body endclass : simple_virt_seq

4.7.3 Controlling Other Sequencers¶

When using a virtual sequencer, the behavioral relationship between the subsequencers and the virtual sequence need to be considered. There are three typical possibilities:

a) Business as usual—The virtual sequencer and the subsequencers generate traffic at the same time, using the built-in capability of the original subsequencers. The data items resulting from the subse- quencers’ default behavior—along with those injected by sequences invoked by the virtual sequencer—will be intermixed and executed in an arbitrary order by the driver. This is the default behavior, so there is no need to do anything to achieve this.

b) Disable the subsequencers—Using the UVMConfigDb::set routines, the default_sequence property of the subsequencers is set to null, disabling their default behavior. The following code snippet disables the default sequences on the subsequencers in the example in Section 4.7.4:

// Configuration: Disable subsequencer sequences.
UVMConfigDb.set(this, "*.cpu_seqr.*_phase",
    "default_sequence", null);;
UVMConfigDb.set(this, "*.eth_seqr.*_phase",
    "default_sequence", null);

c) Use grab()/lock() and ungrab()/unlock()—In this case, a virtual sequence can achieve full control over its subsequencers for a limited time and then let the original sequences continue working. The grab and lock APIs effectively prevent sequence items from being executed on the locked sequencer, unless those items originate from the locking sequence. The grab() method places the lock request at the head of the sequencer’s arbitration queue, allowing the caller to prevent items currently waiting for grant from being processed, whereas the lock() method places the request at the end of the queue, allowing the items to be processed before the lock is granted.

NOTE—Sequences which are executing on the sequencer, but not actively attempting to send a sequence item (or attempting to achieve a lock) will continue to operate as normal. Therefore, it is important to ensure a given subsequencer is not a virtual sequencer before attempting to lock it.

The following example illustrates this using the functions grab() and ungrab() in the sequence consumer interface:

virtual task body();
    // Grab the cpu sequencer if not virtual.
    if (p_sequencer.cpu_seqr != null)
        grab(p_sequencer.cpu_seqr); // Execute a sequence.
    `uvm_do_on(conf_seq, p_sequencer.cpu_seqr) // Ungrab.
    if (p_sequencer.cpu_seqr != null) ungrab(p_sequencer.cpu_seqr);
endtask

NOTE—When grabbing several sequencers, make sure to use some convention to avoid deadlocks. For example, always grab in a standard order.

4.7.4 Connecting a Virtual Sequencer to Subsequencers¶

To connect a virtual sequencer to its subsequencers:

a) Assign the sequencer references specified in the virtual sequencer to instances of the sequencers.

This is a simple reference assignment and should be done only after all components are created:

v_sequencer.cpu_seqr = cpu_seqr
v_sequencer.eth_seqr = eth_seqr

b) Perform the assignment in the connect() phase of the verification environment at the appropriate location in the verification environment hierarchy.

Alternatively, the sequencer pointer could be set as a resource during build, as shown with eth_seqr below.

The following more-complete example shows a top-level environment, which instantiates the ethernet and cpu components and the virtual sequencer that controls the two. At the top-level environment, the path to the sequencers inside the various components is known and that path is used to get a handle to them and connect them to the virtual sequencer:

class simple_tb(UVMEnv):

    cpu_env_c cpu0; // Reuse a cpu verification component.
    eth_env_c eth0;
    // Reuse an ethernet verification component.
    simple_virtual_sequencer v_sequencer; ... // Constructor and UVM automation macros virtual

    function void build_phase(uvm_phase phase);
        super.build_phase(phase);
        // Configuration: Set the default sequence for the virtual sequencer.
        UVMConfigDb.set(this, "v_sequencer.run_phase",
            "default_sequence", simple_virt_seq.type_id::get());
        // Build envs with subsequencers.
        cpu0 = cpu_env_c.type_id.create("cpu0", this);
        eth0 = eth_env_c.type_id.create("eth0", this);

        // Build the virtual sequencer. v_sequencer =
        simple_virtual_sequencer.type_id.create("v_sequencer", this);
    endfunction : build_phase

    // Connect virtual sequencer to subsequencers.

    function void connect_phase();
        v_sequencer.cpu_seqr = cpu0.master[0].sequencer;
        UVMConfigDb.set(this,“v_sequencer”, “eth_seqr”,eth0.tx_rx_agent.sequencer);
    endfunction : connect_phase

endclass: simple_tb

4.8 Checking for DUT Correctness ¶

Getting the device into desired states is a significant part of verification. The environment should verify valid responses from the DUT before a feature is declared verified. Two types of auto-checking mechanisms can be used:

a) Assertions—Derived from the specification or from the implementation and ensure correct timing

behavior. Assertions typically focus on signal-level activity. b) Data checkers—Ensure overall device correctness.

As was mentioned in Section 1.1.9, checking and coverage should be done in the passive domain independently of the active logic. Reusable assertions are part of reusable components. See Chapter 3 for more information. Designers can also place assertions in the DUT RTL. Refer to your ABV documentation for more information.

4.9 Scoreboards ¶

A crucial element of a self-checking environment is the scoreboard. Typically, a scoreboard verifies the proper operation of the design at a functional level. The responsibility of a scoreboard varies greatly depending on the implementation. This section will show an example of a scoreboard that verifies that a given UBus slave interface operates as a simple memory. While the memory operation is critical to the UBus demonstration environment, the focus of this section is on the steps necessary to create and use a scoreboard in an environment so those steps can be repeated for any scoreboard application.

UBus Scoreboard Example

For the UBus demo environment, a scoreboard is necessary to verify the slave agent is operating as a simple memory. The data written to an address should be returned when that address is read. The desired topology is shown in Figure 20.

In this example, the user has created a top-level environment with one UBus environment that contains the bus monitor, one active master agent, and one active slave agent. Every component in the UBus environment is created using the build_phase() methods defined by the IP developer.

4.9.1 Creating the Scoreboard¶

Before the scoreboard can be added to the ubus_example_env, the scoreboard component must be defined.

To define the scoreboard:

a) Add the TLM export necessary to communicate with the environment monitor(s). b) Implement the necessary functions and tasks required by the TLM export. c) Define the action taken when the export is called.

4.9.2 Adding Exports to uvm_scoreboard¶

In the example shown in Figure 20, the scoreboard requires only one port to communicate with the environment. Since the monitors in the environment have provided an analysis port write() interface via the TLM uvm_analysis_port(s), the scoreboard will provide the TLM uvm_analysis_imp.

The ubus_example_scoreboard component derives from the uvm_scoreboard and declares and instantiates an analysis_imp. For more information on TLM interfaces, see “TLM Interfaces” in the UVM 1.2 Class Reference. The declaration and creation is done inside the constructor.

1 class ubus_example_scoreboard extends uvm_scoreboard; 2 uvm_analysis_imp #(ubus_transfer, ubus_example_scoreboard) 3 item_collected_export; 4 … 5 function new (string name, uvm_component parent); 6 super.new(name, parent);

item_collected_export

ubus0

master0

sequencer

P

driver

slave0

P

monitor

sequencer

driver

P

monitor

7 endfunction : new 8 function void build_phase(uvm_phase phase); 9 item_collected_export = new(“item_collected_export”, this); 10 endfunction 11 …

Line 2 declares the uvm_analysis_export. The first parameter, ubus_transfer, defines the uvm_object communicated via this TLM interface. The second parameter defines the type of this implementation’s parent. This is required so that the parent’s write() method can be called by the export.

Line 9 creates the implementation instance. The constructor arguments define the name of this implementation instance and its parent.

4.9.3 Requirements of the TLM Implementation¶

Since the scoreboard provides an uvm_analysis_imp, the scoreboard must implement all interfaces required by that export. This means that the implementation for the write virtual function needs to be defined. For the ubus_example_scoreboard, write() has been defined as:

virtual function void write(ubus_transfer trans);

if (!disable_scoreboard)

memory_verify(trans); endfunction : write

The write() implementation defines what happens when data is provided on this interface. In this case, if disable_scoreboard is 0, the memory_verify() function is called with the transaction as the argument.

4.9.4 Defining the Action Taken¶

When the write port is called via write(), the implementation of write() in the parent of the implementation is called. For more information, see “TLM Interfaces” in the UVM 1.2 Class Reference. As seen in Section 4.9.3, the write() function is defined to called the memory_verify() function if disable_scoreboard is set to 0.

The memory_verify() function makes the appropriate calls and comparisons needed to verify a memory operation. This function is not crucial to the communication of the scoreboard with the rest of the environment and not discussed here. The ubus_example_scoreboard.sv file shows the implementation.

4.9.5 Adding the Scoreboard to the Environment¶

Once the scoreboard is defined, the scoreboard can be added to the UBus example top-level environment. First, declare the ubus_example_scoreboard inside the ubus_example_env class.

ubus_example_scoreboard scoreboard0;

After the scoreboard is declared, it can be constructed inside the build() phase:

function ubus_example_env::build_phase(uvm_phase phase);

… scoreboard0 = ubus_example_scoreboard.type_id.create(“scoreboard0”,

this);

… endfunction

Here, the scoreboard0 of type ubus_example_scoreboard is created using the create() function and given the name scoreboard0. It is then assigned the ubus_example_tb as its parent.

After the scoreboard is created, the ubus_example_env can connect the port on the UBus environment slaves[0] monitor to the export on the scoreboard.

function ubus_example_env::connect_phase(uvm_phase phase);

… ubus0.slaves[0].monitor.item_collected_port.connect(

scoreboard0.item_collected_export); … endfunction

This ubus_example_env’s connect() function code makes the connection, using the TLM ports connect() interface, between the port in the monitor of the slaves[0] agent inside the ubus0 environment and the implementation in the ubus_example_scoreboard called scoreboard0. For more information on the use of binding of TLM ports, see “TLM Interfaces” in the UVM 1.2 Class Reference.

4.9.6 Summary¶

The process for adding a scoreboard in this section can be applied to other scoreboard applications in terms of environment communication. To summarize:

Create the scoreboard component.

1) Add the necessary exports. 2) Implement the required functions and tasks. 3) Create the functions necessary to perform the implementation-specific functionality. b) Add the scoreboard to the environment.

1) Declare and instantiate the scoreboard component. 2) Connect the scoreboard implementation(s) to the environment ports of interest.

The UBus demo has a complete scoreboard example. See Chapter 7 for more information.

4.10 Implementing a Coverage Model

To ensure thorough verification, observers should be used to represent the verification goals. SystemVerilog provides a rich set of functional-coverage features.

4.10.1 Selecting a Coverage Method

No single coverage metric ensures completeness. There are two coverage methods:

a) Explicit coverage—is user-defined coverage. The user specifies the coverage goals, the needed val- ues, and collection time. As such, analyzing these goals is straightforward. Completing all your cov- erage goals should be one of the metrics used to determine the completion of the DUT’s verification. An example of such a metric is SystemVerilog functional coverage. The disadvantage of such met- rics is that missing goals are not taken into account.

b) Implicit coverage—is done with automatic metrics that are driven from the RTL or other metrics already existing in the code. Typically, creating an implicit coverage report is straightforward and does not require a lot of effort. For example, code coverage, expression coverage, and FSM (finite- state machine) coverage are types of implicit coverage. The disadvantage of implicit coverage is it is difficult to map the coverage requirements to the verification goals. It also is difficult to map cover- age holes into unexecuted high-level features. In addition, implicit coverage is not complete, since it does not take into account high-level abstract events and does not create associations between paral- lel threads (that is, two or more events occurring simultaneously).

Starting with explicit coverage is recommended. You should build a coverage model that represents your high-level verification goals. Later, you can use implicit coverage as a “safety net” to check and balance the explicit coverage.

NOTE—Reaching 100% functional coverage with very low code-coverage typically means the functional coverage needs to be refined and enhanced.

4.10.2 Implementing a Functional Coverage Model¶

A verification component should come with a protocol-specific functional-coverage model. You may want to disable some coverage aspects that are not important or do not need to be verified. For example, you might not need to test all types of bus transactions in your system or you might want to remove that goal from the coverage logic that specifies all types of transactions as goals. You might also want to extend the functional-coverage model and create associations between the verification component coverage and other attributes in the system or other interface verification components. For example, you might want to ensure proper behavior when all types of transactions are sent and the FIFO in the system is full. This would translate into crossing the transaction type with the FIFO-status variable. This section describes how to implement this type of functional coverage model.

4.10.3 Enabling and Disabling Coverage

The verification IP developer should provide configuration properties that allow the integrator or test writer to control aspects of the coverage model (see Section 3.12.3). The VIP documentation should cover what properties can be set to affect coverage. The most basic of controls would determine whether coverage is collected at all. The UBus monitors demonstrate this level of control. To disable coverage before the environment is created, use the UVMConfigDb() interface:

UVMConfigDb#(int)::(this,"ubus0.masters[0].monitor", "coverage_enable", 0)

Once the environment is created, the property can be set directly:

ubus0.masters[0].monitor.coverage_enable = 0;

5. Using the Register Layer Classes ¶

5.1 Overview ¶

The UVM register layer classes are used to create a high-level, object-oriented model for memory-mapped registers and memories in a design under verification (DUV). The UVM register layer defines several base classes that, when properly extended, abstract the read/write operations to registers and memories in a DUV. This abstraction mechanism allows the migration of verification environments and tests from block to system levels without any modifications. It also can move uniquely named fields between physical registers without requiring modifications in the verification environment or tests. Finally, UVM provides a register test sequence library containing predefined testcases you can use to verify the correct operation of registers and memories in a DUV.

A register model is typically composed of a hierarchy of blocks that map to the design hierarchy. Blocks can contain registers, register files and memories, as well as other blocks. The register layer classes support front-door and back-door access to provide redundant paths to the register and memory implementation, and verify the correctness of the decoding and access paths, as well as increased performance after the physical access paths have been verified. Designs with multiple physical interfaces, as well as registers, register files, and memories shared across multiple interfaces, are also supported.

Most of the UVM register layer classes must be specialized via extensions to provide an abstract view that corresponds to the actual registers and memories in a design. Due to the large number of registers in a design and the numerous small details involved in properly configuring the UVM register layer classes, this specialization is normally done by a model generator. Model generators work from a specification of the registers and memories in a design and thus are able to provide an up-to-date, correct-by-construction register model. Model generators are outside the scope of the UVM library.

Figure 21 shows how a register model is used in a verification environment.

5.2 Usage Model ¶

A register model is an instance of a register block, which may contain any number of registers, register files, memories, and other blocks. Each register file contains any number of registers and other register files. Each register contains any number of fields, which mirror the values of the corresponding elements in hardware.

For each element in a register model–field, register, register file, memory or block–there is a class instance that abstracts the read and write operations on that element.

Figure 22 shows the class collaboration diagram of the register model.

A block generally corresponds to a design component with its own host processor interface(s), address decoding, and memory-mapped registers and memories. If a memory is physically implemented externally to the block, but accessed through the block as part of the block’s address space, then the memory is considered as part of the block register model.

All data values are modeled as fields. Fields represent a contiguous set of bits. Fields are wholly contained in a register. A register may span multiple addresses. The smallest register model that can be used is a block. A block may contain one register and no memories, or thousands of registers and gigabytes of memory. Repeated structures may be modeled as register arrays, register file arrays, or block arrays.

Figure 23 shows the structure of a sample design block containing two registers, which have two and three fields respectively, an internal memory, and an external memory. Figure 24 shows the structure of the corresponding register model.

When using a register model, fields, registers, and memory locations are accessed through read and write methods in their corresponding abstraction class. It is the responsibility of the register model to turn these abstracted accesses into read and write cycles at the appropriate addresses via the appropriate bus driver. A register model user never needs to track the specific address or location of a field, register, or memory location, only its name.

For example, the field ADDR in the CONFIG register shown in Figure 23 can be accessed through the register model shown in Figure24 using the CODEC.CONFIG.ADDR.read() method. Similarly, location 7 in the BFR memory can be accessed using the CODEC.BFR.write(7,value) method.

The location of fields within a physical register is somewhat arbitrary. If a field name is unique across all registers’ fields within a block, it may also be accessed independently of their register location using an alias handle declared in the block. For example, the same ADDR field, being unique in name to all other fields in the CODEC block, may also be accessed using CODEC.ADDR.read(). Then, if ADDR is relocated from CONFIG to another register, any tests or environments that reference CODEC.ADDR will not be affected. Because a typical design has hundreds if not thousands of fields, the declaration and assignment of field aliases in a block are left as an optional feature in a register model generator.

5.2.1 Sub-register Access¶

When reading or writing a field using uvm_reg_field::read() or uvm_reg_field::write(), what actually happens depends on a lot of factors. If possible, only that field is read or written. Otherwise, the entire register containing that field is read or written, possibly causing unintended side effects to the other fields contained in that same register.

Consider the 128-bit register shown in Figure 25. Assuming a 32-bit data bus with a little-endian layout, accessing this entire register requires four cycles at addresses 0x00, 0x04, 0x08, and 0x0C respectively. However, field D can be accessed using a single cycle at address 0x01. Since this field occupies an entire physical address, accessing it does not pose a challenge.

Similarly, accessing field C can be done using a single access at address 0x00. However, this will also access fields B and A. Accessing field F requires two physical accesses, at addresses 0x02 and 0x03, but this would also access fields E and G at the same time. Accessing adjacent fields might not be an issue, but if the access has a side-effect on any of these fields, such as a clear-on-read field or writable field, this process will have unintended consequences.

When the underlying bus-protocol supports byte-enabling, field C (at address 0x00, lane #2) can be accessed without affecting the other fields at the same address. And since field F is byte-aligned, it can be accessed without side effects by accessing address 0x02, lane #3 and address 0x03, lane #0. However, fields B and A remain inaccessible without mutual side effects as they do not individually occupy an entire byte lane.

Thus, individual field access is supported for fields that are the sole occupant of one or more byte lane(s) if the containing register does not use a user-defined front-door and the underlying bus protocol supports byte enabling. A field may also be individually-accessible if the other fields in the same byte lanes are not affected by read or write operations. Whether a field can be individually accessible (assuming the underlying protocol supports byte-enabling) is specified by the register model generator in the uvm_reg_field::configure() method.

For individual field access to actually occur, two conditions must be met: the field must be identified as being the sole occupant of its byte lane by the register model generator via the uvm_reg_field::configure() method and the bus protocol must report that it supports byte- enables via the uvm_reg_adapter::supports_byte_enable property.

Finally, individual field access is only supported for front-door accesses. When using back-door accesses, the entire register–and thus all the fields it contains–will always be accessed via a peek-modify-poke operation.

5.2.2 Mirroring¶

The register model maintains a mirror of what it thinks the current value of registers is inside the DUT. The mirrored value is not guaranteed to be correct because the only information the register model has is the read and write accesses to those registers. If the DUT internally modifies the content of any field or register through its normal operations (e.g., by setting a status bit or incrementing an accounting counter), the mirrored value becomes outdated.

The register model takes every opportunity to update its mirrored value. For every read operation, the mirror for the read register is updated. For every write operation, the new mirror value for the written register is predicted based on the access modes of the bits in the register (read/write, read-only, write-1- to-clear, etc.). Resetting a register model sets the mirror to the reset value specified in the model. A mirror is not a scoreboard, however; while a mirror can accurately predict the content of registers that are not updated by the design, it cannot determine if an updated value is correct or not.

You can update the mirror value of a register to the value stored in the DUT by using the uvm_reg_field::mirror(), uvm_reg::mirror(), or uvm_reg_block::mirror() methods. Updating the mirror for a field also updates the mirror for all the other fields in the same register. Updating the mirror for a block updates the mirror for all fields and registers it contains. Updating a mirror in a large block may take a lot of simulation time if physical read cycles are used; whereas, updating using back-door access usually takes zero-time.

You can write to mirrored values in the register model in zero-time by using the UVMRegField.set or UVMReg.set methods. Once a mirror value has been overwritten, it no longer reflects the value in the corresponding field or register in the DUT. You can update the DUT to match the mirror values by using the uvm_reg::update() or uvm_reg_block::update() methods. If the new mirrored value matches the old mirrored value, the register is not updated, thus saving unnecessary bus cycles. Updating a block with its mirror updates all the fields and registers the block contains with their corresponding mirror values. Updating a large block may take a lot of simulation time if physical write cycles are used; whereas, updating using back-door access usually takes zero-time. It is recommended you use this update-from-mirror process when configuring the DUT to minimize the number of write operations performed.

To access a field or register’s current mirror value in zero-time, use the UVMRegField.get or UVMReg.get methods. However, if UVMRegField.set or UVMReg.set is used to write a desired value to the DUT, get() only returns the desired value, modified according to the access mode for that field or register, until the actual write to the DUT has taken place via update().

5.2.3 Memories are not Mirrored¶

Memories can be quite large, so they are usually modeled using a sparse-array approach. Only the locations that have been written to are stored and later read back. Any unused memory location is not modeled. Mirroring a memory would require that the same technique be used.

When verifying the correct operations of a memory, it is necessary to read and write all addresses. This negates the memory-saving characteristics of a sparse-array technique, as both the memory model of the DUT and the memory would mirror, become fully populated, and duplicate the same large amount of information.

Unlike bits in fields and registers, the behavior of bits in a memory is very simple: all bits of a memory can either be written to or not. A memory mirror would then be a ROM or RAM memory model–a model that is already being used in the DUT to model the memory being mirrored. The memory mirror can then be replaced by providing back-door access to the memory model.

Therefore, using the uvm_mem::peek() or uvm_mem::poke() methods provide the exact same functionality as a memory mirror. Additionally, unlike a mirror based on observed read and write operations, using back-door accesses instead of a mirror always returns or sets the actual value of a memory location in the DUT.

5.3 Access API ¶

Register and fields have a variety of methods to get the current value of a register or field and modify it. It is important to use the correct API to obtain the desired result.

5.3.1 read / write¶

The normal access API are the read() and write() methods. When using the front-door (path=BFM), one or more physical transactions is executed on the DUT to read or write the register. The mirrored value is then updated to reflect the expected value in the DUT register after the observed transactions.

When using the back-door (path=BACKDOOR), peek or poke operations are executed on the DUT to read or write the register via the back-door mechanism, bypassing the physical interface. The behavior of the registers is mimicked as much as possible to duplicate the effect of reading or writing the same value via the front-door. For example, a read from a clear-on-read field causes 0’s to be poked back into the field after the peek operation. The mirrored value is then updated to reflect the actual sampled or deposited value in the register after the observed transactions.

Desired Value

predict(READ/WRITE)

Mirrored Value

read/write(frontdoor)

Bus Agent

Register in DUT

5.3.2 peek / poke¶

Using the peek() and poke() methods reads or writes directly to the register respectively, which bypasses the physical interface. The mirrored value is then updated to reflect the actual sampled or deposited value in the register after the observed transactions.

5.3.3 get / set¶

Using the get() and set() methods reads or writes directly to the desired mirrored value respectively, without accessing the DUT. The desired value can subsequently be uploaded into the DUT using the update() method.

5.3.4 randomize¶

Using the randomize() method copies the randomized value in the uvm_reg_field::value property into the desired value of the mirror by the post_randomize() method. The desired value can subsequently be uploaded into the DUT using the update() method.

Desired Value

Mirrored Value

read/write(backdoor) Mimic

Register in DUT Desired Value

Mirrored Value

peek/poke()

Register in DUT Desired Value

Mirrored Value

get/set()

Register in DUT

Desired Value

Mirrored Value

randomize() Register in DUT 5.3.5 update

Using the update() method invokes the write() method if the desired value (previously modified using set() or randomize()) is different from the mirrored value. The mirrored value is then updated to reflect the expected value in the register after the executed transactions.

Desired Value

update() If different

Mirrored Value

Register in DUT

5.3.6 mirror¶

Using the mirror() method invokes the read() method to update the mirrored value based on the readback value. mirror() can also compare the readback value with the current mirrored value before updating it.

Desired Value

mirror()

Mirrored Value

Register in DUT

5.3.7 Concurrent Accesses¶

The register model can be accessed from multiple concurrent execution threads. However, it internally serializes the access to the same register to ensure predictability of the implicitly-updated mirrored value and of the other field values in the same register when a individual field is accessed.

A semaphore in each register ensures it can be read or written by only one process at a time. Any other process attempting access will block and not resume until after the current operation completes and after other processes that were blocked before it have completed their operations.

If a thread in the middle of executing a register operation is explicitly killed, it will be necessary to release the semaphore in the register it was accessing by calling the uvm_reg::reset() method.

5.4 Coverage Models ¶

The UVM register library classes do not include any coverage models as a coverage model for a register will depend on the fields it contains and the layout of those fields, and a coverage model for a block will depend on the registers and memories it contains and the addresses where they are located. Since coverage model information is added to the UVM register library classes by the register model generator, that generator needs to include a suitable coverage model. Consequently, the UVM register library classes provide the necessary API for a coverage model to sample the relevant data into a coverage model.

Due to the significant memory and performance impact of including a coverage model in a large register model, the coverage model needs to handle the possibility that specific cover groups will not be instantiated or to turn off coverage measurement even if the cover groups are instantiated. Therefore, the UVM register library classes provide the necessary API to control the instantiation and sampling of various coverage models.

5.4.1 Predefined Coverage Identifiers¶

The UVM library has several predefined functional coverage model identifiers, as shown in Table 6. Each symbolic value specifies a different type of coverage model. The symbolic values use a one-hot encoding. Therefore, multiple coverage models may be specified by OR’ing them. Additional symbolic values may be provided for vendor-specific and user-specific coverage models that fall outside of the pre-defined coverage model types. To avoid collisions with pre-defined UVM, vendor-defined, and user-defined coverage model identifiers, bits 0 through 7 are reserved for UVM, bits 8 through 15 are reserved for vendors, and bits 16 through 23 are reserved for users. Finally, bits 24 and above are reserved for future assignment.

Table 6–Pre-defined Functional Coverage Type Identifiers

Identifier Description

UVM_NO_COVERAGE No coverage models.

UVM_CVR_REG_BITS Coverage models for the bits read or written in registers.

UVM_CVR_ADDR_MAP Coverage models for the addresses read or written in an address map.

UVM_CVR_FIELD_VALS Coverage models for the values of fields.

UVM_CVR_ALL All coverage models.

5.4.2 Controlling Coverage Model Construction and Sampling¶

By default, coverage models are not included in a register model when it is instantiated. To be included, they must be enabled via the uvm_reg::include_coverage() method. It is recommended register-level coverage models are only included in unit-level environments; block-level coverage models may be included in block and system-level environments:

UVMReg.include_coverage(“\*”, UVM_CVR_REG_BITS + UVM_CVR_FIELD_VALS)

Furthermore, the sampling for a coverage model is implicitly disabled by default. To turn the sampling for specific coverage models on or off, use the UVMRegBlock.set_coverage, UVMReg.set_coverage, and UVMMem.set_coverage methods.

5.5 Constructing a Register Model ¶

This section describes how to construct a UVM register model to represent different register and memory access and composition structures. The target audience for this section is generator writers. End users of the register model need not be familiar with the model construction process, only the final structure of the model.

5.5.1 Field Types¶

There is usually no need to construct field types. Fields are simple instantiations of the uvm_reg_field class. A field type may only be needed to specify field-level constraints, which could also be specified in the containing register.

A field type is constructed using a class extended from the uvm_reg_field class. There must be one class per unique field type. The name of the field type is created by the register model generator. The name of the field type class must be unique within the scope of its declaration. The field type class must include an appropriate invocation of the uvm_object_utils():

class my_fld_type(UVMRegField):
    ...
uvm_object_utils(my_fld_type) # Outside the class

Field types are instantiated in the build() method of the containing register types.

5.5.1.1 Class Properties and Constraints¶

A separate constraint block should be defined for each aspect being constrained–e.g., one to keep it valid, one to keep it reasonable–so they can be turned off individually. The name of a constraint block shall be indicative of its purpose. Constraints shall constrain the value class property. Additional state variables may be added to the field class if they facilitate the constraints:

class my_fld_type extends uvm_reg_field;
    constraint valid {
        value inside {0, 1, 2, 4, 8, 16, 32};
    };
endclass

To ensure state variable and constraint block names do not collide with other symbols in uvm_reg_field base class, it is recommended their names be in all UPPERCASE.

If the post_randomize() method is overridden, it must call super.post_randomize().

5.5.1.2 Constructor¶

The constructor must be a valid uvm_object constructor. The constructor shall call the UVMRegField.__init__ method with appropriate argument values for the field type:

class my_fld_type extends uvm_reg_field;
    function new(string name = “my_fld_type”);
        super.new(name);
    endfunction
endclass

5.5.1.3 Predefined Field Access Policies¶

The access policy of a field is specified using the uvm_reg_field::configure() method, called from the build() method of the register that instantiates it.

Table 7 shows the pre-defined access policies for uvm_reg_field. Unless otherwise stated, the effect of a read cycle on the current value is performed after the current value has been sampled for read-back. Additional field access policies may be defined using the uvm_reg_field::define_access() method and by modeling their behavior by extending the uvm_reg_field or uvm_reg_cbs classes.

Table 7–Pre-defined Field Access Policies

**Access Policy Description Effect of a Write **

on Current Field Value

Effect of a Read on Current Field Value

Read- back Value

RO Read Only No effect. No effect. Current

value

RW Read, Write Changed to written value. No effect. Current

value

RC Read Clears All No effect. Sets all bits to 0’s. Current

value

RS Read Sets All No effect. Sets all bits to 1’s. Current

value

WRC Write,

Read Clears All

Changed to written value. Sets all bits to 0’s. Current

value

WRS Write, Read Sets All

Changed to written value. Sets all bits to 1’s. Current

value

WC Write Clears All Sets all bits to 0’s. No effect. Current

value

WS Write Sets All Sets all bits to 1’s. No effect. Current

value

WSRC Write Sets All,

Read Clears All

Sets all bits to 1’s. Sets all bits to 0’s. Current

value

WCRS Write Clears All,

Read Sets All

Sets all bits to 0’s. Sets all bits to 1’s. Current

value

W1C Write 1 to Clear If the bit in the written value is a 1, the cor- responding bit in the field is set to 0. Other- wise, the field bit is not affected.

No effect. Current

value

W1S Write 1 to Set If the bit in the written value is a 1, the cor- responding bit in the field is set to 1. Other- wise, the field bit is not affected.

No effect. Current

value

W1T Write 1 to Toggle If the bit in the written value is a 1, the cor-

responding bit in the field is inverted. Other- wise, the field bit is not affected.

No effect. Current

value

W0C Write 0 to Clear If the bit in the written value is a 0, the cor- responding bit in the field is set to 0. Other- wise, the field bit is not affected.

No effect. Current

value

Table 7–Pre-defined Field Access Policies (Continued)

**Access Policy Description Effect of a Write **

on Current Field Value

5.5.1.4 IP-XACT Field Access Mapping¶

Table 8 – Table 12 show the mapping between the IEEE 1685-2009 field access policies (as specified by the fieldData group, see section 6.10.9.2 of the IEEE 1685-2009 Standard) and the pre-defined access policies for uvm_reg_field. Several combinations of access, modifiedWriteValue, and readAction are specified as n/a because they do not make practical sense. However, they can always be modeled as user-defined fields if they are used in a design.

Effect of a Read on Current Field Value

Read- back Value

W0S Write 0 to Set If the bit in the written value is a 0, the cor- responding bit in the field is set to 1. Other- wise, the field bit is not affected.

No effect. Current

value

W0T Write 0 to Toggle If the bit in the written value is a 0, the cor-

responding bit in the field is inverted. Other- wise, the field bit is not affected.

No effect. Current

value

W1SRC Write 1 to Set,

Read Clears All

If the bit in the written value is a 1, the cor- responding bit in the field is set to 1. Other- wise, the field bit is not affected.

Sets all bits to 0’s. Current

value

W1CRS Write 1 to Clear,

Read Sets All

If the bit in the written value is a 1, the cor- responding bit in the field is set to 0. Other- wise, the field bit is not affected.

Sets all bits to 1’s. Current

value

W0SRC Write 0 to Set,

Read Clears All

If the bit in the written value is a 0, the cor- responding bit in the field is set to 1. Other- wise, the field bit is not affected.

Sets all bits to 0’s. Current

value

W0CRS Write 0 to Clear,

Read Sets All

If the bit in the written value is a 0, the cor- responding bit in the field is set to 0. Other- wise, the field bit is not affected.

Sets all bits to 1’s. Current

value

WO Write Only Changed to written value. No effect. Undefined

WOC Write Only

Clears All

Sets all bits to 0’s. No effect. Undefined

WOS Write Only Sets All

Sets all bits to 1’s. No effect. Undefined

W1 Write Once Changed to written value if this is the first

write operation after a hard reset. Otherwise, has no effect.

No effect. Current

value

WO1 Write Only, Once Changed to written value if this is the first

write operation after a hard reset. Otherwise, has no effect.

No effect. Undefined

Table 8–IP-XACT Mapping for access==read-write

access == read-write

modifiedWriteValue

readAction

Unspecified clear set modify

Unspecified RW WRC WRS User-defined

oneToClear W1C n/a W1CRS User-defined

oneToSet W1S W1SRC n/a User-defined

oneToToggle W1T n/a n/a User-defined

zeroToClear W0C n/a W0CRS User-defined

zeroToSet W0S W0SRC n/a User-defined

zeroToToggle W0T n/a n/a User-defined

clear WC n/a WCRS User-defined

set WS WSRC n/a User-defined

modify User-defined User-defined User-defined User-defined

Table 9–IP-XACT Mapping for access==read-only

access == read-only

modifiedWriteValue

readAction

Unspecified clear set modify

Unspecified RO RC RS User-defined

All others n/a n/a n/a n/a

Table 10–IP-XACT Mapping for access==write-only

access == write-only

modifiedWriteValue

readAction

Unspecified clear set modify

Unspecified WO n/a n/a n/a

clear WOC n/a n/a n/a

set WOS n/a n/a n/a

All others n/a n/a n/a n/a

Table 11–IP-XACT Mapping for access==read-writeOnce

access == read-writeOnce

modifiedWriteValue

5.5.1.5 Reserved Fields¶

There is no pre-defined field access policy for reserved fields. That is because —reserved” is a documentation concept, not a behavioral specification. Reserved fields should be left unmodelled (where they will be assumed to be RO fields filled with 0’s), modeled using an access policy that corresponds to their actual hardware behavior or not be compared using uvm_reg_field::set_compare(0).

If reserved fields are not to be used, they should be identified with the NO_REG_TESTS attribute and have a user-defined behavior extension that will issue an error message if they are used.

5.5.1.6 User-defined Field Access Policy¶

The UVM field abstraction class contains several predefined field access modes. The access modes are used, in combination with observed read and write operations, to determine the expected value of a field. Although most fields fall within one of the predefined access policies, it is possible to design a field that behaves predictably but differently from the predefined ones.

New access policy identifiers can similarly be defined to document the user-defined behavior of the field.

class protected_field extends uvm_reg_field;

local static bit m_protected = define_access(“Protected”); … endclass

The behavior of the user-defined field access policy is implemented by extending the pre/post read/write virtual methods in the field abstraction class or in the field callback methods. For example, a protected field that can only be written if another field is set to a specific value can be modeled as shown below:

class protected_field extends uvm_reg_field;

local uvm_reg_field protect_mode;

readAction

Unspecified clear set modify

Unspecified W1 n/a n/a n/a

All others n/a n/a n/a n/a

Table 12–IP-XACT Mapping for access==writeOnce

access == writeOnce

modifiedWriteValue

readAction

Unspecified clear set modify

Unspecified WO1 n/a n/a n/a

All others n/a n/a n/a n/a

virtual task pre_write(uvm_reg_item rw);

// Prevent the write if protect mode is ON if (protect_mode.get()) begin

rw.value = value; endtask endclass

You can modify the behavior of any field to the user-specified behavior by registering a callback extension with that field. First, define the callback class; then, register an instance of it with an instance of the field whose behavior you want to modify:

class protected_field_cb extends uvm_reg_cbs;

local uvm_reg_field protect_mode; virtual task pre_write(uvm_reg_item rw);

// Prevent the write if protect mode is ON if (protect_mode.get()) begin

uvm_reg_field field; if ($cast(field,rw.element)) rw.value = field.get(); end endtask endclass

protected_field_cb protect_cb = new(“protect_cb”,protect_mode) uvm_callbacks#(my_field_t, uvm_reg_cbs)::add(my_field, protect_cb);

5.5.1.7 Field Usage vs. Field Behavior¶

The access mode of a field is used to specify the physical behavior of the field so the mirror can track, as best as it can, the value of the field. It is not designed to specify whether or not it is suitable or functionally correct to use the field in a particular fashion from the application’s perspective.

For example, a configuration field could be designed to be written only once by the software after the design comes out of reset. If the design does not support dynamic reprovisioning, it may not be proper to subsequently modify the value of that configuration field. Whether the field should be specified as write- once (W1) depends on the hardware functionality. If the hardware does not prevent the subsequent write operation, the field should be specified as read-write as that would accurately reflect the actual behavior of the field.

To include usage assertions to specific fields (e.g., specifying a configuration field is never written to more than once despite the fact that doing so is physically possible), implement that check in an extension of the field abstraction class, but do not prevent the mirror from reflecting that value in the hardware:

class config_once_field extends uvm_reg_field;

    local bit m_written = 0;
    virtual task pre_write(uvm_reg_item rw);
        if (m_written)
            ’uvm_error(field.get_full_name(), "..."); m_written = 1;
    endtask: pre_write

    virtual function reset(string kind = “HARD”);
        if (has_reset(kind)) m_written = 0;
        super.reset(kind);
    endfunction
endclass

5.5.2 Register Types¶

A register type is constructed using a class extended from the uvm_reg class. There must be one class per unique register type. The name of the register type is created by the register model generator. The name of the register type class must be unique within the scope of its declaration. The register type class must include an appropriate invocation of the `uvm_object_utils() macro.

class my_reg_type extends uvm_reg;

`uvm_object_utils(my_reg_type) endclass

Register types are instantiated in the build() method of the block and register file types.

5.5.2.1 Class Properties and Constraints¶

The register type must contain a public class property for each field it contains. The name of the field class property shall be the name of the field. The field class property shall have the rand attribute. Field class properties may be arrays.

class my_reg_type extends uvm_reg;

rand uvm_reg_field F1; rand uvm_reg_field F2[3]; endclass

To ensure field names do not collide with other symbols in the uvm_reg base class, it is recommended their names be in all UPPERCASE.

Constraints, if any, should be defined in separate blocks for each aspect being constrained. This allows them to be turned off individually. The name of a constraint block shall be indicative of its purpose. Constraints shall constrain the value class property of each field in the register. Additional state variables may be added to the register type class if they facilitate the constraints. If the post_randomize() method is overridden, it must call the super.post_randomize() method.

If a register has only one field, then you would not want to have to write:

R.randomize() with (value.value == 5);

Instead, instantiate a private dummy field and include a rand class property named value in the register class. A constraint shall keep the value class property equal to the field’s value class property.

class my_reg_type extends uvm_reg;

rand uvm_reg_data_t value; local rand uvm_reg_field _dummy;

constraint _dummy_is_reg {

_dummy.value == value; } endclass

Then, randomizing an instance of the register looks like the more natural:

R.randomize() with (value == 5);

5.5.2.2 Constructor¶

The constructor must be a valid uvm_object constructor. The constructor shall call the uvm_reg::new() method with appropriate argument values for the register type.

class my_reg_type extends uvm_reg;

function new(string name = “my_reg_type”);

super.new(.name(name), .n_bits(32), .has_coverage(UVM_NO_COVERAGE)); endfunction endclass

5.5.2.3 Build() Method¶

A virtual build() function, with no arguments, shall be implemented.

The build() method shall instantiate all field class properties using the class factory. Because the register model is a uvm_object hierarchy, not a uvm_component hierarchy, no parent reference is specified and the full hierarchical name of the register type instance is specified as the context. The build() method shall call the uvm_reg_field::configure() method for all field class properties with the appropriate argument values for the field instance and specifying this as the field parent.

class my_reg_type extends uvm_reg;

virtual function build();

this.F1 = uvm_reg_field.type_id.create(.name(“F1”),

.parent(null), .contxt(get_full_name())); this.F1.configure(this, …);

endfunction endclass

5.5.2.4 Additional Methods¶

Register model generators are free to add access methods to abstract common operations, For example, a read-modify-write method could be added to a register model:

class my_reg extends uvm_reg;

… task RMW(output uvm_status_e status;

input uvm_reg_data_t data; input uvm_reg_data_t mask; …); uvm_reg_data_t tmp; read(status, tmp, …); tmp &= ~mask; tmp |= data & mask; write(status, tmp, …); endtask endclass

Although allowed by UVM, such additional methods are not part of the standard. To avoid collisions with class members that may be added in the future, the name of these methods should be in UPPERCASE or be given a generator-specific prefix.

5.5.2.5 Coverage Model¶

A register-level coverage model is defined and instantiated in the register type class. It measures the coverage of read and write operations and field values on each instance of that register type. The uvm_reg::sample() or uvm_reg::sample_values() methods shall be used to trigger the sampling of a coverage point based on the data provided as argument or data gathered from the current state of the register type instance. The sampling of the coverage model shall only occur if sampling for the corresponding coverage model has been turned on, as reported by the uvm_reg::get_coverage() method.

class my_reg extends uvm_reg;

protected uvm_reg_data_t m_current; protected uvm_reg_data_t m_data; protected bit m_is_read;

covergroup cg1;

… endgroup … virtual function void sample(uvm_reg_data_t data,

uvm_reg_data_r byte_en, bit is_read, uvm_reg_map map); if (get_coverage(UVM_CVR_REG_BITS)) begin

m_current = get(); m_data = data; m_is_read = is_read; cg1.sample(); endif endfunction endclass

All the coverage models that may be included in the register type shall be reported to the uvm_reg base class using the uvm_reg::build_coverage() method when super.new() is called or the uvm_reg::add_coverage() method. If no functional coverage models are included in the generated register type, UVM_NO_COVERAGE shall be specified. Register-level coverage groups shall only be instantiated in the constructor if the construction of the corresponding coverage model is enabled, as reported by the uvm_reg::has_coverage() method.

class my_reg_typ extends uvm_reg;

… covergroup cg1;

… endgroup

covergroup cg_vendor;

… endgroup

function new(string name = “my_reg_typ”);

super.new(name, 32, build_coverage(UVM_CVR_REG_BITS + VENDOR_CVR_REG)); if (has_coverage(UVM_CVR_REG_BITS))

cg1 = new(); if (has_coverage(VENDOR_CVR_REG))

cg_vendor = new(); endfunction

… endclass

The content, structure, and options of the coverage group is defined by the register model generator and is outside the scope of UVM.

5.5.3 Register File Types¶

A register file type is constructed using a class extended from the uvm_reg_file class. There must be one class per unique register file type. The name of the register file type is created by the register model generator. The name of the register file type class must be unique within the scope of its declaration. The register file type class must include an appropriate invocation of the `uvm_object_utils() macro.

class my_rf_type extends uvm_reg_file;

’uvm_object_utils(my_rf_type) endclass

Register file types are instantiated in the build() method of the block and register file types.

5.5.3.1 Class Properties¶

A register file type must contain a public class property for each register it contains. The name of the register class property shall be the name of the register. The type of the register class property shall be the name of the register type. Each register class property shall have the rand attribute. Register class properties may be arrays.:sub:`class my_rf_type extends uvm_reg_file; `

rand my_reg1_type R1; rand my_reg2_type R2[3]; endclass

Register files can contain other register files. A register file type must contain a public class property for each register file it contains. The name of the register file class property shall be the name of the register file. The type of the register file class property shall be the name of the register file type. The register file class property shall have the rand attribute. Register file class properties may be arrays.

class my_rf_type extends uvm_reg_file;

rand my_regfile1_type RF1; rand my_regfile2_type RF2[3]; endclass

To ensure register and register file names do not collide with other symbols in the register file abstraction base class, it is recommended their name be in all UPPERCASE or prefixed with an underscore (_).

The register file type may contain a constraint block for each cross-register constraint group it contains. The name of the constraint block shall be indicative of its purpose. Constraints shall constraint the uvm_reg_field::value class property of the fields in the registers contained in the register file. Additional state variables may be added to the register field type if they facilitate the constraints.

5.5.3.2 Constructor¶

The constructor must be a valid uvm_object constructor. The constructor shall call the uvm_ reg::configure() method with appropriate argument values for the register type.

class my_rf_type extends uvm_reg_file;

function new(string name = “my_rf_type”);

super.(name); endfunction endclass

5.5.3.3 build() Method¶

A virtual build() function, with no arguments, shall be implemented.

The build() method shall instantiate all register and register file class properties using the class factory. The name of the register or register file instance shall be prefixed with the name of the enclosing register file instance. Because the register model is a uvm_object hierarchy, not a uvm_component hierarchy, no parent reference is specified and the full hierarchical name of the block parent of the register file type instance is specified as the context. The build() method shall call the configure() method for all register and register file class properties, specifying get_block() for the parent block and this for the parent register file. The build() method shall call the build() method for all register and register file class properties.

class my_rf_type extends uvm_reg_file;

virtual function build();

uvm_reg_block blk = get_block();

this.RF1 = my_rf1_type.type_id.create(

.name($psprintf(“%s.rf1”, get_name())), .parent(null), .contxt(blk.get_full_name())); this.RF1.configure(get_block(), this, …); this.RF1.build(); this.RF1.add_hdl_path(); endfunction endclass

5.5.3.4 map() Method¶

A virtual map() function, with uvm_reg_map and address offset arguments, shall be implemented. The map() method shall call uvm_reg_map::add_reg() for all register class properties, adding the value of the address offset argument to the offset of the register in the register file. The map() method shall call the map() method of all register file class properties, adding the value of the address offset argument to the offset of the register file base offset. The map() method may call the add_hdl_path() method for all register or register file class properties with appropriate argument values for the register or register file instance.

class my_rf_type extends uvm_reg_file;

virtual function map(uvm_reg_map mp, uvm_reg_addr_t offset);

mp.add_reg(this.R1, offset + ’h04, …); mp.add_reg(this.R2, offset + ’h08, …); this.RF1.map(mp, offset + ’h200); endfunction endclass

5.5.3.5 set_offset() Method¶

A virtual set_offset() function, with a uvm_reg_map and address offset arguments, may also be implemented. The set_offset() method shall call the set_offset() method for all register and

register file class properties with appropriate argument values for the each instance, adding the value of the address offset argument to the offset of the register and register file base offset.

class my_rf_type extends uvm_reg_file;

virtual function set_offset(uvm_reg_map mp, uvm_reg_addr_t offset);

this.R1.set_offset(mp, offset + ’h04, …); this.R2.set_offset(mp, offset + ’h08, …); this.RF1.set_offset(mp, offset + ’h200); endfunction endclass

5.5.4 Memory Types¶

A memory type is constructed using a class extended from the uvm_mem class. There must be one class per unique memory type. The name of the memory type is created by the register model generator. The name of the memory type class must be unique within the scope of its declaration. The memory type class must include an appropriate invocation of the `uvm_object_utils() macro.

class my_mem_type extends uvm_mem;

`uvm_object_utils(my_mem_type) endclass

Memory types are instantiated in the build() method of the block and register file types.

5.5.4.1 Class Properties¶

The memory type need not contain any class property.

5.5.4.2 Constructor¶

The constructor must be a valid uvm_object constructor. The constructor shall call the uvm_ mem::new() method with appropriate argument values for the memory type.

class my_mem_type extends uvm_mem;

function new(string name = “my_mem_type”);

super.new(name, …); endfunction endclass

5.5.4.3 Coverage Model¶

A memory-level coverage model is defined and instantiated in the memory type class. It measures the coverage of the accessed offsets on each instance of that memory type. The uvm_mem::sample() method shall be used to trigger the sampling of a coverage point, based on the data provided as an argument or gathered from the current state of the memory type instance. The sampling of the coverage model shall only occur if sampling for the corresponding coverage model has been turned on, as reported by the uvm_mem::get_coverage(() method.

class my_mem extends uvm_mem;

local uvm_reg_addr_t m_offset;

covergroup cg_addr;

… endgroup …

virtual function void sample(uvm_reg_addr_t offset,

bit is_read, uvm_reg_map map); if (get_coveragen(UVM_CVR_ADDR_MAP)) begin

m_offset = offset; cg_addr.sample(); end endfunction endclass

All the coverage models that may be included in the memory type shall be reported to the uvm_mem base class using uvm_mem::build_coverage() when super.new() is called or using the uvm_mem::add_coverage() method. If no functional coverage models are included in the generated memory type, UVM_NO_COVERAGE shall be specified. Memory-level coverage groups shall only be instantiated in the constructor if the construction of the corresponding coverage model is enabled, as reported by the uvm_mem::has_coverage() method.

class my_mem extends uvm_mem;

… covergroup cg_addr;

… endgroup

function new(string name = “my_mem”);

super.new(name, …, build_coverage(UVM_CVR_ADDR_MAP)); if (has_coverage(UVM_CVR_ADDR_MAP))

cg_addr = new(); endfunction: new … endclass : my_mem

The content, structure, and options of the coverage group is defined by the register model generator and is outside the scope of UVM.

5.5.5 Block Types¶

A block type is constructed using a class extended from the uvm_reg_block class. There must be one class per unique block type. The name of the block type is created by the register model generator. The name of the block type class must be unique within the scope of its declaration. The block type class must include an appropriate invocation of the `uvm_object_utils() macro.

class my_blk_type extends uvm_reg_block;

`uvm_object_utils(my_blk_type) endclass

Block types are instantiated in the build() method of other block types and in verification environments.

5.5.5.1 Class Properties¶

The block type must contain a class property for each named address map it contains. The name of the address map class property shall be the name of the address map. The type of the address map class property shall be uvm_reg_map. The address map class property shall not have the rand attribute. Address map class properties shall not be arrays.

class my_blk_type extends uvm_reg_block;

uvm_reg_map AHB; uvm_reg_map WSH; endclass

The block type must contain a class property for each register it contains. The name of the register class property shall be the name of the register. The type of the register class property shall be the name of the register type. The register class property shall have the rand attribute. Register class properties may be arrays.:sub:`class my_blk_type extends uvm_reg_block; `

rand my_r1_type R1; rand my_r2_type R2[3]; endclass

The block type must contain a class property for each register file it contains. The name of the register file class property shall be the name of the register file. The type of the register file class property shall be the name of the register file type. The register file class property shall have the rand attribute. Register file class properties may be arrays.

class my_blk_type extends uvm_reg_block;

rand my_rf1_type RF1; rand my_rf2_type RF2[3]; endclass

The block type must contain a class property for each memory it contains. The name of the memory class property shall be the name of the memory. The type of the memory class property shall be the name of the memory type. The memory class property should not have the rand attribute. Memory class properties may be arrays.

class my_blk_type extends uvm_reg_block;

my_mem1_type RAM1; my_mem2_type RAM2[3]; endclass

The block type must contain a class property for each sub-block it contains. The name of the sub-block class property shall be the name of the sub-block. The type of the sub-block class property shall be the name of the sub-block type. The sub-block class property shall have the rand attribute. Sub-block class properties may be arrays.

class my_blk_type extends uvm_reg_block;

rand my_blk1_type BLK1; rand my_blk2_type BLK2[3]; endclass

To ensure register, register file, memory and block names do not collide with other symbols in uvm_reg_block base class, it is recommended their name be in all UPPERCASE or prefixed with an underscore (_).

Constraints, if any, should be defined in separate blocks for each aspect being constrained. This allows them to be turned off individually. The name of a constraint block shall be indicative of its purpose. Constraints shall constraint the uvm_reg_field::value class property of the fields in lower-level registers. Additional state variables may be added to the block type class if they facilitate the constraints. If the post_randomize() method is overridden, it must call the super.post_randomize() method.

5.5.5.2 Constructor¶

The constructor must be a valid uvm_object constructor. The constructor shall call the uvm_reg_block::new() method with appropriate argument values for the block type.

class my_blk_type extends uvm_reg_block;

function new(string name = “my_blk_type”);

super.new(.name(name), .has_coverage(UVM_NO_COVERAGE)); endfunction endclass

5.5.5.3 build() Method¶

A virtual build() function, with no arguments, shall be implemented.

The build() method shall instantiate all named address maps by calling the uvm_reg_block::create_map() method, specifying appropriate argument values for the address map in the block type. One of the named address maps shall be assigned to the uvm_reg_block::default_map class property.

class my_blk_type extends uvm_reg_block;

virtual function build();

this.AHB = create_map(); this.WSH = create_map(); this.default_map = this.AHB; endfunction endclass

If the block does not contain any named address maps, the build() method shall instantiate an anonymous address map by calling the uvm_reg_block::create_map() method, specifying the name of the address map and other appropriate argument values for the block type, and assign those to the uvm_reg_block::default_map class property.

class my_blk_type extends uvm_reg_block;

virtual function build();

this.default_map = create_map(.name(“default_map”,

.base_addr(’h1000), .n_bytes(4), .endian(UVM_LITTLE_ENDIAN)); endfunction endclass

The build() method shall instantiate all register, register file, memory, and sub-block class properties using the class factory. Because the register model is a uvm_object hierarchy, not a uvm_component hierarchy, no parent reference is specified and the full hierarchical name of the block type instance is specified as the context. The build() method shall call the configure() method for all register, register file, memory, and sub-block class properties, specifying this as the block parent and null as the register file parent. The build() method shall call the build() method for all register, register file, and sub-block class properties. The build() method may call the add_hdl_path() method for any register, register file, memory, or sub-block class properties with the appropriate argument values for the register, register file, memory, or sub-block instance.

class my_blk_type extends uvm_reg_block;

virtual function build();

this.BLK1 = my_blk1_type.type_id.create(“BLK1”, null,

get_full_name()); this.BLK1.configure(this, …); this.BLK1.build(); endfunction endclass

After a register or memory has been created, the build() method shall call the appropriate uvm_reg_map::add_*() method for all address maps where the register, register file, or memory is accessible, specifying its offset in that address map.

class my_blk_type extends uvm_reg_block;

virtual function build();

this.R1 = my_reg1_type.type_id.create(“R1”, null, get_full_name()); this.R1.configure(this,…); this.R1.build(); this.default_map.add_reg(this.R1, ’h04, …); endfunction endclass

After a register file has been built, the build() method shall call its map() method for all address maps where the register file is accessible, specifying its offset in that address map.

class my_rf_type extends uvm_reg_regfile;

virtual function build();

this.RF1 = my_rf1_type.type_id.create(“RF1”, null, get_full_name()); this.RF1.build(); this.RF1.map(this.default_map, ’h200, …); endfunction endclass

After a sub-block has been built, for each address map in that sub-block, the build() method shall call the appropriate uvm_reg_map::add_submap() method for all address maps where the sub-block address map is accessible, specifying its offset in that upper address map.

class my_blk_type extends uvm_reg_block;

virtual function build();

this.BLK1.build(); this.default_map.add_submap(this.BLK1.default_map, ’h8000); endfunction endclass

5.5.5.4 Coverage Model¶

A block-level coverage model is defined and instantiated in the block type class. It measures the coverage of the accessed offsets and field values on each instance of that block type. The uvm_reg_block::sample() or uvm_reg_block::sample_values() methods shall be used to trigger the sampling of a coverage point, based on the data provided as an argument or gathered from the current state of the block type instance. The sampling of the coverage model shall only occur if sampling for the corresponding coverage model has been turned on, as reported by the uvm_reg_block::get_coverage() method.

class my_blk extends uvm_reg_block;

covergroup cg_vals;

… endgroup

… virtual function void sample_values();

super.sample_values(); if (get_coverage(UVM_CVR_FIELD_VALS))

cg_vals.sample(); endfunction endclass : my_blk

All the coverage models that may be included in the block type shall be reported to the uvm_reg_mem base class using uvm_reg_block::build_coverage() when super.new() is called or using the uvm_reg_block::add_coverage() method. If no functional coverage models are included in the generated block type, UVM_NO_COVERAGE shall be specified. Block-level coverage groups shall only be instantiated in the constructor if the construction of the corresponding coverage model is enabled, as reported by the uvm_reg_block::has_coverage() method.

class my_blk extends uvm_reg_block;

covergroup cg_vals;

… endgroup … function new(string name = “my_blk”);

super.new(name, build_coverage(UVM_CVR_FIELD_VALS)); if (has_coverage(UVM_CVR_FIELD_VALS))

cg_vals = new(); endfunction: new … endclass : my_blk

The content, structure, and options of the coverage group is defined by the register model generator and is outside the scope of UVM.

5.5.6 Packaging a Register Model¶

The generator is free to structure the generated code into packages and files to facilitate compilation or reuse.

The following practices are recommended, but not required:

a) Block types, and all the register, register file, and memory types they require, should be located in

separate packages. b) Register, register file, and memory types shared by more than one block type should be located in

separate packages. c) A header file, with all the required import statements to use the register model, should be generated. d) A lengthy build() method may be split into several, shorter sub-methods. The sub-methods shall

be declared local and called by the build() method.

5.5.7 Maximum Data Size¶

By default, the maximum size of fields, registers, and memories is 64 bits. This limitation is implemented via the definition of the uvm_reg_data_t type.

typedef bit [63:0] uvm_reg_data_t;

The uvm_reg_data_t type is used in all methods and API that deal with data values to and from the register model. Smaller fields, registers, and memories are intrinsically supported by using the SystemVerilog automatic value extension and truncation.

The maximum data size may be reduced to save memory in large register models. It may also be increased to support larger fields, registers, or memory values. The size of data values may be specified at compile-time by defining the ’UVM_REG_DATA_WIDTH macro.

% … +define+UVM_REG_DATA_WIDTH=256 …

It is recommended register model generator provide a warning message if the maximum data size need to be increased. It is also recommended the register model contain a static initializer check for the required minimum data size and issue a fatal error message when that is not set appropriately:

class my_blk extends uvm_reg_block;

local static bit m_req_data_width = check_data_width(256); … endclass

5.6 Back-door Access ¶

Back-door access to registers and memory locations is an important tool for efficiently verifying their correct operation.

A back-door access can uncover bugs that may be hidden because write and read cycles are performed using the same access path. For example, if the wrong memory is accessed or the data bits are reversed, whatever bug is introduced on the way in (during the write cycle) will be undone on the way out (during the read cycle).

A back-door improves the efficiency of verifying registers and memories since it can access registers and memory locations with little or no simulation time. Later, once the proper operation of the physical interface has been demonstrated, you can use back-door access to completely eliminate the simulation time required to configure the DUT (which can sometimes be a lengthy process).

A back-door access operates by directly accessing the simulation constructs that implement the register or memory model through a hierarchical path within the design hierarchy. The main challenges of implementing a back-door access are the identification and maintenance of that hierarchical path and the nature of the simulation constructs used to implement the register or memory model.

5.6.1 Back-door read/write vs. peek/poke¶

You can perform back-door access to registers and memory by calling the following read/write methods with their path argument as UVM_BACKDOOR:

a) uvm_reg_field::read() or uvm_reg_field::write() b) uvm_reg::read() or uvm_reg::write() c) uvm_mem::read() or uvm_mem::write()

… or by calling the following peek/poke methods:

d) uvm_reg_field::peek() or uvm_reg_field::poke() e) uvm_reg::peek() or uvm_reg::poke() f) uvm_mem::peek() or uvm_mem::poke()

The peek() methods return the raw value read using the back-door without modifying the content of the register or memory. Should the register content be modified upon a normal read operation, such as a clear-on-read field, it will not be modified. Therefore, reading using peek() methods may yield different results than reading through read() methods.

The poke() methods deposit the specified value directly in the register or memory. Should the register contain non-writable bits or bits that do not reflect the exact value written, such as read-only or write- 1-to-clear fields, they will contain a different value than if the same value had been written through normal means. All field values, regardless of their access mode, will be forced to the poked value. Therefore, writing using poke() methods may yield different results than writing through the front-door.

When using the read() methods with a back-door access path, the behavior of the register or memory access mimics the same access performed using a front-door access. For example, reading a register containing a clear-on-read field will cause the field value to be cleared by poking 0’s into it.

When using the write() method with a back-door access path, the behavior of the register or memory access mimics the same access performed using a front-door access. For example, writing to a read-only field using back-door access will cause the field value to be maintained by first peeking its current value then poking it back in instead of the specified value.

5.6.2 Hierarchical HDL Paths¶

To access a register or memory directly into the design, it is necessary to know how to get at it. The UVM register library can specify arbitrary hierarchical path components for blocks, register files, registers and memories that, when strung together, provide a unique hierarchical reference to a register or memory. For example, a register with a hierarchical path component defined as X, inside a block with a hierarchical path component defined as Y, inside a block with a hierarchical path component defined as Z has a full hierarchical path defined as Z.Y.X.

HDL path components are specific to the language used to model the DUT and the structure of the DUT model. They may be individual hierarchical scope names (e.g., decoder), partial dot-separated (.) hierarchical paths (e.g., bus_if.decoder) or empty (i.e., they do not contribute to the overall path). They can also be build-time string expressions, but must be string constants at run-time, where the value of the string must be a constant name or partial path: it cannot be an expression. Each path component must be empty or a valid path component: they cannot start or end with a dot separator (.). They need not be valid SystemVerilog path components, as they may be used to refer to hierarchical paths that cross language boundaries. HDL paths terminate at registers and memories.

Multiple HDL paths may be defined for the same block, register file, register or memory abstraction. This indicates the block, register file, register, or memory is duplicated in the model of the DUT. The value of a duplicated register or memory must be kept coherent across all copies.

For example, assuming the following register model hierarchy and HDL path components:

Block b1 “b1”

Block b2 “b2_a”, “b2_b”

Register r1 “r1” Register r2 {“r2_1”, “r2_0”} Block b3 “”

Register r3 “r3.z”, {“r3_1”, “r3_0”}

The full hierarchical paths would be as follows:

Block b1 “b1”

Block b2 “b1.b2_a”, “b1.b2_b”

Register r1 “b1.b2_a.r1”, “b1.b2_b.r1” Register r2 {“b1.b2_a.r2_1”, “b1.b2_a.r2_0”}, {“b1.b2_b.r2_1”, “b1.b2_b.r2_0”} Block b3 n/a

Register r3 “b1.r3.z”, {” b1.r3_1”, “b1.r3_0”}

HDL path components are specified using the following methods:

a) uvm_reg_block::configure() and uvm_reg_block::add_hdl_path() b) uvm_reg_file::configure() and uvm_reg_file::add_hdl_path() c) uvm_reg::configure() and uvm_reg::add_hdl_path_slice() d) uvm_mem::configure() and uvm_mem::add_hdl_path_slice()

The HDL path for a register or memory may be a concatenation of simple names. This is used when the register is implemented or modeled as a concatenation of individual fields variables, or when the memory is implemented or modeled using vertical banking. When specifying a concatenation, the bits may be left unspecified if they are not physically implemented.

For example, the register with the implementation illustrated in Figure 26 has its HDL path component specified using the following concatenation:

rg.add_hdl_path_slice(“RDY”, 15, 1); rg.add_hdl_path_slice(“ID”, 6, 6); rg.add_hdl_path_slice(“COUNT”, 0, 4);

15 11 6 3 0 RDY ID

HDL paths are created by concatenating each path component from the root block to the leaf register or memory. However, if a block HDL path component is defined has a root HDL path component, the HDL path component of any blocks above it are ignored.

5.6.3 VPI-based Back-door Access¶

The UVM register library provides a default back-door access mechanism that uses the HDL path(s) returned by the uvm_reg::get_full_hdl_path() and uvm_reg::get_full_hdl_path() methods for the default design abstraction. Using standard SystemVerilog VPI routines, it samples or deposits values in the HDL constructs referenced by the resulting hierarchical HDL paths. If the HDL paths are valid hierarchical SystemVerilog variables, including indexing and slicing operators, this should work without any further requirements.

5.6.3.1 Including the DPI C Library¶

The implementation of the default back-door access mechanism requires the inclusion of some DPI C code. Please refer to the supplementary UVM documentation provided by your simulator vendor on how to compile and link the UVM C library.

5.6.3.2 Performance Issues with the VPI-based Back-door Access¶

Enabling the VPI functionality required by the default back-door access mechanism may disable performance optimizations normally done in your simulator. Please refer to the supplementary UVM documentation provided by your simulator vendor for additional or alternative steps that may be taken to improve the performance of your simulation.

5.6.4 User-defined Back-door Access¶

Should the DPI-based back-door access prove to be insufficient, a user-defined back-door access can be used instead. A user-defined back-door access is able to use any SystemVerilog constructs or tool-specific utility to access registers and memories. For example, if a memory or register is located in an encrypted model, a user-defined back-door may be used to peek and poke values directly into the encrypted model via a suitable API.

A user-defined register back-door is provided through an extension of the uvm_reg_backdoor class. A back-door write operation is implemented in the uvm_reg_backdoor::write() virtual method whereas a back-door read operation is implemented in the uvm_reg_backdoor::read() virtual method. This back-door access is then associated with a specific register through the uvm_reg::set_backdoor() method or with all the registers within a block using the uvm_reg_block::set_backdoor() method.

A user-defined memory back-door is provided through an extension of the uvm_mem_backdoor class. A back-door write operation is implemented in the uvm_mem_backdoor::write() virtual method whereas a back-door read operation is implemented in the uvm_mem_backdoor::read() virtual method. This back-door access is then associated with a specific memory through the uvm_mem::set_backdoor() method.

User-defined back-door access mechanisms may be defined by the register model generators. They are instantiated and associated with their corresponding block, register, or memory abstraction class in the implementation of their respective build() method. User-defined back-door access mechanisms may also be registered with their corresponding block, register, or memory after the register model construction, overriding any previously defined (or default) back-door access mechanisms. In the latter case, they are instantiated and associated with their corresponding block, register, or memory abstraction class in the implementation of the environment’s build() method.

function void

tb_env::build_phase(uvm_phase phase); super.build_phase(phase); … begin:sub:`my_mem_backdoor bkdr = new; `

regmodel.mem.set_backdoor(bkdr); end endfunction: build_phase

5.6.5 Back-door Access for Protected Memories¶

The content of memories may be protected using one or more protection schemes. They can vary from simple additional bits providing an Error Correction Code to full encryption of its content.

When performing back-door write operations, it is necessary to correctly protect the memory content to avoid errors when a physical interface subsequently reads these memory locations. It may also be useful or

necessary to have direct access to the protected form because these bits are created and used entirely within the design, and can only be accessed through back-door access. The back-door is the only way protected values can be checked and protection errors injected.

The encode() and decode() callback methods located in the uvm_mem_backdoor_cbs class are designed to handle such data protection. The encode() method is applied on the way in and the decode() method is applied on the way out. But, unlike regular callback methods, the decoding is done in the reverse order of registration. This allows multiple layers of data protections to be implemented in the same memory, each modeled using a single callback extension. The order of registration determines the order in which the various layers of protections are applied–then undone.

For example, ECC bits are located in additional memory bits within the same memory location as the data they protect; they must be generated and set for write accesses, and must checked and masked when read.

class ecc_protected extends uvm_mem_backdoor_cbs;

virtual function uvm_reg_data_t encode(uvm_reg_data_t data);

// Append the ECC bits to the data to write data[35:32] = ecc::compute(data[31:0]); return data; endfunction

virtual function uvm_reg_data_t decode(uvm_reg_data_t data);

// Check and mask the ECC bits to the data to write if (data[35:32] != ecc::compute(data[31:0])) ’uvm_error(…) return data[31:0]; endfunction

endclass

Similarly, data written to an encrypted memory must be ciphered during write accesses and deciphered when read.

class encrypted_mem extends uvm_mem_backdoor_cbs;

virtual function uvm_reg_data_t encode(uvm_reg_data_t data);

return crypt::encrypt(data); endfunction

virtual function uvm_reg_data_t decode(uvm_reg_data_t data);

return crypt::decrypt(data); endfunction

endclass

5.6.6 Active Monitoring¶

The mirrored field values in a register model are updated when the fields are accessed through the register model based on the current mirrored value, the accessed data value, and the access policy of the field. They may also be updated based on observed read and write transactions on the bus interface if the register model is integrated with the bus monitor and explicit monitoring is enabled (see Section 5.9.3). Any changes to the field value performed by the design itself cannot be detected and then mirrored in the register model.

The back-door mechanism can be used to automatically update the mirror value of fields that are modified by the design itself by observing the SystemVerilog constructs which are used to store the field values. When a change of value is detected, the mirrored value can be similarly updated.

Because there is no standard value-change callback VPI or PLI functionality, the automatic update of a field can only be implemented using a user-defined back-door. The active monitoring of a register requires the implementation of the uvm_reg_backdoor::is_auto_updated() and uvm_reg_backdoor::wait_for_change() methods.

uvm_reg_backdoor::is_auto_updated() returns TRUE if the specified named field is actively monitored. All actively-monitored fields have their mirror value updated strictly and only through the active mirroring mechanism. Executed or observed transactions are not used to update their mirrored value.

The uvm_reg_backdoor::wait_for_change() task must return only when a change in any of the actively-monitored fields is observed. For each actively-monitored register, a thread calls this task to wait for any change in any of the fields in the register. As soon as it returns, their values are sampled and their mirror values updated. The implementation of that method should not simply wait for the active edge of the clock signal used to update the field values in the design; for optimal performance, the implementation of that method should only return when an actual change occurs.

class active_monitor_r1 extends uvm_reg_backdoor;

virtual function bit is_auto_updated(string fld_name);

case (fld_name) “f1”: return 1; “f2”: return 1; endcase endfunction

virtual task wait_for_change();

@($root.tb_top.dut.rf.f1 or $root.tb_top.dut.rf.f2); endtask endclass

The active-monitoring thread must be started for each actively-monitored register by invoking the uvm_reg_backdoor::start_update_thread() method of its back-door access class once an instance of that back-door access class is created, as shown in the following example:

class add_active_monitors extends my_blk;

virtual function build();

super.build(); begin

active_monitor_r1 am_r1 = new; r1.set_backdoor(am_r1); am_r1.start_update_thread(r1); end endfunction endclass

5.7 Special Registers ¶

The UVM register library presumes all registers and memories are average registers and memories, they are accessible at a known, constant, unique physical address(es), their behavior is constant throughout the simulation regardless of the physical interface used to access them, and they contain a single value.

Designer creativity, the demands of the application, or implementation constraints often require special behaviors be implemented. Special register behavior can be modeled using any number of extension capabilities provided in the UVM register and field abstraction classes. Pre- and post-read/write callback objects, virtual callback methods, user-defined front-doors, and user-defined back-doors may be used to extend the behavior of the base library. And, if all else fails, it is always possible to override virtual methods that are used to access the register content, i.e., read(), write(), peek(), and poke().

5.7.1 Pre-defined Special Registers¶

The UVM library pre-defines some commonly used special registers. A register model generator is free to provide a library of additional special register models and use them in its generated model.

5.7.1.1 Indirect Indexed Registers¶

Some registers are not directly accessible via a dedicated address. Indirect access of an array of such registers is accomplished by first writing an —index” register with a value that specifies the array’s offset, followed by a read or write of a —data” register to obtain or set the value for the register at that specified offset. The pre-defined uvm_reg_indirect_data class models the behavior the —data” register.

A —data” register type is defined by extending the uvm_reg_indirect_data register class. The —data” register must not contain any fields. The —index” and indirect register array must be built first, as the —index” registers and the register array are specified when the —data” register is configured using the uvm_reg_indirect_data::configure() method. The indirect register array, —index“, and —data” registers are added as members of the containing block. However, only the —index” and —data” registers are added to a map in the containing block. The registers in the indirect register array must be not added to the address map in the containing block because they have no dedicated address.

class my_blk_type extends uvm_reg_block;

ind_idx_reg IND_IDX; ind_data_reg IND_DATA; ind_reg INDIRECT_REG[256];

virtual function build();

foreach (INDIRECT_REG[i]) begin

string name = $sformatf(“INDIRECT_REG[%0d]”,i); INDIRECT_REG[i]=

ind_reg.type_id.create(name,,get_full_name()); INDIRECT_REG[i].configure(this, null, …); INDIRECT_REG[i].build(); end

IND_IDX = ind_idx_reg.type_id.create(“IND_IDX”,,get_full_name()); IND_IDX.configure(this, null, …); IND_IDX.build();

IND_DATA = ind_data_reg.type_id.create(“IND_DATA”,,get_full_name()); IND_DATA.configure(IND_IDX, INDIRECT_REG, this, null); IND_DATA.build();

default_map = create_map(“”, 0, 4, UVM_BIG_ENDIAN); default_map.add_reg(IND_IDX, 0); default_map.add_reg(IND_DATA, 4); endfunction endclass

The registers in the indirect register array cannot be accessed via a back-door access to the —data” register. Back-door access to the register array is provided by performing back-door accesses via the unmapped, indirect register itself.

If a different indirection mechanism is required, a user-defined register extension will be necessary.

5.7.1.2 FIFO (first-in, first-out) Registers¶

A FIFO register is not a register in the usual sense. It is a FIFO whose push and pop operations are mapped to write and read operations at a specific address. Writing to that address causes the data written to be pushed at the end of the FIFO. Reading from that address returns the data that is currently at the head of the FIFO and pops it. Whether the FIFO is full or empty is usually specified via status bits in another register.

To model a FIFO register, the register type shall be extended from the uvm_reg_fifo class. The maximum number of entries in the FIFO and the size of each entry is specified when calling super.new().

class fifo_reg extends uvm_reg_fifo;

function new(string name = “fifo_reg”);

super.new(name,8,32,UVM_NO_COVERAGE); endfunction: new

`uvm_object_utils(fifo_reg)

endclass

Backdoor access to a FIFO register is not allowed.

5.7.2 Unmapped Registers and Memories¶

By default, the entire register or memory is assumed to be linearly mapped into the address space of the block that instantiates it. Each register or location in a memory thus corresponds to a unique address in the block. However, you can use different addressing mechanisms. For example, you could access a large memory in a limited address space using an indexing mechanism: the desired offset within the memory is written into a register, then the data at that memory offset is read or written by reading or writing another register. This memory is effectively unmapped: it does not appear in the linear address space used to access it. See Section 5.7.1.1.

The number of possible access mechanisms is potentially infinite and only limited by the imagination, requirements, and constraints of designers. To support arbitrary access mechanisms, it is possible to replace the default linearly mapped access mechanism with any user-defined access mechanism.

5.7.2.1 User-defined Front-door Access¶

User-defined front-door access is made possible by extending the uvm_reg_frontdoor class and registering an instance of the class with specific registers or memories using the uvm_reg::set_frontdoor() or uvm_mem::set_frontdoor() method. The uvm_reg_frontdoor is a UVMSequence. For each write or read operation, the register model creates a uvm_reg_item object representing the operation, assigns it to the rw_info property of registered front-door sequence, and calls its start method. Ultimately, the front-door’s body task is called, which must be implement to perform the actual operation.

class indexed_reg_frontdoor extends uvm_reg_frontdoor;

local uvm_reg m_idx_reg; local uvm_reg m_data_reg; local bit [7:0] m_addr;

function new(string name=“indexed_reg_frontdoor_inst”);

super.new(name); endfunction

function void configure(uvm_reg idx, uvm_reg data, bit [7:0] addr);

m_idx = idx;; m_data = data; m_addr = addr; endfunction: new

virtual task body(uvm_reg_item rw);

m_idx_reg.write(status, m_addr, …); if (status != UVM_IS_OK)

return; if (rw.kind == UVM_WRITE)

m_data.write(rw.status, data, …); else_{m_data.read(rw.status, data, …);} endtask

endclass

User-defined front-doors are instantiated and associated with their corresponding register or memory abstraction class in the build() method of the block or register file that instantiates them or the build() phase callback of the environment component where the register model is instantiated and built.

virtual function void build();

foreach TABLE[i] begin

indexed_reg_frontdoor idx_frtdr = new(INDEX, DATA, i);

= idx_reg_frontdoor.type_id.create(“idx_frtdr”,,get_full_name()); idx_frntdoor.configure(idx_reg, data_reg, i); regmodel.TABLE[i].set_frontdoor(idx_frontdoor, default_map, …); end endfunction: build

A user-defined front-door is registered on a per-map basis, affecting the access of a register or memory through a specific physical interface. A different front-door mechanism (or the built-in one) can be used for other physical interfaces. For example, a memory could use the indexed addressing scheme described above for one physical interface but be mapped normally within the address map of another physical interface.

5.7.2.2 Mirroring Unmapped Registers¶

When using explicit or passive monitoring to update the mirror value in unmapped registers, it will be necessary to override the uvm_reg::predict() method of the register(s) used to access the unmapped registers, since the observed transactions will be using the address of those access registers, not the unmapped (unaddressable) registers that are ultimately accessed.

In the case of an indirect register, the uvm_reg_indirect_data class extends predict for you and serves as an example of how you do this for your custom unmapped registers.

function bit uvm_reg_indirect_data::predict (uvm_reg_data_t value, …);

if (m_idx.get() >= m_tbl.size()) begin

`uvm_error(“Index reg > than size of indirect register array”) return 0; end return m_tbl[m_idx.get()].predict(value, …); endfunction

5.7.3 Aliased Registers¶

Aliased registers are registers that are accessible from multiple addresses in the same address map. They are different from shared registers as the latter are accessible from multiple address maps. Typically, the fields in aliased registers will have different behavior depending on the address used to access them. For example, the fields in a register may be readable and writable when accessed using one address, but read-only when accessed from another.

Modelling aliased registers in UVM involves more than simply mapping the same register at two different addresses. A UVM register model requires each instance of a uvm_reg class be mapped to a unique address in an address map. For aliased registers, this requires a register class instance for each address. All this enables using a specific register instance to access the aliased register via a specific address.

For example, the (incomplete) register model shown below models a register aliased at two addresses: ’h0100 and ’h0200. Each alias is known under a different instance name, Ra and Rb respectively. To access the aliased register via address ’h0100, the Ra instance would be used.

class my_blk extends uvm_reg_block;

rand my_reg_Ra Ra; rand my_reg_Rb Rb; virtual function build();

… default_map.add_reg(Ra, ’h0100); default_map.add_reg(Rb, ’h0200); endfunction endclass

Each register instance must be of a register type that models the behavior of the register and field it contains of its corresponding alias. For example, a register that contains a field that is RW when accessed via one address, but RO when accessed via another would require two register types: one with a RW field and another one with a RW field, and both using the same field names.

class my_reg_Ra extends uvm_reg;

rand uvm_reg_field F1; … virtual function void build();

F1 = uvm_reg_field.type_id.create(“F1”); F1.configure(this, 8, 0, “RW”, 0, 8’h0, 1, 1, 1); endfunction … endclass

class my_reg_Rb extends uvm_reg;

uvm_reg_field F1; … virtual function void build();

F1 = uvm_reg_field.type_id.create(“F1”); F1.configure(this, 8, 0, “RO”, 0, 8’h0, 1, 0, 1); endfunction

… endclass

The aliasing functionality must be provided in a third class that links the two register type instances. The aliasing class can make use of the pre-defined register and field callback methods to implement the aliasing functionality. It may also make use of additional APIs or functionality created by the register model generator in the different register types that model each alias of the register. The aliasing class should be based on uvm_object to be factory-enabled. The required reference to the various register instance aliases shall be supplied via a configure() method.

class write_also_to_F extends uvm_reg_cbs;

local uvm_reg_field m_toF;

function new(uvm_reg_field toF);

m_toF = toF; endfunction

virtual function void post_predict(uvm_reg_field fld,

uvm_reg_data_t value, uvm_predict_e kind, uvm_path_e path, uvm_reg_map map); if (kind != UVM_PREDICT_WRITE) return;

void’(m_toF.predict(value, -1, UVM_PREDICT_WRITE, path, map)); endfunction

endclass

class alias_RaRb extends uvm_object;

protected reg_Ra m_Ra; protected reg_Rb m_Rb;

`uvm_object_utils(alias_RaRb)

function new(string name = “alias_RaRb”);

super.new(name); endfunction: new

function void configure(reg_Ra Ra, reg_Rb Rb);

write_also_to_F F2F;

m_Ra = Ra; m_Rb = Rb;

F2F = new(Rb.F1); uvm_reg_field_cb::add(Ra.F1, F2F); endfunction : configure endclass : alias_RaRb

The register file or block containing the various register aliases shall also instantiate the aliasing class in its build() method and call the configure() method with appropriate arguments.

class my_blk extends uvm_reg_block;

rand my_reg_Ra Ra; rand my_reg_Rb Rb; …

virtual function build();

default_map = create_map(“”, 0, 4, UVM_BIG_ENDIAN);

Ra = reg_Ra.type_id.create(“Ra”,,get_full_name()); Ra.configure(this, null); Ra.build();

Rb = reg_Rb.type_id.create(“Rb”,,get_full_name()); Rb.configure(this, null); Rb.build();

default_map.add_reg(Ra, ’h0100); default_map.add_reg(Rb, ’h0200);

begin

alias_RaRb RaRb;

RaRb = alias_RaRb.type_id.create(“RaRb”,,get_full_name()); RaRb.configure(Ra, Rb); end endfunction endclass

There are no pre-defined aliasing classes because the nature of the aliasing is highly variable, not just in how the fields provide different behaviors through the various aliases, but potentially in their layout as well.

5.7.4 Unimplemented Registers¶

A UVM register model can model registers that are specified, but have not yet been implemented in the DUV. This allows the verification environment and testcases to make use of these registers before they are available.

Because these registers are unimplemented, there is nothing to actually read or write inside the DUT. Since the mirror in a register abstraction class provides a faithful model of the expected behavior of that register, it can be used to provide a read back value. A yet-to-be-implemented register is thus modeled by writing to and reading from the mirror.

An unimplemented register can be modeled by providing a user-defined front- and back-door that access the mirrored value instead of performing bus transactions.

class not_yet_implemented_fd extends uvm_reg_frontdoor;

… virtual task body();

uvm_reg R;

$cast(R, rw_info.element); if (rw_info.kind == UVM_READ) rw_info.value[0] = R.get();

R.predict(rw_info.value[0], -1,

(rw_info.kind == UVM_READ) ?

UVM_PREDICT_READ : UVM_PREDICT_WRITE, rw_info.path, rw_info.map); endtask endclass

The user-defined front- and back-door can be registered in the environment where the register model is instantiated. As Register model generators will not have access to environment, they can register the back- door and front-door in the build() method of the containing block for the unimplemented register. Additionally, it is recommended to log info/warning messages indicating that the particular register is unimplemented.

virtual function void connect_phase(uvm_phase phase);

… if (get_parent() == null) begin

not_yet_implemented_fd fd; ’uvm_warning(“NotYetImpl”,

“Working around yet-to-be-implemented registers”); fd = new; regmodel.R2.set_frontdoor(fd); end endfunction

5.8 Integrating a Register Model in a Verification Environment ¶

Test sequences, whether pre-defined or user-defined ones, need a verification environment in which to execute. The register model needs to be an integral part of that verification environment to be used by the tests to access registers and memories in the DUT.

An environment must have a reference to the register model that corresponds to the DUT it verifies. It is recommended a class property named regmodel be used for that purpose. To enable vertical reuse of the environment, it must first check if its register model has been defined by a higher-level environment. If not, it must be allocated using the class factory, explicitly built by calling its build() method, then it calls the uvm_reg_block::lock_model() method. After creating any sub-block environments, their register models must then be specified by setting their respective regmodel class properties. All of this must be implemented in the environment’s build phase method.

class block_env(UVMEnv):

block_reg_model regmodel; subblk_env subblk;

virtual function void build_phase(uvm_phase phase);

if (regmodel == null) begin

regmodel = block_reg_model.type_id.create(“regmodel”, this); regmodel.build(); regmodel.lock_model(); end subblk = subblk_env.type_id.create(“subblk”, this); subblk.regmodel = regmodel.subblk; endfunction endclass

If HDL paths are used, the root HDL paths must be specified in the environment that instantiates the register model. The value of that root path will depend on the location of the model for the DUT within the complete simulation model.

class block_env(UVMEnv):

block_reg_model regmodel; virtual function void build_phase(uvm_phase phase);

if (regmodel == null) begin

regmodel = block_reg_model.type_id.create(“regmodel”, this); regmodel.build(); regmodel.set_hdl_path_root(“tb_top.dut”); end endfunction endclass

5.9 Integrating a Register Model ¶

A register model must be integrated with the bus agents that perform and monitor the actual read and write operations. The terms —bus driver“, —bus agent”, —bus interface” and —bus operations” are used to describe the components, protocol, and interface associated with the execution of read and write operations on the DUT. The integration may be established via a non-bus-based interface and protocol.

The integration with the bus agent must only be done on root blocks. Root blocks model the entire DUT and they are the only ones who have access to and knowledge of the externally-visible address maps. Lower- level register models will translate their read and write operations in terms of read and write operations at the root block level, using root-level addresses and bus protocols.

To that end, the integration process must be conditional to the register model being a root register model. This is accomplished by checking if the register model has a parent. If not, it is a root model and integration with the bus agent may proceed. All this must be implemented in the environment’s connect phase.

class block_env(UVMEnv):

block_reg_model regmodel; subblk_env subblk;

virtual function void connect_phase(uvm_phase phase);

if (regmodel.get_parent() == null) begin // Integrate register model with bus agent … end endfunction endclass

There are three structural bus agent integration approaches for keeping the register model’s mirror values in sync with the DUT: implicit prediction, explicit prediction, and passive.

Implicit prediction only requires the integration of the register model with one or more bus sequencers. Updates to the mirror are predicted automatically (i.e., implicitly) by the register model after the completion of each read, write, peek, or poke operation. This integration is the simplest and quickest, but it will fail to observe bus operations that did not originate from the register model (e.g., by a third-party bus agent) and thus fail to appropriately update the corresponding mirror values.

Explicit prediction requires the register model be integrated with both the bus sequencers and corresponding bus monitors. In this mode, implicit prediction is turned off and all updates to the mirror are predicted externally (i.e., explicitly) by a uvm_reg_predictor component, one for each bus interface. The predictor receives the bus operations observed by a connected bus monitor, determines the register being accessed by performing a reverse-lookup using the observed address, and then calls the found register’s predict method explicitly to update the mirror. This integration requires more work, but it will observe all bus operations, whether they originated from the register model or a third-party bus agent, and thus appropriately update the corresponding mirror values.

Passive integration only requires the integration of the register model with the bus monitor as described above. All the monitoring of the register operations is performed externally to (i.e., explicitly) the register model. All bus operations, whether they originated from the register model or a third-party bus agent, are observed and thus appropriately reflected in the corresponding mirror values. Because the register model is not integrated with a bus sequencer, it cannot be used to read and write register and memories in the DUT, only to track and verify their current value.

5.9.1 Transaction Adapter¶

The first step in integrating a register model with a bus agent are the conversion functions between a generic read/write bus operation descriptor, uvm_reg_bus_op, used by the register model and the protocol- specific read/write transaction descriptor used by the bus agent.

The transaction adapter is implemented by extending the uvm_reg_adapter class and implementing the reg2bus() and bus2reg() methods. Being a uvm_object, the bus adapter must implement a suitable uvm_object constructor and use the ’uvm_object_utils() macro to enable it for the class factory.

class reg2apb_adapter extends uvm_reg_adapter;

`uvm_object_utils(reg2apb_adapter)

function new(string name = “reg2apb_adapter”);

super.new(name); endfunction

virtual function UVMSequenceItem reg2bus(const ref uvm_reg_bus_op rw);

apb_rw apb = apb_rw.type_id.create(“apb_rw”); apb.kind = (rw.kind == UVM_READ) ? apb_rw::READ : apb_rw::WRITE; apb.addr = rw.addr; apb.data = rw.data; return apb; endfunction

virtual function void bus2reg(UVMSequenceItem bus_item,

ref uvm_reg_bus_op rw); apb_rw apb; if (!$cast(apb,bus_item)) begin

`uvm_fatal(“NOT_APB_TYPE”,

“Provided bus_item is not of the correct type”) return; end rw.kind = apb.kind ? UVM_READ : UVM_WRITE; rw.addr = apb.addr; rw.data = apb.data; rw.status = UVM_IS_OK; endfunction endclass

If the bus protocol supports byte lane enables (i.e., it is possible to read or write individual bytes in a multi- byte bus), the supports_byte_enable class property should be set to TRUE in the constructor. Similarly, the provides_responses class property should be set to TRUE if the bus driver returns responses, e.g., the result of a read operation, in a separate response descriptor:

class reg2apb_adapter extends uvm_reg_adapter;

function new(string name = “”);

super.new(name); supports_byte_enables = 0; provides_responses = 1; endfunction endclass

Because this transaction adapter is specific to the bus agent, not the register model, it should be provided as part of a UVM-compliant bus UVC.

The transaction adapter is then instantiated in the connect phase of the environments corresponding to root register models:

class block_env(UVMEnv): block_reg_model regmodel; subblk_env subblk;

virtual function void connect_phase(uvm_phase phase);

… if (regmodel.get_parent() == null) begin

reg2apb_adapter reg2apb =

reg2apb_adapter.type_id.create(“reg2apb”,,get_full_name()); …

end endfunction

endclass

5.9.2 Integrating Bus Sequencers¶

All integration approaches require a register model be configured with one or more bus sequencers. The register model becomes a property of a uvm_reg_sequence subtype that executes

a) directly on a bus sequencer, if there is only one bus interface providing access to the DUT registers; b) as a virtual sequence, if there are one or more bus interfaces providing access to the DUT registers; c) as a register sequence running on a generic, bus-independent sequencer, which is layered on top of a

downstream bus sequencer.

Note–To keep the code examples that follow succinct and focused on register model integration, we do not show obtaining handles via configuration or the resources database, or a priori sequence registration to a specific sequencer.

5.9.2.1 Register Sequence Running on the Bus Sequencer¶

The simplest approach is to run register sequences directly on the bus sequencer, competing directly with all other —native” bus sequences concurrently running on the bus sequencer. The register sequence will, via the register model, produce bus sequence stimulus using a preconfigured bus adapter. This approach is suitable for when the registers being accessed by the register sequence are accessible via a single bus interface, as shown in Figure 27.

Implementing this approach is accomplished by registering the bus sequencer and corresponding transaction adapter with the appropriate address map in the register model. The model is registered with the user- defined register sequence and the sequence started on the bus sequencer. As with any other running bus sequence, the register sequence’s is_relevant, pre_do, mid_do, and post_do methods are called during execution of each bus item generated by the model. To gain exclusive access to the bus, the register sequence may also call grab or lock to prevent other bus sequences from running:

class block_env(UVMEnv):

    block_reg_model regmodel;
    apb_agent apb;

    virtual function void connect_phase(uvm_phase phase);
        if (regmodel.get_parent() == null) begin

        reg2apb_adapter reg2apb =

        reg2apb_adapter.type_id.create(“reg2apb”,,get_full_name());
        regmodel.APB.set_sequencer(apb.sequencer, reg2apb);
        regmodel.set_auto_predict(1); end ...
    endfunction ...
endclass

The above example registers an APB bus sequencer and APB-specific bus adapter with the APB address map defined in top-level register model. If the register model defines only a single map, the map may be referenced via the handle default_map.

You define a register sequence by extending uvm_reg_sequence and defining the body() task to use the model property:

class my_reg_sequence(UVMRegSequence):

    # block_reg_model model;

    async def body(self):

        uvm_status_e status;
        uvm_reg_data_t data;

        self.model.A.write(status, 'h33, .parent(this));
        if (status == UVM_NOT_OK)
            ’uvm_error(...) model.A.read(status, data, .parent(this));
        if (data != ’h33)
            uvm_error(...)

uvm_object_utils(my_reg_sequence)

The uvm_reg_sequence class parameterizes its base class. This allows you to splice in any user-defined UVMSequence subtype if needed:

class VIP_sequence extends UVMSequence #(VIP_base_item); class my_reg_sequence extends uvm_reg_sequence (VIP_sequence);

To run the register sequence, assign the sequence model property and start it on the bus sequencer:

class my_test extends UVMTest;

    block_env env;

    virtual function void run_phase(uvm_phase phase);
        my_reg_sequence seq = my_reg_sequence.type_id.create(“seq”,this);
        seq.start(env.apb.master);
    endfunction

endclass

5.9.2.2 Register Sequence Running as a Virtual Sequence¶

When the registers in the DUT become accessible via more than one physical bus interface, the same register sequence may instead be started as a virtual sequence as the sequencer used in each write/read call is not directly referenced. The register model routes the operation to the appropriate sequencer based on which map is in effect.

Consider a register model with two registers accessible via different bus interfaces, as shown in Figure 28. As in the previous example in Section 5.9.2.1, the sequence calls write and read on regA and regB without referring to a map or sequencer.

Note–Write and read calls have an optional map argument, but specifying a map explicitly would limit sequence reuse.

The only difference between this and running directly on a bus sequencer is more than one sequencer/ adapter pair is registered with the register model and the register sequence’s start method is called without specifying a sequencer to run:

class block_env(UVMEnv):

    block_reg_model regmodel; apb_agent apb; wishbone_agent wsh;

    virtual function void connect_phase(uvm_phase phase);

        if (regmodel.get_parent() == null) begin

            reg2apb_adapter reg2apb =

            reg2apb_adapter.type_id.create(“reg2apb”,,get_full_name());
            reg2wsh_adapter reg2wsh =

            reg2wsh_adapter.type_id.create(“reg2wsh”,,get_full_name());
            regmodel.APB.set_sequencer(apb.sequencer, reg2apb);
            regmodel.WSH.set_sequencer(wsh.sequencer, reg2wsh);
            regmodel.set_auto_predict(1);
        end ...
    endfunction ...

endclass

A register model having more than one configured interface offers interesting timing possibilities. For example, if two registers are accessible via different busses, their accesses can be concurrent:

class my_reg_sequence(UVMRegSequence):

    # block_reg_model model;

    async def body(self):
        status = 0
        fork
            model.APB.write(status, 'h33, .parent(this))
            model.WSH.read(status, ’h66, .parent(this))
        join

uvm_object_utils(my_reg_sequence)

To run the register sequence, register the model and start it without specifying a particular sequencer:

class my_test(UVMTest):

    def __init__(self, name="my_test", parent=None):
        super().__init__(name, parent)
        self.env = block_env('block_env', self)

    async def run_phase(self, phase):
        seq = my_reg_sequence.type_id.create(“seq”, self)
        await seq.start(None)

If you needed to grab a particular sequencer for corner-case testing and were not concerned about creating a dependency on a particular sequencer:

self.grab(regmodel.APB.get_sequencer());
...
self.ungrab(regmode.APB.get_sequencer());

5.9.2.3 Register Sequence Running on a Layered Register Sequencer¶

An alternative integration mechanism is to connect the register model with a register sequencer, then layer that register sequencer on top of the bus sequencer, as shown in Figure 29. The register operations will —execute” as abstract sequence items on the register sequencer, allowing central, bus-independent control of the register sequences. However, this also prevents register sequences from competing directly with or having control over concurrently executing bus sequences (i.e., via grab and ungrab), mixing register and bus-specific sequences and sequence item execution within the same sequence, and being notified of bus- specific operations (via pre_do, mid_do, post_do, and is_relevant). This process also only works with a single bus interface, as all register operations are funneled through a single register sequence.

In this scheme, you are effectively moving the built-in register-to-bus item conversion and bus item execution from the register model to an external translation sequence, which can be overridden to perform custom address translation or item conversions. The register model sends abstract uvm_reg_item descriptors to the register sequencer. It is the responsibility of the translation sequence running on the bus sequencer to get these abstract items and convert them to physical bus items and start them on the bus sequencer. The uvm_reg_sequence base class provides this functionality. It parameterizes its base class to enable it to run on bus-specific sequencers.

This is implemented in the connect phase by first registering a uvm_reg_item sequencer and null adapter with the address map corresponding to the bus being targeted. In a single-map model, the default_map is typically used.

You then create an instance of a translation sequence and configure it with the register sequencer handle and bus adapter. The pre-defined layering sequence uvm_reg_sequence, properly parameterized and configured, may be used in this step.

Then, in the run phase, you start the translation sequence on the bus sequencer:

typedef uvm_reg_sequence #(UVMSequence #(apb_rw)) reg2apb_seq_t;

class block_env(UVMEnv):

    block_reg_model regmodel;
    uvm_sequencer#(uvm_reg_item) reg_seqr;
    apb_agent apb;
    reg2apb_seq_t reg2apb_seq;

    virtual function void connect_phase(uvm_phase phase);

        if (regmodel.get_parent() == null) begin

            regmodel.default_map.set_sequencer(reg_seqr,null);
            reg2apb_seq = reg2apb_seq_t.type_id.create(“reg2apb_seq”,,
                self.get_full_name())
            reg2apb_seq.reg_seqr = reg_seqr
            reg2apb_seq.adapter = reg2apb_adapter.type_id.create(“reg2apb”,,
                self.get_full_name())
            regmodel.set_auto_predict(1)
        end
    endfunction

    virtual task run();
        reg2apb_seq.start(apb.sequencer);
    endtask

endclass

To run a register sequence, you register the model and start it on the register sequencer:

class my_test(UVMTest):

    # block_env env;

    async def run_phase(self, phase):
        seq = my_reg_sequence.type_id.create(“seq”,self)
        await seq.start(env.reg_seqr)

5.9.3 Integrating the Register Model with a Bus Monitor¶

By default, the register model updates its mirror copy of the register values implicitly. Every time a register is read or written through the register model, its mirror value is updated. However, if other agents on the bus interface perform read and write transactions outside the context of the register model, the register model must learn of these bus operations to update its mirror accordingly.

Integration of a bus monitor (see Figure30) to make predictions based on observed transactions is independent from how bus sequencers are integrated. All previously described bus sequencer integration approaches may employ explicit, bus monitor-based prediction.

The predictor accepts bus transactions from a connected bus monitor. It uses the preconfigured adapter to obtain the canonical address and data from the bus operation. The map is used to lookup the register object associated with that address. The register’s predict() method is then called with the observed data to update the mirror value. If the register width is wider than the bus, the predictor will collect multiple observed bus operations before calling predict() with the register’s full value. As a final step, a generic uvm_reg_item descriptor representing the abstract register operation is broadcast to subscribers of its analysis port.

Integration is accomplished by first instantiating a uvm_reg_predictor component, parameterized to the bus transaction type, and configuring it with the adapter and address map in the register model that corresponds to the bus being monitored. The uvm_reg_predictor component is then connected to the bus monitor’s analysis port:

class block_env(UVMEnv):

    block_reg_model regmodel;
    uvm_reg_predictor#(apb_rw) apb2reg_predictor;
    apb_agent apb;

    virtual function void build_phase(uvm_phase phase);
        ...
        apb2reg_predictor = new(“apb2reg_predictor”, this);
    endfunction

    virtual function void connect_phase(uvm_phase phase);

        if (regmodel.get_parent() == null) begin
            reg2apb_adapter reg2apb =
            reg2apb_adapter.type_id.create(“reg2apb”,,get_full_name());
            ...
            apb2reg_predictor.map = regmodel.APB;
            apb2reg_predictor.adapter = reg2apb;
            regmodel.APB.set_auto_predict(0);
            apb.monitor.ap.connect(apb2reg_predictor.bus_in);
        end
        ...
    endfunction
    ...

endclass

When explicit prediction is employed, the implicit prediction must be turned off using uvm_reg_map::set_auto_predict(0).

Note–For register models with a single address map, the name of the address map will be default_map.

5.10 Randomizing Field Values

A register model can specify constraints on field values. You can add additional constraints by extending the field, register, register file, or block abstraction class and substituting it in the register model using the factory or by using randomize() with {} when randomizing a field, register, register file, or block. When constraining a field value, the class property to be constrained is named value. This is not the class property that is eventually mirrored or updated and used by the get() and set() methods; it cannot be used for purposes other than random constraints:

ok = regmodel.r1.randomize() with { f1.value <=’hF; };

Once randomized, the selected field values in a register or block may be automatically uploaded to the DUT by using the uvm_reg::update() or uvm_reg_block::update() method. This will upload any randomized value that is different from the current mirrored value to the DUT. If you override the post_randomize() method of a field abstraction class, you must call super.post_randomize() to ensure the randomized value is properly set into the mirror.

You can relax constraints specified in a register model by turning the corresponding constraint block OFF:

regmodel.r1.consistency.constraint_mode(0)

5.11 Pre-defined Sequences

Once a register model has been instantiated in an environment and integrated with the DUT, it is possible to execute any of the predefined register tests sequences to verify the proper operation of the registers and memories in the DUV. It is recommended you start with the simplest test–the hardware reset test–to debug the register model, the environment, the physical transactors, and the DUV to a level where it can be taken through more complicated tests. Some of the predefined test sequences require back-door access be available for registers or memories.

The predefined test sequences in Table 13 are included in the register library. You can combine them in a higher-level virtual sequence to better verify your design. Test sequences are not applied to any block, register, or memory with the NO_REG_TESTS attribute defined. Refer to the UVM 1.2 Class Reference for more details on each pre-defined test sequence.

Table 13–Pre-defined Test Sequences

Sequence Name Description Attributes

uvm_reg_hw_reset_seq Reads all the register in a block and

check their value is the specified reset value.

Skip block or register if any of the following attributes are defined: NO_REG_HW_RESET_TEST :sub:`NO_REG_TESTS `

uvm_reg_single_bit_bash_ seq

Sequentially writes 1’s and 0’s in each bit of the register, checking it is appro- priately set or cleared, based on the field access policy specified for the field containing the target bit.

Skip register if any of the following attributes are defined: NO_REG_BIT_BASH_TEST NO_REG_TESTS

uvm_reg_bit_bash_seq Executes the uvm_reg_single_

bit_bash_seq sequence for all reg- isters in a block and sub-blocks.

Skip block if any of the following attributes are defined: NO_REG_BIT_BASH_TEST NO_REG_TESTS

uvm_reg_single_access_ seq

Requires the back-door be defined for the register. For each address map in which the register is accessible, writes the register then confirms the value was written using the back-door. Sub- sequently writes a value via the back- door and checks the corresponding value can be read through the address map.

Skip register if any of the follow- ing attributes are defined: NO_REG_ACCESS_TEST NO_REG_TESTS

uvm_reg_access_seq Executes the uvm_reg_single_

access_seq sequence for all regis- ters in a block and sub-blocks.

Skip block if any of the following attributes are defined: NO_REG_ACCESS_TEST NO_REG_TESTS

uvm_mem_single_walk_seq Write a walking pattern into the memory then checks it can be read back with the expected value.

Skip memory if any of the follow- ing attributes are defined: NO_MEM_WALK_TEST NO_MEM_TESTS NO_REG_TESTS

uvm_mem_walk_seq Executes the uvm_mem_single_

walk_seq sequence for all memories in a block and sub-blocks.

Skip block if any of the following attributes are defined: NO_MEM_WALK_TEST NO_MEM_TESTS NO_REG_TESTS

uvm_mem_single_access_ seq

Requires the back-door be defined the memory. For each address map in which the memory is accessible, writes the memory locations for each memory then confirms the value was written using the back-door. Subsequently writes a value via the back-door and checks the corresponding value can be read through the address map.

Skip memory if any of the follow- ing attributes are defined: NO_MEM_ACCESS_TEST NO_MEM_TESTS NO_REG_TESTS

Table 13–Pre-defined Test Sequences (Continued)

Sequence Name Description Attributes

uvm_mem_access_seq Executes the uvm_mem_single_

access_seq sequence for all memo- ries in a block and sub-blocks.

Skip block if any of the following attributes are defined: NO_MEM_ACCESS_TEST NO_MEM_TESTS NO_REG_TESTS

uvm_reg_shared_access_ seq

Requires the register be mapped in multiple address maps. For each address map in which the register is accessible, writes the register via one map then confirms the value was writ- ten by reading it from all other address maps.

Skip register if any of the follow- ing attributes are defined: NO_SHARED_ACCESS_TEST NO_REG_TESTS

uvm_mem_shared_access_ seq

Requires the memory be mapped in multiple address maps. For each address map in which the memory is accessible, writes each memory loca- tion via one map then confirms the value was written by reading it from all other address maps.

Skip memory if any of the follow- ing attributes are defined: NO_SHARED_ACCESS_TEST NO_MEM_TESTS NO_REG_TESTS

uvm_reg_mem_shared_ access_seq

Executes the uvm_reg_shared_ access_seq sequence for all regis- ters in a block and sub-blocks. Exe- cutes the uvm_mem_shared_ access_seq sequence for all memo- ries in a block and sub-blocks.

Skip block if any of the following attributes are defined: NO_SHARED_ACCESS_TEST NO_MEM_TESTS NO_REG_TESTS

uvm_reg_mem_built_in_seq Execute all the selected predefined

block-level sequences. By default, all pre-defined block-level sequences are selected.

Applies attributes governing each predefined sequence, as defined above.

uvm_reg_mem_hdl_paths_ seq

Verify the HDL path(s) specified for registers and memories are valid.

Skip register or memory if no HDL path(s) have been specified.

6. Advanced Topics ¶

This chapter discusses UVM topics and capabilities of the UVM Class Library that are beyond the essential material covered in the previous chapters. Consult this chapter as needed.

6.1 The uvm_component Base Class ¶

All the infrastructure components in an UVM verification environment, including environments and tests, are derived either directly or indirectly from the uvm_component class. These components become part of the environment hierarchy that remains in place for the duration of the simulation.¹ Typically, you will derive your classes from the methodology classes, which are themselves extensions of uvm_component. However, understanding the uvm_component is important because many of the facilities that the methodology classes offer are derived from this class. User-defined classes derived from this class inherit built-in automation, although this feature is entirely optional.

The following sections describe some of the capabilities that are provided by the uvm_component base class and how to use them.The key pieces of functionality provided by the uvm_component base class include:

– Phasing and execution control – Configuration methods – Factory convenience methods – Hierarchical reporting control.

6.2 The Built-In Factory and Overrides ¶

6.2.1 About the Factory¶

UVM provides a built-in factory to allow components to create objects without specifying the exact class of the object being creating. The mechanism used is referred to as an override and the override can be by instance or type. This functionality is useful for changing sequence functionality or for changing one version of a component for another. Any components which are to be swapped shall be polymorphically compatible. This includes having all the same TLM interface handles and TLM objects need to be created by the new replacement component. The factory provides this capability with a static allocation function that you can use instead of the built-in new function. The function provided by the factory is:

type_name.type_id.create(string name, uvm_component parent)

Since the create() method is automatically type-specific, it may be used to create components or objects. When creating objects, the second argument, parent, is optional.

A component using the factory to create data objects would execute code like the following:

task mycomponent::run_phase(uvm_phase phase);

mytype data; // Data must be mytype or derivative. data = mytype.type_id.create(“data”); $display(“type of object is: %0s”, data.get_type_name());

… endtask

1_{Contrast to UVMSequence, sequence_item, and transaction, which are transient –they are created, used, and then} garbage collected when dereferenced.

In the code above, the component requests an object from the factory that is of type mytype with an instance name of data.

When the factory creates this object, it will first search for an instance override that matches the full instance name of the object. If no instance-specific override is found, the factory will search for a type-wide override for the type mytype. If no type override is found then the type created will be of type mytype.

6.2.2 Factory Registration¶

You must tell the factory how to generate objects of specific types. In UVM, there are a number of ways to do this allocation.

– The recommended method is to use the `uvm_object_utils(T) or `uvm_compo- nent_utils(T) macro in a derivative uvm_object or uvm_component class declaration, respectively. These macros contain code that will register the given type with the factory. If T is a parameterized type, use `uvm_object_param_utils or `uvm_component_param_utils, e.g.,

`uvm_object_utils(packet) ’uvm_component_param_utils(my_driver) – Use the `uvm_object_registry(T,S) or `uvm_component_registry(T,S) registra- tion macros. These macros can appear anywhere in the declaration space of the class declaration of T and will associate the string S to the object type T. These macros are called by the corresponding uvm_*_utils macros, so you may only use them if you do not use the ’uvm_*_utils macros.

6.2.3 Component Overrides¶

A global factory allows you to substitute a predefined-component type with some other type that is specialized for your needs, without having to derive the container type. The factory can replace a component type within the component hierarchy without changing any other component in the hierarchy. You need to know how to use the factory, but not how the factory works.

NOTE–All type-override code should be executed in a parent prior to building the child(ren). This means that environ- ment overrides should be specified in the test.

Two interfaces exist to replace default components. These interfaces allow overriding by type or instance, and will be examined one at a time.

To override a default component:

a) Define a class that derives from the default component class. b) Execute the override (described in the following sections). c) Build the environment.

6.2.3.1 Type Overrides¶

The first component override replaces all components of the specified type with the new specified type. The prototype is.

<orig_type>.type_id.set_type_override(<override_type>::get_type(),

bit replace = 1);

The set_type_override() method is a static method of the static type_id member of the orig_type class. These elements are defined in the uvm_component_utils and

uvm_object_utils macros. The first argument (override_type::get_type()) is the type that will be returned by the factory in place of the original type. The second argument, replace, determines whether to replace an existing override (replace = 1). If this bit is 0 and an override of the given type does not exist, the override is registered with the factory. If this bit is 0 and an override of the given type does exist, the override is ignored.

If no overrides are specified, the environment will be constructed using the original types. For example, suppose the environment creates an ubus_master_driver type component inside ubus_master_agent.build(). The set_type_override method allows you to override this behavior in order to have an ubus_new_master_driver for all instances of ubus_master_driver.

ubus_master_driver.type_id.set_type_override(ubus_new_master_driver:

get_type);

This overrides the original type (ubus_master_driver) to be the new type (ubus_new_master_driver). In this case, we have overridden the type that is created when the environment should create an ubus_master_driver. The complete hierarchy would now be built as shown in Figure 31. Note that the ubus_master_agent definition does not change, but the factory was used to create the child instances, the type override causes the new type to be created instead.

NOTE–While only one ubus_master_driver instance is replaced in this example, any and all ubus_mas- ter_driver instances would be replaced in an environment containing multiple ubus_master_drivers

6.2.3.2 Instance Overrides¶

The second component override replaces targeted components of the matching instance path with the new specified type. The prototype for uvm_component is:

<orig_type>.type_id.set_inst_override(<override_type>.get_type(),
    string inst_path);

As shown in Section 6.2.3.1, set_inst_override() is a static method of the type_id member of orig_type. The first argument is the type that will be returned in place of the original type when creating the component at the relative path as specified by the second argument. The second argument can be considered the —target” of the override.

Assume the ubus_new_slave_monitor has already been defined. Once the following code is executed, the environment will now create the new type, ubus_new_slave_monitor, for all instances that match the instance path:

ubus_slave_monitor.type_id.set_inst_override(ubus_new_slave_monitor.
    get_type(), “slaves[0].monitor”);

In this case, the type is overridden that is created when the environment should create an ubus_slave_monitor for only the slaves[0].monitor instance that matches the instance path in the override. The complete hierarchy would now be built as shown in Figure 32. For illustration purposes, this hierarchy assumes both overrides have been executed.

NOTE–Instance overrides are used in a first-match order. For each component, the first applicable instance override is used when the environment is constructed. If no instance overrides are found, then the type overrides are searched for any applicable type overrides. The ordering of the instance overrides in your code affects the application of the instance overrides. You should execute more-specific instance overrides first. For example,

mytype.type_id.set_inst_override(newtype::get_type(), “a.b.*”); my_type.type_id.set_inst_override(different_type::get_type(), “a.b.c”);

will create a.b.c with different_type. All other objects under a.b of mytype are created using newtype. If you switch the order of the instance override calls then all of the objects under a.b will get newtype and the instance override a.b.c is ignored.

my_type.type_id.set_inst_override(different_type::get_type(), “a.b.c”); mytype.type_id.set_inst_override(newtype::get_type(), “a.b.*”);

6.3 Callbacks ¶

Callbacks are an optional facility end users can use to augment component behavior. Callbacks may have a noticeable impact on performance, so they should only be used when the desired functionality cannot be achieved in any other way.

6.3.1 Use Model¶

To provide a callback facility to end-users, the component developer needs to:

a) Derive a callback class from the uvm_callback base. It should declare one or more methods that

comprise the —callback interface”. b) Optionally, define a typedef to the uvm_callbacks pool typed to our specific component-call-

back combination. c) Define the component to iteratively invoke the registered implementations of each method defined in the callback class defined in Step (a) at the desired locations within a component main body of code. Whether a function or a task is used is determined by the ability of the component’s code or behavior to be suspended or not during its execution.

To use callbacks, the user needs to:

xserial0 (xserial_env)

tx (xserial_tx_agent)

driver (ubus_new_master_driver)

slaves[0] (ubus_slave_agent)

bus_monitor (xserial_bus_monitor)

monitor (ubus_new_slave_monitor)

driver (xserial_tx_driver)

sequencer (ubus_master_sequencer)

sequencer (xserial_tx_sequencer)

New type created by instance override

bus_monitor (ubus_bus_monitor)

monitor (xserial_tx_monitor)

driver (ubus_slave_driver)

sequencer (ubus_slave_sequencer)

d) Define a new callback class extending from the callback base class provided by the developer, over-

riding one or more of the available callback methods. e) Register one or more instances of the callback with the component(s) you wish to extend.

These steps are illustrated in the following simple example.

6.3.2 Example¶

The example below demonstrates callback usage. The component developer defines a driver component and a driver-specific callback class. The callback class defines the hooks available for users to override. The component using the callbacks (that is, calling the callback methods) also defines corresponding virtual methods for each callback hook. The end-user may then define either a callback or a driver subtype to extend driver’s behavior.

6.3.2.1 Developer Code¶

Define a callback class extending from uvm_callback.

The callback class defines an application-specific interface consisting of one or more function or task prototypes. The signatures of each method have no restrictions. To be able to use the callback invocation macros, functions should return void or bit type. In the example below, a new bus_bfm_cb class extending from uvm_callback is defined. The developer of the bus_bfm component decides to add two hooks for users, trans_received and trans_executed: 1) trans_received–the bus driver calls this after it first receives a new transaction item. It provides a handle to both itself and the new transaction. An additional argument determines whether to drop (1) or execute (0) the transaction. 2) trans_executed–the bus driver calls this after executing the transaction, passing in a han-

dle to itself and the transaction, which may contain read data or other status information.

virtual class bus_driver_cb extends uvm_callback;

virtual function void trans_received(bus_driver driver, bus_tr tr,

ref bit drop); endfunction virtual task trans_executed(bus_driver driver, bus_tr tr);

endtask function new(string name=”bus_driver_cb”);

super.new(name); endfunction endclass b) Define a typedef to the uvm_callbacks pool typed to our specific component-callback com-

bination. UVM callbacks are type-safe, meaning any attempt to register a callback to a component not designed for that callback simply will not compile. In exchange for this type-safety we must endure a bit of parameterized syntax as follows:

typedef uvm_callbacks #(bus_driver, bus_driver_cb) bus_driver_cb_pool; The alias bus_driver_cb_pool can help both the component developer and the end-user pro- duce more readable code. c) Embed the callback feature in the component that will use it.

The developer of the bus_bfm adds the trans_received and trans_executed virtual methods, with empty default implementations and utilizes some macros that implement the most common algorithms

for executing all registered callbacks. With this in place, end-users can now customize component behavior in two ways:

– extend bus_driver and override one or more of the virtual methods trans_received or trans_executed. Then configure the factory to use the new type via a type or instance override. – extend bus_driver_cb and override one or more of the virtual methods trans_received or trans_executed. Then register an instance of the new callback type with all or a specific instance of bus_driver. The latter requires access to the specific instance of the bus_driver.

class bus_driver extends uvm_driver;

function new (string name, uvm_component parent=null);

super.new(name,parent); endfunction `uvm_register_cb(bus_driver, bus_driver_cb) virtual function bit trans_received(bus_tr tr, ref bit drop);

endfunction virtual task trans_executed(bus_tr tr);

endtask virtual task run_phase(uvm_phase phase);

super.run_phase(phase); forever begin bus_tr tr; bit drop = 0; seq_item_port.get_next_item(tr); ’uvm_info(“bus_tr received”,tr.convert2string(), UVM_LOW) trans_received(tr, drop); ’uvm_do_callbacks(bus_driver_cb, bus_driver,

trans_received(this, tr, drop)) if (drop) begin

’uvm_info(“bus_tr dropped”,

“user callback indicated DROPPED”, UVM_HIGH) return; end #100; trans_executed(tt); `uvm_do_callbacks(bus_driver_cb, bus_driver,

trans_executed(this, tr)) ’uvm_info(“bus_tr executed”, tr.convert2string(), UVM_LOW)

seq_item_port.item_done(tr); end // forever endtask endclass

The driver’s put task, which implements the component’s primary functionality, merely calls the virtual methods and callbacks at the appropriate times during execution.

6.3.2.2 End User Code¶

Using the callback feature of a component involves the following steps:

Extend the developer-supplied callback class.

Define a new callback class that extends from the class provided by the component developer, implementing any or all of the methods of the callback interface. In our example, we define both hooks, trans_received and trans_executed. For trans_received, we randomly choose whether to return 0 or 1. When 1, the bus_driver will —drop” the received transaction. For trans_executed, we delay #10 to prevent back-to- back transactions.

class my_bus_bfm_cb extends bus_bfm_cb;

function new(string name=”bus_bfm_cb_inst”);

super.new(name); endfunction `uvm_object_utils(my_bus_bfm_cb) virtual function bit trans_received(bus_bfm driver, bus_tr tr);

’uvm_info_context(“trans_received_cb”,

{” bus_bfm=”,driver.get_full_name(),”

tr=”,tr.convert2string()}, UVM_LOW, driver) return $urandom & 1; endfunction virtual task trans_executed(bus_bfm driver, bus_tr tr);

’uvm_info(“trans_executed_cb”,

{” bus_bfm=”,driver.get_full_name(),”

tr=”,tr.convert2string()}, UVM_LOW, driver) #10; endtask endclass b) Create callback object(s) and register with component you wish to extend.

To keep the example simple and focus on callback usage, we do not show a complete or compliant UVM environment. In the top module, we instantiate the bus_bfm and an instance of our custom callback class. To register the callback object with the driver, we first get a handle to the global callback pool for our specific driver-callback combination. Luckily, the developer provided a convenient typedef in his Step (b) that makes our code a little more readable. Then, we associate (register) the callback object with a driver using the callback pool’s add_cb method. After calling display_cbs to show the registration was successful, we push several transactions into the driver. The output shows that the methods in our custom callback implementa- tion are called for each transaction the driver receives.

module top;

bus_tr tr = new; bus_bfm driver = new(“driver”); my_bus_bfm_cb cb = new(“cb”);

initial begin

bd_cb::add(driver,cb); cbs.display_cbs(); for (int i=1; i<=5; i++) begin

tr.addr = i; tr.data = 6-i; driver.in.put(tr); end end endmodule c) Instance-specific callback registrations can only be performed after the component instance exists. Therefore, those are typically done in the build() and end_of_elaboration() for exten- sions that need to apply for the entire duration of the test and in the run() method for extensions that need to apply for a specific portion of the testcase.

class error_test extends UVMTest;

function new(string name = “error_test”, uvm_component parent = null);

super.new(name, parent); endfunction

virtual task run_phase(uvm_phase phase);

cbs = new;

#1000; bd_cb::add_by_name(“top.bfm”, cbs, this); #100; bd_cb::delete(this, cbs); endtask endclass

6.4 The Sequence Library ¶

In UVM, it is possible to group similar sequences together into a sequence library. The uvm_sequence_library is an extension of the UVMSequence base class.

class uvm_sequence_library #(type REQ=int, RSP=REQ)

extends UVMSequence#(REQ,RSP);

The uvm_sequence_library is a sequence that contains a list of registered sequence types. It can be configured to create and execute these sequences any number of times using one of several modes of operation, including a user-defined mode. When started (as any other sequence) the sequence library will randomly select and execute a sequence from its sequences queue, depending on the selection_mode chosen.

– UVM_SEQ_LIB_RAND: Randomly select from the queue. – UVM_SEQ_LIB_RANDC: Randomly select from the queue without repeating until all sequences

have executed. – UVM_SEQ_LIB_ITEM: Execute a single item. – UVM_SEQ_LIB_USER: Call the select_sequence() method, which the user may override, to

generate an index into the queue to select a sequence to execute.

The selection mode may be set using the configuration database:

UVMConfigDb.set(this, “<sequencer path>”,

“default_sequence.selection_mode”,

MODE);

To create a sequence library, declare your own extension of uvm_sequence_library and initialize it as follows:

class my_seq_lib extends uvm_sequence_library #(my_item);

’uvm_object_utils(my_seq_lib) ’uvm_sequence_library_utils(my_seq_lib) function new(string name=””);

super.new(name); init_sequence_library(); endfunction … endclass

Individual sequence types may then be added to the sequence library, by type, using the ’uvm_add_to_seq_lib macro:

class my_seq1 extends my_seq; ’uvm_object_utils(my_seq1); ’uvm_add_to_seq_lib(my_seq1, my_seq_lib) ’uvm_add_to_seq_lib(my_seq1, my_other_seq_lib) … endclass

A sequence type may be added to more than one sequence library by having multiple ’uvm_add_to_seq_lib calls in the sequence definition. The parameterization of the sequences and the sequence library must be compatible.

Alternatively, sequences can be added to every instance of a particular type of sequence library using the add_typewide_sequence() and/or add_typewide_sequences() methods:

class mem_seq_lib extends uvm_sequence_library #(mem_item);

… function new(string name=”mem_seq_lib”);

super.new(name); //Explicitly add the memory sequences to the library add_typewide_sequences({mem_seq1::get_type(), mem_seq2::get_type(),

mem_seq3::get_type(), mem_seq4::get_type(), mem_seq5::get_type(), mem_seq6::get_type()}); init_sequence_library(); endfunction : new endclass : mem_seq_lib

The most convenient location to place the add_typewide_sequence() and/or add_typewide_sequences() call is in the constructor of the sequence_library type itself. You may also register a sequence with an individual instance of a sequence library by using the add_sequence() or add_sequences() method. This would typically be done in the test where the sequence is instantiated.

class test_seq_lib extends test_base;

… task main_phase(uvm_phase phase);

phase.raise_objection(this, “Raising Main Objection”); //Register another sequence with this sequence library instance seq_lib.add_sequence( mem_error_seq::get_type() ); //Start the mem sequence seq_lib.start(m_mem_sequencer); //This task call is blocking phase.drop_objection(this, “Dropping Main Objection”); endtask : main_phase endclass : test_seq_lib

The sequence library is then just started as any other sequence.

class test_seq_lib extends test_base;

… task main_phase(uvm_phase phase);

phase.raise_objection(this, “Raising Main Objection”); //Configure the constraints for how many sequences are run seq_lib.min_random_count = 5; seq_lib.max_random_count = 12; //Randomize the sequence library if (!seq_lib.randomize()) `uvm_error(report_id, “The mem_seq_lib library failed to randomize()”) //Start the mem sequence seq_lib.start(m_mem_sequencer); //This task call is blocking phase.drop_objection(this, “Dropping Main Objection”); endtask : main_phase endclass : test_seq_lib

6.5 Advanced Sequence Control ¶

This section discusses advanced techniques for sequence control.

6.5.1 Implementing Complex Scenarios¶

This section highlights how to implement various complex scenarios.

6.5.1.1 Executing Multiple Sequences Concurrently¶

There are two ways you can create concurrently-executing sequences: the following subsections show an example of each method.

6.5.1.1.1 Using the uvm_do Macros with fork/join¶

In this example, the sequences are executed with fork/join. The simulator schedules which sequence requests interaction with the sequencer. The sequencer schedules which items are provided to the driver, arbitrating between the sequences that are willing to provide an item for execution and selects them one at a time. The a and b sequences are subsequences of the fork_join_sequence.

class fork_join_sequence extends UVMSequence #(simple_item);

… // Constructor and UVM automation macros go here.

// See Section 4.7.2 a_seq a; b_seq b; virtual task body();

fork_{`uvm_do(a)} `uvm_do(b) join endtask : body endclass : fork_join_sequence

6.5.1.1.2 Starting several Sequences in Parallel¶

In this example, the concurrent_seq sequence activates two sequences in parallel. It waits for the sequences to complete. Instead, it immediately finishes after activating the sequences. Also, the a and b sequences are started as root sequences.

class concurrent_seq extends UVMSequence #(simple_item);

… // Constructor and UVM automation macros go here.

// See Section 4.7.2 a_seq a; b_seq b; virtual task body();

// Initialize the sequence variables with the factory. `a = a_seq.type_id.create(“a”); b = b_seq.type_id.create(“b”); // Start each subsequence as a new thread. fork_{a.start(m_sequencer);} b.start(m_sequencer); join endtask : body endclass : concurrent_seq

NOTE–The sequence.start() method allows the sequence to be started on any sequencer.

See uvm_create in the UVM 1.2 Class Reference for additional information.

6.5.1.1.3 Using the pre_body()/post_body() a¶

pre_start()/post_start() Callbacks**

The UVM Class Library provides additional callback tasks to use with the various flavors of the `uvm_do macros. pre_body() and pre_start() are invoked before the sequence’s body() task, and post_body() and post_start() are invoked after the body() task. The *_body() callbacks are designed to be skipped for child sequences, while *_start() callbacks are executed for all sequences. When calling the start() method of a sequence, the call_pre_post parameter should be set to 0 for child sequences. The library will automatically set this parameter for sequences started via `uvm_do macros.

NOTE–The *_body() callbacks are included primarily for legacy reasons, with the intention that non-root sequences do not execute them.

The following example declares a new sequence type and implements its callback tasks.

class revised_seq extends fork_join_sequence;

… // Constructor and UVM automation macros go here.

// See Section 4.7.2 task pre_start();

super.pre_start(); // Wait until initialization is done. @p_sequencer.initialization_done; endtask : pre_start task post_start();

super.post_start(); do_cleanup(); endtask : post_start endclass : revised_seq

NOTE–The initialization_done event declared in the sequencer can be accessed directly via the p_se- quencer variable. The p_sequencer variable is available if the ’uvm_declare_p_sequencer macro was used.

6.5.1.2 Interrupt Sequences¶

A DUT might include an interrupt option. Typically, an interrupt should be coupled with some response by the agent. Once the interrupt is serviced, activity prior to the interrupt should be resumed from the point where it was interrupted. Your verification environment can support interrupts using sequences.

6.5.1.2.1 Using grab/ungrab¶

To handle interrupts using sequences:

Define an interrupt handler sequence that will do the following:

1) Wait for the interrupt event to occur. 2) Grab the sequencer for exclusive access. 3) Execute the interrupt service operations using the proper items or sequences. 4) Ungrab the sequencer. b) Start the interrupt-handler sequence in the sequencer or in the default sequence. (You can configure

the sequencer to run the default sequence when the simulation begins.)

Example

Define an interrupt handler sequence.

// Upon an interrupt, grab the sequencer, and execute a

// read_status_seq sequence. class interrupt_handler_seq extends UVMSequence #(bus_transfer);

… // Constructor and UVM automation macros here

// See Section 4.7.2 read_status_seq interrupt_clear_seq;

virtual task body();

forever begin

// Initialize the sequence variables with the factory. @p_sequencer.interrupt; grab(p_sequencer); interrupt_clear_seq.start(p_sequencer) ungrab(p_sequencer); end endtask : body endclass : interrupt_handler_seq

Then, start the interrupt handler sequence in the sequencer. The example below does this in the sequencer itself at the run phase:

class my_sequncer extends uvm_sequencer;

… // Constructor and UVM automation macros here

// See Section 4.7.2 interrupt_handler_seq interrupt_seq; virtual task run_phase(uvm_phase phase);

interrupt_seq =

interrupt_handler_seq.type_id.create(“interrupt_seq”); fork_{interrupt_seq.start(this);} join_none super.run(); endtask : run endclass : my_sequncer

NOTE–In this step, we cannot use any of the `uvm_do macros since they can be used only in sequences. Instead, we use utility functions in the sequencer itself to create an instance of the interrupt handler sequence through the common factory.

6.5.1.2.2 Using Sequence Prioritization¶

A less disruptive approach to implementing interrupt handling using sequences is to use sequence prioritization. Here, the interrupt monitoring thread starts the ISR sequence with a priority higher than the main process. This has the benefit of holding off normal —mainline” sequence operations while at the same time allowing still higher priority interrupts to inject themselves and win arbitration on the sequencer.

Since sequences are functor objects rather than simple subroutines, multiple ISR sequences can be active on a sequencer regardless of priority settings. Priority only affects the individual sequences’ ability to send a sequence_item to the driver. Therefore, whatever processing is happening in an ISR sequence can still continue even if a higher priority sequence interrupts it; the only time priority is considered is when multiple sequences have called and are blocked in their start_item() tasks simultaneously.

The following example demonstrates four ISR sequences started at different priorities, allowing a higher priority ISR to execute in preference to a lower ISR.

class int_test_seq extends UVMSequence #(bus_seq_item);

`uvm_object_utils(int_test_seq) function new(string name = “int_test_seq”);

super.new(name); endfunction task body;

set_ints setup_ints; // Main sequence running on the bus isr ISR0,ISR1,ISR2,ISR3; // Interrupt service routines int_config i_cfg; // Config object contains IRQ monitoring tasks setup_ints = set_ints.type_id.create(“setup_ints”); my_sequencer = get_sequencer(); // ISR0 is highest priority ISR0 = isr.type_id.create(“ISR0”); ISR0.id = “ISR0”; ISR0.i = 0; // ISR1 is medium priority ISR1 = isr.type_id.create(“ISR1”); ISR1.id = “ISR1”; ISR1.i = 1; // ISR2 is medium priority ISR2 = isr.type_id.create(“ISR2”); ISR2.id = “ISR2”; ISR2.i = 2; // ISR3 is lowest priority ISR3 = isr.type_id.create(“ISR1”); ISR3.id = “ISR3”; ISR3.i = 3; i_cfg = int_config::get_config(my_sequencer); // Set up sequencer to use priority based on FIFO order my_sequencer.set_arbitration(SEQ_ARB_STRICT_FIFO); // Main thread, plus one for each ISR fork

setup_ints.start(my_sequencer); // Highest priority forever begin

i_cfg.wait_for_IRQ0(); ISR0.isr_no++; ISR0.start(my_sequencer,this,HIGH); end // Medium priority forever begin

i_cfg.wait_for_IRQ1(); ISR1.isr_no++; ISR1.start(my_sequencer,this,MED); end // Medium priority forever begin

i_cfg.wait_for_IRQ2(); ISR2.isr_no++; ISR2.start(my_sequencer,this,MED); end // Lowest priority forever begin

i_cfg.wait_for_IRQ3(); ISR3.isr_no++; ISR3.start(my_sequencer,this,LOW); end

join_any disable fork; endtask endclass

Some items to note from this example:

– The user-selected priority is passed to each sequence’s start(). The default priority is a value of

Higher values denote higher priority.

– The default arbitration scheme for uvm_sequencer is SEQ_ARB_FIFO, which does not take sequence priority into account. It is critical that the sequencer’s arbitration scheme be changed to either SEQ_ARB_STRICT_FIFO, SEQ_ARB_STRICT_RANDOM, or (if a user-defined arbitration scheme is defined that takes priority into consideration) SEQ_ARB_USER.

6.5.1.3 Controlling the Scheduling of Items¶

There might be several sequences doing items concurrently. However, the driver can handle only one item at a time. Therefore, the sequencer maintains a queue of do actions. When the driver requests an item, the sequencer chooses a single do action to perform from the do actions waiting in its queue. Therefore, when a sequence is doing an item, the do action is blocked until the sequencer is ready to choose it.

The scheduling algorithm works on a first-come-first-served basis. You can affect the algorithm using grab(), ungrab(), and is_relevant().

If a sequence is grabbing the sequencer, then the sequencer will choose the first do action that satisfies the following conditions:

– It is done by the grabbing sequence or its descendants.

– The is_relevant() method of the sequence doing it returns 1.

If no sequence is grabbing the sequencer, then the sequencer will choose the first do action that satisfies the following condition:

The is_relevant() method of the sequence doing it returns 1.

If there is no do action to choose, then get_next_item() is blocked. The sequencer will try to choose again (that is, reactivate the scheduling algorithm) when one of the following happens:

Another do action is added to the queue.

b) A new sequence grabs the sequencer, or the current grabber ungrabs the sequencer.

c) Any one of the blocked sequence’s wait_for_relevant() task returns. See Section 6.5.1.4 for

more information.

When calling try_next_item(), if the sequencer does not succeed in choosing a do action before the time specified in uvm_driver::wait_for_sequences(), uvm_driver::try_next_item() returns with NULL.

6.5.1.4 Run-Time Control of Sequence Relevance¶

In some applications, it is useful to invoke sequences concurrently with other sequences and have them execute items under certain conditions. Such a sequence can therefore become relevant or irrelevant, based on the current conditions, which may include the state of the DUT, the state of other components in the verification environment, or both. To implement this, you can use the sequence is_relevant() function. Its effect on scheduling is discussed in Section 6.5.1.3.

If you are using is_relevant(), you must also implement the wait_for_relevant() task to prevent the sequencer from hanging under certain circumstances. The following example illustrates the use of both.

class flow_control_seq extends UVMSequence #(bus_transfer);

… // Constructor and UVM automation macros go here.

// See Section 4.7.2

bit relevant_flag;

function bit is_relevant();

return(relevant_flag); endfunction

// This task is started by the sequencer if none of the running

// sequences is relevant. The task must return when the sequence // becomes relevant again. task wait_for_relevant(); while(!is_relevant())

@(relevant_flag); // Use the appropriate sensitivity list. endtask

task monitor_credits();

… // Logic goes here to monitor available credits, setting // relevant_flag to 1 if enough credits exist to send // count frames, 0 otherwise. endtask : monitor_credits

task send_frames();

my_frame frame; repeat (count) `uvm_do(frame) endtask : send_frames

virtual task body();

fork:sub:`monitor_credits(); `

send_frames(); join_any endtask : body endclass : flow_control_seq

6.5.2 Protocol Layering¶

This section discusses the layering of protocols and how to implement it using sequences. Sequence layering and agent layering are two ways in which UVM components can be composed to create a layered protocol implementation.

6.5.2.1 Introduction to Layering¶

Many protocols are specified and implemented according to layers. A higher-layer protocol is transparently transported on a lower-layer protocol. That lower-layer protocol may in-turn be transparently transported on an even lower-layer protocol. For example, as illustrated in Figure 33, TCP protocol packets can first be segmented into IPv4 frames; the IPv4 frames can then be encapsulated into Ethernet frames, the Ethernet frames can then be transmitted over a XAUI interface onto a fiber optic medium.

Figure 33–Examples of Protocol Layering

USB transfers are composed of USB transactions which are composed of USB packets which are then transmitted serially on the USB cable. On reception, the protocol layers are interpreted and re-assembled in the opposite order. The IP segments are extracted from the Ethernet frame payload from which TCP packets are re-assembled from the payload of individual IP segments. Similarly, a USB device will interpret a received packets into a specific transaction which are then interpreted by the USB endpoint into a specific transfer.

Some protocol layers are intimately related and are designed to work as an integrated stack. For example, USB transfers are transported exclusively by USB transactions which are then exclusively transported by USB packets. Conversely, a USB transaction can only transport a portion of a USB transfer. However, because a lower-layer is essentially transparent to the higher-layers it transports, many higher-layer protocol may be transported on a variety of lower-level protocols. For example, a TCP packet may be transported over an IPv4, IPv6, or PPP protocol. Conversely, a lower-layer protocol may transport different higher-layer protocols, often at the same time. For example, an Ethernet link may transport a mix of IPv4 frames, IPv6 frames, or UDP packets. To complicate matters even further, lower-level protocols can be transparently tunneled through a higher-layer protocol. For example, an IP stream (itself carrying a variety of higher-layer protocols) can be encrypted then wrapped into TCP packets and tunneled through a secure TCP transport layer to be decrypted and processed at the other end as if they had been natively transported.

Individual transactions from one protocol layers may correspond to transactions at a different layer in many ways (see Figure 34).

a) One-to-one–One and only complete higher-layer transaction is transported by a single lower-layer

transaction. b) One-to-many–One higher-layer transaction is broken into many lower-layer transactions. c) Many-to-one–Many higher-layer transactions are combined into one lower-layer transaction. d) Many-to-many–Many higher-layer transactions are combined into multiple lower-layer transac- tions. Higher-layer transactions may not be fully contained into a single lower-layer transaction and may span two or more lower-layer transactions.

Figure 34–UVM Layer Mapping

When implementing protocol verification IP, it is important that the independence of the protocol layering structure be maintained. For integrated protocol layers, such as USB, sequence layering within a single protocol agent should be used. For protocols that can transport (or be transported by) different protocols, agent layering should be used.

6.5.2.2 Layering in UVM¶

This general approach to layering in UVM uses multiple sequencers and monitors as shown in Figure 35.

Figure 35–UVM Layering Architecture

DUT

seq

packet driver

Packet Sequencer

seq

This lower-layer sequence pulls information directly from the higher-layer Multi-Layer sequencer Architecture

Frame Sequencer

seq

On the generation path, a lower-layer sequencer executes lower-layer sequences and a higher-layer sequencer executes higher-layer sequences. A lower-layer sequence pulls data from the higher-layer sequencer and uses the higher-layer sequence to create corresponding lower-layer sequences and then execute them.

On the monitoring path, a lower-layer monitor observes the execution of lower-layer transactions and publishes the observed transactions on its analysis port. A higher-layer monitor observes the execution of higher-layer transactions by listening to the observed lower-layer transactions, ignoring any non-pertinent lower-layer transactions, then publishes the observed higher-layer transactions on its analysis port.

For integrated protocol layers, the different layer sequencers and monitors are encapsulated in the same agent and they are connected internally. For arbitrarily layered protocols, the different layer sequencers and monitors are encapsulated in separate agents and must be explicitly connected according to the required protocol layering structure in the verification environment.

6.5.2.3 Layering Integrated Protocols¶

An integrated protocol is a protocol with a statically-defined layer structure. The nature and topology of the various layers is well-defined and cannot (or should not) be modified. Such protocols should be layered within a single agent, with a sequencer for each protocol layer instance (see Figure 36).

Integrated Agent

High-Level Sequencer _Monitor High-Level Sequence

Lower-Level

Monitor Sequencer

Layering Sequence

Driver

Figure 36–Integrated Layered Agent

The layering of a higher-layer protocol onto a lower-layer protocol is performed in a layering sequence, running on the lower-layer sequencer. The layering sequence must be created and started on the lower sequencer at the beginning of the run phase, usually as the default run_phase sequence. It behaves like a driver, in that its body() method never returns: the layering sequence is a forever loop that pulls higher- layer sequence items from the higher-layer sequencer, translates them into lower-layer sequence items which are then executed normally.

The layering sequencers contain a handle to the higher-layer sequencer(s) from which information must be pulled. The pulled information (a higher-layer item) is put in a property of the sequence and is then used to constrain various properties in the lower-layer item(s). To maximize reuse and minimize complexity, the layering sequence accesses the get_next_item() and other methods of the higher-level sequencer directly, without having to go through a TLM port. The connection itself is done by setting the handle in the layering sequence at the time the containing agent’s connect() method is invoked.

If the protocol layering requires handshaking or feedback from the lower-layer protocol, first declare an uvm_analysis_imp and a corresponding write() method in the lower-layer sequencer that is connected to the delayering monitor. The sequencer’s write() method will have to notify the layering sequence of the relevant protocol state changes.

6.5.2.3.1 Complex Protocols¶

For more complex protocol hierarchies, multiple layering sequences can be concurrently executed on the lower-layer sequencer (as shown in Figure 37).

Integrated Agent

Higher-layer

Higher-Layer Sequencer B

Sequencer C High-Level

Higher-Layer Sequencer A

Driver

Figure 37–Complex Protocol Stack

6.5.2.3.2 Layered Integrated Protocol Example¶

The following example assumes the existence of lower-layer sequence item, higher-layer sequence item and sequencer classes named lower_item, upper_item, and upper_sequencer respectively. The example further assumes a simple one-to-one mapping between the upper- and lower-layer protocol.

Not all code required for creating a UVM-compliant agent (such as utility macro calls and configuration of virtual interfaces) is shown for brevity and to focus on the implementation of the layering.

The layering sequence is a lower-layer sequence that pulls upper-layer items and translates them to lower- layer items in a forever loop. Because there are no guarantees that the upper-layer items will be available whenever a lower-level item is ready to be executed, the layering sequence should pull the upper-level items as required before calling start_item(). Should it be necessary to deal with the response from the execution of the lower-level item, it can be obtained by the get_response() method of the layering sequence.

class upper_to_lower_seq extends UVMSequence#(lower_item);

… uvm_sequencer #(upper_item) upper_sequencer;

High-Level Sequence

Sequence

High-Level Sequence

Sequencer

Layering Sequence

virtual task body();

upper_item u_item; lower_item l_item; forever begin

upper_sequencer.get_next_item(u_item) l_item = upper_to_lower(u_item); start_item(l_item); finish_item(l_item); // Optional: get_response(rsp); upper_sequencer.item_done(); end endtask : body endclass : upper_to_lower_seq

The upper-layer monitor observes lower-layer items on an analysis export and translates them to upper-layer items, then publishes them on its higher-layer analysis port. Any lower-layer item observed by the monitor that is not relevant to the upper-layer protocol is simply ignored.

class upper_monitor extends

UVMSubscriber#(lower_item); //provides analysis_export of //lower_item type

uvm_analysis_port#(upper_item) ap;

function new(string name, uvm_component parent);

super.new(name, parent); ap = new(“ap”, this); endfunction

virtual function void write(lower_item l_item);

upper_item u_item; if (is_relevant(l_item)) begin

u_item = upper_item.type_id.create(“u_item”,,get_fullname()); lower_to_upper(l_item, u_item); ap.write(u_item); end endfunction: write endclass: upper_monitor

The higher- and lower-layer sequencers and monitors are instantiated in the protocol agent, and are connected during that agent’s connect phase.

class integrated_protocol_agent extends uvm_agent;

lower_driver drv; lower_sequencer l_sqr; upper_sequencer u_sqr; lower_monitor l_mon; upper_monitor u_mon; analysis_port#(upper_item) ap;

function new(string name; uvm_component parent);

super.new(name, parent); ap = new(“ap”, this); endfunction

function void build();

super.build(); // Make lower sequencer execute upper-to-lower translation sequence.

task run_phase(uvm_phase phase);

upper_to_lower_seq u2lseq = upper_to_lower_seq.type_id.create(“u2l”, this); u2lseq.start(l_sqr);

drv = lower_driver.type_id.create(“drv”, this); l_sqr = lower_sequencer.type_id.create((“l_sqr”, this); u_sqr = upper_sequencer.type_id.create((“u_sqr”, this); l_mon = lower_monitor.type_id.create((“l_mon”, this); u_mon = upper_monitor.type_id.create((“u_mon”, this); endfunction : build

function void

connect();drv.seq_item_port.connect(l_sqr.seq_item_export); u2lseq.upper_sequencer = u_sqr; l_mon.ap.connect(u_mon.a_export);

ap.connect(u_mon.ap); endfunction : connect

6.5.2.4 Layering Arbitrary Protocols¶

Arbitrarily-layered protocols are protocols with a user-defined layer structure. The nature and topology of the various layers is not defined by the protocol themselves and can be combined as required by the design under verification. Each protocol exists independently from each other and is modeled using its own agent, often written independently. The protocols should be layered within an environment, with an agent instance for each protocol layer instance.

The layering of a higher-layer protocol onto a lower-layer protocol is performed in a layering driver, that replaces the default driver in the higher-layer agent (see Figure 38).

Layering of UVM drivers. — Layering Drivers¶

A layering driver is implemented using the same familiar techniques used to implement a —regular” driver, except that a higher-layer sequence item is executed in terms of lower-level sequence items instead of pin wiggling through a virtual interface. Similarly, the default monitor in the higher-layer agent is replaced with a de-layering monitor.

The layering driver and monitor must supersede the default driver and monitor using the class factory. The layering driver is connected to the lower-layer sequencer using a pass thru sequence. A pass thru sequence is a simple sequence that executes a single sequence item in a directed fashion, without randomizing it. A pass thru sequence is used in preference to using the uvm_sequencer_base::send_request() or uvm_sequencer_base::execute_item() methods so it can be configured with state information, such as priority, to shape the lower-level protocol traffic.

The pass thru sequence is functionally equivalent to the virtual interface of a —regular” driver and thus should similarly be passed by the containing environment the configuration database. The connection between the agent’s monitoring paths is performed at the time the containing environment’s connect() method is invoked, using the same technique as when connecting a subscriber to an analysis port.

6.5.2.4.1 Multiple Pass thru Sequences¶

When creating complex protocol hierarchies (see Figure39), multiple pass thru sequences can be concurrently executed on the lower-layer agent, one for each higher-layer protocol.

6.5.2.4.2 Layered Arbitrary Protocol Example¶

The following example assumes the existence of a lower-layer agent and higher-layer agent named lower_agent and upper_agent respectively, with the usual item, driver and monitor classes named lower_item, lower_driver, lower_monitor, upper_item, upper_driver, and upper_monitor. The example further assumes a simple one-to-one mapping between the upper and lower-layer protocol. A complete example can be found in the UVM distribution under the directory test/examples/simple/layering/agents.

Not all code required for creating UVM-compliant components (such as utility macro calls and configuration of virtual interfaces) is shown for brevity and to focus on the implementation of the layering.

First, a lower-layer pass thru sequence must be defined. It is a trivial sequence that will be used to execute individual lower-level items and optionally get their response. The sequence has no body() method. The items will be executed by explicitly calling its start_item() and finish_item() methods from the layering driver:

class lower_passthru_seq(UVMSequence):

    def __init__(self, name=“lower_passthru_sequence”):
        super().name()

The layering driver is an upper-layer driver that pulls upper-layer items and translates them to lower-layer items in a forever loop. This is accomplished in the driver’s run_phase() method. It is important NOT to call the run_phase() method of the base class as it implements the default driver functionality the layering driver is designed to replace. The lower-level items are executed by directly calling the start_item() and finish_item() methods of the pass thru sequence and specifying the lower-level sequencer where the item is executed. The reference to the lower-level sequencer is obtained from the configuration database at the beginning of the run_phase. Notice it is not necessary to explicitly start the pass thru sequence as it is only a context placeholder for executing the lower-level items. Should it be necessary to deal with the response from the execution of the lower-level item, it can be obtained by similarly directly calling the get_response() method of the pass thru sequence. The layering driver should not raise objections: it is the responsibility of the higher-layer sequence to object as necessary:

class upper_layering_driver(upper_driver):

    ...

    async def run_phase(self, phase):
        # DO NOT CALL super.run_phase()!!
        lower_passthru_seq l_seq
        l_sqr = []
        UVMConfigDb.get(self, "“,”lower_sqr“, l_sqr);
        l_sqr = l_sqr[0]
        l_seq = lower_passthru_seq.type_id.create(”l_seq", self)

        while True:
            u_item = []
            await self.seq_item_port.get_next_item(u_item)
            u_item = u_item[0]
            l_item = upper_to_lower(u_item)
            await l_seq.start_item(l_item, -1, sqr)
            l_seq.finish_item(l_item, -1, sqr)
            # Optional: l_seq.get_response(rsp);
            self.seq_item_port.item_done()

The (de)layering monitor is an upper-layer monitor that observes lower-layer items on an analysis export and translates them to upper-layer items, then publishes them on its higher-layer analysis port. Any lower- layer item observed by the monitor that is not relevant to the upper-layer protocol is simply ignored. The run_phase() method is left empty and does not call super.run_phase() to disable the normal higher-layer monitor functionality:

class upper_layering_monitor(upper_monitor):

    uvm_analysis_imp#(lower_to_upper_monitor, lower_item) a_export;

    def __init__ (self, name, parent):
        super().__init__(name, parent)
        self.a_export = UVMAnalysisImp(“a_export”, self)

    def write(self, l_item):
        # upper_item u_item;
        if self.is_relevant(l_item):
            u_item = self.lower_to_upper(l_item)
            self.ap.write(u_item)


    async def run_phase(self, phase):
        # DO NOT CALL super.run_phase(phase)

If the protocol layering requires handshaking or feedback from the lower-layer protocol, first declare an UVMAnalysisImp and a corresponding write() method in the layering driver that is connected to the delayering monitor. The driver’s write() method will have to notify the layering thread of the relevant protocol state changes.

The higher-layer and lower-layer agents are instantiated in a layered environment. The layering driver and monitor are instantiated in the higher-layer agent using the class factory. The lower-level sequencer instance is configured as the lower-level sequencer for the layering driver using the configuration DB. Note this must be done in the connect phase as the sequencer instance does not yet exists at the completion of the build phase. The analysis port of the lower-layer agent is connected to the analysis export of the higher-layer layering monitor using the usual connection mechanism:

class layered_env(UVMEnv):

    upper_agent u_agent;
    lower_agent l_agent;

    function build_phase(uvm_phase phase);
        l_seq = new();
        l_agent = lower_agent.type_id.create(“l_agent”, this); u_agent =
        upper_agent.type_id.create(“u_agent”, this);
        set_inst_override_by_type(“u_agent.drv”, upper_driver::get_type(),

        upper_layering_driver::get_type());
        set_inst_override_by_type(“u_agent.mon”, upper_monitor::get_type(),

        upper_layering_monitor::get_type()); end


    function void connect_phase(uvm_phase phase);
        upper_layering_monitor u_mon; super.connect_phase(phase);
        UVMConfigDb.set(this, “u_agent.drv”, “lower_sqr”,

        l_agent.sqr); $cast(u_mon, u_agent.mon);

        l_agent.ap.connect(u_mon.a_export);

6.5.2.5 Dynamic Protocol Layering Structure¶

In most networking application, it is necessary to verify the dynamic reprovisioning of components and the ability of components to adapt to changes in the protocol topology. For example, adding and removing devices and hubs from a USB network is a normal operation of that protocol which modifies the structure of the protocol layers. Therefore, it should be possible for the layered protocol structure to adapt to dynamic changes, and be able to add and remove additional protocol layers and sibling protocol streams.

Unfortunately, due to the static nature of UVM components, the protocol layer hierarchy cannot be dynamically modified at run-time; all alternative protocol hierarchies must be created entirely at build time. Different protocol layer topologies can then be created by starting and stopping the corresponding layering or pass thru sequence at the root of the protocol stack that must be introduced or removed.

6.5.3 Generating the Item or Sequence in Advance¶

The various uvm_do functions perform several steps sequentially, including the allocation of an object (sequence or sequence item), synchronization with the driver (if needed), randomization, sending to the driver, and so on. The UVM Class Library provides additional functions that enable finer control of these various steps. This section describes these function variations, which represent various combinations of calling the standard methods on sequence items:

async def body():
    req = request_time.type_id.create(“req”); //set values in ‘req’ here
    await self.start_item(req, priority, sequencer)
    req.randomize(); // or with constraints
    self.finish_item(req, priority)

6.5.3.1 uvm_create¶

This function allocates an object using the common factory and initializes its properties. Its argument is a variable of type UVMSequenceItem or UVMSequence. You can use the function with uvm-python’s sv.constraint_mode() and sv.rand_mode() functions to control subsequent randomization of the sequence or sequence item.

In the following example, my_seq is similar to previous sequences that have been discussed. The main differences involve the use of the uvm_create(item0) call. After the function call, the rand_mode() and constraint_mode() functions are used and some direct assignments to properties of item0 occur. The manipulation of the item0 object is possible since memory has been allocated for it, but randomization has not yet taken place. Subsequent sections will review the possible options for sending this pre-generated item to the driver:

class my_seq extends UVMSequence #(my_item);

    ... // Constructor and UVM automation macros go here.

    // See Section 4.7.2
    virtual task body();
        \`uvm_create(req)
        req.addr.rand_mode(0); // Disables randomization of addr
        req.dc1.constraint_mode(0); // Disables constraint dc1 req.addr = 27;
        ...
    endtask : body

endclass: my_seq

You can also use a sequence variable as an argument to uvm_create.

NOTE–You might need to disable a constraint to avoid a conflict.

6.5.3.2 uvm_send¶

This macro processes the UVMSequenceItem or UVMSequence class handle argument as shown in Figure15 and Figure16, without any allocation or randomization. Sequence items are placed in the sequencer’s queue to await processing while subsequences are processed immediately. The parent pre_do(), mid_do(), and post_do() callbacks still occur as shown.

In the following example, we show the use of uvm_create() to pre-allocate a sequence item along with uvm_send, which processes it as shown in Figure 15, without allocation or randomization:

class my_seq2 extends UVMSequence #(my_item);

    ... // Constructor and UVM automation macros go here.

    // See Section 4.7.2
    virtual task body();
        uvm_create(req)
        req.addr = 27;
        req.data = 4; // No randomization. Use a purely pre-generated item.
        uvm_send(req)
    endtask : body
endclass: my_seq2

Similarly, a sequence variable could be provided to the uvm_create and uvm_send calls above, in which case the sequence would be processed in the manner shown in Figure 16, without allocation or randomization.

6.5.3.3 uvm_rand_send, uvm_rand_send_with¶

These macros are identical to uvm_send (see Section 6.5.3.2), with the single difference of randomizing the given class handle before processing it. This enables you to adjust an object as required while still using class constraints with late randomization, that is, randomization on the cycle that the driver is requesting the item. uvm_rand_send() takes just the object handle. uvm_rand_send_with() takes an extra argument, which can be any valid inline constraints to be used for the randomization.

The following example shows the use of uvm_create to pre-allocate a sequence item along with the uvm_rand_send macros, which process it as shown in Figure15, without allocation. The rand_mode() and constraint_mode() constructs are used to show fine-grain control on the randomization of an object:

class my_seq3 extends UVMSequence #(my_item);

    ...
    // Constructor and UVM automation macros go here.

    // See Section 4.7.2
    virtual task body(); \`uvm_create(req)
        req.addr.rand_mode(0);
        req.dc1.constraint_mode(0);
        req.addr = 27; // Randomize and process the item.
        uvm_rand_send(req) // Randomize and process again, this time with inline constraints.
        uvm_rand_send_with(req, {data < 1000;})
    endtask : body

endclass: my_seq3

6.5.4 Executing Sequences and Items on other Sequencers¶

In the preceding sections, all uvm_do macros (and their variants) execute the specified item or sequence on the current p_sequencer. To allow sequences to execute items or other sequences on specific sequencers, additional macro variants are included that allow specification of the desired sequencer.

“uvm_do_on**,”uvm_do_on_with**, “uvm_do_on_pri**, and”uvm_do_on_pri_with**

All of these macros are exactly the same as their root versions, except they all take an additional argument (always the second argument) that is a reference to a specific sequencer.

’uvm_do_on(s_seq, that_sequencer); ’uvm_do_on_with(s_seq, that_sequencer, {s_seq.foo == 32’h3;})

6.6 Command Line Interface (CLI)¶

6.6.1 Introduction¶

The Command Line Processor class provides a general interface to the command line arguments that are provided for the given simulation. Not only can users retrieve the complete arguments using methods such as ~get_args()~ and ~get_arg_matches()~, but they can also retrieve the suffixes of arguments using ~get_arg_values()~.

The UVMCmdlineProcessor class also provides support for setting various UVM variables from the command line, such as components’ verbosities and configuration settings for integral types and strings. Command line arguments that are in UPPERCASE should only have one setting to invocation. Command line arguments in lowercase can have multiple settings per invocation. All of this is further described in UVMCmdlineProcessor in the UVM 1.2 Class Reference.

6.6.2 Getting Started¶

To start using the UVMCmdlineProcessor, the user needs to first get access to the singleton instance of the UVMCmdlineProcessor:

cmdline_processor = UVMCmdlineProcessor::get_inst();

A common use case involves using the get_arg_value() function to get the value of a specific argument, which is returned through an output argument. The total number of matches returned from this function usually is of interest when there are no matches and no default value. In this case, the user may generate an error. Similar to $test$plusargs, if the command line contains multiple matching arguments, the first value is returned:

my_value = "default_value"
rc = cmdline_processor.get_arg_value("+abc=", my_value)

If the user knows the value is an integer, this string value may be further turned into an integer by calling the SystemVerilog function atoi() as follows:

my_int_value = int(my_value)

If processing multiple values makes sense for a particular option (as opposed to just the first one found), use the get_arg_values() function instead, which returns a queue of all the matches:

string my_value_list[$];
int rc = cmdline_process.get_values("+abc=", my_value_list);

The UVMCmdlineProcessor provides comprehensive access to the command line processing; see Section 6.6.3 and the UVM 1.2 Class Reference for more details.

6.6.3 UVM-aware Command Line Processing¶

This section highlights how to select tests, set verbosity, and control other UVM facilities using the CLI.

6.6.3.1 Selecting Tests¶

The UVMCmdlineProcessor is used to pass the +UVM_TESTNAME option to the UVMRoot.run_test() routine to select which class will get constructed as the top-level testcase.

6.6.3.2 Setting Verbosity¶

The UVMCmdlineProcessor looks for the +UVM_VERBOSITY option to change the verbosity for all UVM components. It is also possible to control the verbosity in a much more granular way by using the +uvm_set_verbosity option. The +uvm_set_verbosity option has a specific format that allows control over the phases where the verbosity change applies, and in the case of time-consuming phases, exactly what time it applies. Typically, verbosity is only turned up during time-consuming phases as the test approaches the time where an error occurs to help in debugging that error. The simulation will run faster if it is not burdened by generating debug messages earlier on where they are not required.

The +uvm_set_verbosity option is used as follows:

sim_cmd +uvm_set_verbosity=component_name,id,verbosity,phase_name,optional_time

In a similar fashion, the severity, and also the action taken, can be modified as follows:

sim_cmd +uvm_set_action=component_name,id,severity,action
sim_cmd +uvm_set_severity=component_name,id,current_severity,new_severity

When using cocotb-Makefiles, add SIMARGS+=”+uvm_set_verbosity…” to Makefile instead.

6.6.3.3 Other UVM facilities that can be Controlled from the Command Line¶

Table 14 shows other UVM options the user can set from the CLI/Makefile.

Table 14–UVM CLI Options

Facility Setting

Instance-specific factory override +uvm_set_inst_override

Type-specific factory override +uvm_set_type_override

Integer configuration +uvm_set_config_int

String configuration +uvm_set_config_string

Timeout +UVM_TIMEOUT

Max quit count +UVM_MAX_QUIT_COUNT

Objection mechanism debug +UVM_OBJECTION_TRACE

Please see the UVM 1.2 Class Reference for more examples of using the UVMCmdlineProcessor class facilities.

6.7 Mixins in UVM ¶

To reduce coding overhead, the UVM library provides a set of mixin functions. These mixins can be used to combine the definition of multiple things into one step (e.g., declare a field and a task, when both are required to exist at the same time). However, you are not required to use these mixins and may instead choose to build the expanded (required) code yourself.

It is important to understand the decision of mixin vs non-mixin usage is typically a —performance” vs —creation speed and/or maintenance effort” decision. A few points might help to find the best choice for your particular usage. If in doubt, try to profile the code before you assume it is too slow.

Such usage can be a personal or project preference.

The UVM mixin can be considered as a kind of —first class citizens”, similar to functions and tasks. However, bug fixes, use model changes, and performance improvements inside the mixins will obviously only reach the users of such mixins.

Performance differences between the two approaches may heavily depend upon the mixin used, exact use model, tool vendor, and tool version.

The benefits of a hand-coded implementation (non-macro) need to be weighed against those of having a shorter user code, but eventually utilizing more generic code.

Neither path (macro or non-macro) is exclusive within a class or project. Special care must be taken when mixing the two within a single class or inheritance tree, as the automation provided by the uvm_field mixins will always execute prior to the handcrafted implementation. The only exception here is “auto configuration” which relies upon the field registration functions, will only execute when super().build_phase() is called. When functions are used incorrectly, error/warning messages might not directly indicate the source of the issue and, during simulation, debugging with functions might be more challenging.

It is also important to understand that different mixin types are used in different contexts.

Registration macros (uvm*utils*, callbacks, constants, etc.) are

typically used once and do not incur a performance penalty.

Sequence macros (uvm_do*)

typically expand in just a few lines and do not create any over- head, but for really fine grained sequence and item control, the underlying sequence functions may be used directly.

Field utils mixin function

(uvm_field_*type*) provide for a number of field types (int, enum, string, object, aa, sarray, queue, etc.), which are generic implementations for that particular field of a print(), compare(), pack(), unpack(), or record() function. Handcrafted implementations for these functions can be supplied instead of the generic code provided by the uvm_field_* macros by implementing the do_print(), do_compare(), do_pack(), do_unpack(), or do_record() functions respectively.

The uvm_field_* macro implementations have a non-programmable execution

order within the compare/copy/etc. methods, whereas the handcrafted implementations provide full con- trol over this order via do_compare/do_copy/etc. methods.

The UVM distribution comes with two examples of small environments illustrating the usage with and without macro usage. The examples are located under test/examples/simple/basic_examples/pkg and test/examples/simple/sequence/basic_read_write_sequence.

Example:

class bus_trans extends UVMSequenceItem;

    bit [11:0] addr; bit [7:0] data; bus_op_t op; \`ifdef USE_FIELD_MACROS

    \`uvm_object_utils_begin(bus_trans)

    \`uvm_field_int(addr,UVM_DEFAULT)
    \`uvm_field_int(data,UVM_DEFAULT|UVM_NORECORD)
    \`uvm_field_enum(bus_op_t,op,UVM_DEFAULT) \`uvm_field_utils_end \`else

    \`uvm_object_utils(bus_trans)

    virtual function void do_copy (uvm_object rhs);
        bus_trans rhs_;
        if(!$cast(rhs_, rhs))
            \`uvm_error("do_copy", "$cast failed, check type compatability")
        super.do_copy(rhs);
        addr = rhs_.addr;
        data = rhs_.data;
        op = rhs_.op;
    endfunction

    virtual function bit do_compare(uvm_object rhs,uvm_comparer comparer);
        bus_trans rhs_;
        if(!$cast(rhs_, rhs))
            \`uvm_fatal("do_compare", "cast failed, check type compatability")
            return ((op == rhs_.op) && (addr == rhs_.addr) && (data == rhs_.data));
    endfunction

    virtual function void do_print(uvm_printer printer);
        super.do_print(printer);
        printer.print_generic("op", "bus_op_t", $size(bus_op_t), op.name());
        printer.print_int("int", addr, $size(addr));
        printer.print_int("int", data, $size(data));
    endfunction

    virtual void do_record(uvm_recorder recorder);
        if (!is_recording_enabled())
            return;
        super.do_record(recorder);
        uvm_record_int("int", addr, $bits(addr))
        uvm_record_int("int", data, $bits(data))
    endfunction

    virtual function void do_pack (uvm_packer packer);
        super.do_pack(packer);
        uvm_pack_enum(op)
        uvm_pack_int(data)
        uvm_pack_int(addr)
    endfunction

    virtual function void do_unpack(uvm_packer packer);
        super.do_unpack(packer);
        uvm_unpack_enum(op,bus_op_t)
        uvm_unpack_int(data)
        uvm_unpack_int(addr)
    endfunction \`endif

    function new(string name="");
        super.new(name);
    endfunction

endclass

7. UBus Verification Component Example ¶

This chapter introduces the basic architecture of the UBus verification component. It also discusses an executable demo you can run to get hands-on experience in simulation. The UBus source code is provided as a further aid to understanding the verification component architecture. When developing your own simulation environment, you should follow the UBus structure and not its protocol-specific functionality.

All UBus verification component subcomponents inherit from some base class in the UVM Class Library, so make sure you have the UVM 1.2 Class Reference available while reading this chapter. It will be important to know, understand, and use the features of these base classes to fully appreciate the rich features you get–with very little added code–right out of the box.

You should also familiarize yourself with the UBus specification in Chapter 8. While not a prerequisite, understanding the UBus protocol will help you distinguish UBus protocol-specific features from verification component protocol-independent architecture.

7.1 UBus Example ¶

The UBus example constructs an verification environment consisting of a master and a slave. In the default test, the UBus slave communicates using the slave_memory sequence and the UBus master sequence read_modify_write validates the behavior of the UBus slave memory device. Instructions for running the UBus example can be found in the readme.txt file in the examples/ubus/examples directory of the UVM kit.

The output from the simulation below shows the UBus testbench topology containing an environment.The environment contains one active master and one active slave agent. The test runs the read_modify_write sequence, which activates the read byte sequence followed by the write byte sequence, followed by another read byte sequence. An assertion verifies the data read in the second read byte sequence is identical to the data written in the write byte sequence. The following output is generated when the test is simulated with UVM_VERBOSITY = UVM_LOW:

# UVM_INFO @ 0: reporter [RNTST] Running test test_read_modify_write...
# UVM_INFO test_lib.sv(55) @ 0: uvm_test_top [test_read_modify_write]
# Printing the test topology :
#--------------------------------------------------------------------
# Name Type Size Value
#--------------------------------------------------------------------
# uvm_test_top test_read_modify_write - @350
# ubus_example_tb0 ubus_example_tb - @372
# scoreboard0 ubus_example_scoreboard - @395
# item_collected_export uvm_analysis_imp - @404
# disable_scoreboard integral 1 'h0
# num_writes integral 32 'd0
# num_init_reads integral 32 'd0
# num_uninit_reads integral 32 'd0
# recording_detail uvm_verbosity 32 UVM_FULL
# ubus0 ubus_env - @386
# bus_monitor ubus_bus_monitor - @419
# masters[0] ubus_master_agent - @454
# slaves[0] ubus_slave_agent - @468
# has_bus_monitor integral 1 'h1
# num_masters integral 32 'h1
# num_slaves integral 32 'h1
# intf_checks_enable integral 1 'h1
# intf_coverage_enable integral 1 'h1
# recording_detail uvm_verbosity 32 UVM_FULL
# recording_detail uvm_verbosity 32 UVM_FULL
# --------------------------------------------------------------------
#
# UVM_INFO ubus_example_scoreboard.sv(100) @ 110:
  uvm_test_top.ubus_example_tb0.scoreboard0 [ubus_example_scoreboard]
  READ to empty address...Updating address : b877 with data : 91
# UVM_INFO ../sv/ubus_bus_monitor.sv(223) @ 110:

Rest of the original output removed.

7.2 UBus Example Architecture ¶

Figure 40 shows the testbench topology of the UBus simulation environment in the UBus example delivered with this release.

7.3 UBus Top Module ¶

The UBus testbench is instantiated in a top-level module to create a class-based simulation environment. The example below uses an example DUT with UBus-specific content. The example is intentionally trivial so the focus is on the UBus verification component environment.

The top module contains the typical HDL constructs and a SystemVerilog interface. This interface is used to connect the class-based testbench to the DUT. The UBus environment inside the testbench uses a virtual interface variable to refer to the SystemVerilog interface. The following example shows the UBus interface (xi0) and the example DUT connected together and passed to the components in the testbench via the resource database (Line 17). The run_test() command used to simulate the DUT and the testbench is covered in the next section.

Example: ubus_tb_top.sv

1 module ubus_tb_top; 2 _{3 import uvm_pkg::*;} 4 import ubus_pkg::; 5 `include “test_lib.sv” 6 :sub:`7 ubus_if vif(); // SystemVerilog interface to the DUT` 8 :sub:`9 dut_dummy dut(` 10 vif.sig_request[0], 11 … 12 vif.sig_error 13 ); 14 :sub:`15 initial begin` 16 automatic uvm_coreservice_t cs_ = uvm_coreservice_t::get(); 17 UVMConfigDb#(virtual ubus_if)::set(cs_.get_root(),””, 18 “vif”,vif); 19 run_test(); 20 end 21 _{22 initial begin} 23 vif.sig_reset <= 1’b1; 24 vif.sig_clock <= 1’b1; 25 #51 vif.sig_reset = 1’b0; 26 end 27 _{28 //Generate clock.} 29 always 30 #5 vif.sig_clock = ~vif.sig_clock; 31 :sub:`32 endmodule `

The UBus SystemVerilog interface is instantiated in the top-level testbench module. The interface uses generally-accepted naming conventions for its signals to allow easy mapping to any naming conventions employed by other implementations of the UBus protocol. The DUT pins connect directly to the signal inside the interface instance. Currently, the signals are simple non-directional variables that are driven either by the DUT or the class-based testbench environment via a virtual interface. The UBus interface contains concurrent assertions to perform physical checks. Refer to Section4.8 and Section7.12 for more information.

7.4 The Test ¶

In UVM, the test is defined in a separate class, test_read_modify_write. It derives from ubus_example_base_test that, in turn, derives from UVMTest. The ubus_example_base_test test builds the ubus_example_tb object and manages the run_phase() phase of the test. Subsequent derived tests, such as test_read_modify_write, can leverage this functionality as shown in the example below.

All classes that use the `uvm_component_utils macros are registered with a common factory, uvm_factory. When the top module calls run_test(test_name), the factory is called upon to create an instance of a test with type test_name and then simulation is started. When run_test is called without an argument, a +UVM_TESTNAME=test_name command-line option is checked and, if it exists, the test with that type name is created and executed. If neither are found, all constructed components will be cycled through their simulation phases. Refer to Section 4.5 for more information.

Example: test_lib.sv

1 `include “ubus_example_tb.sv” 2 _{3 class ubus_example_base_test extends UVMTest;} 4 _{5 ubus_example_tb ubus_example_tb0; // UBus verification environment} 6 uvm_table_printer printer; 7 bit test_pass = 1; 8 _{9 function new(string name = “ubus_example_base_test”,} 10 uvm_component parent=null); 11 super.new(name, parent); 12 endfunction 13 // UVM build_phase() phase 14 virtual function void build_phase(uvm_phase phase); 15 super.build_phase(phase); 16 // Enable transaction recording for everything. 17 UVMConfigDb.set(this, “*”, “recording_detail”, UVM_FULL); 18 // Create the testbench. 19 ubus_example_tb0 =

ubus_example_tb.type_id.create(“ubus_example_tb0”, this); 20 // Create specific-depth printer for printing the created topology. 21 printer = new(); 22 printer.knobs.depth = 3; 23 endfunction: build_phase 24 // Built-in UVM phase 25 function void end_of_elaboration_phase(uvm_phase phase); 26 // Set verbosity for the bus monitor for this demo. 27 if(ubus_example_tb0.ubus0.bus_monitor != null) 28 ubus_example_tb0.ubus0.bus_monitor.set_report_verbosity_level

(UVM_FULL); 29 // Print the test topology. 30 `uvm_info(get_type_name(), 31 $sformatf(“Printing the test topology :n%s”, 32 this.sprint(printer)), UVM_LOW) 33 endfunction : end_of_elaboration_phase(); 34 // UVM run_phase() phase 35 task run_phase(uvm_phase phase); 36 //set a drain-time for the environment if desired 37 phase.phase_done.set_drain_time(this, 50); 38 endtask: run_phase

39 _{40 function void extract_phase(uvm_phase phase);} 41 if(ubus_example_tb0.scoreboard0.sbd_error) 42 test_pass = 1’b0; 43 endfunction // void 44 _{45 function void report_phase(uvm_phase phase);} 46 if(test_pass) begin 47 `uvm_info(get_type_name(), “* UVM TEST PASSED *”, UVM_NONE) 48 end 49 else begin 50 `uvm_error(get_type_name(), “* UVM TEST FAIL *”) 51 end 52 endfunction 53 endclass : ubus_example_base_test

Line1 Include the necessary file for the test. The testbench used in this example is the ubus_example_tb that contains, by default, the bus monitor, one master, and one slave. See Section 7.5.

Line 3 All tests should derive from the UVMTest class and use the uvm_component_utils or the uvm_component_utils_begin/uvm_component_utils_end macros. See the UVM 1.2 Class Reference for more information.

Line 5 Declare the testbench. It will be constructed by the build_phase() function of the test.

Line 6 Declare a printer of type uvm_table_printer, which will be used later to print the topology. This is an optional feature. It is helpful in viewing the relationship of your topology defined in the configuration and the physical testbench created for simulation. Refer to the UVM 1.2 Class Reference for different types of printers available.

Line 14 - Line 23 Specify the build_phase() function for the base test. As required, build first calls the super.build_phase() function in order to update any overridden fields. Then the ubus_example_tb is created using the create() function. The build_phase() function of the ubus_example_tb is executed by the UVM library phasing mechanism during build_phase(). The user is not required to explicitly call ubus_example_tb0.build_phase().

Line 25 - Line 33 Specify the end_of_elaboration_phase() function for the base test. This function is called after all the component’s build_phase() and connect() phases are executed. At this point, the test can assume that the complete testbench hierarchy is created and all testbench connections are made. The test topology is printed.

Line 35 - Line 38 Specify the run_phase() task for the base test. In this case, we set a drain time of 50 micro-seconds. Once all of the end-of-test objections were dropped, a 50 micro-second delay is introduced before the run phase it terminated.

Now that the base test is defined, a derived test will be examined. The following code is a continuation of the test_lib.sv file:

class test_read_modify_write(ubus_example_base_test):

    def __init__(self, name="test_read_modify_write", parent=None):
        super().__init__(name, parent)

    def build_phase(self, phase):
        UVMConfigDb.set(self, "ubus_example_tb0.ubus0.masters[0].sequencer.run_phase",
                "default_sequence", read_modify_write_seq.type_id.get())
        UVMConfigDb.set(self, "ubus_example_tb0.ubus0.slaves[0].sequencer.run_phase",
                "default_sequence", slave_memory_seq.type_id.get())
        #    // Create the tb
        ubus_example_base_test.build_phase(self, phase)


    async def run_phase(self, phase):
        phase.raise_objection(self, "test_read_modify_write OBJECTED")
        master_sqr = self.ubus_example_tb0.ubus0.masters[0].sequencer
        slave_sqr = self.ubus_example_tb0.ubus0.slaves[0].sequencer

        uvm_info("TEST_TOP", "Forking master_proc now", UVM_LOW)
        master_seq = read_modify_write_seq("r_mod_w_seq")
        master_proc = cocotb.fork(master_seq.start(master_sqr))

        slave_seq = slave_memory_seq("mem_seq")
        slave_proc = cocotb.fork(slave_seq.start(slave_sqr))
        #await [slave_proc, master_proc.join()]
        await sv.fork_join_any([slave_proc, master_proc])
        phase.drop_objection(self, "test_read_modify_write drop objection")
        if not master_seq.test_pass:
            self.test_pass = False

The build_phase() function of the derivative test, test_read_modify_write, is of interest. The build_phase() function uses the resource database to set the master agent sequencer’s default sequence for the main() phase to use the read_modify_write_seq sequence type. Similarly, it defines the slave agent sequencer’s default sequence for the run_phase() phase to use the slave_memory_seq sequence type. Once these resources are set, super.build_phase() is called which creates the ubus_example_tb0 as specified in the ubus_example_base_test build function.

The run_phase() task implementation is inherited by test_read_modify_write since this test derives from the ubus_example_base_test. Since that implementation is sufficient for this test, no action is required by you. This greatly simplifies this test.

7.5 Testbench Environment ¶

This section discusses the testbench created in the Example: test_lib.sv in Section 7.4. The code that creates the ubus_example_tb is repeated here:

ubus_example_tb0 = ubus_example_tb.type_id.create("ubus_example_tb0", self)

In general, testbenches can contain any number of envs (verification components) of any type: ubus, pci, ahb, ethernet, and so on. The UBus example creates a simple testbench consisting of a single UBus environment (verification component) with one master agent, slave agent, and bus monitor (see Figure 41).

The following code defines a class that specifies this configuration. The test will create an instance of this class.

Example: ubus_example_tb.sv:

function void ubus_example_tb::build_phase();
super.build_phase();
UVMConfigDb.set(this,".ubus0",
"num_masters", 1);
UVMConfigDb.set(this,".ubus0",
"num_slaves", 1);
ubus0 = ubus_env.type_id.create("ubus0", this);
scoreboard0 = ubus_example_scoreboard.type_id.create("scoreboard0", this);
endfunction : build
10
11 function void ubus_example_tb::connect_phase(uvm_phase phase);
// Connect the slave0 monitor to scoreboard.
ubus0.slaves[0].monitor.item_collected_port.connect(
scoreboard0.item_collected_export);
endfunction : connect_phase
16
17 function void end_of_elaboration_phase();
// Set up slave address map for ubus0 (basic default).
ubus0.set_slave_address_map("slaves[0]", 0, 16'hffff);
endfunction : end_of_elaboration

Line 1 Declare the build_phase() function.

Line 2 Call super.build_phase() in order to update any overridden fields. This is important because the test, which creates the testbench, may register overrides for the testbench. Calling super.build_phase() will ensure that those overrides are updated.

Line3 - Line5 The UVMConfigDb.set calls are adjusting the num_masters and num_slaves configuration fields of the ubus_env. In this case, the ubus0 instance of the ubus_env is being manipulated. Line 3 instructs the ubus0 instance of the ubus_env to contain one master agent. The num_masters property of the ubus_env specifies how many master agents should be created. The same is done for num_slaves.

Line 7 Create the ubus_env instance named ubus0. The create() call specifies that an object of type ubus_env should be created with the instance name ubus0.

Line 7 As with ubus0, the scoreboard is created.

Line 11 Declare the connect_phase() function.

Line 12 Make the connections necessary for the ubus0 environment and the scoreboard0 between the analysis port on the ubus0.slaves[0].monitor and the analysis export on the scoreboard0 instance.

Line 17 Declare the end_of_elaboration_phase() built-in UVM phase.

Line 19 Assign the slave address map for the slaves[0]. Since all components in the complete testbench have been created and connected prior to the start of end_of_elaboration_phase(), the slave instances are guaranteed to exist at this point.

7.6 UBus Environment ¶

The ubus_env component contains any number of UBus master and slave agents. In this demo, the ubus_env (shown in Figure 42) is configured to contain just one master and one slave agent.

NOTE–The bus monitor is created by default.

The build_phase() function of the ubus_env creates the master agents, slave agents, and the bus monitor. Three properties control whether these are created. The source code is shown here:

def build_phase(self, phase):
    inst_name = ""
    super().build_phase(phase)

    arr = []
    if UVMConfigDb.get(None, "*", "vif", arr):
        uvm_info("GOT_VIF", "vif was received from configDb", UVM_HIGH)
        self.vif = arr[0]

    if self.vif is None:
        self.uvm_report_fatal("NOVIF","virtual interface must be set for: " +
                self.get_full_name() + ".vif")

    if self.has_bus_monitor is True:
        self.bus_monitor = ubus_bus_monitor.type_id.create("bus_monitor", self)

    arr = []
    if UVMConfigDb.get(self, "", "num_masters", arr) is True:
        self.num_masters = arr[0]

    for i in range(self.num_masters):
        inst_name = sv.sformatf("masters[%0d]", i)
        master = ubus_master_agent.type_id.create(inst_name, self)
        self.masters.append(master)
        UVMConfigDb.set(self, inst_name + ".monitor", "master_id", i)
        UVMConfigDb.set(self, inst_name + ".driver", "master_id", i)

    arr = []
    if UVMConfigDb.get(self, "", "num_slaves", arr) is True:
        self.num_slaves = arr[0]

    for i in range(self.num_slaves):
        inst_name = sv.sformatf("slaves[%0d]", i)
        self.slaves.append(ubus_slave_agent.type_id.create(inst_name, self))

Line 1 Declare the build_phase() function.

Line 3 Call super.build_phase(). This guarantees that the configuration fields (num_masters, num_slaves, and has_bus_monitor) are updated per any resource settings.

Line 14 - Line 15 Create the bus monitor if the has_bus_monitor control field is set to 1. The create function is used for creation.

Line 17 - Line 26 The master’s dynamic array is sized per the num_masters control field, which is read from the resource database. This allows the for loop to populate the dynamic array according to the num_masters value. The instance name that is used for the master agent instance is built using sv.sformatf so the instance names match the dynamic-array identifiers exactly. The iterator of the for loop is also used to set a resource value for the master_id properties of the master agent and all its children (through the use of the asterisk). This defines which request-grant pair is driven by the master agent.

Line 28 - Line 34 As in the master-agent creation code above, this code creates the slave agents using num_slaves but does not set a resource for the slave agent.

7.7 UBus Master Agent ¶

The ubus_master_agent (shown in Figure 43) and ubus_slave_agent are structured identically; the only difference is the protocol-specific function of its subcomponents.

The UBus master agent contains up to three subcomponents: the sequencer, driver, and monitor. By default, all three are created. However, the configuration can specify the agent as passive (is_active=UVM_PASSIVE), which disables the creation of the sequencer and driver. The ubus_master_agent is derived from uvm_agent.

The build_phase() function of the ubus_master_agent is specified to create the driver, sequencer, and the monitor. The is_active property controls whether the driver and sequencer are created.

Example: ubus_master_agent.sv

def build_phase(self, phase):
    super().build_phase(phase)
    self.monitor = ubus_master_monitor.type_id.create("monitor", self)
    if self.get_is_active() == UVM_ACTIVE:
        self.driver = ubus_master_driver.type_id.create("driver", self)
        self.sequencer = ubus_master_sequencer.type_id.create("sequencer",
                self)

    arr = []
    if UVMConfigDb.get(self, "*", "master_id", arr) is True:
        self.master_id = arr[0]
    UVMConfigDb.set(self, "", "master_id", self.master_id)

def connect_phase(self, phase):
    if self.get_is_active() == UVM_ACTIVE:
        self.driver.seq_item_port.connect(self.sequencer.seq_item_export)
        self.sequencer.addr_ph_port.connect(self.monitor.addr_ph_imp)

Line 1 Declare the build_phase() function.

Line 2 Call super.build_phase(). This guarantees that the configuration field (is_active) is updated per any overrides.

Line 3 Create the monitor. The monitor is always created. Creation is not conditional on a control field.

Line 4 - Line 7 Create the sequencer and driver if the is_active control field is set to UVM_ACTIVE. The create() function is used for creation. Note the use of the base uvm_sequencer.

Line 14 Declare the connect_phase() function.

Line 15 - Line 17 Since the driver expects transactions from the sequencer, the interfaces in both components should be connected using the connect() function. The agent (which creates the monitor, sequencer, and driver) is responsible for connecting the interfaces of its children.

7.8 UBus Master Sequencer ¶

This component controls the flow of sequence items to the driver (see Figure 44).

The sequencer controls which sequence items are provided to the driver. The uvm_sequencer base class will automatically read the sequence resource set for each specific run-time phase and start an instance of that sequence by default.

7.9 UBus Driver ¶

This component drives the UBus bus-signals interface by way of the xmi virtual interface property (see Figure 45). The ubus_master_driver fetches ubus_transfer transactions from the sequencer and processes them based on the physical-protocol definition. In the UBus example, the seq_item_port methods get_next_item() and item_done() are accessed to retrieve transactions from the sequencer.

The primary role of the driver is to drive (in a master) or respond (in a slave) on the UBus bus according to the signal-level protocol. This is done in the run_phase() task that is automatically invoked as part of UVM’s built-in simulation phasing. For the master driver, the core routine is summarized as follows:

async def get_and_drive(self):
    await RisingEdge(self.vif.sig_reset)
    while True:
        await RisingEdge(self.vif.sig_clock)
        req = []
        await self.seq_item_port.get_next_item(req)
        rsp = req[0].clone()
        rsp.set_id_info(req[0])
        await self.drive_transfer(rsp)
        self.seq_item_port.item_done()
        self.seq_item_port.put_response(rsp)

Once the sig_reset signal is deasserted, the driver’s run task runs forever until stopped by having all run_phase objections dropped. You are encouraged to study the UBus driver source code to gain a deeper understanding of the implementation specific to the UBus protocol.

7.10 UBus Agent Monitor

The UBus monitor collects ubus_transfers seen on the UBus signal-level interface (see Figure 46). If the checks and coverage are present, those corresponding functions are performed as well.

The primary role of the UBus master monitor is to sample the activity on the UBus master interface and collect the ubus_transfer transactions that pertain to its parent master agent only. The transactions that are collected are provided to the external world by way of a TLM analysis port. The monitor performs this duty in the run task that is automatically invoked as part of simulation phasing. The run task may fork other processes and call other functions or tasks in performance of its duties. The exact implementation is protocol- and programmer-dependent, but the entry point, the run task, is the same for all components.

The monitor’s functionality is contained in an infinite loop defined with the run_phase() task. Once all of the run_phase objections were dropped, the run_phase() tasks finish, allowing other simulation phases to complete, and the simulation itself to end.

The checks are responsible for enforcing protocol-specific checks, and the coverage is responsible for collecting functional coverage from the collected ubus_transfers.

7.11 UBus Bus Monitor

The UBus bus monitor collects ubus_transfers seen on the UBus signal-level interface and emits status updates via a state transaction, indicating different activity on the bus. The UBus bus monitor has class checks and collects coverage if checks and coverage collection is enabled. The UBus bus monitor is instantiated within the UBus environment.

The ubus_env build_phase() function has a control field called has_bus_monitor, which determines whether the ubus_bus_monitor is created or not. The bus monitor will be created by default since the default value for this control field is 1. You can use the UVMConfigDb interface to override this value:

UVMConfigDb.set(self, "ubus0", "has_bus_monitor", 0)

Here, the ubus0 instance of ubus_env has its has_bus_monitor control field overridden to 0. Therefore, the ubus_bus_monitor in ubus0 will not be present. The build_phase() function for the ubus_env that uses the has_bus_monitor control field can be found in Section 7.6.

7.11.1 Collecting Transfers from the Bus

The UBus bus monitor populates the fields of ubus_transfer, including the master and slave, which indicate which master and slave are performing a transfer on the bus. These fields are required to ensure a slave responds to the appropriate address range when initiated by a master.

In the UBus protocol, each master on the bus has a dedicated request signal and a dedicated grant signal defined by the master agent’s ID. To determine which master is performing a transfer on the bus, the UBus bus monitor checks which grant line is asserted.

To keep the UBus bus monitor example simple, an assumption has been made that the nth master connects to the nth request and grant lines. For example, master[0] is connected to grant0, master[1] is connected to grant1, and so on. Therefore, when the UBus bus monitor sees grant0 is asserted, it assumes master[0] is performing the transfer on the bus.

To determine which slave should respond to the transfer on the bus, the UBus bus monitor needs to know the address range supported by each slave in the environment. The environment developer has created the user interface API, ubus_env::set_slave_address_map(), to set the address map for the slave as well as the bus monitor. The prototype for this function is:

set_slave_address_map(self, slave_name, min_addr, max_addr)

For each slave, call set_slave_address_map() with the minimum and maximum address values to which the slave should respond. This function sets the address map for the slave and provides information to the bus monitor about each slave and its address map.

Using the address map information for each slave and the address that is collected from the bus, the bus monitor determines which slave has responded to the transfer.

7.11.2 Number of Transfers

The bus monitor has a protected field property, num_transactions, which holds the number of transfers that were monitored on the bus.

7.11.3 Notifiers Emitted by the UBus Bus Monitor

The UBus bus monitor contains two analysis ports, which provide information on the different types of activity occurring on the UBus signal-level interface

state_port–This port provides a ubus_status object which contains an
enumerated bus_state property. The bus_state property reflects bus-state changes. For example, when the bus enters reset, the bus_state property is set to RST_START and the ubus_status object is written to the analysis port.
item_collected_port–This port provides the UBus
transfer that is collected from the signal interface after a transfer is complete. This collected transfer is written to the item_collect- ed_port analysis port.

NOTE–Any component provided by the appropriate TLM interfaces can attach to these TLM ports and listen to the information provided.

7.11.4 Checks and Coverage

The UBus bus monitor performs protocol-specific checks using class checks and collects functional coverage from the collected ubus_transfers.

The UVM field coverage_enable and checks_enable are used to control whether coverage and checks, respectively, will be performed or not. Refer to Section 4.10 for more information.

7.12 UBus Interface

The UBus interface is a named bundle of nets and variables such that the master agents, slave agents, and bus monitor can drive or monitor the signals in it. Any physical checks to be performed are placed in the interface. Refer to Section 4.10.

Assertions are added to perform physical checks.The ubus_env field intf_checks_enable controls whether these checks are performed. Refer to Section 4.10 for more information.

The code below is an example of a physical check for the UBus interface, which confirms a valid address is driven during the normal address phase. A concurrent assertion is added to the interface to perform the check and is labeled assertAddrUnknown. This assertion evaluates on every positive edge of sig_clock if has_checks is true. The has_checks bit is controlled by the intf_checks_enable field. If any bit of the address is found to be at an unknown value during the normal address phase, an error message is issued.

always @(posedge sig_clock)

begin

assertAddrUnknown:assert property ( disable iff(!has_checks)

(($onehot(sig_grant) |-> ! $isunknown(sig_addr))) else_{$error(“ERR_ADDR_XZn Address went to X or Z during Address Phase”);} end

8. UBus Specification ¶

8.1 Introduction ¶

8.1.1 Motivation¶

The motivation for the UBus specification is to provide an example of a simple bus standard for demonstration purposes and to illustrate the methodology required for a bus-based verification component. As such, the UBus specification is designed to demonstrate all of the important features of a typical modern bus standard while keeping complexity to a minimum.

8.1.2 Bus Overview¶

The UBus is a simple non-multiplexed, synchronous bus with no pipelining (to ensure simple drivers). The address bus is 16-bits wide and the data bus is byte-wide (so as to avoid alignment issues). Simple burst transfers are allowed and slaves are able to throttle data rates by inserting wait states.

The bus can have any number of masters and slaves (the number of masters is only limited by the arbitration implementation). Masters and slaves are collectively known as —bus agents”.

The transfer of data is split into three phases: Arbitration Phase, Address Phase, and Data Phase. Because no pipelining is allowed, these phases happen sequentially for each burst of data. The Arbitration and Address Phases each take exactly one clock cycle. The Data Phase may take one or more clock cycles.

8.2 Bus Description ¶

8.2.1 Bus Signals¶

The list of bus signals (not including arbitration signals) is shown in Table 15. All control signals are active high.

Table 15–Bus Signals

Signal name | Width (bits) | Driven by | Purpose                                                                                 |
:——         | :——          | :—–        | :——                                                                                     |
clock       | 1                | n/a           | Master clock                                                                                |
reset       | 1                | n/a           | Bus reset                                                                                   |
start       | 1                | arbiter       | This signal is high during the Arbitration Phase and low during the Address and Data Phases |
addr        | 16               | master        | R4C4                                                                                        |
size        | 2                | master        | R5C4                                                                                        |
read        | 1                | master        | R6C4                                                                                        |
write       | 1                | master        | R7C4                                                                                        |
bip         | 1                | master        | R7C4                                                                                        |
data        | 8                | master/slave  | R7C4                                                                                        |
wait        | 1                | slave         | R7C4                                                                                        |
error       | 1                | slave         | R7C4                                                                                        |

addr 16 master Address of first byte of a transfer

size 2 master Indicates how many bytes will be transfers:

00 => 1 byte 01 => 2 bytes 10 => 4 bytes 11 => 8 bytes

read 1 master This signal is high for read transfers (write must be low)

Table 15–Bus Signals (Continued)

Signal Name

**Width (bits) Driven By Purpose **

write 1 master This signal is high for write transfers (read must be low)

bip 1 master Burst In Progress–driven high by master during Data Phase for all bytes, except the last byte of the burst. This signal, when combined with wait and error, can be used by the arbiter to determine if the bus will start a new transfer in the next clock cycle

data 8 master/slave Data for reads and writes

wait 1 slave High if slave needs master to wait for completion of transfer

error 1 slave High if slave error condition applies to this transfer

8.2.2 Clocking¶

All bus agents operate synchronous to the rising edge of the clock signal with the exception of gnt signals (see Section 8.3).

8.2.3 Reset¶

The active high reset signal is synchronous to the rising edge of clock. reset shall be asserted during power up and shall remain asserted for a minimum of five rising edges of clock* after power and clock have stabilized. Thereafter, reset shall be de-asserted synchronous to a rising edge of clock.

reset may be asserted at any time during operation. In such cases, reset must be asserted for at least three clock cycles and must be both asserted and de-asserted synchronous to the rising edge of clock. The assertion of reset cancels any pending transfer at the first rising edge of clock where reset is asserted. Any bytes that have been transferred prior to assertion of reset are considered to have succeeded. Any byte that would have succeeded at the rising edge of clock where reset is first asserted is considered to have failed.

While reset is asserted, all agents should ignore all bus and arbitration signals. While reset is asserted, the arbiter should drive start and all gnt signals low. At the first rising edge of clock where reset is de-asserted, the arbiter should drive start high. Thereafter, the normal bus operation should occur.

8.3 Arbitration Phase ¶

Each UBus shall have a single, central arbiter to perform arbitration and certain other central control functions.

The Arbitration Phase always lasts for one clock cycle. During the Arbitration Phase, the arbiter shall drive the start signal high. At all other times, the arbiter should drive the start signal low. The start signal can therefore be used by slaves to synchronize themselves with the start of each transfer. The arbiter shall always drive start high in the cycle following the last cycle of each Data Phase or in the cycle following a —no operation” (NOP) Address Phase (see Section 8.4.1). The last cycle of a Data Phase is defined as a Data Phase cycle in which the error signal is high, or both the bip and wait signals are low.

Each master on the bus has a dedicated req signal and gnt signal. The arbiter samples all req signals at each falling edge of clock where start is asserted and asserts a single gnt signal based on an unspecified priority system. At all falling edges of clock where start is not asserted, the arbiter shall drive all gnt signals low.

Thus, a master can see assertion of its gnt signal not only as an indication that it has been granted the bus, but also as an indication that it must start an Address Phase. It is not necessary for the master to check the start signal before starting its Address Phase.

Once a master is granted the bus, it must drive a transaction onto the bus immediately. No other master is allowed to drive the bus until the current master has completed its transaction.

NOTE–Only the arbiter is allowed to drive a NOP transfer. This means a master must drive a real transfer if it is granted the bus. Therefore, masters should not request the bus unless they can guarantee they will be ready to do a real transfer.

Arbitration signals shall be active high and shall be named according to a convention whereby the first part of the name is the root signal name (req_ for the request signal; gnt_ for the grant signal) and the second part of the name is the logical name or number of the master. Although the arbitration signals form part of the UBus specification, they are not considered to be —bus” signals as they are not connected to all agents on the bus.

It is up to individual implementations to choose an appropriate arbitration system. Arbiters might allocate different priorities to each master or might choose randomly with each master having equal priority.

8.4 Address Phase ¶

The Address Phase lasts for a single clock cycle and always immediately follows the Arbitration Phase.

8.4.1 NOP Cycle¶

Where no master has requested the bus and the start signal is asserted at the falling edge of clock, no gnt signal is asserted at the start of the Address Phase and the arbiter itself is responsible for driving the bus to a —no operation” (NOP) state. It does this by driving the addr and size signals to all zeroes and both the read and write signals low. A NOP address phase has no associated data phase so the arbiter shall assert the start signal in the following clock cycle.

NOTE–This means the arbiter is connected to certain bus signals in addition to the arbitration signals and behaves as a —default master”.

8.4.2 Normal Address Phase¶

If, at the rising edge of clock, a master sees its gnt signal asserted, then it must drive a valid Address Phase in the following cycle. The master should also de-assert its req signal at this clock edge unless it has a further transfer pending.

During the Address Phase, the granted master should drive the addr and size signals to valid values and should drive either read or write (but not both) high.The address driven on addr represents the address of the first byte of a burst transfer. It is up to the slave to generate subsequent addresses during burst transfers.

The master shall only drive the addr, size, read, and write signals during the Address Phase. During the subsequent Data Phase, the master should not drive these signals.

8.5 Data Phase ¶

The Data Phase may last for one or more clock cycles. The Data Phase follows immediately after the Address Phase (and is immediately followed by the Arbitration Phase).

8.5.1 Write Transfer¶

The master shall drive the first byte of data onto the bus on the clock cycle after driving a write Address Phase. If, at the end of this clock cycle, the slave has asserted the wait signal, then the master shall continue to drive the same data byte for a further clock cycle. The data signal may only change at the end of a cycle where wait is not asserted. Thus, the slave can insert as many wait states as it requires. The master shall drive the bip signal high throughout the Data Phase until the point where the final byte of the transfer is driven onto the bus, at which point it shall be driven low.

At the end of the transfer (the end of the cycle where both bip and wait are low) the master shall cease to drive all bus signals.

8.5.2 Error during Write Transfer¶

The slave shall drive the error throughout the Data Phase. If a slave encounters an error condition at any point during the Data Phase of a write transfer, it may signal this by asserting the error signal. To signal an error condition, the slave must drive the error signal high while driving the wait signal low. This indicates to the master that the associated byte of the transfer failed–any previous bytes in the burst are considered to have succeeded; any subsequent bytes in the burst are abandoned. The assertion of error always terminates the Data Phase even if bip is asserted simultaneously.

8.5.3 Read Transfer¶

On the clock cycle after the master drives a read Address Phase, the slave can take one of two actions: drive the first byte of data onto the bus while driving the wait signal low or drive the wait signal high to indicate it is not yet ready to drive data. Each byte of data is latched only by the master at the end of a cycle where wait is low–thus the slave can insert as many wait states as is required. The master shall drive the bip signal high throughout the Data Phase until the point where the master is ready to receive the final byte of the transfer, at which point it shall be driven low.

At the end of the transfer (the end of the cycle where both bip and wait are low) the master shall cease to drive all bus signals.

8.5.4 Error during Read Transfer¶

The slave shall drive the error throughout the Data Phase. If a slave encounters an error condition at any point during a read transfer, it may signal this by asserting the error signal. To signal an error condition, the slave must drive the error signal high while driving the wait signal low. This indicates to the master that the associated byte of the transfer failed–any previous bytes in the burst are considered to have succeeded; any subsequent bytes in the burst are abandoned. The assertion of error always terminates the Data Phase even if bip is asserted simultaneously.

8.6 How Data is Driven ¶

Table 16 specifies how data is driven in the UBus specification.

Table 16–What Drives What When

**Signal Name Arbitration Phase Address Phase Data Phase **

start Driven to 1 by arbiter Driven to 0 by arbiter Driven to 0 by arbiter

addr Not driven Driven by master (or to 0 by arbiter for NOP)

8.7 Optional Pipelining Scheme ¶

Not driven

As previously stated, the UBus standard does not normally support pipelining. However, pipelining can optionally be implemented.

NOTE–All agents (including arbitration) on a bus must agree either to pipeline or not to pipeline. Mixing pipelined and non-pipelined agents on the same bus is not supported.

Because pipelining overlaps the Arbitration, Address, and Data Phases, two levels of pipelining are provided; i.e., there are a total of three transfers in progress at any one time.

NOTE–Pipelining results in different bus agents driving the same signals in consecutive clock cycles. As such, there is no period where the signal is not driven as part of a change of sequencers. As a result, care is necessary in the physical design of the bus to ensure that bus contention does not occur. A multiplexed approach will be required (in the form of either a ring or a star).

8.7.1 Pipelined Arbitration Phase¶

In a pipelined system, the Arbitration Phase is performed in parallel with the Address and Data Phases. Arbitration is carried out in every clock cycle regardless of whether this is necessary or not. This is because the arbiter cannot predict whether the next clock cycle will mark the start of a new Address Phase.

size Not driven Driven by master (or to 0 by arbiter

for NOP)

Not driven

read Not driven Driven by master (or to 0 by arbiter

for NOP)

Not driven

write Not driven Driven by master (or to 0 by arbiter

for NOP)

Not driven

bip Not driven Not driven Driven to 1 by master for all but

last byte of transfer

data Not driven Not driven Driven by master during writes. Driven by slave during reads in cycles where wait is low; other- wise, don’t care (may be driven to unknown state or not driven at all)

wait Not driven Not driven Driven by slave

error Not driven Not driven Driven by slave

The Arbiter asserts the start signal in the clock cycle after the end of each Data Phase as in the non-pipelined system. However, this start signal marks the start of all three Phases in parallel.

The end of a Data Phase can be recognized by either assertion of error or de-assertion of both bip and wait.

8.7.2 Pipelined Address Phase¶

A master that has its gnt signal asserted at the clock edge where a Data Phase completes is granted the Address Phase of the bus. It must immediately start driving an Address Phase. Unlike in the non-pipelined bus, where the Address Phase lasts a single clock cycle, the Address Phase in a pipelined bus lasts until the end of the next Data Phase.

Where no master requests the bus and, therefore, no master is granted the bus, the arbiter is responsible for driving NOP until the end of the next Data Phase.

8.7.3 Pipelined Data Phase¶

The Data Phase of a pipelined bus is similar to that of a non-pipelined bus. Where the arbiter drives a NOP for the preceding Address Phase, the master must drive error, bip, and wait low during the Data Phase (which will last for a single clock cycle in this case).

8.8 Example Timing Diagrams ¶

Figure 47 and Figure 48 show sample timing diagrams.

Timing diagram for ubus write operation. — Example Write Waveform¶

Timing diagram for ubus read operation. — Example Read Waveform¶