Thursday, November 17, 2016

Writing Verification Testplan for Module Level Design



   Thanks to Abhilash and Akash for contribution
 Verification is an art. You may google this and find scores of engineers agreeing to this. Many of us may not be born artists. But we are engineers and we know science. We would devise systematic approach to verification. In verification, often our first task is to make a test plan.  This feels like having an empty canvass and not knowing where to start and what to do with colours.
I don’t know how to handle a canvass, but can make some progress with testplan. We will focus on developing a testplan for module/IP level verification. We will have some guidelines, some checkpoints that would help us stay on track. 
Writing testplan for module level (IP level) and full chip level are two different things. There may be overlap, but both needs different approach. Full Chip testplan is not the topic for subsequent text.
Coming to the point, we need to draw an outline for our testplan. Let’s create separate sections to cover different aspects. I would like to make following sections. Each section may have further subsection and sub-sub sections.

  1. Sanity
  2. Register and memory access
  3. Features
  4. Applications
  5. Performance

The segregation is a not mutually exclusive. There are lot of areas overlapping with each other. As a verification engineer, we should worry about not leaving a hole, overlaps are fine.

Sanity
The testcases listed here aims at basic functioning of the module. What constitutes ‘basic’ is contextual. These are the baby steps that when perfected would form base for running of other tests. It would be easy to illustrate with examples.
  • A single correct read and write transaction
  • A packet being correctly received or transmitted
The sanity testcases serve as base indicator about the health of the IP. These testcases would be generally used / run throughout the development process. Typically, we run sanity tests a) after bug fixes to ensure that the fix has not broken the basics b) periodically to test the database (RTL + TB) is compiling and running.
Sanity Tests are most directed, quick running tests. We don’t aim to cover too many features here.

Register and memory access
Most of the IPs have some configurations, memories and status registers. This section deals with them. The testcases include default value checks, read and write accesses, front door & back door read write combinations, memory depth checks, etc. With some experience, you would realise that this section can be copy-pasted from other modules along with tests.
It is always beneficial to list the register and memory access testcases in the testplan. These tests should be run before sanity tests. These tests would help uncover the bugs in configurations and status registers. These bugs are tedious to debug in sanity or other feature tests. 

Features
These are the very reasons for the IP or module to exist. These are the properties or functions that the IP needs to fulfil.  We need to list all the features mentioned in IP architecture/micro-architecture document and standard specifications. In this section we would be covering most of the configurations that the design supports along with status and interrupts. We need to include the register configurations, interrupts and status registers getting triggered functionally. We will be carefully listing everything that needs to be put to test. While listing the testcases, do not worry about implementation of tests. Think of it as somebody else’s job  😃 .For two reasons. 1) We should not restrict the thoughts by implementation constraints 2) We should add more details to the test plan about the procedure, configuration sequence, status register checks, etc. It should be easy for somebody else to implement the testplan i.e. write the testcase.

We will divide this section into two categories.
  • Simple features
  • Composite feature
Simple features: These are straight forward functionalities broken to the base level. We find these directly mentioned in the relevant documents. Typical examples of the statements are
  • Supports burst access up to 256 bytes
  • Supports 24 10/100M ports 
  •  Supports clock recovery from input stream
All above are just examples and actual list is going to be bigger, probably the biggest amongst all the sections.
Composite features: The functionality that constitutes of number of simple features can be termed as composite feature. It has multiple features involved yet can be visualized as single functionality. I like to call it alloy features. Let’s go over some examples:
Exclusive access – It involves exclusive read to address A, then an exclusive write to same address A. There can be other accesses happening in between ex-read and ex-write.
Packet Forwarding: This one needs proper forwarding table established by learning or configuration. We need a packet hitting the forwarding table to get forwarded to destined port.

If above two examples do not strike a chord, then here is vague one. Image a robot with functionality to travel in metro. It involves multiple tasks such a buying a ticket, climbing stairs, wait for right metro, boarding and disembarking. Having included individual scenarios in simple features, we should verify the all the features in conjunction with the other ones. We can club two or more features at a time. The composite feature forms base for our application tests. 

Application
These are the functions that the IP is going to execute on field. It is very very important to know and list application scenarios. As we being first person to verify the IP, can reduce the on-field failure risks to great extent. I have emphasized knowing-applications aspect in other blog Tight Rope Walker - A verification Scenario .
It is difficult to imagine all the end applications. So, we must seek to get the inputs from architects, marketing, software team and everybody related to the IP. The existing chips which have this IP (previous version) can also serve as an input.
This category of tests target multiple functions at a time. The goal is to create real-life scenarios as much as possible.
An application testcase for a switch: Getting packet from multiple interfaces with exponential traffic distribution. The destination can be single port or multiple ports (again depends on the application). Some packets being seen for the first time triggering source learning.
Another example, a core is working on custom multi-threading application. It is getting interrupted by DMA, slow peripheral, high speed peripherals. The core gets data from other cores, caches and main memory. The data may be needed from different pages causing page faults intermittently.
Having successfully tested application scenarios, we are sure of somethings that chip would be useful in labs, despite some missed bugs. As mentioned in Chip Design Verification: Test-plan/Coverage Plan the chip should not be DOA. 

Performance tests
These are not a special category of tests. These are partly feature partly application testcases, with more emphasis on advertised numbers. The separate section is created to mark up the importance. The performance parameters can be found in data sheet or some marketing material for the IP.
The examples are
Handling of incoming traffic of n Gbps, average read write latency of n-clks, video processing capacity at 50 fps, etc.
These tests can stress test the IP. These are performed towards the end when all the features are tested ok.

Error cases
This section lists the unhandled but possible error cases. The handled error cases fall under features sections. We can list unhandled error cases in this sections as the expectation is different as compared to features tests. Since the graceful handling mechanism is not specified in the design, the minimum expectation is not to hang on seeing these errors. The design should be able to come to normal working in turn.
The cases could be
-          Packet arrival rate greater than supported
-          Internal memory soft corruption of single and multiple bits
By the way, I have missed to write this section name in the list given in beginning. Smart verification engineers would have figured that out. That’s the bug in the spec.

Table of Contents
When we write down all the sections, I can imagine the testplan should look something like
1.Sanity
                1.1 basic read write
                1.2 …
2. Register and memory accesses
                2.1 Default value
                2.2 FD, BD
               
3. Feature
                3.1 Transmit
                                3.1.1 Data path
                                                3.1.1.1 small packets
                                                3.1.1.2 large packets
                                3.1.2 Control path
                3.2 Receive
3.2.1 Data path
                                                3.2.1.1 small packets
                                                3.2.1.2 large packets
3.2.2 Control path
                3.3 Processor
4.  Application
                4.1 Customer end installation
                4.2 CO installation
5. Performance
                5.1 data throughput
                5.2 Latency guarantees
6 Errors
                6.1 High traffic
                6.2 Memory errors

Do we have the required documents?
All our efforts are in-vain if we are not having complete information. Architecture documents are the primary source of information for verification. Often one document is not sufficient. The architecture document may be implicitly or explicitly referring to some standard documents, previous IP implementations, etc. We may refer design document to find out implementation issues. It is a good idea to read all the relevant documents and think through the implementation of IP. If we can visualize the complete implementation, we probably have sufficient set.

Self-Review
We would definitely go for review with designers, architects and peers, it is important that we do it once our self. Having written the test plan first hand, I find myself biased and running out of patience to go over the same testplan again. Here the check list comes handy.
ð        Check that all the numbers in the documents are put to test. Search only the numbers. e.g.  clk frequency of 125 MHz, 24 ports, latency of 16 clocks, 1000 Mbps, etc. Identify the testcase in the testplan that is covering it. The testcase can be in any section.
ð        All statements in the documents should be looked at suspiciously. We put a question mark after the statement and look for testcase in testplan that tests it. E.g. Document says, “Small packets are dropped”. Let me test it.
ð        Configuration register – I try to map each configuration register field with test case that verifies it.
ð        Status register – Some testcase should be reading and checking the status during or at the end of simulation. Testcases for configuration and status register checking does not include “Section 2 Register and memory access “
ð        Pin functionality – Ensure that we have tests for IP’s interface pin functions

With all necessary goals set properly, we will get ready to fire tests at design to break it to strengthen it.

1 comment:

  1. Some advanced checks that can be included in test plan are, based on IP type and requirement are :

    1. Reset sequence checks cold/warm etc.
    2. Clock checks for when PPM is supported.
    3. Power down state changes in IPs that support Low power states
    4. Rates changes in IPs which support post reset, configurable rate change
    5. Combinations of multiple interrupts
    6. Concurrency tests when multiple masters can drive data path
    7. Stress tests like overflow/underflow of buffers
    8. Coverage plan in terms of code and toggle coverage
    9. Assertion checks for protocol and assertion coverage
    10. Also NLP tests with PST coverage

    ReplyDelete