Sunday, February 8, 2015

Designers should not verify their own code! REALLY?

Around 2 decades back, the demands from designs were relatively simple and the focus was on improving performance. Process nodes had longer life, power optimization wasn’t even discussed and time to market pressure was relatively less given that the end products enjoyed long life. In those days, it was the designer who would first design and later verify his own code usually using the same HDL. Over the years, as complexity accelerated, a new breed of engineers entered the scene, the DV engineers! The rational given was that there is a need for an independent pair of eyes to confirm if the design is meeting the intent! Verification was still sequential to the design in early days of directed verification. Soon, there was a need for constrained random verification (CRV) and additional techniques to contain the growing verification challenge. The test bench development now started in parallel to the design, improving the size & need of verification teams further. With non HDLs i.e. HVLs entering the scene the need for DV engineers was inevitable. All these years, the rational of having an additional pair of eyes continued to be heard to an extent that we have started believing that designers should not verify their own code.

In my last post I emphasized the need for collaboration wherein designers and verification engineers need to come together for faster verification closure. Neil Johnson recently concluded in his post on designers verifying their own code. My 2 cents to whether designers should not or shall I say do not verify their own code –

To start with, let’s look at what all involves verification? The figure adjoining is a summary of efforts spent in verification based on the study conducted by Wilson research in 2012 (commissioned by Mentor). Clubbing some of the activities it is clear that ~40% of the time is spent in Test planning, Test bench development and other activities. The rest ~60% of the effort is spent on Debug & creating + running tests. The DUT here can be an IP or SoC. 

When an IP is under development or SoC is getting integrated, the DV engineers would be involved into the 40% of the activities mentioned above. These are the tasks that actually fall in line with the statement of additional pair of eyes validating design intent. They need to understand the architecture of the DUT and come up with a verification plan, develop verification environment and hooks to monitor progress. At this level, the involvement of the design team starts with activities like test plan review, code coverage review, inputs to corner cases and tests of interest. 

So, once the design is alive on the testbench, do the designers just sit & watch the DV team validate the representation of spec developed by them? NO!

Debugging alone is a single major activity that consumes an equal amount or sometimes more efforts from the designers to root cause the bug. Apart from it, there is significant involvement of the design team during IP & SoC verification.

For IPs, CRV is a usual choice. The power of CRV lies in automating the test generation using the testbench. A little additional automation enables the designers to generate constrained tests themselves. Assertions are another very important aspect in IPs. With introduction of assertion synthesis tools, the designers work on segregating the generated points into assertions or coverage. For SoCs, apart from reuse of CRV, directed verification is an obvious choice. Introduction to new tools on graph based verification help designers to try out tests based on the test plan developed by the DV engineer. Apart from this, corner case analysis and usecase waveform reviews are another time consuming contributions put in by designers for verifying the DUT.

Coming back to the rational on having an independent pair of eyes verify the code, the implication was never that the designers shall not verify their own code. Infact there is no way for the DV team to do it in a disjoint fashion. Today the verification engineer himself is designing a highly sophisticated test bench that is actually equivalent to a designer’s code in complexity.  So it would be rather apt to say that it is the two designs striking each other to enable verification under the collaboration between design & verification teams!

What is your take on this? Drop a note below!