Engineers at a leading IC design services company recently came up with a novel approach that improves verification efficiency by "layering" sequences of transactions to reach higher levels of abstraction. The approach is based on the Universal Verification Methodology (UVM), which defines a reusable UVM Verification Component (UVC) that, among other tasks, provides stimulus to the design under test using sequences.
This methodology was presented at the recent CDNLive India 2013 conference by Gaurav Jalan (right), Director of Engineering at SmartPlay Technologies, a global leader in VLSI design services. The paper, titled "UVM Sequence Layering in Configurable Memory Controller Verification," won a best paper award. Presentation slides are available online, as described at the end of this post. Jalan also writes a functional verification blog.
In an interview, Jalan described the problem they were trying to solve and the reasons for introducing a new methodology. He noted that his team was asked to verify a highly configurable DRAM memory controller with over 400 possible configurations. They were provided with raw RTL and a specification by the customer, and their task was to define a verification strategy using SystemVerilog and UVM, come up with an executable verification plan, develop a verification environment using Cadence verification IP (VIP), and "execute verification from scratch to closure."
Traditional and New Approach
The design under test (DUT) supported multiple protocols on both the processor interface and the DRAM interface. In a conventional verification approach, engineers would develop test sequences targeting each protocol on the processor side, and develop scoreboards for each protocol on the DRAM side. In this approach each configuration works as a separate DUT, and additional time is required to develop all the sequences for different configurations. The conventional approach also poses limitations when it comes to reusability, portability, scalability, and maintainability.
In the conventional approach, Jalan said, "the overall reusability index was low. To avoid this, we decided to move a level up - define tests and sequences based on the test plan at a higher, protocol-agnostic level that would be translated at run time into the desired protocols based on configuration. This concept is called sequence layering."
With sequence layering, engineers can develop test scenarios that are protocol independent. As shown below, the layering is achieved by a "layering agent" derived from uvm_agent. The layering agent has a monitor and a sequencer, and it "bundles" (translates) protocol-level transaction items into higher-level sequence items. A lower level sequencer translates the abstract level sequences into protocol-specific sequences, and pushes them forward to a driver.
"The way it [layering agent] works," Jalan said, "is that there is a high-level sequence item associated with the layering sequencer. It would connect to the sequence of the lower level protocols using the same mechanism as is used by the sequencer and driver in an uvm_agent. The lower level sequence would have only one sequence running as a ‘forever' thread. Inside this sequence we have a get_next_item similar to what we do in a uvm_driver. The item is received from the higher level sequencer. It is translated by the lower level sequencer and given to its driver. The response is then passed back to the layered sequencer, indicating that the lower level sequencer is ready for the next item."
The layering approach allows engineers to create generic sequences that are reusable across protocol interfaces. An example of a higher-level scenario might be, "Burst Read followed by a Burst Write in the same row." Protocol-specific coverage ensures that all scenarios are covered.
Faster Turnaround Time
The net result was a quicker turnaround time for each DUT configuration. Jalan said that the extra effort required to build the layering agent was 6%, but engineers saved almost 25% overall time later when developing sequences and tests. The overall project effort was 100 man-months. The effort included the use of Cadence ePlanner to develop an executable verification plan, Cadence eManager to track regressions and coverage, Cadence VIP for processor interfaces, and Denali models (from Cadence) for memory interfaces. The Cadence VIP team provided "excellent" support, Jalan said.
To view the CDNLive! India paper, click here. You will be asked for your Cadence log-in (or to sign up for one - a quick and easy process). You will then see a menu of proceedings from seven tracks. Choose track 5, "Functional Verification Track II," and scroll down to the paper titled "UVM Sequence Layering in Configurable Memory Controller" by Ghouse Syed Sherraj and Gaurav Jalan.
Richard Goering