Flexible silicon: GUI-programmable audio processors (Part 2)

Part One of this two-part series introduced the concept of flexible silicon—signal processors that are programmable in environments that express functional and parametric behavior in application-relevant terms (Reference 1). Interacting with the signal processor at this elevated level of abstraction frees the OEM designer from many implementation details that regard signal streams in terms of instantaneous voltages or individual digitally encoded samples. Instead, the designer can focus on issues of functional implementation and performance expressed in the same terms as the system-design goals.

To make this distinction more concrete, this study has concentrated on two audio-signal processors—the XS-125-4 Class D power-amplifier module from D2Audio, which was the primary focus in Part One, and the AD1940 SigmaDSP IC from Analog Devices, which is the primary focus of this installment. In both cases, the devices must accommodate the fact that, at their most basic level of implementation, signal processors manipulate audio signals as time-domain phenomena, yet, at the application level, people prefer to consider audio signals in terms of their spectra and signal envelopes. Designers often leave this bridge between the time-domain and application-level models of audio signals to the system's user interface. The fact that the processors bridge the two concepts results in familiar-looking GUI presentations and could simplify implementation of the system interface.

Though the programming environments that accompany these flexible signal processors can greatly speed parts of your design cycle, your design methods may need to exhibit some flexibility, as well, particularly in design verification. Your familiar schematic-capture and simulation tools don't have these devices ready and waiting as drop-in blocks, macros, or behavioral models. Instead, your verification efforts will likely depend to a great extent on direct measurements. But, given the nature of the devices and the means of control they offer, the most meaningful measurements will be those that are in application-relevant terms, as well. Common instrumentation, such as logic analyzers, oscilloscopes, and voltmeters—though helpful for signal tracing—are less helpful for parametric verification in this environment than are more sophisticated signal analyzers. Your verification success, therefore, may critically depend on a solid familiarity with your measurement equipment and your understanding of its limitations (see sidebar "Flying on instruments").

The Analog Devices AD1940 is a flexible, general-purpose, GUI-programmable audio processor that operates ahead of a system's signal-distribution facilities. Bob Adams and his SigmaDSP team developed the family of parts to which the 1940 belongs. To develop those devices, Adams and the team exploited a custom DSP core, which they designed and optimized for audio-signal processing. As a general-purpose audio processor, the 1940 cannot do with a fixed arrangement of processing blocks, as was the case with the XS-125-4. Instead, its GUI programming-and-control environment, SigmaStudio, presents a drawing surface reminiscent of familiar schematic-capture tools, with a library of 61 processing cells in nine categories from which you can choose. Two additional categories provide system resources, such as hierarchal structures and simulation stimuli and probes. They also include signal sources, such as a tone, pink noise, white noise, and beep generators.

The simulation stimuli and probe cells enable real-time interactive spectral simulation. SigmaStudio allows you to develop signal-flow block diagrams offline without attaching the development hardware. The diagram that Figure 1 depicts, for example, is part of a tutorial exercise I completed after dinner one night at the dining-room table. The next morning, I connected my laptop to the AD1940 evaluation board through a USB port, downloaded the file, and ran test sweeps from the AP SYS-2722 through the system. A comparison of the plots in figures 2a and 2b indicates that the SigmaStudio simulation properly accounts for the equalizers' interband interactions and predicts the system response that the SYS-2722 measures. The straight-line plot falling along the –6-dB line in Figure 2b derives from output channels 2 and 3: signals that the equalizers do not process. These curves, which indicate a 0.4-dB rise in the lowest decade and a 0.35-dB drop-off in the top octave, indicate that similar behaviors, which appear on the Channel 0 and Channel 1 outputs, are not due to the equalizers' deviation from their predicted behavior.

The constant 6-dB difference that figures 2a and 2b depict derives from the fact that Figure 2a's plot simulates the equalizer's transfer function—a relative measure—whereas Figure 2b's is an absolute measure of the system's performance with reference to 0 dBFS (decibels full-scale). Though it is reasonable to set an equalizer to provide, say, 6 dB of gain at some frequency, a digital-audio system cannot produce an output of 6 dBFS: Unlike analog systems, digital-audio systems have no linear headroom above full-scale. To prevent the system from clipping, therefore, this measurement requires a stimulus signal from the generator no greater than –6 dBFS. In practical systems, inputs can reach 0 dBFS, so a real implementation of this circuit would need to include volume-control cells to manage the signal amplitude in advance of the equalizer. Analyze your design's gain structure to ensure that you can't clip an internal node. Using an equalizer to boost frequencies complicates the issue, whereas cutting is always safe with respect to clipping hazards. Also note that, unlike observations you might make of signals in the time domain, spectral plots are graceful at the onset of clipping. Raising the input level, say, 2 dB, results in spectra with peaks that appear to have lower Qs than the equalizer's settings indicate, rather than flattop peaks. A test for your system's gain-structure plan would include a THD (total-harmonic-distortion) measurement with a 0-dBFS sweep with spectral shaping cells turned up as far as your application allows. A somewhat more sophisticated approach is to build adaptable structures that exploit the AD1940's dynamics-management cells. Consider the trade-offs when deciding between upstream and downstream sensing.

The measurement of this circuit employed a POF (plastic-optical-fiber) link from the AP SYS-2722's digital generator to the AD1940 evaluation board's Toslink SPDIF input. The analyzer took the device under test's outputs in pairs from 1/8-in. TRS jacks through unbalanced, shielded leads. The AD1940's signal I/O resides strictly in the digital domain. An AD1939 codec provides the evaluation board's analog-I/O facilities; the codec adds greatly to the evaluation board's versatility. Less apparent, however, is the I/O flexibility inherent in the AD1940. The audio processor can receive or transmit audio data in two-channel I²S, left- or right-justified formats, or eight- or 16-channel TDM (time-division-multiplexed) streams.

The AD1940 evaluation hardware and SigmaStudio software form a seamless, reasonably responsive operating environment with a mostly well-behaved user interface. However, precisely setting virtual rotary controls is more difficult than similar controls on other platforms. The wiring paths are not user-adjustable, and symmetrical-cell arrangements can often result in overlapping paths, which, though mostly harmless, are difficult to read and can slow visual verification that the design agrees with the designer's intent. Barring any change in the SigmaStudio software, you can mitigate this effect by imposing spatial offsets between cells in your project work space.

The current SigmaStudio documentation is not up to the quality of the software itself. The library documentation's table of contents, for example, omits some items that are actually documented, though the search facility can usually find those items in short order if you know how to search for them. Explanations of many cells avoid many of the less obvious cell features. Although you can generally ignore this documentary weakness when using cells that implement familiar functions, it can slow down your project development when you need to invoke a less familiar function. The hardware documentation is likewise overly terse and lacks both a map and a narrative explanation of the board's layout and features. The SigmaDSP team is evidently aware of the need for better documentation and is working to soon upgrade it.

Author Information

You can reach Technical Editor Joshua Israelsohn at jisraelsohn@edn.com.