Reduce MIPS Usage in a "complex" ADAU1701 2x4 DSP Project

Good morning :)

For educational purposes, I am trying to reproduce in SigmaStudio the functionality of a device based on the ADAU1701. The project runs at 48 kHz, and the DSP parameters are controlled in real time by an external MCU.

The signal flow is approximately:

2 × INPUT

each input:
→ signal level readout
→ volume control
→ mute
→ 5-band EQ
- PEQ
- low shelf
- high shelf
- direct biquad coefficient import

→ 2×4 routing matrix

4 × OUTPUT

each output:
→ HPF
→ LPF
→ 5-band EQ
- PEQ
- low shelf
- high shelf
- direct biquad coefficient import
→ volume control
→ delay
→ RMS compressor
→ mute
→ polarity inversion
→ signal level readout

The HPF and LPF can use slopes up to 48 dB/oct, which means up to four biquad sections for the HPF and four for the LPF on each output.

This gives a maximum of:

Input EQ: 2 × 5 = 10 biquads
Output EQ: 4 × 5 = 20 biquads
HPF + LPF: 4 × 8 = 32 biquads

Total: 62 biquads

My main problem is fitting all of this functionality within the ADAU1701 instruction budget.

I have already made several optimizations to reduce MIPS usage.

The volume controls use Gain (No Slew). I do not use separate Mute or Invert blocks, because I can implement these functions using the linear gain value:

+gain → normal polarity
-gain → inverted polarity
0 → mute

I also implemented the 2×4 routing matrix using a mixer instead of separate Mute and Add blocks, which resulted in a reduction in instruction usage.

For the EQ and HPF/LPF filters I use General 2nd Order in Double Precision. The filter coefficients are calculated by the MCU and updated in real time.

I tried replacing some filters with Single Precision versions, but at low frequencies and high Q values I observed noticeable problems with the audio signal. This is especially important for HPF/LPF filters that may operate around 25–80 Hz with narrow PEQ filters.

I also experimented with an RMS Compressor variant with Display. I noticed that it exposes additional values in trap.datrelated to its internal level detection and compression.

If one of these values represents the RMS level of the signal after compression, it could potentially allow me to remove the separate output level detector and save some instructions.

However, I would prefer to use trap.dat only as a last resort. I prefer solutions where parameters and addresses can be obtained through the standard files exported by SigmaStudio, such as PARAM.h.

The device I am trying to reproduce uses RMS metering. In my own final design, however, I would prefer Peak metering if it can be implemented with a reasonable instruction cost.

I am therefore looking for ways to further optimize this project without giving up Double Precision where it is really needed.

I am particularly interested in:

a more instruction-efficient way to implement a large number of DP biquads,

sharing computations between the compressor and level meter,

a cheaper way to implement HPF + LPF with slopes up to 48 dB/oct,

an instruction-efficient way to obtain Peak level readout,

any other common ADAU1701 optimization techniques that I may have overlooked.

Data RAM is not currently the main limitation — I can reduce the maximum delay length as needed.

I will attach the complete .dspproj project to the post.

Best regards

Bartek

2in_4out_diy_passfilter_reduce.dspproj

software

processors and dsp

Processors and Microcontrollers

adau1701