Using the Tool

In this section the basic operation of the tools provided by lib_audio_dsp is described.

This document takes the user through three scenarios, illustrated by way of the included example app_simple_audio_dsp_integration, which may be found in the examples directory in lib_audio_dsp.

These scenarios are:

  • Creating a pipeline

  • Tuning and simulating a pipeline

  • Deploying pipeline code onto the xcore.

The steps in this guide should be executed in a Jupyter Notebook.

Creating a Pipeline

A simple yet useful DSP pipeline that could be made is a bass and treble control with output limiter. In this design the product will stream real time audio boosting or suppressing the treble and bass and then limiting the output amplitude to protect the output device.

The DSP pipeline will perform the processes shown in Fig. 5.

../../../_images/bass_treble_limit.drawio.png

Fig. 5 The target pipeline

The first step is to create an instance of the Pipeline class. This is the top level class which will be used to create and tune the pipeline. On creation the number of inputs and sample rate must be specified.

from audio_dsp.design.pipeline import Pipeline

pipeline, inputs = Pipeline.begin(
    1,          # Number of pipeline inputs.
    fs=48000    # Sample rate.
)

The Pipeline object can now be used to add DSP stages. For high shelf and low shelf use Biquad and for the limiter use LimiterPeak. For a full list of available DSP stages, see the DSP Stages List.

from audio_dsp.design.pipeline import Pipeline
from audio_dsp.stages import *

p, inputs = Pipeline.begin(1, fs=48000)

# i is a list of pipeline inputs. "lowshelf" is a label for this instance of Biquad.
# The new variable x is the output of the lowshelf Biquad
x = p.stage(Biquad, inputs, "lowshelf")

# The output of lowshelf "x" is passed as the input to the
# highshelf. The variable x is reassigned to the outputs of the new Biquad.
x = p.stage(Biquad, x, "highshelf")

# Connect highshelf to the limiter. Labels are optional, however they are required
# if the stage will be tuned later.
x = p.stage(LimiterPeak, x)

# Finally connect to the output of the pipeline.
p.set_outputs(x)

p.draw()

Fig. 6 demonstrates the output of the Jupyter Notebook when the above snippet was executed. The Jupyter Notebook will illustrate the designed pipeline. For information on creating more complex pipeline topologies, see Designing Complex Pipelines.

../../../_images/pipeline_diagram.png

Fig. 6 Generated pipeline diagram

Tuning and simulating a pipeline

Each stage contains a number of designer methods which can be identified as they have the make_ prefix. These can be used to configure the stages. The stages also provide a plot_frequency_response() method which shows the magnitude and phase response of the stage with its current configuration Fig. 7. The two biquads created above will have a flat frequency response until they are tuned. The code below shows how to use the designer methods to convert them into the low shelf and high shelf that is desired. The individual stages are accessed using the labels that were assigned to them when the stage was added to the pipeline.

# Make a low shelf with a centre frequency of 200 Hz, q of 0.7 and gain of +6 dB
p["lowshelf"].make_lowshelf(200, 0.7, 6)
p["lowshelf"].plot_frequency_response()

# Make a high shelf with a centre frequency of 4000 Hz, q of 0.7 and gain of +6 dB
p["highshelf"].make_highshelf(4000, 0.7, 6)
p["highshelf"].plot_frequency_response()
../../../_images/frequency_response.png

Fig. 7 Frequency response of the biquads (low shelf left, high shelf right)

For this tutorial the default settings for the limiter will provide adequate performance.

Code Generation

With an initial pipeline complete, it is time to generate the xcore source code and run it on a device. The code can be generated using the generate_dsp_main() function.

from audio_dsp.design.pipeline import generate_dsp_main
generate_dsp_main(p)

The reference application should then provide instructions for compiling the application and running it on the target device.

Note

Application Note AN02014 discusses integrating a DSP pipeline into the XMOS USB Reference Design.

The generate_dsp_main() function will cause the tuned DSP pipeline to run on the xcore device, where it can be used to stream audio. The next step is to iterate on the design and tune it to perfection. One option is to repeat the steps described above, regenerating the code with new tuning values until the performance requirements are satisfied.

Designing Complex Pipelines

The audio dsp library is not limited to the simple linear pipelines shown above. Stages can scale to take an arbitrary number of inputs, and the outputs of each stage can be split and joined arbitrarily.

When creating a new DSP pipeline, the initialiser returns the pipeline input channels as an instance of StageOutputList, a list-like container of StageOutput. When adding a new stage to the pipeline, a StageOutputList is used to pass the stage inputs. The stage initialiser returns a new instance of StageOutputList containing its outputs.

To select specific channels from a StageOutputList to pass to another stage, standard Python indexing can be used. Channels from multiple instances of StageOutputList can be combined by using the + operator.

The below shows an example of how this could work with a pipeline with 7 inputs. The pipeline graph is shown in Fig. 8.

from audio_dsp.design.pipeline import Pipeline
from audio_dsp.stages import Biquad

# start with 7 input channels
p, inputs = Pipeline.begin(7, fs=48000)

# pass the first 2 inputs to a 2-channel Biquad
i0 = p.stage(Biquad, inputs[0:2])

# pass the third input (index 2) to a 1-channel biquad
i1 = p.stage(Biquad, inputs[2])

# pass the inputs at index 3, 5, and 6 to a 3 channel biquad
i2 = p.stage(Biquad, inputs[3, 5, 6])

# pass all of i0 and i1, as well as the first channel in i2
# to create a 4 channel biquad
i3 = p.stage(Biquad, i0 + i1 + i2[0])

# The pipeline output has 6 channels:
# - all four i3 channels
# - the 2nd and 3rd channel from i2
p.set_outputs(i3 + i2[1:])
../../../_images/7_chan_biquad_pipeline.svg

Fig. 8 A pipeline splitting and combining inputs.

In order to split a signal path, a Fork stage should be used. This takes a count parameter that specifies how many times to duplicate each input to the Fork. The code block below shows how the signal chain can be forked. The pipeline graph is shown in Fig. 9.

from audio_dsp.design.pipeline import Pipeline
from audio_dsp.stages import Fork

p, inputs = Pipeline.begin(1, fs=48000)

# fork the input to create a 2 channel signal
x = p.stage(Fork, inputs, count=2)

# fork again to create a 4 channel signal
x = p.stage(Fork, x, count=2)

# there are now 4 channels in the pipeline output
p.set_outputs(x)
../../../_images/fork_pipeline.svg

Fig. 9 A pipeline with Fork stages.

As the pipeline grows it may end up consuming more MIPS than are available on a single xcore thread. The pipeline design interface allows adding additional threads using the next_thread() method of the Pipeline instance. Each thread in the pipeline represents an xcore hardware thread. Do not add more threads than are available in your application. The maximum number of threads that should be used, if available, is five. This limitation is due to the architecture of the xcore processor. The pipeline graph is shown in Fig. 10.

# thread 0
i = p.stage(Biquad, i)

# thread 1
p.next_thread()
i = p.stage(Biquad, i)

# thread 2
p.next_thread()
i = p.stage(Biquad, i)
../../../_images/multi_thread_pipeline.svg

Fig. 10 A multi-threaded DSP pipeline.

When using multiple threads, the signal paths must cross the same number of threads to reach the output. Failure to do this will result in threads blocking one another, and a reduced pipeline throughput.

The pipeline below is not acceptable, as the first channel crosses 2 threads, whilst the second channel only crosses one. The pipeline graph is shown in Fig. 11.

p, inputs = Pipeline.begin(2, fs=48000)

# inputs[1] is not used on thread 0
x1 = p.stage(Biquad, inputs[0])

p.next_thread()

# inputs[1] first used on thread 1
x = p.stage(Biquad, x1 + inputs[1])

p.set_outputs(x)
../../../_images/thread_crossings_bad.svg

Fig. 11 A bad DSP pipeline with differing thread crossings. The first biquad input crosses 2 threads, the second biquad input only crosses 1 thread.

To ensure signal paths cross the same number of threads, Bypass Stages can be used, as shown in the example below. The pipeline graph is shown in Fig. 12.

p, inputs = Pipeline.begin(2, fs=48000)

# both inputs are not used on this thread
x1 = p.stage(Biquad, inputs[0])
x2 = p.stage(Bypass, inputs[1])

p.next_thread()

x = p.stage(Biquad, x1 + x2)

p.set_outputs(x)
../../../_images/thread_crossings_bypass.svg

Fig. 12 A good DSP pipeline with thread crossings using Bypass stages. Both biquad inputs cross 2 threads.

Parallel thread paths are also permitted. In the below case, the inputs are passed to separate threads before being used in a third. The pipeline graph is shown in Fig. 13.

p, inputs = Pipeline.begin(2, fs=48000)

x1 = p.stage(Biquad, inputs[0])
p.next_thread()

x2 = p.stage(Bypass, inputs[1])
p.next_thread()

# x1 and x2 have both crossed 1 thread already
x = p.stage(Biquad, x1 + x2)

p.set_outputs(x)
../../../_images/thread_crossings_parallel.svg

Fig. 13 A good DSP pipeline with parallel thread paths. Both biquad inputs cross 1 thread before being combined in a third.

DSP Pipeline JSON Format

For integrating with external tools, such as GUIs, the DSP pipeline can be exported to a JSON format. To define and enforce the JSON schema, pydantic models of each stage are used. Pydantic provides type hinting and JSON schema verification to ensure that DSP pipelines are valid.

To convert a pipeline to JSON, the following steps can be followed:

from audio_dsp.design.pipeline import Pipeline
from audio_dsp.design.parse_json import DspJson, pipeline_to_dspjson, make_pipeline
from pathlib import Path
from audio_dsp.stages import Biquad

p, inputs = Pipeline.begin(1, fs=48000)
x = p.stage(Biquad, inputs)
p.set_outputs(x)

# Generate the pydantic model of the pipeline
json_data = pipeline_to_dspjson(p)

# write the JSON data to a file
filepath = Path("pipeline.json")
filepath.write_text(json_data.model_dump_xdsp())

To read a pipeline from a JSON file, the following steps can be followed:

# read the JSON data back to a DSP pipeline
json_text = filepath.read_text()
json_data = DspJson.model_validate_json(json_text)
p = make_pipeline(json_data)

The JSON file can be edited in an external editor. If the edited JSON file is not valid, an exception will be raised during model_validate_json with a description of the error.