AN02014: Integrating a Generated Audio DSP Pipeline into a USB Audio Application£££index.html#an02014-integrating-a-generated-audio-dsp-pipeline-into-a-usb-audio-application

AN02014: Integrating a Generated Audio DSP Pipeline into a USB Audio Application$$$Introduction£££index.html#introduction

This application note explains how to integrate custom DSP into the XMOS USB Audio Reference Design (sw_usb_audio) by utilising the XMOS Audio DSP solution (lib_audio_dsp). An application showcasing this integration accompanies this document and the steps taken to create that application are described below.

The XMOS audio DSP solution offers a low-code approach to designing custom audio DSP through a Python library, which generates a multithreaded, pipelined DSP process for the xcore. The XMOS USB Audio Reference Design serves as a highly configurable audio IO platform. By following the process outlined in this document, it is possible to create highly specific audio DSP applications with less than 100 lines of additional embedded source code.

Both of the repositories discussed above contain detailed documentation on their use which should be consulted when modifying this application to any specific product. The scope of this note is limited to a single configuration of the XMOS USB Audio Reference Design and a simple DSP pipeline. It explains the steps required to add DSP to any configuration of the XMOS USB Audio Reference Design.

This app note is part one of two that discusses adding DSP to a USB application. Part two, AN02015, covers adding run-time control and includes DSP pipeline that showcases more of the DSP library’s features.

AN02014: Integrating a Generated Audio DSP Pipeline into a USB Audio Application$$$Getting Started£££index.html#getting-started

AN02014: Integrating a Generated Audio DSP Pipeline into a USB Audio Application$$$Getting Started$$$Running the example£££index.html#running-the-example

First, connect the XK-AUDIO-316-MC-AB to your computer with both the “DEBUG” and “USB DEVICE” Micro-USB ports as shown in Fig. 1.

Once connected follow these steps:

1. Open a terminal and activate the XTC enviroment. Optionally, create a Python virtual environment and activate it.

2. Get the source code for this app note from https://www.xmos.com/application-notes/

3. Navigate to the root directory of this app note and install the Python requirements:

   pip install -Ur requirements.txt
   

4. Start the Jupyter notebook from the app_dsp_and_usb directory. Jupyter Notebook is an interactive Python editor which was installed via the pip command in the previous step.

   cd app_dsp_and_usb
   jupyter notebook
   

5. If this does not automatically open a browser window, then copy the url from the output of jupyter that starts with http://127.0.0.1 and navigate to it in your web browser.

6. Open dsp.ipynb on the web interface by double clicking on the file name.

7. Execute all the cells in the notebook by selecting “Run all cells” from the “Run” menu.

This final step will display a diagram that represents the provided simple DSP pipeline. It will then generate the xcore source code, build the 8-channel application, and run it on the connected device. The device will appear on the connected computer as an 8-input, 8-output USB audio device named “XMOS xCORE.ai MC (UAC2.0)”. The provided DSP design simply reduces the input signal by 6 dB in both directions. When audio is played through the USB device, the signal will be processed by the DSP and then output through the 3.5 mm jacks on the board.

The DSP pipeline in dsp.ipynb provides a template that can be adapted to other needs and can form the base for the user’s own specific application requirements. The notebook can be updated and rerun to try different designs and iterate rapidly to find the best solution.

AN02014: Integrating a Generated Audio DSP Pipeline into a USB Audio Application$$$Application Overview£££index.html#application-overview

The XMOS USB Audio 2.0 Reference Design, sw_usb_audio, offers a versatile infrastructure for transmitting audio between various audio interfaces, including USB, I²S, ADAT, and SPDIF. It’s highly adaptable, supporting diverse combinations of input and output interfaces. Although the given configurations will transfer audio between interfaces without modification (excluding its built-in mixer module), it does offer callback hooks for application code to intercept the data flow. The pertinent callback, UserBufferManagement, is invoked once per sample period with the most recent data from each interface.

The core functionality of sw_usb_audio is provided by lib_xua; this is also the case for the application associated with this note. lib_xua provides the ability to configure which features are enabled and which tiles different features run on. The ports available in the end product will significantly effect the choice of tile for each feature. Consult the sw_usb_audio documents for more details. For this example application the chosen values for XUD_TILE and AUDIO_IO_TILE are 0 and 1 respectively. This leads to the software structure shown in Fig. 2.

The UserBufferManagement callback is executed on the audiohub thread on tile 1. This is the only thread in use on that tile, so there are 7 threads left for new functionality. It is important to understand the thread usage of the tile that will execute the DSP in order to know how many threads of DSP the pipeline design may use.

To reduce the complexity of this application, there is one provided build configuration and support for a single board. This has facilitated the removal of most files present in sw_usb_audio. However, the correlation between the remaining source files and those found in sw_usb_audio should be evident. All changes discussed in this document can also be applied to a full sw_usb_audio based design to integrate the generated pipeline.

This application has been configured to operate at a single sample rate of 48 kHz (using lib_xua configuration macros). The DSP pipeline must be designed for a fixed sample rate to allow tuning parameters to be determined. lib_audio_dsp supports rates other than 48 kHz, but the rate set when designing the DSP must align with the rate set for the rest of the application to get the expected performance.

AN02014: Integrating a Generated Audio DSP Pipeline into a USB Audio Application$$$Implementing UserBufferManagement£££index.html#implementing-userbuffermanagement

This application note includes a modified version of the UserBufferManagement function normally found in sw_usb_audio. It has the following prototype:

void UserBufferManagement(unsigned* sampsFromUsbToAudio, unsigned* sampsFromAudioToUsb);

The two arguments contain the input samples and must be filled with the output samples. The first argument takes channels from the USB host as an input and outputs to the other interfaces. The second argument outputs to the USB host and takes channels from the other audio interfaces as an input.

To pass audio through the DSP pipeline two functions must be called. These are defined in adsp_pipeline.h from lib_audio_dsp:

static inline void adsp_pipeline_source(adsp_pipeline_t *adsp, int32_t **data);
static inline void adsp_pipeline_sink(adsp_pipeline_t *adsp, int32_t **data);

The first passes samples to the pipeline, and the second reads processed samples from the pipeline. Both of these use chanends stored within an instance of adsp_pipeline_t that must be initialised by functions from the generated pipeline. Both adsp_pipeline_source and adsp_pipeline_sink block on a chanend until the DSP threads are available to process the sample exchange. This can lead to issues if the generated DSP cannot meet the real time requirements of the system.

The data parameter expects an array containing blocks of samples for each input and output channel. For this application the block size will be 1; therefore, we can construct both data parameters by initialising a new array of pointers that reference the correct elements in sampsFromUsbToAudio and sampsFromAudioToUsb. Fig. 3 shows how the dsp_input and dsp_output arrays are constructed in this application.

It is important to note that the size of sampsFromUsbToAudio and sampsFromAudioToUsb depend on the application configuration of lib_xua. In this application there are 8 USB OUT channels and 8 ADC channels, totalling 16 DSP inputs. There are also 8 USB IN channels and 8 DAC channels, totalling 16 channels of DSP outputs. The provided application will adapt to different I²S and USB configurations but will need updating when other lib_xua interfaces are enabled.

It is also important to note that the channel indices in dsp_input and dsp_output will be used later when defining the DSP pipeline.

Once constructed, the dsp_input can be passed into the “adsp_pipeline_source” function, and the dsp_output can be passed to the “adsp_pipeline_sink” function. An example implementation of UserBufferManagement can be found in app_dsp.c from the application provided with this note.

AN02014: Integrating a Generated Audio DSP Pipeline into a USB Audio Application$$$Generating the DSP pipeline£££index.html#generating-the-dsp-pipeline

Adding DSP to the project requires an initial DSP design, which is best done in a Jupyter notebook. This requires Python and some Python packages, which the application specifies in requirements.txt. This application’s source directory contains the notebook used at app_dsp_and_usb/dsp.ipynb. The Jupyter documentation covers the details of creating, modifying, and executing a Jupyter notebook.

Upon opening dsp.ipynb, you will find a simple DSP design that processes the 16 pipeline inputs. The design of a DSP pipeline is covered thoroughly in the user guide associated with lib_audio_dsp. The indices of the pipeline inputs match the indices of the channels in the dsp_input array discussed in the Implementing UserBufferManagement. The outputs of the pipeline, set when pipeline.set_outputs() is called, have indices that align with the dsp_output array.

After defining the DSP pipeline the notebook will proceed to generate the xcore source code:

generate_dsp_main(pipeline, out_dir="src/generated_dsp")

This function takes the pipeline and generates the source code in the provided out_dir, relative to the parent folder of dsp.ipynb. Consequently, this action creates the following files:

app_dsp_and_usb/src/generated_dsp
├── adsp_generated_auto.c
├── adsp_generated_auto.h
└── adsp_instance_id_auto.h

To include these files in the build, the CMakeLists.txt has been updated. No change was required to add the C files as XCommon CMake will automatically find them. In order for the application to include the header files, APP_INCLUDES has been appended with src/generated_dsp (see app_dsp_and_usb/CMakelists.txt for this change).

With the generated files included in the build it is possible to start the DSP threads in the application. The following function is defined in app_dsp.c and called in user_main.h on tile 1.

void dsp_thread(void) {
    // Initialise the DSP instance and enter the generated DSP main function.
    // This will never return.
    m_dsp = adsp_auto_pipeline_init();
    adsp_auto_pipeline_main(m_dsp);
}

With these changes, the application can be run on the board. The supplied Jupyter notebook will automatically do this as the final step of execution.

The custom DSP application is now ready for the development of a more complex DSP pipeline, such as the example described in AN02015.

AN02014: Integrating a Generated Audio DSP Pipeline into a USB Audio Application$$$References£££index.html#references

• sw_usb_audio

• lib_audio_dsp

• lib_xua

• XTC

• XK-AUDIO-316-MC-AB

• XCommon CMake

• Jupyter

• Graphviz

AN02014: Integrating a Generated Audio DSP Pipeline into a USB Audio Application$$$Support£££index.html#support

For all support issues please visit http://www.xmos.com/support