Speaker A
We are here at Linaro Connect in San Francisco, and who are you?
Overview of Cadence HiFi 3 DSP SDK for Android on HiKey 960 board, highlighting low power audio processing and developer tools.
Key Takeaways
- Cadence HiFi 3 DSP enables highly efficient, low-power audio processing on Android devices.
- The SDK abstracts DSP complexity with a simple API, making it accessible to a broad range of developers.
- The DSP’s capabilities extend beyond audio to general signal processing and AI applications.
- Open source contributions to AOSP facilitate integration and community development.
- Future enhancements will focus on advanced audio features like multi-mic input and noise cancellation.
Summary
- Introduction to the HiKey 960 board featuring the Hi3660 SoC with integrated Cadence HiFi 3 DSP.
- HiFi 3 DSP is optimized for low-power audio processing, enabling over 30 hours of music playback with minimal battery drain.
- The DSP operates independently with other SoC components powered off to save energy.
- Software architecture includes a HiFi driver in the Linux kernel and an Xtensa audio framework for inter-processor communication.
- Developers are provided with an audio framework and a simple app developer API with fewer than 10 calls for easy DSP integration.
- The HiFi 3 DSP supports generic signal processing tasks beyond audio, such as neural network processing.
- The DSP architecture is heterogeneous and multi-core, widely used in billions of devices across various SoCs.
- The SDK and integration code are open source and contributed to AOSP, with ongoing development and future AI-focused features.
- Plans include multi-microphone input support for far-field voice commands and active noise cancellation.
- The project is supported by a global team including members based in Dallas, San Jose, and Pune.
Chapters
- 00:00Introduction and HiKey 960 Board Overview
- 00:35Low Power Audio Playback Benefits
- 01:01Power Efficiency and DSP Operation
- 01:22Software Architecture and Inter-Processor Communication
- 01:47Developer Tools and Audio Framework
- 02:10App Developer API and Ease of Use
- 02:58Team and Development Locations
- 04:06DSP Capabilities Beyond Audio
- 04:51HiFi DSP Family and Compatibility
- 06:59Open Source Contributions and Future Plans
Full Transcript — Download SRT & Markdown
Speaker A
I am Raj Powar, and I work for Cadence. And what are you showing here with the 96Boards, right?
Speaker A
Yes, that's right. So, what we have here, as you may see, is a HiKey 960 board, and it's got an SoC on it. That particular SoC is called Hi3660, which you can see on the board. Integrated on that SoC is a HiFi 3 DSP from Cadence.
Speaker A
This is a very popular, widely used low-power DSP. What it is good at is running many audio processing algorithms, like MP3 decode, AC3, and many other such codecs, but it runs them at very low power.
Speaker A
The benefit to the end user is that they can play music without draining their battery.
Speaker A
How low power does it go?
Speaker A
Oh, I think we have tested where more than 30 hours of audio playing can be done without having any effect on the battery. So, everything else in the SoC is powered off. It's only the DSP that needs power.
Speaker A
Exactly, exactly. And how do you integrate that in all the Linux and Android environments?
Speaker A
So, that's what I mean. Basically, we have something called inter-processor communication. If you look at the software architecture, the kernel has a HiFi driver.
Speaker A
The applications at the top can talk to that HiFi driver, and the HiFi driver talks to the DSP using what we call the Xtensa audio framework.
Speaker A
We are also providing developers with an audio framework available on the DSP that should make their life easier from both development and integration perspectives.
Speaker A
Right now, it's playing some music here, and you have this kind of stuff going on here, like how to create an audio chain.
Speaker A
Yes, exactly. There are different kinds of developers. Some might be very aware of how to do DSP programming, while others might not want to spend time learning about it.
Speaker A
So, what we have done is abstracted the usage of the DSP through some APIs that we call the app developers API.
Speaker A
If you see here, the app developers API is also shown in the software architecture. People can just write to that interface, and there are very few, less than 10 APIs they have to learn.
Speaker A
If they know those APIs, they can leverage the power and other benefits of the HiFi 3 DSP.
Speaker A
What has been your involvement here with Linaro? Do you have a team?
Speaker A
Yes. Where are you based?
Speaker A
I am based in Dallas, Texas.
Speaker A
We have a team working on this. There is a set of people available in San Jose.
Speaker A
This is Niranjan Yadla working on it, and Sachin Ganekar heads up the Pune development organization. Underneath him, there are two gentlemen, Vijay Pawar, Ajay Homkar, and Girish, who are working very actively on these boards.
Speaker A
They provide open source access to use the DSP for audio, but can it be used for something else?
Speaker A
Exactly. Our HiFi 3 DSP is very good at performing math operations and other signal processing capabilities. We have also provided a generic interface where you can pass input buffers to the DSP, and it will process them and give you output buffers.
Speaker A
You can use that for neural net processing or any compute-intensive algorithms you want to implement on the HiFi 3 DSP.
Speaker A
It’s a heterogeneous multi-core processing architecture with a Cadence DSP that is on many different SoCs worldwide, not only on HiSilicon.
Speaker A
Yes, it is a very popular DSP that has been integrated into millions of devices.
Speaker A
Yes, billions of devices.
Speaker A
So, it’s actually been the same architecture of the DSP for many years?
Speaker A
There is a whole family of HiFi DSPs. It started with HiFi 2, then HiFi 2 EP, then HiFi 3, HiFi 4, and there is also a Fusion DSP optimized for IoT applications.
Speaker A
Everything you are doing here can be used for the whole family of DSPs?
Speaker A
Right now?
Speaker A
Right now only for this one?
Speaker A
No, our interface is generic. It is agnostic of the members of the family. There is nothing unique that makes it only for HiFi 3, but the SoC on this board right now is HiFi 3.
Speaker A
In theory, this work can be done for all current and future DSPs.
Speaker A
Exactly.
Speaker A
To get open source to work with it?
Speaker A
Yes, we are providing the basic glue logic that connects and shows how our i53 is connected to Android, and we are providing open source code. We are contributing to AOSP today.
Speaker A
How much work is to be done in the future, and what is the strategy?
Speaker A
There are a lot of possibilities. What you are seeing is fresh from the oven, if I may say so. There are many applications, particularly in the future AI space.
Speaker A
In particular, I would like to point out we are going to be working on providing multi-microphone input so people can process audio from an array of microphones.
Speaker A
You can do voice commands from far across the room, and it will still work.
Speaker A
You can do active noise cancellation and many other things. We want to provide a platform for developers so they can implement these easily on a HiFi 3 DSP.
Speaker A
I would like to acknowledge the support we have from our management: Draw Madan, who heads up the software group, and Larry Prezewara, who is our marketing director.
Topics:CadenceHiFi 3 DSPAndroidHiKey 960Hi3660low power audioDSP SDKopen sourceAOSPaudio processing
