How to use TI DSP libraries and OpenCL on C7x DSP and C66x DSP (TDA4VM / BeagleBone AI-64)

To-do Projects
1

Hi BeagleBoard community,

I’m using the BeagleBone AI-64 with a custom Debian 12 minimal image (aarch64) and want to take advantage of the C7x DSP (TDA4VM SoC) and C66 DSP for signal processing.

Specifically:

  • I want to use TI DSP libraries (e.g., DSPLIB, FFTLIB)
  • I want to run OpenCL on the DSP to offload compute from the Cortex-A72

My setup:

  • BBAI-64 with Debian 12 minimal, booted from SD card
  • Kernel 5.x with remoteproc, rpmsg, etc.
  • I have extracted the Processor SDK and RTOS SDK contents

Could someone guide me on:

  • Whether TI’s OpenCL runtime (ti-opencl) can work with this Debian setup?
  • Where to place firmware blobs to boot DSP via remoteproc?
  • Whether TI OpenCL runtime can auto-boot the DSPs or I must do it manually?
  • Any success stories running OpenCL on the DSP (not GPU) on BBAI-64?

Any help or pointers to example setups would be great!

Thanks

2

It depends on you end game.

If you need hard deterministic control you will need to use the Ti product.
If you need process video, signals, and research use the Nvidia product.

CUDA is mature and user friendly, Ti is a manipulative and controlling. Did I have any luck getting the Ti product up, no, too many intentional voids in the information. So it comes down to your use case and available time.

If you do find a real working solution please post it, the AI64 board is rock solid and would be a good fit in some industrial applications.

1 Like
3

I would sooo love to do a machine that uses the image processing capabilities of this thing for targeting. Items on a conveyor belt, defects on plywood, roaches in the kitchen, so on.

If only there was just one solid guide out there of someone successfully doing something of this nature. Once one person figures it out, the river starts flowing. The board would have done a lot better if it was released with a full stack example showing it’s power. Been years, still nobody has done it.

Laser weed killing machine? :point_right: :point_left:

1 Like
4

The Jetson orin nano is the clear winner for that application. It leaves every SBC in the dust. They too have some issues, one is NOT very cool at all. They seem to be slow walking bringing back to market so I assume they found the issue and are fixing the silicon or the SDK.

I got tired of waiting for them and now use a big box with GPU. Nano is great for inference but you cannot train on it due to so little Ram.

Pretty sure the DSP works fine, I have seen Sitara chips in Fanuc controls, also sure if Fanuc has a problem with the code that Ti would assign a dedicated team to them. They are not about to let the public know where the “gold” is buried. Typical industry paradigm with this, overwhelming majority of makers/hobbists do not have a clue in the world they are getting played like fools on this. Typical ground cover is, “It is complicated”, “We are not here to teach you how to develop code”.

That is all “double talk” for if you pay us or a preferred partner we can bring your “dream” to reality…