IWR6843AOP: Executing HWA-accelerated object detection control process on DSP?

Hello.

Aubrey Jones said:

I'm drawing this conclusion from the different DPU docs each saying the "HWA" version will run on the R4F or the DSP cores. Am I understanding that correctly?

Your understanding is mostly correct.  Just to clarify, the AOP version of the project only uses the HWA core as opposed to the using the HWA and DSP jointly, so you can use both as well as just the HWA. 

Aubrey Jones said:

(What am I trying to do? Send back the largest, densest pointcloud possible to the host over UART. After switching to external 2Mbps UART and compressing coordinates, my bottleneck now is simply inter-frame processing time in the AOP demo. I want to try moving some or all processing to the DSP, as it has a faster core clock.)

If you want to use the AOP and use the DSP with the HWA, you can simply use the ISK project and swap out the antenna_geometry.c file with the corresponding file in the AOP project.  You will also have to change the pinmuxing done in the ISK project to match the pinmuxing done in the AOP project as these devices are not pin-to-pin compatible.

Aubrey Jones said:

Which documentation would help me in understanding the software development model used for building/linking combined MSS+DSS programs?

Looking through the ISK project should help with understanding the building/linking/post-build steps to generate an image for an MSS+DSS program.  At a high level, you want to build the DSS project first, then the MSS, and building the MSS project will also generate a single multi-core image that you can run.

Sincerely,

Santosh

Thank you for the reply!

Santosh Krishnan said:

If you want to use the AOP and use the DSP with the HWA, you can simply use the ISK project and swap out the antenna_geometry.c file with the corresponding file in the AOP project. 

I had originally thought I could do this (in fact the antenna geometry file is the same, and there is a define of ISK vs AOP that determines which is used)... but is the ISK example actually using the HWA? Or is it using the DSP implementation for each of the Procs?

I'm looking at the document file `mmwave_sdk_03_06_02_00-LTS/packages/ti/datapath/dpc/dpu/aoaproc/docs/doxygen/html/index.html`.

It says "Data processing is split in two paths, ping and pong path, one detected object per path. In each path processing is split between HWA and local core and it is divided in four stages: 2D-FFT calculation by HWA, Doppler compensation by local core, 3D-FFT calculation by HWA and XYZ estimation by local core." At the top of the page, it says that it can run on either the R4F or the DSP, which means either could be the "local core", right? And the DSP is faster than the ARM, so it's more performant to run on the DSP?

But looking at the code in `out_of_the_box_6843_isk_dss`, the string "hwa" doesn't appear anywhere in the code. In the AOP example, it appears dozens of times. So it would seem that the ISK is using the actual DSP implementation of each of the DPC Procs, not the DSP-hosted HW accelerated path described in the document string I quoted above.

I thought it would be easy to experiment and profile performance of HWA vs DSP, but I guess a good question now is whether I'm even on the right path? I want maximum performance from CFAR and AOA calculations. I want to process the most possible points in range and doppler space without locking up, compress the coordinates, and stream them over UART. I don't care about power. I only care about processing the most points I can before the next frame interrupt. Is DSP + HWA even the right way? Or is pure DSP faster?

Hello.

Let me look into this and provide an update by the end of the day tomorrow.

Sincerely,

Santosh

thank you!

> DSP is much faster, so all the signal processing is done there.

To be clear, you're saying that the DSP alone is faster than DSP+HWA?

Thank you! I believe I am starting to understand.

I think my misunderstanding was around the HWA's performance. I had mistakenly thought it was also clocked at 600MHz. But looking at block diagrams from several different data sheets for products not containing the DSP, and from what you're saying, it would seem that it's actually clocked at a much slower speed.

Thank you very much for the insight you've offered!

Can you give any insight on the 2D AOA calculation? It seems this is a key part of the AOP demo which is missing from the ISK demo. How does the 2D AOA calculation differ from the standard AOA calculation?

Hello Aubrey.

This question has started to branch from the original question; please create a new post on the E2E forums and we will provide a response as soon as possible.

Sincerely,

Santosh