Getting Ryzen AI Software

[hungryDodo]

Hi team,

I'm an individual developer interested in experimenting with the AMD Ryzen™ AI platform. I've been following the mlir-aie repository and came across the following information:

Supporting AMD Ryzen™ AI with AMD XDNA™/AIE-ML (AIE2) and AMD XDNA™ 2 (AIE2P): Install AMD Vitis™ AIE Essentials
Install Vitis™ AIE Essentials from Ryzen AI Software 1.3 Early Access. We will assume you use the installation directory, /tools/ryzen_ai-1.3.0/vitis_aie_essentials.

The link provided for the Ryzen AI Software 1.3 Early Access program seems to be for corporate or partner applications, and as an individual developer, I'm unable to submit a request.

Could you please provide guidance on how an individual developer can obtain the Vitis AIE Essentials software and the ryzen_ai-1.3.0 package? Are there any alternative download links or different access channels for the open-source community?

Your assistance would be greatly appreciated.

[jgmelber]

Hi! Is there a specific reason you would like RAI SW? You should be able to use the project with Peano (llvm-aie) as the single core AIE compiler.

[hungryDodo]

Hi there,

Thank you for your prompt and helpful reply!

I would like to provide some more context regarding my request. My specific reason for wanting the RAI SW is that I am attempting to port the Aries project to run on an HX370 device.

The original project targets a VCK190 and uses the Vitis toolchain, alongside a custom-compiled LLVM and mlir-aie. Based on the README for mlir-aie, I speculated that the RAI SW might be a drop-in replacement for the original Vitis toolchain, which would make it easier to adapt the project's Makefile.

Since I am new to the RAI SW stack, I'm still trying to understand the differences between these toolchains. I was also wondering if the compiler Peano could achieve the same results.

I've already installed Vitis 2024.02 on my HX370, and most of the programming_examples in mlir-aie compile and run successfully. However, the color_detect example that uses OpenCV fail, which I raised in a previous issue #2554. As you noted in your reply to #2554, support for the aie2p architecture is not provided in the traditional Vitis tool downloads like 2024.2.

This led me to believe that the RAI SW might be the key, as I assumed it would contain the appropriate aie2p backend support for the HX370.

I will definitely spend more time looking into how to use Peano as a viable alternative. I would be very grateful for any insights or analysis you could offer on this. Thank you for your time and expertise!

[hungryDodo]

Hi there,

Good news! I've been digging into this more and realized the compilation issue isn't a problem with the toolchain itself—it looks like the code is simply missing a kernel for the aie2p architecture, only for aie1 and aie2.

I'm going to take a shot at writing some kernel code, based on the examples in mlir-aie, to see if I can get things working.

I'll be sure to come back and update this issue with my findings, whether it works or I run into new challenges.

Thanks again for all the help and insights!