mlcommons tiny performance benchmark on EBAZ4205 (former bitcoin mining board built around Xilinx Zynq-7000 SoC)

ARCHIVED

Xilinx's DPU requires more LUT than EBAZ can provide. Accelarating the benchmark on the PS without that IP is more challenging.

Reproducing

The final executable is expected to run on the ARM CPU accelterated by the PL side of the chip. The compilation of the kernel is a step which is not included in the root makefile yet and should be done separately.

First, run make (GNU) within the root folder to:

• build a cross-compiling toochain for ARM target. We use builroot for that (available in buildroot/host/bin)
• build the mlcommons tiny benchmarking tool (available in bin/runner)

Now enter the buildroot folder and run make to build the linux kernel. Everything in output/images should be copied in a suitable sd card (TODO: doc), together with the ../br2/ebaz4205/uEnv.txt file AND a suitable bitstream file (more on this later).

I am powering the board connected via ethernet cable to a network that configures the board using DHCP. I then ssh into the board for running the target side of the benchmark. The serial UART is reserved for the host part of the eembc framework.

./deploy.sh is used to transfer a new build of the tool to the board. Run the executable with a suitable console device (i.e. the serial port connected to your host machine that runs the benchmark framework). I run it like

% ./runner-a2ed450 /dev/ttyPS0

On the other side, run the framework following their intructions which you can find https://github.com/eembc/energyrunner.

Gotchas

• this repo uses submodules. You should clone it with git clone git@github.com:jecoz/mlperf-tiny-ebaz4205.git --recurse-submodules. If you forget to do so, you can later initialize submodules issuing a git submodule update --init --recursive

Notes

From https://www.xilinx.com/html_docs/vitis_ai/1_3/qbq1603967854451.html

• On the software side, the platform needs to provide the XRT and ZOCL packages. The host application can use the XRT OpenCL™ API to control the kernel. Vitis AI Runtime can control the DPU with XRT. ZOCL is the kernel module that talks to acceleration kernels. It needs a device tree node which has to be added.

• https://embed-me.com/from-hdl-to-linux-userland/

• https://developer.xilinx.com/en/articles/implementing-ai-in-an-embedded-project.html

• https://github.com/Xilinx/Vitis-Tutorials/blob/master/Vitis_Platform_Creation/Introduction/02-Edge-AI-ZCU104/step1.md