elastic-ai.Experiment-Framework

A framework to support you in performing experiments with the elastic-ai.Hardware (eai.Hardware) and related setups.

Features

• Stream data to elastic-ai.Hardware
• Stream data from elastic-ai.Hardware
• Parse Vivado reports into python dictionaries
• Call custom functions on the elastic-ai.Hardware from python scripts
• Measure Power/Latency for different workloads
• Perform different types of inferences on the HW accelerators from the elastic-ai.Creator via python API
• Push multiple bitstreams with FPGA configurations to eai.Hardware flash chips
• Load specific bitstreams from flash to FPGA
• Communication via USB/UART
• Communication via TCP/Lightweight IP
• Parse headers with c structs to automatically generate corresponding python types and methods to construct them from received data
• Parse headers with c functions to automatically generate corresponding python stubs for remote procedure calls to the eai.Hardware
• Semi-automatic synthesis (info via eaixp synth --help) with
  • Vivado local
  • Vivado remote
  • Cached Synthesis

Quick Start

$ uvx --from "git+https://github.com/es-ude/elastic-ai.experiment-framework.git" eaixp --help

or

$ uv tool install "git+https://github.com/es-ude/elastic-ai.experiment-framework.git"
$ uv tool run eaixp --help

or

$ pip install "git+https://github.com/es-ude/elastic-ai.experiment-framework.git"
$ eaixp --help

Communicating with the elastic node via elasticai runtime

Ensure you have the following tools installed

• cmake
• gcc-arm-embedded-none-eabi (version 13 or higher)

Then you can start off with the example in this repository under example-firmware/env5. Build the firmware using

 $ cmake --preset env5_rev2_release
 $ cmake --build --preset env5_rev2_release

and copy to the device like so

$ cp build/env5_rev2_release\
  /experiments/env5_experiment_firmware.uf2\
  /var/Volume/RP2

Extend the remote control CLI

The remote control lives in elasticai.experiment_framework.remote_control. The CLI can be easily extended as follows:

import elasticai.experiment_framework.remote_control as rc
import click

@rc.main.command
@click.pass_obj
@click.argument("data", type=str)
def my_custom_command(obj, data):
  rc_handle = rc.RemoteControl(obj)
  my_cmd_id = 250
  result = rc_handle.send_command(my_cmd_id, data, len(data))
  print(result)

If you need more control, you can use the elasticai.experiment_framework.remote_control_protocol directly in your command.

Synthesis

To use the synthesis feature, you have to provide information on where and how to run the synthesis. You can do that explicitly from your python script. But in most cases this will be user specific, so it is better to store that information in environment variables. If you're using devenv, you can create a file called devenv.local.nix with content like this:

{pkgs, ...}:

{
  env = {
    SYNTH_HOST = "192.168.1.34";
    SYNTH_SSH_USER = "elasticai";
    SYNTH_TARGET = "env5";
    SYNTH_WORKING_DIR = "/home/ies/synthesis";
    SYNTH_KEY = "hw_testing";
  };
}

Make sure to gitignore that file so you don't accidentally push it. Also note that the exact variables may also depend on the concrete synthesis you are running.