renesas-rz/rzv_drp-ai_tvm

Extension package of Apache TVM (Machine Learning Compiler) for Renesas DRP-AI accelerators powered by Edgecortix MERA(TM) Based Apache TVM version: v0.11.1

Extension package of TVM Deep Learning Compiler for Renesas DRP-AI accelerators powered by EdgeCortix MERA™ (DRP-AI TVM)

TVM Documentation | TVM Community | TVM github |

DRP-AI TVM ¹ is Machine Learning Compiler plugin for Apache TVM with AI accelerator DRP-AI provided by Renesas Electronics Corporation.

License

(C) Copyright EdgeCortix, Inc. 2022
(C) Copyright Renesas Electronics Corporation 2022
Contributors Licensed under an Apache-2.0 license.

Supported Embedded Platforms

• Renesas RZ/V2L Evaluation Board Kit (How to get)
• Renesas RZ/V2M Evaluation Board Kit (How to get)
• Renesas RZ/V2MA Evaluation Board Kit (How to get)
• Renesas RZ/V2H Evaluation Board Kit

Introduction

Overview

This compiler stack is an extension of the DRP-AI Translator to the TVM backend. CPU and DRP-AI can work together for the inference processing of the AI models.

File Configuration

Directory	Details
tutorials	Sample compile script
apps	Sample inference application on the target board
setup	Setup scripts for building a TVM environment
obj	Pre-build runtime binaries
docs	Documents, i.e., Model list and API list
img	Image files used in this document
tvm	TVM repository from GitHub
3rdparty	3rd party tools
how-to	Sample to solve specific problems, i.e., How to run validation between x86 and DRP-AI
pruning	Sample scripts to prune the models with DRP-AI Extension Package

Installation

• Installing DRP-AI TVM (RZ/V2L, RZ/V2M, RZ/V2MA)
• Installing DRP-AI TVM (RZ/V2H)

Deploy AI models on DRP-AI

Video

Following video shows brief tutorial for how to deploy ONNX model on RZ/V series board.
RZ/V DRP-AI TVM Tutorial - How to Run ONNX Model (YouTube)

Overview

To deploy the AI model to DRP-AI on the target board, you need to compile the model with DRP-AI TVM¹ to generate Runtime Model Data (Compile).
SDK generated from RZ/V Linux Package and DRP-AI Support Package is required to compile the model.

After compiled the model, you need to copy the file to the target board (Deploy).
You also need to copy the C++ inference application and DRP-AI TVM¹ Runtime Library to run the AI model inference.

RZ/V2L, RZ/V2M, RZ/V2MA

RZ/V2H

Following pages show the example to compile the ResNet18 model and run it on the target board.

[Option] Prune the model with DRP-AI Extension Package

• What is Pruning and Quantization?
• Installing DRP-AI Extention Package
• Pruning and retraining by DRP-AI Extension Package

Compile model with DRP-AI TVM¹

• Compiling model for DRP-AI TVM (RZ/V2L, RZ/V2M, RZ/V2MA)
• Compiling model for DRP-AI TVM (RZ/V2H)

Run inference on board

• Application Example for DRP-AI TVM (RZ/V2L, RZ/V2M, RZ/V2MA)
• Application Example for DRP-AI TVM (RZ/V2H)

Sample Application

To find more AI examples, please see How-to page.

Confirmed AI Model

If you want to know which models have been tested by Renesas, please refer to Model List.

Support

Error List

If error occurred at compile/runtime operation, please refer error list.

Tips

How-to page includes following explanation.

• profiler;
• validation between x86 and DRP-AI;
• etc.

Enquiries

If you have any questions, please contact Renesas Technical Support.

Footnotes

1. DRP-AI TVM is powered by EdgeCortix MERA™ Compiler Framework. ↩ ↩² ↩³ ↩⁴