Does LCE support deployment on ARM32 Cortex-M7?

[Oslomayor]

Hi all,

I am urgently seeking to deploy the TFLite models converted using Larq Compute Engine (LCE) on an ARM32 device, specifically a Cortex-M7 CPU, the STM32F7 series MCU.

I have seen some related issues, such as #647, #688.
I have also noticed the "arm32" keyword in the LCE source code, kernels_arm32.h.

Does LCE support deployment on ARM32 Cortex-M7 processors?
Is there some relevant documentation or sample code for me to reference, or are there any other solutions that would allow me to deploy BNN TFLite models on the Cortex-M7?

[Tombana]

Hi @Oslomayor.
Plumerai (the company that created Larq Compute Engine) offers an inference engine for Cortex-M CPUs, but this is a paid product and not open source. You can contact Plumerai through this form https://plumerai.com/contact_us, and the documentation of the inference engine can be found at https://docs.plumerai.com/1.16/inference_engine/ .

[Oslomayor]

Hi @Tombana
Thank you for the information you provided. Plumerai's inference engine seems to support int8 and int16 quantized models, just as it states in the documentation: "Plumerai's inference engine provides fast execution of INT8 quantized or INT16 quantized deep learning models in tiny memory spaces." However, what I need is to deploy int1 models on Cortex-M CPUs.

[Tombana]

You're correct, support for 1-bit layers is not mentioned in the public documentation. However, 2D binary convolutions are supported with optimized implementations for Cortex-M4 and Cortex-M7.