i.MX ProcessorsのトピックRe: iMX 8M Plus: NPU info and Arm Compute Library

iMX 8M Plus: NPU info and Arm Compute Library

NeuralBlue — Tue, 02 Mar 2021 22:28:59 GMT

Currently,

NXPs BSP is still based on the v20.02.1 of ARM Compute Library:
https://source.codeaurora.org/external/imx/arm-computelibrary-imx/log/?h=imx_5.4.70_2.3.0

There are several advantages to upgrade to Compute Library >20.05 (ideally v. > 21). One of these advantages are related to QASYMM8_SIGNED (alias int8) for certain quantized layers; infact CL versions <20.05 does not fully support it and use only uint8 (unsigned).

How to install newer versions of Compute Library on top of BSP?

Furtheremore are there technical info about NPU and haw it handles int8, uint8 and the relative accumulations int8xint8 and uint8xuint8?

Thanks,
S.

Re: iMX 8M Plus: NPU info and Arm Compute Library

Zhiming_Liu — Wed, 03 Mar 2021 02:59:55 GMT

I am trying to change version to 21

About the NPU detail,that is unpublic

Zhiming

Re: iMX 8M Plus: NPU info and Arm Compute Library

NeuralBlue — Sat, 06 Mar 2021 12:30:13 GMT

Hi Liu,
NPU details are not necessary if we could change version to at least 21.
Signed integer Int8 is our preferred chioce for quantized layers.
Any progress or update on this topic?
Best,
V.

Re: iMX 8M Plus: NPU info and Arm Compute Library

NeuralBlue — Sat, 29 May 2021 13:39:39 GMT

OK Zhiming,
thanks.

We're now under NDA with NXP. Please give details, because we continue to suffer this issue:
https://community.nxp.com/t5/i-MX-Processors/iMX8-MPlus-NNAPI-Delegate-on-NPU-has-different-accuracy/m-p/1284374#M174691

Thanks,
S.