Forums - Regarding 16bit Fixed Point mode on DSP

1 post / 0 new
or Register
Regarding 16bit Fixed Point mode on DSP
idpcom
Join Date: 21 Dec 20
Posts: 1
Posted: Sun, 2021-03-07 21:10
Top

Hi,

Because of quantization noise, the w8a8 (fixed point, weight 8 / activation 8) is not suitable for an image to image regression.

So I want to use fixed point of w8a16 or w16a16.

I succeeded in quantizing w8a16 according to the guide below.

https://developer.qualcomm.com/docs/snpe/tools.html#tools_snpe-dlc-quantize

However, in the implementation process, I found that there is only DSP_FIXED8_TF option for DSP in Runtime_t enum class

($SNPE_ROOT/include/zdl/DlSystem/DlEnums.hpp)

How can I implement w8a16?

 

In addition, I got the following information in the quantization process.

[INFO] Setting activation for layer: input_image:0 and buffer: input_image:0

[INFO] bw: 16, min: -0.065696, max: 0.999997, delta: 0.000016, offset: -4040.000000

Do I need to convert input data to 16bit to use it? (The input of the current model is Float32.)
 
  • Up0
  • Down0

Opinions expressed in the content posted here are the personal opinions of the original authors, and do not necessarily reflect those of Qualcomm Incorporated or its subsidiaries (“Qualcomm”). The content is provided for informational purposes only and is not meant to be an endorsement or representation by Qualcomm or any other party. This site may also provide links or references to non-Qualcomm sites and resources. Qualcomm makes no representations, warranties, or other commitments whatsoever about any non-Qualcomm sites or third-party resources that may be referenced, accessible from, or linked to this site.