Hi,
I'm confused about the data padding.
In the previous TDA2 version TIDL has all 4 padding in all directions and it's easy to understand.
According to the documents "ti_dl/docs/user_guide_html/md_tidl_fsg_io_tensors_format.html" , it indicates that input and output will have pad{T , L ,R ,B}.
But when I' m checking some value in custom layer example, it uses padH and padW.
Here's the first question:
Is padH = pad Top , pad Bot?
Is padW = pad Left , pad Right?
Also I check the value of some variables in the custom layer I made, founding that variables have different value when running quantools(TIDL Process with REF_ONLY FLOW) and inference phase.
For example :
In the quantools phase:
A inData has width and height with 128 ,96 respectively , and its padW and padH equal to 1.
InData pitch equal to 12771 (Channel Pitch) and 129(Line Pitch). (12771 = 129 x 99).
Why the Channel Pitch is 129 x 99?
In inference phase :
InData has {w , h , padW , padH } = {128, 96 ,0 ,0 }.
InData pitch equal to 12416 (Channel Pitch) and 128(Line Pitch). (12416 = 128 x 97).
Why the Channel Pitch is 128 x 97 not 128 x 96?
There is a same question about outData.
The outData dims should be 64x48.
outData channel pitch is 3840 and line pitch is 72.
3840 isn't divisible by 72.
Could you please explain how the padding method works in TIDL or some customer layer import setting I probably miss to cause this problem.
Thank you,
Kevin
