Hi!
I found a next problem during inference of my net with TIDL:
I have input shape 1x1x360x640 and certain amount of MaxPool layers.
There is one MaxPool layer which produces wrong output (while ONNXRuntime gives the right one).
Params of the layer:
Kernel: 2x2
Ceil mode: 0
Pads: 0, 0, 0, 0
Strides: 2, 2
Input shape: 1x48x45x80
Output shape: 1x48x22x40.
The difference between this "error" layer and other MaxPools is that the input shape % 2 != 0 (45 % 2 != 0)
Other layers is divisible by 2.
If I increase input shape to 1x1x640x640 - I got Input shape 1x48x80x80 and Output shape 1x48x40x40 - and outputs from ONNX and TIDL are the same.
Unfortunatelly I can't provide a neural net for example because of NDA but it seems that it gonna happened in other cases if shape % 2 != 0.
I will try to shrink net and provide a simple one for example.
Is this a bug you found before? Will it be fixed in next releases?
