Unintended clipping of values during output quantization

[nhma20]

Hi all,

I am trying to quantize the output of a FINN accelerator as follows:

self.qnt_output = QuantIdentity(
quant_type=QuantType.INT,
bit_width=8,
signed=False,
min_val=0.0,
max_val=255.0)

x = self.qnt_output(x)

I specify signed=False as well as min_val=0.0 and max_val=255 to use the whole range of 8 bits on the output. However, I observe that the output values are clipped between {0,127} such that e.g. 198 becomes 127. When I do not specify signed but keep min_val=0.0 and max_val=255, output values are divided by 2 such that e.g. 198 becomes 99.
The behavior is observed when simulating the stitched IP in Vivado.

Am I doing something wrong? I hope someone can help.

Thanks,
Nicolaj