Redefine symmetry as offset=-round((qmin + qmax) / 2)

[quic-kyunggeu]

Problem Statement

Given an unsigned asymmetric quantizer, setting quantizer.symmetric = True only turns it into an unsigned symmetric quantizer.
This is problematic for three reasons

• Blindly setting symmetric=True often results in catastrophic accuracy drop
• It's a bit counter-intuitive that int8 qdq and uint8 qdq produces radically different outputs
• It's vastly different from v1 behavior where the users could set symmetric=True without thinking about int8 vs. uint8

Proposal

I propose redefining symmetry as below:
offset = -round((qmin+qmax) / 2)

In a sense, this is a generalized version of the current definition of symmetry (offset = 0) because, with the new definition, floating point 0.0 will always line up with the midpoint of the integer grid -- 0 if int8, 128 if uint8.

This new definition has the following pros and cons

Pros:

• Setting quantizer.symmetric = True/False doesn't affect the qdq simulation accuracy.
• For most users who NEVER use unsigned symmetric, this is the most transparent solution (requires no user code change) to prevent catastrophic accuracy drop

Cons:

• The meaning of quantizer.signed and quantizer.symmetric slightly diverges from our day-to-day usage of "signed" and "symmetric".
  Particularly, quantizer.signed = False and quantizer.symmetric = True doesn't mean what we've been conventionally referring to as "unsigned symmetry".