A zero-point Z is the integer code that represents real zero in affine quantization. Together with scale S, it positions a tensor’s real-valued range on an integer grid: r = S(q − Z). Exact representation of zero preserves ReLU sparsity and supports correct zero-padding. section 2.2 section 2.6
For real bounds r_min and r_max and integer bounds q_min and q_max, first extend the real range to include zero. Calculate S = (r_max − r_min)/(q_max − q_min), then Z = round(q_min − r_min/S). Quantize a value with q = clip(round(r/S + Z), q_min, q_max); reconstruct it with r = S(q − Z). Rounding introduces granular error, while clipping introduces overload error. section 2.2 section 2.6
Unsigned 8-bit ReLU activations
For values from 0 to 6 and integer codes 0 to 255, S = 6/255 ≈ 0.023529 and Z = 0. The value 0.25 maps to 11 and reconstructs to approximately 0.258824; zero maps exactly to 0. section 2.2
A nonzero zero-point adds correction terms to integer dot products. Symmetric weights set the weight zero-point to zero, removing the input-sum correction; weight-dependent corrections can be precomputed. Asymmetric activations can instead use the grid more efficiently for one-sided distributions, at the cost of correction overhead. Check the target hardware and runtime when choosing the scheme. section 2.2 section 2.6
Table 2.3 Affine dot product expansion terms
|
Term
|
Formula
|
Description
|
Computational Characteristic
|
|
1. Core Product
|
∑(qwqx)
|
Standard integer dot product
|
Expensive. Main O(N) MAC operation.
|
|
2. Weight Sum
|
−Zx∑qw
|
Weights scaled by input zero-point
|
Pre-computable.
∑qw is constant for trained model.
|
|
3. Input Sum
|
−Zw∑qx
|
Inputs scaled by weight zero-point
|
Runtime
Cost. ∑qx depends on input.
|
|
4. Offset Constant
|
+N⋅Zw⋅Zx
|
Constant offset
|
Pre-computable. Can fold into bias.
|
The dot-product expansion identifies which zero-point corrections can be precomputed and which depend on runtime inputs. section 2.2
Calibration determines the range used to calculate S and Z. Weight ranges may be known, but activation ranges vary with inputs; percentile clipping or entropy minimization can select more robust boundaries, and distribution drift may require recalibration. section 2.6

