OpenVINO Version

2024.6.0

Operating System

Ubuntu 20.04 (LTS)

Device used for inference

CPU

Framework

None

Model used

No response

Issue description

I have identified a behavior divergence between OpenVINO and standard NumPy arithmetic regarding uint8 subtraction underflow.

When performing subtraction A - B where A < B using uint8 data types:

NumPy (Oracle) performs modular arithmetic (wrap-around). Result: (3 - 4) mod 256 = 255.

OpenVINO (CPU) performs saturation arithmetic. Result: max(0, 3 - 4) = 0.

While saturation is common in image processing, the standard Subtract operator in generic computational graphs typically expects standard integer arithmetic (wrap-around) unless specified as a specific "QLinear" or "Saturated" op. This inconsistency causes accuracy drops when porting models that rely on bitwise/integer arithmetic logic.

Step-by-step reproduction

Construct a minimal OpenVINO graph with two uint8 Inputs and a Subtract node.

Feed inputs A=3 and B=4.

Compare the OpenVINO inference result with NumPy.

bug-1.py

Relevant log output

--- Reproducing uint8 Sub Arithmetic Underflow (Saturation) BUG ---

Input A (uint8): 3
Input B (uint8): 4

--- Oracle (Numpy) Calculation ---
np.uint8(3) - np.uint8(4) = 255

--- OpenVINO Calculation ---
OpenVINO Sub(3, 4) = 0

--- Conclusion ---
Oracle (Expected, Wrap-around): 255
OpenVINO (Actual): 0

✅ BUG REPRODUCED!
   Oracle (uint8) 3 - 4 = 255 (Wrap-around)
   OpenVINO (uint8) 3 - 4 = 0 (Saturation)
   This matches the reported behavior: Oracle: 255 vs OpenVINO: 0

Issue submission checklist

• I'm reporting an issue. It's not a question.
• I checked the problem with the documentation, FAQ, open issues, Stack Overflow, etc., and have not found a solution.
• There is reproducer code and related data files such as images, videos, models, etc.