VectorBlox vnnx_tflite.py Fix: Problem 4 - INT8 Constant ADD/SUB Operation Issue Resolution

Problem 4: INT8 Constant ADD/SUB Operation

Compilation Failure Log

Assertion failed: (activation_min != 0)

File "vbx_cnn_model.c", line XXXX

Root Cause

In the SPNv2 model, there are cases where one side of ADD or SUB operations is a constant. For example, bias addition like x + 0.5 or normalization like x - mean.

In TFLite, such operations are represented with one input as a constant tensor. VectorBlox SDK determines this as multi_input=False.

On the other hand, when both are variables like x + y, VectorBlox SDK determines this as multi_input=True.

The problem is that when multi_input=False, VectorBlox SDK does not initialize quantization parameters (activation_min/max, multiplier, shift, etc.) for ADD/SUB. Therefore, when executing VNNX, the C layer finds that one of the quantization parameters, activation_min, is 0, causing an assertion failure during INT8 full quantization.

Solution Process

We modified the code to initialize all quantization parameters even when multi_input=False.

Before modification, parameters were only set when multi_input=True:

if subcode in ['ADD', 'SUB']:
    if multi_input:
        # Set quantization parameters
        sn.activation_min = -128
        sn.activation_max = 127
        # Set multiplier, shift, etc.
    # else: do nothing → error in C layer!

After modification, parameters are always set regardless of multi_input:

if subcode in ['ADD', 'SUB']:
    # Always initialize quantization parameters
    sn.activation_min = -128
    sn.activation_max = 127
    
    # Set input/output scale, offset
    input_scale = i_tensor['quantization']['scale']
    output_scale = o_tensor['quantization']['scale']
    input_offset = i_tensor['quantization']['zero_point'][0]
    output_offset = o_tensor['quantization']['zero_point'][0]
    
    # Calculate and set multiplier, shift
    input_multiplier, input_shift = get_quantized_multiplier(input_scale)
    output_multiplier, output_shift = get_quantized_multiplier(output_scale)
    sn.input_multiplier = len(weights)
    weights += struct.pack(...)
    # ... (initialize all fields)
    
    # Additional settings for multi_input=False
    if not multi_input:
        # Also set filter_data, bias_data, etc.
        sn.eltwise8.left_shift = left_shift
        sn.eltwise8.swap_inputs = 0
        # ...

We ensured that all ELTWISE operations have the structure fields required by the C layer.

Solution Result

After modification, ADD/SUB with constant RHS also execute normally:

[ADD with constant RHS]
All quantization parameters initialized
VNNX execution successful ✓
C layer assertion passed ✓

Modification Summary

This modification enables all ADD/SUB operations, including constant operations, to execute stably.

Series Posts

• Previous: Problem 3 - SLICE/STRIDED_SLICE 5D Tensor Processing
• Next: Problem 5 - Constant Tensor Buffer Structure Compatibility

Language: 한국어 (Korean)