Module quantize 

INT8 Quantization with π-Based Calibration

Implements efficient INT8 quantization for CNN inference using π-derived constants to avoid quantization boundary resonance artifacts.

Why π?

In low-precision quantization, values tend to collapse into repeating buckets when scale factors align with powers of two. Using π-derived constants breaks this symmetry:

• π is irrational (non-repeating, infinite structure)
• Avoids power-of-2 boundary alignment
• Provides deterministic anti-resonance offsets

Quantization Schemes

• Symmetric: For weights (zero-centered distributions)
• Asymmetric: For activations (ReLU outputs are non-negative)
• Per-channel: Different scale per output channel (higher accuracy)
• Per-tensor: Single scale for entire tensor (faster)

Performance

INT8 inference provides:

• 4x memory reduction vs FP32
• 2-3x speedup on AVX2/AVX-512 (VNNI)
• 2-4x speedup on ARM NEON (SDOT)

Modules

pi_constants

π-based scale factors to avoid power-of-2 resonance

Structs

PerChannelQuantParams

Per-channel quantization parameters

QuantParams

Quantization parameters for a tensor or channel

QuantizedTensor

Quantized INT8 tensor storage

Enums

QuantizationType

Quantization type

Functions

dequantize_batch

Batch dequantize i8 to f32

dequantize_batch_avx2^⚠

AVX2 batch dequantization

dequantize_simd

SIMD-dispatched dequantization

quantize_batch

Batch quantize f32 to i8 using π-calibration

quantize_batch_avx2^⚠

AVX2 batch quantization (8 values at a time)

quantize_simd

SIMD-dispatched quantization