Mixed-Precision MLP (int8 + Q-format)

A hybrid MLP that runs an int8 affine frontend and classifier around a Q8.8 fixed-point hidden tier, connected by the Phase 17 pure-integer qbridge converters — the entire inference path has no floating-point math at runtime.

How it works

Pipeline: QDense int8 frontend → qrelu → affineToQValueIntBuffer bridge → Q8.8 dense matvec (int32 accumulator) → qValueToAffineIntBuffer bridge → QDense int8 classifier. Both bridges use the same gemmlowp Q0.31 (multiplier, shift) primitive that Requantizer already relies on, so no <cmath>, <type_traits>, or float is needed at runtime.
Demonstrates the Phase 17 pure-integer mixed-precision qbridge (cpp/qbridge.hpp): the affineToQValueInt / qValueToAffineInt converters let an int8 affine model and a TinyMind Q-format layer interoperate at the deployable shape QUANTIZATION=1, FLOAT=0, STD=0. The conversion triples are built host-side once at calibration time by buildAffineToQValueIntParams / buildQValueToAffineIntParams.
The int8 logits land around 0.005 max-abs error vs the float reference on the bundled synthetic dataset, well inside the exemplar’s tolerance.

Build and run

cd examples/mixed_precision_mlp_int8_qformat
make release
make run
make plot      # needs matplotlib; a venv/pyenv works if it is not already in your Python

The Makefile builds hosted (FLOAT=1 STD=1 QUANTIZATION=1) for the parity report, but the inference path itself is freestanding-clean for the deployable FLOAT=0 STD=0 shape. make bench writes a CSV cycle/byte report and make golden emits the int8 byte stream wired into the integration suite. apps/import_pytorch/tinymind_import.py carries QFormatDense / HybridBoundary descriptors that emit the same weights.hpp layout for importing real PyTorch / TensorFlow models.

Output

int8 vs float parity for the hybrid int8 + Q-format MLP

The left panel overlays the int8-dequantized output on the float reference; the curves track tightly across all four outputs. The right panel’s per-element error stays under ~0.0043, showing the round trip through the Q8.8 hidden tier and the two pure-integer bridges loses only a few thousandths versus all-float math.

Source on GitHub