PWPolyF Piecewise Polynomial Activation

Overview

PWPolyF is a hardware activation layer that approximates nonlinear functions (GELU, SiLU, Sigmoid, Tanh) using piecewise polynomials evaluated with Horner’s method on a chain of DSPFP32 FMA units. It is RTL-only and targets Versal devices exclusively (requires the DSPFP32 primitive).

With the default degree=2, this uses two cascaded DSPs and one RAMB18 coefficient ROM per PE, giving single-cycle-per-element throughput.

The input domain is partitioned into 1 + 2*5*(2^K) segments. With the default K=3 this gives 81 segments. Polynomial coefficients are generated at HDL build time by fitting to PyTorch reference functions.

Usage

Use PWPolyFActivation from finn.util.torch_hw_modules as a drop-in replacement for nn.GELU, nn.SiLU, nn.Sigmoid, or nn.Tanh:

from finn.util.torch_hw_modules import PWPolyFActivation

# In your model
self.act = PWPolyFActivation(func="gelu", K=3, degree=2)

Export via Brevitas export_qonnx. The resulting PWPolyFunction node is converted to a PWPolyF HW op by InferPWPolyFLayer.

Path A: PWPolyFActivation           Path B: nn.GELU / nn.SiLU / etc.
    |  (util/torch_hw_modules.py)        |  torch.onnx.export
    |  export_qonnx                      |  (dynamo=True or False)
    v                                    v
PWPolyFunction node                 Standard ONNX ops (Gelu, Sigmoid,
(qonnx.custom_op.general)           Tanh, Sigmoid+Mul for SiLU,
                                    Div+Erf+Add+Mul+Mul for GELU)
    |                                      |
    +------------- both paths -------------+
                      |
                InferPWPolyFLayer
                      v
            PWPolyF HW op (custom_op/fpgadataflow)
                      |  SpecializeLayers
                      v
            PWPolyF_rtl (custom_op/fpgadataflow/rtl)
                      |  generate_hdl
                      v
            finn-rtllib/pwpolyf/hdl/ SystemVerilog IP

Standard ONNX Op Inference

InferPWPolyFLayer also recognises standard ONNX activation ops, allowing models using nn.GELU, nn.SiLU, nn.Sigmoid, or nn.Tanh to be automatically converted to PWPolyF HW layers with default K=3, degree=2.

Supported patterns:

• Gelu (opset 20+) or Erf-based GELU pattern → func="gelu"

• Sigmoid → func="sigmoid"

• Tanh → func="tanh"

• Sigmoid + Mul (SiLU pattern) → func="silu"

Only FLOAT32 inputs are converted.

Folding

PWPolyF uses PE parallelism. NumChannels % PE == 0 must hold. SetFolding handles PE selection automatically.