Accelerating Indirect Learning Architecture (ILA) Digital Predistortion and LEO Optical Link Simulations via 128-Bit Vector SIMD Execution in Hybrid WebAssembly and C++23 Environments
Physical-layer simulation of low Earth orbit (LEO) satellite systems involves two distinctcomputational workloads: the identification of nonlinear behavioral models of power amplifiers(PAs) for digital predistortion (DPD), and the large-scale evaluation of radio-frequency andoptical atmospheric propagation models. In current practice, both workloads run either incommercially licensed environments or in Python prototypes whose performance degradesunder iterative parameter sweeps. This paper presents PhasorWave, a single-source C++23computational core that compiles both to native code vectorized with AVX2/FMA3 and toWebAssembly with 128-bit SIMD extensions (WASM SIMD128). The core implements threekernels: (i) Generalized Memory Polynomial (GMP) coefficient extraction under the IndirectLearning Architecture (ILA) using QR decomposition via modified Gram–Schmidt withDaniel–Gragg–Kaufman–Stewart (DGKS) re-orthogonalization and explicit back-substitution,avoiding both the normal equations and the singular value decomposition (SVD); (ii) theITU-R P.618-14 rain attenuation procedure in branch-free vector form; and (iii) the HufnagelValley (HV-5/7) atmospheric turbulence model for computing the Fried parameter (r0) andthe isoplanatic angle (θ0). Native Python integration uses a zero-copy bridge via pybind11 andthe memoryview buffer protocol. Benchmarked on batches of N = 50,000 elements on an IntelCore i5-10300H host, the WASM SIMD128 module (88620bytes, ≈ 89KiB) achieves a medianlatency of 1.63ms for ITU-R P.618-14 rain attenuation, 1.36ms for HV-5/7 optical turbulence,and 101.24ms for combined ILA-GMP extraction and application (P = 39 coefficients),representing speedups of 3.1× to 3.7× over vectorized Python/NumPy while the nativeAVX2/FMA3 path reaches 5.4× to 16.2× speedups. Against double-precision references, thesingle-precision SIMD kernels maintain a maximum absolute error of 4.36 × 10−5 dB in Ap, amaximum relative error of 3.83 × 10−7 in r0, a GMP modeling NMSE of −73.02dB, and anACPR improvement of 39.66dB over three ILA iterations.
Authors
- Alejandro Vega Morilla
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-17
- DOI
- https://doi.org/10.5281/zenodo.24326095
- Primary Topic
- Parallel Computing and Optimization Techniques
- Type
- preprint