Introduction
AscendSiPBoost (SiP) provides a series of high-performance signal processing operators. These operators can be called via PyTorch for AI models or directly through C++ for signal processing tasks. It includes fused operators for Basic Linear Algebra Subprograms (BLAS), Fast Fourier Transforms (FFT), complex number calculation, and signal processing.
This document describes how to install AscendSiPBoost and provides sample code for typical use cases, helping developers quickly get started.
Architecture
Figure 1 shows the position of AscendSiPBoost in the Ascend operator technology stack.
- The AscendSiPBoost framework manages all operators. On the host side, it handles the tiling function, while on the device side, it supports binary loading. It also provides APIs for the upper layer to call a single operator or multiple operators in batches.
- FFT operators consist of a dedicated NPU Kernel and a PLAN framework. It provides external APIs for C2C, C2R, and R2C transformations to support developer implementation.
- BLAS operators provide a dedicated kernel based on BLAS-related standard definitions and APIs from level 1 to level 3 for developers.
- The basic complex computing library provides fundamental operators supporting complex number types, enabling flexible combination and usage on the user side. These operators are not provided in this version.
- The signal-field fusion operator library provides fusion operators such as PC, MTD, CFAR, and Interpolation. It is tailored for scenarios such as pulse signal analysis, dynamic target detection, and constant false alarm. Some interpolation operators are provided in this version.
- Solver operators provide complex linear algebra functions based on BLAS, such as matrix decomposition and eigenvalue solving. These operators are not provided in this version.