CANN/asnumpy-docs 架构设计

Architecture

【免费下载链接】asnumpy-docs项目地址: https://gitcode.com/cann/asnumpy-docs

This document describes the internal architecture of AsNumpy, including the three-layer design, the coreNPUArraydata structure, the API module layout, and the NPU extension strategy via OpenBOAT.

Three-Layer Architecture

AsNumpy is built on three layers that cleanly separate concerns:

+------------------------------------------+ | Python Frontend (asnumpy/*.py) | | - __init__.py 122+ exported symbols | | - array.py array creation (10 fn) | | - math.py math ops (80+ fn) | | - linalg/ linear algebra | | - random/ random sampling | | - logic.py logic functions (16 fn) | | - nn.py neural network (softmax) | | - statistics.py statistics (mean) | +------------------------------------------+ | pybind11 binding layer (python/*.cpp) PYBIND11_MODULE(asnumpy_core, ...) | +------------------------------------------+ | C++ Core (src/, include/) | | - NPUArray core data structure | | - namespace asnumpy operator impls | | - Ascend ACL / ACLNN operator wrappers | | - CANN Runtime device management | +------------------------------------------+ | Huawei CANN (ascendcl, runtime, nnopbase, opapi) | +------------------------------------------+ | Ascend NPU Hardware (910B) | +------------------------------------------+

NPUArray Design

NPUArrayis the core data structure of AsNumpy. It is designed around three principles:

Compatibility

NPUArrayexposes the same interface asnumpy.ndarrayat the Python level. Functions are named identically (add,multiply,matmul, …) so existing NumPy code can be migrated by changing the import and adding data transfer calls.

Encapsulation

Internally,NPUArrayholds:

• dtype— element type (aclDtype)
• shape— dimension sizes
• strides— memory layout
• tensorPtr— pointer to the underlyingaclTensor
• devicePtr(private) — raw device memory address, exposed viadevice_address()

Users never interact with these fields directly; the Python layer presents a clean ndarray-like interface.

Resource Management

NPUArrayfollows RAII: the destructor automatically callsaclDestroyTensorandaclrtFree, eliminating manual memory management. All four C++ value semantics are implemented (copy constructor, move constructor, copy assignment, move assignment).

Data transfer:

• FromNumpy— usesACL_MEMCPY_HOST_TO_DEVICE
• ToNumpy— usesACL_MEMCPY_DEVICE_TO_HOST;float16/BF16require specialuint16_tunpacking

API Architecture

AsNumpy's API is divided intofunctional modulesandfoundation modules:

Functional modulescover the primary scientific computing domains:

Foundation modulessupport the functional layer:

NPU Extension Module

CANN's built-in operators are primarily designed for deep learning (training and inference). AsNumpy targets general scientific computing — data analysis, numerical methods, signal processing — which requires a broader operator set than CANN alone provides.

The gap:CANN built-in operators cannot cover all of NumPy's API surface.

The strategy:

• Wrap existing CANN built-in operators directly
• Develop missing operators manually (Ascend C)
• Supplement with the OpenBOAT open-source operator library

The value:This three-pronged approach progressively closes the compatibility gap toward the goal of covering the top 100 most-used NumPy APIs by v1.0.

OpenBOAT

OpenBOATis an open-source Ascend C operator library built and maintained by theAISS Group, School of Computer Science, Harbin Institute of Technology(led by Prof. Su Tonghua).

AsNumpy and OpenBOAT are developed by the same research group, ensuring tight integration and coordinated roadmap planning.

【免费下载链接】asnumpy-docs项目地址: https://gitcode.com/cann/asnumpy-docs