NeutronSparseLite

NeutronSparseLite accelerates sparse matrix-dense matrix multiplication (SpMM) on Ascend NPUs. The runtime is built on MindSpore, Huawei's Ascend-native AI framework, so the main program runs through the Ascend backend and dispatches to compiled NPU operators supplied by MindSpore/CANN whenever the corresponding tensor operations are executed.

The project uses a hybrid heterogeneous execution strategy.

In short, MindSpore is the primary Ascend-native runtime layer for this repository, and NeutronSparseLite relies on its prebuilt Ascend operator stack instead of implementing every NPU kernel directly in Python.

0. Environment Setup

We use openEuler 2.10.11 as the operating system and CANN Toolkit 8.2.rc1 for the Ascend backend. MindSpore must be installed with Ascend support; CPU-only MindSpore builds are not enough for the benchmark path in NeutronSparseLiteRun.py.

The Python packages and corresponding versions required for NeutronSparse are as follows:

Package	Version
Python	3.11
MindSpore	2.7.0.rc1

Why MindSpore matters here:

• Ascend-native: MindSpore is designed to target Ascend hardware through CANN and the Ascend runtime.
• Compiled NPU operators: calls such as ops.matmul, COOTensor.to_dense, and TensorScatterAdd are dispatched to compiled Ascend kernels when running on device_target="Ascend".
• Python orchestration, NPU execution: the Python code prepares data, chooses dense or sparse paths, and launches tensor operations; heavy numerical work is handled by the compiled NPU operator stack.

1. Compiling

Before running the project, you must compile the custom C++ preprocessing extension. This extension prepares graph metadata on CPU; runtime SpMM execution still goes through MindSpore's Ascend backend and its compiled NPU operators.

Enter the compilation directory

cd prepare/sgt_cpp

Clean previous build artifacts

rm -rf build

Compile the extension in-place

python setup.py build_ext --inplace

2. Get Dataset

Navigate to the data directory and extract the dataset archive.

Go to the data directory (assuming you are currently in prepare/sgt_cpp)

cd ../../origin_data/data_mtx

Extract the data

tar -zxvf data.tar.gz

3. Preprocess Data

After extracting the data, navigate to the prepare folder to run the preprocessing scripts.

Go back to the prepare folder

cd ../../prepare

1. Convert MTX format to COO format

   python mtx2coo.py

2. Reorder the graph (Replace 'reddit' with your dataset name)

   bash graph_reorder.sh reddit

3. Generate CSR format

   python generate_csr.py

4. Prepare data for NeutronSparse

   python NeutronSparsePrepare.py

4. Run NeutronSparse

Once compilation and data preparation are complete, you can run the main program.

Go back to the project root, then enter the run folder

cd ../NeutronSparseLite

Quick start the training/inference

python NeutronSparseLiteRun.py

At startup, the program sets:

context.set_context(mode=context.PYNATIVE_MODE, device_target="Ascend", device_id=0)

This makes MindSpore launch tensor operations on Ascend. If 3rdparty/ops-nn is present, the program also reports the optional ops-nn helper/source discovery status before falling back to the stable MindSpore operator path for execution.