MatXtract: Sparsity-Aware Matrix Transformation via Cascaded Compute Density Extraction for SpMV

This repository provides a Sparse Matrix-Vector Multiplication (SpMV) computation library optimized for GPU architectures, leveraging Tensor Cores and CUDA Cores to achieve high performance through automated techniques.

Features

• Efficient SpMV computation on GPUs.
• Optimized utilization of tensor cores and CUDA cores.
• Support for both FP16 and FP64 precision.
• Easy-to-use build and execution process.

Getting Started

Hardware Requirements

• CPU: AMD EPYC 7V13 64-Core Processor
• GPU: NVIDIA A100 80GB PCIe (GPU driver version 570.124.06 or later)
• Disk Space: At least 300GB (required to store the sparse matrix dataset)

Software Requirements

• CUDA: CUDA-12.8 (tested). Lower versions (down to CUDA 11.0) are supported but may negatively affect performance.
• Compiler: GCC-11.4.0 or newer (tested)

Build Instructions

1. Clone the repository:

   git clone <repository_url>
   cd <repository_name>

2. Build the project:

   mkdir -p build
   cd build
   cmake ..
   make -j2

3. Configure FP64 of FP16 support (optional):

• To enable FP64 precision, modify the CMakeLists.txt file before building:

  option(USE_FP64 "Enable fp64 support" ON)
  

• The default configuration uses FP16 precision (OFF).

Dataset Preparation

To prepare the testing dataset, execute the following script in the data directory under the project root:

• Full Dataset: To prepare all matrices (~400GB):

bash prepare_all_dataset.sh

• Sample Dataset: To quickly test the execution on a few representative matrices:

bash prepare_sample_dataset.sh

The dataset will be generated in:data/mtx.

Running the Program

After compilation, the executable files are located in the build/ directory.

MatXtract Performance Test

Run MatXtract with specific crux parameters (global_col, local_row):

./matxtract_perftest (global_col) (local_row) <path_to_matrixA.mtx>

Note: global_col refers to the column-wise threshold $\tau_c$, and local_row refers to the row-wise threshold $\tau_r$ as described in the paper.

To use default crux parameters (global_col = 0, local_row = 0):

./matxtract_perftest <path_to_matrixA.mtx>

Bayesian Optimization

To identify approximately optimal crux parameters, use Bayesian optimization:

cd ML
bash ml_install.sh
source ml_venv/bin/activate
(ml_venv) python bayes_opt.py <path_to_matrixA.mtx>

For batch processing multiple matrices:

Set the matrix directory in batch_bayes_opt.py:

MATRIX_ROOT_DIR = "path_to_mtx_dir"

Then run:

(ml_venv) python batch_bayes_opt.py

Example Console Output (FP64):

[INFO] Testing matrix: ../data/mtx/cnr-2000/cnr-2000.mtx
Init0 (0,0) Time = 0.047995 ms
Init1 (1,1) Time = 0.080856 ms
===========================================
        Bayesian Optimization Result       
===========================================
Best col_frac  = 0.4592
Best hot_frac  = 0.3337
Min Time (ms)  = 0.0393

Baseline Comparisons

• cuSPARSE: To measure cuSPARSE's SpMV performance:

./cuda_perftest <path_to_matrixA.mtx>

• CSR5 and Merge-SpMV: Both are integrated into /baselines.