This is the source code repository of the ISCA'25 paper HYTE: Flexible Tiling for Sparse Accelerators via Hybrid Static-Dynamic Approaches.

$ g++ -O3 -march=native  src/config.cpp src/data.cpp src/estimation.cpp src/parameters.cpp src/util.cpp src/statistics.cpp src/cache.cpp src/dynamic.cpp src/simulator.cpp src/main.cpp -o hyte

The scheduler and simulator accept sparse matrices from MatrixMarket (.mtx). The folder containing these matrices is under data.

The following command simulates multiplication of matrix1 and matrix2 with the configuration specified in config/config.json:

$ ./hyte matrix1 matrix2 config/config.json

Here is a sample json configuration:

{
    "dataflow": 2,
    "format": 0,
    "transpose": 0,
    "minBlock": 2,
    "cachesize": 4,
    "memorybandwidth": 68,
    "PEcnt": 32,
    "srambank": 32,
    "iscache": 0,
    "cacheScheme": 0,
    "baselinetest": 0,
    "outputDir": "./output/"
}

The code contains both baseline test and HYTE test. You can change the execution mode through the baselinetest parameter in the config.json, in which 0 represents HYTE and 1 represents the three baselines (run in turn).

The static process and dynamic process of HYTE run together in ./hyte. During the HYTE execution in main.cpp, firstly getParameterSample() from estimation.cpp is called to get the requisite parameters of the specific matrix as described in Section 5.1 in the paper. Then, we search for the best tiling scheme as described in Section 5.2 in the paper. The cost model described in Section 5.3 is called through the function runTile() from simulator.cpp, with the first parameter isest = 1, representing the analytical cost model. After obtaining the optimized static tiling scheme, runTile() from simulator.cpp is called with the first parameter isest = 0, representing the simulated performance. During the simulation, update_T() in dynamic.cpp is called for dynamic tuning as described in Section 6.4 in the paper.

If you use this tool in your research, please kindly cite the following paper.

Xintong Li, Zhiyao Li, and Mingyu Gao. HYTE: Flexible Tiling for Sparse Accelerators via Hybrid Static-Dynamic Approaches. In Proceedings of the 52nd Annual International Symposium on Computer Architecture (ISCA), 2025.