What question did this study set out to answer?

The aim is to create fast CUDA kernels for Strassen-like matrix multiplication algorithms using SubCuber.

Compiling Strassen-like Matrix Multiplication Algorithms to Fast CUDA Kernels

View Full Paper

Ask AI

Bookmark

Authors

Randomized trial evaluates SubCuber’s performance in matrix multiplication with GPU optimization, highlighting efficiency gains.

The aim is to create fast CUDA kernels for Strassen-like matrix multiplication algorithms using SubCuber.
Developed a domain-specific compiler, SubCuber, to generate CUDA kernels.
Constructed dependency graphs and selected efficient kernel schedules for Strassen-like algorithms.
Applied fusion strategies based on recursion levels and matrix sizes.
SubCuber's code outperformed CUDA implementations by up to 12% for one recursion level and 22% for two levels.
State-of-the-art implementations only improved performance by up to 8% and 16% respectively.
In language models, SubCuber enhanced inference speed by up to 16%.

Abhinav Jangda (2025) studied this question.

View Full Paper

Ask AI

Bookmark