SparseZETA: Intelligent Auto-tuner for Designing High-Performance SpMV Programs
summary
SparseZETA generates SpMV kernels by predicting and applying sequences of 24 concrete operators (compress, row/column blocking, padding, mapping, reduction choices) in one pass, casting autotuning as behavior-cloning over an intermediate representation state. A multi-stage self-training pipeline produces ground-truth action sequences without exhaustive search. On NVIDIA A100 and RTX 2080 Ti it reports 1.27×–15.66× (1.44×–19.07×) speedups and reduces design time from hours/days to ~82.5 ms per matrix via CPU inference.
AI-written summary · about
chapters (11)
- 0:00SparseZETA and SPMV challenges
- 1:26Template-based autotuners limits
- 2:55AlphaSparse: Template-free generation
- 4:43Iterative search unnecessary
- 6:50Behavior cloning framework
- 8:14Action and state spaces
- 9:58Online inference process
- 11:45Multi-stage self-training
- 13:12Performance evaluation
- 14:18Summary and future
- 14:38Q&A
related talks
19mOptimizations and abstractions for sparse machine learning
24mMoving to RISC-V Vector: A Practical Journey of AI Operator Optimization
19mGeneralSparse: Bridging the Gap in SpMM for Pruned Large Language Model Inference..
28mFunctional Sparse Tensor Compilation
40mSparse models are fast models: Improving DNN inference performance by over 10X
16mAccelerate HPC and AI Applications with RVV Auto-Vectorization
23mSparse Computing Drives Energy-Efficient Computation for Artificial Intelligence
20mBoosting Video Codec With RISC-V Vector Extension