Triton-Seq
A compiler-extended Triton framework for GPU-accelerated Smith-Waterman sequence alignment with custom MLIR passes for automatic shared memory optimization.
Cornell University · Nov. 2024 – Apr. 2026
Research Intern with Prof. Zhiru Zhang
A high-performance GPU-accelerated framework for Smith-Waterman sequence alignment, featuring optimized Triton kernels and custom compiler extensions for automatic shared memory optimization.
Highlights
- Optimized Triton kernel with wavefront parallelism, a ring-buffer design, and 4-bit DNA sequence encoding.
- Custom MLIR passes for automatic shared-memory promotion, reproducing a manually validated optimization of approximately 35%.
- Asynchronous double-buffered execution that overlaps CPU preparation and GPU execution across two CUDA streams.
- Bit-exact correctness verification against AGAThA and ksw2.
