Please join me in welcoming Zihao Ye (@yzh119) as a new committer in TVM.
Zihao has made significant improvements on TensorIR, TVMScript and CUDA backends, Including but not limited to:
- Introduce new features and improve performance on the CUDA backend.
- Enable L2 cache flush during CUDA evaluation
- Robustify TensorIR schedule primitive
- Add new TensorIR primitive:
set_dtype
Additionally, Zihao extend the TensorIR to the sparse area in his new work (https://arxiv.org/pdf/2207.04606.pdf), which is accepted by ASPLOS 2023. The RFC is here (apache/tvm-rfcs#100).
His activities: