Feature description
Add TileLang + TVM Backend Support for High-Performance Tensor Core/Kernel Optimization
Feature motivation
TileLang is a state-of-the-art tile-based DSL built on TVM that provides:
- Cross-hardware kernel generation (NVIDIA / AMD / Ascend / CPU)
- Explicit tile abstraction
- Shared memory / pipelining / tensor core support
- Automatic performance tuning via TVM
- Near-CUDA performance with Python/Rust-like syntax
By integrating TileLang + TVM, Burn could:
- Significantly improve compute-bound kernel performance (GEMM, Attention, Conv)
- Gain cross-hardware portability beyond native backends
- Support user-defined high-performance custom kernels
- Complement existing backends without breaking changes
Feature description
Add TileLang + TVM Backend Support for High-Performance Tensor Core/Kernel Optimization
Feature motivation
TileLang is a state-of-the-art tile-based DSL built on TVM that provides:
By integrating TileLang + TVM, Burn could: