tilelang-cuda-patterns

You can implement CUDA kernels for matmul and GEMV using TileLang by selecting a predefined template from this pattern library and adapting the block and grid sizes to your problem dimensions. These templates leverage shared memory and efficient memory access strategies.

Available CUDA kernel templates for TileLang include elementwise computations, reduction aggregations, matmul, and GEMV. They provide ready-to-use structures for memory tiling and shared memory strategies to standardize and accelerate kernel development.

Yes, you need a CUDA runtime and the TileLang CUDA DSL to use these kernel patterns. The patterns map computations directly to grid, block, and memory primitives, requiring this specific environment to generate and execute the kernels correctly.

The best way to scaffold a reduction kernel in TileLang is to select the built-in reduction pattern provided here, which offers built-in strategies for aggregations like sum and max, and then adapt the block and grid configurations to fit your data size.

Memory tiling in TileLang CUDA patterns works by mapping computations to grid, block, and memory primitives. It utilizes shared memory strategies within the provided matmul and GEMV templates to ensure efficient memory access during kernel execution.