Heterogeneous Computing

From Threads to Tiles: T2T, a Compiler for CUDA-to-NPU Translation via 2D Vectorization