Kernel fusion combines multiple GPU operations into a single device kernel to reduce global memory traffic and eliminate intermediate buffers. Using sum(abs(x)) as a running example, three fusion approaches are compared: manual CUDA C++ kernel fusion (full control, highest maintenance), implicit fusion via torch.compile (automatic but non-deterministic), and explicit fusion via cuda.compute using TransformIterator with reduce_into (Python ergonomics with deterministic, CUB-backed results). All three fused approaches achieve ~3x speedup over the naive two-kernel baseline by reducing memory traffic from 3 GiB to 1 GiB on an RTX 4090. The cuda.compute approach is highlighted as offering the best balance of productivity and predictability for Python developers.
