124x Slower: What PyTorch DataLoader Actually Does at the Kernel Level

TL;DR: PyTorch's DataLoader can be 50-124x slower than direct tensor indexing for in-memory GPU...