We demonstrate the efficiency and performance-portability of encoding sparse ten- sors in a linearized format. This encoding scheme allows a unified algorithm for matricized tensor times Khatri-Rao product (MTTKRP) that achieves high perfor- mance on both CPUs and GPUs, using a single, compact, and mode-agnostic copy of the tensor. Our new linearized sparse tensor format achieves geometric mean speedup of 4.3x and 2.6x over the state-of-the-art on the latest CPU and GPU architectures, respectively.