Extending Tensor Virtual Machine to Support Deep-Learning Accelerators with Convolution Cores

Deep-learning accelerators are increasingly popular. There are two prevalent accelerator architectures: one based on general matrix multiplication units and the other on convolution cores. However, Tensor Virtual Machine (TVM), a widely used deep-learning compiler stack, does not support the latter. This paper proposes a general framework for extending TVM to support deep-learning accelerators with convolution cores. We have applied it to two well-known accelerators: Nvidia's NVDLA and Bitmain's BM1880 successfully. Deep-learning workloads can now be readily deployed to these accelerators through TVM and executed efficiently. This framework can extend TVM to other accelerators with minimum effort.