Over the last few years, optical computing has become a potential solution to computationally heavy convolution, aimed at accelerating various arti cial intelligence applications. However, past schemes have never e ciently realized fully parallel optical convolution. Here, we propose a new paradigm for a universal convolution accelerator with truly massive parallelism and high precision based on optical multi-imaging-casting architecture. Speci cally, a two-dimensional Dammann grating is adopted for the generation of multiple displaced images of the kernel, which is the core process for kernel sliding on the convolved matrix. Our experimental results indicate that the computing accuracy is typically close to 8-bit, and this accuracy can be improved further by using hybrid analog-digital coding method. In addition, a convolutional neural network for the standard MNIST dataset is demonstrated, and the recognition accuracy for inference is up to 97.3%. The paradigm reported here will open new opportunities for high-throughput universal convolution accelerators for real-time or quasi-real-time AI applications.