Accelerating HDC-CNN Hybrid Models Using Custom Instructions on RISC-V GPUs