AKGF: Automatic Kernel Generation for DNN on CPU-FPGA.

This article presents the Automatic Kernel Generation for DNN on CPU-FPGA (AKGF) framework, designed to optimize the deployment of deep neural networks (DNN) on heterogeneous CPU-FPGA platforms. AKGF leverages TVM's Halide intermediate representation (IR) and the polyhedral model to generate and optimize code that efficiently partitions computation between CPU and FPGA cores, improving inference speed and power consumption. Experimental evaluations on Xilinx ZCU102 and other FPGA boards demonstrate that AKGF achieves significant performance gains—up to 6.7 times speedup—and reduces power consumption by half compared to state-of-the-art accelerators across various DNN models including GEMM, AlexNet, VGG19, and YOLOv5. The framework's optimizations include legality checks, loop transformations, loop fusion and tiling, latency hiding, and an auto-tuner with a semi-supervised cost model tailored for CPU-FPGA architectures. Limitations include its current dependence on Xilinx FPGA toolchains and the need for hardware-specific code generation support.