high performance computing on graphics processing units: hgpu.org

This work explores the viability of end-to-end convolutional neural network inference using OpenCL HLS kernels generated from TVM on Intel FPGAs. We explore layer-pipelined execution for small networks and time-multiplexed kernels for larger CNNs. Naively generated kernels do not produce efficient hardware. We propose a set of optimizations to increase parallelism, resource utilization, and more efficiently use memory bandwidth. They include loop unrolling, tiling, fusion, invariant code motion, cached writes, CL channels, autorun kernels, concurrent execution, and parameterized kernels. These optimizations improve performance up to a factor of 1150x over the naive baseline implementation generated by TVM. Compared to Keras/Tensorflow on a 56-core Xeon 8280, we observe performance improvements up to 4.57x and 1.4x over LeNet and MobileNet but has a slowdown at 0.43x for ResNet18/34.