high performance computing on graphics processing units: hgpu.org

In this report, we analyze the computational and performance aspects of current state-of-the-art object recognition and optical flow algorithms. First, we identify important algorithms for object recognition and optical flow, then we perform a pattern decomposition to identify key computations. We include profiles of the runtime and energy efficiency (GFLOPS/W) for our implementation of these applications on a commercial architecture. Finally, we include an analysis of memory-bandwidth boundedness for optical flow to identify opportunities for communication-avoiding algorithms. Our results were measured on an Intel i7-4770K (Haswell) reference platform. A five-layer convolutional neural network used for object classification achieves 0.70 GFLOPS/W, which is 21% of the theoretical compute bound for this Haswell processor. On the Horn-Schunck, Lucas-Kanade, and Brox optical flow methods our implementations achieve 0.0338, 0.0103, and 0.0203 GFLOPS/W respectively. Our implementation achieves 7.9% of the theoretical bandwidth bound, assuming no cross-iteration memory optimization, for Horn-Schunk optical flow using the Jacobi solver, and 9.7% of the bandwidth bound for the conjugate-gradient solver. To improve performance, we will focus first on increasing bandwidth utilization, then on doing cross-iteration memory optimizations such as blocking and tiling the Jacobi solver and employing communication-avoiding linear solvers. We also compare the runtime-accuracy tradeoffs for each optical flow method. We find that each method has distinct advantages over the other methods in terms of the runtime-accuracy tradeoff, so we will continue to develop and support all three methods in the future.