基于OpenCL的自动微分并行实现及其应用

Automatic Differentiation Based on OpenCL Parallel Computing and Its Application

摘要:

针对如光束平差这样的大规模优化问题，实现基于OpenCL的并行化自动微分。采用更有效的反向计算模式，实现对多参数函数的导数计算。在OpenCL框架下，主机端完成C/C++形式的函数构建以及基于拓扑排序的计算序列生成，设备端按照计算序列完成函数值以及导数的并行计算。测试结果表明，将实现的自动微分应用于光束平差的雅可比矩阵计算后，相比于采用OpenMP的Ceres Solver，运行速度提高了约3.6倍。

Abstract:

A parallelized implementation of automatic differentiation that derives from the problem of bundle adjustment is proposed, which is based on OpenCL parallel computing framework. Reverse mode of automatic differentiation is more efficient to compute the derivatives of functions with multiple parameters, which is the case of computing the Jacobian matrix in bundle adjustment problem. Under the framework of OpenCL, C/C++ style function construction and topological sorting based computational sequence generation are implemented on the host side. On the device side, function values and derivatives are computed in parallel according to computational sequence. Large scale bundle adjustment datasets are used to evaluate the proposed implementation. The result shows that our implementation runs about 3.6 times faster than Ceres Solver which utilizes OpenMP parallel programming model.