神经网络前向传播在GPU上的实现

基于CUDA架构在GPU上实现了神经网络前向传播算法,该算法利用神经网络各层内神经元计算的并行性,每层使用一个Kernel函数来并行计算该层神经元的值,每个Kernel函数都根据神经网络的特性和CUDA架构的特点进行优化。实验表明,该算法比普通的CPU上的算法快了约7倍。研究结果对于提高神经网络的运算速度以及CUDA的适用场合都有参考价值。

A forward propagation implementation of neural network on GPU

A neural network forward propagation algorithm based on CUDA is implemented. This algorithm utilizes the parallelism of each layer, uses a kernel function to handle the parallelism computation of the neuron values of that level, and each kernel function is optimized according the characteristic of network and CUDA architecture. Experiment shows that this method can achieve up to 7 times of speedup over the ordinary method on CPU. This research has reference value for both increasing neural network computing speed and suitability of CUDA.