Input size limit for backprop about gpu-rodinia

[yuhc]

Hi, I'm not an expert in neural network, but I think it's possible to tune the code to use larger grid and thread size. Then you'll be able to use larger dataset.

(Related files:
https://github.com/yuhc/gpu-rodinia/blob/master/cuda/backprop/backprop_cuda.cu#L80
https://github.com/yuhc/gpu-rodinia/blob/master/cuda/backprop/facetrain.c#L19)

[mahmoodn]

Actually, 1048576 is divided by 16. So, I think it is a valid input size. However I don't know if the original author has limited his code to such size.