MNNConvert converts all the fully connection layers into 1x1 convolutions.

But I found that when the input count is large (about 7x7x512=25088), then opencl may throw CL_INVALID_KERNEL_ARGS.

Not sure if some of kernel configs need to be updated to handle such cases. But I am curious why you convert all the fully connection layers into 1x1 convolutions. Any reason for doing this?