Rewrite the DNN in a way that all the memory are passed from outside

[syifan]

[JunjoFor]

I have seen that dnn benchmarks don't have the option to use unified-gpu-memory. Is this related to that? I am interested in execute dnn workloads in unified gpus with unified memory and virtual memory. Could you give me more information about this topic?

[syifan]

@JunjoFor -unified-gpus and -use-unified-memory should be available to all the benchmarks.

[JunjoFor]

@syifan You can't run unified memory in dnn workloads, there is a panic in lenet, minerva, vgg16 and xor benchmarks. It is not implemented.

// SetUnifiedMemory asks the benchmark to use unified memory.
func (b *Benchmark) SetUnifiedMemory() {
	panic("unified memory is not supported by dnn workloads")
}