Reuse buffer in concatenation

Questions
1

I’m implementing a DenseNet and there’s an optimization I’d like to make, I’m curious if it’s expressible in TVM. Naively implementing a DenseNet requires a large number of concatenations, which consume a large amount of memory and bandwidth. The output of each conv2d layer of a DenseNet is concatenated with its inputs and fed to the next layer.

If we look at a single pixel, the operations are as follows:

RGB . kernel1 -> M
RGBM . kernel2 -> N
RGBMN . kernel3 -> O
RGBMNO . kernel4 -> … and so on…

Ideally, I’d like to set up my convolution layers so that they write into slices of the target buffer, which I can then use as-is for the next convolution.

Is there some way to implement this correctly?

2

no, that’s not possible. But I think NNVM already does some kind of in-place op detection. You can read the contents of nnvm/src/pass/plan_memory.cc and see if it suits your need.

3

One way to solve the problem is to just build kernel as you normally do, when you pass in output array structure pass in slice of a big array.