Open hseok-oh opened 7 months ago
Due to recommendation from https://github.com/Samsung/ONE/issues/12814#issuecomment-2316619907 I'm assigning myself to this issue ;)
@hseok-oh Could you please take a look at https://github.com/Samsung/ONE/pull/14057? I am not sure if you plan implementation similar to my draft or maybe more global approach (like creating common buffer during tensor creation). The second one approach has much larger affected area.
We are always using
memcpy
forReshape
andExpandDims
operation. If these operation's input is used only once and updated shape does not require padding, we can skip copy operation.For this, we need to modify memory allocation and tensor handling implementation.