Add interface to create handle tensor using user-allocated buffer about tim-vx HOT 7 CLOSED

It could be like this:

  bool isHandleTensor = (spec_.attr_ & TensorAttribute::INPUT) ||
                        (spec_.attr_ & TensorAttribute::OUTPUT);
  
  if (isHandleTensor) {
    id_ = vsi_nn_AddTensorFromHandle(graph_->graph(), VSI_NN_TENSOR_ID_AUTO,
                                     &attr, reinterpret_cast<uint8_t*>(data_));
  } else {
    id_ = vsi_nn_AddTensor(graph_->graph(), VSI_NN_TENSOR_ID_AUTO, &attr,
                           nullptr);
  }

  if (VSI_NN_TENSOR_ID_NA == id_) {
    VSILOGE("Create tensor fail!");
    return false;
  }

  if (data_ != nullptr && !isHandleTensor) {
    if (!CopyDataToTensor(data_, 0)) {
      VSILOGE("Copy data to tensor fail!");
      return false;
    }
  }

however, the data_ member variable is const void*, maybe using const_cast or change the class interface (e.g., add a Graph::CreateTensorFromHandle method to avoid confusion?)

from tim-vx.

@Goose-Bomb , yes, standalone API is prefered in this case.

Graph::CreateTensorFromHandle() is reasonable.

from tim-vx.

@Goose-Bomb,

Are you still active on this item?

from tim-vx.

@Goose-Bomb,

Are you still active on this item?

Yeah, and I'm willing to contribute to this, but I don't know if this will break the current interface. Similar PR like #181 which replaces some public fields of classes like TensorSpec and Quantization to setter/getter is not merged. So I'm expecting maintainers from VeriSilicon to make this change. If you guys have no plans on doing this, I can try to contribute on this.

from tim-vx.

#181 already get merged. It just supports the getter with the const modifier.

We don't have a plan to do this in short term. We are welcome and appreciate it if you can contribute to this.

from tim-vx.

@Goose-Bomb I would like to close this issue because we decide to not support wrap user memory in TIM-VX. We already support import DMA buffer file descriptor in Tim-vx. We think it should be able to meet most requirements.

from tim-vx.

@Goose-Bomb I would like to close this issue because we decide to not support wrap user memory in TIM-VX. We already support import DMA buffer file descriptor in Tim-vx. We think it should be able to meet most requirements.

Alright, it makes sense.

from tim-vx.