Method Forward

src/model/layer/opencl_convolution.cc:28–75

\copydoc Layer::Forward(int flag, const Tensor&)

/// \copydoc Layer::Forward(int flag, const Tensor&)
const Tensor OpenclConvolution::Forward(int flag, const Tensor &input) {
  CHECK(buf_.empty());
  CHECK_EQ(input.device()->lang(), kOpencl);
  CHECK_EQ(input.nDim(), 4u);

  if (flag & kTrain) buf_.push(input);

  auto batchsize = input.shape(0);
  auto imagesize = input.Size() / batchsize;
  auto data_type = input.data_type();
  auto device = input.device();

  // TODO(wangwei) update the layer config if the input sample shape changes
  CHECK(input.shape(1) == channels_ && input.shape(2) == height_ &&
        input.shape(3) == width_) << "input sample shape should not change";

  Shape shape{batchsize, num_filters_, conv_height_, conv_width_};
  Tensor output(shape, device, data_type);
  Tensor col_data(Shape{col_height_, col_width_}, device, data_type);

  for (size_t b = 0; b < batchsize; b++) {
    int offset = b * imagesize;

    col_data.device()->Exec([input, offset, col_data, this](Context * ctx) mutable {

      this->Im2Col(input.block(), offset,
      height_, width_,
      kernel_h_, kernel_w_,
      pad_h_, pad_w_,
      stride_h_, stride_w_,
      conv_height_, conv_width_,
      0, channels_,
      col_data.block(), ctx);
    },
    {input.block()},
    {col_data.block()});

    Tensor each = Mult(weight_, col_data);

    if (bias_term_) {
      AddColumn(bias_, &each);
    }

    CopyDataToFrom(&output, each, each.Size(), b * each.Size());
  }

  return output;
}


/// \copydoc Layer::Backward(int, const Tensor&, const Tensor&);