Method DoXYBroadcast

tensorflow/stream_executor/dnn.h:2403–2409

Grows the input tensor by replicating the X and Y dimensions. The batch and depth/feature_map dimensions are unchanged. Currently, the input tensor is limited to X=1 and Y=1. For example, the input has dimensions x=2, y=3, and replicate_x=3, replicate_y=2. The diagonal elements of the output would be: [x0y0, x1y1, x0y2, x1y0, x0y1, x1y2]. Here is the example as a picture. input: AB CD EF broadcas

  //  output_data: un-owned device memory region in which to place the
  //    padded result.
  virtual bool DoXYBroadcast(Stream* stream,
                             const dnn::BatchDescriptor& dimensions,
                             const DeviceMemory<float>& input_data,
                             int64 replicate_x, int64 replicate_y,
                             DeviceMemory<float>* output_data) {
    return false;
  }

  // Enqueues an asynchronous memcpy of the *quantized* output of a layer (that
  // is, bytes instead of scaled floats) into 'host_dst' if they are available