deeplearning4j/deeplearning4j

/* ******************************************************************************
 *
 *
 * This program and the accompanying materials are made available under the
 * terms of the Apache License, Version 2.0 which is available at
 * https://www.apache.org/licenses/LICENSE-2.0.
 *
 *  See the NOTICE file distributed with this work for additional
 *  information regarding copyright ownership.
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
 * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
 * License for the specific language governing permissions and limitations
 * under the License.
 *
 * SPDX-License-Identifier: Apache-2.0
 ******************************************************************************/

//
// @author Yurii Shyrma (iuriish@yahoo.com)
//

#include <system/op_boilerplate.h>
#if NOT_EXCLUDED(OP_broadcast_dynamic_shape)

#include <ops/declarable/CustomOperations.h>

namespace sd {
namespace ops {

//////////////////////////////////////////////////////////////////////////
CUSTOM_OP_IMPL(broadcast_dynamic_shape, 2, 1, false, 0, 0) {
  auto x = INPUT_VARIABLE(0);
  auto y = INPUT_VARIABLE(1);

  auto z = OUTPUT_VARIABLE(0);

  REQUIRE_TRUE(x->rankOf() == 1, 0,
               "BROADCAST_DYNAMIC_SHAPE OP: the first input array must have rank = 1, but got %i instead!",
               x->rankOf());
  REQUIRE_TRUE(y->rankOf() == 1, 0,
               "BROADCAST_DYNAMIC_SHAPE OP: the second input array must have rank = 1, but got %i instead!",
               y->rankOf());
  REQUIRE_TRUE(x->dataType() == y->dataType(), 0,
               "BROADCAST_DYNAMIC_SHAPE OP: both input arrays must have the same integer type !");

  // contract shapeInfos, neglect and don't fill strides, ews, order
  // shapes are of interest only
  std::vector<sd::LongType> xShapeInfo(shape::shapeInfoLength(x->lengthOf()));
  std::vector<sd::LongType> yShapeInfo(shape::shapeInfoLength(y->lengthOf()));

  // fill rank and data type
  xShapeInfo[0] = x->lengthOf();
  yShapeInfo[0] = y->lengthOf();
  ArrayOptions::setDataType(
      xShapeInfo.data(),
      sd::DataType::INT64);  // fill with some data type, it doesn't matter what type exactly to choose
  ArrayOptions::setDataType(yShapeInfo.data(), sd::DataType::INT64);

  for (sd::LongType i = 0; i < x->lengthOf(); ++i) xShapeInfo[i + 1] = x->e<sd::LongType>(i);

  for (sd::LongType i = 0; i < y->lengthOf(); ++i) yShapeInfo[i + 1] = y->e<sd::LongType>(i);

  const sd::LongType* poinerOnOutShapeInfo = nullptr;

  const bool isBroadcastPossible = ShapeUtils::evalBroadcastShapeInfo(
      xShapeInfo.data(), yShapeInfo.data(), true, poinerOnOutShapeInfo, block.launchContext()->getWorkspace());

  REQUIRE_TRUE(
      isBroadcastPossible, 0,
      "BROADCAST_DYNAMIC_SHAPE OP: the shapes of two input arrays %s and %s are not suitable for broadcast operation !",
      ShapeUtils::shapeAsString(xShapeInfo.data()).c_str(), ShapeUtils::shapeAsString(yShapeInfo.data()).c_str());

  for (sd::LongType i = 0; i < z->lengthOf(); ++i) z->p<sd::LongType>(i, poinerOnOutShapeInfo[i + 1]);

  return sd::Status::OK;
}

DECLARE_TYPES(broadcast_dynamic_shape) {
  getOpDescriptor()->setAllowedOutputTypes({ALL_INTS})->setAllowedInputTypes({ALL_INTS});
}

//////////////////////////////////////////////////////////////////////////
DECLARE_SHAPE_FN(broadcast_dynamic_shape) {
  const int xRank = INPUT_VARIABLE(0)->lengthOf();
  const int yRank = INPUT_VARIABLE(1)->lengthOf();

  const int maxRank = xRank > yRank ? xRank : yRank;

  auto outputShapeInfo =
      ConstantShapeHelper::getInstance().vectorShapeInfo(maxRank, ArrayOptions::dataType(inputShape->at(0)));

  return SHAPELIST(outputShapeInfo);
}

}  // namespace ops
}  // namespace sd

#endif