cavis/libnd4j/include/ops/declarable/impl/LegacyIndexReduceOp.cpp

/* ******************************************************************************
 *
 *
 * 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
 ******************************************************************************/

//
// Created by raver119 on 16.10.2017.
//

#include <ops/declarable/LegacyIndexReduceOp.h>
#include <helpers/ShapeUtils.h>
#include <helpers/TAD.h>
#include <graph/Status.h>
#include <helpers/ConstantTadHelper.h>


namespace sd {
    namespace ops {
        LegacyIndexReduceOp::LegacyIndexReduceOp() : LegacyOp::LegacyOp(1){
            //
        }

        LegacyIndexReduceOp::LegacyIndexReduceOp(int opNum) : LegacyOp::LegacyOp(1, opNum) {
            //
        }

        LegacyOp* LegacyIndexReduceOp::clone() {
            return new LegacyIndexReduceOp(this->_opNum);
        }

        ShapeList *LegacyIndexReduceOp::calculateOutputShape(ShapeList *inputShape, sd::graph::Context &block) {
            auto inShape = inputShape->at(0);

            if (block.getAxis()->size() == 0 && block.width() == 1) {
                Nd4jLong *newShape;
                // in this case we just return scalar
                ALLOCATE(newShape, block.getWorkspace(), shape::shapeInfoLength(2), Nd4jLong);
                newShape[0] = 2;
                newShape[1] = 1;
                newShape[2] = 1;
                newShape[3] = 1;
                newShape[4] = 1;
                newShape[6] = 1;
                newShape[7] = 99;

                auto result = ConstantShapeHelper::getInstance().createShapeInfo(ShapeDescriptor(newShape, DataType::INT64));
                RELEASE(newShape, block.getWorkspace());
                return SHAPELIST(result);
            } else if (block.getAxis()->size()){
                // in this case we're building proper shape for reduction
                auto array = INPUT_VARIABLE(0); //new NDArray(nullptr, inShape, block.getWorkspace());

                auto newShape = ShapeUtils::evalReduceShapeInfo('c', *block.getAxis(), *array, DataType::INT64, false, true, block.workspace());
                return SHAPELIST(newShape);
            }
            else {
                bool allAxes = false;
                auto indices = INPUT_VARIABLE(1);
                Nd4jLong rank = shape::rank(inShape);
                if (indices->lengthOf() == rank)
                    allAxes = true;

                std::vector<int> axis(indices->lengthOf());
                for (int e = 0; e < indices->lengthOf(); e++) {
                    // lol otherwise we segfault on macOS
                    int f = indices->e<int>(e);
                    axis[e] = f >= 0 ? f : f += rank;
                }
                if (allAxes){
                    Nd4jLong *newShape;
                        // in this case we just return scalar
                        ALLOCATE(newShape, block.getWorkspace(), shape::shapeInfoLength(2), Nd4jLong);
                        newShape[0] = 2;
                        newShape[1] = 1;
                        newShape[2] = 1;
                        newShape[3] = 1;
                        newShape[4] = 1;
                        newShape[6] = 1;
                        newShape[7] = 99;

                        auto result = ConstantShapeHelper::getInstance().createShapeInfo(ShapeDescriptor(newShape, DataType::INT64));
                        RELEASE(newShape, block.getWorkspace());
                        return SHAPELIST(result);
                } else {
                    // in this case we're building proper shape for reduction
                    auto array = INPUT_VARIABLE(0); //new NDArray(nullptr, inShape, block.getWorkspace());
                    return SHAPELIST(ShapeUtils::evalReduceShapeInfo('c', axis, *array, DataType::INT64, false, true, block.workspace()));
                }
            }
        }

        /**
        *   For all reductions rules are simple: either you return scalar, or you return reduced NDArray.
        *   It solely depends on input shape, and requested dimensions
        */
        Nd4jStatus LegacyIndexReduceOp::validateAndExecute(Context &block) {
            auto x = INPUT_VARIABLE(0);
            auto z = OUTPUT_VARIABLE(0);

            NDArray::prepareSpecialUse({z}, {x});

            if (z->dataType() != INT64) {
                throw std::runtime_error("IndexReduce operations require output to be INT64");
            }

            int opNum = block.opNum() < 0 ? this->_opNum : block.opNum();

            bool allAxes = false;

            ExtraArguments extras(*block.getTArguments());
            PointersManager manager(block.launchContext(), "LegacyIndexReduceOp");

            if (block.width() == 1) {
                if (block.getAxis()->size() == 0) {
                    // scalar
                    NativeOpExecutioner::execIndexReduceScalar(block.launchContext(), opNum, x->buffer(), x->shapeInfo(),
                                                                         x->specialBuffer(), x->specialShapeInfo(),
                                                                         extras.argumentsAsT(x->dataType()),
                                                                         z->buffer(), z->shapeInfo(),
                                                                         z->specialBuffer(), z->specialShapeInfo());
                } else {
                    // TAD
                    std::vector<int> dims(block.getAxis()->size());
                    for (size_t e = 0; e < dims.size(); e++) {
                        auto axe = block.getAxis()->at(e);
                        dims[e] = axe < 0 ? axe + x->rankOf(): axe;
                    }
                    if (dims.size() > 1)
                        std::sort(dims.begin(), dims.end());

                    auto tadPack = sd::ConstantTadHelper::getInstance().tadForDimensions(x->shapeInfo(), dims);

                    NativeOpExecutioner::execIndexReduce(block.launchContext(), opNum, x->buffer(), x->shapeInfo(),
                                                        x->specialBuffer(), x->specialShapeInfo(),
                                                        extras.argumentsAsT(x->dataType()),
                                                        reinterpret_cast<Nd4jLong *>(z->buffer()), z->shapeInfo(),
                                                        z->specialBuffer(), z->specialShapeInfo(),
                                                        nullptr, (int) dims.size(),
                                                        Environment::getInstance().isCPU() ? tadPack.primaryShapeInfo() : tadPack.specialShapeInfo(), Environment::getInstance().isCPU() ? tadPack.primaryOffsets() : tadPack.specialOffsets());
                }
            } else {
                // TF mode
                auto indices = INPUT_VARIABLE(1);
                if (indices->lengthOf() == x->rankOf())
                    allAxes = true;

                std::vector<int> axis(indices->lengthOf());
                for (int e = 0; e < indices->lengthOf(); e++) {
                    // lol otherwise we segfault on macOS
                    int f = indices->e<int>(e);
                    axis[e] = f >= 0 ? f : f += x->rankOf();
                }

                if (allAxes) {
                    NativeOpExecutioner::execIndexReduceScalar(block.launchContext(), opNum, x->buffer(), x->shapeInfo(),
                                                              x->specialBuffer(), x->specialShapeInfo(),
                                                              extras.argumentsAsT(x->dataType()),
                                                              z->buffer(), z->shapeInfo(), z->specialBuffer(),
                                                              z->specialShapeInfo());

                } else {
                    if (indices->lengthOf() > 1)
                        std::sort(axis.begin(), axis.end());

                    REQUIRE_TRUE(axis.size() > 0, 0, "Some dimensions required for reduction!");

                    auto tadPack = sd::ConstantTadHelper::getInstance().tadForDimensions(x->shapeInfo(), axis);

                    NativeOpExecutioner::execIndexReduce(block.launchContext(), opNum,
                            x->buffer(), x->shapeInfo(), x->specialBuffer(), x->specialShapeInfo(),
                            extras.argumentsAsT(x->dataType()),
                            reinterpret_cast<Nd4jLong *>(z->buffer()),
                            z->shapeInfo(), z->specialBuffer(), z->specialShapeInfo(),
                            nullptr, (int) axis.size(),
                            Environment::getInstance().isCPU() ? tadPack.primaryShapeInfo() : tadPack.specialShapeInfo(),
                            Environment::getInstance().isCPU() ? tadPack.primaryOffsets() : tadPack.specialOffsets());
                }
            }

            manager.synchronize();
            STORE_RESULT(*z);

            return Status::OK();
        }
    }
}