cavis/libnd4j/include/ops/declarable/helpers/cuda/flatten.cu

/*******************************************************************************
 * Copyright (c) 2015-2018 Skymind, Inc.
 *
 * 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.
 *
 * 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 raver119@gmail.com
//

#include <ops/declarable/helpers/flatten.h>
#include <helpers/PointersManager.h>

namespace sd {
    namespace ops {
        namespace helpers {
            template <typename T>
            void _CUDA_G flattenKernel(void **xBuffers, Nd4jLong **xShapeInfos, Nd4jLong *offsets, Nd4jLong numInputs, void *zBuffer, const Nd4jLong *zShapeInfo, char order) {

                int xCoord[MAX_RANK];

                // each block of threads works on 1 input array
                for (Nd4jLong e = blockIdx.x; e < numInputs; e += gridDim.x) {
                    auto z = reinterpret_cast<T*>(zBuffer) + offsets[e];

                    auto xBuffer = reinterpret_cast<T*>(xBuffers[e]);
                    auto xShapeInfo = xShapeInfos[e];
                    auto xLength = shape::length(xShapeInfo);

                    // each element of this input array has own place within common output array
                    for (uint i = threadIdx.x; i < xLength; i += blockDim.x)
                        z[i] = xBuffer[getIndexOffsetOrdered(i, xShapeInfo, order)];
                }
            }

            template <typename T>
            void flatten_(sd::LaunchContext *context, std::vector<NDArray*> &inputs, NDArray *output, char order) {
                PointersManager pm(context, "flatten");

                std::vector<const void*> hdBuffers(inputs.size());
                std::vector<Nd4jLong> hOffsets(inputs.size());
                std::vector<const Nd4jLong *> hdShapes(inputs.size());
                Nd4jLong cOffset = 0;

                // calculating offsets in output
                for (int e = 0; e < inputs.size(); e++) {
                    hOffsets[e] = cOffset;
                    cOffset += inputs[e]->lengthOf();

                    hdBuffers[e] = inputs[e]->specialBuffer();
                    hdShapes[e] = inputs[e]->specialShapeInfo();
                }

                // copying pointers to device
                auto dBuffers = (void **) pm.replicatePointer(hdBuffers.data(), inputs.size() * sizeof(void*));
                auto dShapes = (Nd4jLong **)pm.replicatePointer(hdShapes.data(), inputs.size() * sizeof(Nd4jLong*));
                auto dOffsets = (Nd4jLong *) pm.replicatePointer(hOffsets.data(), inputs.size() * sizeof(Nd4jLong));


                flattenKernel<T><<<256, 512, 8192, *context->getCudaStream()>>>(dBuffers, dShapes, dOffsets, inputs.size(), output->specialBuffer(), output->specialShapeInfo(), order);

                pm.synchronize();
            }

            void flatten(sd::LaunchContext *context, std::vector<NDArray*> &inputs, NDArray *output, char order) {
                // FIXME: we want NDArrayFactory::prepareSpecialUse here eventually
                for (auto v:inputs)
                    v->syncToDevice();

                BUILD_SINGLE_SELECTOR(output->dataType(), flatten_, (context, inputs, output, order), LIBND4J_TYPES);
                NDArray::registerSpecialUse({output}, {});
            }
        }
    }
}