cavis/libnd4j/include/ops/declarable/helpers/cuda/prefix.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 Yurii Shyrma (iuriish@yahoo.com), created on 12.06.2019
//

#include <ops/ops.h>
#include <ConstantTadHelper.h>
#include <PointersManager.h>
#include <ShapeUtils.h>
#include <ops/declarable/helpers/prefix.h>

namespace nd4j {
namespace ops {
namespace helpers {

///////////////////////////////////////////////////////////////////
template <typename T>
__global__ static void prefixPerBlockCuda(scalar::Ops op,
                                        const void* vx, const Nd4jLong* xTadShapeInfo, const Nd4jLong* xTadOffsets,
                                              void* vz, const Nd4jLong* zTadShapeInfo, const Nd4jLong* zTadOffsets,
                                        const Nd4jLong numTads, const Nd4jLong tadLen,
                                        const bool exclusive, const bool reverse) {

    __shared__ T *shared, lastElemInChunk;
    __shared__ uint numTadChunks, blockDim2;

    if (threadIdx.x == 0) {
        extern __shared__ unsigned char shmem[];
        shared       = reinterpret_cast<T*>(shmem);
        blockDim2    = 2 * blockDim.x;
        numTadChunks = (tadLen +  blockDim2 - 1) / blockDim2;  // ceil
    }
    __syncthreads();

    const auto xTad = reinterpret_cast<const T*>(vx) + xTadOffsets[blockIdx.x];
          auto zTad = reinterpret_cast<T*>(vz)       + zTadOffsets[blockIdx.x];

    Nd4jLong sharedInd(2 * threadIdx.x), leftArrInd, rightArrInd, step;
    T xLeft, xRight;

    for (uint i = 0; i < numTadChunks; ++i) {

        leftArrInd  = sharedInd + i * blockDim2;
        rightArrInd = leftArrInd + 1;

        if(reverse) {
            if(rightArrInd < tadLen) {
                rightArrInd = tadLen - 1 - rightArrInd;
                leftArrInd  = tadLen - 1 - leftArrInd;
            }
            else if(leftArrInd < tadLen)
                leftArrInd = tadLen - 1 - leftArrInd;
        }

        if(leftArrInd < tadLen)
            shared[sharedInd] = xLeft = xTad[shape::getIndexOffset(leftArrInd, xTadShapeInfo, tadLen)];
        // else
        //     shared[sharedInd] = (op == scalar::Add) ? 0 : 1;

        if(rightArrInd < tadLen)
            shared[sharedInd + 1] = xRight = xTad[shape::getIndexOffset(rightArrInd, xTadShapeInfo, tadLen)];
        // else
        //     shared[sharedInd + 1] = (op == scalar::Add) ? 0 : 1;


        step = 1;

        for (uint d = blockDim.x; d > 0; d /= 2) {

            __syncthreads();
            if(threadIdx.x < d) {
                uint left  = step * (sharedInd + 1) - 1;
                uint right = step * (sharedInd + 2) - 1;
                shared[right] = (op == scalar::Add) ? (shared[right] + shared[left]) : (shared[right] * shared[left]);
            }
            step *= 2;
        }

        if (threadIdx.x == 0)
            shared[blockDim2 - 1] = (op == scalar::Add) ? 0 : 1;

        for (uint d = 1; d < blockDim2; d *= 2) {

            step /= 2;

            __syncthreads();
            if(threadIdx.x < d) {
                uint left  = step * (sharedInd + 1) - 1;
                uint right = step * (sharedInd + 2) - 1;
                T temp = shared[left];
                shared[left]  = shared[right];
                shared[right] = (op == scalar::Add) ? (shared[right] + temp) : (shared[right] * temp);
            }
        }

        __syncthreads();

        if(leftArrInd < tadLen) {
            T result = shared[sharedInd];
            if(!exclusive)
                result = (op == scalar::Add) ? result + xLeft : result * xLeft;
            if(i > 0)
                result = (op == scalar::Add) ? result + lastElemInChunk : result * lastElemInChunk;
            zTad[shape::getIndexOffset(leftArrInd, zTadShapeInfo, tadLen)] = result;
        }

        if(rightArrInd < tadLen) {
            T result = shared[sharedInd + 1];
            if(!exclusive)
                result = (op == scalar::Add) ? result + xRight : result * xRight;
            if(i > 0)
                result = (op == scalar::Add) ? result + lastElemInChunk : result * lastElemInChunk;
            if(i < numTadChunks - 1 && threadIdx.x == blockDim.x - 1)    // last element in chunk
                lastElemInChunk = !exclusive ? result : (op == scalar::Add) ? result + xRight : result * xRight;
            zTad[shape::getIndexOffset(rightArrInd, zTadShapeInfo, tadLen)] = result;
        }
    }
}

///////////////////////////////////////////////////////////////////
template<typename X>
static void prefixPerBlockCudaLauncher(const int blocksPerGrid, const int threadsPerBlock, const int sharedMem, const cudaStream_t *stream,
                                        scalar::Ops op,
                                        const void* vx, const Nd4jLong* xTadShapeInfo, const Nd4jLong* xTadOffsets,
                                              void* vz, const Nd4jLong* zTadShapeInfo, const Nd4jLong* zTadOffsets,
                                        const Nd4jLong numTads, const Nd4jLong tadLen,
                                        const bool exclusive, const bool reverse) {

    prefixPerBlockCuda<X><<<blocksPerGrid, threadsPerBlock, sharedMem, *stream>>>(op, vx, xTadShapeInfo, xTadOffsets, vz, zTadShapeInfo, zTadOffsets, numTads, tadLen, exclusive, reverse);
}

///////////////////////////////////////////////////////////////////
void prefix(nd4j::LaunchContext * context, scalar::Ops op, const NDArray* x, NDArray* z, const std::vector<int>& dims, bool exclusive, bool reverse) {

    auto packX = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(x->getShapeInfo(), dims);
    auto packZ = nd4j::ConstantTadHelper::getInstance()->tadForDimensions(z->getShapeInfo(), dims);

    const Nd4jLong numTads = packX.numberOfTads();
    const Nd4jLong tadLen = x->lengthOf() / numTads;

    const int threadsPerBlock = MAX_NUM_THREADS / 2;
    const int blocksPerGrid = numTads;
    const int sharedMem = 2 * threadsPerBlock * x->sizeOfT() + 128;

    PointersManager manager(context, "prefix");

    NDArray::prepareSpecialUse({z}, {x});
    BUILD_SINGLE_SELECTOR(x->dataType(), prefixPerBlockCudaLauncher, (blocksPerGrid, threadsPerBlock, sharedMem, context->getCudaStream(), op, x->getSpecialBuffer(), packX.platformShapeInfo(), packX.platformOffsets(), z->specialBuffer(), packZ.platformShapeInfo(), packZ.platformOffsets(), numTads, tadLen, exclusive, reverse), LIBND4J_TYPES);
    NDArray::registerSpecialUse({z}, {x});

    manager.synchronize();
}

///////////////////////////////////////////////////////////////////
void prefix(nd4j::LaunchContext * context, scalar::Ops op, const NDArray* x, NDArray* z, bool exclusive, bool reverse) {
    prefix(context, op, x, z, {}, exclusive, reverse);
}

}
}
}