cavis/libnd4j/include/ops/declarable/helpers/cuda/addBias.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)
//


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

namespace sd    {
namespace ops     {
namespace helpers {

//////////////////////////////////////////////////////////////////////
template<typename X, typename Y>
__global__ static void addBiasCuda( const void* vx, const Nd4jLong* xShapeInfo,
                                    const void* vy, const Nd4jLong* yShapeInfo,
                                          void* vz, const Nd4jLong* zShapeInfo,
                                    const bool isNCHW) {

    // bias [oC]

    // if(input_rank == 4)
        // input and output have same shapes: [bS, oH, oW, oC] (NHWC) or [bS, oC, oH, oW] (NCHW)
    // if(input_rank == 5)
        // input and output have same shapes: [bS, oD, oH, oW, oC] (NHWC) or [bS, oD, oC, oH, oW] (NCHW)

    const X* x = reinterpret_cast<const X*>(vx);
    const Y* y = reinterpret_cast<const Y*>(vy);
          X* z = reinterpret_cast<X*>(vz);

    __shared__ int rank, channelPosition, posOfNonUnityDim;
    __shared__ Nd4jLong len, *sharedMem;
    __shared__ bool xzSameOffsets, xzAreSame;

    if (threadIdx.x == 0) {

        extern __shared__ unsigned char shmem[];
        sharedMem = reinterpret_cast<Nd4jLong*>(shmem);

        rank = shape::rank(xShapeInfo);     // xRank == zRank
        xzSameOffsets = shape::haveSameShapeAndStrides(xShapeInfo, zShapeInfo);
        len = shape::length(xShapeInfo);
        channelPosition = isNCHW ? 1 : rank - 1;        // second or last
        xzAreSame = x == z;

        shape::isCommonVector(yShapeInfo, posOfNonUnityDim);
    }
    __syncthreads();

    auto coords = sharedMem + threadIdx.x * rank;

    for (Nd4jLong i = blockIdx.x * blockDim.x + threadIdx.x; i < len; i += blockDim.x * gridDim.x) {

        shape::index2coords(i, xShapeInfo, coords);

        const auto xOffsets = shape::getOffset(xShapeInfo, coords);
        const auto zOffsets = xzSameOffsets ? xOffsets : shape::getOffset(zShapeInfo, coords);
        const auto yOffsets = coords[channelPosition] * shape::stride(yShapeInfo)[posOfNonUnityDim];

        if(xzAreSame)
            z[zOffsets] += static_cast<X>(y[yOffsets]);
        else
            z[zOffsets] = x[xOffsets] + static_cast<X>(y[yOffsets]);
    }
}

//////////////////////////////////////////////////////////////////////////
template<typename X, typename Y>
static void addBiasCudaLauncher(const int blocksPerGrid, const int threadsPerBlock, const int sharedMem, const cudaStream_t *stream,
                                         const void* vx, const Nd4jLong* xShapeInfo,
                                         const void* vy, const Nd4jLong* yShapeInfo,
                                               void* vz, const Nd4jLong* zShapeInfo,
                                         const bool isNCHW) {

    addBiasCuda<X,Y><<<blocksPerGrid, threadsPerBlock, sharedMem, *stream>>>(vx, xShapeInfo, vy, yShapeInfo, vz, zShapeInfo, isNCHW);
}

template<typename X, typename Y>
__global__ static void addBias2DCuda( const void* vx,
                                        const void* vy,
                                        void* vz,
                                        uint32_t blocks, uint32_t length) {

    auto y = reinterpret_cast<const Y*>(vy);

    for (uint32_t b = blockIdx.x; b < blocks; b += gridDim.x) {
        auto x = reinterpret_cast<const X*>(vx) + length * b;
        auto z = reinterpret_cast<X*>(vz) + length * b;

        for (uint32_t e = threadIdx.x; e < length; e += blockDim.x) {
            z[e] = x[e] + y[e];
        }
    }
}

template<typename X, typename Y>
static void addBias2DCudaLauncher(const cudaStream_t *stream, const void* vx,
                                  const void* vy,
                                  void* vz,
                                  uint32_t blocks, uint32_t length) {

    addBias2DCuda<X,Y><<<256, 1024, 128, *stream>>>(vx, vy, vz, blocks, length);
}

//////////////////////////////////////////////////////////////////////////
void addBias(sd::graph::Context& block, const NDArray& input, const NDArray& bias, NDArray& output, const bool isNCHW) {

    PointersManager manager(block.launchContext(), "addBias");
    NDArray::prepareSpecialUse({&output}, {&input, &bias});

    if (input.rankOf() == 2 && bias.rankOf() == 1 && input.ordering() == 'c' && output.ordering() == 'c' && input.ews() == 1 && bias.ews() == 1 && input.sizeAt(1) == bias.sizeAt(0)) {
        BUILD_DOUBLE_SELECTOR(input.dataType(), bias.dataType(), addBias2DCudaLauncher,
                              (block.launchContext()->getCudaStream(), input.specialBuffer(), bias.specialBuffer(), output.specialBuffer(), input.sizeAt(0), bias.sizeAt(0)),
                              FLOAT_TYPES, FLOAT_TYPES);
    } else {
        // default case
        const int threadsPerBlock = MAX_NUM_THREADS / 4;
        const int blocksPerGrid = (input.lengthOf() + threadsPerBlock - 1) / threadsPerBlock;
        const int sharedMem = input.rankOf() * sizeof(Nd4jLong) * threadsPerBlock + 128;


        BUILD_DOUBLE_SELECTOR(input.dataType(), bias.dataType(), addBiasCudaLauncher,
                              (blocksPerGrid, threadsPerBlock, sharedMem, block.launchContext()->getCudaStream(), input.specialBuffer(), input.specialShapeInfo(), bias.specialBuffer(), bias.specialShapeInfo(), output.specialBuffer(), output.specialShapeInfo(), isNCHW),
                              FLOAT_TYPES, FLOAT_TYPES);
    }
    NDArray::registerSpecialUse({&output}, {&input, &bias});
    manager.synchronize();
}

}
}
}