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

#ifndef CUDA_LAMBDA_HELPER
#define CUDA_LAMBDA_HELPER

#include <cuda.h>
#include <cuda_runtime.h>
#include <helpers/shape.h>
#include <system/op_boilerplate.h>



static sd::LongType SD_DEVICE  length(const sd::LongType *shapeInfo) { return shape::length(shapeInfo); }

template <typename T, typename Lambda>
static SD_KERNEL void lambdaKernel(const void *vx, const sd::LongType *xShapeInfo, void *vz,
                                   const sd::LongType *zShapeInfo, Lambda lambda);
template <typename T, typename Lambda>
static SD_KERNEL void lambdaIndexedKernel(const void *vx, const sd::LongType *xShapeInfo, void *vz,
                                          const sd::LongType *zShapeInfo, Lambda lambda);
template <typename T, typename Lambda>
static SD_KERNEL void lambdaIndexedPairwiseKernel(const void *vx, const sd::LongType *xShapeInfo, const void *vy,
                                                  const sd::LongType *yShapeInfo, void *vz,
                                                  const sd::LongType *zShapeInfo, Lambda lambda);
template <typename T, typename Lambda>
static SD_KERNEL void lambdaPairwiseKernel(const void *vx, const sd::LongType *xShapeInfo, const void *vy,
                                           const sd::LongType *yShapeInfo, void *vz, const sd::LongType *zShapeInfo,
                                           Lambda lambda);
template <typename T, typename Lambda>
static SD_KERNEL void lambdaPairwiseKernel(const void *scalarPtr, const void *vx, const sd::LongType *xShapeInfo, void *vz, const sd::LongType *zShapeInfo,
                                           Lambda lambda);
template <typename T, typename Lambda>
static SD_KERNEL void lambdaTriplewiseKernel(const void *vw, const sd::LongType *wShapeInfo, const void *vx,
                                             const sd::LongType *xShapeInfo, const void *vy,
                                             const sd::LongType *yShapeInfo, void *vz, const sd::LongType *zShapeInfo,
                                             Lambda lambda);

template <typename T>
class LambdaHelper {
 public:
  template <typename Lambda>
  SD_INLINE static void lambdaLauncher(cudaStream_t *stream, const void *vx, const sd::LongType *xShapeInfo, void *vz,
                                       const sd::LongType *zShapeInfo, Lambda lambda) {
    lambdaKernel<T, Lambda><<<256, 512, 1024, *stream>>>(vx, xShapeInfo, vz, zShapeInfo, lambda);
    auto err = cudaStreamSynchronize(*stream);
    if (err != 0) THROW_EXCEPTION("NDArray::applyLambda execution failed");
  }

  template <typename Lambda>
  SD_INLINE static void lambdaIndexedLauncher(cudaStream_t *stream, const void *vx, const sd::LongType *xShapeInfo,
                                              void *vz, const sd::LongType *zShapeInfo, Lambda lambda) {
    lambdaIndexedKernel<T, Lambda><<<256, 512, 1024, *stream>>>(vx, xShapeInfo, vz, zShapeInfo, lambda);
    auto err = cudaStreamSynchronize(*stream);
    if (err != 0) THROW_EXCEPTION("NDArray::applyIndexedLambda execution failed");
  }

  template <typename Lambda>
  SD_INLINE static void lambdaPairwiseLauncher(cudaStream_t *stream, const void *vx, const sd::LongType *xShapeInfo, bool otherIsScalar,
                                               const void *vy, const sd::LongType *yShapeInfo, void *vz,
                                               const sd::LongType *zShapeInfo, Lambda lambda) {
    if (otherIsScalar) {
      lambdaPairwiseKernel<T, Lambda><<<256, 512, 1024, *stream>>>(vy, vx, xShapeInfo, vz, zShapeInfo, lambda);
    } else {
      lambdaPairwiseKernel<T, Lambda>
      <<<256, 512, 1024, *stream>>>(vx, xShapeInfo, vy, yShapeInfo, vz, zShapeInfo, lambda);
    }
    auto err = cudaStreamSynchronize(*stream);
    if (err != 0) THROW_EXCEPTION("NDArray::applyPairwiseLambda execution failed");
  }

  template <typename Lambda>
  SD_INLINE static void lambdaIndexedPairwiseLauncher(cudaStream_t *stream, const void *vx,
                                                      const sd::LongType *xShapeInfo, const void *vy,
                                                      const sd::LongType *yShapeInfo, void *vz,
                                                      const sd::LongType *zShapeInfo, Lambda lambda) {
    lambdaIndexedPairwiseKernel<T, Lambda>
    <<<256, 512, 1024, *stream>>>(vx, xShapeInfo, vy, yShapeInfo, vz, zShapeInfo, lambda);
    auto err = cudaStreamSynchronize(*stream);
    if (err != 0) THROW_EXCEPTION("NDArray::applyIndexedPairwiseLambda execution failed");
  }

  template <typename Lambda>
  SD_INLINE static void lambdaTriplewiseLauncher(cudaStream_t *stream, const void *vw, const sd::LongType *wShapeInfo,
                                                 const void *vx, const sd::LongType *xShapeInfo, const void *vy,
                                                 const sd::LongType *yShapeInfo, void *vz,
                                                 const sd::LongType *zShapeInfo, Lambda lambda) {
    lambdaTriplewiseKernel<T, Lambda>
    <<<256, 512, 1024, *stream>>>(vw, wShapeInfo, vx, xShapeInfo, vy, yShapeInfo, vz, zShapeInfo, lambda);
    auto err = cudaStreamSynchronize(*stream);
    if (err != 0) THROW_EXCEPTION("NDArray::applyTriplewiseLambda execution failed");
  }
};

////////////////////////////////////////////////////////////////////////
template <typename T, typename Lambda>
static SD_KERNEL void lambdaKernel(const void *vx, const sd::LongType *xShapeInfo, void *vz,
                                   const sd::LongType *zShapeInfo, Lambda lambda) {
  auto x = reinterpret_cast<const T *>(vx);
  auto z = reinterpret_cast<T *>(vz);

  auto zLength = length(zShapeInfo);

  auto tid = threadIdx.x + blockIdx.x * blockDim.x;
  __shared__ sd::LongType  xRank;
  __shared__ sd::LongType *xShape;
  __shared__ sd::LongType *xStride;
  __shared__ sd::LongType  zRank;
  __shared__ sd::LongType *zShape;
  __shared__ sd::LongType *zStride;
  if(threadIdx.x == 0) {
    xRank = shape::rank(xShapeInfo);
    xShape = shape::shapeOf(xShapeInfo);
    xStride = shape::stride(xShapeInfo);
    zRank = shape::rank(zShapeInfo);
    zShape = shape::shapeOf(zShapeInfo);
    zStride = shape::stride(zShapeInfo);
  }

  __syncthreads();

  for (sd::LongType e = tid; e < zLength; e += blockDim.x * gridDim.x) {
    sd::LongType xCoords[SD_MAX_RANK];
    sd::LongType zCoords[SD_MAX_RANK];
    sd::LongType xOffset;
    sd::LongType zOffset;

    INDEX2COORDS(e, xRank, xShape, xCoords);
    COORDS2INDEX(xRank,xStride, xCoords, xOffset);
    INDEX2COORDS(e, zRank, zShape, zCoords);
    COORDS2INDEX(zRank, zStride, zCoords, zOffset);

    z[zOffset] = lambda(x[xOffset]);
  }
}
////////////////////////////////////////////////////////////////////////
template <typename T, typename Lambda>
static SD_KERNEL void lambdaIndexedKernel(const void *vx, const sd::LongType *xShapeInfo, void *vz,
                                          const sd::LongType *zShapeInfo, Lambda lambda) {
  auto x = reinterpret_cast<const T *>(vx);
  auto z = reinterpret_cast<T *>(vz);

  auto zLength = length(zShapeInfo);

  auto tid = threadIdx.x + blockIdx.x * blockDim.x;

  __shared__ sd::LongType  xRank;
  __shared__ sd::LongType *xShape;
  __shared__ sd::LongType *xStride;
  __shared__ sd::LongType  zRank;
  __shared__ sd::LongType *zShape;
  __shared__ sd::LongType *zStride;
  if(threadIdx.x == 0) {
    xRank = shape::rank(xShapeInfo);
    xShape = shape::shapeOf(xShapeInfo);
    xStride = shape::stride(xShapeInfo);
    zRank = shape::rank(zShapeInfo);
    zShape = shape::shapeOf(zShapeInfo);
    zStride = shape::stride(zShapeInfo);
  }

  for (sd::LongType e = tid; e < zLength; e += blockDim.x * gridDim.x) {
    sd::LongType xCoords[SD_MAX_RANK];
    sd::LongType zCoords[SD_MAX_RANK];
    sd::LongType xOffset;
    sd::LongType zOffset;

    INDEX2COORDS(e, xRank, xShape, xCoords);
    COORDS2INDEX(xRank,xStride, xCoords, xOffset);
    INDEX2COORDS(e, zRank, zShape, zCoords);
    COORDS2INDEX(zRank, zStride, zCoords, zOffset);

    z[zOffset] = lambda(e, x[xOffset]);
  }
}

////////////////////////////////////////////////////////////////////////
template <typename T, typename Lambda>
static SD_KERNEL void lambdaIndexedPairwiseKernel(const void *vx, const sd::LongType *xShapeInfo, const void *vy,
                                                  const sd::LongType *yShapeInfo, void *vz,
                                                  const sd::LongType *zShapeInfo, Lambda lambda) {
  auto x = reinterpret_cast<const T *>(vx);
  auto y = reinterpret_cast<const T *>(vy);
  auto z = reinterpret_cast<T *>(vz);

  auto zLength = length(zShapeInfo);

  auto tid = threadIdx.x + blockIdx.x * blockDim.x;

  __shared__ sd::LongType  xRank;
  __shared__ sd::LongType *xShape;
  __shared__ sd::LongType *xStride;
  __shared__ sd::LongType  yRank;
  __shared__ sd::LongType *yShape;
  __shared__ sd::LongType *yStride;
  __shared__ sd::LongType  zRank;
  __shared__ sd::LongType *zShape;
  __shared__ sd::LongType *zStride;
  if(threadIdx.x == 0) {
    xRank = shape::rank(xShapeInfo);
    xShape = shape::shapeOf(xShapeInfo);
    xStride = shape::stride(xShapeInfo);
    yRank = shape::rank(yShapeInfo);
    yShape = shape::shapeOf(yShapeInfo);
    yStride = shape::stride(yShapeInfo);
    zRank = shape::rank(zShapeInfo);
    zShape = shape::shapeOf(zShapeInfo);
    zStride = shape::stride(zShapeInfo);
  }

  __syncthreads();


  for (sd::LongType e = tid; e < zLength; e += blockDim.x * gridDim.x) {
    sd::LongType xCoords[SD_MAX_RANK];
    sd::LongType yCoords[SD_MAX_RANK];
    sd::LongType zCoords[SD_MAX_RANK];
    sd::LongType xOffset;
    sd::LongType yOffset;
    sd::LongType zOffset;

    INDEX2COORDS(e, xRank,xShape, xCoords);
    COORDS2INDEX(xRank, xStride, xCoords, xOffset);
    INDEX2COORDS(e, yRank, yShape, yCoords);
    COORDS2INDEX(yRank, yStride, yCoords, yOffset);
    INDEX2COORDS(e, zRank, zShape, zCoords);
    COORDS2INDEX(zRank,zStride, zCoords, zOffset);

    z[zOffset] = lambda(e, x[xOffset], y[yOffset]);
  }
}

////////////////////////////////////////////////////////////////////////
template <typename T, typename Lambda>
static SD_KERNEL void lambdaPairwiseKernel(const void *vx, const sd::LongType *xShapeInfo, const void *vy,
                                           const sd::LongType *yShapeInfo, void *vz, const sd::LongType *zShapeInfo,
                                           Lambda lambda) {
  auto x = reinterpret_cast<const T *>(vx);
  auto y = reinterpret_cast<const T *>(vy);
  auto z = reinterpret_cast<T *>(vz);

  auto zLength = length(zShapeInfo);

  auto tid = threadIdx.x + blockIdx.x * blockDim.x;

  __shared__ sd::LongType  xRank;
  __shared__ sd::LongType *xShape;
  __shared__ sd::LongType *xStride;
  __shared__ sd::LongType  yRank;
  __shared__ sd::LongType *yShape;
  __shared__ sd::LongType *yStride;
  __shared__ sd::LongType  zRank;
  __shared__ sd::LongType *zShape;
  __shared__ sd::LongType *zStride;
  if(threadIdx.x == 0) {
    xRank = shape::rank(xShapeInfo);
    xShape = shape::shapeOf(xShapeInfo);
    xStride = shape::stride(xShapeInfo);
    yRank = shape::rank(yShapeInfo);
    yShape = shape::shapeOf(yShapeInfo);
    yStride = shape::stride(yShapeInfo);
    zRank = shape::rank(zShapeInfo);
    zShape = shape::shapeOf(zShapeInfo);
    zStride = shape::stride(zShapeInfo);
  }

  __syncthreads();


  for (sd::LongType e = tid; e < zLength; e += blockDim.x * gridDim.x) {
    sd::LongType xCoords[SD_MAX_RANK];
    sd::LongType yCoords[SD_MAX_RANK];
    sd::LongType zCoords[SD_MAX_RANK];
    sd::LongType xOffset;
    sd::LongType yOffset;
    sd::LongType zOffset;

    INDEX2COORDS(e, xRank, xShape, xCoords);
    COORDS2INDEX(xRank, xStride, xCoords, xOffset);
    INDEX2COORDS(e, yRank, yShape, yCoords);
    COORDS2INDEX(yRank, yStride, yCoords, yOffset);
    INDEX2COORDS(e, zRank, zShape, zCoords);
    COORDS2INDEX(zRank,zStride, zCoords, zOffset);

    z[zOffset] = lambda(x[xOffset], y[yOffset]);
  }
}
///////////////////////////////////////////////////////////////////////
template <typename T, typename Lambda>
static SD_KERNEL void lambdaPairwiseKernel(const void *scalarPtr, const void *vx, const sd::LongType *xShapeInfo,
                                           void *vz, const sd::LongType *zShapeInfo, Lambda lambda) {
  auto x = reinterpret_cast<const T *>(vx);
  auto y = reinterpret_cast<const T *>(scalarPtr);
  auto z = reinterpret_cast<T *>(vz);

  auto yVal = *y;

  auto zLength = length(zShapeInfo);

  auto tid = threadIdx.x + blockIdx.x * blockDim.x;

  __shared__ sd::LongType  xRank;
  __shared__ sd::LongType *xShape;
  __shared__ sd::LongType *xStride;
  __shared__ sd::LongType  yRank;
  __shared__ sd::LongType *yShape;
  __shared__ sd::LongType *yStride;
  __shared__ sd::LongType  zRank;
  __shared__ sd::LongType *zShape;
  __shared__ sd::LongType *zStride;
  if(threadIdx.x == 0) {
    xRank = shape::rank(xShapeInfo);
    xShape = shape::shapeOf(xShapeInfo);
    xStride = shape::stride(xShapeInfo);
    yRank = shape::rank(yShapeInfo);
    yShape = shape::shapeOf(yShapeInfo);
    yStride = shape::stride(yShapeInfo);
    zRank = shape::rank(zShapeInfo);
    zShape = shape::shapeOf(zShapeInfo);
    zStride = shape::stride(zShapeInfo);
  }

  __syncthreads();



  for (sd::LongType e = tid; e < zLength; e += blockDim.x * gridDim.x) {
    sd::LongType xCoords[SD_MAX_RANK];
    sd::LongType zCoords[SD_MAX_RANK];
    sd::LongType xOffset;
    sd::LongType zOffset;

    INDEX2COORDS(e,xRank,xShape, xCoords);
    COORDS2INDEX(xRank, xStride, xCoords, xOffset);
    INDEX2COORDS(e, zRank, zShape, zCoords);
    COORDS2INDEX(zRank, zStride, zCoords, zOffset);

    z[zOffset] = lambda(x[xOffset], yVal);
  }
}
////////////////////////////////////////////////////////////////////////
template <typename T, typename Lambda>
static SD_KERNEL void lambdaTriplewiseKernel(const void *vw, const sd::LongType *wShapeInfo, const void *vx,
                                             const sd::LongType *xShapeInfo, const void *vy,
                                             const sd::LongType *yShapeInfo, void *vz, const sd::LongType *zShapeInfo,
                                             Lambda lambda) {
  auto w = reinterpret_cast<const T *>(vw);
  auto x = reinterpret_cast<const T *>(vx);
  auto y = reinterpret_cast<const T *>(vy);
  auto z = reinterpret_cast<T *>(vz);

  auto zLength = length(zShapeInfo);

  auto tid = threadIdx.x + blockIdx.x * blockDim.x;
  __shared__ sd::LongType  xRank;
  __shared__ sd::LongType *xShape;
  __shared__ sd::LongType *xStride;
  __shared__ sd::LongType  yRank;
  __shared__ sd::LongType *yShape;
  __shared__ sd::LongType *yStride;
  __shared__ sd::LongType  zRank;
  __shared__ sd::LongType *zShape;
  __shared__ sd::LongType *zStride;
  __shared__ sd::LongType wRank;
  __shared__ sd::LongType *wShape;
  __shared__ sd::LongType *wStride;

  if(threadIdx.x == 0) {
    xRank = shape::rank(xShapeInfo);
    xShape = shape::shapeOf(xShapeInfo);
    xStride = shape::stride(xShapeInfo);
    yRank = shape::rank(yShapeInfo);
    yShape = shape::shapeOf(yShapeInfo);
    yStride = shape::stride(yShapeInfo);
    zRank = shape::rank(zShapeInfo);
    zShape = shape::shapeOf(zShapeInfo);
    zStride = shape::stride(zShapeInfo);
    wRank = shape::rank(wShapeInfo);
    wShape = shape::shapeOf(wShapeInfo);
    wStride = shape::stride(wShapeInfo);
  }

  __syncthreads();


  for (sd::LongType e = tid; e < zLength; e += blockDim.x * gridDim.x) {
    sd::LongType wCoords[SD_MAX_RANK];
    sd::LongType xCoords[SD_MAX_RANK];
    sd::LongType yCoords[SD_MAX_RANK];
    sd::LongType zCoords[SD_MAX_RANK];
    sd::LongType wOffset;
    sd::LongType xOffset;
    sd::LongType yOffset;
    sd::LongType zOffset;

    INDEX2COORDS(e, wRank, wShape, wCoords);
    COORDS2INDEX(wRank, wStride, wCoords, wOffset);
    INDEX2COORDS(e, xRank,xShape, xCoords);
    COORDS2INDEX(xRank, xStride, xCoords, xOffset);
    INDEX2COORDS(e, yRank, yStride, yCoords);
    COORDS2INDEX(yRank,yStride, yCoords, yOffset);
    INDEX2COORDS(e, zRank, zShape, zCoords);
    COORDS2INDEX(zRank, zStride, zCoords, zOffset);

    z[zOffset] = lambda(w[wOffset], x[xOffset], y[yOffset]);
  }
}
#endif

//////////////////////////////////////////////////////////////////////////
template <typename Lambda>
void NDArray::applyLambda(Lambda func, NDArray *target) {
  auto dtype = this->dataType();

  if (dtype != target->dataType()) THROW_EXCEPTION("NDArray::applyLambda X/Z data types must be the same");
  prepareSpecialUse({&target}, {this});
  BUILD_SINGLE_SELECTOR(
      dtype, LambdaHelper,
      ::lambdaLauncher(this->_context->getCudaStream(), this->specialBuffer(), this->specialShapeInfo(),
                       target->specialBuffer(), target->specialShapeInfo(), func),
      SD_COMMON_TYPES);
  registerSpecialUse({&target}, {this});
}

//////////////////////////////////////////////////////////////////////////
template <typename Lambda>
void NDArray::applyPairwiseLambda(NDArray *other, Lambda func, NDArray *target) {
  auto dtype = this->dataType();

  if (dtype != target->dataType() || dtype != other->dataType())
    THROW_EXCEPTION("NDArray::applyPairwiseLambda X/Y/Z data types must be the same");
  bool otherIsScalar = other->isScalar();
  prepareSpecialUse({&target}, {this, &other});
  BUILD_SINGLE_SELECTOR(
      dtype, LambdaHelper,
      ::lambdaPairwiseLauncher(this->_context->getCudaStream(), this->specialBuffer(), this->specialShapeInfo(), otherIsScalar,
                               other->specialBuffer(), other->specialShapeInfo(), target->specialBuffer(),
                               target->specialShapeInfo(), func),
      SD_COMMON_TYPES);
  registerSpecialUse({&target}, {this, &other});
}

//////////////////////////////////////////////////////////////////////////
template <typename Lambda>
void NDArray::applyIndexedLambda(Lambda func, NDArray *target) {
  auto dtype = this->dataType();
  if (dtype != target->dataType())
    THROW_EXCEPTION("NDArray::applyIndexedLambda X/Z data types must be the same");

  prepareSpecialUse({&target}, {this});
  BUILD_SINGLE_SELECTOR(
      dtype, LambdaHelper,
      ::lambdaIndexedLauncher(this->_context->getCudaStream(), this->specialBuffer(), this->specialShapeInfo(),
                              target->specialBuffer(), target->specialShapeInfo(), func),
      SD_COMMON_TYPES);
  registerSpecialUse({&target}, {this});
}

//////////////////////////////////////////////////////////////////////////
template <typename Lambda>
void NDArray::applyIndexedPairwiseLambda(NDArray *other, Lambda func, NDArray *target) {
  auto dtype = this->dataType();
  if (dtype != target->dataType() || dtype != other->dataType())
    THROW_EXCEPTION("NDArray::applyIndexedPairwiseLambda X/Y/Z data types must be the same");

  prepareSpecialUse({&target}, {this, &other});
  BUILD_SINGLE_SELECTOR(
      dtype, LambdaHelper,
      ::lambdaIndexedPairwiseLauncher(this->_context->getCudaStream(), this->specialBuffer(), this->specialShapeInfo(),
                                      other->specialBuffer(), other->specialShapeInfo(), target->specialBuffer(),
                                      target->specialShapeInfo(), func),
      SD_COMMON_TYPES);
  registerSpecialUse({&target}, {this, &other});
}

//////////////////////////////////////////////////////////////////////////
template <typename Lambda>
void NDArray::applyTriplewiseLambda(NDArray *second, NDArray *third, Lambda func, NDArray *target) {
  auto dtype = this->dataType();

  if (dtype != target->dataType() || dtype != second.dataType() || dtype != third.dataType())
    THROW_EXCEPTION("NDArray::applyTriplewiseLambda X/Y/Z data types must be the same");

  prepareSpecialUse({&target}, {this, &second, &third});
  BUILD_SINGLE_SELECTOR(
      dtype, LambdaHelper,
      ::lambdaTriplewiseLauncher(this->_context->getCudaStream(), this->specialBuffer(), this->specialShapeInfo(),
                                 second.specialBuffer(), second.specialShapeInfo(), third.specialBuffer(),
                                 third.specialShapeInfo(), target->specialBuffer(), target->specialShapeInfo(), func),
      SD_COMMON_TYPES);
  registerSpecialUse({&target}, {this, &second, &third});
}