/******************************************************************************
 *
 * Project:  High Performance Image Reprojector
 * Purpose:  Implementation of the GDALWarpKernel class.  Implements the actual
 *           image warping for a "chunk" of input and output imagery already
 *           loaded into memory.
 * Author:   Frank Warmerdam, warmerdam@pobox.com
 *
 ******************************************************************************
 * Copyright (c) 2003, Frank Warmerdam <warmerdam@pobox.com>
 * Copyright (c) 2008-2013, Even Rouault <even dot rouault at mines-paris dot org>
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included
 * in all copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
 * DEALINGS IN THE SOFTWARE.
 ****************************************************************************/

#include "cpl_port.h"
#include "gdalwarper.h"

#include <cfloat>
#include <cmath>
#include <cstddef>
#include <cstdlib>
#include <cstring>

#include <algorithm>
#include <limits>
#include <map>
#include <new>
#include <vector>

#include "cpl_atomic_ops.h"
#include "cpl_conv.h"
#include "cpl_error.h"
#include "cpl_multiproc.h"
#include "cpl_progress.h"
#include "cpl_string.h"
#include "cpl_vsi.h"
#include "cpl_worker_thread_pool.h"
#include "gdal.h"
#include "gdal_alg.h"
#include "gdal_alg_priv.h"
#include "gdalwarpkernel_opencl.h"

// We restrict to 64bit processors because they are guaranteed to have SSE2.
// Could possibly be used too on 32bit, but we would need to check at runtime.
#if defined(__x86_64) || defined(_M_X64)
#include "gdalsse_priv.h"

#if __SSE4_1__
#include <smmintrin.h>
#endif

#if __SSE3__
#include <pmmintrin.h>
#endif

#endif

CPL_CVSID("$Id: gdalwarpkernel.cpp 6e83a202bef2283a9f88fe03fe2c349b56d98243 2019-11-13 12:52:10 +0100 Even Rouault $")

constexpr double BAND_DENSITY_THRESHOLD = 0.0000000001;
constexpr float SRC_DENSITY_THRESHOLD =  0.000000001f;

// #define INSTANTIATE_FLOAT64_SSE2_IMPL

static const int anGWKFilterRadius[] =
{
    0,  // Nearest neighbour
    1,  // Bilinear
    2,  // Cubic Convolution (Catmull-Rom)
    2,  // Cubic B-Spline
    3,  // Lanczos windowed sinc
    0,  // Average
    0,  // Mode
    0,  // Reserved GRA_Gauss=7
    0,  // Max
    0,  // Min
    0,  // Med
    0,  // Q1
    0,  // Q3
};

static double GWKBilinear(double dfX);
static double GWKCubic(double dfX);
static double GWKBSpline(double dfX);
static double GWKLanczosSinc(double dfX);

static const FilterFuncType apfGWKFilter[] =
{
    nullptr,            // Nearest neighbour
    GWKBilinear,     // Bilinear
    GWKCubic,        // Cubic Convolution (Catmull-Rom)
    GWKBSpline,      // Cubic B-Spline
    GWKLanczosSinc,  // Lanczos windowed sinc
    nullptr,  // Average
    nullptr,  // Mode
    nullptr,  // Reserved GRA_Gauss=7
    nullptr,  // Max
    nullptr,  // Min
    nullptr,  // Med
    nullptr,  // Q1
    nullptr,  // Q3
};

// TODO(schwehr): Can we make these functions have a const * const arg?
static double GWKBilinear4Values(double* padfVals);
static double GWKCubic4Values(double* padfVals);
static double GWKBSpline4Values(double* padfVals);
static double GWKLanczosSinc4Values(double* padfVals);

static const FilterFunc4ValuesType apfGWKFilter4Values[] =
{
    nullptr,  // Nearest neighbour
    GWKBilinear4Values,     // Bilinear
    GWKCubic4Values,        // Cubic Convolution (Catmull-Rom)
    GWKBSpline4Values,      // Cubic B-Spline
    GWKLanczosSinc4Values,  // Lanczos windowed sinc
    nullptr,  // Average
    nullptr,  // Mode
    nullptr,  // Reserved GRA_Gauss=7
    nullptr,  // Max
    nullptr,  // Min
    nullptr,  // Med
    nullptr,  // Q1
    nullptr,  // Q3
};

int GWKGetFilterRadius(GDALResampleAlg eResampleAlg)
{
    return anGWKFilterRadius[eResampleAlg];
}

FilterFuncType GWKGetFilterFunc(GDALResampleAlg eResampleAlg)
{
    return apfGWKFilter[eResampleAlg];
}

FilterFunc4ValuesType GWKGetFilterFunc4Values(GDALResampleAlg eResampleAlg)
{
    return apfGWKFilter4Values[eResampleAlg];
}

#ifdef HAVE_OPENCL
static CPLErr GWKOpenCLCase( GDALWarpKernel * );
#endif

static CPLErr GWKGeneralCase( GDALWarpKernel * );
static CPLErr GWKRealCase( GDALWarpKernel *poWK );
static CPLErr GWKNearestNoMasksOrDstDensityOnlyByte( GDALWarpKernel *poWK );
static CPLErr GWKBilinearNoMasksOrDstDensityOnlyByte( GDALWarpKernel *poWK );
static CPLErr GWKCubicNoMasksOrDstDensityOnlyByte( GDALWarpKernel *poWK );
static CPLErr GWKCubicNoMasksOrDstDensityOnlyFloat( GDALWarpKernel *poWK );
#ifdef INSTANTIATE_FLOAT64_SSE2_IMPL
static CPLErr GWKCubicNoMasksOrDstDensityOnlyDouble( GDALWarpKernel *poWK );
#endif
static CPLErr GWKCubicSplineNoMasksOrDstDensityOnlyByte( GDALWarpKernel *poWK );
static CPLErr GWKNearestByte( GDALWarpKernel *poWK );
static CPLErr GWKNearestNoMasksOrDstDensityOnlyShort( GDALWarpKernel *poWK );
static CPLErr GWKBilinearNoMasksOrDstDensityOnlyShort( GDALWarpKernel *poWK );
static CPLErr GWKBilinearNoMasksOrDstDensityOnlyFloat( GDALWarpKernel *poWK );
#ifdef INSTANTIATE_FLOAT64_SSE2_IMPL
static CPLErr GWKBilinearNoMasksOrDstDensityOnlyDouble( GDALWarpKernel *poWK );
#endif
static CPLErr GWKCubicNoMasksOrDstDensityOnlyShort( GDALWarpKernel *poWK );
static CPLErr GWKCubicSplineNoMasksOrDstDensityOnlyShort( GDALWarpKernel *poWK );
static CPLErr GWKNearestShort( GDALWarpKernel *poWK );
static CPLErr GWKNearestNoMasksOrDstDensityOnlyFloat( GDALWarpKernel *poWK );
static CPLErr GWKNearestFloat( GDALWarpKernel *poWK );
static CPLErr GWKAverageOrMode( GDALWarpKernel * );
static CPLErr GWKCubicNoMasksOrDstDensityOnlyUShort( GDALWarpKernel * );
static CPLErr GWKCubicSplineNoMasksOrDstDensityOnlyUShort( GDALWarpKernel * );
static CPLErr GWKBilinearNoMasksOrDstDensityOnlyUShort( GDALWarpKernel * );

/************************************************************************/
/*                           GWKJobStruct                               */
/************************************************************************/

typedef struct _GWKJobStruct GWKJobStruct;

struct _GWKJobStruct
{
    GDALWarpKernel *poWK;
    int             iYMin;
    int             iYMax;
    volatile int   *pnCounter;
    volatile int   *pbStop;
    CPLCond        *hCond;
    CPLMutex       *hCondMutex;
    int           (*pfnProgress)(GWKJobStruct* psJob);
    void           *pTransformerArg;

    void           (*pfnFunc)(void*); // used by GWKRun() to assign the proper pTransformerArg
} ;

struct GWKThreadInitData
{
    GDALTransformerFunc pfnTransformerInit = nullptr;
    void               *pTransformerArgInit = nullptr;

    void               *pTransformerArg = nullptr;
    GIntBig             nThreadId = 0;
};

struct GWKThreadData
{
    CPLWorkerThreadPool* poThreadPool = nullptr;
    GWKJobStruct* pasThreadJob = nullptr;
    CPLCond* hCond = nullptr;
    CPLMutex* hCondMutex = nullptr;
    void* pTransformerArgInput = nullptr; // owned by calling layer. Not to be destroyed
    std::map<GIntBig, void*> mapThreadToTransformerArg{};
};

/************************************************************************/
/*                        GWKProgressThread()                           */
/************************************************************************/

// Return TRUE if the computation must be interrupted.
static int GWKProgressThread( GWKJobStruct* psJob )
{
    CPLAcquireMutex(psJob->hCondMutex, 1.0);
    (*(psJob->pnCounter))++;
    CPLCondSignal(psJob->hCond);
    int bStop = *(psJob->pbStop);
    CPLReleaseMutex(psJob->hCondMutex);

    return bStop;
}

/************************************************************************/
/*                      GWKProgressMonoThread()                         */
/************************************************************************/

// Return TRUE if the computation must be interrupted.
static int GWKProgressMonoThread( GWKJobStruct* psJob )
{
    GDALWarpKernel *poWK = psJob->poWK;
    int nCounter = ++(*(psJob->pnCounter));
    if( !poWK->pfnProgress( poWK->dfProgressBase + poWK->dfProgressScale *
                            (nCounter / static_cast<double>(psJob->iYMax)),
                            "", poWK->pProgress ) )
    {
        CPLError( CE_Failure, CPLE_UserInterrupt, "User terminated" );
        *(psJob->pbStop) = TRUE;
        return TRUE;
    }
    return FALSE;
}

/************************************************************************/
/*                       GWKGenericMonoThread()                         */
/************************************************************************/

static CPLErr GWKGenericMonoThread( GDALWarpKernel *poWK,
                                    void (*pfnFunc) (void *pUserData) )
{
    volatile int bStop = FALSE;
    volatile int nCounter = 0;

    GWKJobStruct sThreadJob;
    sThreadJob.poWK = poWK;
    sThreadJob.pnCounter = &nCounter;
    sThreadJob.iYMin = 0;
    sThreadJob.iYMax = poWK->nDstYSize;
    sThreadJob.pbStop = &bStop;
    sThreadJob.hCond = nullptr;
    sThreadJob.hCondMutex = nullptr;
    sThreadJob.pfnProgress = GWKProgressMonoThread;
    sThreadJob.pTransformerArg = poWK->pTransformerArg;

    pfnFunc(&sThreadJob);

    return !bStop ? CE_None : CE_Failure;
}

/************************************************************************/
/*                     GWKThreadInitTransformer()                       */
/************************************************************************/

static void GWKThreadInitTransformer(void* pData)
{
    GWKThreadInitData* psInitData = static_cast<GWKThreadInitData*>(pData);
    if( psInitData->pTransformerArg == nullptr )
        psInitData->pTransformerArg =
            GDALCloneTransformer(psInitData->pTransformerArgInit);
    if( psInitData->pTransformerArg != nullptr )
    {
        // In case of lazy opening (for example RPCDEM), do a dummy
        // transformation to be sure that the DEM is really opened with the
        // context of this thread.
        double dfX = 0.5;
        double dfY = 0.5;
        double dfZ = 0.0;
        int bSuccess = FALSE;
        CPLPushErrorHandler(CPLQuietErrorHandler);
        psInitData->pfnTransformerInit(psInitData->pTransformerArg, TRUE, 1,
                                  &dfX, &dfY, &dfZ, &bSuccess );
        CPLPopErrorHandler();
    }
    psInitData->nThreadId = CPLGetPID();
}

/************************************************************************/
/*                          GWKThreadsCreate()                          */
/************************************************************************/

void* GWKThreadsCreate( char** papszWarpOptions,
                        GDALTransformerFunc pfnTransformer,
                        void* pTransformerArg )
{
    const char* pszWarpThreads =
        CSLFetchNameValue(papszWarpOptions, "NUM_THREADS");
    if( pszWarpThreads == nullptr )
        pszWarpThreads = CPLGetConfigOption("GDAL_NUM_THREADS", "1");

    int nThreads = 0;
    if( EQUAL(pszWarpThreads, "ALL_CPUS") )
        nThreads = CPLGetNumCPUs();
    else
        nThreads = atoi(pszWarpThreads);
    if( nThreads <= 1 )
        nThreads = 0;
    if( nThreads > 128 )
        nThreads = 128;

    GWKThreadData* psThreadData = new GWKThreadData();
    CPLCond* hCond = nullptr;
    if( nThreads )
        hCond = CPLCreateCond();
    if( nThreads && hCond )
    {
/* -------------------------------------------------------------------- */
/*      Duplicate pTransformerArg per thread.                           */
/* -------------------------------------------------------------------- */
        bool bTransformerCloningSuccess = true;

        psThreadData->hCond = hCond;
        psThreadData->pasThreadJob = static_cast<GWKJobStruct *>(
            VSI_CALLOC_VERBOSE(sizeof(GWKJobStruct), nThreads));
        if( psThreadData->pasThreadJob == nullptr )
        {
            GWKThreadsEnd(psThreadData);
            return nullptr;
        }

        psThreadData->hCondMutex = CPLCreateMutex();
        if( psThreadData->hCondMutex == nullptr )
        {
            GWKThreadsEnd(psThreadData);
            return nullptr;
        }
        CPLReleaseMutex(psThreadData->hCondMutex);

        std::vector<std::unique_ptr<GWKThreadInitData>> apoInitData;
        std::vector<void*> apInitData;
        for( int i = 0; i < nThreads; i++ )
        {
            psThreadData->pasThreadJob[i].hCond = psThreadData->hCond;
            psThreadData->pasThreadJob[i].hCondMutex = psThreadData->hCondMutex;

            std::unique_ptr<GWKThreadInitData> poInitData(new GWKThreadInitData());
            poInitData->pfnTransformerInit = pfnTransformer;
            poInitData->pTransformerArgInit = pTransformerArg;
            if( i == 0 )
                poInitData->pTransformerArg = pTransformerArg;
            else
                poInitData->pTransformerArg = nullptr;
            apoInitData.push_back(std::move(poInitData));
            apInitData.push_back(apoInitData.back().get());
        }

        psThreadData->poThreadPool = new (std::nothrow) CPLWorkerThreadPool();
        if( psThreadData->poThreadPool == nullptr ||
            !psThreadData->poThreadPool->Setup(nThreads,
                                          GWKThreadInitTransformer,
                                          &apInitData[0]) )
        {
            GWKThreadsEnd(psThreadData);
            return nullptr;
        }

        for( int i = 1; i < nThreads; i++ )
        {
            if( apoInitData[i]->pTransformerArg == nullptr )
            {
                CPLDebug("WARP", "Cannot deserialize transformer");
                bTransformerCloningSuccess = false;
                break;
            }
        }

        if( bTransformerCloningSuccess )
        {
            psThreadData->pTransformerArgInput = pTransformerArg;
            for( int i = 0; i < nThreads; i++ )
            {
                const auto nThreadId = apoInitData[i]->nThreadId;
                CPLAssert(psThreadData->mapThreadToTransformerArg.find(nThreadId) ==
                    psThreadData->mapThreadToTransformerArg.end());
                psThreadData->mapThreadToTransformerArg[nThreadId] =
                    apoInitData[i]->pTransformerArg;
            }
        }
        else
        {
            for( int i = 1; i < nThreads; i++ )
            {
                if( apoInitData[i]->pTransformerArg )
                    GDALDestroyTransformer(apoInitData[i]->pTransformerArg);
            }
            CPLFree(psThreadData->pasThreadJob);
            psThreadData->pasThreadJob = nullptr;
            delete psThreadData->poThreadPool;
            psThreadData->poThreadPool = nullptr;

            CPLDebug("WARP", "Cannot duplicate transformer function. "
                     "Falling back to mono-thread computation");
        }
    }

    return psThreadData;
}

/************************************************************************/
/*                             GWKThreadsEnd()                          */
/************************************************************************/

void GWKThreadsEnd( void* psThreadDataIn )
{
    if( psThreadDataIn == nullptr )
        return;

    GWKThreadData* psThreadData = static_cast<GWKThreadData *>(psThreadDataIn);
    if( psThreadData->poThreadPool )
    {
        for( auto& pair: psThreadData->mapThreadToTransformerArg )
        {
            if( pair.second != psThreadData->pTransformerArgInput )
                GDALDestroyTransformer(pair.second);
        }
        delete psThreadData->poThreadPool;
    }
    CPLFree(psThreadData->pasThreadJob);
    if( psThreadData->hCond )
        CPLDestroyCond(psThreadData->hCond);
    if( psThreadData->hCondMutex )
        CPLDestroyMutex(psThreadData->hCondMutex);
    delete psThreadData;
}

/************************************************************************/
/*                         ThreadFuncAdapter()                          */
/************************************************************************/

static void ThreadFuncAdapter(void* pData)
{
    // Assign the pTransformerArg created in the current thread to this job
    // This workarounds the PROJ bug fixed in https://github.com/OSGeo/PROJ/pull/1726
    GWKJobStruct* psJob = static_cast<GWKJobStruct *>(pData);
    const GWKThreadData* psThreadData =
        static_cast<const GWKThreadData*>(psJob->poWK->psThreadData);
    auto oIter = psThreadData->mapThreadToTransformerArg.find(CPLGetPID());
    CPLAssert(oIter != psThreadData->mapThreadToTransformerArg.end());
    psJob->pTransformerArg = oIter->second;
    psJob->pfnFunc(pData);
}


/************************************************************************/
/*                                GWKRun()                              */
/************************************************************************/

static CPLErr GWKRun( GDALWarpKernel *poWK,
                      const char* pszFuncName,
                      void (*pfnFunc) (void *pUserData) )

{
    const int nDstYSize = poWK->nDstYSize;

    CPLDebug("GDAL", "GDALWarpKernel()::%s() "
             "Src=%d,%d,%dx%d Dst=%d,%d,%dx%d",
             pszFuncName,
             poWK->nSrcXOff, poWK->nSrcYOff,
             poWK->nSrcXSize, poWK->nSrcYSize,
             poWK->nDstXOff, poWK->nDstYOff,
             poWK->nDstXSize, poWK->nDstYSize);

    if( !poWK->pfnProgress( poWK->dfProgressBase, "", poWK->pProgress ) )
    {
        CPLError( CE_Failure, CPLE_UserInterrupt, "User terminated" );
        return CE_Failure;
    }

    GWKThreadData* psThreadData =
        static_cast<GWKThreadData*>(poWK->psThreadData);
    if( psThreadData == nullptr || psThreadData->poThreadPool == nullptr )
    {
        return GWKGenericMonoThread(poWK, pfnFunc);
    }

    int nThreads =
        std::min(psThreadData->poThreadPool->GetThreadCount(), nDstYSize / 2);
    // Config option mostly useful for tests to be able to test multithreading
    // with small rasters
    const int nWarpChunkSize = atoi(
        CPLGetConfigOption("WARP_THREAD_CHUNK_SIZE", "65536"));
    if( nWarpChunkSize > 0 )
    {
        GIntBig nChunks =
            static_cast<GIntBig>(nDstYSize) * poWK->nDstXSize / nWarpChunkSize;
        if( nThreads > nChunks )
            nThreads = static_cast<int>(nChunks);
    }
    if( nThreads <= 0 )
        nThreads = 1;

    CPLDebug("WARP", "Using %d threads", nThreads);

    volatile int bStop = FALSE;
    volatile int nCounter = 0;

    CPLAcquireMutex(psThreadData->hCondMutex, 1000);

/* -------------------------------------------------------------------- */
/*      Submit jobs                                                     */
/* -------------------------------------------------------------------- */
    for( int i = 0; i < nThreads; i++ )
    {
        psThreadData->pasThreadJob[i].poWK = poWK;
        psThreadData->pasThreadJob[i].pnCounter = &nCounter;
        psThreadData->pasThreadJob[i].iYMin =
            static_cast<int>((static_cast<GIntBig>(i)) * nDstYSize / nThreads);
        psThreadData->pasThreadJob[i].iYMax =
            static_cast<int>((static_cast<GIntBig>(i + 1)) *
                             nDstYSize / nThreads);
        psThreadData->pasThreadJob[i].pbStop = &bStop;
        if( poWK->pfnProgress != GDALDummyProgress )
            psThreadData->pasThreadJob[i].pfnProgress = GWKProgressThread;
        else
            psThreadData->pasThreadJob[i].pfnProgress = nullptr;
        psThreadData->pasThreadJob[i].pfnFunc = pfnFunc;
        psThreadData->poThreadPool->SubmitJob( ThreadFuncAdapter,
                            static_cast<void*>(&psThreadData->pasThreadJob[i]) );
    }

/* -------------------------------------------------------------------- */
/*      Report progress.                                                */
/* -------------------------------------------------------------------- */
    if( poWK->pfnProgress != GDALDummyProgress )
    {
        while( nCounter < nDstYSize )
        {
            CPLCondWait(psThreadData->hCond, psThreadData->hCondMutex);

            if( !poWK->pfnProgress(
                    poWK->dfProgressBase + poWK->dfProgressScale *
                    (nCounter / static_cast<double>(nDstYSize)),
                    "", poWK->pProgress ) )
            {
                CPLError( CE_Failure, CPLE_UserInterrupt, "User terminated" );
                bStop = TRUE;
                break;
            }
        }
    }

    /* Release mutex before joining threads, otherwise they will dead-lock */
    /* forever in GWKProgressThread() */
    CPLReleaseMutex(psThreadData->hCondMutex);

/* -------------------------------------------------------------------- */
/*      Wait for all jobs to complete.                                  */
/* -------------------------------------------------------------------- */
    psThreadData->poThreadPool->WaitCompletion();

    return !bStop ? CE_None : CE_Failure;
}

/************************************************************************/
/* ==================================================================== */
/*                            GDALWarpKernel                            */
/* ==================================================================== */
/************************************************************************/

/**
 * \class GDALWarpKernel "gdalwarper.h"
 *
 * Low level image warping class.
 *
 * This class is responsible for low level image warping for one
 * "chunk" of imagery.  The class is essentially a structure with all
 * data members public - primarily so that new special-case functions
 * can be added without changing the class declaration.
 *
 * Applications are normally intended to interactive with warping facilities
 * through the GDALWarpOperation class, though the GDALWarpKernel can in
 * theory be used directly if great care is taken in setting up the
 * control data.
 *
 * <h3>Design Issues</h3>
 *
 * The intention is that PerformWarp() would analyze the setup in terms
 * of the datatype, resampling type, and validity/density mask usage and
 * pick one of many specific implementations of the warping algorithm over
 * a continuum of optimization vs. generality.  At one end there will be a
 * reference general purpose implementation of the algorithm that supports
 * any data type (working internally in double precision complex), all three
 * resampling types, and any or all of the validity/density masks.  At the
 * other end would be highly optimized algorithms for common cases like
 * nearest neighbour resampling on GDT_Byte data with no masks.
 *
 * The full set of optimized versions have not been decided but we should
 * expect to have at least:
 *  - One for each resampling algorithm for 8bit data with no masks.
 *  - One for each resampling algorithm for float data with no masks.
 *  - One for each resampling algorithm for float data with any/all masks
 *    (essentially the generic case for just float data).
 *  - One for each resampling algorithm for 8bit data with support for
 *    input validity masks (per band or per pixel).  This handles the common
 *    case of nodata masking.
 *  - One for each resampling algorithm for float data with support for
 *    input validity masks (per band or per pixel).  This handles the common
 *    case of nodata masking.
 *
 * Some of the specializations would operate on all bands in one pass
 * (especially the ones without masking would do this), while others might
 * process each band individually to reduce code complexity.
 *
 * <h3>Masking Semantics</h3>
 *
 * A detailed explanation of the semantics of the validity and density masks,
 * and their effects on resampling kernels is needed here.
 */

/************************************************************************/
/*                     GDALWarpKernel Data Members                      */
/************************************************************************/

/**
 * \var GDALResampleAlg GDALWarpKernel::eResample;
 *
 * Resampling algorithm.
 *
 * The resampling algorithm to use.  One of GRA_NearestNeighbour, GRA_Bilinear,
 * GRA_Cubic, GRA_CubicSpline, GRA_Lanczos, GRA_Average, or GRA_Mode.
 *
 * This field is required. GDT_NearestNeighbour may be used as a default
 * value.
 */

/**
 * \var GDALDataType GDALWarpKernel::eWorkingDataType;
 *
 * Working pixel data type.
 *
 * The datatype of pixels in the source image (papabySrcimage) and
 * destination image (papabyDstImage) buffers.  Note that operations on
 * some data types (such as GDT_Byte) may be much better optimized than other
 * less common cases.
 *
 * This field is required.  It may not be GDT_Unknown.
 */

/**
 * \var int GDALWarpKernel::nBands;
 *
 * Number of bands.
 *
 * The number of bands (layers) of imagery being warped.  Determines the
 * number of entries in the papabySrcImage, papanBandSrcValid,
 * and papabyDstImage arrays.
 *
 * This field is required.
 */

/**
 * \var int GDALWarpKernel::nSrcXSize;
 *
 * Source image width in pixels.
 *
 * This field is required.
 */

/**
 * \var int GDALWarpKernel::nSrcYSize;
 *
 * Source image height in pixels.
 *
 * This field is required.
 */

/**
 * \var double GDALWarpKernel::dfSrcXExtraSize;
 *
 * Number of pixels included in nSrcXSize that are present on the edges of
 * the area of interest to take into account the width of the kernel.
 *
 * This field is required.
 */

/**
 * \var double GDALWarpKernel::dfSrcYExtraSize;
 *
 * Number of pixels included in nSrcYExtraSize that are present on the edges of
 * the area of interest to take into account the height of the kernel.
 *
 * This field is required.
 */

/**
 * \var int GDALWarpKernel::papabySrcImage;
 *
 * Array of source image band data.
 *
 * This is an array of pointers (of size GDALWarpKernel::nBands) pointers
 * to image data.  Each individual band of image data is organized as a single
 * block of image data in left to right, then bottom to top order.  The actual
 * type of the image data is determined by GDALWarpKernel::eWorkingDataType.
 *
 * To access the pixel value for the (x=3, y=4) pixel (zero based) of
 * the second band with eWorkingDataType set to GDT_Float32 use code like
 * this:
 *
 * \code
 *   float dfPixelValue;
 *   int   nBand = 1;  // Band indexes are zero based.
 *   int   nPixel = 3; // Zero based.
 *   int   nLine = 4;  // Zero based.
 *
 *   assert( nPixel >= 0 && nPixel < poKern->nSrcXSize );
 *   assert( nLine >= 0 && nLine < poKern->nSrcYSize );
 *   assert( nBand >= 0 && nBand < poKern->nBands );
 *   dfPixelValue = ((float *) poKern->papabySrcImage[nBand-1])
 *                                  [nPixel + nLine * poKern->nSrcXSize];
 * \endcode
 *
 * This field is required.
 */

/**
 * \var GUInt32 **GDALWarpKernel::papanBandSrcValid;
 *
 * Per band validity mask for source pixels.
 *
 * Array of pixel validity mask layers for each source band.   Each of
 * the mask layers is the same size (in pixels) as the source image with
 * one bit per pixel.  Note that it is legal (and common) for this to be
 * NULL indicating that none of the pixels are invalidated, or for some
 * band validity masks to be NULL in which case all pixels of the band are
 * valid.  The following code can be used to test the validity of a particular
 * pixel.
 *
 * \code
 *   int   bIsValid = TRUE;
 *   int   nBand = 1;  // Band indexes are zero based.
 *   int   nPixel = 3; // Zero based.
 *   int   nLine = 4;  // Zero based.
 *
 *   assert( nPixel >= 0 && nPixel < poKern->nSrcXSize );
 *   assert( nLine >= 0 && nLine < poKern->nSrcYSize );
 *   assert( nBand >= 0 && nBand < poKern->nBands );
 *
 *   if( poKern->papanBandSrcValid != NULL
 *       && poKern->papanBandSrcValid[nBand] != NULL )
 *   {
 *       GUInt32 *panBandMask = poKern->papanBandSrcValid[nBand];
 *       int    iPixelOffset = nPixel + nLine * poKern->nSrcXSize;
 *
 *       bIsValid = panBandMask[iPixelOffset>>5]
 *                  & (0x01 << (iPixelOffset & 0x1f));
 *   }
 * \endcode
 */

/**
 * \var GUInt32 *GDALWarpKernel::panUnifiedSrcValid;
 *
 * Per pixel validity mask for source pixels.
 *
 * A single validity mask layer that applies to the pixels of all source
 * bands.  It is accessed similarly to papanBandSrcValid, but without the
 * extra level of band indirection.
 *
 * This pointer may be NULL indicating that all pixels are valid.
 *
 * Note that if both panUnifiedSrcValid, and papanBandSrcValid are available,
 * the pixel isn't considered to be valid unless both arrays indicate it is
 * valid.
 */

/**
 * \var float *GDALWarpKernel::pafUnifiedSrcDensity;
 *
 * Per pixel density mask for source pixels.
 *
 * A single density mask layer that applies to the pixels of all source
 * bands.  It contains values between 0.0 and 1.0 indicating the degree to
 * which this pixel should be allowed to contribute to the output result.
 *
 * This pointer may be NULL indicating that all pixels have a density of 1.0.
 *
 * The density for a pixel may be accessed like this:
 *
 * \code
 *   float fDensity = 1.0;
 *   int nPixel = 3;  // Zero based.
 *   int nLine = 4;   // Zero based.
 *
 *   assert( nPixel >= 0 && nPixel < poKern->nSrcXSize );
 *   assert( nLine >= 0 && nLine < poKern->nSrcYSize );
 *   if( poKern->pafUnifiedSrcDensity != NULL )
 *     fDensity = poKern->pafUnifiedSrcDensity
 *                                  [nPixel + nLine * poKern->nSrcXSize];
 * \endcode
 */

/**
 * \var int GDALWarpKernel::nDstXSize;
 *
 * Width of destination image in pixels.
 *
 * This field is required.
 */

/**
 * \var int GDALWarpKernel::nDstYSize;
 *
 * Height of destination image in pixels.
 *
 * This field is required.
 */

/**
 * \var GByte **GDALWarpKernel::papabyDstImage;
 *
 * Array of destination image band data.
 *
 * This is an array of pointers (of size GDALWarpKernel::nBands) pointers
 * to image data.  Each individual band of image data is organized as a single
 * block of image data in left to right, then bottom to top order.  The actual
 * type of the image data is determined by GDALWarpKernel::eWorkingDataType.
 *
 * To access the pixel value for the (x=3, y=4) pixel (zero based) of
 * the second band with eWorkingDataType set to GDT_Float32 use code like
 * this:
 *
 * \code
 *   float dfPixelValue;
 *   int   nBand = 1;  // Band indexes are zero based.
 *   int   nPixel = 3; // Zero based.
 *   int   nLine = 4;  // Zero based.
 *
 *   assert( nPixel >= 0 && nPixel < poKern->nDstXSize );
 *   assert( nLine >= 0 && nLine < poKern->nDstYSize );
 *   assert( nBand >= 0 && nBand < poKern->nBands );
 *   dfPixelValue = ((float *) poKern->papabyDstImage[nBand-1])
 *                                  [nPixel + nLine * poKern->nSrcYSize];
 * \endcode
 *
 * This field is required.
 */

/**
 * \var GUInt32 *GDALWarpKernel::panDstValid;
 *
 * Per pixel validity mask for destination pixels.
 *
 * A single validity mask layer that applies to the pixels of all destination
 * bands.  It is accessed similarly to papanUnitifiedSrcValid, but based
 * on the size of the destination image.
 *
 * This pointer may be NULL indicating that all pixels are valid.
 */

/**
 * \var float *GDALWarpKernel::pafDstDensity;
 *
 * Per pixel density mask for destination pixels.
 *
 * A single density mask layer that applies to the pixels of all destination
 * bands.  It contains values between 0.0 and 1.0.
 *
 * This pointer may be NULL indicating that all pixels have a density of 1.0.
 *
 * The density for a pixel may be accessed like this:
 *
 * \code
 *   float fDensity = 1.0;
 *   int   nPixel = 3; // Zero based.
 *   int   nLine = 4;  // Zero based.
 *
 *   assert( nPixel >= 0 && nPixel < poKern->nDstXSize );
 *   assert( nLine >= 0 && nLine < poKern->nDstYSize );
 *   if( poKern->pafDstDensity != NULL )
 *     fDensity = poKern->pafDstDensity[nPixel + nLine * poKern->nDstXSize];
 * \endcode
 */

/**
 * \var int GDALWarpKernel::nSrcXOff;
 *
 * X offset to source pixel coordinates for transformation.
 *
 * See pfnTransformer.
 *
 * This field is required.
 */

/**
 * \var int GDALWarpKernel::nSrcYOff;
 *
 * Y offset to source pixel coordinates for transformation.
 *
 * See pfnTransformer.
 *
 * This field is required.
 */

/**
 * \var int GDALWarpKernel::nDstXOff;
 *
 * X offset to destination pixel coordinates for transformation.
 *
 * See pfnTransformer.
 *
 * This field is required.
 */

/**
 * \var int GDALWarpKernel::nDstYOff;
 *
 * Y offset to destination pixel coordinates for transformation.
 *
 * See pfnTransformer.
 *
 * This field is required.
 */

/**
 * \var GDALTransformerFunc GDALWarpKernel::pfnTransformer;
 *
 * Source/destination location transformer.
 *
 * The function to call to transform coordinates between source image
 * pixel/line coordinates and destination image pixel/line coordinates.
 * See GDALTransformerFunc() for details of the semantics of this function.
 *
 * The GDALWarpKern algorithm will only ever use this transformer in
 * "destination to source" mode (bDstToSrc=TRUE), and will always pass
 * partial or complete scanlines of points in the destination image as
 * input.  This means, among other things, that it is safe to the the
 * approximating transform GDALApproxTransform() as the transformation
 * function.
 *
 * Source and destination images may be subsets of a larger overall image.
 * The transformation algorithms will expect and return pixel/line coordinates
 * in terms of this larger image, so coordinates need to be offset by
 * the offsets specified in nSrcXOff, nSrcYOff, nDstXOff, and nDstYOff before
 * passing to pfnTransformer, and after return from it.
 *
 * The GDALWarpKernel::pfnTransformerArg value will be passed as the callback
 * data to this function when it is called.
 *
 * This field is required.
 */

/**
 * \var void *GDALWarpKernel::pTransformerArg;
 *
 * Callback data for pfnTransformer.
 *
 * This field may be NULL if not required for the pfnTransformer being used.
 */

/**
 * \var GDALProgressFunc GDALWarpKernel::pfnProgress;
 *
 * The function to call to report progress of the algorithm, and to check
 * for a requested termination of the operation.  It operates according to
 * GDALProgressFunc() semantics.
 *
 * Generally speaking the progress function will be invoked for each
 * scanline of the destination buffer that has been processed.
 *
 * This field may be NULL (internally set to GDALDummyProgress()).
 */

/**
 * \var void *GDALWarpKernel::pProgress;
 *
 * Callback data for pfnProgress.
 *
 * This field may be NULL if not required for the pfnProgress being used.
 */

/************************************************************************/
/*                           GDALWarpKernel()                           */
/************************************************************************/

GDALWarpKernel::GDALWarpKernel() :
    papszWarpOptions(nullptr),
    eResample(GRA_NearestNeighbour),
    eWorkingDataType(GDT_Unknown),
    nBands(0),
    nSrcXSize(0),
    nSrcYSize(0),
    dfSrcXExtraSize(0.0),
    dfSrcYExtraSize(0.0),
    papabySrcImage(nullptr),
    papanBandSrcValid(nullptr),
    panUnifiedSrcValid(nullptr),
    pafUnifiedSrcDensity(nullptr),
    nDstXSize(0),
    nDstYSize(0),
    papabyDstImage(nullptr),
    panDstValid(nullptr),
    pafDstDensity(nullptr),
    dfXScale(1.0),
    dfYScale(1.0),
    dfXFilter(0.0),
    dfYFilter(0.0),
    nXRadius(0),
    nYRadius(0),
    nFiltInitX(0),
    nFiltInitY(0),
    nSrcXOff(0),
    nSrcYOff(0),
    nDstXOff(0),
    nDstYOff(0),
    pfnTransformer(nullptr),
    pTransformerArg(nullptr),
    pfnProgress(GDALDummyProgress),
    pProgress(nullptr),
    dfProgressBase(0.0),
    dfProgressScale(1.0),
    padfDstNoDataReal(nullptr),
    psThreadData(nullptr)
{}

/************************************************************************/
/*                          ~GDALWarpKernel()                           */
/************************************************************************/

GDALWarpKernel::~GDALWarpKernel() {}

/************************************************************************/
/*                            PerformWarp()                             */
/************************************************************************/

/**
 * \fn CPLErr GDALWarpKernel::PerformWarp();
 *
 * This method performs the warp described in the GDALWarpKernel.
 *
 * @return CE_None on success or CE_Failure if an error occurs.
 */

CPLErr GDALWarpKernel::PerformWarp()

{
    const CPLErr eErr = Validate();

    if( eErr != CE_None )
        return eErr;

    // See #2445 and #3079.
    if( nSrcXSize <= 0 || nSrcYSize <= 0 )
    {
        if( !pfnProgress( dfProgressBase + dfProgressScale,
                          "", pProgress ) )
        {
            CPLError( CE_Failure, CPLE_UserInterrupt, "User terminated" );
            return CE_Failure;
        }
        return CE_None;
    }

/* -------------------------------------------------------------------- */
/*      Pre-calculate resampling scales and window sizes for filtering. */
/* -------------------------------------------------------------------- */

    dfXScale = static_cast<double>(nDstXSize) / (nSrcXSize - dfSrcXExtraSize);
    dfYScale = static_cast<double>(nDstYSize) / (nSrcYSize - dfSrcYExtraSize);
    if( nSrcXSize >= nDstXSize && nSrcXSize <= nDstXSize + dfSrcXExtraSize )
        dfXScale = 1.0;
    if( nSrcYSize >= nDstYSize && nSrcYSize <= nDstYSize + dfSrcYExtraSize )
        dfYScale = 1.0;
    if( dfXScale < 1.0 )
    {
        double dfXReciprocalScale =  1.0 / dfXScale;
        const int nXReciprocalScale =
            static_cast<int>(dfXReciprocalScale + 0.5);
        if( fabs(dfXReciprocalScale-nXReciprocalScale) < 0.05 )
            dfXScale = 1.0 / nXReciprocalScale;
    }
    if( dfYScale < 1.0 )
    {
        double dfYReciprocalScale =  1.0 / dfYScale;
        const int nYReciprocalScale =
            static_cast<int>(dfYReciprocalScale + 0.5);
        if( fabs(dfYReciprocalScale-nYReciprocalScale) < 0.05 )
            dfYScale = 1.0 / nYReciprocalScale;
    }

    // XSCALE and YSCALE undocumented for now. Can help in some cases.
    // Best would probably be a per-pixel scale computation.
    const char* pszXScale = CSLFetchNameValue(papszWarpOptions, "XSCALE");
    if( pszXScale != nullptr )
        dfXScale = CPLAtof(pszXScale);
    const char* pszYScale = CSLFetchNameValue(papszWarpOptions, "YSCALE");
    if( pszYScale != nullptr )
        dfYScale = CPLAtof(pszYScale);

#if DEBUG_VERBOSE
    CPLDebug("WARP", "dfXScale = %f, dfYScale = %f", dfXScale, dfYScale);
#endif

    const int bUse4SamplesFormula = dfXScale >= 0.95 && dfYScale >= 0.95;

    // Safety check for callers that would use GDALWarpKernel without using
    // GDALWarpOperation.
    if( (eResample == GRA_CubicSpline || eResample == GRA_Lanczos ||
         ((eResample == GRA_Cubic || eResample == GRA_Bilinear) &&
          !bUse4SamplesFormula)) &&
        atoi(CSLFetchNameValueDef(papszWarpOptions,
                                  "EXTRA_ELTS", "0") ) != WARP_EXTRA_ELTS )
    {
        CPLError( CE_Failure, CPLE_AppDefined,
                  "Source arrays must have WARP_EXTRA_ELTS extra elements at "
                  "their end. "
                  "See GDALWarpKernel class definition. If this condition is "
                  "fulfilled, define a EXTRA_ELTS=%d warp options",
                  WARP_EXTRA_ELTS);
        return CE_Failure;
    }

    dfXFilter = anGWKFilterRadius[eResample];
    dfYFilter = anGWKFilterRadius[eResample];

    nXRadius =
        dfXScale < 1.0
        ? static_cast<int>(ceil(dfXFilter / dfXScale))
        : static_cast<int>(dfXFilter);
    nYRadius =
        dfYScale < 1.0
        ? static_cast<int>(ceil(dfYFilter / dfYScale))
        : static_cast<int>(dfYFilter);

    // Filter window offset depends on the parity of the kernel radius.
    nFiltInitX = ((anGWKFilterRadius[eResample] + 1) % 2) - nXRadius;
    nFiltInitY = ((anGWKFilterRadius[eResample] + 1) % 2) - nYRadius;

/* -------------------------------------------------------------------- */
/*      Set up resampling functions.                                    */
/* -------------------------------------------------------------------- */
    if( CPLFetchBool( papszWarpOptions, "USE_GENERAL_CASE", false ) )
        return GWKGeneralCase( this );

#if defined(HAVE_OPENCL)
    if( (eWorkingDataType == GDT_Byte
         || eWorkingDataType == GDT_CInt16
         || eWorkingDataType == GDT_UInt16
         || eWorkingDataType == GDT_Int16
         || eWorkingDataType == GDT_CFloat32
         || eWorkingDataType == GDT_Float32) &&
        (eResample == GRA_Bilinear
         || eResample == GRA_Cubic
         || eResample == GRA_CubicSpline
         || eResample == GRA_Lanczos) &&
        CPLFetchBool( papszWarpOptions, "USE_OPENCL", true ) )
    {
        const CPLErr eResult = GWKOpenCLCase( this );

        // CE_Warning tells us a suitable OpenCL environment was not available
        // so we fall through to other CPU based methods.
        if( eResult != CE_Warning )
            return eResult;
    }
#endif  // defined HAVE_OPENCL

    const bool bNoMasksOrDstDensityOnly =
        papanBandSrcValid == nullptr
        && panUnifiedSrcValid == nullptr
        && pafUnifiedSrcDensity == nullptr
        && panDstValid == nullptr;

    if( eWorkingDataType == GDT_Byte
        && eResample == GRA_NearestNeighbour
        && bNoMasksOrDstDensityOnly )
        return GWKNearestNoMasksOrDstDensityOnlyByte( this );

    if( eWorkingDataType == GDT_Byte
        && eResample == GRA_Bilinear
        && bNoMasksOrDstDensityOnly )
        return GWKBilinearNoMasksOrDstDensityOnlyByte( this );

    if( eWorkingDataType == GDT_Byte
        && eResample == GRA_Cubic
        && bNoMasksOrDstDensityOnly )
        return GWKCubicNoMasksOrDstDensityOnlyByte( this );

    if( eWorkingDataType == GDT_Byte
        && eResample == GRA_CubicSpline
        && bNoMasksOrDstDensityOnly )
        return GWKCubicSplineNoMasksOrDstDensityOnlyByte( this );

    if( eWorkingDataType == GDT_Byte
        && eResample == GRA_NearestNeighbour )
        return GWKNearestByte( this );

    if( (eWorkingDataType == GDT_Int16 || eWorkingDataType == GDT_UInt16)
        && eResample == GRA_NearestNeighbour
        && bNoMasksOrDstDensityOnly )
        return GWKNearestNoMasksOrDstDensityOnlyShort( this );

    if( (eWorkingDataType == GDT_Int16 )
        && eResample == GRA_Cubic
        && bNoMasksOrDstDensityOnly )
        return GWKCubicNoMasksOrDstDensityOnlyShort( this );

    if( (eWorkingDataType == GDT_Int16 )
        && eResample == GRA_CubicSpline
        && bNoMasksOrDstDensityOnly )
        return GWKCubicSplineNoMasksOrDstDensityOnlyShort( this );

    if( (eWorkingDataType == GDT_Int16 )
        && eResample == GRA_Bilinear
        && bNoMasksOrDstDensityOnly )
        return GWKBilinearNoMasksOrDstDensityOnlyShort( this );

    if( (eWorkingDataType == GDT_UInt16 )
        && eResample == GRA_Cubic
        && bNoMasksOrDstDensityOnly )
        return GWKCubicNoMasksOrDstDensityOnlyUShort( this );

    if( (eWorkingDataType == GDT_UInt16 )
        && eResample == GRA_CubicSpline
        && bNoMasksOrDstDensityOnly )
        return GWKCubicSplineNoMasksOrDstDensityOnlyUShort( this );

    if( (eWorkingDataType == GDT_UInt16 )
        && eResample == GRA_Bilinear
        && bNoMasksOrDstDensityOnly )
        return GWKBilinearNoMasksOrDstDensityOnlyUShort( this );

    if( (eWorkingDataType == GDT_Int16 || eWorkingDataType == GDT_UInt16)
        && eResample == GRA_NearestNeighbour )
        return GWKNearestShort( this );

    if( eWorkingDataType == GDT_Float32
        && eResample == GRA_NearestNeighbour
        && bNoMasksOrDstDensityOnly )
        return GWKNearestNoMasksOrDstDensityOnlyFloat( this );

    if( eWorkingDataType == GDT_Float32
        && eResample == GRA_NearestNeighbour )
        return GWKNearestFloat( this );

    if( eWorkingDataType == GDT_Float32
        && eResample == GRA_Bilinear
        && bNoMasksOrDstDensityOnly )
        return GWKBilinearNoMasksOrDstDensityOnlyFloat( this );

    if( eWorkingDataType == GDT_Float32
        && eResample == GRA_Cubic
        && bNoMasksOrDstDensityOnly )
        return GWKCubicNoMasksOrDstDensityOnlyFloat( this );

#ifdef INSTANTIATE_FLOAT64_SSE2_IMPL
    if( eWorkingDataType == GDT_Float64
        && eResample == GRA_Bilinear
        && bNoMasksOrDstDensityOnly )
        return GWKBilinearNoMasksOrDstDensityOnlyDouble( this );

    if( eWorkingDataType == GDT_Float64
        && eResample == GRA_Cubic
        && bNoMasksOrDstDensityOnly )
        return GWKCubicNoMasksOrDstDensityOnlyDouble( this );
#endif

    if( eResample == GRA_Average )
        return GWKAverageOrMode( this );

    if( eResample == GRA_Mode )
        return GWKAverageOrMode( this );

    if( eResample == GRA_Max )
        return GWKAverageOrMode( this );

    if( eResample == GRA_Min )
        return GWKAverageOrMode( this );

    if( eResample == GRA_Med )
        return GWKAverageOrMode( this );

    if( eResample == GRA_Q1 )
        return GWKAverageOrMode( this );

    if( eResample == GRA_Q3 )
        return GWKAverageOrMode( this );

    if (!GDALDataTypeIsComplex(eWorkingDataType))
    {
        return GWKRealCase( this );
    }

    return GWKGeneralCase( this );
}

/************************************************************************/
/*                              Validate()                              */
/************************************************************************/

/**
 * \fn CPLErr GDALWarpKernel::Validate()
 *
 * Check the settings in the GDALWarpKernel, and issue a CPLError()
 * (and return CE_Failure) if the configuration is considered to be
 * invalid for some reason.
 *
 * This method will also do some standard defaulting such as setting
 * pfnProgress to GDALDummyProgress() if it is NULL.
 *
 * @return CE_None on success or CE_Failure if an error is detected.
 */

CPLErr GDALWarpKernel::Validate()

{
    if( static_cast<size_t>(eResample) >=
        (sizeof(anGWKFilterRadius) / sizeof(anGWKFilterRadius[0])) )
    {
        CPLError( CE_Failure, CPLE_AppDefined,
                  "Unsupported resampling method %d.",
                  static_cast<int>(eResample) );
        return CE_Failure;
    }

    return CE_None;
}

/************************************************************************/
/*                         GWKOverlayDensity()                          */
/*                                                                      */
/*      Compute the final density for the destination pixel.  This      */
/*      is a function of the overlay density (passed in) and the        */
/*      original density.                                               */
/************************************************************************/

static void GWKOverlayDensity( GDALWarpKernel *poWK, int iDstOffset,
                               double dfDensity )
{
    if( dfDensity < 0.0001 || poWK->pafDstDensity == nullptr )
        return;

    poWK->pafDstDensity[iDstOffset] =  static_cast<float>
        ( 1.0 - (1.0 - dfDensity) * (1.0 - poWK->pafDstDensity[iDstOffset]) );
}

/************************************************************************/
/*                          GWKRoundValueT()                            */
/************************************************************************/

template<class T, EMULATED_BOOL is_signed> struct sGWKRoundValueT
{
    static T eval(double);
};

template<class T> struct sGWKRoundValueT<T, true> /* signed */
{
    static T eval(double dfValue) { return static_cast<T>(floor(dfValue + 0.5)); }
};

template<class T> struct sGWKRoundValueT<T, false> /* unsigned */
{
    static T eval(double dfValue) { return static_cast<T>(dfValue + 0.5); }
};

template<class T> static T GWKRoundValueT(double dfValue)
{
    return sGWKRoundValueT<T, std::numeric_limits<T>::is_signed>::eval(dfValue);
}

template<> float GWKRoundValueT<float>(double dfValue)
{
    return static_cast<float>(dfValue);
}

#ifdef notused
template<> double GWKRoundValueT<double>(double dfValue)
{
    return dfValue;
}
#endif

/************************************************************************/
/*                            GWKClampValueT()                          */
/************************************************************************/

template<class T>
static CPL_INLINE T GWKClampValueT(double dfValue)
{
    if( dfValue < std::numeric_limits<T>::min() )
        return std::numeric_limits<T>::min();
    else if( dfValue > std::numeric_limits<T>::max() )
        return std::numeric_limits<T>::max();
    else
        return GWKRoundValueT<T>(dfValue);
}

template<> float GWKClampValueT<float>(double dfValue)
{
    return static_cast<float>(dfValue);
}

#ifdef notused
template<> double GWKClampValueT<double>(double dfValue)
{
    return dfValue;
}
#endif

/************************************************************************/
/*                         GWKSetPixelValueRealT()                      */
/************************************************************************/

template<class T>
static bool GWKSetPixelValueRealT( GDALWarpKernel *poWK, int iBand,
                                   int iDstOffset, double dfDensity,
                                   T value)
{
    T *pDst = reinterpret_cast<T*>(poWK->papabyDstImage[iBand]);

/* -------------------------------------------------------------------- */
/*      If the source density is less than 100% we need to fetch the    */
/*      existing destination value, and mix it with the source to       */
/*      get the new "to apply" value.  Also compute composite           */
/*      density.                                                        */
/*                                                                      */
/*      We avoid mixing if density is very near one or risk mixing      */
/*      in very extreme nodata values and causing odd results (#1610)   */
/* -------------------------------------------------------------------- */
    if( dfDensity < 0.9999 )
    {
        if( dfDensity < 0.0001 )
            return true;

        double dfDstDensity = 1.0;

        if( poWK->pafDstDensity != nullptr )
            dfDstDensity = poWK->pafDstDensity[iDstOffset];
        else if( poWK->panDstValid != nullptr
                 && !((poWK->panDstValid[iDstOffset>>5]
                       & (0x01 << (iDstOffset & 0x1f))) ) )
            dfDstDensity = 0.0;

        // It seems like we also ought to be testing panDstValid[] here!

        const double dfDstReal = pDst[iDstOffset];

        // The destination density is really only relative to the portion
        // not occluded by the overlay.
        const double dfDstInfluence = (1.0 - dfDensity) * dfDstDensity;

        const double dfReal =
            (value * dfDensity + dfDstReal * dfDstInfluence)
            / (dfDensity + dfDstInfluence);

/* -------------------------------------------------------------------- */
/*      Actually apply the destination value.                           */
/*                                                                      */
/*      Avoid using the destination nodata value for integer datatypes  */
/*      if by chance it is equal to the computed pixel value.           */
/* -------------------------------------------------------------------- */
        pDst[iDstOffset] = GWKClampValueT<T>(dfReal);
    }
    else
    {
        pDst[iDstOffset] = value;
    }

    if( poWK->padfDstNoDataReal != nullptr &&
        poWK->padfDstNoDataReal[iBand] ==
        static_cast<double>(pDst[iDstOffset]) )
    {
        if( pDst[iDstOffset] == std::numeric_limits<T>::min() )
            pDst[iDstOffset] = std::numeric_limits<T>::min() + 1;
        else
            pDst[iDstOffset]--;
    }

    return true;
}

/************************************************************************/
/*                          GWKSetPixelValue()                          */
/************************************************************************/

static bool GWKSetPixelValue( GDALWarpKernel *poWK, int iBand,
                              int iDstOffset, double dfDensity,
                              double dfReal, double dfImag )

{
    GByte *pabyDst = poWK->papabyDstImage[iBand];

/* -------------------------------------------------------------------- */
/*      If the source density is less than 100% we need to fetch the    */
/*      existing destination value, and mix it with the source to       */
/*      get the new "to apply" value.  Also compute composite           */
/*      density.                                                        */
/*                                                                      */
/*      We avoid mixing if density is very near one or risk mixing      */
/*      in very extreme nodata values and causing odd results (#1610)   */
/* -------------------------------------------------------------------- */
    if( dfDensity < 0.9999 )
    {
        if( dfDensity < 0.0001 )
            return true;

        double dfDstDensity = 1.0;
        if( poWK->pafDstDensity != nullptr )
            dfDstDensity = poWK->pafDstDensity[iDstOffset];
        else if( poWK->panDstValid != nullptr
                 && !((poWK->panDstValid[iDstOffset>>5]
                       & (0x01 << (iDstOffset & 0x1f))) ) )
            dfDstDensity = 0.0;

        double dfDstReal = 0.0;
        double dfDstImag = 0.0;
        // It seems like we also ought to be testing panDstValid[] here!

        // TODO(schwehr): Factor out this repreated type of set.
        switch( poWK->eWorkingDataType )
        {
          case GDT_Byte:
            dfDstReal = pabyDst[iDstOffset];
            dfDstImag = 0.0;
            break;

          case GDT_Int16:
            dfDstReal = reinterpret_cast<GInt16 *>(pabyDst)[iDstOffset];
            dfDstImag = 0.0;
            break;

          case GDT_UInt16:
            dfDstReal = reinterpret_cast<GUInt16*>(pabyDst)[iDstOffset];
            dfDstImag = 0.0;
            break;

          case GDT_Int32:
            dfDstReal = reinterpret_cast<GInt32*>(pabyDst)[iDstOffset];
            dfDstImag = 0.0;
            break;

          case GDT_UInt32:
            dfDstReal = reinterpret_cast<GUInt32*>(pabyDst)[iDstOffset];
            dfDstImag = 0.0;
            break;

          case GDT_Float32:
            dfDstReal = reinterpret_cast<float*>(pabyDst)[iDstOffset];
            dfDstImag = 0.0;
            break;

          case GDT_Float64:
            dfDstReal = reinterpret_cast<double*>(pabyDst)[iDstOffset];
            dfDstImag = 0.0;
            break;

          case GDT_CInt16:
            dfDstReal = reinterpret_cast<GInt16*>(pabyDst)[iDstOffset*2];
            dfDstImag = reinterpret_cast<GInt16*>(pabyDst)[iDstOffset*2+1];
            break;

          case GDT_CInt32:
            dfDstReal = reinterpret_cast<GInt32*>(pabyDst)[iDstOffset*2];
            dfDstImag = reinterpret_cast<GInt32*>(pabyDst)[iDstOffset*2+1];
            break;

          case GDT_CFloat32:
            dfDstReal = reinterpret_cast<float*>(pabyDst)[iDstOffset*2];
            dfDstImag = reinterpret_cast<float*>(pabyDst)[iDstOffset*2+1];
            break;

          case GDT_CFloat64:
            dfDstReal = reinterpret_cast<double*>(pabyDst)[iDstOffset*2];
            dfDstImag = reinterpret_cast<double*>(pabyDst)[iDstOffset*2+1];
            break;

          default:
            CPLAssert( false );
            dfDstDensity = 0.0;
            return false;
        }

        // The destination density is really only relative to the portion
        // not occluded by the overlay.
        const double dfDstInfluence = (1.0 - dfDensity) * dfDstDensity;

        dfReal =
            (dfReal * dfDensity + dfDstReal * dfDstInfluence)
            / (dfDensity + dfDstInfluence);

        dfImag =
            (dfImag * dfDensity + dfDstImag * dfDstInfluence)
            / (dfDensity + dfDstInfluence);
    }

/* -------------------------------------------------------------------- */
/*      Actually apply the destination value.                           */
/*                                                                      */
/*      Avoid using the destination nodata value for integer datatypes  */
/*      if by chance it is equal to the computed pixel value.           */
/* -------------------------------------------------------------------- */

// TODO(schwehr): Can we make this a template?
#define CLAMP(type) \
    do { \
    type* _pDst = reinterpret_cast<type*>(pabyDst); \
    if( dfReal < static_cast<double>(std::numeric_limits<type>::min()) ) \
        _pDst[iDstOffset] = static_cast<type>(std::numeric_limits<type>::min()); \
    else if( dfReal > static_cast<double>(std::numeric_limits<type>::max()) ) \
        _pDst[iDstOffset] = static_cast<type>(std::numeric_limits<type>::max()); \
    else \
        _pDst[iDstOffset] = \
            (std::numeric_limits<type>::is_signed) ? \
                static_cast<type>(floor(dfReal + 0.5)) : \
                static_cast<type>(dfReal + 0.5);  \
    if( poWK->padfDstNoDataReal != nullptr && \
        poWK->padfDstNoDataReal[iBand] == static_cast<double>(_pDst[iDstOffset]) ) \
    { \
        if( _pDst[iDstOffset] == static_cast<type>(std::numeric_limits<type>::min()) )  \
            _pDst[iDstOffset] = static_cast<type>(std::numeric_limits<type>::min() + 1); \
        else \
            _pDst[iDstOffset]--; \
    } } while( false )

    switch( poWK->eWorkingDataType )
    {
      case GDT_Byte:
        CLAMP(GByte);
        break;

      case GDT_Int16:
        CLAMP(GInt16);
        break;

      case GDT_UInt16:
        CLAMP(GUInt16);
        break;

      case GDT_UInt32:
        CLAMP(GUInt32);
        break;

      case GDT_Int32:
        CLAMP(GInt32);
        break;

      case GDT_Float32:
        reinterpret_cast<float*>(pabyDst)[iDstOffset] = static_cast<float>(dfReal);
        break;

      case GDT_Float64:
        reinterpret_cast<double*>(pabyDst)[iDstOffset] = dfReal;
        break;

      case GDT_CInt16:
      {
        typedef GInt16 T;
        if( dfReal < static_cast<double>(std::numeric_limits<T>::min()) )
            reinterpret_cast<T*>(pabyDst)[iDstOffset*2] = std::numeric_limits<T>::min();
        else if( dfReal > static_cast<double>(std::numeric_limits<T>::max()) )
            reinterpret_cast<T*>(pabyDst)[iDstOffset*2] = std::numeric_limits<T>::max();
        else
            reinterpret_cast<T*>(pabyDst)[iDstOffset*2] = static_cast<T>(floor(dfReal+0.5));
        if( dfImag < static_cast<double>(std::numeric_limits<T>::min()) )
            reinterpret_cast<T*>(pabyDst)[iDstOffset*2+1] =  std::numeric_limits<T>::min();
        else if( dfImag > static_cast<double>(std::numeric_limits<T>::max()) )
            reinterpret_cast<T*>(pabyDst)[iDstOffset*2+1] = std::numeric_limits<T>::max();
        else
            reinterpret_cast<T*>(pabyDst)[iDstOffset*2+1] = static_cast<T>(floor(dfImag+0.5));
        break;
      }

      case GDT_CInt32:
      {
        typedef GInt32 T;
        if( dfReal < static_cast<double>(std::numeric_limits<T>::min()) )
            reinterpret_cast<T*>(pabyDst)[iDstOffset*2] = std::numeric_limits<T>::min();
        else if( dfReal > static_cast<double>(std::numeric_limits<T>::max()) )
            reinterpret_cast<T*>(pabyDst)[iDstOffset*2] = std::numeric_limits<T>::max();
        else
            reinterpret_cast<T*>(pabyDst)[iDstOffset*2] = static_cast<T>(floor(dfReal+0.5));
        if( dfImag < static_cast<double>(std::numeric_limits<T>::min()) )
            reinterpret_cast<T*>(pabyDst)[iDstOffset*2+1] = std::numeric_limits<T>::min();
        else if( dfImag > static_cast<double>(std::numeric_limits<T>::max()) )
            reinterpret_cast<T*>(pabyDst)[iDstOffset*2+1] = std::numeric_limits<T>::max();
        else
            reinterpret_cast<T*>(pabyDst)[iDstOffset*2+1] = static_cast<T>(floor(dfImag+0.5));
        break;
      }

      case GDT_CFloat32:
        reinterpret_cast<float*>(pabyDst)[iDstOffset*2] =  static_cast<float>(dfReal);
        reinterpret_cast<float*>(pabyDst)[iDstOffset*2+1] = static_cast<float>(dfImag);
        break;

      case GDT_CFloat64:
        reinterpret_cast<double*>(pabyDst)[iDstOffset*2] = dfReal;
        reinterpret_cast<double*>(pabyDst)[iDstOffset*2+1] = dfImag;
        break;

      default:
        return false;
    }

    return true;
}

/************************************************************************/
/*                       GWKSetPixelValueReal()                         */
/************************************************************************/

static bool GWKSetPixelValueReal( GDALWarpKernel *poWK, int iBand,
                                  int iDstOffset, double dfDensity,
                                  double dfReal )

{
    GByte *pabyDst = poWK->papabyDstImage[iBand];

/* -------------------------------------------------------------------- */
/*      If the source density is less than 100% we need to fetch the    */
/*      existing destination value, and mix it with the source to       */
/*      get the new "to apply" value.  Also compute composite           */
/*      density.                                                        */
/*                                                                      */
/*      We avoid mixing if density is very near one or risk mixing      */
/*      in very extreme nodata values and causing odd results (#1610)   */
/* -------------------------------------------------------------------- */
    if( dfDensity < 0.9999 )
    {
        if( dfDensity < 0.0001 )
            return true;

        double dfDstReal = 0.0;
        double dfDstDensity = 1.0;

        if( poWK->pafDstDensity != nullptr )
            dfDstDensity = poWK->pafDstDensity[iDstOffset];
        else if( poWK->panDstValid != nullptr
                 && !((poWK->panDstValid[iDstOffset>>5]
                       & (0x01 << (iDstOffset & 0x1f))) ) )
            dfDstDensity = 0.0;

        // It seems like we also ought to be testing panDstValid[] here!

        switch( poWK->eWorkingDataType )
        {
          case GDT_Byte:
            dfDstReal = pabyDst[iDstOffset];
            break;

          case GDT_Int16:
            dfDstReal = reinterpret_cast<GInt16*>(pabyDst)[iDstOffset];
            break;

          case GDT_UInt16:
            dfDstReal = reinterpret_cast<GUInt16*>(pabyDst)[iDstOffset];
            break;

          case GDT_Int32:
            dfDstReal = reinterpret_cast<GInt32*>(pabyDst)[iDstOffset];
            break;

          case GDT_UInt32:
            dfDstReal = reinterpret_cast<GUInt32*>(pabyDst)[iDstOffset];
            break;

          case GDT_Float32:
            dfDstReal = reinterpret_cast<float*>(pabyDst)[iDstOffset];
            break;

          case GDT_Float64:
            dfDstReal = reinterpret_cast<double*>(pabyDst)[iDstOffset];
            break;

          default:
            CPLAssert( false );
            dfDstDensity = 0.0;
            return false;
        }

        // The destination density is really only relative to the portion
        // not occluded by the overlay.
        const double dfDstInfluence = (1.0 - dfDensity) * dfDstDensity;

        dfReal =
            (dfReal * dfDensity + dfDstReal * dfDstInfluence)
            / (dfDensity + dfDstInfluence);
    }

/* -------------------------------------------------------------------- */
/*      Actually apply the destination value.                           */
/*                                                                      */
/*      Avoid using the destination nodata value for integer datatypes  */
/*      if by chance it is equal to the computed pixel value.           */
/* -------------------------------------------------------------------- */

    switch( poWK->eWorkingDataType )
    {
      case GDT_Byte:
        CLAMP(GByte);
        break;

      case GDT_Int16:
        CLAMP(GInt16);
        break;

      case GDT_UInt16:
        CLAMP(GUInt16);
        break;

      case GDT_UInt32:
        CLAMP(GUInt32);
        break;

      case GDT_Int32:
        CLAMP(GInt32);
        break;

      case GDT_Float32:
        reinterpret_cast<float*>(pabyDst)[iDstOffset] = static_cast<float>(dfReal);
        break;

      case GDT_Float64:
        reinterpret_cast<double*>(pabyDst)[iDstOffset] = dfReal;
        break;

      default:
        return false;
    }

    return true;
}

/************************************************************************/
/*                          GWKGetPixelValue()                          */
/************************************************************************/

/* It is assumed that panUnifiedSrcValid has been checked before */

static bool GWKGetPixelValue( GDALWarpKernel *poWK, int iBand,
                              int iSrcOffset, double *pdfDensity,
                              double *pdfReal, double *pdfImag )

{
    GByte *pabySrc = poWK->papabySrcImage[iBand];

    if( poWK->papanBandSrcValid != nullptr
        && poWK->papanBandSrcValid[iBand] != nullptr
        && !((poWK->papanBandSrcValid[iBand][iSrcOffset>>5]
              & (0x01 << (iSrcOffset & 0x1f)))) )
    {
        *pdfDensity = 0.0;
        return false;
    }

    // TODO(schwehr): Fix casting.
    switch( poWK->eWorkingDataType )
    {
      case GDT_Byte:
        *pdfReal = pabySrc[iSrcOffset];
        *pdfImag = 0.0;
        break;

      case GDT_Int16:
        *pdfReal = reinterpret_cast<GInt16*>(pabySrc)[iSrcOffset];
        *pdfImag = 0.0;
        break;

      case GDT_UInt16:
        *pdfReal = reinterpret_cast<GUInt16*>(pabySrc)[iSrcOffset];
        *pdfImag = 0.0;
        break;

      case GDT_Int32:
        *pdfReal = reinterpret_cast<GInt32*>(pabySrc)[iSrcOffset];
        *pdfImag = 0.0;
        break;

      case GDT_UInt32:
        *pdfReal = reinterpret_cast<GUInt32*>(pabySrc)[iSrcOffset];
        *pdfImag = 0.0;
        break;

      case GDT_Float32:
        *pdfReal = reinterpret_cast<float*>(pabySrc)[iSrcOffset];
        *pdfImag = 0.0;
        break;

      case GDT_Float64:
        *pdfReal = reinterpret_cast<double*>(pabySrc)[iSrcOffset];
        *pdfImag = 0.0;
        break;

      case GDT_CInt16:
        *pdfReal = reinterpret_cast<GInt16*>(pabySrc)[iSrcOffset*2];
        *pdfImag = reinterpret_cast<GInt16*>(pabySrc)[iSrcOffset*2+1];
        break;

      case GDT_CInt32:
        *pdfReal = reinterpret_cast<GInt32*>(pabySrc)[iSrcOffset*2];
        *pdfImag = reinterpret_cast<GInt32*>(pabySrc)[iSrcOffset*2+1];
        break;

      case GDT_CFloat32:
        *pdfReal = reinterpret_cast<float*>(pabySrc)[iSrcOffset*2];
        *pdfImag = reinterpret_cast<float*>(pabySrc)[iSrcOffset*2+1];
        break;

      case GDT_CFloat64:
        *pdfReal = reinterpret_cast<double*>(pabySrc)[iSrcOffset*2];
        *pdfImag = reinterpret_cast<double*>(pabySrc)[iSrcOffset*2+1];
        break;

      default:
        *pdfDensity = 0.0;
        return false;
    }

    if( poWK->pafUnifiedSrcDensity != nullptr )
        *pdfDensity = poWK->pafUnifiedSrcDensity[iSrcOffset];
    else
        *pdfDensity = 1.0;

    return *pdfDensity != 0.0;
}

/************************************************************************/
/*                       GWKGetPixelValueReal()                         */
/************************************************************************/

static bool GWKGetPixelValueReal( GDALWarpKernel *poWK, int iBand,
                                  int iSrcOffset, double *pdfDensity,
                                  double *pdfReal )

{
    GByte *pabySrc = poWK->papabySrcImage[iBand];

    if( poWK->papanBandSrcValid != nullptr
        && poWK->papanBandSrcValid[iBand] != nullptr
        && !((poWK->papanBandSrcValid[iBand][iSrcOffset>>5]
              & (0x01 << (iSrcOffset & 0x1f)))) )
    {
        *pdfDensity = 0.0;
        return false;
    }

    switch( poWK->eWorkingDataType )
    {
      case GDT_Byte:
        *pdfReal = pabySrc[iSrcOffset];
        break;

      case GDT_Int16:
        *pdfReal = reinterpret_cast<GInt16*>(pabySrc)[iSrcOffset];
        break;

      case GDT_UInt16:
        *pdfReal = reinterpret_cast<GUInt16*>(pabySrc)[iSrcOffset];
        break;

      case GDT_Int32:
        *pdfReal = reinterpret_cast<GInt32*>(pabySrc)[iSrcOffset];
        break;

      case GDT_UInt32:
        *pdfReal = reinterpret_cast<GUInt32*>(pabySrc)[iSrcOffset];
        break;

      case GDT_Float32:
        *pdfReal = reinterpret_cast<float*>(pabySrc)[iSrcOffset];
        break;

      case GDT_Float64:
        *pdfReal = reinterpret_cast<double*>(pabySrc)[iSrcOffset];
        break;

      default:
        CPLAssert(false);
        return false;
    }

    if( poWK->pafUnifiedSrcDensity != nullptr )
        *pdfDensity = poWK->pafUnifiedSrcDensity[iSrcOffset];
    else
        *pdfDensity = 1.0;

    return *pdfDensity != 0.0;
}

/************************************************************************/
/*                          GWKGetPixelRow()                            */
/************************************************************************/

/* It is assumed that adfImag[] is set to 0 by caller code for non-complex */
/* data-types. */

static bool GWKGetPixelRow( GDALWarpKernel *poWK, int iBand,
                            int iSrcOffset, int nHalfSrcLen,
                            double* padfDensity,
                            double adfReal[],
                            double* padfImag )
{
    // We know that nSrcLen is even, so we can *always* unroll loops 2x.
    const int nSrcLen = nHalfSrcLen * 2;
    bool bHasValid = false;

    if( padfDensity != nullptr )
    {
        // Init the density.
        for( int i = 0; i < nSrcLen; i += 2 )
        {
            padfDensity[i] = 1.0;
            padfDensity[i+1] = 1.0;
        }

        if( poWK->panUnifiedSrcValid != nullptr )
        {
            for( int i = 0; i < nSrcLen; i += 2 )
            {
                if( poWK->panUnifiedSrcValid[(iSrcOffset+i)>>5]
                    & (0x01 << ((iSrcOffset+i) & 0x1f)) )
                    bHasValid = true;
                else
                    padfDensity[i] = 0.0;

                if( poWK->panUnifiedSrcValid[(iSrcOffset+i+1)>>5]
                    & (0x01 << ((iSrcOffset+i+1) & 0x1f)) )
                    bHasValid = true;
                else
                    padfDensity[i+1] = 0.0;
            }

            // Reset or fail as needed.
            if( bHasValid )
                bHasValid = false;
            else
                return false;
        }

        if( poWK->papanBandSrcValid != nullptr
            && poWK->papanBandSrcValid[iBand] != nullptr )
        {
            for( int i = 0; i < nSrcLen; i += 2 )
            {
                if( poWK->papanBandSrcValid[iBand][(iSrcOffset+i)>>5]
                    & (0x01 << ((iSrcOffset+i) & 0x1f)) )
                    bHasValid = true;
                else
                    padfDensity[i] = 0.0;

                if( poWK->papanBandSrcValid[iBand][(iSrcOffset+i+1)>>5]
                    & (0x01 << ((iSrcOffset+i+1) & 0x1f)) )
                    bHasValid = true;
                else
                    padfDensity[i+1] = 0.0;
            }

            // Reset or fail as needed.
            if( bHasValid )
                bHasValid = false;
            else
                return false;
        }
    }

    // TODO(schwehr): Fix casting.
    // Fetch data.
    switch( poWK->eWorkingDataType )
    {
        case GDT_Byte:
        {
            GByte* pSrc = reinterpret_cast<GByte*>(poWK->papabySrcImage[iBand]);
            pSrc += iSrcOffset;
            for( int i = 0; i < nSrcLen; i += 2 )
            {
                adfReal[i] = pSrc[i];
                adfReal[i+1] = pSrc[i+1];
            }
            break;
        }

        case GDT_Int16:
        {
            GInt16* pSrc = reinterpret_cast<GInt16*>(poWK->papabySrcImage[iBand]);
            pSrc += iSrcOffset;
            for( int i = 0; i < nSrcLen; i += 2 )
            {
                adfReal[i] = pSrc[i];
                adfReal[i+1] = pSrc[i+1];
            }
            break;
        }

         case GDT_UInt16:
         {
            GUInt16* pSrc = reinterpret_cast<GUInt16*>(poWK->papabySrcImage[iBand]);
            pSrc += iSrcOffset;
            for( int i = 0; i < nSrcLen; i += 2 )
            {
                adfReal[i] = pSrc[i];
                adfReal[i+1] = pSrc[i+1];
            }
            break;
         }

        case GDT_Int32:
        {
            GInt32* pSrc = reinterpret_cast<GInt32*>(poWK->papabySrcImage[iBand]);
            pSrc += iSrcOffset;
            for( int i = 0; i < nSrcLen; i += 2 )
            {
                adfReal[i] = pSrc[i];
                adfReal[i+1] = pSrc[i+1];
            }
            break;
        }

        case GDT_UInt32:
        {
            GUInt32* pSrc = reinterpret_cast<GUInt32*>(poWK->papabySrcImage[iBand]);
            pSrc += iSrcOffset;
            for( int i = 0; i < nSrcLen; i += 2 )
            {
                adfReal[i] = pSrc[i];
                adfReal[i+1] = pSrc[i+1];
            }
            break;
        }

        case GDT_Float32:
        {
            float* pSrc = reinterpret_cast<float*>(poWK->papabySrcImage[iBand]);
            pSrc += iSrcOffset;
            for( int i = 0; i < nSrcLen; i += 2 )
            {
                adfReal[i] = pSrc[i];
                adfReal[i+1] = pSrc[i+1];
            }
            break;
        }

       case GDT_Float64:
       {
            double* pSrc = reinterpret_cast<double*>(poWK->papabySrcImage[iBand]);
            pSrc += iSrcOffset;
            for( int i = 0; i < nSrcLen; i += 2 )
            {
                adfReal[i] = pSrc[i];
                adfReal[i+1] = pSrc[i+1];
            }
            break;
       }

        case GDT_CInt16:
        {
            GInt16* pSrc = reinterpret_cast<GInt16*>(poWK->papabySrcImage[iBand]);
            pSrc += 2 * iSrcOffset;
            for( int i = 0; i < nSrcLen; i += 2 )
            {
                adfReal[i] = pSrc[2*i];
                padfImag[i] = pSrc[2*i+1];

                adfReal[i+1] = pSrc[2*i+2];
                padfImag[i+1] = pSrc[2*i+3];
            }
            break;
        }

        case GDT_CInt32:
        {
            GInt32* pSrc = reinterpret_cast<GInt32*>(poWK->papabySrcImage[iBand]);
            pSrc += 2 * iSrcOffset;
            for( int i = 0; i < nSrcLen; i += 2 )
            {
                adfReal[i] = pSrc[2*i];
                padfImag[i] = pSrc[2*i+1];

                adfReal[i+1] = pSrc[2*i+2];
                padfImag[i+1] = pSrc[2*i+3];
            }
            break;
        }

        case GDT_CFloat32:
        {
            float* pSrc = reinterpret_cast<float*>(poWK->papabySrcImage[iBand]);
            pSrc += 2 * iSrcOffset;
            for( int i = 0; i < nSrcLen; i += 2 )
            {
                adfReal[i] = pSrc[2*i];
                padfImag[i] = pSrc[2*i+1];

                adfReal[i+1] = pSrc[2*i+2];
                padfImag[i+1] = pSrc[2*i+3];
            }
            break;
        }

        case GDT_CFloat64:
        {
            double* pSrc = reinterpret_cast<double*>(poWK->papabySrcImage[iBand]);
            pSrc += 2 * iSrcOffset;
            for( int i = 0; i < nSrcLen; i += 2 )
            {
                adfReal[i] = pSrc[2*i];
                padfImag[i] = pSrc[2*i+1];

                adfReal[i+1] = pSrc[2*i+2];
                padfImag[i+1] = pSrc[2*i+3];
            }
            break;
        }

        default:
            CPLAssert(false);
            if( padfDensity )
                memset( padfDensity, 0, nSrcLen * sizeof(double) );
            return false;
    }

    if( padfDensity == nullptr )
        return true;

    if( poWK->pafUnifiedSrcDensity == nullptr )
    {
        for( int i = 0; i < nSrcLen; i += 2 )
        {
            // Take into account earlier calcs.
            if( padfDensity[i] > SRC_DENSITY_THRESHOLD )
            {
                padfDensity[i] = 1.0;
                bHasValid = true;
            }

            if( padfDensity[i+1] > SRC_DENSITY_THRESHOLD )
            {
                padfDensity[i+1] = 1.0;
                bHasValid = true;
            }
        }
    }
    else
    {
        for( int i = 0; i < nSrcLen; i += 2 )
        {
            if( padfDensity[i] > SRC_DENSITY_THRESHOLD )
                padfDensity[i] = poWK->pafUnifiedSrcDensity[iSrcOffset+i];
            if( padfDensity[i] > SRC_DENSITY_THRESHOLD )
                bHasValid = true;

            if( padfDensity[i+1] > SRC_DENSITY_THRESHOLD )
                padfDensity[i+1] = poWK->pafUnifiedSrcDensity[iSrcOffset+i+1];
            if( padfDensity[i+1] > SRC_DENSITY_THRESHOLD )
                bHasValid = true;
        }
    }

    return bHasValid;
}

/************************************************************************/
/*                          GWKGetPixelT()                              */
/************************************************************************/

template<class T>
static bool GWKGetPixelT( GDALWarpKernel *poWK, int iBand,
                          int iSrcOffset, double *pdfDensity,
                          T *pValue )

{
    T *pSrc = reinterpret_cast<T *>(poWK->papabySrcImage[iBand]);

    if( ( poWK->panUnifiedSrcValid != nullptr
          && !((poWK->panUnifiedSrcValid[iSrcOffset>>5]
                & (0x01 << (iSrcOffset & 0x1f))) ) )
        || ( poWK->papanBandSrcValid != nullptr
             && poWK->papanBandSrcValid[iBand] != nullptr
             && !((poWK->papanBandSrcValid[iBand][iSrcOffset>>5]
                   & (0x01 << (iSrcOffset & 0x1f)))) ) )
    {
        *pdfDensity = 0.0;
        return false;
    }

    *pValue = pSrc[iSrcOffset];

    if( poWK->pafUnifiedSrcDensity == nullptr )
        *pdfDensity = 1.0;
    else
        *pdfDensity = poWK->pafUnifiedSrcDensity[iSrcOffset];

    return *pdfDensity != 0.0;
}

/************************************************************************/
/*                        GWKBilinearResample()                         */
/*     Set of bilinear interpolators                                    */
/************************************************************************/

static bool GWKBilinearResample4Sample( GDALWarpKernel *poWK, int iBand,
                                double dfSrcX, double dfSrcY,
                                double *pdfDensity,
                                double *pdfReal, double *pdfImag )

{
    // Save as local variables to avoid following pointers.
    const int nSrcXSize = poWK->nSrcXSize;
    const int nSrcYSize = poWK->nSrcYSize;

    int iSrcX = static_cast<int>(floor(dfSrcX - 0.5));
    int iSrcY = static_cast<int>(floor(dfSrcY - 0.5));
    double dfRatioX = 1.5 - (dfSrcX - iSrcX);
    double dfRatioY = 1.5 - (dfSrcY - iSrcY);
    bool bShifted = false;

    if( iSrcX == -1 )
    {
        iSrcX = 0;
        dfRatioX = 1;
    }
    if( iSrcY == -1 )
    {
        iSrcY = 0;
        dfRatioY = 1;
    }
    int iSrcOffset = iSrcX + iSrcY * nSrcXSize;

    // Shift so we don't overrun the array.
    if( nSrcXSize * nSrcYSize == iSrcOffset + 1
        || nSrcXSize * nSrcYSize == iSrcOffset + nSrcXSize + 1 )
    {
        bShifted = true;
        --iSrcOffset;
    }

    double adfDensity[2] = { 0.0, 0.0 };
    double adfReal[2] = { 0.0, 0.0 };
    double adfImag[2] = { 0.0, 0.0 };
    double dfAccumulatorReal = 0.0;
    double dfAccumulatorImag = 0.0;
    double dfAccumulatorDensity = 0.0;
    double dfAccumulatorDivisor = 0.0;

    // Get pixel row.
    if( iSrcY >= 0 && iSrcY < nSrcYSize
        && iSrcOffset >= 0 && iSrcOffset < nSrcXSize * nSrcYSize
        && GWKGetPixelRow( poWK, iBand, iSrcOffset, 1,
                           adfDensity, adfReal, adfImag ) )
    {
        double dfMult1 = dfRatioX * dfRatioY;
        double dfMult2 = (1.0-dfRatioX) * dfRatioY;

        // Shifting corrected.
        if( bShifted )
        {
            adfReal[0] = adfReal[1];
            adfImag[0] = adfImag[1];
            adfDensity[0] = adfDensity[1];
        }

        // Upper Left Pixel.
        if( iSrcX >= 0 && iSrcX < nSrcXSize
            && adfDensity[0] > SRC_DENSITY_THRESHOLD )
        {
            dfAccumulatorDivisor += dfMult1;

            dfAccumulatorReal += adfReal[0] * dfMult1;
            dfAccumulatorImag += adfImag[0] * dfMult1;
            dfAccumulatorDensity += adfDensity[0] * dfMult1;
        }

        // Upper Right Pixel.
        if( iSrcX+1 >= 0 && iSrcX+1 < nSrcXSize
            && adfDensity[1] > SRC_DENSITY_THRESHOLD )
        {
            dfAccumulatorDivisor += dfMult2;

            dfAccumulatorReal += adfReal[1] * dfMult2;
            dfAccumulatorImag += adfImag[1] * dfMult2;
            dfAccumulatorDensity += adfDensity[1] * dfMult2;
        }
    }

    // Get pixel row.
    if( iSrcY+1 >= 0 && iSrcY+1 < nSrcYSize
        && iSrcOffset+nSrcXSize >= 0
        && iSrcOffset+nSrcXSize < nSrcXSize * nSrcYSize
        && GWKGetPixelRow( poWK, iBand, iSrcOffset+nSrcXSize, 1,
                           adfDensity, adfReal, adfImag ) )
    {
        double dfMult1 = dfRatioX * (1.0-dfRatioY);
        double dfMult2 = (1.0-dfRatioX) * (1.0-dfRatioY);

        // Shifting corrected
        if ( bShifted )
        {
            adfReal[0] = adfReal[1];
            adfImag[0] = adfImag[1];
            adfDensity[0] = adfDensity[1];
        }

        // Lower Left Pixel
        if ( iSrcX >= 0 && iSrcX < nSrcXSize
             && adfDensity[0] > SRC_DENSITY_THRESHOLD )
        {
            dfAccumulatorDivisor += dfMult1;

            dfAccumulatorReal += adfReal[0] * dfMult1;
            dfAccumulatorImag += adfImag[0] * dfMult1;
            dfAccumulatorDensity += adfDensity[0] * dfMult1;
        }

        // Lower Right Pixel.
        if( iSrcX+1 >= 0 && iSrcX+1 < nSrcXSize
            && adfDensity[1] > SRC_DENSITY_THRESHOLD )
        {
            dfAccumulatorDivisor += dfMult2;

            dfAccumulatorReal += adfReal[1] * dfMult2;
            dfAccumulatorImag += adfImag[1] * dfMult2;
            dfAccumulatorDensity += adfDensity[1] * dfMult2;
        }
    }

/* -------------------------------------------------------------------- */
/*      Return result.                                                  */
/* -------------------------------------------------------------------- */
    if( dfAccumulatorDivisor == 1.0 )
    {
        *pdfReal = dfAccumulatorReal;
        *pdfImag = dfAccumulatorImag;
        *pdfDensity = dfAccumulatorDensity;
        return false;
    }
    else if( dfAccumulatorDivisor < 0.00001 )
    {
        *pdfReal = 0.0;
        *pdfImag = 0.0;
        *pdfDensity = 0.0;
        return false;
    }
    else
    {
        *pdfReal = dfAccumulatorReal / dfAccumulatorDivisor;
        *pdfImag = dfAccumulatorImag / dfAccumulatorDivisor;
        *pdfDensity = dfAccumulatorDensity / dfAccumulatorDivisor;
        return true;
    }
}

template<class T>
static bool GWKBilinearResampleNoMasks4SampleT( GDALWarpKernel *poWK, int iBand,
                                        double dfSrcX, double dfSrcY,
                                        T *pValue )

{

    const int iSrcX = static_cast<int>(floor(dfSrcX - 0.5));
    const int iSrcY = static_cast<int>(floor(dfSrcY - 0.5));
    const int iSrcOffset = iSrcX + iSrcY * poWK->nSrcXSize;
    const double dfRatioX = 1.5 - (dfSrcX - iSrcX);
    const double dfRatioY = 1.5 - (dfSrcY - iSrcY);

    T* pSrc = reinterpret_cast<T *>(poWK->papabySrcImage[iBand]);


    if( iSrcX >= 0 && iSrcX+1 < poWK->nSrcXSize
        && iSrcY >= 0 && iSrcY+1 < poWK->nSrcYSize )
    {
        // TODO(schwehr): Should be able to remove these casts.
        const double dfAccumulator =
            (static_cast<double>(pSrc[iSrcOffset]) * dfRatioX +
             static_cast<double>(pSrc[iSrcOffset+1]) * (1.0 - dfRatioX)) *
            dfRatioY +
            (static_cast<double>(pSrc[iSrcOffset+poWK->nSrcXSize]) * dfRatioX +
             static_cast<double>(pSrc[iSrcOffset+1+poWK->nSrcXSize]) *
             (1.0 - dfRatioX)) * (1.0-dfRatioY);

        *pValue = GWKRoundValueT<T>(dfAccumulator);

        return true;
    }

    double dfMult = 0.0;
    double dfAccumulatorDivisor = 0.0;
    double dfAccumulator = 0.0;

    // Upper Left Pixel.
    if( iSrcX >= 0 && iSrcX < poWK->nSrcXSize
        && iSrcY >= 0 && iSrcY < poWK->nSrcYSize )
    {
        dfMult = dfRatioX * dfRatioY;

        dfAccumulatorDivisor += dfMult;

        dfAccumulator += pSrc[iSrcOffset] * dfMult;
    }

    // Upper Right Pixel.
    if( iSrcX+1 >= 0 && iSrcX+1 < poWK->nSrcXSize
        && iSrcY >= 0 && iSrcY < poWK->nSrcYSize )
    {
        dfMult = (1.0 - dfRatioX) * dfRatioY;

        dfAccumulatorDivisor += dfMult;

        dfAccumulator += pSrc[iSrcOffset+1] * dfMult;
    }

    // Lower Right Pixel.
    if( iSrcX+1 >= 0 && iSrcX+1 < poWK->nSrcXSize
        && iSrcY+1 >= 0 && iSrcY+1 < poWK->nSrcYSize )
    {
        dfMult = (1.0 - dfRatioX) * (1.0 - dfRatioY);

        dfAccumulatorDivisor += dfMult;

        dfAccumulator += pSrc[iSrcOffset+1+poWK->nSrcXSize] * dfMult;
    }

    // Lower Left Pixel.
    if( iSrcX >= 0 && iSrcX < poWK->nSrcXSize
        && iSrcY+1 >= 0 && iSrcY+1 < poWK->nSrcYSize )
    {
        dfMult = dfRatioX * (1.0 - dfRatioY);

        dfAccumulatorDivisor += dfMult;

        dfAccumulator += pSrc[iSrcOffset+poWK->nSrcXSize] * dfMult;
    }

/* -------------------------------------------------------------------- */
/*      Return result.                                                  */
/* -------------------------------------------------------------------- */
    double dfValue = 0.0;

    if( dfAccumulatorDivisor < 0.00001 )
    {
        *pValue = 0;
        return false;
    }
    else if( dfAccumulatorDivisor == 1.0 )
    {
        dfValue = dfAccumulator;
    }
    else
    {
        dfValue = dfAccumulator / dfAccumulatorDivisor;
    }

    *pValue = GWKRoundValueT<T>(dfValue);

    return true;
}

/************************************************************************/
/*                        GWKCubicResample()                            */
/*     Set of bicubic interpolators using cubic convolution.            */
/************************************************************************/

// http://verona.fi-p.unam.mx/boris/practicas/CubConvInterp.pdf Formula 18
// or http://en.wikipedia.org/wiki/Cubic_Hermite_spline : CINTx(p_1,p0,p1,p2)
// http://en.wikipedia.org/wiki/Bicubic_interpolation: matrix notation

// TODO(schwehr): Use an inline function.
#define CubicConvolution(distance1,distance2,distance3,f0,f1,f2,f3) \
     (             f1                                               \
      + 0.5 * (distance1*(f2 - f0)                                  \
             + distance2*(2.0*f0 - 5.0*f1 + 4.0*f2 - f3)            \
             + distance3*(3.0*(f1 - f2) + f3 - f0)))

/************************************************************************/
/*                       GWKCubicComputeWeights()                       */
/************************************************************************/

// adfCoeffs[2] = 1.0 - (adfCoeffs[0] + adfCoeffs[1] - adfCoeffs[3]);

// TODO(schwehr): Use an inline function.
#define GWKCubicComputeWeights(dfX_, adfCoeffs) \
{ \
    const double dfX = dfX_; \
    const double dfHalfX = 0.5 * dfX; \
    const double dfThreeX = 3.0 * dfX; \
    const double dfHalfX2 = dfHalfX * dfX; \
 \
    adfCoeffs[0] = dfHalfX * (-1 + dfX * (2 - dfX)); \
    adfCoeffs[1] = 1 + dfHalfX2 * (-5 + dfThreeX); \
    adfCoeffs[2] = dfHalfX * (1 + dfX * (4 - dfThreeX)); \
    adfCoeffs[3] = dfHalfX2 * (-1 + dfX); \
}

// TODO(schwehr): Use an inline function.
#define CONVOL4(v1, v2) ((v1)[0] * (v2)[0] + (v1)[1] * (v2)[1] + \
                         (v1)[2] * (v2)[2] + (v1)[3] * (v2)[3])

#if 0
// Optimal (in theory...) for max 2 convolutions: 14 multiplications
// instead of 17.
// TODO(schwehr): Use an inline function.
#define GWKCubicComputeWeights_Optim2MAX(dfX_, adfCoeffs, dfHalfX) \
{ \
    const double dfX = dfX_; \
    dfHalfX = 0.5 * dfX; \
    const double dfThreeX = 3.0 * dfX; \
    const double dfXMinus1 = dfX - 1; \
 \
    adfCoeffs[0] = -1 + dfX * (2 - dfX); \
    adfCoeffs[1] = dfX * (-5 + dfThreeX); \
    /*adfCoeffs[2] = 1 + dfX * (4 - dfThreeX);*/ \
    adfCoeffs[2] = -dfXMinus1 - adfCoeffs[1]; \
    /*adfCoeffs[3] = dfX * (-1 + dfX); */ \
    adfCoeffs[3] = dfXMinus1 - adfCoeffs[0]; \
}

// TODO(schwehr): Use an inline function.
#define CONVOL4_Optim2MAX(adfCoeffs, v, dfHalfX) \
      ((v)[1] + (dfHalfX) * (\
          (adfCoeffs)[0] * (v)[0] + (adfCoeffs)[1] * (v)[1] + \
          (adfCoeffs)[2] * (v)[2] + (adfCoeffs)[3] * (v)[3]))
#endif

static bool GWKCubicResample4Sample( GDALWarpKernel *poWK, int iBand,
                                     double dfSrcX, double dfSrcY,
                                     double *pdfDensity,
                                     double *pdfReal, double *pdfImag )

{
    const int iSrcX = static_cast<int>(dfSrcX - 0.5);
    const int iSrcY = static_cast<int>(dfSrcY - 0.5);
    const int  iSrcOffset = iSrcX + iSrcY * poWK->nSrcXSize;
    const double dfDeltaX = dfSrcX - 0.5 - iSrcX;
    const double dfDeltaY = dfSrcY - 0.5 - iSrcY;
    double adfDensity[4] = {};
    double adfReal[4] = {};
    double adfImag[4] = {};
    int i;

    // Get the bilinear interpolation at the image borders.
    if( iSrcX - 1 < 0 || iSrcX + 2 >= poWK->nSrcXSize
        || iSrcY - 1 < 0 || iSrcY + 2 >= poWK->nSrcYSize )
        return GWKBilinearResample4Sample( poWK, iBand, dfSrcX, dfSrcY,
                                           pdfDensity, pdfReal, pdfImag );

    double adfValueDens[4] = {};
    double adfValueReal[4] = {};
    double adfValueImag[4] = {};

    double adfCoeffsX[4] = {};
    GWKCubicComputeWeights(dfDeltaX, adfCoeffsX);

    for( i = -1; i < 3; i++ )
    {
        if( !GWKGetPixelRow(poWK, iBand, iSrcOffset + i * poWK->nSrcXSize - 1,
                            2, adfDensity, adfReal, adfImag)
            || adfDensity[0] < SRC_DENSITY_THRESHOLD
            || adfDensity[1] < SRC_DENSITY_THRESHOLD
            || adfDensity[2] < SRC_DENSITY_THRESHOLD
            || adfDensity[3] < SRC_DENSITY_THRESHOLD )
        {
            return GWKBilinearResample4Sample( poWK, iBand, dfSrcX, dfSrcY,
                                       pdfDensity, pdfReal, pdfImag );
        }

        adfValueDens[i + 1] = CONVOL4(adfCoeffsX, adfDensity);
        adfValueReal[i + 1] = CONVOL4(adfCoeffsX, adfReal);
        adfValueImag[i + 1] = CONVOL4(adfCoeffsX, adfImag);
    }

/* -------------------------------------------------------------------- */
/*      For now, if we have any pixels missing in the kernel area,      */
/*      we fallback on using bilinear interpolation.  Ideally we        */
/*      should do "weight adjustment" of our results similarly to       */
/*      what is done for the cubic spline and lanc. interpolators.      */
/* -------------------------------------------------------------------- */

    double adfCoeffsY[4] = {};
    GWKCubicComputeWeights(dfDeltaY, adfCoeffsY);

    *pdfDensity = CONVOL4(adfCoeffsY, adfValueDens);
    *pdfReal    = CONVOL4(adfCoeffsY, adfValueReal);
    *pdfImag    = CONVOL4(adfCoeffsY, adfValueImag);

    return true;
}

// We do not define USE_SSE_CUBIC_IMPL since in practice, it gives zero
// perf benefit.

#if defined(USE_SSE_CUBIC_IMPL) && (defined(__x86_64) || defined(_M_X64))

/************************************************************************/
/*                           XMMLoad4Values()                           */
/*                                                                      */
/*  Load 4 packed byte or uint16, cast them to float and put them in a  */
/*  m128 register.                                                      */
/************************************************************************/

static CPL_INLINE __m128 XMMLoad4Values(const GByte* ptr)
{
#ifdef CPL_CPU_REQUIRES_ALIGNED_ACCESS
    unsigned int i;
    memcpy(&i, ptr, 4);
    __m128i xmm_i = _mm_cvtsi32_si128(s);
#else
    __m128i xmm_i = _mm_cvtsi32_si128(*(unsigned int*)(ptr));
#endif
    // Zero extend 4 packed unsigned 8-bit integers in a to packed
    // 32-bit integers.
#if __SSE4_1__
    xmm_i = _mm_cvtepu8_epi32(xmm_i);
#else
    xmm_i = _mm_unpacklo_epi8(xmm_i, _mm_setzero_si128());
    xmm_i = _mm_unpacklo_epi16(xmm_i, _mm_setzero_si128());
#endif
    return _mm_cvtepi32_ps(xmm_i);
}

static CPL_INLINE __m128 XMMLoad4Values(const GUInt16* ptr)
{
#ifdef CPL_CPU_REQUIRES_ALIGNED_ACCESS
    GUInt64 i;
    memcpy(&i, ptr, 8);
    __m128i xmm_i =  _mm_cvtsi64_si128(s);
#else
    __m128i xmm_i =  _mm_cvtsi64_si128(*(GUInt64*)(ptr));
#endif
    // Zero extend 4 packed unsigned 16-bit integers in a to packed
    // 32-bit integers.
#if __SSE4_1__
    xmm_i = _mm_cvtepu16_epi32(xmm_i);
#else
    xmm_i = _mm_unpacklo_epi16(xmm_i, _mm_setzero_si128());
#endif
    return _mm_cvtepi32_ps(xmm_i);
}

/************************************************************************/
/*                           XMMHorizontalAdd()                         */
/*                                                                      */
/*  Return the sum of the 4 floating points of the register.            */
/************************************************************************/

#if __SSE3__
static CPL_INLINE float XMMHorizontalAdd(__m128 v)
{
    __m128 shuf = _mm_movehdup_ps(v);        // (v3   , v3   , v1   , v1)
    __m128 sums = _mm_add_ps(v, shuf);       // (v3+v3, v3+v2, v1+v1, v1+v0)
    shuf        = _mm_movehl_ps(shuf, sums); // (v3   , v3   , v3+v3, v3+v2)
    sums        = _mm_add_ss(sums, shuf);    // (v1+v0)+(v3+v2)
    return        _mm_cvtss_f32(sums);
}
#else
static CPL_INLINE float XMMHorizontalAdd(__m128 v)
{
    __m128 shuf = _mm_movehl_ps(v, v);           // (v3   , v2   , v3   , v2)
    __m128 sums = _mm_add_ps(v, shuf);           // (v3+v3, v2+v2, v3+v1, v2+v0)
    shuf        = _mm_shuffle_ps(sums, sums, 1); // (v2+v0, v2+v0, v2+v0, v3+v1)
    sums        = _mm_add_ss(sums, shuf);        // (v2+v0)+(v3+v1)
    return _mm_cvtss_f32(sums);
}
#endif

#endif // defined(USE_SSE_CUBIC_IMPL) && (defined(__x86_64) || defined(_M_X64))

/************************************************************************/
/*            GWKCubicResampleSrcMaskIsDensity4SampleRealT()            */
/************************************************************************/

// Note: if USE_SSE_CUBIC_IMPL, only instantiate that for Byte and UInt16,
// because there are a few assumptions above those types.

template<class T>
static CPL_INLINE bool GWKCubicResampleSrcMaskIsDensity4SampleRealT(
    GDALWarpKernel *poWK, int iBand,
    double dfSrcX, double dfSrcY,
    double *pdfDensity,
    double *pdfReal )
{
    const int iSrcX = static_cast<int>(dfSrcX - 0.5);
    const int iSrcY = static_cast<int>(dfSrcY - 0.5);
    const int iSrcOffset = iSrcX + iSrcY * poWK->nSrcXSize;

    // Get the bilinear interpolation at the image borders.
    if( iSrcX - 1 < 0 || iSrcX + 2 >= poWK->nSrcXSize
        || iSrcY - 1 < 0 || iSrcY + 2 >= poWK->nSrcYSize )
    {
        double adfImagIgnored[4] = {};
        return GWKBilinearResample4Sample(poWK, iBand, dfSrcX, dfSrcY,
                                          pdfDensity, pdfReal, adfImagIgnored);
    }

#if defined(USE_SSE_CUBIC_IMPL) && (defined(__x86_64) || defined(_M_X64))
    const float fDeltaX = static_cast<float>(dfSrcX) - 0.5f - iSrcX;
    const float fDeltaY = static_cast<float>(dfSrcY) - 0.5f - iSrcY;

    // TODO(schwehr): Explain the magic numbers.
    float  afTemp[4 + 4 + 4 + 1];
    float* pafAligned =
        reinterpret_cast<float*>(afTemp + ((size_t)afTemp & 0xf));
    float* pafCoeffs = pafAligned;
    float* pafDensity = pafAligned + 4;
    float* pafValue = pafAligned + 8;

    const float fHalfDeltaX = 0.5f * fDeltaX;
    const float fThreeDeltaX = 3.0f * fDeltaX;
    const float fHalfDeltaX2 = fHalfDeltaX * fDeltaX;

    pafCoeffs[0] = fHalfDeltaX * (-1 + fDeltaX * (2 - fDeltaX));
    pafCoeffs[1] = 1 + fHalfDeltaX2 * (-5 + fThreeDeltaX);
    pafCoeffs[2] = fHalfDeltaX * (1 + fDeltaX * (4 - fThreeDeltaX));
    pafCoeffs[3] = fHalfDeltaX2 * (-1 + fDeltaX);
    __m128 xmmCoeffs = _mm_load_ps(pafCoeffs);
    const __m128 xmmThreshold = _mm_load1_ps(&SRC_DENSITY_THRESHOLD);

    __m128 xmmMaskLowDensity = _mm_setzero_ps();
    for( int i = -1, iOffset = iSrcOffset - poWK->nSrcXSize - 1;
         i < 3; i++, iOffset += poWK->nSrcXSize )
    {
        const __m128 xmmDensity = _mm_loadu_ps(
                                    poWK->pafUnifiedSrcDensity + iOffset);
        xmmMaskLowDensity = _mm_or_ps(xmmMaskLowDensity,
                                      _mm_cmplt_ps(xmmDensity, xmmThreshold));
        pafDensity[i + 1] = XMMHorizontalAdd(_mm_mul_ps(xmmCoeffs, xmmDensity));

        const __m128 xmmValues = XMMLoad4Values(
                              ((T*) poWK->papabySrcImage[iBand]) + iOffset);
        pafValue[i + 1] = XMMHorizontalAdd(_mm_mul_ps(xmmCoeffs, xmmValues));
    }
    if( _mm_movemask_ps(xmmMaskLowDensity) )
    {
        double adfImagIgnored[4] = {};
        return GWKBilinearResample4Sample( poWK, iBand, dfSrcX, dfSrcY,
                                  pdfDensity, pdfReal, adfImagIgnored );
    }

    const float fHalfDeltaY = 0.5f * fDeltaY;
    const float fThreeDeltaY = 3.0f * fDeltaY;
    const float fHalfDeltaY2 = fHalfDeltaY * fDeltaY;

    pafCoeffs[0] = fHalfDeltaY * (-1 + fDeltaY * (2 - fDeltaY));
    pafCoeffs[1] = 1 + fHalfDeltaY2 * (-5 + fThreeDeltaY);
    pafCoeffs[2] = fHalfDeltaY * (1 + fDeltaY * (4 - fThreeDeltaY));
    pafCoeffs[3] = fHalfDeltaY2 * (-1 + fDeltaY);

    xmmCoeffs = _mm_load_ps(pafCoeffs);

    const __m128 xmmDensity = _mm_load_ps(pafDensity);
    const __m128 xmmValue = _mm_load_ps(pafValue);
    *pdfDensity = XMMHorizontalAdd(_mm_mul_ps(xmmCoeffs, xmmDensity));
    *pdfReal = XMMHorizontalAdd(_mm_mul_ps(xmmCoeffs, xmmValue));

    // We did all above computations on float32 whereas the general case is
    // float64. Not sure if one is fundamentally more correct than the other
    // one, but we want our optimization to give the same result as the
    // general case as much as possible, so if the resulting value is
    // close to some_int_value + 0.5, redo the computation with the general
    // case.
    // Note: If other types than Byte or UInt16, will need changes.
    if( fabs(*pdfReal - static_cast<int>(*pdfReal) - 0.5) > .007 )
        return true;

#endif // defined(USE_SSE_CUBIC_IMPL) && (defined(__x86_64) || defined(_M_X64))

    const double  dfDeltaX = dfSrcX - 0.5 - iSrcX;
    const double  dfDeltaY = dfSrcY - 0.5 - iSrcY;

    double adfValueDens[4] = {};
    double adfValueReal[4] = {};

    double adfCoeffsX[4] = {};
    GWKCubicComputeWeights(dfDeltaX, adfCoeffsX);

    double adfCoeffsY[4] = {};
    GWKCubicComputeWeights(dfDeltaY, adfCoeffsY);

    for( int i = -1; i < 3; i++ )
    {
        const int iOffset = iSrcOffset+i*poWK->nSrcXSize - 1;
#if !(defined(USE_SSE_CUBIC_IMPL) && (defined(__x86_64) || defined(_M_X64)))
        if( poWK->pafUnifiedSrcDensity[iOffset + 0] < SRC_DENSITY_THRESHOLD ||
            poWK->pafUnifiedSrcDensity[iOffset + 1] < SRC_DENSITY_THRESHOLD ||
            poWK->pafUnifiedSrcDensity[iOffset + 2] < SRC_DENSITY_THRESHOLD ||
            poWK->pafUnifiedSrcDensity[iOffset + 3] < SRC_DENSITY_THRESHOLD )
        {
            double adfImagIgnored[4] = {};
            return GWKBilinearResample4Sample( poWK, iBand, dfSrcX, dfSrcY,
                                      pdfDensity, pdfReal, adfImagIgnored );
        }
#endif

        adfValueDens[i + 1] = CONVOL4(
            adfCoeffsX,
            poWK->pafUnifiedSrcDensity + iOffset);

        adfValueReal[i + 1] = CONVOL4(
            adfCoeffsX,
            reinterpret_cast<T*>(poWK->papabySrcImage[iBand]) + iOffset);
    }

    *pdfDensity = CONVOL4(adfCoeffsY, adfValueDens);
    *pdfReal    = CONVOL4(adfCoeffsY, adfValueReal);

    return true;
}

/************************************************************************/
/*              GWKCubicResampleSrcMaskIsDensity4SampleReal()             */
/*     Bi-cubic when source has and only has pafUnifiedSrcDensity.      */
/************************************************************************/

static bool GWKCubicResampleSrcMaskIsDensity4SampleReal(
                             GDALWarpKernel *poWK, int iBand,
                             double dfSrcX, double dfSrcY,
                             double *pdfDensity,
                             double *pdfReal )

{
    const int iSrcX = static_cast<int>(dfSrcX - 0.5);
    const int iSrcY = static_cast<int>(dfSrcY - 0.5);
    const int iSrcOffset = iSrcX + iSrcY * poWK->nSrcXSize;
    const double dfDeltaX = dfSrcX - 0.5 - iSrcX;
    const double dfDeltaY = dfSrcY - 0.5 - iSrcY;

    // Get the bilinear interpolation at the image borders.
    if( iSrcX - 1 < 0 || iSrcX + 2 >= poWK->nSrcXSize
        || iSrcY - 1 < 0 || iSrcY + 2 >= poWK->nSrcYSize )
    {
        double adfImagIgnored[4] = {};
        return GWKBilinearResample4Sample(poWK, iBand, dfSrcX, dfSrcY,
                                          pdfDensity, pdfReal, adfImagIgnored);
    }

    double adfCoeffsX[4] = {};
    GWKCubicComputeWeights(dfDeltaX, adfCoeffsX);

    double adfCoeffsY[4] = {};
    GWKCubicComputeWeights(dfDeltaY, adfCoeffsY);

    double adfValueDens[4] = {};
    double adfValueReal[4] = {};
    double adfDensity[4] = {};
    double adfReal[4] = {};
    double adfImagIgnored[4] = {};

    for( int i = -1; i < 3; i++ )
    {
        if( !GWKGetPixelRow(poWK, iBand, iSrcOffset + i * poWK->nSrcXSize - 1,
                            2, adfDensity, adfReal, adfImagIgnored)
            || adfDensity[0] < SRC_DENSITY_THRESHOLD
            || adfDensity[1] < SRC_DENSITY_THRESHOLD
            || adfDensity[2] < SRC_DENSITY_THRESHOLD
            || adfDensity[3] < SRC_DENSITY_THRESHOLD )
        {
            return GWKBilinearResample4Sample( poWK, iBand, dfSrcX, dfSrcY,
                                      pdfDensity, pdfReal, adfImagIgnored );
        }

        adfValueDens[i + 1] = CONVOL4(adfCoeffsX, adfDensity);
        adfValueReal[i + 1] = CONVOL4(adfCoeffsX, adfReal);
    }

    *pdfDensity = CONVOL4(adfCoeffsY, adfValueDens);
    *pdfReal    = CONVOL4(adfCoeffsY, adfValueReal);

    return true;
}

template<class T>
static bool GWKCubicResampleNoMasks4SampleT( GDALWarpKernel *poWK, int iBand,
                                     double dfSrcX, double dfSrcY,
                                     T *pValue )

{
    const int iSrcX = static_cast<int>(dfSrcX - 0.5);
    const int iSrcY = static_cast<int>(dfSrcY - 0.5);
    const int iSrcOffset = iSrcX + iSrcY * poWK->nSrcXSize;
    const double dfDeltaX = dfSrcX - 0.5 - iSrcX;
    const double dfDeltaY = dfSrcY - 0.5 - iSrcY;
    const double dfDeltaY2 = dfDeltaY * dfDeltaY;
    const double dfDeltaY3 = dfDeltaY2 * dfDeltaY;

    // Get the bilinear interpolation at the image borders.
    if( iSrcX - 1 < 0 || iSrcX + 2 >= poWK->nSrcXSize
        || iSrcY - 1 < 0 || iSrcY + 2 >= poWK->nSrcYSize )
        return GWKBilinearResampleNoMasks4SampleT ( poWK, iBand, dfSrcX, dfSrcY,
                                             pValue );

    double adfCoeffs[4] = {};
    GWKCubicComputeWeights(dfDeltaX, adfCoeffs);

    double adfValue[4] = {};

    for( int i = -1; i < 3; i++ )
    {
        const int iOffset = iSrcOffset + i * poWK->nSrcXSize - 1;

        adfValue[i + 1] = CONVOL4(
            adfCoeffs,
            reinterpret_cast<T*>(poWK->papabySrcImage[iBand]) + iOffset);
    }

    const double dfValue = CubicConvolution(
        dfDeltaY, dfDeltaY2, dfDeltaY3,
        adfValue[0], adfValue[1], adfValue[2], adfValue[3]);

    *pValue = GWKClampValueT<T>(dfValue);

    return true;
}

/************************************************************************/
/*                          GWKLanczosSinc()                            */
/************************************************************************/

/*
 * Lanczos windowed sinc interpolation kernel with radius r.
 *        /
 *        | sinc(x) * sinc(x/r), if |x| < r
 * L(x) = | 1, if x = 0                     ,
 *        | 0, otherwise
 *        \
 *
 * where sinc(x) = sin(PI * x) / (PI * x).
 */

static double GWKLanczosSinc( double dfX )
{
    if( dfX == 0.0 )
        return 1.0;

    const double dfPIX = M_PI * dfX;
    const double dfPIXoverR = dfPIX / 3;
    const double dfPIX2overR = dfPIX * dfPIXoverR;
    return sin(dfPIX) * sin(dfPIXoverR) / dfPIX2overR;
}

static double GWKLanczosSinc4Values( double* padfValues )
{
    for( int i = 0; i < 4; i++ )
    {
        if( padfValues[i] == 0.0 )
        {
            padfValues[i] = 1.0;
        }
        else
        {
            const double dfPIX = M_PI * padfValues[i];
            const double dfPIXoverR = dfPIX / 3;
            const double dfPIX2overR = dfPIX * dfPIXoverR;
            padfValues[i] = sin(dfPIX) * sin(dfPIXoverR) / dfPIX2overR;
        }
    }
    return padfValues[0] + padfValues[1] + padfValues[2] + padfValues[3];
}

/************************************************************************/
/*                           GWKBilinear()                              */
/************************************************************************/

static double GWKBilinear( double dfX )
{
    double dfAbsX = fabs(dfX);
    if( dfAbsX <= 1.0 )
        return 1 - dfAbsX;
    else
        return 0.0;
}

static double GWKBilinear4Values( double* padfValues )
{
    double dfAbsX0 = fabs(padfValues[0]);
    double dfAbsX1 = fabs(padfValues[1]);
    double dfAbsX2 = fabs(padfValues[2]);
    double dfAbsX3 = fabs(padfValues[3]);
    if( dfAbsX0 <= 1.0 )
        padfValues[0] = 1 - dfAbsX0;
    else
        padfValues[0] = 0.0;
    if( dfAbsX1 <= 1.0 )
        padfValues[1] = 1 - dfAbsX1;
    else
        padfValues[1] = 0.0;
    if( dfAbsX2 <= 1.0 )
        padfValues[2] = 1 - dfAbsX2;
    else
        padfValues[2] = 0.0;
    if( dfAbsX3 <= 1.0 )
        padfValues[3] = 1 - dfAbsX3;
    else
        padfValues[3] = 0.0;
    return  padfValues[0] +  padfValues[1] + padfValues[2] + padfValues[3];
}

/************************************************************************/
/*                            GWKCubic()                                */
/************************************************************************/

static double GWKCubic( double dfX )
{
    // http://en.wikipedia.org/wiki/Bicubic_interpolation#Bicubic_convolution_algorithm
    // W(x) formula with a = -0.5 (cubic hermite spline )
    // or https://www.cs.utexas.edu/~fussell/courses/cs384g-fall2013/lectures/mitchell/Mitchell.pdf
    // k(x) (formula 8) with (B,C)=(0,0.5) the Catmull-Rom spline
    double dfAbsX = fabs(dfX);
    if( dfAbsX <= 1.0 )
    {
        double dfX2 = dfX * dfX;
        return dfX2 * (1.5 * dfAbsX - 2.5) + 1;
    }
    else if( dfAbsX <= 2.0 )
    {
        double dfX2 = dfX * dfX;
        return dfX2 * (-0.5 * dfAbsX + 2.5) - 4 * dfAbsX + 2;
    }
    else
        return 0.0;
}

static double GWKCubic4Values( double* padfValues )
{
    const double dfAbsX_0 = fabs(padfValues[0]);
    const double dfAbsX_1 = fabs(padfValues[1]);
    const double dfAbsX_2 = fabs(padfValues[2]);
    const double dfAbsX_3 = fabs(padfValues[3]);
    const double dfX2_0 = padfValues[0] * padfValues[0];
    const double dfX2_1 = padfValues[1] * padfValues[1];
    const double dfX2_2 = padfValues[2] * padfValues[2];
    const double dfX2_3 = padfValues[3] * padfValues[3];

    double dfVal0 = 0.0;
    if( dfAbsX_0 <= 1.0 )
        dfVal0 = dfX2_0 * (1.5 * dfAbsX_0 - 2.5) + 1.0;
    else if( dfAbsX_0 <= 2.0 )
        dfVal0 = dfX2_0 * (-0.5 * dfAbsX_0 + 2.5) - 4.0 * dfAbsX_0 + 2.0;

    double dfVal1 = 0.0;
    if( dfAbsX_1 <= 1.0 )
        dfVal1 = dfX2_1 * (1.5 * dfAbsX_1 - 2.5) + 1.0;
    else if( dfAbsX_1 <= 2.0 )
        dfVal1 = dfX2_1 * (-0.5 * dfAbsX_1 + 2.5) - 4.0 * dfAbsX_1 + 2.0;

    double dfVal2 = 0.0;
    if( dfAbsX_2 <= 1.0 )
        dfVal2 = dfX2_2 * (1.5 * dfAbsX_2 - 2.5) + 1.0;
    else if( dfAbsX_2 <= 2.0 )
        dfVal2 = dfX2_2 * (-0.5 * dfAbsX_2 + 2.5) - 4.0 * dfAbsX_2 + 2.0;

    double dfVal3 = 0.0;
    if( dfAbsX_3 <= 1.0 )
        dfVal3 = dfX2_3 * (1.5 * dfAbsX_3 - 2.5) + 1.0;
    else if( dfAbsX_3 <= 2.0 )
        dfVal3 = dfX2_3 * (-0.5 * dfAbsX_3 + 2.5) - 4.0 * dfAbsX_3 + 2.0;

    padfValues[0] = dfVal0;
    padfValues[1] = dfVal1;
    padfValues[2] = dfVal2;
    padfValues[3] = dfVal3;
    return dfVal0 + dfVal1 + dfVal2 + dfVal3;
}

/************************************************************************/
/*                           GWKBSpline()                               */
/************************************************************************/

// https://www.cs.utexas.edu/~fussell/courses/cs384g-fall2013/lectures/mitchell/Mitchell.pdf
// Equation 8 with (B,C)=(1,0)
// 1/6 * ( 3 * |x|^3 -  6 * |x|^2 + 4) |x| < 1
// 1/6 * ( -|x|^3 + 6 |x|^2  - 12|x| + 8) |x| >= 1 and |x| < 2

static double GWKBSpline( double x )
{
    const double xp2 = x + 2.0;
    const double xp1 = x + 1.0;
    const double xm1 = x - 1.0;

    // This will most likely be used, so we'll compute it ahead of time to
    // avoid stalling the processor.
    const double xp2c = xp2 * xp2 * xp2;

    // Note that the test is computed only if it is needed.
    // TODO(schwehr): Make this easier to follow.
    return
        xp2 > 0.0
        ? ((xp1 > 0.0)?((x > 0.0)?((xm1 > 0.0)?
                                   -4.0 * xm1*xm1*xm1:0.0) +
                        6.0 * x*x*x:0.0) +
           -4.0 * xp1*xp1*xp1:0.0) + xp2c
        : 0.0;  // * 0.166666666666666666666
}

static double GWKBSpline4Values( double* padfValues )
{
    for( int i = 0; i < 4; i++ )
    {
        const double x = padfValues[i];
        const double xp2 = x + 2.0;
        const double xp1 = x + 1.0;
        const double xm1 = x - 1.0;

        // This will most likely be used, so we'll compute it ahead of time to
        // avoid stalling the processor.
        const double xp2c = xp2 * xp2 * xp2;

        // Note that the test is computed only if it is needed.
        // TODO(schwehr): Make this easier to follow.
        padfValues[i] = (xp2 > 0.0)?((xp1 > 0.0)?((x > 0.0)?((xm1 > 0.0)?
                                                    -4.0 * xm1*xm1*xm1:0.0) +
                                        6.0 * x*x*x:0.0) +
                            -4.0 * xp1*xp1*xp1:0.0) +
                xp2c:0.0;  // * 0.166666666666666666666
    }
    return padfValues[0] + padfValues[1] + padfValues[2] + padfValues[3];
}
/************************************************************************/
/*                       GWKResampleWrkStruct                           */
/************************************************************************/

typedef struct _GWKResampleWrkStruct GWKResampleWrkStruct;

typedef bool (*pfnGWKResampleType) ( GDALWarpKernel *poWK, int iBand,
                                     double dfSrcX, double dfSrcY,
                                     double *pdfDensity,
                                     double *pdfReal, double *pdfImag,
                                     GWKResampleWrkStruct* psWrkStruct );

struct _GWKResampleWrkStruct
{
    pfnGWKResampleType pfnGWKResample;

    // Space for saved X weights.
    double  *padfWeightsX;
    bool    *pabCalcX;

    double  *padfWeightsY; // Only used by GWKResampleOptimizedLanczos.
    int      iLastSrcX; // Only used by GWKResampleOptimizedLanczos.
    int      iLastSrcY; // Only used by GWKResampleOptimizedLanczos.
    double   dfLastDeltaX; // Only used by GWKResampleOptimizedLanczos.
    double   dfLastDeltaY; // Only used by GWKResampleOptimizedLanczos.

    // Space for saving a row of pixels.
    double  *padfRowDensity;
    double  *padfRowReal;
    double  *padfRowImag;
};

/************************************************************************/
/*                    GWKResampleCreateWrkStruct()                      */
/************************************************************************/

static bool GWKResample( GDALWarpKernel *poWK, int iBand,
                        double dfSrcX, double dfSrcY,
                        double *pdfDensity,
                        double *pdfReal, double *pdfImag,
                        GWKResampleWrkStruct* psWrkStruct );

static bool GWKResampleOptimizedLanczos( GDALWarpKernel *poWK, int iBand,
                                        double dfSrcX, double dfSrcY,
                                        double *pdfDensity,
                                        double *pdfReal, double *pdfImag,
                                        GWKResampleWrkStruct* psWrkStruct );

static GWKResampleWrkStruct* GWKResampleCreateWrkStruct(GDALWarpKernel *poWK)
{
    const int nXDist = ( poWK->nXRadius + 1 ) * 2;
    const int nYDist = ( poWK->nYRadius + 1 ) * 2;

    GWKResampleWrkStruct* psWrkStruct = static_cast<GWKResampleWrkStruct *>(
        CPLMalloc(sizeof(GWKResampleWrkStruct)));

    // Alloc space for saved X weights.
    psWrkStruct->padfWeightsX =
        static_cast<double *>(CPLCalloc( nXDist, sizeof(double)));
    psWrkStruct->pabCalcX =
        static_cast<bool *>(CPLMalloc(nXDist * sizeof(bool)));

    psWrkStruct->padfWeightsY =
        static_cast<double *>(CPLCalloc(nYDist, sizeof(double)));
    psWrkStruct->iLastSrcX = -10;
    psWrkStruct->iLastSrcY = -10;
    psWrkStruct->dfLastDeltaX = -10;
    psWrkStruct->dfLastDeltaY = -10;

    // Alloc space for saving a row of pixels.
    if( poWK->pafUnifiedSrcDensity == nullptr &&
        poWK->panUnifiedSrcValid == nullptr &&
        poWK->papanBandSrcValid == nullptr )
    {
        psWrkStruct->padfRowDensity = nullptr;
    }
    else
    {
        psWrkStruct->padfRowDensity =
            static_cast<double *>(CPLCalloc(nXDist, sizeof(double)));
    }
    psWrkStruct->padfRowReal =
        static_cast<double *>(CPLCalloc(nXDist, sizeof(double)));
    psWrkStruct->padfRowImag =
        static_cast<double *>(CPLCalloc(nXDist, sizeof(double)));

    if( poWK->eResample == GRA_Lanczos )
    {
        psWrkStruct->pfnGWKResample = GWKResampleOptimizedLanczos;

        const double dfXScale = poWK->dfXScale;
        if( dfXScale < 1.0 )
        {
            int iMin = poWK->nFiltInitX;
            int iMax = poWK->nXRadius;
            while( iMin * dfXScale < -3.0 )
                iMin++;
            while( iMax * dfXScale > 3.0 )
                iMax--;

            for( int i = iMin; i <= iMax; ++i )
            {
                psWrkStruct->padfWeightsX[i-poWK->nFiltInitX] =
                    GWKLanczosSinc(i * dfXScale);
            }
        }

        const double dfYScale = poWK->dfYScale;
        if( dfYScale < 1.0 )
        {
            int jMin = poWK->nFiltInitY;
            int jMax = poWK->nYRadius;
            while( jMin * dfYScale < -3.0 )
                jMin++;
            while( jMax * dfYScale > 3.0 )
                jMax--;

            for( int j = jMin; j <= jMax; ++j )
            {
                psWrkStruct->padfWeightsY[j-poWK->nFiltInitY] =
                    GWKLanczosSinc(j * dfYScale);
            }
        }
    }
    else
        psWrkStruct->pfnGWKResample = GWKResample;

    return psWrkStruct;
}

/************************************************************************/
/*                    GWKResampleDeleteWrkStruct()                      */
/************************************************************************/

static void GWKResampleDeleteWrkStruct(GWKResampleWrkStruct* psWrkStruct)
{
    CPLFree( psWrkStruct->padfWeightsX );
    CPLFree( psWrkStruct->padfWeightsY );
    CPLFree( psWrkStruct->pabCalcX );
    CPLFree( psWrkStruct->padfRowDensity );
    CPLFree( psWrkStruct->padfRowReal );
    CPLFree( psWrkStruct->padfRowImag );
    CPLFree( psWrkStruct );
}

/************************************************************************/
/*                           GWKResample()                              */
/************************************************************************/

static bool GWKResample( GDALWarpKernel *poWK, int iBand,
                        double dfSrcX, double dfSrcY,
                        double *pdfDensity,
                        double *pdfReal, double *pdfImag,
                        GWKResampleWrkStruct* psWrkStruct )

{
    // Save as local variables to avoid following pointers in loops.
    const int nSrcXSize = poWK->nSrcXSize;
    const int nSrcYSize = poWK->nSrcYSize;

    double dfAccumulatorReal = 0.0;
    double dfAccumulatorImag = 0.0;
    double dfAccumulatorDensity = 0.0;
    double dfAccumulatorWeight = 0.0;
    const int iSrcX = static_cast<int>(floor(dfSrcX - 0.5));
    const int iSrcY = static_cast<int>(floor(dfSrcY - 0.5));
    const int iSrcOffset = iSrcX + iSrcY * nSrcXSize;
    const double dfDeltaX = dfSrcX - 0.5 - iSrcX;
    const double dfDeltaY = dfSrcY - 0.5 - iSrcY;

    const double dfXScale = poWK->dfXScale;
    const double dfYScale = poWK->dfYScale;

    const int nXDist = ( poWK->nXRadius + 1 ) * 2;

    // Space for saved X weights.
    double *padfWeightsX = psWrkStruct->padfWeightsX;
    bool *pabCalcX = psWrkStruct->pabCalcX;

    // Space for saving a row of pixels.
    double *padfRowDensity = psWrkStruct->padfRowDensity;
    double *padfRowReal = psWrkStruct->padfRowReal;
    double *padfRowImag = psWrkStruct->padfRowImag;

    // Mark as needing calculation (don't calculate the weights yet,
    // because a mask may render it unnecessary).
    memset( pabCalcX, false, nXDist * sizeof(bool) );

    FilterFuncType pfnGetWeight = apfGWKFilter[poWK->eResample];
    CPLAssert(pfnGetWeight);

    // Skip sampling over edge of image.
    int j = poWK->nFiltInitY;
    int jMax= poWK->nYRadius;
    if( iSrcY + j < 0 )
        j = -iSrcY;
    if( iSrcY + jMax >= nSrcYSize )
        jMax = nSrcYSize - iSrcY - 1;

    int iMin = poWK->nFiltInitX;
    int iMax = poWK->nXRadius;
    if( iSrcX + iMin < 0 )
        iMin = -iSrcX;
    if( iSrcX + iMax >= nSrcXSize )
        iMax = nSrcXSize - iSrcX - 1;

    const int bXScaleBelow1 = ( dfXScale < 1.0 );
    const int bYScaleBelow1 = ( dfYScale < 1.0 );

    int iRowOffset = iSrcOffset + (j - 1) * nSrcXSize + iMin;

    // Loop over pixel rows in the kernel.
    for( ; j <= jMax; ++j )
    {
        iRowOffset += nSrcXSize;

        // Get pixel values.
        // We can potentially read extra elements after the "normal" end of the
        // source arrays, but the contract of papabySrcImage[iBand],
        // papanBandSrcValid[iBand], panUnifiedSrcValid and pafUnifiedSrcDensity
        // is to have WARP_EXTRA_ELTS reserved at their end.
        if( !GWKGetPixelRow( poWK, iBand, iRowOffset, (iMax-iMin+2)/2,
                             padfRowDensity, padfRowReal, padfRowImag ) )
            continue;

        // Calculate the Y weight.
        double dfWeight1 = ( bYScaleBelow1 ) ?
                pfnGetWeight((j - dfDeltaY) * dfYScale):
                pfnGetWeight(j - dfDeltaY);

        // Iterate over pixels in row.
        double dfAccumulatorRealLocal = 0.0;
        double dfAccumulatorImagLocal = 0.0;
        double dfAccumulatorDensityLocal = 0.0;
        double dfAccumulatorWeightLocal = 0.0;

        for( int i = iMin; i <= iMax; ++i )
        {
            // Skip sampling if pixel has zero density.
            if( padfRowDensity != nullptr &&
                padfRowDensity[i-iMin] < SRC_DENSITY_THRESHOLD )
                continue;

            double dfWeight2 = 0.0;

            // Make or use a cached set of weights for this row.
            if( pabCalcX[i-iMin] )
            {
                // Use saved weight value instead of recomputing it.
                dfWeight2 = padfWeightsX[i-iMin];
            }
            else
            {
                // Calculate & save the X weight.
                padfWeightsX[i-iMin] = dfWeight2 = ( bXScaleBelow1 ) ?
                        pfnGetWeight((i - dfDeltaX) * dfXScale):
                        pfnGetWeight(i - dfDeltaX);

                pabCalcX[i-iMin] = true;
            }

            // Accumulate!
            dfAccumulatorRealLocal += padfRowReal[i-iMin] * dfWeight2;
            dfAccumulatorImagLocal += padfRowImag[i-iMin] * dfWeight2;
            if( padfRowDensity != nullptr )
                dfAccumulatorDensityLocal += padfRowDensity[i-iMin] * dfWeight2;
            dfAccumulatorWeightLocal += dfWeight2;
        }

        dfAccumulatorReal += dfAccumulatorRealLocal * dfWeight1;
        dfAccumulatorImag += dfAccumulatorImagLocal * dfWeight1;
        dfAccumulatorDensity += dfAccumulatorDensityLocal * dfWeight1;
        dfAccumulatorWeight += dfAccumulatorWeightLocal * dfWeight1;
    }

    if( dfAccumulatorWeight < 0.000001 ||
        (padfRowDensity != nullptr && dfAccumulatorDensity < 0.000001) )
    {
        *pdfDensity = 0.0;
        return false;
    }

    // Calculate the output taking into account weighting.
    if( dfAccumulatorWeight < 0.99999 || dfAccumulatorWeight > 1.00001 )
    {
        *pdfReal = dfAccumulatorReal / dfAccumulatorWeight;
        *pdfImag = dfAccumulatorImag / dfAccumulatorWeight;
        if( padfRowDensity != nullptr )
            *pdfDensity = dfAccumulatorDensity / dfAccumulatorWeight;
        else
            *pdfDensity = 1.0;
    }
    else
    {
        *pdfReal = dfAccumulatorReal;
        *pdfImag = dfAccumulatorImag;
        if( padfRowDensity != nullptr )
            *pdfDensity = dfAccumulatorDensity;
        else
            *pdfDensity = 1.0;
    }

    return true;
}

/************************************************************************/
/*                      GWKResampleOptimizedLanczos()                   */
/************************************************************************/

static bool GWKResampleOptimizedLanczos( GDALWarpKernel *poWK, int iBand,
                        double dfSrcX, double dfSrcY,
                        double *pdfDensity,
                        double *pdfReal, double *pdfImag,
                        GWKResampleWrkStruct* psWrkStruct )

{
    // Save as local variables to avoid following pointers in loops.
    const int nSrcXSize = poWK->nSrcXSize;
    const int nSrcYSize = poWK->nSrcYSize;

    double dfAccumulatorReal = 0.0;
    double dfAccumulatorImag = 0.0;
    double dfAccumulatorDensity = 0.0;
    double dfAccumulatorWeight = 0.0;
    const int     iSrcX = static_cast<int>(floor( dfSrcX - 0.5 ));
    const int     iSrcY = static_cast<int>(floor( dfSrcY - 0.5 ));
    const int     iSrcOffset = iSrcX + iSrcY * nSrcXSize;
    const double  dfDeltaX = dfSrcX - 0.5 - iSrcX;
    const double  dfDeltaY = dfSrcY - 0.5 - iSrcY;

    const double dfXScale = poWK->dfXScale;
    const double dfYScale = poWK->dfYScale;

    // Space for saved X weights.
    double *padfWeightsX = psWrkStruct->padfWeightsX;
    double *padfWeightsY = psWrkStruct->padfWeightsY;

    // Space for saving a row of pixels.
    double *padfRowDensity = psWrkStruct->padfRowDensity;
    double *padfRowReal = psWrkStruct->padfRowReal;
    double *padfRowImag = psWrkStruct->padfRowImag;

    // Skip sampling over edge of image.
    int jMin = poWK->nFiltInitY;
    int jMax = poWK->nYRadius;
    if( iSrcY + jMin < 0 )
        jMin = -iSrcY;
    if( iSrcY + jMax >= nSrcYSize )
        jMax = nSrcYSize - iSrcY - 1;

    int iMin = poWK->nFiltInitX;
    int iMax = poWK->nXRadius;
    if( iSrcX + iMin < 0 )
        iMin = -iSrcX;
    if( iSrcX + iMax >= nSrcXSize )
        iMax = nSrcXSize - iSrcX - 1;

    if( dfXScale < 1.0 )
    {
        while( iMin * dfXScale < -3.0 )
            iMin++;
        while( iMax * dfXScale > 3.0 )
            iMax--;
        // padfWeightsX computed in GWKResampleCreateWrkStruct.
    }
    else
    {
        while( iMin - dfDeltaX < -3.0 )
            iMin++;
        while( iMax - dfDeltaX > 3.0 )
            iMax--;

        if( iSrcX != psWrkStruct->iLastSrcX ||
            dfDeltaX != psWrkStruct->dfLastDeltaX )
        {
            // Optimisation of GWKLanczosSinc(i - dfDeltaX) based on the
            // following trigonometric formulas.

// TODO(schwehr): Move this somewhere where it can be rendered at LaTeX.
// sin(M_PI * (dfBase + k)) = sin(M_PI * dfBase) * cos(M_PI * k) + cos(M_PI * dfBase) * sin(M_PI * k)
// sin(M_PI * (dfBase + k)) = dfSinPIBase * cos(M_PI * k) + dfCosPIBase * sin(M_PI * k)
// sin(M_PI * (dfBase + k)) = dfSinPIBase * cos(M_PI * k)
// sin(M_PI * (dfBase + k)) = dfSinPIBase * (((k % 2) == 0) ? 1 : -1)

// sin(M_PI / dfR * (dfBase + k)) = sin(M_PI / dfR * dfBase) * cos(M_PI / dfR * k) + cos(M_PI / dfR * dfBase) * sin(M_PI / dfR * k)
// sin(M_PI / dfR * (dfBase + k)) = dfSinPIBaseOverR * cos(M_PI / dfR * k) + dfCosPIBaseOverR * sin(M_PI / dfR * k)

            const double dfSinPIDeltaXOver3 = sin((-M_PI / 3.0) * dfDeltaX);
            const double dfSin2PIDeltaXOver3 =
                dfSinPIDeltaXOver3 * dfSinPIDeltaXOver3;
            // Ok to use sqrt(1-sin^2) since M_PI / 3 * dfDeltaX < PI/2.
            const double dfCosPIDeltaXOver3 = sqrt(1.0 - dfSin2PIDeltaXOver3);
            const double dfSinPIDeltaX =
                (3.0 - 4 * dfSin2PIDeltaXOver3) * dfSinPIDeltaXOver3;
            const double dfInvPI2Over3 = 3.0 / (M_PI * M_PI);
            const double dfInvPI2Over3xSinPIDeltaX =
                dfInvPI2Over3 * dfSinPIDeltaX;
            const double dfInvPI2Over3xSinPIDeltaXxm0d5SinPIDeltaXOver3 =
                -0.5 * dfInvPI2Over3xSinPIDeltaX * dfSinPIDeltaXOver3;
            const double dfSinPIOver3 = 0.8660254037844386;
            const double dfInvPI2Over3xSinPIDeltaXxSinPIOver3xCosPIDeltaXOver3 =
                dfSinPIOver3 * dfInvPI2Over3xSinPIDeltaX * dfCosPIDeltaXOver3;
            const double padfCst[] = {
                dfInvPI2Over3xSinPIDeltaX * dfSinPIDeltaXOver3,
                dfInvPI2Over3xSinPIDeltaXxm0d5SinPIDeltaXOver3 -
                        dfInvPI2Over3xSinPIDeltaXxSinPIOver3xCosPIDeltaXOver3,
                dfInvPI2Over3xSinPIDeltaXxm0d5SinPIDeltaXOver3 +
                        dfInvPI2Over3xSinPIDeltaXxSinPIOver3xCosPIDeltaXOver3 };

            for( int i = iMin; i <= iMax; ++i )
            {
                const double dfX = i - dfDeltaX;
                if( dfX == 0.0 )
                    padfWeightsX[i-poWK->nFiltInitX] = 1.0;
                else
                    padfWeightsX[i-poWK->nFiltInitX] =
                                            padfCst[(i + 3) % 3] / (dfX * dfX);
#if DEBUG_VERBOSE
                // TODO(schwehr): AlmostEqual.
                //CPLAssert(fabs(padfWeightsX[i-poWK->nFiltInitX] -
                //               GWKLanczosSinc(dfX, 3.0)) < 1e-10);
#endif
            }

            psWrkStruct->iLastSrcX = iSrcX;
            psWrkStruct->dfLastDeltaX = dfDeltaX;
        }
    }

    if( dfYScale < 1.0 )
    {
        while( jMin * dfYScale < -3.0 )
            jMin++;
        while( jMax * dfYScale > 3.0 )
            jMax--;
        // padfWeightsY computed in GWKResampleCreateWrkStruct.
    }
    else
    {
        while( jMin - dfDeltaY < -3.0 )
            jMin++;
        while( jMax - dfDeltaY > 3.0 )
            jMax--;

        if( iSrcY != psWrkStruct->iLastSrcY ||
            dfDeltaY != psWrkStruct->dfLastDeltaY )
        {
            const double dfSinPIDeltaYOver3 = sin((-M_PI / 3.0) * dfDeltaY);
            const double dfSin2PIDeltaYOver3 =
                dfSinPIDeltaYOver3 * dfSinPIDeltaYOver3;
            // Ok to use sqrt(1-sin^2) since M_PI / 3 * dfDeltaY < PI/2.
            const double dfCosPIDeltaYOver3 = sqrt(1.0 - dfSin2PIDeltaYOver3);
            const double dfSinPIDeltaY =
                (3.0 - 4.0 * dfSin2PIDeltaYOver3) * dfSinPIDeltaYOver3;
            const double dfInvPI2Over3 = 3.0 / (M_PI * M_PI);
            const double dfInvPI2Over3xSinPIDeltaY =
                dfInvPI2Over3 * dfSinPIDeltaY;
            const double dfInvPI2Over3xSinPIDeltaYxm0d5SinPIDeltaYOver3 =
                -0.5 * dfInvPI2Over3xSinPIDeltaY * dfSinPIDeltaYOver3;
            const double dfSinPIOver3 = 0.8660254037844386;
            const double dfInvPI2Over3xSinPIDeltaYxSinPIOver3xCosPIDeltaYOver3 =
                dfSinPIOver3 * dfInvPI2Over3xSinPIDeltaY * dfCosPIDeltaYOver3;
            const double padfCst[] = {
                dfInvPI2Over3xSinPIDeltaY * dfSinPIDeltaYOver3,
                dfInvPI2Over3xSinPIDeltaYxm0d5SinPIDeltaYOver3 -
                        dfInvPI2Over3xSinPIDeltaYxSinPIOver3xCosPIDeltaYOver3,
                dfInvPI2Over3xSinPIDeltaYxm0d5SinPIDeltaYOver3 +
                        dfInvPI2Over3xSinPIDeltaYxSinPIOver3xCosPIDeltaYOver3 };

            for( int j = jMin; j <= jMax; ++j )
            {
                const double dfY = j - dfDeltaY;
                if( dfY == 0.0 )
                    padfWeightsY[j-poWK->nFiltInitY] = 1.0;
                else
                    padfWeightsY[j-poWK->nFiltInitY] =
                                            padfCst[(j + 3) % 3] / (dfY * dfY);
#if DEBUG_VERBOSE
                // TODO(schwehr): AlmostEqual.
                //CPLAssert(fabs(padfWeightsY[j-poWK->nFiltInitY] -
                //               GWKLanczosSinc(dfY, 3.0)) < 1e-10);
#endif
            }

            psWrkStruct->iLastSrcY = iSrcY;
            psWrkStruct->dfLastDeltaY = dfDeltaY;
        }
    }

    int iRowOffset = iSrcOffset + (jMin - 1) * nSrcXSize + iMin;

    // If we have no density information, we can simply compute the
    // accumulated weight.
    if( padfRowDensity == nullptr )
    {
        double dfRowAccWeight = 0.0;
        for( int i = iMin; i <= iMax; ++i )
        {
            dfRowAccWeight += padfWeightsX[i-poWK->nFiltInitX];
        }
        double dfColAccWeight = 0.0;
        for( int j = jMin; j <= jMax; ++j )
        {
            dfColAccWeight += padfWeightsY[j-poWK->nFiltInitY];
        }
        dfAccumulatorWeight = dfRowAccWeight * dfColAccWeight;
    }

    const bool bIsNonComplex = !GDALDataTypeIsComplex(poWK->eWorkingDataType);

    // Loop over pixel rows in the kernel.
    int nCountValid = 0;
    for( int j = jMin; j <= jMax; ++j )
    {
        iRowOffset += nSrcXSize;

        // Get pixel values.
        // We can potentially read extra elements after the "normal" end of the
        // source arrays, but the contract of papabySrcImage[iBand],
        // papanBandSrcValid[iBand], panUnifiedSrcValid and pafUnifiedSrcDensity
        // is to have WARP_EXTRA_ELTS reserved at their end.
        if( !GWKGetPixelRow( poWK, iBand, iRowOffset, (iMax-iMin+2)/2,
                             padfRowDensity, padfRowReal, padfRowImag ) )
            continue;

        const double dfWeight1 = padfWeightsY[j-poWK->nFiltInitY];

        // Iterate over pixels in row.
        if( padfRowDensity != nullptr )
        {
            for( int i = iMin; i <= iMax; ++i )
            {
                // Skip sampling if pixel has zero density.
                if( padfRowDensity[i - iMin] < SRC_DENSITY_THRESHOLD )
                    continue;

                nCountValid ++;

                //  Use a cached set of weights for this row.
                const double dfWeight2 =
                    dfWeight1 * padfWeightsX[i- poWK->nFiltInitX];

                // Accumulate!
                dfAccumulatorReal += padfRowReal[i - iMin] * dfWeight2;
                dfAccumulatorImag += padfRowImag[i - iMin] * dfWeight2;
                dfAccumulatorDensity += padfRowDensity[i - iMin] * dfWeight2;
                dfAccumulatorWeight += dfWeight2;
            }
        }
        else if( bIsNonComplex )
        {
            double dfRowAccReal = 0.0;
            for( int i = iMin; i <= iMax; ++i )
            {
                const double dfWeight2 = padfWeightsX[i- poWK->nFiltInitX];

                // Accumulate!
                dfRowAccReal += padfRowReal[i - iMin] * dfWeight2;
            }

            dfAccumulatorReal += dfRowAccReal * dfWeight1;
        }
        else
        {
            double dfRowAccReal = 0.0;
            double dfRowAccImag = 0.0;
            for( int i = iMin; i <= iMax; ++i )
            {
                const double dfWeight2 = padfWeightsX[i- poWK->nFiltInitX];

                // Accumulate!
                dfRowAccReal += padfRowReal[i - iMin] * dfWeight2;
                dfRowAccImag += padfRowImag[i - iMin] * dfWeight2;
            }

            dfAccumulatorReal += dfRowAccReal * dfWeight1;
            dfAccumulatorImag += dfRowAccImag * dfWeight1;
        }
    }

    if( dfAccumulatorWeight < 0.000001 ||
         (padfRowDensity != nullptr && (
             dfAccumulatorDensity < 0.000001 ||
             nCountValid < (jMax - jMin + 1) * (iMax - iMin + 1)/2)) )
    {
        *pdfDensity = 0.0;
        return false;
    }

    // Calculate the output taking into account weighting.
    if( dfAccumulatorWeight < 0.99999 || dfAccumulatorWeight > 1.00001 )
    {
        const double dfInvAcc = 1.0 / dfAccumulatorWeight;
        *pdfReal = dfAccumulatorReal * dfInvAcc;
        *pdfImag = dfAccumulatorImag * dfInvAcc;
        if( padfRowDensity != nullptr )
            *pdfDensity = dfAccumulatorDensity * dfInvAcc;
        else
            *pdfDensity = 1.0;
    }
    else
    {
        *pdfReal = dfAccumulatorReal;
        *pdfImag = dfAccumulatorImag;
        if( padfRowDensity != nullptr )
            *pdfDensity = dfAccumulatorDensity;
        else
            *pdfDensity = 1.0;
    }

    return true;
}

/************************************************************************/
/*                        GWKResampleNoMasksT()                         */
/************************************************************************/

template <class T>
static bool GWKResampleNoMasksT( GDALWarpKernel *poWK, int iBand,
                                double dfSrcX, double dfSrcY,
                                T *pValue, double *padfWeight )

{
    // Commonly used; save locally.
    const int nSrcXSize = poWK->nSrcXSize;
    const int nSrcYSize = poWK->nSrcYSize;

    const int iSrcX = static_cast<int>(floor(dfSrcX - 0.5));
    const int iSrcY = static_cast<int>(floor(dfSrcY - 0.5));
    const int iSrcOffset = iSrcX + iSrcY * nSrcXSize;

    const int nXRadius = poWK->nXRadius;
    const int nYRadius = poWK->nYRadius;

    // Politely refuse to process invalid coordinates or obscenely small image.
    if( iSrcX >= nSrcXSize || iSrcY >= nSrcYSize
        || nXRadius > nSrcXSize || nYRadius > nSrcYSize )
        return GWKBilinearResampleNoMasks4SampleT(poWK, iBand, dfSrcX, dfSrcY,
                                                  pValue);

    T* pSrcBand = reinterpret_cast<T*>(poWK->papabySrcImage[iBand]);
    const double dfDeltaX = dfSrcX - 0.5 - iSrcX;
    const double dfDeltaY = dfSrcY - 0.5 - iSrcY;

    const FilterFuncType pfnGetWeight = apfGWKFilter[poWK->eResample];
    CPLAssert(pfnGetWeight);
    const FilterFunc4ValuesType pfnGetWeight4Values =
        apfGWKFilter4Values[poWK->eResample];
    CPLAssert(pfnGetWeight4Values);

    const double dfXScale = std::min(poWK->dfXScale, 1.0);
    const double dfYScale = std::min(poWK->dfYScale, 1.0);

    // Loop over all rows in the kernel.
    double dfAccumulatorWeightHorizontal = 0.0;
    double dfAccumulatorWeightVertical = 0.0;

    int iMin = 1 - nXRadius;
    if( iSrcX + iMin < 0 )
        iMin = -iSrcX;
    int iMax = nXRadius;
    if( iSrcX + iMax >= nSrcXSize-1 )
        iMax = nSrcXSize-1 - iSrcX;
    int i = iMin;  // Used after for.
    int iC = 0;  // Used after for.
    for( ; i+2 < iMax; i += 4, iC += 4 )
    {
        padfWeight[iC] = (i - dfDeltaX) * dfXScale;
        padfWeight[iC+1] = padfWeight[iC] + dfXScale;
        padfWeight[iC+2] = padfWeight[iC+1] + dfXScale;
        padfWeight[iC+3] = padfWeight[iC+2] + dfXScale;
        dfAccumulatorWeightHorizontal += pfnGetWeight4Values(padfWeight+iC);
    }
    for( ; i <= iMax; ++i, ++iC )
    {
        const double dfWeight = pfnGetWeight((i - dfDeltaX) * dfXScale);
        padfWeight[iC] = dfWeight;
        dfAccumulatorWeightHorizontal += dfWeight;
    }

    int j = 1 - nYRadius;
    if(  iSrcY + j < 0 )
        j = -iSrcY;
    int jMax = nYRadius;
    if( iSrcY + jMax >= nSrcYSize-1 )
        jMax = nSrcYSize-1 - iSrcY;

    double dfAccumulator = 0.0;

    for( ; j <= jMax; ++j )
    {
        const int iSampJ = iSrcOffset + j * nSrcXSize;

        // Loop over all pixels in the row.
        double dfAccumulatorLocal = 0.0;
        double dfAccumulatorLocal2 = 0.0;
        iC = 0;
        i = iMin;
        // Process by chunk of 4 cols.
        for( ; i+2 < iMax; i += 4, iC += 4 )
        {
            // Retrieve the pixel & accumulate.
            dfAccumulatorLocal += pSrcBand[i+iSampJ] * padfWeight[iC];
            dfAccumulatorLocal += pSrcBand[i+1+iSampJ] * padfWeight[iC+1];
            dfAccumulatorLocal2 += pSrcBand[i+2+iSampJ] * padfWeight[iC+2];
            dfAccumulatorLocal2 += pSrcBand[i+3+iSampJ] * padfWeight[iC+3];
        }
        dfAccumulatorLocal += dfAccumulatorLocal2;
        if( i < iMax )
        {
            dfAccumulatorLocal += pSrcBand[i+iSampJ] * padfWeight[iC];
            dfAccumulatorLocal += pSrcBand[i+1+iSampJ] * padfWeight[iC+1];
            i += 2;
            iC += 2;
        }
        if( i == iMax )
        {
            dfAccumulatorLocal += pSrcBand[i+iSampJ] * padfWeight[iC];
        }

        // Calculate the Y weight.
        const double dfWeight = pfnGetWeight((j - dfDeltaY) * dfYScale);
        dfAccumulator += dfWeight * dfAccumulatorLocal;
        dfAccumulatorWeightVertical += dfWeight;
    }

    const double dfAccumulatorWeight =
        dfAccumulatorWeightHorizontal * dfAccumulatorWeightVertical;

    *pValue = GWKClampValueT<T>(dfAccumulator / dfAccumulatorWeight);

    return true;
}

/* We restrict to 64bit processors because they are guaranteed to have SSE2 */
/* Could possibly be used too on 32bit, but we would need to check at runtime */
#if defined(__x86_64) || defined(_M_X64)

/************************************************************************/
/*                    GWKResampleNoMasks_SSE2_T()                       */
/************************************************************************/

template<class T>
static bool GWKResampleNoMasks_SSE2_T( GDALWarpKernel *poWK, int iBand,
                                      double dfSrcX, double dfSrcY,
                                      T *pValue, double *padfWeight )
{
    // Commonly used; save locally.
    const int nSrcXSize = poWK->nSrcXSize;
    const int nSrcYSize = poWK->nSrcYSize;

    const int iSrcX = static_cast<int>(floor(dfSrcX - 0.5));
    const int iSrcY = static_cast<int>(floor(dfSrcY - 0.5));
    const int iSrcOffset = iSrcX + iSrcY * nSrcXSize;
    const int nXRadius = poWK->nXRadius;
    const int nYRadius = poWK->nYRadius;

    // Politely refuse to process invalid coordinates or obscenely small image.
    if( iSrcX >= nSrcXSize || iSrcY >= nSrcYSize ||
        nXRadius > nSrcXSize || nYRadius > nSrcYSize )
        return GWKBilinearResampleNoMasks4SampleT(poWK, iBand, dfSrcX, dfSrcY,
                                                  pValue);

    const T* pSrcBand = reinterpret_cast<const T*>(poWK->papabySrcImage[iBand]);

    const FilterFuncType pfnGetWeight = apfGWKFilter[poWK->eResample];
    CPLAssert(pfnGetWeight);
    const FilterFunc4ValuesType pfnGetWeight4Values =
        apfGWKFilter4Values[poWK->eResample];
    CPLAssert(pfnGetWeight4Values);

    const double dfDeltaX = dfSrcX - 0.5 - iSrcX;
    const double dfDeltaY = dfSrcY - 0.5 - iSrcY;
    const double dfXScale = std::min(poWK->dfXScale, 1.0);
    const double dfYScale = std::min(poWK->dfYScale, 1.0);

    // Loop over all rows in the kernel.
    double dfAccumulatorWeightHorizontal = 0.0;
    double dfAccumulatorWeightVertical = 0.0;
    double dfAccumulator = 0.0;

    int iMin = 1 - nXRadius;
    if( iSrcX + iMin < 0 )
        iMin = -iSrcX;
    int iMax = nXRadius;
    if( iSrcX + iMax >= nSrcXSize-1 )
        iMax = nSrcXSize-1 - iSrcX;
    int i, iC;
    for( iC = 0, i = iMin; i+2 < iMax; i+=4, iC+=4 )
    {
        padfWeight[iC] = (i - dfDeltaX) * dfXScale;
        padfWeight[iC+1] = padfWeight[iC] + dfXScale;
        padfWeight[iC+2] = padfWeight[iC+1] + dfXScale;
        padfWeight[iC+3] = padfWeight[iC+2] + dfXScale;
        dfAccumulatorWeightHorizontal += pfnGetWeight4Values(padfWeight+iC);
    }
    for( ; i <= iMax; ++i, ++iC )
    {
        double dfWeight = pfnGetWeight((i - dfDeltaX) * dfXScale);
        padfWeight[iC] = dfWeight;
        dfAccumulatorWeightHorizontal += dfWeight;
    }

    int j = 1 - nYRadius;
    if(  iSrcY + j < 0 )
        j = -iSrcY;
    int jMax = nYRadius;
    if( iSrcY + jMax >= nSrcYSize-1 )
        jMax = nSrcYSize-1 - iSrcY;

    // Process by chunk of 4 rows.
    for( ; j+2 < jMax; j+=4 )
    {
        int     iSampJ = iSrcOffset + j * nSrcXSize;

        // Loop over all pixels in the row.
        iC = 0;
        i = iMin;
        // Process by chunk of 4 cols.
        XMMReg4Double v_acc_1 = XMMReg4Double::Zero();
        XMMReg4Double v_acc_2 = XMMReg4Double::Zero();
        XMMReg4Double v_acc_3 = XMMReg4Double::Zero();
        XMMReg4Double v_acc_4 = XMMReg4Double::Zero();
        for( ; i+2 < iMax; i+=4, iC+=4 )
        {
            // Retrieve the pixel & accumulate.
            XMMReg4Double v_pixels_1 =
                XMMReg4Double::Load4Val(pSrcBand+i+iSampJ);
            XMMReg4Double v_pixels_2 =
                XMMReg4Double::Load4Val(pSrcBand+i+iSampJ+nSrcXSize);
            XMMReg4Double v_pixels_3 =
                XMMReg4Double::Load4Val(pSrcBand+i+iSampJ+2*nSrcXSize);
            XMMReg4Double v_pixels_4 =
                XMMReg4Double::Load4Val(pSrcBand+i+iSampJ+3*nSrcXSize);

            XMMReg4Double v_padfWeight =
                XMMReg4Double::Load4Val(padfWeight + iC);

            v_acc_1 += v_pixels_1 * v_padfWeight;
            v_acc_2 += v_pixels_2 * v_padfWeight;
            v_acc_3 += v_pixels_3 * v_padfWeight;
            v_acc_4 += v_pixels_4 * v_padfWeight;
        }

        if( i < iMax )
        {
            XMMReg2Double v_pixels_1 =
                XMMReg2Double::Load2Val(pSrcBand+i+iSampJ);
            XMMReg2Double v_pixels_2 =
                XMMReg2Double::Load2Val(pSrcBand+i+iSampJ+nSrcXSize);
            XMMReg2Double v_pixels_3 =
                XMMReg2Double::Load2Val(pSrcBand+i+iSampJ+2*nSrcXSize);
            XMMReg2Double v_pixels_4 =
                XMMReg2Double::Load2Val(pSrcBand+i+iSampJ+3*nSrcXSize);

            XMMReg2Double v_padfWeight =
                XMMReg2Double::Load2Val(padfWeight + iC);

            v_acc_1.AddToLow( v_pixels_1 * v_padfWeight );
            v_acc_2.AddToLow( v_pixels_2 * v_padfWeight );
            v_acc_3.AddToLow( v_pixels_3 * v_padfWeight );
            v_acc_4.AddToLow( v_pixels_4 * v_padfWeight );

            i+=2;
            iC+=2;
        }

        double dfAccumulatorLocal_1 = v_acc_1.GetHorizSum();
        double dfAccumulatorLocal_2 = v_acc_2.GetHorizSum();
        double dfAccumulatorLocal_3 = v_acc_3.GetHorizSum();
        double dfAccumulatorLocal_4 = v_acc_4.GetHorizSum();

        if( i == iMax )
        {
            dfAccumulatorLocal_1 +=
                static_cast<double>(pSrcBand[i+iSampJ]) * padfWeight[iC];
            dfAccumulatorLocal_2 +=
                static_cast<double>(pSrcBand[i+iSampJ + nSrcXSize]) *
                padfWeight[iC];
            dfAccumulatorLocal_3 +=
                static_cast<double>(pSrcBand[i+iSampJ + 2 * nSrcXSize]) *
                padfWeight[iC];
            dfAccumulatorLocal_4 +=
                static_cast<double>(pSrcBand[i+iSampJ + 3 * nSrcXSize]) *
                padfWeight[iC];
        }

        // Calculate the Y weight.
        const double dfWeight0 = (j - dfDeltaY) * dfYScale;
        const double dfWeight1 = dfWeight0 + dfYScale;
        const double dfWeight2 = dfWeight1 + dfYScale;
        const double dfWeight3 = dfWeight2 + dfYScale;
        double adfWeight[4] = { dfWeight0, dfWeight1, dfWeight2, dfWeight3 };

        dfAccumulatorWeightVertical += pfnGetWeight4Values(adfWeight);
        dfAccumulator += adfWeight[0] * dfAccumulatorLocal_1;
        dfAccumulator += adfWeight[1] * dfAccumulatorLocal_2;
        dfAccumulator += adfWeight[2] * dfAccumulatorLocal_3;
        dfAccumulator += adfWeight[3] * dfAccumulatorLocal_4;
    }
    for( ; j <= jMax; ++j )
    {
        int iSampJ = iSrcOffset + j * nSrcXSize;

        // Loop over all pixels in the row.
        iC = 0;
        i = iMin;
        // Process by chunk of 4 cols.
        XMMReg4Double v_acc = XMMReg4Double::Zero();
        for( ; i+2 < iMax; i+=4, iC+=4 )
        {
            // Retrieve the pixel & accumulate.
            XMMReg4Double v_pixels= XMMReg4Double::Load4Val(pSrcBand+i+iSampJ);
            XMMReg4Double v_padfWeight =
                XMMReg4Double::Load4Val(padfWeight + iC);

            v_acc += v_pixels * v_padfWeight;
        }

        double dfAccumulatorLocal = v_acc.GetHorizSum();

        if( i < iMax )
        {
            dfAccumulatorLocal += pSrcBand[i+iSampJ] * padfWeight[iC];
            dfAccumulatorLocal += pSrcBand[i+1+iSampJ] * padfWeight[iC+1];
            i+=2;
            iC+=2;
        }
        if( i == iMax )
        {
            dfAccumulatorLocal += static_cast<double>(pSrcBand[i+iSampJ]) * padfWeight[iC];
        }

        // Calculate the Y weight.
        double dfWeight = pfnGetWeight((j - dfDeltaY) * dfYScale);
        dfAccumulator += dfWeight * dfAccumulatorLocal;
        dfAccumulatorWeightVertical += dfWeight;
    }

    const double dfAccumulatorWeight =
        dfAccumulatorWeightHorizontal * dfAccumulatorWeightVertical;

    *pValue = GWKClampValueT<T>(dfAccumulator / dfAccumulatorWeight);

    return true;
}

/************************************************************************/
/*                     GWKResampleNoMasksT<GByte>()                     */
/************************************************************************/

template<>
bool GWKResampleNoMasksT<GByte>( GDALWarpKernel *poWK, int iBand,
                                 double dfSrcX, double dfSrcY,
                                 GByte *pValue, double *padfWeight )
{
    return GWKResampleNoMasks_SSE2_T(poWK, iBand, dfSrcX, dfSrcY,
                                     pValue, padfWeight);
}

/************************************************************************/
/*                     GWKResampleNoMasksT<GInt16>()                    */
/************************************************************************/

template<>
bool GWKResampleNoMasksT<GInt16>( GDALWarpKernel *poWK, int iBand,
                                  double dfSrcX, double dfSrcY,
                                  GInt16 *pValue, double *padfWeight )
{
    return GWKResampleNoMasks_SSE2_T(poWK, iBand, dfSrcX, dfSrcY,
                                     pValue, padfWeight);
}

/************************************************************************/
/*                     GWKResampleNoMasksT<GUInt16>()                   */
/************************************************************************/

template<>
bool GWKResampleNoMasksT<GUInt16>( GDALWarpKernel *poWK, int iBand,
                                   double dfSrcX, double dfSrcY,
                                   GUInt16 *pValue, double *padfWeight )
{
    return GWKResampleNoMasks_SSE2_T(poWK, iBand, dfSrcX, dfSrcY,
                                     pValue, padfWeight);
}

/************************************************************************/
/*                     GWKResampleNoMasksT<float>()                     */
/************************************************************************/

template<>
bool GWKResampleNoMasksT<float>( GDALWarpKernel *poWK, int iBand,
                                 double dfSrcX, double dfSrcY,
                                 float *pValue, double *padfWeight )
{
    return GWKResampleNoMasks_SSE2_T(poWK, iBand, dfSrcX, dfSrcY,
                                     pValue, padfWeight);
}

#ifdef INSTANTIATE_FLOAT64_SSE2_IMPL

/************************************************************************/
/*                     GWKResampleNoMasksT<double>()                    */
/************************************************************************/

template<>
bool GWKResampleNoMasksT<double>( GDALWarpKernel *poWK, int iBand,
                                  double dfSrcX, double dfSrcY,
                                  double *pValue, double *padfWeight )
{
    return GWKResampleNoMasks_SSE2_T(poWK, iBand, dfSrcX, dfSrcY,
                                     pValue, padfWeight);
}

#endif /* INSTANTIATE_FLOAT64_SSE2_IMPL */

#endif /* defined(__x86_64) || defined(_M_X64) */

/************************************************************************/
/*                     GWKRoundSourceCoordinates()                      */
/************************************************************************/

static void GWKRoundSourceCoordinates( int nDstXSize,
                                       double* padfX,
                                       double* padfY,
                                       double* padfZ,
                                       int* pabSuccess,
                                       double dfSrcCoordPrecision,
                                       double dfErrorThreshold,
                                       GDALTransformerFunc pfnTransformer,
                                       void* pTransformerArg,
                                       double dfDstXOff,
                                       double dfDstY )
{
    double dfPct = 0.8;
    if( dfErrorThreshold > 0 && dfSrcCoordPrecision / dfErrorThreshold >= 10.0 )
    {
        dfPct = 1.0 - 2 * 1.0 / (dfSrcCoordPrecision / dfErrorThreshold);
    }
    const double dfExactTransformThreshold = 0.5 * dfPct * dfSrcCoordPrecision;

    for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
    {
        const double dfXBefore = padfX[iDstX];
        const double dfYBefore = padfY[iDstX];
        padfX[iDstX] =
            floor(padfX[iDstX] / dfSrcCoordPrecision + 0.5) *
            dfSrcCoordPrecision;
        padfY[iDstX] =
            floor(padfY[iDstX] / dfSrcCoordPrecision + 0.5) *
            dfSrcCoordPrecision;

        // If we are in an uncertainty zone, go to non-approximated
        // transformation.
        // Due to the 80% of half-precision threshold, dfSrcCoordPrecision must
        // be at least 10 times greater than the approximation error.
        if( fabs(dfXBefore - padfX[iDstX]) > dfExactTransformThreshold ||
            fabs(dfYBefore - padfY[iDstX]) > dfExactTransformThreshold )
        {
            padfX[iDstX] = iDstX + dfDstXOff;
            padfY[iDstX] = dfDstY;
            padfZ[iDstX] = 0.0;
            pfnTransformer( pTransformerArg, TRUE, 1,
                            padfX + iDstX, padfY + iDstX,
                            padfZ + iDstX, pabSuccess + iDstX );
            padfX[iDstX] =
                floor(padfX[iDstX] / dfSrcCoordPrecision + 0.5) *
                dfSrcCoordPrecision;
            padfY[iDstX] =
                floor(padfY[iDstX] / dfSrcCoordPrecision + 0.5) *
                dfSrcCoordPrecision;
        }
    }
}

/************************************************************************/
/*                           GWKOpenCLCase()                            */
/*                                                                      */
/*      This is identical to GWKGeneralCase(), but functions via        */
/*      OpenCL. This means we have vector optimization (SSE) and/or     */
/*      GPU optimization depending on our prefs. The code itself is     */
/*      general and not optimized, but by defining constants we can     */
/*      make some pretty darn good code on the fly.                     */
/************************************************************************/

#if defined(HAVE_OPENCL)
static CPLErr GWKOpenCLCase( GDALWarpKernel *poWK )
{
    const int nDstXSize = poWK->nDstXSize;
    const int nDstYSize = poWK->nDstYSize;
    const int nSrcXSize = poWK->nSrcXSize;
    const int nSrcYSize = poWK->nSrcYSize;
    const int nDstXOff = poWK->nDstXOff;
    const int nDstYOff = poWK->nDstYOff;
    const int nSrcXOff = poWK->nSrcXOff;
    const int nSrcYOff = poWK->nSrcYOff;
    bool bUseImag = false;

    cl_channel_type imageFormat;
    switch( poWK->eWorkingDataType )
    {
      case GDT_Byte:
        imageFormat = CL_UNORM_INT8;
        break;
      case GDT_UInt16:
        imageFormat = CL_UNORM_INT16;
        break;
      case GDT_CInt16:
        bUseImag = true;
        CPL_FALLTHROUGH
      case GDT_Int16:
        imageFormat = CL_SNORM_INT16;
        break;
      case GDT_CFloat32:
        bUseImag = true;
        CPL_FALLTHROUGH
      case GDT_Float32:
        imageFormat = CL_FLOAT;
        break;
      default:
        // No support for higher precision formats.
        CPLDebug("OpenCL",
                 "Unsupported resampling OpenCL data type %d.",
                 static_cast<int>(poWK->eWorkingDataType));
        return CE_Warning;
    }

    OCLResampAlg resampAlg;
    switch( poWK->eResample )
    {
      case GRA_Bilinear:
        resampAlg = OCL_Bilinear;
        break;
      case GRA_Cubic:
        resampAlg = OCL_Cubic;
        break;
      case GRA_CubicSpline:
        resampAlg = OCL_CubicSpline;
        break;
      case GRA_Lanczos:
        resampAlg = OCL_Lanczos;
        break;
      default:
        // No support for higher precision formats.
        CPLDebug("OpenCL",
                 "Unsupported resampling OpenCL resampling alg %d.",
                 static_cast<int>(poWK->eResample));
        return CE_Warning;
    }

    struct oclWarper *warper = nullptr;
    cl_int err;
    CPLErr eErr = CE_None;

    // TODO(schwehr): Fix indenting.
  try
  {

    // Using a factor of 2 or 4 seems to have much less rounding error
    // than 3 on the GPU.
    // Then the rounding error can cause strange artifacts under the
    // right conditions.
    warper =
        GDALWarpKernelOpenCL_createEnv(nSrcXSize, nSrcYSize,
                                       nDstXSize, nDstYSize,
                                       imageFormat, poWK->nBands, 4,
                                       bUseImag,
                                       poWK->papanBandSrcValid != nullptr,
                                       poWK->pafDstDensity,
                                       poWK->padfDstNoDataReal,
                                       resampAlg, &err);

    if( err != CL_SUCCESS || warper == nullptr )
    {
        eErr = CE_Warning;
        if( warper != nullptr )
            throw eErr;
        return eErr;
    }

    CPLDebug("GDAL", "GDALWarpKernel()::GWKOpenCLCase() "
             "Src=%d,%d,%dx%d Dst=%d,%d,%dx%d",
             nSrcXOff, nSrcYOff, nSrcXSize, nSrcYSize,
             nDstXOff, nDstYOff, nDstXSize, nDstYSize);

    if( !poWK->pfnProgress( poWK->dfProgressBase, "", poWK->pProgress ) )
    {
        CPLError( CE_Failure, CPLE_UserInterrupt, "User terminated" );
        eErr = CE_Failure;
        throw eErr;
    }

    /* ==================================================================== */
    /*      Loop over bands.                                                */
    /* ==================================================================== */
    for( int iBand = 0; iBand < poWK->nBands; iBand++ )
    {
        if( poWK->papanBandSrcValid != nullptr &&
            poWK->papanBandSrcValid[iBand] != nullptr)
        {
            GDALWarpKernelOpenCL_setSrcValid(
                warper, reinterpret_cast<int *>(poWK->papanBandSrcValid[iBand]), iBand);
            if( err != CL_SUCCESS )
            {
                CPLError( CE_Failure, CPLE_AppDefined,
                          "OpenCL routines reported failure (%d) on line %d.",
                          static_cast<int>(err), __LINE__ );
                eErr = CE_Failure;
                throw eErr;
            }
        }

        err = GDALWarpKernelOpenCL_setSrcImg(warper,
                                             poWK->papabySrcImage[iBand],
                                             iBand);
        if( err != CL_SUCCESS )
        {
            CPLError(CE_Failure, CPLE_AppDefined,
                     "OpenCL routines reported failure (%d) on line %d.",
                     static_cast<int>(err), __LINE__);
            eErr = CE_Failure;
            throw eErr;
        }

        err = GDALWarpKernelOpenCL_setDstImg(warper,
                                             poWK->papabyDstImage[iBand],
                                             iBand);
        if( err != CL_SUCCESS )
        {
            CPLError(CE_Failure, CPLE_AppDefined,
                     "OpenCL routines reported failure (%d) on line %d.",
                     static_cast<int>(err), __LINE__);
            eErr = CE_Failure;
            throw eErr;
        }
    }

    /* -------------------------------------------------------------------- */
    /*      Allocate x,y,z coordinate arrays for transformation ... one     */
    /*      scanlines worth of positions.                                   */
    /* -------------------------------------------------------------------- */

    // For x, 2 *, because we cache the precomputed values at the end.
    double *padfX =
        static_cast<double *>(CPLMalloc(2 * sizeof(double) * nDstXSize));
    double * padfY =
        static_cast<double *>(CPLMalloc(sizeof(double) * nDstXSize));
    double *padfZ =
        static_cast<double *>(CPLMalloc(sizeof(double) * nDstXSize));
    int *pabSuccess = static_cast<int *>(CPLMalloc(sizeof(int) * nDstXSize));
    const double dfSrcCoordPrecision = CPLAtof(
        CSLFetchNameValueDef(poWK->papszWarpOptions,
                             "SRC_COORD_PRECISION", "0"));
    const double dfErrorThreshold = CPLAtof(
        CSLFetchNameValueDef(poWK->papszWarpOptions, "ERROR_THRESHOLD", "0"));

    // Precompute values.
    for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
        padfX[nDstXSize + iDstX] = iDstX + 0.5 + poWK->nDstXOff;

    /* ==================================================================== */
    /*      Loop over output lines.                                         */
    /* ==================================================================== */
    for( int iDstY = 0; iDstY < nDstYSize && eErr == CE_None; ++iDstY )
    {
        /* ---------------------------------------------------------------- */
        /*      Setup points to transform to source image space.            */
        /* ---------------------------------------------------------------- */
        memcpy( padfX, padfX + nDstXSize, sizeof(double) * nDstXSize );
        const double dfYConst = iDstY + 0.5 + poWK->nDstYOff;
        for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
            padfY[iDstX] = dfYConst;
        memset( padfZ, 0, sizeof(double) * nDstXSize );

        /* ---------------------------------------------------------------- */
        /*      Transform the points from destination pixel/line coordinates*/
        /*      to source pixel/line coordinates.                           */
        /* ---------------------------------------------------------------- */
        poWK->pfnTransformer( poWK->pTransformerArg, TRUE, nDstXSize,
                              padfX, padfY, padfZ, pabSuccess );
        if( dfSrcCoordPrecision > 0.0 )
        {
            GWKRoundSourceCoordinates(nDstXSize, padfX, padfY, padfZ,
                                      pabSuccess,
                                      dfSrcCoordPrecision,
                                      dfErrorThreshold,
                                      poWK->pfnTransformer,
                                      poWK->pTransformerArg,
                                      0.5 + nDstXOff,
                                      iDstY + 0.5 + nDstYOff);
        }

        err = GDALWarpKernelOpenCL_setCoordRow(warper, padfX, padfY,
                                               nSrcXOff, nSrcYOff,
                                               pabSuccess, iDstY);
        if( err != CL_SUCCESS )
        {
            CPLError(CE_Failure, CPLE_AppDefined,
                     "OpenCL routines reported failure (%d) on line %d.",
                     static_cast<int>(err), __LINE__);
            eErr = CE_Failure;
            break;
        }

        // Update the valid & density masks because we don't do so in the
        // kernel.
        for( int iDstX = 0; iDstX < nDstXSize && eErr == CE_None; iDstX++ )
        {
            const double dfX = padfX[iDstX];
            const double dfY = padfY[iDstX];
            const int iDstOffset = iDstX + iDstY * nDstXSize;

            // See GWKGeneralCase() for appropriate commenting.
            if( !pabSuccess[iDstX] || dfX < nSrcXOff || dfY < nSrcYOff )
                continue;

            int iSrcX = static_cast<int>(dfX) - nSrcXOff;
            int iSrcY = static_cast<int>(dfY) - nSrcYOff;

            if( iSrcX < 0 || iSrcX >= nSrcXSize ||
                iSrcY < 0 || iSrcY >= nSrcYSize )
                continue;

            int iSrcOffset = iSrcX + iSrcY * nSrcXSize;
            double dfDensity = 1.0;

            if( poWK->pafUnifiedSrcDensity != nullptr
                && iSrcX >= 0 && iSrcY >= 0
                && iSrcX < nSrcXSize && iSrcY < nSrcYSize )
                dfDensity = poWK->pafUnifiedSrcDensity[iSrcOffset];

            GWKOverlayDensity( poWK, iDstOffset, dfDensity );

            // Because this is on the bit-wise level, it can't be done well in
            // OpenCL.
            if( poWK->panDstValid != nullptr )
                poWK->panDstValid[iDstOffset>>5] |= 0x01 << (iDstOffset & 0x1f);
        }
    }

    CPLFree( padfX );
    CPLFree( padfY );
    CPLFree( padfZ );
    CPLFree( pabSuccess );

    if( eErr != CE_None )
        throw eErr;

    err = GDALWarpKernelOpenCL_runResamp(warper,
                                         poWK->pafUnifiedSrcDensity,
                                         poWK->panUnifiedSrcValid,
                                         poWK->pafDstDensity,
                                         poWK->panDstValid,
                                         poWK->dfXScale, poWK->dfYScale,
                                         poWK->dfXFilter, poWK->dfYFilter,
                                         poWK->nXRadius, poWK->nYRadius,
                                         poWK->nFiltInitX, poWK->nFiltInitY);

    if( err != CL_SUCCESS )
    {
        CPLError(CE_Failure, CPLE_AppDefined,
                 "OpenCL routines reported failure (%d) on line %d.",
                 static_cast<int>(err), __LINE__);
        eErr = CE_Failure;
        throw eErr;
    }

    /* ==================================================================== */
    /*      Loop over output lines.                                         */
    /* ==================================================================== */
    for( int iDstY = 0; iDstY < nDstYSize && eErr == CE_None; iDstY++ )
    {
        for( int iBand = 0; iBand < poWK->nBands; iBand++ )
        {
            void *rowReal = nullptr;
            void *rowImag = nullptr;
            GByte *pabyDst = poWK->papabyDstImage[iBand];

            err = GDALWarpKernelOpenCL_getRow(warper, &rowReal, &rowImag,
                                              iDstY, iBand);
            if( err != CL_SUCCESS )
            {
                CPLError( CE_Failure, CPLE_AppDefined,
                          "OpenCL routines reported failure (%d) on line %d.",
                          static_cast<int>(err), __LINE__ );
                eErr = CE_Failure;
                throw eErr;
            }

            // Copy the data from the warper to GDAL's memory.
            switch( poWK->eWorkingDataType )
            {
              case GDT_Byte:
                memcpy(&(pabyDst[iDstY*nDstXSize]),
                       rowReal, sizeof(GByte)*nDstXSize);
                break;
              case GDT_Int16:
                memcpy(&(reinterpret_cast<GInt16 *>(pabyDst)[iDstY*nDstXSize]),
                       rowReal, sizeof(GInt16)*nDstXSize);
                break;
              case GDT_UInt16:
                memcpy(&(reinterpret_cast<GUInt16 *>(pabyDst)[iDstY*nDstXSize]),
                       rowReal, sizeof(GUInt16)*nDstXSize);
                break;
              case GDT_Float32:
                memcpy(&(reinterpret_cast<float *>(pabyDst)[iDstY*nDstXSize]),
                       rowReal, sizeof(float)*nDstXSize);
                break;
              case GDT_CInt16:
              {
                  GInt16 *pabyDstI16 = &(reinterpret_cast<GInt16 *>(pabyDst)[iDstY*nDstXSize]);
                  for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
                  {
                      pabyDstI16[iDstX*2  ] = static_cast<GInt16 *>(rowReal)[iDstX];
                      pabyDstI16[iDstX*2+1] = static_cast<GInt16 *>(rowImag)[iDstX];
                  }
              }
              break;
              case GDT_CFloat32:
              {
                  float *pabyDstF32 = &(reinterpret_cast<float *>(pabyDst)[iDstY*nDstXSize]);
                  for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
                  {
                      pabyDstF32[iDstX*2  ] = static_cast<float *>(rowReal)[iDstX];
                      pabyDstF32[iDstX*2+1] = static_cast<float *>(rowImag)[iDstX];
                  }
              }
              break;
              default:
                // No support for higher precision formats.
                CPLError(CE_Failure, CPLE_AppDefined,
                         "Unsupported resampling OpenCL data type %d.",
                         static_cast<int>(poWK->eWorkingDataType) );
                eErr = CE_Failure;
                throw eErr;
            }
        }
    }
  }
  catch( const CPLErr& )
  {
  }

    if( (err = GDALWarpKernelOpenCL_deleteEnv(warper)) != CL_SUCCESS )
    {
        CPLError(CE_Failure, CPLE_AppDefined,
                 "OpenCL routines reported failure (%d) on line %d.",
                 static_cast<int>(err), __LINE__ );
        return CE_Failure;
    }

    return eErr;
}
#endif /* defined(HAVE_OPENCL) */

/************************************************************************/
/*                     GWKCheckAndComputeSrcOffsets()                   */
/************************************************************************/
static CPL_INLINE bool GWKCheckAndComputeSrcOffsets(
    const int* _pabSuccess,
    int _iDstX,
    const double* _padfX,
    const double* _padfY,
    const GDALWarpKernel* _poWK,
    int _nSrcXSize,
    int _nSrcYSize,
    int& iSrcOffset)
{
    if( !_pabSuccess[_iDstX] )
        return false;

    // If this happens this is likely the symptom of a bug somewhere.
    if( CPLIsNan(_padfX[_iDstX]) || CPLIsNan(_padfY[_iDstX]) )
    {
        static bool bNanCoordFound = false;
        if( !bNanCoordFound )
        {
            CPLDebug("WARP", "NaN coordinate found.");
            bNanCoordFound = true;
        }
        return false;
    }

/* -------------------------------------------------------------------- */
/*      Figure out what pixel we want in our source raster, and skip    */
/*      further processing if it is well off the source image.          */
/* -------------------------------------------------------------------- */
    /* We test against the value before casting to avoid the */
    /* problem of asymmetric truncation effects around zero.  That is */
    /* -0.5 will be 0 when cast to an int. */
    if( _padfX[_iDstX] < _poWK->nSrcXOff
        || _padfY[_iDstX] < _poWK->nSrcYOff )
        return false;

    // Check for potential overflow when casting from float to int, (if
    // operating outside natural projection area, padfX/Y can be a very huge
    // positive number before doing the actual conversion), as such cast is
    // undefined behaviour that can trigger exception with some compilers
    // (see #6753)
    if( _padfX[_iDstX] + 1e-10 > _nSrcXSize + _poWK->nSrcXOff ||
        _padfY[_iDstX] + 1e-10 > _nSrcYSize + _poWK->nSrcYOff )
    {
        return false;
    }

    const int iSrcX =
        static_cast<int>(_padfX[_iDstX] + 1.0e-10) - _poWK->nSrcXOff;
    const int iSrcY =
        static_cast<int>(_padfY[_iDstX] + 1.0e-10) - _poWK->nSrcYOff;

    // Those checks should normally be OK given the previous ones.
    CPLAssert( iSrcX >= 0 );
    CPLAssert( iSrcY >= 0 );
    CPLAssert( iSrcX < _nSrcXSize );
    CPLAssert( iSrcY < _nSrcYSize );

    iSrcOffset = iSrcX + iSrcY * _nSrcXSize;

    return true;
}

/************************************************************************/
/*                           GWKGeneralCase()                           */
/*                                                                      */
/*      This is the most general case.  It attempts to handle all       */
/*      possible features with relatively little concern for            */
/*      efficiency.                                                     */
/************************************************************************/

static void GWKGeneralCaseThread( void* pData)

{
    GWKJobStruct* psJob = reinterpret_cast<GWKJobStruct *>(pData);
    GDALWarpKernel *poWK = psJob->poWK;
    const int iYMin = psJob->iYMin;
    const int iYMax = psJob->iYMax;

    int nDstXSize = poWK->nDstXSize;
    int nSrcXSize = poWK->nSrcXSize;
    int nSrcYSize = poWK->nSrcYSize;

/* -------------------------------------------------------------------- */
/*      Allocate x,y,z coordinate arrays for transformation ... one     */
/*      scanlines worth of positions.                                   */
/* -------------------------------------------------------------------- */
    // For x, 2 *, because we cache the precomputed values at the end.
    double *padfX =
        static_cast<double *>(CPLMalloc(2 * sizeof(double) * nDstXSize));
    double *padfY =
        static_cast<double *>(CPLMalloc(sizeof(double) * nDstXSize));
    double *padfZ =
        static_cast<double *>(CPLMalloc(sizeof(double) * nDstXSize));
    int *pabSuccess = static_cast<int *>(CPLMalloc(sizeof(int) * nDstXSize));

    const bool bUse4SamplesFormula =
        poWK->dfXScale >= 0.95 && poWK->dfYScale >= 0.95;

    GWKResampleWrkStruct* psWrkStruct = nullptr;
    if( poWK->eResample != GRA_NearestNeighbour )
    {
        psWrkStruct = GWKResampleCreateWrkStruct(poWK);
    }
    const double dfSrcCoordPrecision = CPLAtof(
        CSLFetchNameValueDef(poWK->papszWarpOptions,
                             "SRC_COORD_PRECISION", "0"));
    const double dfErrorThreshold = CPLAtof(
        CSLFetchNameValueDef(poWK->papszWarpOptions, "ERROR_THRESHOLD", "0"));

    // Precompute values.
    for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
        padfX[nDstXSize + iDstX] = iDstX + 0.5 + poWK->nDstXOff;

/* ==================================================================== */
/*      Loop over output lines.                                         */
/* ==================================================================== */
    for( int iDstY = iYMin; iDstY < iYMax; iDstY++ )
    {
/* -------------------------------------------------------------------- */
/*      Setup points to transform to source image space.                */
/* -------------------------------------------------------------------- */
        memcpy( padfX, padfX + nDstXSize, sizeof(double) * nDstXSize );
        const double dfY = iDstY + 0.5 + poWK->nDstYOff;
        for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
            padfY[iDstX] = dfY;
        memset( padfZ, 0, sizeof(double) * nDstXSize );

/* -------------------------------------------------------------------- */
/*      Transform the points from destination pixel/line coordinates    */
/*      to source pixel/line coordinates.                               */
/* -------------------------------------------------------------------- */
        poWK->pfnTransformer( psJob->pTransformerArg, TRUE, nDstXSize,
                              padfX, padfY, padfZ, pabSuccess );
        if( dfSrcCoordPrecision > 0.0 )
        {
            GWKRoundSourceCoordinates(nDstXSize, padfX, padfY, padfZ,
                                      pabSuccess,
                                      dfSrcCoordPrecision,
                                      dfErrorThreshold,
                                      poWK->pfnTransformer,
                                      psJob->pTransformerArg,
                                      0.5 + poWK->nDstXOff,
                                      iDstY + 0.5 + poWK->nDstYOff);
        }

/* ==================================================================== */
/*      Loop over pixels in output scanline.                            */
/* ==================================================================== */
        for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
        {
            int iSrcOffset = 0;
            if( !GWKCheckAndComputeSrcOffsets(pabSuccess, iDstX, padfX, padfY,
                                    poWK, nSrcXSize, nSrcYSize, iSrcOffset) )
                continue;

/* -------------------------------------------------------------------- */
/*      Do not try to apply transparent/invalid source pixels to the    */
/*      destination.  This currently ignores the multi-pixel input      */
/*      of bilinear and cubic resamples.                                */
/* -------------------------------------------------------------------- */
            double dfDensity = 1.0;

            if( poWK->pafUnifiedSrcDensity != nullptr )
            {
                dfDensity = poWK->pafUnifiedSrcDensity[iSrcOffset];
                if( dfDensity < SRC_DENSITY_THRESHOLD )
                    continue;
            }

            if( poWK->panUnifiedSrcValid != nullptr
                && !(poWK->panUnifiedSrcValid[iSrcOffset>>5]
                     & (0x01 << (iSrcOffset & 0x1f))) )
                continue;

/* ==================================================================== */
/*      Loop processing each band.                                      */
/* ==================================================================== */
            bool bHasFoundDensity = false;

            const int iDstOffset = iDstX + iDstY * nDstXSize;
            for( int iBand = 0; iBand < poWK->nBands; iBand++ )
            {
                double dfBandDensity = 0.0;
                double dfValueReal = 0.0;
                double dfValueImag = 0.0;

/* -------------------------------------------------------------------- */
/*      Collect the source value.                                       */
/* -------------------------------------------------------------------- */
                if( poWK->eResample == GRA_NearestNeighbour ||
                    nSrcXSize == 1 || nSrcYSize == 1 )
                {
                    // FALSE is returned if dfBandDensity == 0, which is
                    // checked below.
                    CPL_IGNORE_RET_VAL(
                        GWKGetPixelValue( poWK, iBand, iSrcOffset,
                                          &dfBandDensity,
                                          &dfValueReal, &dfValueImag ));
                }
                else if( poWK->eResample == GRA_Bilinear &&
                         bUse4SamplesFormula )
                {
                    GWKBilinearResample4Sample( poWK, iBand,
                                         padfX[iDstX]-poWK->nSrcXOff,
                                         padfY[iDstX]-poWK->nSrcYOff,
                                         &dfBandDensity,
                                         &dfValueReal, &dfValueImag );
                }
                else if( poWK->eResample == GRA_Cubic &&
                         bUse4SamplesFormula )
                {
                    GWKCubicResample4Sample( poWK, iBand,
                                        padfX[iDstX]-poWK->nSrcXOff,
                                        padfY[iDstX]-poWK->nSrcYOff,
                                        &dfBandDensity,
                                        &dfValueReal, &dfValueImag );
                }
                else
#ifdef DEBUG
                // Only useful for clang static analyzer.
                if( psWrkStruct != nullptr )
#endif
                {
                    psWrkStruct->pfnGWKResample( poWK, iBand,
                                 padfX[iDstX]-poWK->nSrcXOff,
                                 padfY[iDstX]-poWK->nSrcYOff,
                                 &dfBandDensity,
                                 &dfValueReal, &dfValueImag, psWrkStruct );
                }

                // If we didn't find any valid inputs skip to next band.
                if( dfBandDensity < BAND_DENSITY_THRESHOLD )
                    continue;

                bHasFoundDensity = true;

/* -------------------------------------------------------------------- */
/*      We have a computed value from the source.  Now apply it to      */
/*      the destination pixel.                                          */
/* -------------------------------------------------------------------- */
                GWKSetPixelValue( poWK, iBand, iDstOffset,
                                  dfBandDensity,
                                  dfValueReal, dfValueImag );
            }

            if( !bHasFoundDensity )
              continue;

/* -------------------------------------------------------------------- */
/*      Update destination density/validity masks.                      */
/* -------------------------------------------------------------------- */
            GWKOverlayDensity( poWK, iDstOffset, dfDensity );

            if( poWK->panDstValid != nullptr )
            {
                poWK->panDstValid[iDstOffset>>5] |=
                    0x01 << (iDstOffset & 0x1f);
            }
        } /* Next iDstX */

/* -------------------------------------------------------------------- */
/*      Report progress to the user, and optionally cancel out.         */
/* -------------------------------------------------------------------- */
        if( psJob->pfnProgress && psJob->pfnProgress(psJob) )
            break;
    }

/* -------------------------------------------------------------------- */
/*      Cleanup and return.                                             */
/* -------------------------------------------------------------------- */
    CPLFree( padfX );
    CPLFree( padfY );
    CPLFree( padfZ );
    CPLFree( pabSuccess );
    if( psWrkStruct )
        GWKResampleDeleteWrkStruct(psWrkStruct);
}

static CPLErr GWKGeneralCase( GDALWarpKernel *poWK )
{
    return GWKRun( poWK, "GWKGeneralCase", GWKGeneralCaseThread );
}

/************************************************************************/
/*                            GWKRealCase()                             */
/*                                                                      */
/*      General case for non-complex data types.                        */
/************************************************************************/

static void GWKRealCaseThread( void* pData)

{
    GWKJobStruct* psJob = static_cast<GWKJobStruct*>(pData);
    GDALWarpKernel *poWK = psJob->poWK;
    const int iYMin = psJob->iYMin;
    const int iYMax = psJob->iYMax;

    const int nDstXSize = poWK->nDstXSize;
    const int nSrcXSize = poWK->nSrcXSize;
    const int nSrcYSize = poWK->nSrcYSize;

/* -------------------------------------------------------------------- */
/*      Allocate x,y,z coordinate arrays for transformation ... one     */
/*      scanlines worth of positions.                                   */
/* -------------------------------------------------------------------- */

    // For x, 2 *, because we cache the precomputed values at the end.
    double *padfX =
        static_cast<double *>(CPLMalloc(2 * sizeof(double) * nDstXSize));
    double *padfY =
        static_cast<double *>(CPLMalloc(sizeof(double) * nDstXSize));
    double *padfZ =
        static_cast<double *>(CPLMalloc(sizeof(double) * nDstXSize));
    int *pabSuccess = static_cast<int *>(CPLMalloc(sizeof(int) * nDstXSize));

    const bool bUse4SamplesFormula =
        poWK->dfXScale >= 0.95 && poWK->dfYScale >= 0.95;

    GWKResampleWrkStruct* psWrkStruct = nullptr;
    if( poWK->eResample != GRA_NearestNeighbour )
    {
        psWrkStruct = GWKResampleCreateWrkStruct(poWK);
    }
    const double dfSrcCoordPrecision = CPLAtof(
        CSLFetchNameValueDef(poWK->papszWarpOptions,
                             "SRC_COORD_PRECISION", "0"));
    const double dfErrorThreshold = CPLAtof(
        CSLFetchNameValueDef(poWK->papszWarpOptions, "ERROR_THRESHOLD", "0"));

    const bool bSrcMaskIsDensity =
        poWK->panUnifiedSrcValid == nullptr &&
        poWK->papanBandSrcValid == nullptr &&
        poWK->pafUnifiedSrcDensity != nullptr;

    // Precompute values.
    for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
        padfX[nDstXSize + iDstX] = iDstX + 0.5 + poWK->nDstXOff;

/* ==================================================================== */
/*      Loop over output lines.                                         */
/* ==================================================================== */
    for( int iDstY = iYMin; iDstY < iYMax; iDstY++ )
    {
/* -------------------------------------------------------------------- */
/*      Setup points to transform to source image space.                */
/* -------------------------------------------------------------------- */
        memcpy( padfX, padfX + nDstXSize, sizeof(double) * nDstXSize );
        const double dfY = iDstY + 0.5 + poWK->nDstYOff;
        for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
            padfY[iDstX] = dfY;
        memset( padfZ, 0, sizeof(double) * nDstXSize );

/* -------------------------------------------------------------------- */
/*      Transform the points from destination pixel/line coordinates    */
/*      to source pixel/line coordinates.                               */
/* -------------------------------------------------------------------- */
        poWK->pfnTransformer( psJob->pTransformerArg, TRUE, nDstXSize,
                              padfX, padfY, padfZ, pabSuccess );
        if( dfSrcCoordPrecision > 0.0 )
        {
            GWKRoundSourceCoordinates(nDstXSize, padfX, padfY, padfZ,
                                      pabSuccess,
                                      dfSrcCoordPrecision,
                                      dfErrorThreshold,
                                      poWK->pfnTransformer,
                                      psJob->pTransformerArg,
                                      0.5 + poWK->nDstXOff,
                                      iDstY + 0.5 + poWK->nDstYOff);
        }

/* ==================================================================== */
/*      Loop over pixels in output scanline.                            */
/* ==================================================================== */
        for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
        {
            int iSrcOffset = 0;
            if( !GWKCheckAndComputeSrcOffsets(pabSuccess, iDstX, padfX, padfY,
                                    poWK, nSrcXSize, nSrcYSize, iSrcOffset) )
                continue;

/* -------------------------------------------------------------------- */
/*      Do not try to apply transparent/invalid source pixels to the    */
/*      destination.  This currently ignores the multi-pixel input      */
/*      of bilinear and cubic resamples.                                */
/* -------------------------------------------------------------------- */
            double dfDensity = 1.0;

            if( poWK->pafUnifiedSrcDensity != nullptr )
            {
                dfDensity = poWK->pafUnifiedSrcDensity[iSrcOffset];
                if( dfDensity < SRC_DENSITY_THRESHOLD )
                    continue;
            }

            if( poWK->panUnifiedSrcValid != nullptr
                && !(poWK->panUnifiedSrcValid[iSrcOffset>>5]
                     & (0x01 << (iSrcOffset & 0x1f))) )
                continue;

/* ==================================================================== */
/*      Loop processing each band.                                      */
/* ==================================================================== */
            bool bHasFoundDensity = false;

            const int iDstOffset = iDstX + iDstY * nDstXSize;
            for( int iBand = 0; iBand < poWK->nBands; iBand++ )
            {
                double dfBandDensity = 0.0;
                double dfValueReal = 0.0;

/* -------------------------------------------------------------------- */
/*      Collect the source value.                                       */
/* -------------------------------------------------------------------- */
                if( poWK->eResample == GRA_NearestNeighbour ||
                    nSrcXSize == 1 || nSrcYSize == 1 )
                {
                    // FALSE is returned if dfBandDensity == 0, which is
                    // checked below.
                    CPL_IGNORE_RET_VAL(
                        GWKGetPixelValueReal( poWK, iBand, iSrcOffset,
                                              &dfBandDensity, &dfValueReal ));
                }
                else if( poWK->eResample == GRA_Bilinear &&
                         bUse4SamplesFormula )
                {
                    double dfValueImagIgnored = 0.0;
                    GWKBilinearResample4Sample( poWK, iBand,
                                         padfX[iDstX]-poWK->nSrcXOff,
                                         padfY[iDstX]-poWK->nSrcYOff,
                                         &dfBandDensity,
                                         &dfValueReal, &dfValueImagIgnored );
                }
                else if( poWK->eResample == GRA_Cubic &&
                         bUse4SamplesFormula )
                {
                    if( bSrcMaskIsDensity )
                    {
                        if( poWK->eWorkingDataType == GDT_Byte )
                        {
                            GWKCubicResampleSrcMaskIsDensity4SampleRealT<GByte>(
                                                poWK, iBand,
                                                padfX[iDstX]-poWK->nSrcXOff,
                                                padfY[iDstX]-poWK->nSrcYOff,
                                                &dfBandDensity,
                                                &dfValueReal );
                        }
                        else if( poWK->eWorkingDataType == GDT_UInt16 )
                        {
                            GWKCubicResampleSrcMaskIsDensity4SampleRealT<GUInt16>(
                                                poWK, iBand,
                                                padfX[iDstX]-poWK->nSrcXOff,
                                                padfY[iDstX]-poWK->nSrcYOff,
                                                &dfBandDensity,
                                                &dfValueReal );
                        }
                        else
                        {
                            GWKCubicResampleSrcMaskIsDensity4SampleReal( poWK, iBand,
                                    padfX[iDstX]-poWK->nSrcXOff,
                                    padfY[iDstX]-poWK->nSrcYOff,
                                    &dfBandDensity,
                                    &dfValueReal );
                        }
                    }
                    else
                    {
                        double dfValueImagIgnored = 0.0;
                        GWKCubicResample4Sample( poWK, iBand,
                                            padfX[iDstX]-poWK->nSrcXOff,
                                            padfY[iDstX]-poWK->nSrcYOff,
                                            &dfBandDensity,
                                            &dfValueReal,
                                            &dfValueImagIgnored);
                    }
                }
                else
#ifdef DEBUG
                // Only useful for clang static analyzer.
                if( psWrkStruct != nullptr )
#endif
                {
                    double dfValueImagIgnored = 0.0;
                    psWrkStruct->pfnGWKResample(
                        poWK, iBand,
                        padfX[iDstX]-poWK->nSrcXOff,
                        padfY[iDstX]-poWK->nSrcYOff,
                        &dfBandDensity,
                        &dfValueReal, &dfValueImagIgnored, psWrkStruct);
                }

                // If we didn't find any valid inputs skip to next band.
                if( dfBandDensity < BAND_DENSITY_THRESHOLD )
                    continue;

                bHasFoundDensity = true;

/* -------------------------------------------------------------------- */
/*      We have a computed value from the source.  Now apply it to      */
/*      the destination pixel.                                          */
/* -------------------------------------------------------------------- */
                GWKSetPixelValueReal(poWK, iBand, iDstOffset,
                                     dfBandDensity,
                                     dfValueReal);
            }

            if( !bHasFoundDensity )
              continue;

/* -------------------------------------------------------------------- */
/*      Update destination density/validity masks.                      */
/* -------------------------------------------------------------------- */
            GWKOverlayDensity( poWK, iDstOffset, dfDensity );

            if( poWK->panDstValid != nullptr )
            {
                poWK->panDstValid[iDstOffset>>5] |=
                    0x01 << (iDstOffset & 0x1f);
            }
        }  // Next iDstX.

/* -------------------------------------------------------------------- */
/*      Report progress to the user, and optionally cancel out.         */
/* -------------------------------------------------------------------- */
        if( psJob->pfnProgress && psJob->pfnProgress(psJob) )
            break;
    }

/* -------------------------------------------------------------------- */
/*      Cleanup and return.                                             */
/* -------------------------------------------------------------------- */
    CPLFree( padfX );
    CPLFree( padfY );
    CPLFree( padfZ );
    CPLFree( pabSuccess );
    if( psWrkStruct )
        GWKResampleDeleteWrkStruct(psWrkStruct);
}

static CPLErr GWKRealCase( GDALWarpKernel *poWK )
{
    return GWKRun( poWK, "GWKRealCase", GWKRealCaseThread );
}

/************************************************************************/
/*                GWKResampleNoMasksOrDstDensityOnlyThreadInternal()    */
/************************************************************************/

template<class T, GDALResampleAlg eResample, int bUse4SamplesFormula>
static void GWKResampleNoMasksOrDstDensityOnlyThreadInternal( void* pData )

{
    GWKJobStruct* psJob = static_cast<GWKJobStruct *>(pData);
    GDALWarpKernel *poWK = psJob->poWK;
    const int iYMin = psJob->iYMin;
    const int iYMax = psJob->iYMax;

    const int nDstXSize = poWK->nDstXSize;
    const int nSrcXSize = poWK->nSrcXSize;
    const int nSrcYSize = poWK->nSrcYSize;

/* -------------------------------------------------------------------- */
/*      Allocate x,y,z coordinate arrays for transformation ... one     */
/*      scanlines worth of positions.                                   */
/* -------------------------------------------------------------------- */

    // For x, 2 *, because we cache the precomputed values at the end.
    double *padfX =
        static_cast<double *>(CPLMalloc(2 * sizeof(double) * nDstXSize));
    double *padfY =
        static_cast<double *>(CPLMalloc(sizeof(double) * nDstXSize));
    double *padfZ =
        static_cast<double *>(CPLMalloc(sizeof(double) * nDstXSize));
    int *pabSuccess = static_cast<int *>(CPLMalloc(sizeof(int) * nDstXSize));

    const int nXRadius = poWK->nXRadius;
    double  *padfWeight =
      static_cast<double *>(CPLCalloc(1 + nXRadius * 2, sizeof(double)));
    const double dfSrcCoordPrecision = CPLAtof(
        CSLFetchNameValueDef(poWK->papszWarpOptions,
                             "SRC_COORD_PRECISION", "0"));
    const double dfErrorThreshold = CPLAtof(
        CSLFetchNameValueDef(poWK->papszWarpOptions, "ERROR_THRESHOLD", "0"));

    // Precompute values.
    for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
        padfX[nDstXSize + iDstX] = iDstX + 0.5 + poWK->nDstXOff;

/* ==================================================================== */
/*      Loop over output lines.                                         */
/* ==================================================================== */
    for( int iDstY = iYMin; iDstY < iYMax; iDstY++ )
    {
/* -------------------------------------------------------------------- */
/*      Setup points to transform to source image space.                */
/* -------------------------------------------------------------------- */
        memcpy( padfX, padfX + nDstXSize, sizeof(double) * nDstXSize );
        const double dfY = iDstY + 0.5 + poWK->nDstYOff;
        for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
            padfY[iDstX] = dfY;
        memset( padfZ, 0, sizeof(double) * nDstXSize );

/* -------------------------------------------------------------------- */
/*      Transform the points from destination pixel/line coordinates    */
/*      to source pixel/line coordinates.                               */
/* -------------------------------------------------------------------- */
        poWK->pfnTransformer( psJob->pTransformerArg, TRUE, nDstXSize,
                              padfX, padfY, padfZ, pabSuccess );
        if( dfSrcCoordPrecision > 0.0 )
        {
            GWKRoundSourceCoordinates(nDstXSize, padfX, padfY, padfZ, pabSuccess,
                                      dfSrcCoordPrecision,
                                      dfErrorThreshold,
                                      poWK->pfnTransformer,
                                      psJob->pTransformerArg,
                                      0.5 + poWK->nDstXOff,
                                      iDstY + 0.5 + poWK->nDstYOff);
        }

/* ==================================================================== */
/*      Loop over pixels in output scanline.                            */
/* ==================================================================== */
        for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
        {
            int iSrcOffset = 0;
            if( !GWKCheckAndComputeSrcOffsets(pabSuccess, iDstX, padfX, padfY,
                                              poWK, nSrcXSize, nSrcYSize,
                                              iSrcOffset) )
                continue;

/* ==================================================================== */
/*      Loop processing each band.                                      */
/* ==================================================================== */
            const int iDstOffset = iDstX + iDstY * nDstXSize;

            for( int iBand = 0; iBand < poWK->nBands; iBand++ )
            {
                T value = 0;
                if( eResample == GRA_NearestNeighbour )
                {
                    value = reinterpret_cast<T *>(
                        poWK->papabySrcImage[iBand])[iSrcOffset];
                }
                else if( bUse4SamplesFormula )
                {
                    if( eResample == GRA_Bilinear )
                        GWKBilinearResampleNoMasks4SampleT( poWK, iBand,
                                                padfX[iDstX]-poWK->nSrcXOff,
                                                padfY[iDstX]-poWK->nSrcYOff,
                                                &value );
                    else
                        GWKCubicResampleNoMasks4SampleT( poWK, iBand,
                                              padfX[iDstX]-poWK->nSrcXOff,
                                              padfY[iDstX]-poWK->nSrcYOff,
                                              &value );
                }
                else
                {
                    GWKResampleNoMasksT( poWK, iBand,
                                    padfX[iDstX]-poWK->nSrcXOff,
                                    padfY[iDstX]-poWK->nSrcYOff,
                                    &value,
                                    padfWeight);
                }
                reinterpret_cast<T *>(poWK->papabyDstImage[iBand])[iDstOffset] =
                    value;
            }

            if( poWK->pafDstDensity )
                poWK->pafDstDensity[iDstOffset] = 1.0f;
        }

/* -------------------------------------------------------------------- */
/*      Report progress to the user, and optionally cancel out.         */
/* -------------------------------------------------------------------- */
        if( psJob->pfnProgress && psJob->pfnProgress(psJob) )
            break;
    }

/* -------------------------------------------------------------------- */
/*      Cleanup and return.                                             */
/* -------------------------------------------------------------------- */
    CPLFree( padfX );
    CPLFree( padfY );
    CPLFree( padfZ );
    CPLFree( pabSuccess );
    CPLFree( padfWeight );
}

template<class T, GDALResampleAlg eResample>
static void GWKResampleNoMasksOrDstDensityOnlyThread( void* pData )
{
    GWKResampleNoMasksOrDstDensityOnlyThreadInternal<T, eResample,
                                                     FALSE>(pData);
}

template<class T, GDALResampleAlg eResample>
static void GWKResampleNoMasksOrDstDensityOnlyHas4SampleThread( void* pData )

{
    GWKJobStruct* psJob = static_cast<GWKJobStruct *>(pData);
    GDALWarpKernel *poWK = psJob->poWK;
    CPLAssert(eResample == GRA_Bilinear || eResample == GRA_Cubic);
    const bool bUse4SamplesFormula =
        poWK->dfXScale >= 0.95 && poWK->dfYScale >= 0.95;
    if( bUse4SamplesFormula )
        GWKResampleNoMasksOrDstDensityOnlyThreadInternal<T, eResample,
                                                         TRUE>(pData);
    else
        GWKResampleNoMasksOrDstDensityOnlyThreadInternal<T, eResample,
                                                         FALSE>(pData);
}

static CPLErr GWKNearestNoMasksOrDstDensityOnlyByte( GDALWarpKernel *poWK )
{
    return GWKRun(
        poWK, "GWKNearestNoMasksOrDstDensityOnlyByte",
        GWKResampleNoMasksOrDstDensityOnlyThread<GByte, GRA_NearestNeighbour>);
}

static CPLErr GWKBilinearNoMasksOrDstDensityOnlyByte( GDALWarpKernel *poWK )
{
    return GWKRun(
        poWK, "GWKBilinearNoMasksOrDstDensityOnlyByte",
        GWKResampleNoMasksOrDstDensityOnlyHas4SampleThread<GByte,
                                                           GRA_Bilinear>);
}

static CPLErr GWKCubicNoMasksOrDstDensityOnlyByte( GDALWarpKernel *poWK )
{
    return GWKRun(
        poWK, "GWKCubicNoMasksOrDstDensityOnlyByte",
        GWKResampleNoMasksOrDstDensityOnlyHas4SampleThread<GByte, GRA_Cubic>);
}

static CPLErr GWKCubicNoMasksOrDstDensityOnlyFloat( GDALWarpKernel *poWK )
{
    return GWKRun(
        poWK, "GWKCubicNoMasksOrDstDensityOnlyFloat",
        GWKResampleNoMasksOrDstDensityOnlyHas4SampleThread<float, GRA_Cubic>);
}

#ifdef INSTANTIATE_FLOAT64_SSE2_IMPL

static CPLErr GWKCubicNoMasksOrDstDensityOnlyDouble( GDALWarpKernel *poWK )
{
    return GWKRun(
        poWK, "GWKCubicNoMasksOrDstDensityOnlyDouble",
        GWKResampleNoMasksOrDstDensityOnlyHas4SampleThread<double, GRA_Cubic>);
}
#endif

static CPLErr GWKCubicSplineNoMasksOrDstDensityOnlyByte( GDALWarpKernel *poWK )
{
    return GWKRun(
        poWK, "GWKCubicSplineNoMasksOrDstDensityOnlyByte",
        GWKResampleNoMasksOrDstDensityOnlyThread<GByte, GRA_CubicSpline>);
}

/************************************************************************/
/*                          GWKNearestByte()                            */
/*                                                                      */
/*      Case for 8bit input data with nearest neighbour resampling      */
/*      using valid flags. Should be as fast as possible for this       */
/*      particular transformation type.                                 */
/************************************************************************/

template<class T>
static void GWKNearestThread( void* pData )

{
    GWKJobStruct* psJob = static_cast<GWKJobStruct *>(pData);
    GDALWarpKernel *poWK = psJob->poWK;
    const int iYMin = psJob->iYMin;
    const int iYMax = psJob->iYMax;

    const int nDstXSize = poWK->nDstXSize;
    const int nSrcXSize = poWK->nSrcXSize;
    const int nSrcYSize = poWK->nSrcYSize;

/* -------------------------------------------------------------------- */
/*      Allocate x,y,z coordinate arrays for transformation ... one     */
/*      scanlines worth of positions.                                   */
/* -------------------------------------------------------------------- */

    // For x, 2 *, because we cache the precomputed values at the end.
    double *padfX =
        static_cast<double *>(CPLMalloc(2 * sizeof(double) * nDstXSize));
    double *padfY =
        static_cast<double *>(CPLMalloc(sizeof(double) * nDstXSize));
    double *padfZ =
       static_cast<double *>(CPLMalloc(sizeof(double) * nDstXSize));
    int *pabSuccess = static_cast<int *>(CPLMalloc(sizeof(int) * nDstXSize));

    const double dfSrcCoordPrecision = CPLAtof(
        CSLFetchNameValueDef(poWK->papszWarpOptions,
                             "SRC_COORD_PRECISION", "0"));
    const double dfErrorThreshold = CPLAtof(
        CSLFetchNameValueDef(poWK->papszWarpOptions, "ERROR_THRESHOLD", "0"));

    // Precompute values.
    for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
        padfX[nDstXSize + iDstX] = iDstX + 0.5 + poWK->nDstXOff;

/* ==================================================================== */
/*      Loop over output lines.                                         */
/* ==================================================================== */
    for( int iDstY = iYMin; iDstY < iYMax; iDstY++ )
    {

/* -------------------------------------------------------------------- */
/*      Setup points to transform to source image space.                */
/* -------------------------------------------------------------------- */
        memcpy( padfX, padfX + nDstXSize, sizeof(double) * nDstXSize );
        const double dfY = iDstY + 0.5 + poWK->nDstYOff;
        for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
            padfY[iDstX] = dfY;
        memset( padfZ, 0, sizeof(double) * nDstXSize );

/* -------------------------------------------------------------------- */
/*      Transform the points from destination pixel/line coordinates    */
/*      to source pixel/line coordinates.                               */
/* -------------------------------------------------------------------- */
        poWK->pfnTransformer( psJob->pTransformerArg, TRUE, nDstXSize,
                              padfX, padfY, padfZ, pabSuccess );
        if( dfSrcCoordPrecision > 0.0 )
        {
            GWKRoundSourceCoordinates(nDstXSize, padfX, padfY, padfZ,
                                      pabSuccess,
                                      dfSrcCoordPrecision,
                                      dfErrorThreshold,
                                      poWK->pfnTransformer,
                                      psJob->pTransformerArg,
                                      0.5 + poWK->nDstXOff,
                                      iDstY + 0.5 + poWK->nDstYOff);
        }
/* ==================================================================== */
/*      Loop over pixels in output scanline.                            */
/* ==================================================================== */
        for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
        {
            int iSrcOffset = 0;
            if( !GWKCheckAndComputeSrcOffsets(pabSuccess, iDstX, padfX, padfY,
                                    poWK, nSrcXSize, nSrcYSize, iSrcOffset) )
                continue;

/* -------------------------------------------------------------------- */
/*      Do not try to apply invalid source pixels to the dest.          */
/* -------------------------------------------------------------------- */
            if( poWK->panUnifiedSrcValid != nullptr
                && !(poWK->panUnifiedSrcValid[iSrcOffset>>5]
                     & (0x01 << (iSrcOffset & 0x1f))) )
                continue;

/* -------------------------------------------------------------------- */
/*      Do not try to apply transparent source pixels to the destination.*/
/* -------------------------------------------------------------------- */
            double dfDensity = 1.0;

            if( poWK->pafUnifiedSrcDensity != nullptr )
            {
                dfDensity = poWK->pafUnifiedSrcDensity[iSrcOffset];
                if( dfDensity < SRC_DENSITY_THRESHOLD )
                    continue;
            }

/* ==================================================================== */
/*      Loop processing each band.                                      */
/* ==================================================================== */

            const int iDstOffset = iDstX + iDstY * nDstXSize;

            for( int iBand = 0; iBand < poWK->nBands; iBand++ )
            {
                T value = 0;
                double dfBandDensity = 0.0;

/* -------------------------------------------------------------------- */
/*      Collect the source value.                                       */
/* -------------------------------------------------------------------- */
                if( GWKGetPixelT(poWK, iBand, iSrcOffset,
                                 &dfBandDensity, &value) )
                {
                    if( dfBandDensity < 1.0 )
                    {
                        if( dfBandDensity == 0.0 )
                        {
                            // Do nothing.
                        }
                        else
                        {
                            // Let the general code take care of mixing.
                            GWKSetPixelValueRealT( poWK, iBand, iDstOffset,
                                          dfBandDensity, value );
                        }
                    }
                    else
                    {
                        reinterpret_cast<T *>(
                            poWK->papabyDstImage[iBand])[iDstOffset] = value;
                    }
                }
            }

/* -------------------------------------------------------------------- */
/*      Mark this pixel valid/opaque in the output.                     */
/* -------------------------------------------------------------------- */
            GWKOverlayDensity( poWK, iDstOffset, dfDensity );

            if( poWK->panDstValid != nullptr )
            {
                poWK->panDstValid[iDstOffset>>5] |=
                    0x01 << (iDstOffset & 0x1f);
            }
        } /* Next iDstX */

/* -------------------------------------------------------------------- */
/*      Report progress to the user, and optionally cancel out.         */
/* -------------------------------------------------------------------- */
        if( psJob->pfnProgress && psJob->pfnProgress(psJob) )
            break;
    }

/* -------------------------------------------------------------------- */
/*      Cleanup and return.                                             */
/* -------------------------------------------------------------------- */
    CPLFree( padfX );
    CPLFree( padfY );
    CPLFree( padfZ );
    CPLFree( pabSuccess );
}

static CPLErr GWKNearestByte( GDALWarpKernel *poWK )
{
    return GWKRun( poWK, "GWKNearestByte", GWKNearestThread<GByte> );
}

static CPLErr GWKNearestNoMasksOrDstDensityOnlyShort( GDALWarpKernel *poWK )
{
    return GWKRun(
        poWK, "GWKNearestNoMasksOrDstDensityOnlyShort",
        GWKResampleNoMasksOrDstDensityOnlyThread<GInt16, GRA_NearestNeighbour>);
}

static CPLErr GWKBilinearNoMasksOrDstDensityOnlyShort( GDALWarpKernel *poWK )
{
    return GWKRun(
        poWK, "GWKBilinearNoMasksOrDstDensityOnlyShort",
        GWKResampleNoMasksOrDstDensityOnlyHas4SampleThread<GInt16,
                                                           GRA_Bilinear>);
}

static CPLErr GWKBilinearNoMasksOrDstDensityOnlyUShort( GDALWarpKernel *poWK )
{
    return GWKRun(
        poWK, "GWKBilinearNoMasksOrDstDensityOnlyUShort",
        GWKResampleNoMasksOrDstDensityOnlyHas4SampleThread<GUInt16,
                                                           GRA_Bilinear>);
}

static CPLErr GWKBilinearNoMasksOrDstDensityOnlyFloat( GDALWarpKernel *poWK )
{
    return GWKRun(
        poWK, "GWKBilinearNoMasksOrDstDensityOnlyFloat",
        GWKResampleNoMasksOrDstDensityOnlyHas4SampleThread<float,
                                                           GRA_Bilinear>);
}

#ifdef INSTANTIATE_FLOAT64_SSE2_IMPL

static CPLErr GWKBilinearNoMasksOrDstDensityOnlyDouble( GDALWarpKernel *poWK )
{
    return GWKRun(
        poWK, "GWKBilinearNoMasksOrDstDensityOnlyDouble",
        GWKResampleNoMasksOrDstDensityOnlyHas4SampleThread<double,
                                                           GRA_Bilinear>);
}
#endif

static CPLErr GWKCubicNoMasksOrDstDensityOnlyShort( GDALWarpKernel *poWK )
{
    return GWKRun(
        poWK, "GWKCubicNoMasksOrDstDensityOnlyShort",
        GWKResampleNoMasksOrDstDensityOnlyHas4SampleThread<GInt16, GRA_Cubic>);
}

static CPLErr GWKCubicNoMasksOrDstDensityOnlyUShort( GDALWarpKernel *poWK )
{
    return GWKRun(
        poWK, "GWKCubicNoMasksOrDstDensityOnlyUShort",
        GWKResampleNoMasksOrDstDensityOnlyHas4SampleThread<GUInt16, GRA_Cubic>);
}

static CPLErr GWKCubicSplineNoMasksOrDstDensityOnlyShort( GDALWarpKernel *poWK )
{
    return GWKRun(
        poWK, "GWKCubicSplineNoMasksOrDstDensityOnlyShort",
        GWKResampleNoMasksOrDstDensityOnlyThread<GInt16, GRA_CubicSpline>);
}

static CPLErr GWKCubicSplineNoMasksOrDstDensityOnlyUShort(
    GDALWarpKernel *poWK )
{
    return GWKRun(
        poWK, "GWKCubicSplineNoMasksOrDstDensityOnlyUShort",
        GWKResampleNoMasksOrDstDensityOnlyThread<GUInt16, GRA_CubicSpline>);
}

static CPLErr GWKNearestShort( GDALWarpKernel *poWK )
{
    return GWKRun(poWK, "GWKNearestShort", GWKNearestThread<GInt16>);
}

static CPLErr GWKNearestNoMasksOrDstDensityOnlyFloat( GDALWarpKernel *poWK )
{
    return GWKRun(
        poWK, "GWKNearestNoMasksOrDstDensityOnlyFloat",
        GWKResampleNoMasksOrDstDensityOnlyThread<float, GRA_NearestNeighbour>);
}

static CPLErr GWKNearestFloat( GDALWarpKernel *poWK )
{
    return GWKRun(poWK, "GWKNearestFloat", GWKNearestThread<float>);
}

/************************************************************************/
/*                           GWKAverageOrMode()                         */
/*                                                                      */
/************************************************************************/

static void GWKAverageOrModeThread(void* pData);

static CPLErr GWKAverageOrMode( GDALWarpKernel *poWK )
{
    return GWKRun( poWK, "GWKAverageOrMode", GWKAverageOrModeThread );
}

// Overall logic based on GWKGeneralCaseThread().
static void GWKAverageOrModeThread( void* pData)
{
    GWKJobStruct* psJob = static_cast<GWKJobStruct *>(pData);
    GDALWarpKernel *poWK = psJob->poWK;
    const int iYMin = psJob->iYMin;
    const int iYMax = psJob->iYMax;

    const int nDstXSize = poWK->nDstXSize;
    const int nSrcXSize = poWK->nSrcXSize;
    const int nSrcYSize = poWK->nSrcYSize;

/* -------------------------------------------------------------------- */
/*      Find out which algorithm to use (small optim.)                  */
/* -------------------------------------------------------------------- */
    int nAlgo = 0;

    // These vars only used with nAlgo == 3.
    int *panVals = nullptr;
    int nBins = 0;
    int nBinsOffset = 0;

    // Only used with nAlgo = 2.
    float* pafRealVals = nullptr;
    float* pafImagVals = nullptr;
    int* panRealSums = nullptr;
    int* panImagSums = nullptr;

    // Only used with nAlgo = 6.
    float quant = 0.5;

    //To control array allocation only when data type is complex
    const bool bIsComplex = GDALDataTypeIsComplex(poWK->eWorkingDataType)!=0;

    if( poWK->eResample == GRA_Average )
    {
        nAlgo = GWKAOM_Average;
    }
    else if( poWK->eResample == GRA_Mode )
    {
        // TODO check color table count > 256.
        if( poWK->eWorkingDataType == GDT_Byte ||
            poWK->eWorkingDataType == GDT_UInt16 ||
            poWK->eWorkingDataType == GDT_Int16 )
        {
            nAlgo = GWKAOM_Imode;

            // In the case of a paletted or non-paletted byte band,
            // Input values are between 0 and 255.
            if( poWK->eWorkingDataType == GDT_Byte )
            {
                nBins = 256;
            }
            // In the case of Int16, input values are between -32768 and 32767.
            else if( poWK->eWorkingDataType == GDT_Int16 )
            {
                nBins = 65536;
                nBinsOffset = 32768;
            }
            // In the case of UInt16, input values are between 0 and 65537.
            else if( poWK->eWorkingDataType == GDT_UInt16 )
            {
                nBins = 65536;
            }
            panVals =
                static_cast<int *>(VSI_MALLOC_VERBOSE(nBins * sizeof(int)));
            if( panVals == nullptr )
                return;
        }
        else
        {
            nAlgo = GWKAOM_Fmode;

            if( nSrcXSize > 0 && nSrcYSize > 0 )
            {
                pafRealVals = static_cast<float *>(
                    VSI_MALLOC3_VERBOSE(nSrcXSize, nSrcYSize, sizeof(float)));
                panRealSums = static_cast<int *>(
                    VSI_MALLOC3_VERBOSE(nSrcXSize, nSrcYSize, sizeof(int)));
                if( pafRealVals == nullptr || panRealSums == nullptr )
                {
                    VSIFree(pafRealVals);
                    VSIFree(panRealSums);
                    return;
                }
            }
        }
    }
    else if( poWK->eResample == GRA_Max )
    {
        nAlgo = GWKAOM_Max;
    }
    else if( poWK->eResample == GRA_Min )
    {
        nAlgo = GWKAOM_Min;
    }
    else if( poWK->eResample == GRA_Med )
    {
        nAlgo = GWKAOM_Quant;
        quant = 0.5;
    }
    else if( poWK->eResample == GRA_Q1 )
    {
        nAlgo = GWKAOM_Quant;
        quant = 0.25;
    }
    else if( poWK->eResample == GRA_Q3 )
    {
        nAlgo = GWKAOM_Quant;
        quant = 0.75;
    }
    else
    {
        // Other resample algorithms not permitted here.
        CPLDebug("GDAL",
                 "GDALWarpKernel():GWKAverageOrModeThread() ERROR, "
                 "illegal resample" );
        return;
    }

    CPLDebug("GDAL",
             "GDALWarpKernel():GWKAverageOrModeThread() using algo %d", nAlgo);

/* -------------------------------------------------------------------- */
/*      Allocate x,y,z coordinate arrays for transformation ... two     */
/*      scanlines worth of positions.                                   */
/* -------------------------------------------------------------------- */

    double *padfX =
        static_cast<double *>(CPLMalloc(sizeof(double) * nDstXSize));
    double *padfY =
        static_cast<double *>(CPLMalloc(sizeof(double) * nDstXSize));
    double *padfZ =
        static_cast<double *>(CPLMalloc(sizeof(double) * nDstXSize));
    double *padfX2 =
        static_cast<double *>(CPLMalloc(sizeof(double) * nDstXSize));
    double *padfY2 =
        static_cast<double *>(CPLMalloc(sizeof(double) * nDstXSize));
    double *padfZ2 =
        static_cast<double *>(CPLMalloc(sizeof(double) * nDstXSize));
    int *pabSuccess = static_cast<int *>(CPLMalloc(sizeof(int) * nDstXSize));
    int *pabSuccess2 = static_cast<int *>(CPLMalloc(sizeof(int) * nDstXSize));

    const double dfSrcCoordPrecision = CPLAtof(
        CSLFetchNameValueDef(poWK->papszWarpOptions,
                             "SRC_COORD_PRECISION", "0"));
    const double dfErrorThreshold = CPLAtof(
        CSLFetchNameValueDef(poWK->papszWarpOptions, "ERROR_THRESHOLD", "0"));

/* ==================================================================== */
/*      Loop over output lines.                                         */
/* ==================================================================== */
    for( int iDstY = iYMin; iDstY < iYMax; iDstY++ )
    {

/* -------------------------------------------------------------------- */
/*      Setup points to transform to source image space.                */
/* -------------------------------------------------------------------- */
        for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
        {
            padfX[iDstX] = iDstX + poWK->nDstXOff;
            padfY[iDstX] = iDstY + poWK->nDstYOff;
            padfZ[iDstX] = 0.0;
            padfX2[iDstX] = iDstX + 1.0 + poWK->nDstXOff;
            padfY2[iDstX] = iDstY + 1.0 + poWK->nDstYOff;
            padfZ2[iDstX] = 0.0;
        }

/* -------------------------------------------------------------------- */
/*      Transform the points from destination pixel/line coordinates    */
/*      to source pixel/line coordinates.                               */
/* -------------------------------------------------------------------- */
        poWK->pfnTransformer( psJob->pTransformerArg, TRUE, nDstXSize,
                              padfX, padfY, padfZ, pabSuccess );
        poWK->pfnTransformer( psJob->pTransformerArg, TRUE, nDstXSize,
                              padfX2, padfY2, padfZ2, pabSuccess2 );

        if( dfSrcCoordPrecision > 0.0 )
        {
            GWKRoundSourceCoordinates(nDstXSize, padfX, padfY, padfZ,
                                      pabSuccess,
                                      dfSrcCoordPrecision,
                                      dfErrorThreshold,
                                      poWK->pfnTransformer,
                                      psJob->pTransformerArg,
                                      poWK->nDstXOff,
                                      iDstY + poWK->nDstYOff);
            GWKRoundSourceCoordinates(nDstXSize, padfX2, padfY2, padfZ2,
                                      pabSuccess2,
                                      dfSrcCoordPrecision,
                                      dfErrorThreshold,
                                      poWK->pfnTransformer,
                                      psJob->pTransformerArg,
                                      1.0 + poWK->nDstXOff,
                                      iDstY + 1.0 + poWK->nDstYOff);
        }

/* ==================================================================== */
/*      Loop over pixels in output scanline.                            */
/* ==================================================================== */
        for( int iDstX = 0; iDstX < nDstXSize; iDstX++ )
        {
            int iSrcOffset = 0;
            double dfDensity = 1.0;
            bool bHasFoundDensity = false;

            if( !pabSuccess[iDstX] || !pabSuccess2[iDstX] )
                continue;
            const int iDstOffset = iDstX + iDstY * nDstXSize;

/* ==================================================================== */
/*      Loop processing each band.                                      */
/* ==================================================================== */

            for( int iBand = 0; iBand < poWK->nBands; iBand++ )
            {
                double dfBandDensity = 0.0;
                double dfValueReal = 0.0;
                double dfValueImag = 0.0;
                double dfValueRealTmp = 0.0;
                double dfValueImagTmp = 0.0;

/* -------------------------------------------------------------------- */
/*      Collect the source value.                                       */
/* -------------------------------------------------------------------- */
                double dfTotalReal = 0.0;
                double dfTotalImag = 0.0;
                // Count of pixels used to compute average/mode.
                int nCount = 0;
                // Count of all pixels sampled, including nodata.
                int nCount2 = 0;

                // Compute corners in source crs.

                // The transformation might not have preserved ordering of
                // coordinates so do the necessary swapping (#5433).
                // NOTE: this is really an approximative fix. To do something
                // more precise we would for example need to compute the
                // transformation of coordinates in the
                // [iDstX,iDstY]x[iDstX+1,iDstY+1] square back to source
                // coordinates, and take the bounding box of the got source
                // coordinates.
                int iSrcXMin =
                    std::max(static_cast<int>(floor(std::min(padfX[iDstX],padfX2[iDstX]) + 1e-10)) -
                             poWK->nSrcXOff, 0);
                int iSrcXMax =
                    std::min(static_cast<int>(ceil(std::max(padfX[iDstX],padfX2[iDstX]) - 1e-10)) -
                             poWK->nSrcXOff, nSrcXSize);
                int iSrcYMin =
                    std::max(static_cast<int>(floor(std::min(padfY[iDstX],padfY2[iDstX]) + 1e-10)) -
                             poWK->nSrcYOff, 0);
                int iSrcYMax =
                    std::min(static_cast<int>(ceil(std::max(padfY[iDstX],padfY2[iDstX]) - 1e-10)) -
                             poWK->nSrcYOff, nSrcYSize);

                if( iSrcXMin == iSrcXMax && iSrcXMax < nSrcXSize )
                    iSrcXMax++;
                if( iSrcYMin == iSrcYMax && iSrcYMax < nSrcYSize )
                    iSrcYMax++;

                // Loop over source lines and pixels - 3 possible algorithms.

                // poWK->eResample == GRA_Average.
                if( nAlgo == GWKAOM_Average )
                {
                    // This code adapted from GDALDownsampleChunk32R_AverageT()
                    // in gcore/overview.cpp.
                    for( int iSrcY = iSrcYMin; iSrcY < iSrcYMax; iSrcY++ )
                    {
                        for( int iSrcX = iSrcXMin; iSrcX < iSrcXMax; iSrcX++ )
                        {
                            iSrcOffset = iSrcX + iSrcY * nSrcXSize;

                            if( poWK->panUnifiedSrcValid != nullptr
                                && !(poWK->panUnifiedSrcValid[iSrcOffset>>5]
                                     & (0x01 << (iSrcOffset & 0x1f))) )
                            {
                                continue;
                            }

                            nCount2++;
                            if( GWKGetPixelValue(
                                    poWK, iBand, iSrcOffset,
                                    &dfBandDensity, &dfValueRealTmp,
                                    &dfValueImagTmp ) &&
                                dfBandDensity > BAND_DENSITY_THRESHOLD )
                            {
                                nCount++;
                                dfTotalReal += dfValueRealTmp;
                                if (bIsComplex)
                                {
                                    dfTotalImag += dfValueImagTmp;
                                }
                            }
                        }
                    }

                    if( nCount > 0 )
                    {
                        dfValueReal = dfTotalReal / nCount;
                        if (bIsComplex)
                        {
                            dfValueImag = dfTotalImag / nCount;
                        }
                        dfBandDensity = 1;
                        bHasFoundDensity = true;
                    }
                }  // GRA_Average.
                else if( nAlgo == GWKAOM_Imode || nAlgo == GWKAOM_Fmode )
                // poWK->eResample == GRA_Mode
                {
                    // This code adapted from GDALDownsampleChunk32R_Mode() in
                    // gcore/overview.cpp.
                    if( nAlgo == GWKAOM_Fmode ) // int32 or float.
                    {
                        // Does it make sense it makes to run a
                        // majority filter on floating point data? But, here it
                        // is for the sake of compatibility. It won't look
                        // right on RGB images by the nature of the filter.
                        int iMaxInd = 0;
                        int iMaxVal = -1;
                        int i = 0;

                        for( int iSrcY = iSrcYMin; iSrcY < iSrcYMax; iSrcY++ )
                        {
                            for( int iSrcX = iSrcXMin;
                                 iSrcX < iSrcXMax;
                                 iSrcX++ )
                            {
                                iSrcOffset = iSrcX + iSrcY * nSrcXSize;

                                if( poWK->panUnifiedSrcValid != nullptr
                                    && !(poWK->panUnifiedSrcValid[iSrcOffset>>5]
                                         & (0x01 << (iSrcOffset & 0x1f))) )
                                    continue;

                                nCount2++;
                                if( GWKGetPixelValue(
                                        poWK, iBand, iSrcOffset,
                                        &dfBandDensity, &dfValueRealTmp,
                                        &dfValueImagTmp ) &&
                                    dfBandDensity > BAND_DENSITY_THRESHOLD )
                                {
                                    nCount++;

                                    const float fVal =
                                        static_cast<float>(dfValueRealTmp);

                                    // Check array for existing entry.
                                    for( i = 0; i < iMaxInd; ++i )
                                        if( pafRealVals[i] == fVal
                                            && ++panRealSums[i] > panRealSums[iMaxVal] )
                                        {
                                            iMaxVal = i;
                                            break;
                                        }

                                    // Add to arr if entry not already there.
                                    if( i == iMaxInd )
                                    {
                                        pafRealVals[iMaxInd] = fVal;
                                        panRealSums[iMaxInd] = 1;

                                        if( iMaxVal < 0 )
                                            iMaxVal = iMaxInd;

                                        ++iMaxInd;
                                    }
                                }
                            }
                        }

                        if( iMaxVal != -1 )
                        {
                            dfValueReal = pafRealVals[iMaxVal];
                            dfBandDensity = 1;
                            bHasFoundDensity = true;
                        }
                    }
                    else // byte or int16.
                    {
                        int nMaxVal = 0;
                        int iMaxInd = -1;

                        memset(panVals, 0, nBins*sizeof(int));

                        for( int iSrcY = iSrcYMin; iSrcY < iSrcYMax; iSrcY++ )
                        {
                            for( int iSrcX = iSrcXMin;
                                 iSrcX < iSrcXMax;
                                 iSrcX++ )
                            {
                                iSrcOffset = iSrcX + iSrcY * nSrcXSize;

                                if( poWK->panUnifiedSrcValid != nullptr
                                    && !(poWK->panUnifiedSrcValid[iSrcOffset>>5]
                                         & (0x01 << (iSrcOffset & 0x1f))) )
                                    continue;

                                nCount2++;
                                if( GWKGetPixelValue(
                                        poWK, iBand, iSrcOffset,
                                        &dfBandDensity, &dfValueRealTmp,
                                        &dfValueImagTmp ) &&
                                    dfBandDensity > BAND_DENSITY_THRESHOLD )
                                {
                                    nCount++;

                                    const int nVal =
                                        static_cast<int>(dfValueRealTmp);
                                    if( ++panVals[nVal+nBinsOffset] > nMaxVal )
                                    {
                                        // Sum the density.
                                        // Is it the most common value so far?
                                        iMaxInd = nVal;
                                        nMaxVal = panVals[nVal+nBinsOffset];
                                    }
                                }
                            }
                        }

                        if( iMaxInd != -1 )
                        {
                            dfValueReal = iMaxInd;
                            dfBandDensity = 1;
                            bHasFoundDensity = true;
                        }
                    }
                }  // GRA_Mode.
                else if( nAlgo == GWKAOM_Max )
                // poWK->eResample == GRA_Max.
                {
                    dfTotalReal = std::numeric_limits<double>::lowest();
                    // This code adapted from nAlgo 1 method, GRA_Average.
                    for( int iSrcY = iSrcYMin; iSrcY < iSrcYMax; iSrcY++ )
                    {
                        for( int iSrcX = iSrcXMin; iSrcX < iSrcXMax; iSrcX++ )
                        {
                            iSrcOffset = iSrcX + iSrcY * nSrcXSize;

                            if( poWK->panUnifiedSrcValid != nullptr
                                && !(poWK->panUnifiedSrcValid[iSrcOffset>>5]
                                     & (0x01 << (iSrcOffset & 0x1f))) )
                            {
                                continue;
                            }

                            // Returns pixel value if it is not no data.
                            if( GWKGetPixelValue(
                                    poWK, iBand, iSrcOffset,
                                    &dfBandDensity, &dfValueRealTmp,
                                    &dfValueImagTmp ) &&
                                dfBandDensity > BAND_DENSITY_THRESHOLD )
                            {
                                nCount++;
                                if( dfTotalReal < dfValueRealTmp )
                                {
                                    dfTotalReal = dfValueRealTmp;
                                }
                            }
                        }
                    }

                    if( nCount > 0 )
                    {
                        dfValueReal = dfTotalReal;
                        dfBandDensity = 1;
                        bHasFoundDensity = true;
                    }
                }  // GRA_Max.
                else if( nAlgo == GWKAOM_Min )
                // poWK->eResample == GRA_Min.
                {
                    dfTotalReal = std::numeric_limits<double>::max();
                    // This code adapted from nAlgo 1 method, GRA_Average.
                    for( int iSrcY = iSrcYMin; iSrcY < iSrcYMax; iSrcY++ )
                    {
                        for( int iSrcX = iSrcXMin; iSrcX < iSrcXMax; iSrcX++ )
                        {
                            iSrcOffset = iSrcX + iSrcY * nSrcXSize;

                            if( poWK->panUnifiedSrcValid != nullptr
                                && !(poWK->panUnifiedSrcValid[iSrcOffset>>5]
                                     & (0x01 << (iSrcOffset & 0x1f))) )
                            {
                                continue;
                            }

                            // Returns pixel value if it is not no data.
                            if( GWKGetPixelValue(
                                    poWK, iBand, iSrcOffset,
                                    &dfBandDensity, &dfValueRealTmp,
                                    &dfValueImagTmp ) &&
                                dfBandDensity > BAND_DENSITY_THRESHOLD )
                            {
                                nCount++;
                                if( dfTotalReal > dfValueRealTmp )
                                {
                                    dfTotalReal = dfValueRealTmp;
                                }
                            }
                        }
                    }

                    if( nCount > 0 )
                    {
                        dfValueReal = dfTotalReal;
                        dfBandDensity = 1;
                        bHasFoundDensity = true;
                    }
                }  // GRA_Min.
                else if( nAlgo == GWKAOM_Quant )
                // poWK->eResample == GRA_Med | GRA_Q1 | GRA_Q3.
                {
                    std::vector<double> dfRealValuesTmp;

                    // This code adapted from nAlgo 1 method, GRA_Average.
                    for( int iSrcY = iSrcYMin; iSrcY < iSrcYMax; iSrcY++ )
                    {
                        for( int iSrcX = iSrcXMin; iSrcX < iSrcXMax; iSrcX++ )
                        {
                            iSrcOffset = iSrcX + iSrcY * nSrcXSize;

                            if( poWK->panUnifiedSrcValid != nullptr
                                && !(poWK->panUnifiedSrcValid[iSrcOffset>>5]
                                     & (0x01 << (iSrcOffset & 0x1f))) )
                            {
                                continue;
                            }

                            // Returns pixel value if it is not no data.
                            if( GWKGetPixelValue(
                                    poWK, iBand, iSrcOffset,
                                    &dfBandDensity, &dfValueRealTmp,
                                    &dfValueImagTmp ) &&
                                dfBandDensity > BAND_DENSITY_THRESHOLD )
                            {
                                nCount++;
                                dfRealValuesTmp.push_back(dfValueRealTmp);
                            }
                        }
                    }

                    if( nCount > 0 )
                    {
                        std::sort(dfRealValuesTmp.begin(), dfRealValuesTmp.end());
                        int quantIdx = static_cast<int>(
                            std::ceil(quant * dfRealValuesTmp.size() - 1));
                        dfValueReal = dfRealValuesTmp[quantIdx];

                        dfBandDensity = 1;
                        bHasFoundDensity = true;
                        dfRealValuesTmp.clear();
                    }
                }  // Quantile.

/* -------------------------------------------------------------------- */
/*      We have a computed value from the source.  Now apply it to      */
/*      the destination pixel.                                          */
/* -------------------------------------------------------------------- */
                if( bHasFoundDensity )
                {
                    // TODO: Should we compute dfBandDensity in fct of
                    // nCount/nCount2, or use as a threshold to set the dest
                    // value?
                    // dfBandDensity = (float) nCount / nCount2;
                    // if( (float) nCount / nCount2 > 0.1 )
                    // or fix gdalwarp crop_to_cutline to crop partially
                    // overlapping pixels.
                    GWKSetPixelValue( poWK, iBand, iDstOffset,
                                      dfBandDensity,
                                      dfValueReal, dfValueImag );
                }
            }

            if( !bHasFoundDensity )
                continue;

/* -------------------------------------------------------------------- */
/*      Update destination density/validity masks.                      */
/* -------------------------------------------------------------------- */
            GWKOverlayDensity( poWK, iDstOffset, dfDensity );

            if( poWK->panDstValid != nullptr )
            {
                poWK->panDstValid[iDstOffset>>5] |=
                    0x01 << (iDstOffset & 0x1f);
            }
        } /* Next iDstX */

/* -------------------------------------------------------------------- */
/*      Report progress to the user, and optionally cancel out.         */
/* -------------------------------------------------------------------- */
        if( psJob->pfnProgress && psJob->pfnProgress(psJob) )
            break;
    }

/* -------------------------------------------------------------------- */
/*      Cleanup and return.                                             */
/* -------------------------------------------------------------------- */
    CPLFree( padfX );
    CPLFree( padfY );
    CPLFree( padfZ );
    CPLFree( padfX2 );
    CPLFree( padfY2 );
    CPLFree( padfZ2 );
    CPLFree( pabSuccess );
    CPLFree( pabSuccess2 );
    VSIFree( panVals );
    VSIFree(pafRealVals);
    VSIFree(panRealSums);
    if (bIsComplex)
    {
        VSIFree(pafImagVals);
        VSIFree(panImagSums);
    }
}