Пример #1
0
int
NI_GeometricTransform(PyArrayObject *input, int (*map)(npy_intp*, double*,
                int, int, void*), void* map_data, PyArrayObject* matrix_ar,
                PyArrayObject* shift_ar, PyArrayObject *coordinates,
                PyArrayObject *output, int order, int mode, double cval)
{
    char *po, *pi, *pc = NULL;
    npy_intp **edge_offsets = NULL, **data_offsets = NULL, filter_size;
    npy_intp ftmp[MAXDIM], *fcoordinates = NULL, *foffsets = NULL;
    npy_intp cstride = 0, kk, hh, ll, jj;
    npy_intp size;
    double **splvals = NULL, icoor[MAXDIM];
    npy_intp idimensions[MAXDIM], istrides[MAXDIM];
    NI_Iterator io, ic;
    Float64 *matrix = matrix_ar ? (Float64*)PyArray_DATA(matrix_ar) : NULL;
    Float64 *shift = shift_ar ? (Float64*)PyArray_DATA(shift_ar) : NULL;
    int irank = 0, orank, qq;

    for(kk = 0; kk < input->nd; kk++) {
        idimensions[kk] = input->dimensions[kk];
        istrides[kk] = input->strides[kk];
    }
    irank = input->nd;
    orank = output->nd;

    /* if the mapping is from array coordinates: */
    if (coordinates) {
        /* initialze a line iterator along the first axis: */
        if (!NI_InitPointIterator(coordinates, &ic))
            goto exit;
        cstride = ic.strides[0];
        if (!NI_LineIterator(&ic, 0))
            goto exit;
        pc = (void *)(PyArray_DATA(coordinates));
    }

    /* offsets used at the borders: */
    edge_offsets = (npy_intp**)malloc(irank * sizeof(npy_intp*));
    data_offsets = (npy_intp**)malloc(irank * sizeof(npy_intp*));
    if (!edge_offsets || !data_offsets) {
        PyErr_NoMemory();
        goto exit;
    }
    for(jj = 0; jj < irank; jj++)
        data_offsets[jj] = NULL;
    for(jj = 0; jj < irank; jj++) {
        data_offsets[jj] = (npy_intp*)malloc((order + 1) * sizeof(npy_intp));
        if (!data_offsets[jj]) {
            PyErr_NoMemory();
            goto exit;
        }
    }
    /* will hold the spline coefficients: */
    splvals = (double**)malloc(irank * sizeof(double*));
    if (!splvals) {
        PyErr_NoMemory();
        goto exit;
    }
    for(jj = 0; jj < irank; jj++)
        splvals[jj] = NULL;
    for(jj = 0; jj < irank; jj++) {
        splvals[jj] = (double*)malloc((order + 1) * sizeof(double));
        if (!splvals[jj]) {
            PyErr_NoMemory();
            goto exit;
        }
    }

    filter_size = 1;
    for(jj = 0; jj < irank; jj++)
        filter_size *= order + 1;

    /* initialize output iterator: */
    if (!NI_InitPointIterator(output, &io))
        goto exit;

    /* get data pointers: */
    pi = (void *)PyArray_DATA(input);
    po = (void *)PyArray_DATA(output);

    /* make a table of all possible coordinates within the spline filter: */
    fcoordinates = (npy_intp*)malloc(irank * filter_size * sizeof(npy_intp));
    /* make a table of all offsets within the spline filter: */
    foffsets = (npy_intp*)malloc(filter_size * sizeof(npy_intp));
    if (!fcoordinates || !foffsets) {
        PyErr_NoMemory();
        goto exit;
    }
    for(jj = 0; jj < irank; jj++)
        ftmp[jj] = 0;
    kk = 0;
    for(hh = 0; hh < filter_size; hh++) {
        for(jj = 0; jj < irank; jj++)
            fcoordinates[jj + hh * irank] = ftmp[jj];
        foffsets[hh] = kk;
        for(jj = irank - 1; jj >= 0; jj--) {
            if (ftmp[jj] < order) {
                ftmp[jj]++;
                kk += istrides[jj];
                break;
            } else {
                ftmp[jj] = 0;
                kk -= istrides[jj] * order;
            }
        }
    }

    size = 1;
    for(qq = 0; qq < output->nd; qq++)
        size *= output->dimensions[qq];
    for(kk = 0; kk < size; kk++) {
        double t = 0.0;
        int constant = 0, edge = 0, offset = 0;
        if (map) {
            /* call mappint functions: */
            if (!map(io.coordinates, icoor, orank, irank, map_data)) {
                if (!PyErr_Occurred())
                    PyErr_SetString(PyExc_RuntimeError,
                                                    "unknown error in mapping function");
                goto exit;
            }
        } else if (matrix) {
            /* do an affine transformation: */
            Float64 *p = matrix;
            for(hh = 0; hh < irank; hh++) {
                icoor[hh] = 0.0;
                for(ll = 0; ll < orank; ll++)
                    icoor[hh] += io.coordinates[ll] * *p++;
                icoor[hh] += shift[hh];
            }
        } else if (coordinates) {
            /* mapping is from an coordinates array: */
            char *p = pc;
            switch (NI_NormalizeType(coordinates->descr->type_num)) {
                CASE_MAP_COORDINATES(p, icoor, irank, cstride, Bool);
                CASE_MAP_COORDINATES(p, icoor, irank, cstride, UInt8);
                CASE_MAP_COORDINATES(p, icoor, irank, cstride, UInt16);
                CASE_MAP_COORDINATES(p, icoor, irank, cstride, UInt32);
#if HAS_UINT64
                CASE_MAP_COORDINATES(p, icoor, irank, cstride, UInt64);
#endif
                CASE_MAP_COORDINATES(p, icoor, irank, cstride, Int8);
                CASE_MAP_COORDINATES(p, icoor, irank, cstride, Int16);
                CASE_MAP_COORDINATES(p, icoor, irank, cstride, Int32);
                CASE_MAP_COORDINATES(p, icoor, irank, cstride, Int64);
                CASE_MAP_COORDINATES(p, icoor, irank, cstride, Float32);
                CASE_MAP_COORDINATES(p, icoor, irank, cstride, Float64);
            default:
                PyErr_SetString(PyExc_RuntimeError,
                                                "coordinate array data type not supported");
                goto exit;
            }
        }
        /* iterate over axes: */
        for(hh = 0; hh < irank; hh++) {
            /* if the input coordinate is outside the borders, map it: */
            double cc = map_coordinate(icoor[hh], idimensions[hh], mode);
            if (cc > -1.0) {
                /* find the filter location along this axis: */
                int start;
                if (order & 1) {
                    start = (int)floor(cc) - order / 2;
                } else {
                    start = (int)floor(cc + 0.5) - order / 2;
                }
                /* get the offset to the start of the filter: */
                offset += istrides[hh] * start;
                if (start < 0 || start + order >= idimensions[hh]) {
                    /* implement border mapping, if outside border: */
                    edge = 1;
                    edge_offsets[hh] = data_offsets[hh];
                    for(ll = 0; ll <= order; ll++) {
                        int idx = start + ll;
                        int len = idimensions[hh];
                        if (len <= 1) {
                            idx = 0;
                        } else {
                            int s2 = 2 * len - 2;
                            if (idx < 0) {
                                idx = s2 * (int)(-idx / s2) + idx;
                                idx = idx <= 1 - len ? idx + s2 : -idx;
                            } else if (idx >= len) {
                                idx -= s2 * (int)(idx / s2);
                                if (idx >= len)
                                    idx = s2 - idx;
                            }
                        }
                        /* calculate and store the offests at this edge: */
                        edge_offsets[hh][ll] = istrides[hh] * (idx - start);
                    }
                } else {
                    /* we are not at the border, use precalculated offsets: */
                    edge_offsets[hh] = NULL;
                }
                spline_coefficients(cc, order, splvals[hh]);
            } else {
                /* we use the constant border condition: */
                constant = 1;
                break;
            }
        }

        if (!constant) {
            npy_intp *ff = fcoordinates;
            const int type_num = NI_NormalizeType(input->descr->type_num);
            t = 0.0;
            for(hh = 0; hh < filter_size; hh++) {
                double coeff = 0.0;
                int idx = 0;

                if (NI_UNLIKELY(edge)) {
                    for(ll = 0; ll < irank; ll++) {
                        if (edge_offsets[ll])
                            idx += edge_offsets[ll][ff[ll]];
                        else
                            idx += ff[ll] * istrides[ll];
                    }
                } else {
                    idx = foffsets[hh];
                }
                idx += offset;
                switch (type_num) {
                    CASE_INTERP_COEFF(coeff, pi, idx, Bool);
                    CASE_INTERP_COEFF(coeff, pi, idx, UInt8);
                    CASE_INTERP_COEFF(coeff, pi, idx, UInt16);
                    CASE_INTERP_COEFF(coeff, pi, idx, UInt32);
#if HAS_UINT64
                    CASE_INTERP_COEFF(coeff, pi, idx, UInt64);
#endif
                    CASE_INTERP_COEFF(coeff, pi, idx, Int8);
                    CASE_INTERP_COEFF(coeff, pi, idx, Int16);
                    CASE_INTERP_COEFF(coeff, pi, idx, Int32);
                    CASE_INTERP_COEFF(coeff, pi, idx, Int64);
                    CASE_INTERP_COEFF(coeff, pi, idx, Float32);
                    CASE_INTERP_COEFF(coeff, pi, idx, Float64);
                default:
                    PyErr_SetString(PyExc_RuntimeError,
                                                    "data type not supported");
                    goto exit;
                }
                /* calculate the interpolated value: */
                for(ll = 0; ll < irank; ll++)
                    if (order > 0)
                        coeff *= splvals[ll][ff[ll]];
                t += coeff;
                ff += irank;
            }
        } else {
            t = cval;
        }
        /* store output value: */
        switch (NI_NormalizeType(output->descr->type_num)) {
            CASE_INTERP_OUT(po, t, Bool);
            CASE_INTERP_OUT_UINT(po, t, UInt8, 0, MAX_UINT8);
            CASE_INTERP_OUT_UINT(po, t, UInt16, 0, MAX_UINT16);
            CASE_INTERP_OUT_UINT(po, t, UInt32, 0, MAX_UINT32);
#if HAS_UINT64
            /* There was a bug in numpy as of (at least) <= 1.6.1 such that
             * MAX_UINT64 was incorrectly defined, leading to a compiler error.
             * NPY_MAX_UINT64 is correctly defined
             */
            CASE_INTERP_OUT_UINT(po, t, UInt64, 0, NPY_MAX_UINT64);
#endif
            CASE_INTERP_OUT_INT(po, t, Int8, MIN_INT8, MAX_INT8);
            CASE_INTERP_OUT_INT(po, t, Int16, MIN_INT16, MAX_INT16);
            CASE_INTERP_OUT_INT(po, t, Int32, MIN_INT32, MAX_INT32);
            CASE_INTERP_OUT_INT(po, t, Int64, MIN_INT64, MAX_INT64);
            CASE_INTERP_OUT(po, t, Float32);
            CASE_INTERP_OUT(po, t, Float64);
        default:
            PyErr_SetString(PyExc_RuntimeError, "data type not supported");
            goto exit;
        }
        if (coordinates) {
            NI_ITERATOR_NEXT2(io, ic, po, pc);
        } else {
            NI_ITERATOR_NEXT(io, po);
        }
    }

 exit:
    if (edge_offsets)
        free(edge_offsets);
    if (data_offsets) {
        for(jj = 0; jj < irank; jj++)
            free(data_offsets[jj]);
        free(data_offsets);
    }
    if (splvals) {
        for(jj = 0; jj < irank; jj++)
            free(splvals[jj]);
        free(splvals);
    }
    if (foffsets)
        free(foffsets);
    if (fcoordinates)
        free(fcoordinates);
    return PyErr_Occurred() ? 0 : 1;
}
Пример #2
0
int NI_ZoomShift(PyArrayObject *input, PyArrayObject* zoom_ar,
                                 PyArrayObject* shift_ar, PyArrayObject *output,
                                 int order, int mode, double cval)
{
    char *po, *pi;
    npy_intp **zeros = NULL, **offsets = NULL, ***edge_offsets = NULL;
    npy_intp ftmp[MAXDIM], *fcoordinates = NULL, *foffsets = NULL;
    npy_intp jj, hh, kk, filter_size, odimensions[MAXDIM];
    npy_intp idimensions[MAXDIM], istrides[MAXDIM];
    npy_intp size;
    double ***splvals = NULL;
    NI_Iterator io;
    Float64 *zooms = zoom_ar ? (Float64*)PyArray_DATA(zoom_ar) : NULL;
    Float64 *shifts = shift_ar ? (Float64*)PyArray_DATA(shift_ar) : NULL;
    int rank = 0, qq;

    for(kk = 0; kk < input->nd; kk++) {
        idimensions[kk] = input->dimensions[kk];
        istrides[kk] = input->strides[kk];
        odimensions[kk] = output->dimensions[kk];
    }
    rank = input->nd;

    /* if the mode is 'constant' we need some temps later: */
    if (mode == NI_EXTEND_CONSTANT) {
        zeros = (npy_intp**)malloc(rank * sizeof(npy_intp*));
        if (!zeros) {
            PyErr_NoMemory();
            goto exit;
        }
        for(jj = 0; jj < rank; jj++)
            zeros[jj] = NULL;
        for(jj = 0; jj < rank; jj++) {
            zeros[jj] = (npy_intp*)malloc(odimensions[jj] * sizeof(npy_intp));
            if(!zeros[jj]) {
                PyErr_NoMemory();
                goto exit;
            }
        }
    }

    /* store offsets, along each axis: */
    offsets = (npy_intp**)malloc(rank * sizeof(npy_intp*));
    /* store spline coefficients, along each axis: */
    splvals = (double***)malloc(rank * sizeof(double**));
    /* store offsets at all edges: */
    edge_offsets = (npy_intp***)malloc(rank * sizeof(npy_intp**));
    if (!offsets || !splvals || !edge_offsets) {
        PyErr_NoMemory();
        goto exit;
    }
    for(jj = 0; jj < rank; jj++) {
        offsets[jj] = NULL;
        splvals[jj] = NULL;
        edge_offsets[jj] = NULL;
    }
    for(jj = 0; jj < rank; jj++) {
        offsets[jj] = (npy_intp*)malloc(odimensions[jj] * sizeof(npy_intp));
        splvals[jj] = (double**)malloc(odimensions[jj] * sizeof(double*));
        edge_offsets[jj] = (npy_intp**)malloc(odimensions[jj] * sizeof(npy_intp*));
        if (!offsets[jj] || !splvals[jj] || !edge_offsets[jj]) {
            PyErr_NoMemory();
            goto exit;
        }
        for(hh = 0; hh < odimensions[jj]; hh++) {
            splvals[jj][hh] = NULL;
            edge_offsets[jj][hh] = NULL;
        }
    }

    /* precalculate offsets, and offsets at the edge: */
    for(jj = 0; jj < rank; jj++) {
        double shift = 0.0, zoom = 0.0;
        if (shifts)
            shift = shifts[jj];
        if (zooms)
            zoom = zooms[jj];
        for(kk = 0; kk < odimensions[jj]; kk++) {
            double cc = (double)kk;
            if (shifts)
                cc += shift;
            if (zooms)
                cc *= zoom;
            cc = map_coordinate(cc, idimensions[jj], mode);
            if (cc > -1.0) {
                int start;
                if (zeros && zeros[jj])
                    zeros[jj][kk] = 0;
                if (order & 1) {
                    start = (int)floor(cc) - order / 2;
                } else {
                    start = (int)floor(cc + 0.5) - order / 2;
                }
                offsets[jj][kk] = istrides[jj] * start;
                if (start < 0 || start + order >= idimensions[jj]) {
                    edge_offsets[jj][kk] = (npy_intp*)malloc((order + 1) * sizeof(npy_intp));
                    if (!edge_offsets[jj][kk]) {
                        PyErr_NoMemory();
                        goto exit;
                    }
                    for(hh = 0; hh <= order; hh++) {
                        int idx = start + hh;
                         int len = idimensions[jj];
                        if (len <= 1) {
                            idx = 0;
                        } else {
                            int s2 = 2 * len - 2;
                            if (idx < 0) {
                                idx = s2 * (int)(-idx / s2) + idx;
                                idx = idx <= 1 - len ? idx + s2 : -idx;
                            } else if (idx >= len) {
                                idx -= s2 * (int)(idx / s2);
                                if (idx >= len)
                                    idx = s2 - idx;
                            }
                        }
                        edge_offsets[jj][kk][hh] = istrides[jj] * (idx - start);
                    }
                }
                if (order > 0) {
                    splvals[jj][kk] = (double*)malloc((order + 1) * sizeof(double));
                    if (!splvals[jj][kk]) {
                        PyErr_NoMemory();
                        goto exit;
                    }
                    spline_coefficients(cc, order, splvals[jj][kk]);
                }
            } else {
                zeros[jj][kk] = 1;
            }
        }
    }

    filter_size = 1;
    for(jj = 0; jj < rank; jj++)
        filter_size *= order + 1;

    if (!NI_InitPointIterator(output, &io))
        goto exit;

    pi = (void *)PyArray_DATA(input);
    po = (void *)PyArray_DATA(output);

    /* store all coordinates and offsets with filter: */
    fcoordinates = (npy_intp*)malloc(rank * filter_size * sizeof(npy_intp));
    foffsets = (npy_intp*)malloc(filter_size * sizeof(npy_intp));
    if (!fcoordinates || !foffsets) {
        PyErr_NoMemory();
        goto exit;
    }

    for(jj = 0; jj < rank; jj++)
        ftmp[jj] = 0;
    kk = 0;
    for(hh = 0; hh < filter_size; hh++) {
        for(jj = 0; jj < rank; jj++)
            fcoordinates[jj + hh * rank] = ftmp[jj];
        foffsets[hh] = kk;
        for(jj = rank - 1; jj >= 0; jj--) {
            if (ftmp[jj] < order) {
                ftmp[jj]++;
                kk += istrides[jj];
                break;
            } else {
                ftmp[jj] = 0;
                kk -= istrides[jj] * order;
            }
        }
    }
    size = 1;
    for(qq = 0; qq < output->nd; qq++)
        size *= output->dimensions[qq];
    for(kk = 0; kk < size; kk++) {
        double t = 0.0;
        int edge = 0, oo = 0, zero = 0;

        for(hh = 0; hh < rank; hh++) {
            if (zeros && zeros[hh][io.coordinates[hh]]) {
                /* we use constant border condition */
                zero = 1;
                break;
            }
            oo += offsets[hh][io.coordinates[hh]];
            if (edge_offsets[hh][io.coordinates[hh]])
                edge = 1;
        }

        if (!zero) {
            npy_intp *ff = fcoordinates;
            const int type_num = NI_NormalizeType(input->descr->type_num);
            t = 0.0;
            for(hh = 0; hh < filter_size; hh++) {
                int idx = 0;
                double coeff = 0.0;

                if (NI_UNLIKELY(edge)) {
                        /* use precalculated edge offsets: */
                    for(jj = 0; jj < rank; jj++) {
                        if (edge_offsets[jj][io.coordinates[jj]])
                            idx += edge_offsets[jj][io.coordinates[jj]][ff[jj]];
                        else
                            idx += ff[jj] * istrides[jj];
                    }
                    idx += oo;
                } else {
                    /* use normal offsets: */
                    idx += oo + foffsets[hh];
                }
                switch (type_num) {
                    CASE_INTERP_COEFF(coeff, pi, idx, Bool);
                    CASE_INTERP_COEFF(coeff, pi, idx, UInt8);
                    CASE_INTERP_COEFF(coeff, pi, idx, UInt16);
                    CASE_INTERP_COEFF(coeff, pi, idx, UInt32);
#if HAS_UINT64
                    CASE_INTERP_COEFF(coeff, pi, idx, UInt64);
#endif
                    CASE_INTERP_COEFF(coeff, pi, idx, Int8);
                    CASE_INTERP_COEFF(coeff, pi, idx, Int16);
                    CASE_INTERP_COEFF(coeff, pi, idx, Int32);
                    CASE_INTERP_COEFF(coeff, pi, idx, Int64);
                    CASE_INTERP_COEFF(coeff, pi, idx, Float32);
                    CASE_INTERP_COEFF(coeff, pi, idx, Float64);
                default:
                    PyErr_SetString(PyExc_RuntimeError,
                                                    "data type not supported");
                    goto exit;
                }
                /* calculate interpolated value: */
                for(jj = 0; jj < rank; jj++)
                    if (order > 0)
                        coeff *= splvals[jj][io.coordinates[jj]][ff[jj]];
                t += coeff;
                ff += rank;
            }
        } else {
            t = cval;
        }
        /* store output: */
        switch (NI_NormalizeType(output->descr->type_num)) {
            CASE_INTERP_OUT(po, t, Bool);
            CASE_INTERP_OUT_UINT(po, t, UInt8, 0, MAX_UINT8);
            CASE_INTERP_OUT_UINT(po, t, UInt16, 0, MAX_UINT16);
            CASE_INTERP_OUT_UINT(po, t, UInt32, 0, MAX_UINT32);
#if HAS_UINT64
            /* There was a bug in numpy as of (at least) <= 1.6.1 such that
             * MAX_UINT64 was incorrectly defined, leading to a compiler error.
             * NPY_MAX_UINT64 is correctly defined
             */
            CASE_INTERP_OUT_UINT(po, t, UInt64, 0, NPY_MAX_UINT64);
#endif
            CASE_INTERP_OUT_INT(po, t, Int8, MIN_INT8, MAX_INT8);
            CASE_INTERP_OUT_INT(po, t, Int16, MIN_INT16, MAX_INT16);
            CASE_INTERP_OUT_INT(po, t, Int32, MIN_INT32, MAX_INT32);
            CASE_INTERP_OUT_INT(po, t, Int64, MIN_INT64, MAX_INT64);
            CASE_INTERP_OUT(po, t, Float32);
            CASE_INTERP_OUT(po, t, Float64);
        default:
            PyErr_SetString(PyExc_RuntimeError, "data type not supported");
            goto exit;
        }
        NI_ITERATOR_NEXT(io, po);
    }

 exit:
    if (zeros) {
        for(jj = 0; jj < rank; jj++)
            if (zeros[jj])
                free(zeros[jj]);
        free(zeros);
    }
    if (offsets) {
        for(jj = 0; jj < rank; jj++)
            if (offsets[jj])
                free(offsets[jj]);
        free(offsets);
    }
    if (splvals) {
        for(jj = 0; jj < rank; jj++) {
            if (splvals[jj]) {
                for(hh = 0; hh < odimensions[jj]; hh++)
                    if (splvals[jj][hh])
                        free(splvals[jj][hh]);
                free(splvals[jj]);
            }
        }
        free(splvals);
    }
    if (edge_offsets) {
        for(jj = 0; jj < rank; jj++) {
            if (edge_offsets[jj]) {
                for(hh = 0; hh < odimensions[jj]; hh++)
                    if (edge_offsets[jj][hh])
                        free(edge_offsets[jj][hh]);
                free(edge_offsets[jj]);
            }
        }
        free(edge_offsets);
    }
    if (foffsets)
        free(foffsets);
    if (fcoordinates)
        free(fcoordinates);
    return PyErr_Occurred() ? 0 : 1;
}
Пример #3
0
int NI_WatershedIFT(PyArrayObject* input, PyArrayObject* markers,
                                        PyArrayObject* strct, PyArrayObject* output)
{
    char *pl, *pm, *pi;
    int ll;
    npy_intp size, jj, hh, kk, maxval;
    npy_intp strides[WS_MAXDIM], coordinates[WS_MAXDIM];
    npy_intp *nstrides = NULL, nneigh, ssize;
    int i_contiguous, o_contiguous;
    NI_WatershedElement *temp = NULL, **first = NULL, **last = NULL;
    Bool *ps = NULL;
    NI_Iterator mi, ii, li;
    NPY_BEGIN_THREADS_DEF;

    i_contiguous = PyArray_ISCONTIGUOUS(input);
    o_contiguous = PyArray_ISCONTIGUOUS(output);
    ssize = 1;
    for(ll = 0; ll < strct->nd; ll++)
        ssize *= strct->dimensions[ll];
    if (input->nd > WS_MAXDIM) {
        PyErr_SetString(PyExc_RuntimeError, "too many dimensions");
        goto exit;
    }
    size = 1;
    for(ll = 0; ll < input->nd; ll++)
        size *= input->dimensions[ll];
    /* Storage for the temporary queue data. */
    temp = (NI_WatershedElement*)malloc(size * sizeof(NI_WatershedElement));
    if (!temp) {
        PyErr_NoMemory();
        goto exit;
    }

    NPY_BEGIN_THREADS;

    pi = (void *)PyArray_DATA(input);
    if (!NI_InitPointIterator(input, &ii))
        goto exit;
    /* Initialization and find the maximum of the input. */
    maxval = 0;
    for(jj = 0; jj < size; jj++) {
        npy_intp ival = 0;
        switch(NI_NormalizeType(input->descr->type_num)) {
        CASE_GET_INPUT(ival, pi, UInt8);
        CASE_GET_INPUT(ival, pi, UInt16);
        default:
            NPY_END_THREADS;
            PyErr_SetString(PyExc_RuntimeError, "data type not supported");
            goto exit;
        }
        temp[jj].index = jj;
        temp[jj].done = 0;
        if (ival > maxval)
            maxval = ival;
        NI_ITERATOR_NEXT(ii, pi);
    }
    pi = (void *)PyArray_DATA(input);
    /* Allocate and initialize the storage for the queue. */
    first = (NI_WatershedElement**)malloc((maxval + 1) *
                                                                                sizeof(NI_WatershedElement*));
    last = (NI_WatershedElement**)malloc((maxval + 1) *
                                                                             sizeof(NI_WatershedElement*));
    if (NI_UNLIKELY(!first || !last)) {
        NPY_END_THREADS;
        PyErr_NoMemory();
        goto exit;
    }
    for(hh = 0; hh <= maxval; hh++) {
        first[hh] = NULL;
        last[hh] = NULL;
    }
    if (!NI_InitPointIterator(markers, &mi))
        goto exit;
    if (!NI_InitPointIterator(output, &li))
        goto exit;
    pm = (void *)PyArray_DATA(markers);
    pl = (void *)PyArray_DATA(output);
    /* initialize all nodes */
    for(ll = 0; ll < input->nd; ll++)
        coordinates[ll] = 0;
    for(jj = 0; jj < size; jj++) {
        /* get marker */
        int label = 0;
        switch(NI_NormalizeType(markers->descr->type_num)) {
        CASE_GET_LABEL(label, pm, UInt8);
        CASE_GET_LABEL(label, pm, UInt16);
        CASE_GET_LABEL(label, pm, UInt32);
#if HAS_UINT64
        CASE_GET_LABEL(label, pm, UInt64);
#endif
        CASE_GET_LABEL(label, pm, Int8);
        CASE_GET_LABEL(label, pm, Int16);
        CASE_GET_LABEL(label, pm, Int32);
        CASE_GET_LABEL(label, pm, Int64);
        default:
            NPY_END_THREADS;
            PyErr_SetString(PyExc_RuntimeError, "data type not supported");
            goto exit;
        }
        switch(NI_NormalizeType(output->descr->type_num)) {
        CASE_PUT_LABEL(label, pl, UInt8);
        CASE_PUT_LABEL(label, pl, UInt16);
        CASE_PUT_LABEL(label, pl, UInt32);
#if HAS_UINT64
        CASE_PUT_LABEL(label, pl, UInt64);
#endif
        CASE_PUT_LABEL(label, pl, Int8);
        CASE_PUT_LABEL(label, pl, Int16);
        CASE_PUT_LABEL(label, pl, Int32);
        CASE_PUT_LABEL(label, pl, Int64);
        default:
            NPY_END_THREADS;
            PyErr_SetString(PyExc_RuntimeError, "data type not supported");
            goto exit;
        }
        NI_ITERATOR_NEXT2(mi, li, pm, pl);
        if (label != 0) {
            /* This node is a marker */
            temp[jj].cost = 0;
            if (!first[0]) {
                first[0] = &(temp[jj]);
                first[0]->next = NULL;
                first[0]->prev = NULL;
                last[0] = first[0];
            } else {
                if (label > 0) {
                    /* object markers are enqueued at the beginning, so they
                       are processed first. */
                    temp[jj].next = first[0];
                    temp[jj].prev = NULL;
                    first[0]->prev = &(temp[jj]);
                    first[0] = &(temp[jj]);
                } else {
                    /* background markers are enqueued at the end, so they are
                         processed after the object markers. */
                    temp[jj].next = NULL;
                    temp[jj].prev = last[0];
                    last[0]->next = &(temp[jj]);
                    last[0] = &(temp[jj]);
                }
            }
        } else {
            /* This node is not a marker */
            temp[jj].cost = maxval + 1;
            temp[jj].next = NULL;
            temp[jj].prev = NULL;
        }
        for(ll = input->nd - 1; ll >= 0; ll--)
            if (coordinates[ll] < input->dimensions[ll] - 1) {
                coordinates[ll]++;
                break;
            } else {
                coordinates[ll] = 0;
            }
    }

    pl = (void *)PyArray_DATA(output);
    ps = (Bool*)PyArray_DATA(strct);
    nneigh = 0;
    for (kk = 0; kk < ssize; kk++)
        if (ps[kk] && kk != (ssize / 2))
            ++nneigh;
    nstrides = (npy_intp*)malloc(nneigh * sizeof(npy_intp));
    if (NI_UNLIKELY(!nstrides)) {
        NPY_END_THREADS;
        PyErr_NoMemory();
        goto exit;
    }
    strides[input->nd - 1] = 1;
    for(ll = input->nd - 2; ll >= 0; ll--)
        strides[ll] = input->dimensions[ll + 1] * strides[ll + 1];
    for(ll = 0; ll < input->nd; ll++)
        coordinates[ll] = -1;
    for(kk = 0; kk < nneigh; kk++)
        nstrides[kk] = 0;
    jj = 0;
    for(kk = 0; kk < ssize; kk++) {
        if (ps[kk]) {
            int offset = 0;
            for(ll = 0; ll < input->nd; ll++)
                offset += coordinates[ll] * strides[ll];
            if (offset != 0)
                nstrides[jj++] += offset;
        }
        for(ll = input->nd - 1; ll >= 0; ll--)
            if (coordinates[ll] < 1) {
                coordinates[ll]++;
                break;
            } else {
                coordinates[ll] = -1;
            }
    }
    /* Propagation phase: */
    for(jj = 0; jj <= maxval; jj++) {
        while (first[jj]) {
            /* dequeue first element: */
            NI_WatershedElement *v = first[jj];
            first[jj] = first[jj]->next;
            if (first[jj])
                first[jj]->prev = NULL;
            v->prev = NULL;
            v->next = NULL;
            /* Mark element as done: */
            v->done = 1;
            /* Iterate over the neighbors of the element: */
            for(hh = 0; hh < nneigh; hh++) {
                npy_intp v_index = v->index, p_index = v->index, idx, cc;
                int qq, outside = 0;
                p_index += nstrides[hh];
                /* check if the neighbor is within the extent of the array: */
                idx = p_index;
                for (qq = 0; qq < input->nd; qq++) {
                    cc = idx / strides[qq];
                    if (cc < 0 || cc >= input->dimensions[qq]) {
                        outside = 1;
                        break;
                    }
                    idx -= cc * strides[qq];
                }
                if (!outside) {
                    NI_WatershedElement *p = &(temp[p_index]);
                    if (!(p->done)) {
                        /* If the neighbor was not processed yet: */
                        int max, pval, vval, wvp, pcost, label, p_idx, v_idx;
                        switch(NI_NormalizeType(input->descr->type_num)) {
                        CASE_WINDEX1(v_index, p_index, strides, input->strides,
                                                 input->nd, i_contiguous, p_idx, v_idx, pi,
                                                 vval, pval, UInt8);
                        CASE_WINDEX1(v_index, p_index, strides, input->strides,
                                                 input->nd, i_contiguous, p_idx, v_idx, pi,
                                                 vval, pval, UInt16);
                        default:
                            NPY_END_THREADS;
                            PyErr_SetString(PyExc_RuntimeError,
                                                            "data type not supported");
                            goto exit;
                        }
                        /* Calculate cost: */
                        wvp = pval - vval;
                        if (wvp < 0)
                            wvp = -wvp;
                        /* Find the maximum of this cost and the current
                             element cost: */
                        pcost = p->cost;
                        max = v->cost > wvp ? v->cost : wvp;
                        if (max < pcost) {
                            /* If this maximum is less than the neighbors cost,
                                 adapt the cost and the label of the neighbor: */
                            int idx;
                            p->cost = max;
                            switch(NI_NormalizeType(output->descr->type_num)) {
                            CASE_WINDEX2(v_index, strides, output->strides, input->nd,
                                                     idx, o_contiguous, label, pl, UInt8);
                            CASE_WINDEX2(v_index, strides, output->strides, input->nd,
                                                     idx, o_contiguous, label, pl, UInt16);
                            CASE_WINDEX2(v_index, strides, output->strides, input->nd,
                                                     idx, o_contiguous, label, pl, UInt32);
#if HAS_UINT64
                            CASE_WINDEX2(v_index, strides, output->strides, input->nd,
                                                     idx, o_contiguous, label, pl, UInt64);
#endif
                            CASE_WINDEX2(v_index, strides, output->strides, input->nd,
                                                     idx, o_contiguous, label, pl, Int8);
                            CASE_WINDEX2(v_index, strides, output->strides, input->nd,
                                                     idx, o_contiguous, label, pl, Int16);
                            CASE_WINDEX2(v_index, strides, output->strides, input->nd,
                                                     idx, o_contiguous, label, pl, Int32);
                            CASE_WINDEX2(v_index, strides, output->strides, input->nd,
                                                     idx, o_contiguous, label, pl, Int64);
                            default:
                                NPY_END_THREADS;
                                PyErr_SetString(PyExc_RuntimeError,
                                                                "data type not supported");
                                goto exit;
                            }
                            switch(NI_NormalizeType(output->descr->type_num)) {
                            CASE_WINDEX3(p_index, strides, output->strides, input->nd,
                                                     idx, o_contiguous, label, pl, UInt8);
                            CASE_WINDEX3(p_index, strides, output->strides, input->nd,
                                                     idx, o_contiguous, label, pl, UInt16);
                            CASE_WINDEX3(p_index, strides, output->strides, input->nd,
                                                     idx, o_contiguous, label, pl, UInt32);
#if HAS_UINT64
                            CASE_WINDEX3(p_index, strides, output->strides, input->nd,
                                                     idx, o_contiguous, label, pl, UInt64);
#endif
                            CASE_WINDEX3(p_index, strides, output->strides, input->nd,
                                                     idx, o_contiguous, label, pl, Int8);
                            CASE_WINDEX3(p_index, strides, output->strides, input->nd,
                                                     idx, o_contiguous, label, pl, Int16);
                            CASE_WINDEX3(p_index, strides, output->strides, input->nd,
                                                     idx, o_contiguous, label, pl, Int32);
                            CASE_WINDEX3(p_index, strides, output->strides, input->nd,
                                                     idx, o_contiguous, label, pl, Int64);
                            default:
                                NPY_END_THREADS;
                                PyErr_SetString(PyExc_RuntimeError,
                                                                "data type not supported");
                                goto exit;
                            }
                            /* If the neighbor is in a queue, remove it: */
                            if (p->next || p->prev) {
                                NI_WatershedElement *prev = p->prev, *next = p->next;
                                if (first[pcost] == p)
                                    first[pcost] = next;
                                if (last[pcost] == p)
                                    last[pcost] = prev;
                                if (prev)
                                    prev->next = next;
                                if (next)
                                    next->prev = prev;
                            }
                            /* Insert the neighbor in the appropiate queue: */
                            if (label < 0) {
                                p->prev = last[max];
                                p->next = NULL;
                                if (last[max])
                                    last[max]->next = p;
                                last[max] = p;
                                if (!first[max])
                                    first[max] = p;
                            } else {
                                p->next = first[max];
                                p->prev = NULL;
                                if (first[max])
                                    first[max]->prev = p;
                                first[max] = p;
                                if (!last[max])
                                    last[max] = p;
                            }
                        }
                    }
                }
            }
        }
    }
 exit:
    NPY_END_THREADS;
    free(temp);
    free(first);
    free(last);
    free(nstrides);
    return PyErr_Occurred() ? 0 : 1;
}