static bool
paste_ (ImageBuf &dst, ROI dstroi,
        const ImageBuf &src, ROI srcroi, int nthreads)
{
    // N.B. Punt on parallelizing because of the subtle interplay
    // between srcroi and dstroi, the parallel_image idiom doesn't
    // handle that especially well. And it's not worth customizing for
    // this function which is inexpensive and not commonly used, and so
    // would benefit little from parallelizing. We can always revisit
    // this later. But in the mean time, we maintain the 'nthreads'
    // parameter for uniformity with the rest of IBA.
    int src_nchans = src.nchannels ();
    int dst_nchans = dst.nchannels ();
    ImageBuf::ConstIterator<S,D> s (src, srcroi);
    ImageBuf::Iterator<D,D> d (dst, dstroi);
    for ( ;  ! s.done();  ++s, ++d) {
        if (! d.exists())
            continue;  // Skip paste-into pixels that don't overlap dst's data
        for (int c = srcroi.chbegin, c_dst = dstroi.chbegin;
             c < srcroi.chend;  ++c, ++c_dst) {
            if (c_dst >= 0 && c_dst < dst_nchans)
                d[c_dst] = c < src_nchans ? s[c] : D(0);
        }
    }
    return true;
}

inline void
compare_value (ImageBuf::ConstIterator<BUFT,float> &a, int chan,
               float aval, float bval, ImageBufAlgo::CompareResults &result,
               float &maxval, double &batcherror, double &batch_sqrerror,
               bool &failed, bool &warned, float failthresh, float warnthresh)
{
    maxval = std::max (maxval, std::max (aval, bval));
    double f = fabs (aval - bval);
    batcherror += f;
    batch_sqrerror += f*f;
    if (f > result.maxerror) {
        result.maxerror = f;
        result.maxx = a.x();
        result.maxy = a.y();
        result.maxz = a.z();
        result.maxc = chan;
    }
    if (! warned && f > warnthresh) {
        ++result.nwarn;
        warned = true;
    }
    if (! failed && f > failthresh) {
        ++result.nfail;
        failed = true;
    }
}

static bool
circular_shift_ (ImageBuf &dst, const ImageBuf &src,
                 int xshift, int yshift, int zshift,
                 ROI dstroi, ROI roi, int nthreads)
{
    if (nthreads != 1 && roi.npixels() >= 1000) {
        // Possible multiple thread case -- recurse via parallel_image
        ImageBufAlgo::parallel_image (
            boost::bind(circular_shift_<DSTTYPE,SRCTYPE>,
                        boost::ref(dst), boost::cref(src),
                        xshift, yshift, zshift,
                        dstroi, _1 /*roi*/, 1 /*nthreads*/),
            roi, nthreads);
        return true;
    }

    // Serial case
    int width = dstroi.width(), height = dstroi.height(), depth = dstroi.depth();
    ImageBuf::ConstIterator<SRCTYPE,DSTTYPE> s (src, roi);
    ImageBuf::Iterator<DSTTYPE,DSTTYPE> d (dst);
    for (  ;  ! s.done();  ++s) {
        int dx = s.x() + xshift;  OIIO::wrap_periodic (dx, dstroi.xbegin, width);
        int dy = s.y() + yshift;  OIIO::wrap_periodic (dy, dstroi.ybegin, height);
        int dz = s.z() + zshift;  OIIO::wrap_periodic (dz, dstroi.zbegin, depth);
        d.pos (dx, dy, dz);
        if (! d.exists())
            continue;
        for (int c = roi.chbegin;  c < roi.chend;  ++c)
            d[c] = s[c];
    }
    return true;
}

static inline bool
isConstantColor_ (const ImageBuf &src, float *color,
                  ROI roi, int nthreads)
{
    // Iterate using the native typing (for speed).
    std::vector<T> constval (roi.nchannels());
    ImageBuf::ConstIterator<T,T> s (src, roi);
    for (int c = roi.chbegin;  c < roi.chend;  ++c)
        constval[c] = s[c];

    // Loop over all pixels ...
    for ( ; ! s.done();  ++s) {
        for (int c = roi.chbegin;  c < roi.chend;  ++c)
            if (constval[c] != s[c])
                return false;
    }
    
    if (color) {
        ImageBuf::ConstIterator<T,float> s (src, roi);
        for (int c = 0;  c < roi.chbegin; ++c)
            color[c] = 0.0f;
        for (int c = roi.chbegin; c < roi.chend; ++c)
            color[c] = s[c];
        for (int c = roi.chend;  c < src.nchannels(); ++c)
            color[c] = 0.0f;
    }

    return true;
}

static int
action_flip (int argc, const char *argv[])
{
    if (ot.postpone_callback (1, action_flip, argc, argv))
        return 0;

    ot.read ();
    ImageRecRef A = ot.pop();
    ot.push (new ImageRec (*A, ot.allsubimages ? -1 : 0,
                           ot.allsubimages ? -1 : 0, true, false));

    int subimages = ot.curimg->subimages();
    for (int s = 0;  s < subimages;  ++s) {
        int miplevels = ot.curimg->miplevels(s);
        for (int m = 0;  m < miplevels;  ++m) {
            const ImageBuf &Aib ((*A)(s,m));
            ImageBuf &Rib ((*ot.curimg)(s,m));
            ImageBuf::ConstIterator<float> a (Aib);
            ImageBuf::Iterator<float> r (Rib);
            int nchans = Rib.nchannels();
            int firstscanline = Rib.ymin();
            int lastscanline = Rib.ymax();
            for ( ; ! r.done(); ++r) {
                a.pos (r.x(), lastscanline - (r.y() - firstscanline));
                for (int c = 0;  c < nchans;  ++c)
                    r[c] = a[c];
            }
        }
    }
             
    return 0;
}

static inline void 
copy_pixels_ (const ImageBuf &buf, int xbegin, int xend,
              int ybegin, int yend, D *r)
{
    int w = (xend-xbegin);
    for (ImageBuf::ConstIterator<S,D> p (buf, xbegin, xend, ybegin, yend);
         p.valid(); ++p) { 
        imagesize_t offset = ((p.y()-ybegin)*w + (p.x()-xbegin)) * buf.nchannels();
        for (int c = 0;  c < buf.nchannels();  ++c)
            r[offset+c] = p[c];
    }
}

static inline void
getpixel_ (const ImageBuf &buf, int x, int y, int z, float *result, int chans)
{
    ImageBuf::ConstIterator<T> pixel (buf, x, y, z);
    if (pixel.valid()) {
        for (int i = 0;  i < chans;  ++i)
            result[i] = pixel[i];
    } else {
        for (int i = 0;  i < chans;  ++i)
            result[i] = 0.0f;
    }
}

static bool
rotate270_ (ImageBuf &dst, const ImageBuf &src, ROI dst_roi, int nthreads)
{
    ROI dst_roi_full = dst.roi_full();
    ImageBuf::ConstIterator<S, D> s (src);
    ImageBuf::Iterator<D, D> d (dst, dst_roi);
    for ( ; ! d.done(); ++d) {
        s.pos (dst_roi_full.yend - d.y() - 1,
               d.x(),
               d.z());
        for (int c = dst_roi.chbegin; c < dst_roi.chend; ++c)
            d[c] = s[c];
    }
    return true;
}

static bool
compare_ (const ImageBuf &A, const ImageBuf &B,
          float failthresh, float warnthresh,
          ImageBufAlgo::CompareResults &result,
          ROI roi, int nthreads)
{
    imagesize_t npels = roi.npixels();
    imagesize_t nvals = npels * roi.nchannels();
    int Achannels = A.nchannels(), Bchannels = B.nchannels();

    // Compare the two images.
    //
    double totalerror = 0;
    double totalsqrerror = 0;
    result.maxerror = 0;
    result.maxx=0, result.maxy=0, result.maxz=0, result.maxc=0;
    result.nfail = 0, result.nwarn = 0;
    float maxval = 1.0;  // max possible value

    ImageBuf::ConstIterator<Atype> a (A, roi, ImageBuf::WrapBlack);
    ImageBuf::ConstIterator<Btype> b (B, roi, ImageBuf::WrapBlack);
    bool deep = A.deep();
    // Break up into batches to reduce cancelation errors as the error
    // sums become too much larger than the error for individual pixels.
    const int batchsize = 4096;   // As good a guess as any
    for ( ;  ! a.done();  ) {
        double batcherror = 0;
        double batch_sqrerror = 0;
        if (deep) {
            for (int i = 0;  i < batchsize && !a.done();  ++i, ++a, ++b) {
                bool warned = false, failed = false;  // For this pixel
                for (int c = roi.chbegin;  c < roi.chend;  ++c)
                    for (int s = 0, e = a.deep_samples(); s < e;  ++s) {
                        compare_value (a, c, a.deep_value(c,s),
                                       b.deep_value(c,s), result, maxval,
                                       batcherror, batch_sqrerror,
                                       failed, warned, failthresh, warnthresh);
                    }
            }
        } else {  // non-deep
            for (int i = 0;  i < batchsize && !a.done();  ++i, ++a, ++b) {
                bool warned = false, failed = false;  // For this pixel
                for (int c = roi.chbegin;  c < roi.chend;  ++c)
                    compare_value (a, c, c < Achannels ? a[c] : 0.0f,
                                   c < Bchannels ? b[c] : 0.0f,
                                   result, maxval, batcherror, batch_sqrerror,
                                   failed, warned, failthresh, warnthresh);
            }
        }
        totalerror += batcherror;
        totalsqrerror += batch_sqrerror;
    }
    result.meanerror = totalerror / nvals;
    result.rms_error = sqrt (totalsqrerror / nvals);
    result.PSNR = 20.0 * log10 (maxval / result.rms_error);
    return result.nfail == 0;
}

static inline void 
get_pixel_channels_ (const ImageBuf &buf, int xbegin, int xend,
                     int ybegin, int yend, int zbegin, int zend, 
                     int chbegin, int chend, D *r,
                     stride_t xstride, stride_t ystride, stride_t zstride)
{
    int w = (xend-xbegin), h = (yend-ybegin);
    int nchans = chend - chbegin;
    ImageSpec::auto_stride (xstride, ystride, zstride, sizeof(D), nchans, w, h);
    for (ImageBuf::ConstIterator<S,D> p (buf, xbegin, xend, ybegin, yend, zbegin, zend);
         !p.done(); ++p) {
        imagesize_t offset = (p.z()-zbegin)*zstride + (p.y()-ybegin)*ystride
                           + (p.x()-xbegin)*xstride;
        D *rc = (D *)((char *)r + offset);
        for (int c = 0;  c < nchans;  ++c)
            rc[c] = p[c+chbegin];
    }
}

static int
action_sub (int argc, const char *argv[])
{
    if (ot.postpone_callback (2, action_sub, argc, argv))
        return 0;

    ImageRecRef B (ot.pop());
    ImageRecRef A (ot.pop());
    ot.read (A);
    ot.read (B);
    ot.push (new ImageRec (*A, ot.allsubimages ? -1 : 0,
                           ot.allsubimages ? -1 : 0, true, false));

    int subimages = ot.curimg->subimages();
    for (int s = 0;  s < subimages;  ++s) {
        int miplevels = ot.curimg->miplevels(s);
        for (int m = 0;  m < miplevels;  ++m) {
            const ImageBuf &Aib ((*A)(s,m));
            const ImageBuf &Bib ((*B)(s,m));
            if (! same_size (Aib, Bib)) {
                // FIXME: some day, there should be options of combining
                // differing images somehow.
                std::cerr << "oiiotool: " << argv[0] << " could not combine images of differing sizes\n";
                continue;
            }
            ImageBuf &Rib ((*ot.curimg)(s,m));
            ImageBuf::ConstIterator<float> a (Aib);
            ImageBuf::ConstIterator<float> b (Bib);
            ImageBuf::Iterator<float> r (Rib);
            int nchans = Rib.nchannels();
            for ( ; ! r.done(); ++r) {
                a.pos (r.x(), r.y());
                b.pos (r.x(), r.y());
                for (int c = 0;  c < nchans;  ++c)
                    r[c] = a[c] - b[c];
            }
        }
    }
             
    return 0;
}

static bool
histogram_impl (const ImageBuf &A, int channel,
                std::vector<imagesize_t> &histogram, int bins,
                float min, float max, imagesize_t *submin,
                imagesize_t *supermax, ROI roi)
{
    // Double check A's type.
    if (A.spec().format != BaseTypeFromC<Atype>::value) {
        A.error ("Unsupported pixel data format '%s'", A.spec().format);
        return false;
    }

    // Initialize.
    ImageBuf::ConstIterator<Atype, float> a (A, roi);
    float ratio = bins / (max-min);
    int bins_minus_1 = bins-1;
    bool submin_ok = submin != NULL;
    bool supermax_ok = supermax != NULL;
    if (submin_ok)
        *submin = 0;
    if (supermax_ok)
        *supermax = 0;
    histogram.assign(bins, 0);

    // Compute histogram.
    for ( ; ! a.done(); a++) {
        float c = a[channel];
        if (c >= min && c < max) {
            // Map range min->max to 0->(bins-1).
            histogram[ (int) ((c-min) * ratio) ]++;
        } else if (c == max) {
            histogram[bins_minus_1]++;
        } else {
            if (submin_ok && c < min)
                (*submin)++;
            else if (supermax_ok)
                (*supermax)++;
        }
    }
    return true;
}

static bool
convolve_ (ImageBuf &dst, const ImageBuf &src, const ImageBuf &kernel,
           bool normalize, ROI roi, int nthreads)
{
    if (nthreads != 1 && roi.npixels() >= 1000) {
        // Lots of pixels and request for multi threads? Parallelize.
        ImageBufAlgo::parallel_image (
            boost::bind(convolve_<DSTTYPE,SRCTYPE>, boost::ref(dst),
                        boost::cref(src), boost::cref(kernel), normalize,
                        _1 /*roi*/, 1 /*nthreads*/),
            roi, nthreads);
        return true;
    }

    // Serial case

    float scale = 1.0f;
    if (normalize) {
        scale = 0.0f;
        for (ImageBuf::ConstIterator<float> k (kernel); ! k.done(); ++k)
            scale += k[0];
        scale = 1.0f / scale;
    }

    float *sum = ALLOCA (float, roi.chend);
    ROI kroi = get_roi (kernel.spec());
    ImageBuf::Iterator<DSTTYPE> d (dst, roi);
    ImageBuf::ConstIterator<SRCTYPE> s (src, roi, ImageBuf::WrapClamp);
    for ( ; ! d.done();  ++d) {

        for (int c = roi.chbegin; c < roi.chend; ++c)
            sum[c] = 0.0f;

        for (ImageBuf::ConstIterator<float> k (kernel, kroi); !k.done(); ++k) {
            float kval = k[0];
            s.pos (d.x() + k.x(), d.y() + k.y(), d.z() + k.z());
            for (int c = roi.chbegin; c < roi.chend; ++c)
                sum[c] += kval * s[c];
        }
        
        for (int c = roi.chbegin; c < roi.chend; ++c)
            d[c] = scale * sum[c];
    }

    return true;
}

static bool
color_count_ (const ImageBuf &src, atomic_ll *count,
              int ncolors, const float *color, const float *eps,
              ROI roi, int nthreads)
{
    if (nthreads != 1 && roi.npixels() >= 1000) {
        // Lots of pixels and request for multi threads? Parallelize.
        ImageBufAlgo::parallel_image (
            boost::bind(color_count_<T>, boost::ref(src),
                        count, ncolors, color, eps,
                        _1 /*roi*/, 1 /*nthreads*/),
            roi, nthreads);
        return true;
    }

    // Serial case
    int nchannels = src.nchannels();
    long long *n = ALLOCA (long long, ncolors);
    for (int col = 0;  col < ncolors;  ++col)
        n[col] = 0;
    for (ImageBuf::ConstIterator<T> p (src, roi);  !p.done();  ++p) {
        int coloffset = 0;
        for (int col = 0;  col < ncolors;  ++col, coloffset += nchannels) {
            int match = 1;
            for (int c = roi.chbegin;  c < roi.chend;  ++c) {
                if (fabsf(p[c] - color[coloffset+c]) > eps[c]) {
                    match = 0;
                    break;
                }
            }
            n[col] += match;
        }
    }

    for (int col = 0;  col < ncolors;  ++col)
        count[col] += n[col];
    return true;
}

inline void
compare_value (ImageBuf::ConstIterator<BUFT,float> &a, int chan,
               float aval, float bval, ImageBufAlgo::CompareResults &result,
               float &maxval, double &batcherror, double &batch_sqrerror,
               bool &failed, bool &warned, float failthresh, float warnthresh)
{
    if (!isfinite(aval) || !isfinite(bval)) {
        if (isnan(aval) == isnan(bval) && isinf(aval) == isinf(bval))
            return; // NaN may match NaN, Inf may match Inf
        if (isfinite(result.maxerror)) {
            // non-finite errors trump finite ones
            result.maxerror = std::numeric_limits<float>::infinity();
            result.maxx = a.x();
            result.maxy = a.y();
            result.maxz = a.z();
            result.maxc = chan;
            return;
        }
    }
    maxval = std::max (maxval, std::max (aval, bval));
    double f = fabs (aval - bval);
    batcherror += f;
    batch_sqrerror += f*f;
    // We use the awkward '!(a<=threshold)' construct so that we have
    // failures when f is a NaN (since all comparisons involving NaN will
    // return false).
    if (!(f <= result.maxerror)) {
        result.maxerror = f;
        result.maxx = a.x();
        result.maxy = a.y();
        result.maxz = a.z();
        result.maxc = chan;
    }
    if (! warned && !(f <= warnthresh)) {
        ++result.nwarn;
        warned = true;
    }
    if (! failed && !(f <= failthresh)) {
        ++result.nfail;
        failed = true;
    }
}

// DEPRECATED version
bool
ImageBufAlgo::add (ImageBuf &dst, const ImageBuf &A, const ImageBuf &B,
                   int options)
{
    // Sanity checks
    
    // dst must be distinct from A and B
    if ((const void *)&A == (const void *)&dst ||
        (const void *)&B == (const void *)&dst) {
        dst.error ("destination image must be distinct from source");
        return false;
    }
    
    // all three images must have the same number of channels
    if (A.spec().nchannels != B.spec().nchannels) {
        dst.error ("channel number mismatch: %d vs. %d", 
                   A.spec().nchannels, B.spec().nchannels);
        return false;
    }
    
    // If dst has not already been allocated, set it to the right size,
    // make it unconditinally float
    if (! dst.pixels_valid()) {
        ImageSpec dstspec = A.spec();
        dstspec.set_format (TypeDesc::TypeFloat);
        dst.alloc (dstspec);
    }
    // Clear dst pixels if instructed to do so
    if (options & ADD_CLEAR_DST) {
        zero (dst);
    }
      
    ASSERT (A.spec().format == TypeDesc::FLOAT &&
            B.spec().format == TypeDesc::FLOAT &&
            dst.spec().format == TypeDesc::FLOAT);
    
    ImageBuf::ConstIterator<float,float> a (A);
    ImageBuf::ConstIterator<float,float> b (B);
    ImageBuf::Iterator<float> d (dst);
    int nchannels = A.nchannels();
    // Loop over all pixels in A
    for ( ; a.valid();  ++a) {  
        // Point the iterators for B and dst to the corresponding pixel
        if (options & ADD_RETAIN_WINDOWS) {
            b.pos (a.x(), a.y());
        } else {
            // ADD_ALIGN_WINDOWS: make B line up with A
            b.pos (a.x()-A.xbegin()+B.xbegin(), a.y()-A.ybegin()+B.ybegin());
        }
        d.pos (a.x(), b.y());
        
        if (! b.valid() || ! d.valid())
            continue;   // Skip pixels that don't align
        
        // Add the pixel
        for (int c = 0;  c < nchannels;  ++c)
              d[c] = a[c] + b[c];
    }
    
    return true;
}

static bool
colorconvert_impl (ImageBuf &R, const ImageBuf &A,
                   const ColorProcessor* processor, bool unpremult,
                   ROI roi, int nthreads)
{
    if (nthreads != 1 && roi.npixels() >= 1000) {
        // Possible multiple thread case -- recurse via parallel_image
        ImageBufAlgo::parallel_image (
            OIIO::bind(colorconvert_impl<Rtype,Atype>,
                        OIIO::ref(R), OIIO::cref(A), processor, unpremult,
                        _1 /*roi*/, 1 /*nthreads*/),
            roi, nthreads);
        return true;
    }

    // Serial case

    int width = roi.width();
    // Temporary space to hold one RGBA scanline
    std::vector<float> scanline(width*4, 0.0f);
    
    // Only process up to, and including, the first 4 channels.  This
    // does let us process images with fewer than 4 channels, which is
    // the intent.
    // FIXME: Instead of loading the first 4 channels, obey
    //        Rspec.alpha_channel index (but first validate that the
    //        index is set properly for normal formats)
    
    int channelsToCopy = std::min (4, roi.nchannels());
    
    // Walk through all data in our buffer. (i.e., crop or overscan)
    // FIXME: What about the display window?  Should this actually promote
    // the datawindow to be union of data + display? This is useful if
    // the color of black moves.  (In which case non-zero sections should
    // now be promoted).  Consider the lin->log of a roto element, where
    // black now moves to non-black.
    
    float * dstPtr = NULL;
    const float fltmin = std::numeric_limits<float>::min();
    
    // If the processor has crosstalk, and we'll be using it, we should
    // reset the channels to 0 before loading each scanline.
    bool clearScanline = (channelsToCopy<4 && 
                          (processor->hasChannelCrosstalk() || unpremult));
    
    ImageBuf::ConstIterator<Atype> a (A, roi);
    ImageBuf::Iterator<Rtype> r (R, roi);
    for (int k = roi.zbegin; k < roi.zend; ++k) {
        for (int j = roi.ybegin; j < roi.yend; ++j) {
            // Clear the scanline
            if (clearScanline)
                memset (&scanline[0], 0, sizeof(float)*scanline.size());
            
            // Load the scanline
            dstPtr = &scanline[0];
            a.rerange (roi.xbegin, roi.xend, j, j+1, k, k+1);
            for ( ; !a.done(); ++a, dstPtr += 4)
                for (int c = 0; c < channelsToCopy; ++c)
                    dstPtr[c] = a[c];

            // Optionally unpremult
            if ((channelsToCopy >= 4) && unpremult) {
                for (int i = 0; i < width; ++i) {
                    float alpha = scanline[4*i+3];
                    if (alpha > fltmin) {
                        scanline[4*i+0] /= alpha;
                        scanline[4*i+1] /= alpha;
                        scanline[4*i+2] /= alpha;
                    }
                }
            }
            
            // Apply the color transformation in place
            processor->apply (&scanline[0], width, 1, 4,
                              sizeof(float), 4*sizeof(float),
                              width*4*sizeof(float));
            
            // Optionally premult
            if ((channelsToCopy >= 4) && unpremult) {
                for (int i = 0; i < width; ++i) {
                    float alpha = scanline[4*i+3];
                    if (alpha > fltmin) {
                        scanline[4*i+0] *= alpha;
                        scanline[4*i+1] *= alpha;
                        scanline[4*i+2] *= alpha;
                    }
                }
            }

            // Store the scanline
            dstPtr = &scanline[0];
            r.rerange (roi.xbegin, roi.xend, j, j+1, k, k+1);
            for ( ; !r.done(); ++r, dstPtr += 4)
                for (int c = 0; c < channelsToCopy; ++c)
                    r[c] = dstPtr[c];
        }
    }
    return true;
}

//.........这里部分代码省略.........
        // If using separable filters, our vertical set of filter tap
        // weights will be the same for the whole scanline we're on.  Just
        // compute and normalize them once.
        float totalweight_y = 0.0f;
        if (separable) {
            for (int j = 0;  j < ytaps;  ++j) {
                float w = filter->yfilt (yratio * (j-radj-(src_yf_frac-0.5f)));
                yfiltval[j] = w;
                totalweight_y += w;
            }
            for (int i = 0;  i <= ytaps;  ++i)
                yfiltval[i] /= totalweight_y;
        }

        for (int x = roi.xbegin;  x < roi.xend;  ++x) {
            float s = (x-dstfx+0.5f)*dstpixelwidth;
            float src_xf = srcfx + s * srcfw;
            int src_x;
            float src_xf_frac = floorfrac (src_xf, &src_x);
            for (int c = 0;  c < nchannels;  ++c)
                pel[c] = 0.0f;
            if (separable) {
                // Cache and normalize the horizontal filter tap weights
                // just once for this (x,y) position, reuse for all vertical
                // taps.
                float totalweight_x = 0.0f;
                for (int i = 0;  i < xtaps;  ++i) {
                    float w = filter->xfilt (xratio * (i-radi-(src_xf_frac-0.5f)));
                    xfiltval[i] = w;
                    totalweight_x += w;
                }

                if (totalweight_x != 0.0f) {
                    for (int i = 0;  i < xtaps;  ++i)  // normalize x filter
                        xfiltval[i] /= totalweight_x;  // weights
                    ImageBuf::ConstIterator<SRCTYPE> srcpel (src, src_x-radi, src_x+radi+1,
                                                             src_y-radj, src_y+radj+1,
                                                             0, 1, ImageBuf::WrapClamp);
                    for (int j = -radj;  j <= radj;  ++j) {
                        float wy = yfiltval[j+radj];
                        if (wy == 0.0f) {
                            // 0 weight for this y tap -- move to next line
                            srcpel.pos (srcpel.x(), srcpel.y()+1, srcpel.z());
                            continue;
                        }
                        for (int i = 0;  i < xtaps; ++i, ++srcpel) {
                            float w = wy * xfiltval[i];
                            for (int c = 0;  c < nchannels;  ++c)
                                pel[c] += w * srcpel[c];
                        }
                    }
                }
                // Copy the pixel value (already normalized) to the output.
                DASSERT (out.x() == x && out.y() == y);
                if (totalweight_y == 0.0f) {
                    // zero it out
                    for (int c = 0;  c < nchannels;  ++c)
                        out[c] = 0.0f;
                } else {
                    for (int c = 0;  c < nchannels;  ++c)
                        out[c] = pel[c];
                }
            } else {
                // Non-separable
                float totalweight = 0.0f;
                ImageBuf::ConstIterator<SRCTYPE> srcpel (src, src_x-radi, src_x+radi+1,
                                                       src_y-radi, src_y+radi+1,
                                                       0, 1, ImageBuf::WrapClamp);
                for (int j = -radj;  j <= radj;  ++j) {
                    for (int i = -radi;  i <= radi;  ++i, ++srcpel) {
                        float w = (*filter)(xratio * (i-(src_xf_frac-0.5f)),
                                            yratio * (j-(src_yf_frac-0.5f)));
                        totalweight += w;
                        if (w == 0.0f)
                            continue;
                        DASSERT (! srcpel.done());
                        for (int c = 0;  c < nchannels;  ++c)
                            pel[c] += w * srcpel[c];
                    }
                }
                DASSERT (srcpel.done());
                // Rescale pel to normalize the filter and write it to the
                // output image.
                DASSERT (out.x() == x && out.y() == y);
                if (totalweight == 0.0f) {
                    // zero it out
                    for (int c = 0;  c < nchannels;  ++c)
                        out[c] = 0.0f;
                } else {
                    for (int c = 0;  c < nchannels;  ++c)
                        out[c] = pel[c] / totalweight;
                }
            }

            ++out;
        }
    }

    return true;
}

void 
IvImage::pixel_transform(bool srgb_to_linear, int color_mode, int select_channel)
{
    /// This table obeys the following function:
    ///
    ///   unsigned char srgb2linear(unsigned char x)
    ///   {
    ///       float x_f = x/255.0;
    ///       float x_l = 0.0;
    ///       if (x_f <= 0.04045)
    ///           x_l = x_f/12.92;
    ///       else
    ///           x_l = powf((x_f+0.055)/1.055,2.4);
    ///       return (unsigned char)(x_l * 255 + 0.5)
    ///   }
    /// 
    ///  It's used to transform from sRGB color space to linear color space.
    static const unsigned char srgb_to_linear_lut[256] = {
        0, 0, 0, 0, 0, 0, 0, 1,
        1, 1, 1, 1, 1, 1, 1, 1,
        1, 1, 2, 2, 2, 2, 2, 2,
        2, 2, 3, 3, 3, 3, 3, 3,
        4, 4, 4, 4, 4, 5, 5, 5,
        5, 6, 6, 6, 6, 7, 7, 7,
        8, 8, 8, 8, 9, 9, 9, 10,
        10, 10, 11, 11, 12, 12, 12, 13,
        13, 13, 14, 14, 15, 15, 16, 16,
        17, 17, 17, 18, 18, 19, 19, 20,
        20, 21, 22, 22, 23, 23, 24, 24,
        25, 25, 26, 27, 27, 28, 29, 29,
        30, 30, 31, 32, 32, 33, 34, 35,
        35, 36, 37, 37, 38, 39, 40, 41,
        41, 42, 43, 44, 45, 45, 46, 47,
        48, 49, 50, 51, 51, 52, 53, 54,
        55, 56, 57, 58, 59, 60, 61, 62,
        63, 64, 65, 66, 67, 68, 69, 70,
        71, 72, 73, 74, 76, 77, 78, 79,
        80, 81, 82, 84, 85, 86, 87, 88,
        90, 91, 92, 93, 95, 96, 97, 99,
        100, 101, 103, 104, 105, 107, 108, 109,
        111, 112, 114, 115, 116, 118, 119, 121,
        122, 124, 125, 127, 128, 130, 131, 133,
        134, 136, 138, 139, 141, 142, 144, 146,
        147, 149, 151, 152, 154, 156, 157, 159,
        161, 163, 164, 166, 168, 170, 171, 173,
        175, 177, 179, 181, 183, 184, 186, 188,
        190, 192, 194, 196, 198, 200, 202, 204,
        206, 208, 210, 212, 214, 216, 218, 220,
        222, 224, 226, 229, 231, 233, 235, 237,
        239, 242, 244, 246, 248, 250, 253, 255
    };
    unsigned char correction_table[256];
    int total_channels = spec().nchannels;
    int color_channels = spec().nchannels;
    int max_channels = m_corrected_image.nchannels();

    // FIXME: Now with the iterator and data proxy in place, it should be
    // trivial to apply the transformations to any kind of data, not just
    // UINT8.
    if (spec().format != TypeDesc::UINT8 || ! m_corrected_image.localpixels()) {
        return;
    }

    if (color_channels > 3) {
        color_channels = 3;
    } else if (color_channels == 2) {
        color_channels = 1;
    }

    // This image is Luminance or Luminance + Alpha, and we are asked to show
    // luminance.
    if (color_channels == 1 && color_mode == 3) {
        color_mode = 0; // Just copy as usual.
    }

    // Happy path:
    if (! srgb_to_linear && color_mode <= 1 && m_gamma == 1.0 && m_exposure == 0.0) {
        ImageBuf::ConstIterator<unsigned char, unsigned char> src (*this);
        ImageBuf::Iterator<unsigned char, unsigned char> dst (m_corrected_image);
        for ( ; src.valid (); ++src) {
            dst.pos (src.x(), src.y());
            for (int i = 0; i < max_channels; i++)
                dst[i] = src[i];
        }
        return;
    }

    // fill the correction_table
    if (gamma() == 1.0 && exposure() == 0.0) {
        for (int pixelvalue = 0; pixelvalue < 256; ++pixelvalue) {
            correction_table[pixelvalue] = pixelvalue;
        }
    } else {
        float inv_gamma = 1.0/gamma();
        float gain = powf (2.0f, exposure());
        for (int pixelvalue = 0; pixelvalue < 256; ++pixelvalue) {
            float pv_f = converter (pixelvalue);
            pv_f = clamp (calc_exposure (pv_f, gain, inv_gamma),
                          0.0f, 1.0f);
            correction_table[pixelvalue] = (unsigned char) (pv_f*255 + 0.5);
//.........这里部分代码省略.........

static bool
resample_ (ImageBuf &dst, const ImageBuf &src, bool interpolate,
           ROI roi, int nthreads)
{
    if (nthreads != 1 && roi.npixels() >= 1000) {
        // Lots of pixels and request for multi threads? Parallelize.
        ImageBufAlgo::parallel_image (
            boost::bind(resample_<DSTTYPE,SRCTYPE>, boost::ref(dst),
                        boost::cref(src), interpolate,
                        _1 /*roi*/, 1 /*nthreads*/),
            roi, nthreads);
        return true;
    }

    // Serial case

    const ImageSpec &srcspec (src.spec());
    const ImageSpec &dstspec (dst.spec());
    int nchannels = src.nchannels();

    // Local copies of the source image window, converted to float
    float srcfx = srcspec.full_x;
    float srcfy = srcspec.full_y;
    float srcfw = srcspec.full_width;
    float srcfh = srcspec.full_height;

    float dstfx = dstspec.full_x;
    float dstfy = dstspec.full_y;
    float dstfw = dstspec.full_width;
    float dstfh = dstspec.full_height;
    float dstpixelwidth = 1.0f / dstfw;
    float dstpixelheight = 1.0f / dstfh;
    float *pel = ALLOCA (float, nchannels);

    ImageBuf::Iterator<DSTTYPE> out (dst, roi);
    ImageBuf::ConstIterator<SRCTYPE> srcpel (src);
    for (int y = roi.ybegin;  y < roi.yend;  ++y) {
        // s,t are NDC space
        float t = (y-dstfy+0.5f)*dstpixelheight;
        // src_xf, src_xf are image space float coordinates
        float src_yf = srcfy + t * srcfh - 0.5f;
        // src_x, src_y are image space integer coordinates of the floor
        int src_y;
        (void) floorfrac (src_yf, &src_y);
        for (int x = roi.xbegin;  x < roi.xend;  ++x) {
            float s = (x-dstfx+0.5f)*dstpixelwidth;
            float src_xf = srcfx + s * srcfw - 0.5f;
            int src_x;
            (void) floorfrac (src_xf, &src_x);

            if (interpolate) {
                src.interppixel (src_xf, src_yf, pel);
                for (int c = roi.chbegin; c < roi.chend; ++c)
                    out[c] = pel[c];
            } else {
                srcpel.pos (src_x, src_y, 0);
                for (int c = roi.chbegin; c < roi.chend; ++c)
                    out[c] = srcpel[c];
            }
            ++out;
        }
    }

    return true;
}