colormapping.c

Go to the documentation of this file.

/*
    This file is part of darktable,
    Copyright (C) 2013-2016 Roman Lebedev.
    Copyright (C) 2013-2017, 2019 Tobias Ellinghaus.
    Copyright (C) 2013-2014, 2016-2017 Ulrich Pegelow.
    Copyright (C) 2014 parafin.
    Copyright (C) 2015 Pedro Côrte-Real.
    Copyright (C) 2016 johannes hanika.
    Copyright (C) 2017, 2019-2020 Heiko Bauke.
    Copyright (C) 2018, 2020, 2022-2023, 2025-2026 Aurélien PIERRE.
    Copyright (C) 2018 Edgardo Hoszowski.
    Copyright (C) 2018 Maurizio Paglia.
    Copyright (C) 2018-2020, 2022 Pascal Obry.
    Copyright (C) 2018 rawfiner.
    Copyright (C) 2019 Andreas Schneider.
    Copyright (C) 2019-2020, 2022 Diederik Ter Rahe.
    Copyright (C) 2019 Diederik ter Rahe.
    Copyright (C) 2019 Jacques Le Clerc.
    Copyright (C) 2020 Aldric Renaudin.
    Copyright (C) 2020-2021 Hubert Kowalski.
    Copyright (C) 2020-2021 Ralf Brown.
    Copyright (C) 2022 Hanno Schwalm.
    Copyright (C) 2022 Martin Bařinka.
    Copyright (C) 2022 Philipp Lutz.
    
    darktable is free software: you can redistribute it and/or modify
    it under the terms of the GNU General Public License as published by
    the Free Software Foundation, either version 3 of the License, or
    (at your option) any later version.
    
    darktable is distributed in the hope that it will be useful,
    but WITHOUT ANY WARRANTY; without even the implied warranty of
    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
    GNU General Public License for more details.
    
    You should have received a copy of the GNU General Public License
    along with darktable.  If not, see <http://www.gnu.org/licenses/>.
*/
#ifdef HAVE_CONFIG_H
#include "caches/pixelpipe_cache_alloc.h"
#include "config.h"
#endif
#include "widgets/bauhaus.h"
#include "system/macros.h"
#include "system/openmp.h"
#include "system/target_clones.h"
#include "system/mem_alloc.h"
#include "system/simd.h"
#include "common/module_versioning.h"
#include <glib/gstdio.h>
#include "pixel/bilateral.h"
#include "colorprofiles/colorspaces.h"
#include "common/imagebuf.h"
#include "common/points.h"
#include "control/control.h"
#include "develop/develop.h"
#include "develop/imageop.h"
#include "develop/imageop_gui.h"
#include "develop/tiling.h"
#include "widgets/drawingarea.h"
 
#include "iop/iop_api.h"
 
#include <gtk/gtk.h>
#include <inttypes.h>
#include <math.h>
#include <stdlib.h>
#include <string.h>
#include "widgets/label.h"
#include "gui/screen_metrics.h"
#include "control/signal.h"
 
DT_MODULE_INTROSPECTION(1, dt_iop_colormapping_params_t)
 
#define HISTN (1 << 11)
#define MAXN 5
 
typedef float float2[2];
 
typedef enum dt_iop_colormapping_flags_t
{
  NEUTRAL = 0,
  HAS_SOURCE = 1 << 0,
  HAS_TARGET = 1 << 1,
  HAS_SOURCE_TARGET = HAS_SOURCE | HAS_TARGET,
  ACQUIRE = 1 << 2,
  GET_SOURCE = 1 << 3,
  GET_TARGET = 1 << 4
} dt_iop_colormapping_flags_t;
 
typedef struct dt_iop_colormapping_flowback_t
{
  float hist[HISTN];
  // n-means (max 5?) with mean/variance
  float2 mean[MAXN];
  float2 var[MAXN];
  float weight[MAXN];
  // number of gaussians used.
  int n; // $MIN: 1 $MAX: 5 $DEFAULT: 1 $DESCRIPTION: "number of clusters"
} dt_iop_colormapping_flowback_t;
 
typedef struct dt_iop_colormapping_params_t
{
  dt_iop_colormapping_flags_t flag; // $DEFAULT: NEUTRAL
  // number of gaussians used.
  int n; // $MIN: 1 $MAX: 5 $DEFAULT: 3 $DESCRIPTION: "number of clusters"
 
  // relative importance of color dominance vs. color proximity
  float dominance; // $MIN: 0.0 $MAX: 100.0 $DEFAULT: 100.0 $DESCRIPTION: "color dominance"
 
  // level of histogram equalization
  float equalization; // $MIN: 0.0 $MAX: 100.0 $DEFAULT: 50.0 $DESCRIPTION: "histogram equalization"
 
  // hist matching table for source image
  float source_ihist[HISTN];
  // n-means (max 5) with mean/variance for source image
  float2 source_mean[MAXN];
  float2 source_var[MAXN];
  float source_weight[MAXN];
 
  // hist matching table for destination image
  int target_hist[HISTN];
  // n-means (max 5) with mean/variance for source image
  float2 target_mean[MAXN];
  float2 target_var[MAXN];
  float target_weight[MAXN];
} dt_iop_colormapping_params_t;
 
typedef struct dt_iop_colormapping_params_t dt_iop_colormapping_data_t;
 
 
typedef struct dt_iop_colormapping_gui_data_t
{
  int flag;
  float *buffer;
  int width;
  int height;
  int ch;
  int flowback_set;
  dt_iop_colormapping_flowback_t flowback;
  GtkWidget *acquire_source_button;
  GtkWidget *acquire_target_button;
  GtkWidget *source_area;
  GtkWidget *target_area;
  GtkWidget *clusters;
  GtkWidget *dominance;
  GtkWidget *equalization;
  cmsHTRANSFORM xform;
} dt_iop_colormapping_gui_data_t;
 
const char *name()
{
  return _("color mapping");
}
 
const char **description(struct dt_iop_module_t *self)
{
  return dt_iop_set_description(self, _("transfer a color palette and tonal repartition from one image to another"),
                                      _("creative"),
                                      _("linear or non-linear, Lab, display-referred"),
                                      _("non-linear, Lab"),
                                      _("non-linear, Lab, display-referred"));
}
 
int default_group()
{
  return IOP_GROUP_COLOR;
}
 
int flags()
{
  return IOP_FLAGS_ONE_INSTANCE | IOP_FLAGS_SUPPORTS_BLENDING | IOP_FLAGS_DEPRECATED;
}
 
int default_colorspace(dt_iop_module_t *self, dt_dev_pixelpipe_t *pipe, const dt_dev_pixelpipe_iop_t *piece)
{
  return IOP_CS_LAB;
}
 
static void capture_histogram(const float *col, const int width, const int height, int *hist)
{
  // build separate histogram
  memset(hist, 0, sizeof(int) * HISTN);
  for(int k = 0; k < height; k++)
    for(int i = 0; i < width; i++)
    {
      const int bin = CLAMP(HISTN * col[4 * (k * width + i) + 0] / 100.0, 0, HISTN - 1);
      hist[bin]++;
    }
 
  // accumulated start distribution of G1 G2
  for(int k = 1; k < HISTN; k++) hist[k] += hist[k - 1];
  for(int k = 0; k < HISTN; k++)
    hist[k] = (int)CLAMP(hist[k] * (HISTN / (float)hist[HISTN - 1]), 0, HISTN - 1);
  // for(int i=0;i<100;i++) printf("#[%d] %d \n", (int)CLAMP(HISTN*i/100.0, 0, HISTN-1),
  // hist[(int)CLAMP(HISTN*i/100.0, 0, HISTN-1)]);
}
 
static void invert_histogram(const int *hist, float *inv_hist)
{
// invert non-normalised accumulated hist
#if 0
  int last = 0;
  for(int i=0; i<HISTN; i++) for(int k=last; k<HISTN; k++)
      if(hist[k] >= i)
      {
        last = k;
        inv_hist[i] = 100.0*k/(float)HISTN;
        break;
      }
#else
  int last = 31;
  for(int i = 0; i <= last; i++) inv_hist[i] = 100.0f * i / (float)HISTN;
  for(int i = last + 1; i < HISTN; i++)
    for(int k = last; k < HISTN; k++)
      if(hist[k] >= i)
      {
        last = k;
        inv_hist[i] = 100.0f * k / (float)HISTN;
        break;
      }
#endif
 
  // printf("inv histogram debug:\n");
  // for(int i=0;i<100;i++) printf("%d => %f\n", i, inv_hist[hist[(int)CLAMP(HISTN*i/100.0, 0, HISTN-1)]]);
  // for(int i=0;i<100;i++) printf("[%d] %f => %f\n", (int)CLAMP(HISTN*i/100.0, 0, HISTN-1),
  // hist[(int)CLAMP(HISTN*i/100.0, 0, HISTN-1)]/(float)HISTN, inv_hist[(int)CLAMP(HISTN*i/100.0, 0,
  // HISTN-1)]);
}
 
__DT_CLONE_TARGETS__
static void get_cluster_mapping(const int n, float2 *mi, const float *wi, float2 *mo, const float *wo,
                                const float dominance, int *mapio)
{
  const float weightscale = 10000.0f;
 
  for(int ki = 0; ki < n; ki++)
  {
    // for each input cluster
    float mdist = FLT_MAX;
    for(int ko = 0; ko < n; ko++)
    {
      // find the best target cluster (the same could be used more than once)
      const float colordist = (mo[ko][0] - mi[ki][0]) * (mo[ko][0] - mi[ki][0])
                              + (mo[ko][1] - mi[ki][1]) * (mo[ko][1] - mi[ki][1]);
      const float weightdist = weightscale * (wo[ko] - wi[ki]) * (wo[ko] - wi[ki]);
      const float dist = colordist * (1.0f - dominance) + weightdist * dominance;
      if(dist < mdist)
      {
        // printf("[%d] => [%d] dominance: %f, colordist: %f, weightdist: %f, dist: %f\n", ki, ko, dominance,
        // colordist, weightdist, dist);
        mdist = dist;
        mapio[ki] = ko;
      }
    }
  }
 
  // printf("cluster mapping:\n");
  // for(int i=0;i<n;i++) printf("[%d] => [%d]\n", i, mapio[i]);
}
 
 
// inverse distant weighting according to D. Shepard's method; with power parameter 2.0
__DT_CLONE_TARGETS__
static void get_clusters(const float *col, const int n, float2 *mean, float *weight)
{
  float mdist = FLT_MAX;
  for(int k = 0; k < n; k++)
  {
    const float dist2 = (col[1] - mean[k][0]) * (col[1] - mean[k][0])
                        + (col[2] - mean[k][1]) * (col[2] - mean[k][1]); // dist^2
    weight[k] = dist2 > 1.0e-6f ? 1.0f / dist2 : -1.0f;                  // direct hits marked as -1
    if(dist2 < mdist) mdist = dist2;
  }
  if(mdist < 1.0e-6f)
    for(int k = 0; k < n; k++)
      weight[k] = weight[k] < 0.0f ? 1.0f : 0.0f; // correction in case of direct hits
  float sum = 0.0f;
  for(int k = 0; k < n; k++) sum += weight[k];
  if(sum > 0.0f)
    for(int k = 0; k < n; k++) weight[k] /= sum;
}
 
 
static int get_cluster(const float *col, const int n, float2 *mean)
{
  float mdist = FLT_MAX;
  int cluster = 0;
  for(int k = 0; k < n; k++)
  {
    const float dist = (col[1] - mean[k][0]) * (col[1] - mean[k][0])
                       + (col[2] - mean[k][1]) * (col[2] - mean[k][1]);
    if(dist < mdist)
    {
      mdist = dist;
      cluster = k;
    }
  }
  return cluster;
}
 
static void kmeans(const float *col, const int width, const int height, const int n, float2 *mean_out,
                   float2 *var_out, float *weight_out)
{
  const int nit = 40;                       // number of iterations
  const int samples = width * height * 0.2; // samples: only a fraction of the buffer.
 
  float2 *const mean = malloc(sizeof(float2) * n);
  float2 *const var = malloc(sizeof(float2) * n);
  int *const cnt = malloc(sizeof(int) * n);
  int count;
 
  float a_min = FLT_MAX, b_min = FLT_MAX, a_max = FLT_MIN, b_max = FLT_MIN;
 
  for(int s = 0; s < samples; s++)
  {
    const int j = CLAMP(dt_points_get() * height, 0, height - 1);
    const int i = CLAMP(dt_points_get() * width, 0, width - 1);
 
    const float a = col[4 * (width * j + i) + 1];
    const float b = col[4 * (width * j + i) + 2];
 
    a_min = fminf(a, a_min);
    a_max = fmaxf(a, a_max);
    b_min = fminf(b, b_min);
    b_max = fmaxf(b, b_max);
  }
 
  // init n clusters for a, b channels at random
  for(int k = 0; k < n; k++)
  {
    mean_out[k][0] = 0.9f * (a_min + (a_max - a_min) * dt_points_get());
    mean_out[k][1] = 0.9f * (b_min + (b_max - b_min) * dt_points_get());
    var_out[k][0] = var_out[k][1] = weight_out[k] = 0.0f;
    mean[k][0] = mean[k][1] = var[k][0] = var[k][1] = 0.0f;
  }
  for(int it = 0; it < nit; it++)
  {
    for(int k = 0; k < n; k++) cnt[k] = 0;
// randomly sample col positions inside roi
    __OMP_PARALLEL_FOR__()
    for(int s = 0; s < samples; s++)
    {
      const int j = CLAMP(dt_points_get() * height, 0, height - 1);
      const int i = CLAMP(dt_points_get() * width, 0, width - 1);
      // for each sample: determine cluster, update new mean, update var
      for(int k = 0; k < n; k++)
      {
        const float L = col[4 * (width * j + i)];
        const dt_aligned_pixel_t Lab = { L, col[4 * (width * j + i) + 1], col[4 * (width * j + i) + 2] };
        // determine dist to mean_out
        const int c = get_cluster(Lab, n, mean_out);
#ifdef _OPENMP
#pragma omp atomic
#endif
        cnt[c]++;
// update mean, var
#ifdef _OPENMP
#pragma omp atomic
#endif
        var[c][0] += Lab[1] * Lab[1];
#ifdef _OPENMP
#pragma omp atomic
#endif
        var[c][1] += Lab[2] * Lab[2];
#ifdef _OPENMP
#pragma omp atomic
#endif
        mean[c][0] += Lab[1];
#ifdef _OPENMP
#pragma omp atomic
#endif
        mean[c][1] += Lab[2];
      }
    }
    // swap old/new means
    for(int k = 0; k < n; k++)
    {
      if(cnt[k] == 0) continue;
      mean_out[k][0] = mean[k][0] / cnt[k];
      mean_out[k][1] = mean[k][1] / cnt[k];
      var_out[k][0] = var[k][0] / cnt[k] - mean_out[k][0] * mean_out[k][0];
      var_out[k][1] = var[k][1] / cnt[k] - mean_out[k][1] * mean_out[k][1];
      mean[k][0] = mean[k][1] = var[k][0] = var[k][1] = 0.0f;
    }
 
    // determine weight of clusters
    count = 0;
    for(int k = 0; k < n; k++) count += cnt[k];
    for(int k = 0; k < n; k++) weight_out[k] = (count > 0) ? (float)cnt[k] / count : 0.0f;
 
    // printf("it %d  %d means:\n", it, n);
    // for(int k=0;k<n;k++) printf("mean %f %f -- var %f %f -- weight %f\n", mean_out[k][0], mean_out[k][1],
    // var_out[k][0], var_out[k][1], weight_out[k]);
  }
 
  dt_free(cnt);
  dt_free(var);
  dt_free(mean);
 
  for(int k = 0; k < n; k++)
  {
    // "eliminate" clusters with a variance of zero
    if(var_out[k][0] == 0.0f || var_out[k][1] == 0.0f)
      mean_out[k][0] = mean_out[k][1] = var_out[k][0] = var_out[k][1] = weight_out[k] = 0;
 
    // we actually want the std deviation.
    var_out[k][0] = sqrtf(var_out[k][0]);
    var_out[k][1] = sqrtf(var_out[k][1]);
  }
 
  // simple bubblesort of clusters in order of ascending weight: just a convenience for the user to keep
  // cluster display a bit more consistent in GUI
  for(int i = 0; i < n - 1; i++)
  {
    for(int j = 0; j < n - 1 - i; j++)
    {
      if(weight_out[j] > weight_out[j + 1])
      {
        float2 temp_mean = { mean_out[j + 1][0], mean_out[j + 1][1] };
        float2 temp_var = { var_out[j + 1][0], var_out[j + 1][1] };
        float temp_weight = weight_out[j + 1];
 
        mean_out[j + 1][0] = mean_out[j][0];
        mean_out[j + 1][1] = mean_out[j][1];
        var_out[j + 1][0] = var_out[j][0];
        var_out[j + 1][1] = var_out[j][1];
        weight_out[j + 1] = weight_out[j];
 
        mean_out[j][0] = temp_mean[0];
        mean_out[j][1] = temp_mean[1];
        var_out[j][0] = temp_var[0];
        var_out[j][1] = temp_var[1];
        weight_out[j] = temp_weight;
      }
    }
  }
}
 
__DT_CLONE_TARGETS__
int process(struct dt_iop_module_t *self, const dt_dev_pixelpipe_t *pipe, const dt_dev_pixelpipe_iop_t *piece, const void *const ivoid,
             void *const ovoid)
{
  const dt_iop_roi_t *const roi_in = &piece->roi_in;
  const dt_iop_roi_t *const roi_out = &piece->roi_out;
  dt_iop_colormapping_data_t *const restrict data = (dt_iop_colormapping_data_t *)piece->data;
  dt_iop_colormapping_gui_data_t *const restrict g = (dt_iop_colormapping_gui_data_t *)dt_iop_gui_data(self);
  float *const restrict in = (float *)ivoid;
  float *const restrict out = (float *)ovoid;
 
  const int width = roi_in->width;
  const int height = roi_in->height;
 
  const float scale = dt_dev_get_module_scale(pipe, roi_in);
  const float sigma_s = 50.0f / scale;
  const float sigma_r = 8.0f; // does not depend on scale
 
  // save a copy of preview input buffer so we can get histogram and color statistics out of it
  if(self->dev->gui_attached && !IS_NULL_PTR(g) && dt_dev_pixelpipe_has_preview_output(self->dev, pipe, roi_out)
     && (data->flag & ACQUIRE))
  {
    dt_iop_gui_enter_critical_section(self);
    dt_free_align(g->buffer);
    g->buffer = NULL;
 
    g->buffer = dt_iop_image_alloc(width, height, 4);
    g->width = width;
    g->height = height;
    g->ch = 4;
 
    if(IS_NULL_PTR(g->buffer))
    {
      dt_iop_gui_leave_critical_section(self);
      return 1;
    }
 
    dt_iop_image_copy_by_size(g->buffer, in, width, height, 4);
 
    dt_iop_gui_leave_critical_section(self);
  }
 
  // process image if all mapping information is present in the parameter set
  if(data->flag & HAS_TARGET && data->flag & HAS_SOURCE)
  {
    // for all pixels: find input cluster, transfer to mapped target cluster and apply histogram
 
    const float dominance = data->dominance / 100.0f;
    const float equalization = data->equalization / 100.0f;
 
    // get mapping from input clusters to target clusters
    int *const mapio = malloc(sizeof(int) * data->n);
    if(IS_NULL_PTR(mapio)) return 1;
 
    get_cluster_mapping(data->n, data->target_mean, data->target_weight, data->source_mean,
                        data->source_weight, dominance, mapio);
 
    float2 *const var_ratio = malloc(sizeof(float2) * data->n);
    if(IS_NULL_PTR(var_ratio))
    {
      dt_free(mapio);
      return 1;
    }
 
    for(int i = 0; i < data->n; i++)
    {
      var_ratio[i][0]
          = (data->target_var[i][0] > 0.0f) ? data->source_var[mapio[i]][0] / data->target_var[i][0] : 0.0f;
      var_ratio[i][1]
          = (data->target_var[i][1] > 0.0f) ? data->source_var[mapio[i]][1] / data->target_var[i][1] : 0.0f;
    }
 
    const size_t npixels = (size_t)height * width;
// first get delta L of equalized L minus original image L, scaled to fit into [0 .. 100]
    __OMP_PARALLEL_FOR__()
    for(size_t k = 0; k < npixels * 4; k += 4)
    {
      const float L = in[k];
      out[k] = 0.5f * ((L * (1.0f - equalization)
                        + data->source_ihist[data->target_hist[(int)CLAMP(
                              HISTN * L / 100.0f, 0.0f, (float)HISTN - 1.0f)]] * equalization) - L) + 50.0f;
      out[k] = CLAMP(out[k], 0.0f, 100.0f);
    }
 
    if(equalization > 0.001f)
    {
      // bilateral blur of delta L to avoid artifacts caused by limited histogram resolution
      dt_bilateral_t *b = dt_bilateral_init(width, height, sigma_s, sigma_r);
      if(IS_NULL_PTR(b))
      {
        dt_free(var_ratio);
        dt_free(mapio);
        return 1;
      }
      dt_bilateral_splat(b, out);
      dt_bilateral_blur(b);
      dt_bilateral_slice(b, out, out, -1.0f);
      dt_bilateral_free(b);
    }
 
    size_t allocsize;
    float *const weight_buf = dt_pixelpipe_cache_alloc_perthread(data->n, sizeof(float), &allocsize);
    if(IS_NULL_PTR(weight_buf))
    {
      dt_free(var_ratio);
      dt_free(mapio);
      return 1;
    }
    __OMP_PARALLEL__()
    {
      // get a thread-private scratch buffer; do this before the actual loop so we don't have to look it up for
      // every single pixel
      float *const restrict weight = dt_get_perthread(weight_buf,allocsize);
      __OMP_FOR__()
      for(size_t j = 0; j < 4*npixels; j += 4)
      {
        const float L = in[j];
        const dt_aligned_pixel_t Lab = { L, in[j + 1], in[j + 2] };
 
        // transfer back scaled and blurred delta L to output L
        out[j] = 2.0f * (out[j] - 50.0f) + L;
        out[j] = CLAMP(out[j], 0.0f, 100.0f);
 
        get_clusters(in + j, data->n, data->target_mean, weight);
        // zero the 'a' and 'b' channels
        out[j + 1] = out[j + 2] = 0.0f;
        // then accumulate a weighted average for a and b
        for(int c = 0; c < data->n; c++)
        {
          out[j + 1] += weight[c] * ((Lab[1] - data->target_mean[c][0]) * var_ratio[c][0]
                                     + data->source_mean[mapio[c]][0]);
          out[j + 2] += weight[c] * ((Lab[2] - data->target_mean[c][1]) * var_ratio[c][1]
                                     + data->source_mean[mapio[c]][1]);
        }
        // pass through the alpha channel
        out[j + 3] = in[j + 3];
      }
    }
 
    dt_pixelpipe_cache_free_align(weight_buf);
    dt_free(var_ratio);
    dt_free(mapio);
  }
  // incomplete parameter set -> do nothing
  else
  {
    dt_iop_image_copy_by_size(out, in, width, height, 4);
  }
 
  return 0;
}
 
void tiling_callback(struct dt_iop_module_t *self, const struct dt_dev_pixelpipe_t *pipe, const struct dt_dev_pixelpipe_iop_t *piece, struct dt_develop_tiling_t *tiling)
{
  const dt_iop_roi_t *const roi_in = &piece->roi_in;
  const float scale = dt_dev_get_module_scale(pipe, roi_in);
  const float sigma_s = 50.0f / scale;
  const float sigma_r = 8.0f; // does not depend on scale
 
  const int width = roi_in->width;
  const int height = roi_in->height;
  const int channels = piece->dsc_in.channels;
 
  const size_t basebuffer = sizeof(float) * channels * width * height;
 
  tiling->factor = 3.0f + (float)dt_bilateral_memory_use(width, height, sigma_s, sigma_r) / basebuffer;
  tiling->maxbuf
      = fmaxf(1.0f, (float)dt_bilateral_singlebuffer_size(width, height, sigma_s, sigma_r) / basebuffer);
  tiling->overhead = 0;
  tiling->overlap = ceilf(4 * sigma_s);
  tiling->xalign = 1;
  tiling->yalign = 1;
}
 
void commit_params(struct dt_iop_module_t *self, dt_iop_params_t *p1, dt_dev_pixelpipe_t *pipe,
                   dt_dev_pixelpipe_iop_t *piece)
{
  dt_iop_colormapping_params_t *p = (dt_iop_colormapping_params_t *)p1;
  dt_iop_colormapping_data_t *d = (dt_iop_colormapping_data_t *)piece->data;
 
  memcpy(d, p, sizeof(dt_iop_colormapping_params_t));
}
 
void gui_changed(dt_iop_module_t *self, GtkWidget *w, void *previous)
{
  dt_iop_colormapping_params_t *p = (dt_iop_colormapping_params_t *)self->params;
  dt_iop_colormapping_gui_data_t *g = (dt_iop_colormapping_gui_data_t *)dt_iop_gui_data(self);
 
  if(w == g->clusters)
  {
    // only reset source/target when changing number of clusters
    memset(p->source_ihist, 0, sizeof(float) * HISTN);
    memset(p->source_mean, 0, sizeof(float) * MAXN * 2);
    memset(p->source_var, 0, sizeof(float) * MAXN * 2);
    memset(p->source_weight, 0, sizeof(float) * MAXN);
    memset(p->target_hist, 0, sizeof(int) * HISTN);
    memset(p->target_mean, 0, sizeof(float) * MAXN * 2);
    memset(p->target_var, 0, sizeof(float) * MAXN * 2);
    memset(p->target_weight, 0, sizeof(float) * MAXN);
    p->flag = NEUTRAL;
    dt_control_queue_redraw_widget(g->source_area);
    dt_control_queue_redraw_widget(g->target_area);
  }
}
 
static void acquire_source_button_pressed(GtkButton *button, dt_iop_module_t *self)
{
  if(dt_gui_widgets_suppressed()) return;
  dt_iop_colormapping_params_t *p = (dt_iop_colormapping_params_t *)self->params;
  p->flag |= ACQUIRE;
  p->flag |= GET_SOURCE;
  p->flag &= ~HAS_SOURCE;
  dt_iop_request_focus(self);
  dt_dev_add_history_item(self->dev, self, TRUE, TRUE);
}
 
static void acquire_target_button_pressed(GtkButton *button, dt_iop_module_t *self)
{
  if(dt_gui_widgets_suppressed()) return;
  dt_iop_colormapping_params_t *p = (dt_iop_colormapping_params_t *)self->params;
  p->flag |= ACQUIRE;
  p->flag |= GET_TARGET;
  p->flag &= ~HAS_TARGET;
  dt_iop_request_focus(self);
  dt_dev_add_history_item(self->dev, self, TRUE, TRUE);
}
 
void init_pipe(struct dt_iop_module_t *self, dt_dev_pixelpipe_t *pipe, dt_dev_pixelpipe_iop_t *piece)
{
  piece->data = dt_calloc_align(sizeof(dt_iop_colormapping_data_t));
  piece->data_size = sizeof(dt_iop_colormapping_data_t);
}
 
void cleanup_pipe(struct dt_iop_module_t *self, dt_dev_pixelpipe_t *pipe, dt_dev_pixelpipe_iop_t *piece)
{
  dt_free_align(piece->data);
  piece->data = NULL;
}
 
void reload_defaults(dt_iop_module_t *module)
{
  dt_iop_colormapping_params_t *d = module->default_params;
 
  dt_iop_colormapping_gui_data_t *g = (dt_iop_colormapping_gui_data_t *)dt_iop_gui_data(module);
  if(module->dev->gui_attached && !IS_NULL_PTR(g) && g->flowback_set)
  {
    memcpy(d->source_ihist, g->flowback.hist, sizeof(float) * HISTN);
    memcpy(d->source_mean, g->flowback.mean, sizeof(float) * MAXN * 2);
    memcpy(d->source_var, g->flowback.var, sizeof(float) * MAXN * 2);
    memcpy(d->source_weight, g->flowback.weight, sizeof(float) * MAXN);
    d->n = g->flowback.n;
    d->flag = HAS_SOURCE;
  }
}
 
 
static gboolean cluster_preview_draw(GtkWidget *widget, cairo_t *crf, dt_iop_module_t *self)
{
  dt_iop_colormapping_params_t *p = (dt_iop_colormapping_params_t *)self->params;
  dt_iop_colormapping_gui_data_t *g = (dt_iop_colormapping_gui_data_t *)dt_iop_gui_data(self);
 
  float2 *mean;
  float2 *var;
 
  if(widget == g->source_area)
  {
    mean = p->source_mean;
    var = p->source_var;
  }
  else
  {
    mean = p->target_mean;
    var = p->target_var;
  }
 
 
  GtkAllocation allocation;
  gtk_widget_get_allocation(widget, &allocation);
  const int inset = 5;
  int width = allocation.width, height = allocation.height;
  cairo_surface_t *cst = dt_cairo_image_surface_create(CAIRO_FORMAT_ARGB32, width, height);
  cairo_t *cr = cairo_create(cst);
  cairo_set_source_rgb(cr, .2, .2, .2);
  cairo_paint(cr);
 
  cairo_translate(cr, inset, inset);
  width -= 2 * inset;
  height -= 2 * inset;
 
 
  const float sep = DT_PIXEL_APPLY_DPI(2.0);
  const float qwd = (width - (p->n - 1) * sep) / (float)p->n;
  for(int cl = 0; cl < p->n; cl++)
  {
    // draw cluster
    for(int j = -1; j <= 1; j++)
      for(int i = -1; i <= 1; i++)
      {
        // draw 9x9 grid showing mean and variance of this cluster.
        double rgb[3] = { 0.5, 0.5, 0.5 };
        cmsCIELab Lab;
        Lab.L = 53.390011;
        Lab.a = (mean[cl][0] + i * var[cl][0]);
        Lab.b = (mean[cl][1] + j * var[cl][1]);
        cmsDoTransform(g->xform, &Lab, rgb, 1);
        cairo_set_source_rgb(cr, rgb[0], rgb[1], rgb[2]);
        cairo_rectangle(cr, qwd * (i + 1) / 3.0, height * (j + 1) / 3.0, qwd / 3.0 - DT_PIXEL_APPLY_DPI(.5),
                        height / 3.0 - DT_PIXEL_APPLY_DPI(.5));
        cairo_fill(cr);
      }
    cairo_translate(cr, qwd + sep, 0);
  }
 
  cairo_destroy(cr);
  cairo_set_source_surface(crf, cst, 0, 0);
  cairo_paint(crf);
  cairo_surface_destroy(cst);
  return TRUE;
}
 
 
static void process_clusters(gpointer instance, gpointer user_data)
{
  dt_iop_module_t *self = (dt_iop_module_t *)user_data;
  dt_iop_colormapping_params_t *p = (dt_iop_colormapping_params_t *)self->params;
  dt_iop_colormapping_gui_data_t *g = (dt_iop_colormapping_gui_data_t *)dt_iop_gui_data(self);
  int new_source_clusters = 0;
 
  if(IS_NULL_PTR(g) || IS_NULL_PTR(g->buffer)) return;
  if(!(p->flag & ACQUIRE)) return;
 
  {
    // Scope guard rather than a begin/end pair: this span has an early return on buffer
    // allocation failure, which would leak the freeze depth with a raw bracket. The guard
    // releases the freeze on every exit path, including that return.
    dt_gui_widget_freeze();
 
    dt_iop_gui_enter_critical_section(self);
    const int width = g->width;
    const int height = g->height;
    const int ch = g->ch;
    float *const restrict buffer = dt_iop_image_alloc(width, height, ch);
    if(IS_NULL_PTR(buffer))
    {
      dt_iop_gui_leave_critical_section(self);
      return;
    }
    dt_iop_image_copy_by_size(buffer, g->buffer, width, height, ch);
    dt_iop_gui_leave_critical_section(self);
 
    if(p->flag & GET_SOURCE)
    {
      int hist[HISTN];
 
      // get histogram of L
      capture_histogram(buffer, width, height, hist);
 
      // invert histogram
      invert_histogram(hist, p->source_ihist);
 
      // get n color clusters
      kmeans(buffer, width, height, p->n, p->source_mean, p->source_var, p->source_weight);
 
      p->flag |= HAS_SOURCE;
      new_source_clusters = 1;
 
      dt_control_queue_redraw_widget(g->source_area);
    }
    else if(p->flag & GET_TARGET)
    {
      // get histogram of L
      capture_histogram(buffer, width, height, p->target_hist);
 
      // get n color clusters
      kmeans(buffer, width, height, p->n, p->target_mean, p->target_var, p->target_weight);
 
      p->flag |= HAS_TARGET;
 
      dt_control_queue_redraw_widget(g->target_area);
    }
 
    dt_free_align(buffer);
 
    if(new_source_clusters)
    {
      memcpy(g->flowback.hist, p->source_ihist, sizeof(float) * HISTN);
      memcpy(g->flowback.mean, p->source_mean, sizeof(float) * MAXN * 2);
      memcpy(g->flowback.var, p->source_var, sizeof(float) * MAXN * 2);
      memcpy(g->flowback.weight, p->source_weight, sizeof(float) * MAXN);
      g->flowback.n = p->n;
      g->flowback_set = 1;
      FILE *f = g_fopen("/tmp/dt_colormapping_loaded", "wb");
      if(f)
      {
        if(fwrite(&g->flowback, sizeof(g->flowback), 1, f) < 1)
          fprintf(stderr, "[colormapping] could not write flowback file /tmp/dt_colormapping_loaded\n");
        fclose(f);
      }
    }
 
    p->flag &= ~(GET_TARGET | GET_SOURCE | ACQUIRE);
  }
 
  if(p->flag & HAS_SOURCE) dt_dev_add_history_item(self->dev, self, TRUE, TRUE);
 
  dt_control_queue_redraw();
}
 
 
void gui_init(struct dt_iop_module_t *self)
{
  dt_iop_colormapping_gui_data_t *g = IOP_GUI_ALLOC(colormapping);
 
  g->flag = NEUTRAL;
  g->flowback_set = 0;
  cmsHPROFILE hsRGB = dt_colorspaces_get_profile(DT_COLORSPACE_SRGB, "", DT_PROFILE_ROLE_INPUT)->profile;
  cmsHPROFILE hLab = dt_colorspaces_get_profile(DT_COLORSPACE_LAB, "", DT_PROFILE_ROLE_ANY)->profile;
  g->xform = cmsCreateTransform(hLab, TYPE_Lab_DBL, hsRGB, TYPE_RGB_DBL, INTENT_PERCEPTUAL, 0);
  g->buffer = NULL;
 
  self->gui->widget = GTK_WIDGET(gtk_box_new(GTK_ORIENTATION_VERTICAL, DT_GUI_BOX_SPACING));
 
  gtk_box_pack_start(GTK_BOX(self->gui->widget), dt_ui_label_new(_("source clusters:")), TRUE, TRUE, 0);
 
  g->source_area = dtgtk_drawing_area_new_with_aspect_ratio(1.0 / 3.0);
  gtk_box_pack_start(GTK_BOX(self->gui->widget), g->source_area, TRUE, TRUE, 0);
  g_signal_connect(G_OBJECT(g->source_area), "draw", G_CALLBACK(cluster_preview_draw), self);
 
  gtk_box_pack_start(GTK_BOX(self->gui->widget), dt_ui_label_new(_("target clusters:")), TRUE, TRUE, 0);
 
  g->target_area = dtgtk_drawing_area_new_with_aspect_ratio(1.0 / 3.0);
  gtk_box_pack_start(GTK_BOX(self->gui->widget), g->target_area, TRUE, TRUE, 0);
  g_signal_connect(G_OBJECT(g->target_area), "draw", G_CALLBACK(cluster_preview_draw), self);
 
  GtkWidget *box = gtk_box_new(GTK_ORIENTATION_HORIZONTAL, DT_GUI_BOX_SPACING);
  gtk_box_pack_start(GTK_BOX(self->gui->widget), box, TRUE, TRUE, 0);
 
  g->acquire_source_button = dt_iop_button_new(self, N_("acquire as source"),
                                               G_CALLBACK(acquire_source_button_pressed), FALSE, 0, 0,
                                               NULL, 0, box);
  gtk_label_set_ellipsize(GTK_LABEL(gtk_bin_get_child(GTK_BIN(g->acquire_source_button))), PANGO_ELLIPSIZE_START);
  gtk_widget_set_tooltip_text(g->acquire_source_button, _("analyze this image as a source image"));
 
  g->acquire_target_button = dt_iop_button_new(self, N_("acquire as target"),
                                               G_CALLBACK(acquire_target_button_pressed), FALSE, 0, 0,
                                               NULL, 0, box);
  gtk_label_set_ellipsize(GTK_LABEL(gtk_bin_get_child(GTK_BIN(g->acquire_target_button))), PANGO_ELLIPSIZE_START);
  gtk_widget_set_tooltip_text(g->acquire_target_button, _("analyze this image as a target image"));
 
  g->clusters = dt_bauhaus_slider_from_params(self, "n");
  gtk_widget_set_tooltip_text(g->clusters, _("number of clusters to find in image. value change resets all clusters"));
 
  g->dominance = dt_bauhaus_slider_from_params(self, "dominance");
  gtk_widget_set_tooltip_text(g->dominance, _("how clusters are mapped. low values: based on color "
                                              "proximity, high values: based on color dominance"));
  dt_bauhaus_slider_set_format(g->dominance, "%");
 
  g->equalization = dt_bauhaus_slider_from_params(self, "equalization");
  gtk_widget_set_tooltip_text(g->equalization, _("level of histogram equalization"));
  dt_bauhaus_slider_set_format(g->equalization, "%");
 
  /* add signal handler for preview pipe finished: process clusters if requested */
  DT_DEBUG_CONTROL_SIGNAL_CONNECT(dt_control_signal_get_global(), DT_SIGNAL_DEVELOP_PREVIEW_PIPE_FINISHED,
                            G_CALLBACK(process_clusters), self);
 
  FILE *f = g_fopen("/tmp/dt_colormapping_loaded", "rb");
  if(f)
  {
    if(fread(&g->flowback, sizeof(g->flowback), 1, f) > 0) g->flowback_set = 1;
    fclose(f);
  }
}
 
void gui_cleanup(struct dt_iop_module_t *self)
{
  dt_iop_colormapping_gui_data_t *g = (dt_iop_colormapping_gui_data_t *)dt_iop_gui_data(self);
 
  DT_DEBUG_CONTROL_SIGNAL_DISCONNECT(dt_control_signal_get_global(), G_CALLBACK(process_clusters), self);
 
  cmsDeleteTransform(g->xform);
  dt_free_align(g->buffer);
  g->buffer = NULL;
 
  IOP_GUI_FREE;
}
 
// clang-format off
// modelines: These editor modelines have been set for all relevant files by tools/update_modelines.py
// vim: shiftwidth=2 expandtab tabstop=2 cindent
// kate: tab-indents: off; indent-width 2; replace-tabs on; indent-mode cstyle; remove-trailing-spaces modified;
// clang-format on