rawprepare.c

Go to the documentation of this file.

/*
    This file is part of darktable,
    Copyright (C) 2014-2017, 2020 Roman Lebedev.
    Copyright (C) 2016 Dan Torop.
    Copyright (C) 2016, 2018 johannes hanika.
    Copyright (C) 2016, 2018-2019 Tobias Ellinghaus.
    Copyright (C) 2016 Ulrich Pegelow.
    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) 2019-2022 Hanno Schwalm.
    Copyright (C) 2020 Aldric Renaudin.
    Copyright (C) 2020-2021, 2023, 2025-2026 Aurélien PIERRE.
    Copyright (C) 2020, 2022 Diederik Ter Rahe.
    Copyright (C) 2020 Hubert Kowalski.
    Copyright (C) 2020, 2022 Ralf Brown.
    Copyright (C) 2021 luzpaz.
    Copyright (C) 2022 Martin Bařinka.
    Copyright (C) 2022 paolodepetrillo.
    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 "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 "common/logging.h"
#include "common/module_versioning.h"
#include "database/database.h"
#include "imageio/imageio_rawspeed.h" // for dt_rawspeed_crop_dcraw_filters
#include "common/opencl.h"
#include "common/imagebuf.h"
#include "common/image.h"
#include "develop/geometry/geometry.h"
#include "develop/imageop.h"
#include "develop/imageop_gui.h"
#include "caches/image_cache.h"
 
#include "gui/presets.h"
#include "iop/iop_api.h"
#include "common/dng_opcode.h"
 
#include <gtk/gtk.h>
#include <stdint.h>
#include <stdlib.h>
#include "widgets/label.h"
#include "control/signal.h"
 
DT_MODULE_INTROSPECTION(2, dt_iop_rawprepare_params_t)
 
typedef enum dt_iop_rawprepare_flat_field_t
{
  FLAT_FIELD_OFF = 0,     // $DESCRIPTION: "disabled"
  FLAT_FIELD_EMBEDDED = 1 // $DESCRIPTION: "embedded GainMap"
} dt_iop_rawprepare_flat_field_t;
 
typedef struct dt_iop_rawprepare_params_t
{
  int32_t x; // $MIN: 0 $MAX: UINT16_MAX $DESCRIPTION: "crop left"
  int32_t y; // $MIN: 0 $MAX: UINT16_MAX $DESCRIPTION: "crop top"
  int32_t width; // $MIN: 0 $MAX: UINT16_MAX $DESCRIPTION: "crop right"
  int32_t height; // $MIN: 0 $MAX: UINT16_MAX $DESCRIPTION: "crop bottom"
  uint16_t raw_black_level_separate[4]; // $MIN: 0 $MAX: UINT16_MAX $DESCRIPTION: "black level"
  uint16_t raw_white_point; // $MIN: 0 $MAX: UINT16_MAX $DESCRIPTION: "white point"
  dt_iop_rawprepare_flat_field_t flat_field; // $DEFAULT: FLAT_FIELD_OFF $DESCRIPTION: "flat field correction"
} dt_iop_rawprepare_params_t;
 
typedef struct dt_iop_rawprepare_gui_data_t
{
  GtkWidget *black_level_separate[4];
  GtkWidget *white_point;
  GtkWidget *x, *y, *width, *height;
  GtkWidget *flat_field;
} dt_iop_rawprepare_gui_data_t;
 
typedef struct dt_iop_rawprepare_data_t
{
  int32_t x, y, width, height; // crop, now unused, for future expansion
  float sub[4];
  float div[4];
 
  // cached for dt_iop_buffer_dsc_t::rawprepare
  struct
  {
    uint16_t raw_black_level;
    uint16_t raw_white_point;
  } rawprepare;
 
  // image contains GainMaps that should be applied
  gboolean apply_gainmaps;
  // GainMap for each filter of RGGB Bayer pattern
  dt_dng_gain_map_t *gainmaps[4];
} dt_iop_rawprepare_data_t;
 
typedef struct dt_iop_rawprepare_global_data_t
{
  int kernel_rawprepare_1f;
  int kernel_rawprepare_1f_gainmap;
  int kernel_rawprepare_1f_unnormalized;
  int kernel_rawprepare_1f_unnormalized_gainmap;
  int kernel_rawprepare_4f;
} dt_iop_rawprepare_global_data_t;
 
 
const char *name()
{
  return C_("modulename", "Raw settings");
}
 
int operation_tags()
{
  return IOP_TAG_DISTORT;
}
 
int flags()
{
  return IOP_FLAGS_ALLOW_TILING | IOP_FLAGS_TILING_FULL_ROI | IOP_FLAGS_ONE_INSTANCE
    | IOP_FLAGS_UNSAFE_COPY;
}
 
int default_group()
{
  return IOP_GROUP_TECHNICAL;
}
 
int default_colorspace(dt_iop_module_t *self, dt_dev_pixelpipe_t *pipe, const dt_dev_pixelpipe_iop_t *piece)
{
  if(piece && piece->dsc_in.cst != IOP_CS_RAW)
    return IOP_CS_RGB;
  return IOP_CS_RAW;
}
 
int legacy_params(dt_iop_module_t *self, const void *const old_params, const int old_version,
                  void *new_params, const int new_version)
{
  typedef struct dt_iop_rawprepare_params_t dt_iop_rawprepare_params_v2_t;
  typedef struct dt_iop_rawprepare_params_v1_t
  {
    int32_t x;
    int32_t y;
    int32_t width;
    int32_t height;
    uint16_t raw_black_level_separate[4];
    uint16_t raw_white_point;
  } dt_iop_rawprepare_params_v1_t;
 
  if(old_version == 1 && new_version == 2)
  {
    dt_iop_rawprepare_params_v1_t *o = (dt_iop_rawprepare_params_v1_t *)old_params;
    dt_iop_rawprepare_params_v2_t *n = (dt_iop_rawprepare_params_v2_t *)new_params;
    memcpy(n, o, sizeof *o);
    n->flat_field = FLAT_FIELD_OFF;
    return 0;
  }
 
  return 1;
}
 
const char **description(struct dt_iop_module_t *self)
{
  return dt_iop_set_description(self, _("sets technical specificities of the raw sensor.\n"
                                        "touch with great care!"),
                                      _("mandatory"),
                                      _("linear, raw, scene-referred"),
                                      _("linear, raw"),
                                      _("linear, raw, scene-referred"));
}
 
void init_presets(dt_iop_module_so_t *self)
{
  dt_database_start_transaction();
 
  dt_gui_presets_add_generic(_("passthrough"), self->op, self->version(),
                             &(dt_iop_rawprepare_params_t){.x = 0,
                                                           .y = 0,
                                                           .width = 0,
                                                           .height = 0,
                                                           .raw_black_level_separate[0] = 0,
                                                           .raw_black_level_separate[1] = 0,
                                                           .raw_black_level_separate[2] = 0,
                                                           .raw_black_level_separate[3] = 0,
                                                           .raw_white_point = UINT16_MAX },
                             sizeof(dt_iop_rawprepare_params_t), 1);
 
  dt_database_release_transaction();
}
 
static inline __attribute__((always_inline)) int compute_proper_crop(const dt_dev_pixelpipe_iop_t *piece, const dt_iop_roi_t *const roi_in, int value)
{
  const double scale = roi_in->scale;
  return (int)roundf((double)value * scale);
}
 
static void _update_output_cfa_descriptor(const dt_dev_pixelpipe_t *pipe,
                                          const dt_dev_pixelpipe_iop_t *piece,
                                          const dt_iop_roi_t *const roi_in,
                                          dt_iop_buffer_dsc_t *dsc)
{
  if(IS_NULL_PTR(pipe) || IS_NULL_PTR(piece) || IS_NULL_PTR(roi_in) || IS_NULL_PTR(dsc)) return;
 
  dsc->filters = pipe->dev->image_storage.dsc.filters;
  memcpy(dsc->xtrans, pipe->dev->image_storage.dsc.xtrans, sizeof(dsc->xtrans));
 
  if(!pipe->dev->image_storage.dsc.filters) return;
 
  /* Rawprepare is the stage that converts the immutable sensor-aligned RAW descriptor
   * attached to the input image into the runtime descriptor seen by downstream RAW
   * modules. Rebuild that contract from the pipe image each time instead of chaining
   * shifts from `piece->dsc_in`, otherwise repeated ROI planning can compound the
   * Bayer/X-Trans phase offset.
   *
   * Only the fixed sensor border trim (d->x/d->y) is folded in here: that offset is
   * constant for a given image, regardless of how the user pans/zooms/crops downstream.
   * The dynamic, ROI-dependent part of the phase shift is applied downstream, fresh on
   * every process()/process_cl() call, from the module's own current roi_in (see the
   * matching comment at the top of demosaic.c's process()) — not here, and not from any
   * ROI callback, since piece->dsc_in is not guaranteed to be populated yet at that point. */
 
  const dt_iop_rawprepare_data_t *d = (dt_iop_rawprepare_data_t *)piece->data;
  const uint32_t crop_x = compute_proper_crop(piece, roi_in, d->x);
  const uint32_t crop_y = compute_proper_crop(piece, roi_in, d->y);
 
  dsc->filters = dt_rawspeed_crop_dcraw_filters(pipe->dev->image_storage.dsc.filters, crop_x, crop_y);
 
  dt_print(DT_DEBUG_DEMOSAIC,
           "[rawprepare] pipe %p (%s) thread %lu piece %p dsc %p roi_in=(%d,%d) scale=%f d->x,y=(%d,%d) -> crop=(%u,%u) filters=0x%x -> 0x%x\n",
           (void *)pipe, pipe->type == DT_DEV_PIXELPIPE_FULL ? "full" :
                          pipe->type == DT_DEV_PIXELPIPE_PREVIEW ? "preview" :
                          pipe->type == DT_DEV_PIXELPIPE_EXPORT ? "export" :
                          pipe->type == DT_DEV_PIXELPIPE_THUMBNAIL ? "thumbnail" : "none",
           (unsigned long)pthread_self(), (void *)piece, (void *)dsc,
           roi_in->x, roi_in->y, roi_in->scale, d->x, d->y, crop_x, crop_y,
           pipe->dev->image_storage.dsc.filters, dsc->filters);
 
  if(pipe->dev->image_storage.dsc.filters != 9u) return;
 
  for(int i = 0; i < 6; ++i)
  {
    for(int j = 0; j < 6; ++j)
    {
      dsc->xtrans[j][i] = pipe->dev->image_storage.dsc.xtrans[(j + crop_y) % 6][(i + crop_x) % 6];
    }
  }
}
 
/* --- the shared geometry core ----------------------------------------------------------
 *
 * rawprepare's geometry is the fixed sensor border trim: a translation by the top/left margin,
 * and an output shrunk by the total margin. Expressed once here for the pipeline callbacks and
 * for geometry_record() (develop/geometry/geometry.h) alike.
 */
 
typedef struct dt_iop_rawprepare_geometry_t
{
  int32_t x, y, width, height;   
} dt_iop_rawprepare_geometry_t;
 
static void _rawprepare_offset(const int32_t crop_x, const int32_t crop_y, const double scale, double *dx,
                               double *dy)
{
  *dx = (double)crop_x * scale;
  *dy = (double)crop_y * scale;
}
 
static void _rawprepare_map_size(const dt_iop_rawprepare_geometry_t *const g,
                                 const dt_iop_roi_t *const roi_in, dt_iop_roi_t *roi_out)
{
  *roi_out = *roi_in;
  roi_out->x = roi_out->y = 0;
 
  const double x = g->x + g->width;
  const double y = g->y + g->height;
  const double scale = roi_in->scale;
  roi_out->width = (int)round((double)roi_out->width - x * scale);
  roi_out->height = (int)round((double)roi_out->height - y * scale);
}
 
int distort_transform(dt_iop_module_t *self, const dt_dev_pixelpipe_t *pipe, const dt_dev_pixelpipe_iop_t *piece,
                      float *const restrict points, size_t points_count)
{
  (void)self;
  (void)pipe;
  dt_iop_rawprepare_data_t *d = (dt_iop_rawprepare_data_t *)piece->data;
 
  // nothing to be done if parameters are set to neutral values (no top/left crop)
  if (d->x == 0 && d->y == 0) return 1;
 
  double x = 0.0, y = 0.0;
  _rawprepare_offset(d->x, d->y, piece->buf_in.scale, &x, &y);
  __OMP_PARALLEL_FOR_SIMD__(aligned(points:64) if(points_count > 100))
  for(size_t i = 0; i < points_count * 2; i += 2)
  {
    points[i] -= x;
    points[i + 1] -= y;
  }
  
 
  return 1;
}
 
int distort_backtransform(dt_iop_module_t *self, const dt_dev_pixelpipe_t *pipe, const dt_dev_pixelpipe_iop_t *piece,
                          float *const restrict points, size_t points_count)
{
  (void)self;
  (void)pipe;
  dt_iop_rawprepare_data_t *d = (dt_iop_rawprepare_data_t *)piece->data;
 
  // nothing to be done if parameters are set to neutral values (no top/left crop)
  if (d->x == 0 && d->y == 0) return 1;
 
  double x = 0.0, y = 0.0;
  _rawprepare_offset(d->x, d->y, piece->buf_in.scale, &x, &y);
  __OMP_PARALLEL_FOR_SIMD__(aligned(points:64) if(points_count > 100))
  for(size_t i = 0; i < points_count * 2; i += 2)
  {
    points[i] += x;
    points[i + 1] += y;
  }
  
 
  return 1;
}
 
void distort_mask(struct dt_iop_module_t *self, const struct dt_dev_pixelpipe_t *pipe,
                  struct dt_dev_pixelpipe_iop_t *piece, const float *const in, float *const out,
                  const dt_iop_roi_t *const roi_in, const dt_iop_roi_t *const roi_out)
{
  (void)self;
  (void)pipe;
  (void)piece;
  dt_iop_copy_image_roi(out, in, 1, roi_in, roi_out, TRUE);
}
 
// we're not scaling here (bayer input), so just crop borders
// see ../../doc/resizing-scaling.md for details
void modify_roi_out(dt_iop_module_t *self, const dt_dev_pixelpipe_t *pipe, dt_dev_pixelpipe_iop_t *piece,
                    dt_iop_roi_t *roi_out,
                    const dt_iop_roi_t *const roi_in)
{
  dt_iop_rawprepare_data_t *d = (dt_iop_rawprepare_data_t *)piece->data;
 
  const dt_iop_rawprepare_geometry_t geo = { d->x, d->y, d->width, d->height };
  _rawprepare_map_size(&geo, roi_in, roi_out);
 
  /* Rawprepare changes the CFA phase according to the effective crop on the current ROI scale.
   * That contract cannot be authored at history resync time because `piece->roi_in` is not known yet.
   * Bind the crop-dependent descriptor fields here, when ROI planning has provided the real input ROI. */
  _update_output_cfa_descriptor(pipe, piece, roi_in, &piece->dsc_out);
}
 
// see ../../doc/resizing-scaling.md for details
void modify_roi_in(dt_iop_module_t *self, const dt_dev_pixelpipe_t *pipe, dt_dev_pixelpipe_iop_t *piece,
                   const dt_iop_roi_t *const roi_out,
                   dt_iop_roi_t *roi_in)
{
  *roi_in = *roi_out;
  dt_iop_rawprepare_data_t *d = (dt_iop_rawprepare_data_t *)piece->data;
 
  const double x = d->x + d->width;
  const double y = d->y + d->height;
  const double scale = roi_in->scale;
  roi_in->width = (int)round((double)roi_in->width + x * scale);
  roi_in->height = (int)round((double)roi_in->height + y * scale);
 
  /* Same reasoning as in modify_roi_out(): the CFA/X-Trans descriptor depends on the input ROI scale,
   * so finalize it here once the upstream ROI has been computed. */
  _update_output_cfa_descriptor(pipe, piece, roi_in, &piece->dsc_out);
}
 
void output_format(dt_iop_module_t *self, dt_dev_pixelpipe_t *pipe, dt_dev_pixelpipe_iop_t *piece,
                   dt_iop_buffer_dsc_t *dsc)
{
  default_output_format(self, pipe, piece, dsc);
  _update_output_cfa_descriptor(pipe, piece, &piece->roi_in, &piece->dsc_out);
}
 
 
void input_format(dt_iop_module_t *self, dt_dev_pixelpipe_t *pipe, dt_dev_pixelpipe_iop_t *piece,
                   dt_iop_buffer_dsc_t *dsc)
{
  memcpy(dsc, &pipe->dev->image_storage.dsc, sizeof(dt_iop_buffer_dsc_t));
}
 
 
static inline __attribute__((always_inline)) int BL(const dt_iop_roi_t *const roi_out,
                                                    const int row, const int col, 
                                                    const int32_t x, const int32_t y)
{
  return ((((row + roi_out->y + y) & 1) << 1) + ((col + roi_out->x + x) & 1));
}
 
void autoset(struct dt_iop_module_t *self, const struct dt_dev_pixelpipe_t *pipe,
             const struct dt_dev_pixelpipe_iop_t *piece, const void *input)
{
  dt_iop_rawprepare_params_t *p = (dt_iop_rawprepare_params_t *)self->params;
  const dt_iop_roi_t *const roi_out = &piece->roi_out;
  const dt_iop_roi_t *const roi_in = &piece->roi_in;
  const int csx = compute_proper_crop(piece, roi_in, p->x);
  const int csy = compute_proper_crop(piece, roi_in, p->y);
 
  if(piece->dsc_in.filters && piece->dsc_in.channels == 1 && piece->dsc_in.datatype == TYPE_UINT16)
  {
    const uint16_t *const restrict in = (const uint16_t *const restrict)input;
    // 4 channels : R, G sampled on R rows, G sampled on B rows, B.
    int max_RGB[4] = { 0 };
 
    __OMP_PARALLEL_FOR__(reduction(max: max_RGB[:4]) collapse(2))
    for(int i = 0; i < roi_out->height; i++)
      for(int j = 0; j < roi_out->width; j++)
      {
        const size_t channel = BL(roi_out, i, j, p->x, p->y);
        const size_t pin = roi_in->width * (i + csy) + j + csx;
        const int pixel = in[pin];
        max_RGB[channel] = MAX(max_RGB[channel], pixel);
      }
 
    p->raw_white_point = MAX(MAX(max_RGB[0], MAX(max_RGB[1], MAX(max_RGB[2], max_RGB[3]))), pipe->dev->image_storage.raw_white_point);
  }
  // Do we need to handle float mosaiced images and non-mosaiced (sRAW) images too ?
}
 
/* Some comments about the cpu code path; tests with gcc 10.x show a clear performance gain for the
   compile generated code vs SSE specific code. This depends slightly on the cpu but it's 1.2 to 3-fold
   better for all tested cases.
*/
__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;
  const dt_iop_rawprepare_data_t *const d = (dt_iop_rawprepare_data_t *)piece->data;
  const int width = roi_out->width;
  const int height = roi_out->height;
  const int input_width = roi_in->width;
  const int roi_x = roi_out->x;
  const int roi_y = roi_out->y;
  const int cfa_x = roi_x + d->x;
  const int cfa_y = roi_y + d->y;
  // fprintf(stderr, "roi in %d %d %d %d\n", roi_in->x, roi_in->y, roi_in->width, roi_in->height);
  // fprintf(stderr, "roi out %d %d %d %d\n", roi_out->x, roi_out->y, roi_out->width, roi_out->height);
 
  const int csx = compute_proper_crop(piece, roi_in, d->x);
  const int csy = compute_proper_crop(piece, roi_in, d->y);
 
  if(piece->dsc_in.filters && piece->dsc_in.channels == 1
     && piece->dsc_in.datatype == TYPE_UINT16)
  { // raw mosaic
 
    const uint16_t *const in = (const uint16_t *const)ivoid;
    float *const out = (float *const)ovoid;
    const int x_phase = cfa_x & 1;
    float inv_div[4];
    for(int k = 0; k < 4; k++) inv_div[k] = 1.0f / d->div[k];
    __OMP_PARALLEL_FOR__()
    for(int j = 0; j < height; j++)
    {
      /* Keep the 2-sensel Bayer period explicit per row, but use scalar
         normalization so the compiler can choose its own vectorization. */
      const size_t pin = (size_t)input_width * (j + csy) + csx;
      const size_t pout = (size_t)j * width;
      const int row_phase = ((j + cfa_y) & 1) << 1;
      const int id0 = row_phase + x_phase;
      const int id1 = row_phase + (x_phase ^ 1);
      const float sub0 = d->sub[id0];
      const float sub1 = d->sub[id1];
      const float inv0 = inv_div[id0];
      const float inv1 = inv_div[id1];
      int i = 0;
 
      for(; i + 1 < width; i += 2)
      {
        out[pout + i + 0] = ((float)in[pin + i + 0] - sub0) * inv0;
        out[pout + i + 1] = ((float)in[pin + i + 1] - sub1) * inv1;
      }
 
      for(; i < width; i++)
      {
        const int id = row_phase + ((x_phase + i) & 1);
        out[pout + i] = ((float)in[pin + i] - d->sub[id]) * inv_div[id];
      }
    }
    
  }
  else if(piece->dsc_in.filters && piece->dsc_in.channels == 1
          && piece->dsc_in.datatype == TYPE_FLOAT)
  { // raw mosaic, fp, unnormalized
 
    const float *const in = (const float *const)ivoid;
    float *const out = (float *const)ovoid;
    const int x_phase = cfa_x & 1;
    float inv_div[4];
    for(int k = 0; k < 4; k++) inv_div[k] = 1.0f / d->div[k];
 
    __OMP_PARALLEL_FOR__()
    for(int j = 0; j < height; j++)
    {
      /* Keep the 2-sensel Bayer period explicit per row, but use scalar
         normalization so the compiler can choose its own vectorization. */
      const size_t pin = (size_t)input_width * (j + csy) + csx;
      const size_t pout = (size_t)j * width;
      const int row_phase = ((j + cfa_y) & 1) << 1;
      const int id0 = row_phase + x_phase;
      const int id1 = row_phase + (x_phase ^ 1);
      const float sub0 = d->sub[id0];
      const float sub1 = d->sub[id1];
      const float inv0 = inv_div[id0];
      const float inv1 = inv_div[id1];
      int i = 0;
 
      for(; i + 1 < width; i += 2)
      {
        out[pout + i + 0] = (in[pin + i + 0] - sub0) * inv0;
        out[pout + i + 1] = (in[pin + i + 1] - sub1) * inv1;
      }
 
      for(; i < width; i++)
      {
        const int id = row_phase + ((x_phase + i) & 1);
        out[pout + i] = (in[pin + i] - d->sub[id]) * inv_div[id];
      }
    }
    
  }
  else
  { // pre-downsampled buffer that needs black/white scaling
 
    const float *const in = (const float *const)ivoid;
    float *const out = (float *const)ovoid;
 
    const float sub = d->sub[0];
    const float div = d->div[0];
 
    const int ch = piece->dsc_in.channels;
    __OMP_PARALLEL_FOR__(collapse(3))
    for(int j = 0; j < height; j++)
    {
      for(int i = 0; i < width; i++)
      {
        for(int c = 0; c < ch; c++)
        {
          const size_t pin = (size_t)ch * (input_width * (j + csy) + csx + i) + c;
          const size_t pout = (size_t)ch * (j * width + i) + c;
 
          out[pout] = (in[pin] - sub) / div;
        }
      }
    }
    
  }
 
  if(piece->dsc_in.filters && piece->dsc_in.channels == 1 && d->apply_gainmaps)
  {
    const uint32_t map_w = d->gainmaps[0]->map_points_h;
    const uint32_t map_h = d->gainmaps[0]->map_points_v;
    const float im_to_rel_x = 1.0f / piece->buf_in.width;
    const float im_to_rel_y = 1.0f / piece->buf_in.height;
    const float rel_to_map_x = 1.0f / d->gainmaps[0]->map_spacing_h;
    const float rel_to_map_y = 1.0f / d->gainmaps[0]->map_spacing_v;
    const float map_origin_h = d->gainmaps[0]->map_origin_h;
    const float map_origin_v = d->gainmaps[0]->map_origin_v;
    float *const out = (float *const)ovoid;
    __OMP_PARALLEL_FOR__()
    for(int j = 0; j < height; j++)
    {
      const float y_map = CLAMP(((roi_y + csy + j) * im_to_rel_y - map_origin_v) * rel_to_map_y, 0, map_h);
      const uint32_t y_i0 = MIN(y_map, map_h - 1);
      const uint32_t y_i1 = MIN(y_i0 + 1, map_h - 1);
      const float y_frac = y_map - y_i0;
      const float * restrict map_row0[4];
      const float * restrict map_row1[4];
      for(int f = 0; f < 4; f++)
      {
        map_row0[f] = &d->gainmaps[f]->map_gain[y_i0 * map_w];
        map_row1[f] = &d->gainmaps[f]->map_gain[y_i1 * map_w];
      }
      for(int i = 0; i < width; i++)
      {
        const int id = BL(roi_out, j, i, d->x, d->y);
        const float x_map = CLAMP(((roi_x + csx + i) * im_to_rel_x - map_origin_h) * rel_to_map_x, 0, map_w);
        const uint32_t x_i0 = MIN(x_map, map_w - 1);
        const uint32_t x_i1 = MIN(x_i0 + 1, map_w - 1);
        const float x_frac = x_map - x_i0;
        const float gain_top = (1.0f - x_frac) * map_row0[id][x_i0] + x_frac * map_row0[id][x_i1];
        const float gain_bottom = (1.0f - x_frac) * map_row1[id][x_i0] + x_frac * map_row1[id][x_i1];
        out[j * width + i] *= (1.0f - y_frac) * gain_top + y_frac * gain_bottom;
      }
    }
    
  }
 
  return 0;
}
 
#ifdef HAVE_OPENCL
int process_cl(dt_iop_module_t *self, const dt_dev_pixelpipe_t *pipe, const dt_dev_pixelpipe_iop_t *piece,
               cl_mem dev_in, cl_mem dev_out)
{
  const dt_iop_roi_t *const roi_in = &piece->roi_in;
  const dt_iop_roi_t *const roi_out = &piece->roi_out;
  dt_iop_rawprepare_data_t *d = (dt_iop_rawprepare_data_t *)piece->data;
  dt_iop_rawprepare_global_data_t *gd = (dt_iop_rawprepare_global_data_t *)self->global_data;
 
  // Scanner DNGs and already-demosaiced files have no CFA; OpenCL path assumes CFA.
  if(piece->dsc_in.filters == 0) return FALSE;
 
  const int devid = pipe->devid;
  cl_mem dev_sub = NULL;
  cl_mem dev_div = NULL;
  cl_mem dev_gainmap[4] = {0};
  cl_int err = -999;
 
  int kernel = -1;
  gboolean gainmap_args = FALSE;
 
  // Monochrome raws (no CFA) also come in as 1 channel; use the 1f kernels regardless of filters.
  if(piece->dsc_in.channels == 1 && piece->dsc_in.datatype == TYPE_UINT16)
  {
    if(d->apply_gainmaps)
    {
      kernel = gd->kernel_rawprepare_1f_gainmap;
      gainmap_args = TRUE;
    }
    else
    {
      kernel = gd->kernel_rawprepare_1f;
    }
  }
  else if(piece->dsc_in.channels == 1 && piece->dsc_in.datatype == TYPE_FLOAT)
  {
    if(d->apply_gainmaps)
    {
      kernel = gd->kernel_rawprepare_1f_unnormalized_gainmap;
      gainmap_args = TRUE;
    }
    else
    {
      kernel = gd->kernel_rawprepare_1f_unnormalized;
    }
  }
  else
  {
    kernel = gd->kernel_rawprepare_4f;
  }
 
  const int csx = compute_proper_crop(piece, roi_in, d->x), csy = compute_proper_crop(piece, roi_in, d->y);
 
  dev_sub = dt_opencl_copy_host_to_device_constant(devid, sizeof(float) * 4, d->sub);
  if(IS_NULL_PTR(dev_sub)) goto error;
 
  dev_div = dt_opencl_copy_host_to_device_constant(devid, sizeof(float) * 4, d->div);
  if(IS_NULL_PTR(dev_div)) goto error;
 
  const int width = roi_out->width;
  const int height = roi_out->height;
 
  size_t sizes[] = { ROUNDUPDWD(width, devid), ROUNDUPDHT(height, devid), 1 };
  dt_opencl_set_kernel_arg(devid, kernel, 0, sizeof(cl_mem), (void *)&dev_in);
  dt_opencl_set_kernel_arg(devid, kernel, 1, sizeof(cl_mem), (void *)&dev_out);
  dt_opencl_set_kernel_arg(devid, kernel, 2, sizeof(int), (void *)&(width));
  dt_opencl_set_kernel_arg(devid, kernel, 3, sizeof(int), (void *)&(height));
  dt_opencl_set_kernel_arg(devid, kernel, 4, sizeof(int), (void *)&csx);
  dt_opencl_set_kernel_arg(devid, kernel, 5, sizeof(int), (void *)&csy);
  dt_opencl_set_kernel_arg(devid, kernel, 6, sizeof(cl_mem), (void *)&dev_sub);
  dt_opencl_set_kernel_arg(devid, kernel, 7, sizeof(cl_mem), (void *)&dev_div);
  dt_opencl_set_kernel_arg(devid, kernel, 8, sizeof(uint32_t), (void *)&roi_out->x);
  dt_opencl_set_kernel_arg(devid, kernel, 9, sizeof(uint32_t), (void *)&roi_out->y);
  if(gainmap_args)
  {
    const int map_size[2] = { d->gainmaps[0]->map_points_h, d->gainmaps[0]->map_points_v };
    const float im_to_rel[2] = { 1.0 / piece->buf_in.width, 1.0 / piece->buf_in.height };
    const float rel_to_map[2] = { 1.0 / d->gainmaps[0]->map_spacing_h, 1.0 / d->gainmaps[0]->map_spacing_v };
    const float map_origin[2] = { d->gainmaps[0]->map_origin_h, d->gainmaps[0]->map_origin_v };
 
    for(int i = 0; i < 4; i++)
    {
      dev_gainmap[i] = dt_opencl_alloc_device(devid, map_size[0], map_size[1], sizeof(float));
      if(dev_gainmap[i] == NULL) goto error;
      err = dt_opencl_write_host_to_device(devid, d->gainmaps[i]->map_gain, dev_gainmap[i],
                                           map_size[0], map_size[1], sizeof(float));
      if(err != CL_SUCCESS) goto error;
    }
 
    dt_opencl_set_kernel_arg(devid, kernel, 10, sizeof(cl_mem), &dev_gainmap[0]);
    dt_opencl_set_kernel_arg(devid, kernel, 11, sizeof(cl_mem), &dev_gainmap[1]);
    dt_opencl_set_kernel_arg(devid, kernel, 12, sizeof(cl_mem), &dev_gainmap[2]);
    dt_opencl_set_kernel_arg(devid, kernel, 13, sizeof(cl_mem), &dev_gainmap[3]);
    dt_opencl_set_kernel_arg(devid, kernel, 14, 2 * sizeof(int), &map_size);
    dt_opencl_set_kernel_arg(devid, kernel, 15, 2 * sizeof(float), &im_to_rel);
    dt_opencl_set_kernel_arg(devid, kernel, 16, 2 * sizeof(float), &rel_to_map);
    dt_opencl_set_kernel_arg(devid, kernel, 17, 2 * sizeof(float), &map_origin);
  }
  err = dt_opencl_enqueue_kernel_2d(devid, kernel, sizes);
  if(err != CL_SUCCESS) goto error;
 
  dt_opencl_release_mem_object(dev_sub);
  dt_opencl_release_mem_object(dev_div);
  for(int i = 0; i < 4; i++) dt_opencl_release_mem_object(dev_gainmap[i]);
 
  return TRUE;
 
error:
  dt_opencl_release_mem_object(dev_sub);
  dt_opencl_release_mem_object(dev_div);
  for(int i = 0; i < 4; i++) dt_opencl_release_mem_object(dev_gainmap[i]);
  dt_print(DT_DEBUG_OPENCL, "[opencl_rawprepare] couldn't enqueue kernel! %d\n", err);
  return FALSE;
}
#endif
 
static int image_is_normalized(const dt_image_t *const image)
{
  // if raw with floating-point data, if not special magic whitelevel, then it needs normalization
  if(dt_image_is_hdr(image))
  {
    union {
        float f;
        uint32_t u;
    } normalized;
    normalized.f = 1.0f;
 
    // dng spec is just broken here.
    return image->raw_white_point == normalized.u;
  }
 
  // else, assume normalized
  return image->dsc.channels == 1 && image->dsc.datatype == TYPE_FLOAT;
}
 
static gboolean image_set_rawcrops(const int32_t imgid, int dx, int dy)
{
  if(imgid <= 0) return FALSE;
 
  dt_image_t *img = NULL;
  img = dt_image_cache_get(imgid, 'r');
  if(IS_NULL_PTR(img)) return FALSE;
  const gboolean test = (img->p_width == img->width - dx)
                     && (img->p_height == img->height - dy);
 
  dt_image_cache_read_release(img);
  if(test) return FALSE;
 
  img = dt_image_cache_get(imgid, 'w');
  if(IS_NULL_PTR(img)) return FALSE;
  img->p_width = img->width - dx;
  img->p_height = img->height - dy;
  dt_image_cache_write_release(img, DT_IMAGE_CACHE_RELAXED);
  return TRUE;
}
 
// Check whether `image` carries GainMaps of the exact type we can apply here. We may reject some
// GainMaps that are valid per the Adobe DNG spec but that we do not support.
//
// The caller MUST hold an image-cache read lock on `image` for the whole call, and for as long as it
// then uses `gainmaps_out`: both the list and the maps it points at belong to the cache entry, which
// frees them in `dt_image_cache_deallocate()` on eviction. Never pass `&dev->image_storage` here --
// that is a SHALLOW struct copy taken by `_dt_dev_refresh_image_storage()`, which drops the cache
// lock on the line after copying, so its `dng_gain_maps` is a borrowed pointer with no owner. Walking
// it while a thumbnail run evicted the entry is the segfault in Sentry 140771917.
static gboolean _validate_gain_maps(const dt_image_t *const image, dt_dng_gain_map_t **gainmaps_out)
{
  dt_dng_gain_map_t *gainmaps[4] = {0};
 
  if(g_list_length(image->dng_gain_maps) != 4)
    return FALSE;
 
  for(int i = 0; i < 4; i++)
  {
    // check that each GainMap applies to one filter of a Bayer image,
    // covers the entire image, and is not a 1x1 no-op
    dt_dng_gain_map_t *g = (dt_dng_gain_map_t *)g_list_nth_data(image->dng_gain_maps, i);
    if(IS_NULL_PTR(g) ||
      g->plane != 0 || g->planes != 1 || g->map_planes != 1 ||
      g->row_pitch != 2 || g->col_pitch != 2 ||
      g->map_points_v < 2 || g->map_points_h < 2 ||
      g->top > 1 || g->left > 1 ||
      g->bottom != image->height || g->right != image->width)
      return FALSE;
    /* The grid dimensions come from the file and are not cross-checked against the number of gain
     * values the opcode actually carried, so a truncated or malformed DNG can declare a grid larger
     * than its payload. Everything downstream then indexes map_gain[] as map_points_h * map_points_v
     * -- process() reads row map_points_v - 1, and _own_gainmaps() memcpy's that many floats out of a
     * source block that only ever held map_gain_count. Reject the map instead of reading past it. */
    if((uint64_t)g->map_points_h * (uint64_t)g->map_points_v > (uint64_t)g->map_gain_count)
      return FALSE;
    uint32_t filter = ((g->top & 1) << 1) + (g->left & 1);
    gainmaps[filter] = g;
  }
 
  // check that there is a GainMap for each filter of the Bayer pattern
  if(gainmaps[0] == NULL || gainmaps[1] == NULL || gainmaps[2] == NULL || gainmaps[3] == NULL)
    return FALSE;
 
  // check that each GainMap has the same shape
  for(int i = 1; i < 4; i++)
  {
    if(gainmaps[i]->map_points_h != gainmaps[0]->map_points_h ||
      gainmaps[i]->map_points_v != gainmaps[0]->map_points_v ||
      gainmaps[i]->map_spacing_h != gainmaps[0]->map_spacing_h ||
      gainmaps[i]->map_spacing_v != gainmaps[0]->map_spacing_v ||
      gainmaps[i]->map_origin_h != gainmaps[0]->map_origin_h ||
      gainmaps[i]->map_origin_v != gainmaps[0]->map_origin_v)
      return FALSE;
  }
 
  if(gainmaps_out)
    memcpy(gainmaps_out, gainmaps, sizeof(gainmaps));
 
  return TRUE;
}
 
// Does this image carry GainMaps we can apply? Answers against the cache entry, under its read lock,
// and lets no pointer escape.
gboolean check_gain_maps(const int32_t imgid)
{
  const dt_image_t *const image = dt_image_cache_get(imgid, 'r');
  if(IS_NULL_PTR(image)) return FALSE;
 
  const gboolean supported = _validate_gain_maps(image, NULL);
  dt_image_cache_read_release(image);
  return supported;
}
 
// Free the pipe-owned deep copies of the GainMaps held in piece->data (see commit_params).
static void _free_owned_gainmaps(dt_iop_rawprepare_data_t *d)
{
  for(int i = 0; i < 4; i++)
  {
    g_free(d->gainmaps[i]);
    d->gainmaps[i] = NULL;
  }
}
 
// Build pipe-owned deep copies of the image's GainMaps into d->gainmaps[]. The maps belong to the
// image-cache entry, which frees them on eviction (thumbnail-generation pressure is enough), so
// `process()` must not read them directly -- that was Sentry #129880848, fixed by owning a private
// copy. The validation and the copy must BOTH happen under the cache read lock taken here: doing the
// copy after releasing it just moves the same use-after-free into the memcpy, and validating through
// `dev->image_storage` -- a shallow copy whose lock was dropped when it was made -- moves it into the
// list walk, which is Sentry 140771917.
// Caller must have freed/NULLed d->gainmaps[] first. Returns TRUE only if all four were copied.
static gboolean _own_gainmaps(const int32_t imgid, dt_iop_rawprepare_data_t *d)
{
  const dt_image_t *const image = dt_image_cache_get(imgid, 'r');
  if(IS_NULL_PTR(image)) return FALSE;
 
  dt_dng_gain_map_t *src[4] = { 0 };
  gboolean owned = _validate_gain_maps(image, src);
 
  for(int i = 0; owned && i < 4; i++)
  {
    // map_planes == 1 is enforced by _validate_gain_maps(), so the gain payload is
    // map_points_h * map_points_v floats -- exactly what process() indexes.
    const size_t sz = sizeof(dt_dng_gain_map_t)
                      + (size_t)src[i]->map_points_h * src[i]->map_points_v * sizeof(float);
    d->gainmaps[i] = g_malloc(sz);
    if(IS_NULL_PTR(d->gainmaps[i]))
    {
      _free_owned_gainmaps(d); // partial copy (OOM): drop everything, don't apply
      owned = FALSE;
      break;
    }
    memcpy(d->gainmaps[i], src[i], sz);
  }
 
  dt_image_cache_read_release(image);
  return owned;
}
 
void commit_params(dt_iop_module_t *self, dt_iop_params_t *params, dt_dev_pixelpipe_t *pipe,
                   dt_dev_pixelpipe_iop_t *piece)
{
  const dt_iop_rawprepare_params_t *const p = (dt_iop_rawprepare_params_t *)params;
  dt_iop_rawprepare_data_t *d = (dt_iop_rawprepare_data_t *)piece->data;
  const dt_image_t *const img = &pipe->dev->image_storage;
 
  d->x = p->x;
  d->y = p->y;
  d->width = p->width;
  d->height = p->height;
 
  if(img->dsc.filters)
  {
    const float white = (float)p->raw_white_point;
 
    for(int i = 0; i < 4; i++)
    {
      d->sub[i] = (float)p->raw_black_level_separate[i];
      d->div[i] = (white - d->sub[i]);
    }
  }
  else
  {
    const float normalizer
        = dt_image_is_hdr(&pipe->dev->image_storage) ? 1.0f : (float)UINT16_MAX;
    const float white = (float)p->raw_white_point / normalizer;
    float black = 0;
    for(int i = 0; i < 4; i++)
    {
      black += p->raw_black_level_separate[i] / normalizer;
    }
    black /= 4.0f;
 
    for(int i = 0; i < 4; i++)
    {
      d->sub[i] = black;
      d->div[i] = (white - black);
    }
  }
 
  float black = 0.0f;
  for(uint8_t i = 0; i < 4; i++)
  {
    black += (float)p->raw_black_level_separate[i];
  }
  d->rawprepare.raw_black_level = (uint16_t)(black / 4.0f);
  d->rawprepare.raw_white_point = p->raw_white_point;
  piece->dsc_out.rawprepare.raw_black_level = d->rawprepare.raw_black_level;
  piece->dsc_out.rawprepare.raw_white_point = d->rawprepare.raw_white_point;
  for(int k = 0; k < 4; k++) piece->dsc_out.processed_maximum[k] = 1.0f;
 
  // Refresh our private, pipe-owned copy of the GainMaps (see _own_gainmaps() for why we must own
  // them rather than borrow image_storage.dng_gain_maps). Free the previous copy in all cases.
  _free_owned_gainmaps(d);
  d->apply_gainmaps = (p->flat_field == FLAT_FIELD_EMBEDDED) && _own_gainmaps(img->id, d);
 
  if(image_set_rawcrops(pipe->dev->image_storage.id, d->x + d->width, d->y + d->height))
    DT_DEBUG_CONTROL_SIGNAL_RAISE(dt_control_signal_get_global(), DT_SIGNAL_METADATA_UPDATE);
 
  // Image-type gating (needs_rawprepare && !normalized) is handled at history level by
  // enable()/force_enable()/reload_defaults(); it is no longer duplicated here.
 
  // OpenCL path only for RAW, single-channel, CFA images (runtime/format decision).
  const gboolean cl_ok = (img->dsc.cst == IOP_CS_RAW && img->dsc.channels == 1 && img->dsc.filters);
  if(!cl_ok) piece->process_cl_ready = FALSE;
 
  dt_iop_fmt_log(self, "commit: class=%s needs_rawprepare=%d normalized=%d cl_ok=%d -> enabled=%d",
                 dt_image_pipe_class_name(dt_image_pipe_class(img)),
                 dt_image_needs_rawprepare(img), image_is_normalized(img), cl_ok, piece->enabled);
}
 
void init_pipe(dt_iop_module_t *self, dt_dev_pixelpipe_t *pipe, dt_dev_pixelpipe_iop_t *piece)
{
  piece->data = dt_calloc_align(sizeof(dt_iop_rawprepare_data_t));
  piece->data_size = sizeof(dt_iop_rawprepare_data_t);
}
 
/* --- the geometry service's view of this module (develop/geometry/geometry.h) --------- */
 
static void _rawprepare_geometry_map_size(const void *data, const dt_iop_roi_t *const in, dt_iop_roi_t *out)
{
  _rawprepare_map_size((const dt_iop_rawprepare_geometry_t *)data, in, out);
}
 
static int _rawprepare_geometry_transform(const void *data, const dt_geometry_record_t *const record,
                                          dt_geometry_chain_t *chain, float *points, size_t points_count)
{
  const dt_iop_rawprepare_geometry_t *const g = (const dt_iop_rawprepare_geometry_t *)data;
  if(g->x == 0 && g->y == 0) return 1;
 
  double x = 0.0, y = 0.0;
  _rawprepare_offset(g->x, g->y, record->in.scale, &x, &y);
  for(size_t i = 0; i < points_count * 2; i += 2)
  {
    points[i] -= x;
    points[i + 1] -= y;
  }
  return 1;
}
 
static int _rawprepare_geometry_backtransform(const void *data, const dt_geometry_record_t *const record,
                                              dt_geometry_chain_t *chain, float *points, size_t points_count)
{
  const dt_iop_rawprepare_geometry_t *const g = (const dt_iop_rawprepare_geometry_t *)data;
  if(g->x == 0 && g->y == 0) return 1;
 
  double x = 0.0, y = 0.0;
  _rawprepare_offset(g->x, g->y, record->in.scale, &x, &y);
  for(size_t i = 0; i < points_count * 2; i += 2)
  {
    points[i] += x;
    points[i + 1] += y;
  }
  return 1;
}
 
static const dt_geometry_vtable_t _rawprepare_geometry_vtable = {
  .map_size = _rawprepare_geometry_map_size,
  .transform = _rawprepare_geometry_transform,
  .backtransform = _rawprepare_geometry_backtransform,
};
 
gboolean geometry_record(dt_iop_module_t *self, const void *params, dt_geometry_record_t *record)
{
  const dt_iop_rawprepare_params_t *const p = (const dt_iop_rawprepare_params_t *)params;
 
  dt_iop_rawprepare_geometry_t *data
      = (dt_iop_rawprepare_geometry_t *)g_malloc0(sizeof(dt_iop_rawprepare_geometry_t));
  if(IS_NULL_PTR(data)) return FALSE;
 
  /* commit_params() copies these four verbatim; nothing else about them is derived. */
  data->x = p->x;
  data->y = p->y;
  data->width = p->width;
  data->height = p->height;
 
  record->data = data;
  record->free_data = dt_free_gpointer;
  record->vtable = &_rawprepare_geometry_vtable;
  return TRUE;
}
 
void cleanup_pipe(dt_iop_module_t *self, dt_dev_pixelpipe_t *pipe, dt_dev_pixelpipe_iop_t *piece)
{
  dt_iop_rawprepare_data_t *d = (dt_iop_rawprepare_data_t *)piece->data;
  if(d) _free_owned_gainmaps(d);
  dt_free_align(piece->data);
  piece->data = NULL;
}
 
 
static gboolean enable(const dt_image_t *image)
{
  // rawprepare (black/white normalization) applies to any raw colorimetry: mosaiced raw and
  // already-demosaiced sRAW / linear DNG alike, unless the buffer is already normalized.
  return dt_image_needs_rawprepare(image)
          && !image_is_normalized(image);
}
 
gboolean force_enable(struct dt_iop_module_t *self, const gboolean current_state)
{
  // History sanitization: rawprepare must be off for non-raw or already-normalized buffers,
  // decided here from image metadata using the same enable() rule as reload_defaults().
  const gboolean active = enable(&self->dev->image_storage);
  const gboolean state = active && current_state;
  dt_iop_fmt_log(self, "force_enable: class=%s supported=%d current=%d -> %d",
                 dt_image_pipe_class_name(dt_image_pipe_class(&self->dev->image_storage)),
                 active, current_state, state);
  return state;
}
 
void reload_defaults(dt_iop_module_t *self)
{
  dt_iop_rawprepare_params_t *d = self->default_params;
  const dt_image_t *const image = &(self->dev->image_storage);
 
  // if there are embedded GainMaps, they should be applied by default to avoid uneven color cast
  gboolean has_gainmaps = check_gain_maps(self->dev->image_storage.id);
 
  *d = (dt_iop_rawprepare_params_t){.x = image->crop_x,
                                    .y = image->crop_y,
                                    .width = image->crop_width,
                                    .height = image->crop_height,
                                    .raw_black_level_separate[0] = image->raw_black_level_separate[0],
                                    .raw_black_level_separate[1] = image->raw_black_level_separate[1],
                                    .raw_black_level_separate[2] = image->raw_black_level_separate[2],
                                    .raw_black_level_separate[3] = image->raw_black_level_separate[3],
                                    .raw_white_point = image->raw_white_point,
                                    .flat_field = has_gainmaps ? FLAT_FIELD_EMBEDDED : FLAT_FIELD_OFF };
 
  self->hide_enable_button = 1;
  self->default_enabled = enable(image);
 
  dt_image_print_debug_info(image, "rawprepare.reload_defaults");
  // Stack visibility (raw vs non_raw) is applied from default_enabled in gui_update(), on the GUI
  // thread when the widget exists; reload_defaults() stays params-only.
}
 
void init_global(dt_iop_module_so_t *self)
{
  const int program = 2; // basic.cl, from programs.conf
  self->data = malloc(sizeof(dt_iop_rawprepare_global_data_t));
 
  dt_iop_rawprepare_global_data_t *gd = self->data;
  gd->kernel_rawprepare_1f = dt_opencl_create_kernel(program, "rawprepare_1f");
  gd->kernel_rawprepare_1f_gainmap = dt_opencl_create_kernel(program, "rawprepare_1f_gainmap");
  gd->kernel_rawprepare_1f_unnormalized = dt_opencl_create_kernel(program, "rawprepare_1f_unnormalized");
  gd->kernel_rawprepare_1f_unnormalized_gainmap = dt_opencl_create_kernel(program, "rawprepare_1f_unnormalized_gainmap");
  gd->kernel_rawprepare_4f = dt_opencl_create_kernel(program, "rawprepare_4f");
}
 
void cleanup_global(dt_iop_module_so_t *self)
{
  dt_iop_rawprepare_global_data_t *gd = (dt_iop_rawprepare_global_data_t *)self->data;
  dt_opencl_free_kernel(gd->kernel_rawprepare_4f);
  dt_opencl_free_kernel(gd->kernel_rawprepare_1f_unnormalized);
  dt_opencl_free_kernel(gd->kernel_rawprepare_1f);
  dt_free(self->data);
}
 
void gui_update(dt_iop_module_t *self)
{
  dt_iop_rawprepare_gui_data_t *g = (dt_iop_rawprepare_gui_data_t *)dt_iop_gui_data(self);
  dt_iop_rawprepare_params_t *p = (dt_iop_rawprepare_params_t *)self->params;
 
  const gboolean is_monochrome = (self->dev->image_storage.flags & (DT_IMAGE_MONOCHROME | DT_IMAGE_MONOCHROME_BAYER)) != 0;
  if(is_monochrome)
  {
    // we might have to deal with old edits, so get average first
    int av = 2; // for rounding
    for(int i = 0; i < 4; i++)
      av += p->raw_black_level_separate[i];
 
    for(int i = 0; i < 4; i++)
      dt_bauhaus_slider_set(g->black_level_separate[i], av / 4);
  }
 
  // don't show upper three black levels for monochromes
  for(int i = 1; i < 4; i++)
    gtk_widget_set_visible(g->black_level_separate[i], !is_monochrome);
 
  gtk_widget_set_visible(g->flat_field, check_gain_maps(self->dev->image_storage.id));
  dt_bauhaus_combobox_set(g->flat_field, p->flat_field);
 
  // raw vs non_raw page, from the per-image default computed by reload_defaults()
  gtk_stack_set_visible_child_name(GTK_STACK(self->gui->widget), self->default_enabled ? "raw" : "non_raw");
}
 
void gui_changed(dt_iop_module_t *self, GtkWidget *w, void *previous)
{
  dt_iop_rawprepare_gui_data_t *g = (dt_iop_rawprepare_gui_data_t *)dt_iop_gui_data(self);
  dt_iop_rawprepare_params_t *p = (dt_iop_rawprepare_params_t *)self->params;
 
  const gboolean is_monochrome = (self->dev->image_storage.flags & (DT_IMAGE_MONOCHROME | DT_IMAGE_MONOCHROME_BAYER)) != 0;
  if(is_monochrome)
  {
    if(w == g->black_level_separate[0])
    {
      const int val = p->raw_black_level_separate[0];
      for(int i = 1; i < 4; i++)
        dt_bauhaus_slider_set(g->black_level_separate[i], val);
    }
  }
}
 
const gchar *black_label[]
  =  { N_("black level 0"),
       N_("black level 1"),
       N_("black level 2"),
       N_("black level 3") };
 
void gui_init(dt_iop_module_t *self)
{
  dt_iop_rawprepare_gui_data_t *g = IOP_GUI_ALLOC(rawprepare);
 
  GtkWidget *box_raw = self->gui->widget = gtk_box_new(GTK_ORIENTATION_VERTICAL, DT_GUI_BOX_SPACING);
 
  for(int i = 0; i < 4; i++)
  {
    gchar *par = g_strdup_printf("raw_black_level_separate[%i]", i);
 
    g->black_level_separate[i] = dt_bauhaus_slider_from_params(self, par);
    dt_bauhaus_widget_set_label(g->black_level_separate[i], black_label[i]);
    gtk_widget_set_tooltip_text(g->black_level_separate[i], _(black_label[i]));
    dt_bauhaus_slider_set_soft_max(g->black_level_separate[i], 16384);
 
    dt_free(par);
  }
 
  g->white_point = dt_bauhaus_slider_from_params(self, "raw_white_point");
  gtk_widget_set_tooltip_text(g->white_point, _("white point"));
  dt_bauhaus_slider_set_soft_max(g->white_point, 16384);
 
  g->flat_field = dt_bauhaus_combobox_from_params(self, "flat_field");
  gtk_widget_set_tooltip_text(g->flat_field, _("flat field correction to compensate for lens shading"));
 
  gtk_box_pack_start(GTK_BOX(self->gui->widget),
                      dt_ui_section_label_new(_("In-camera crop")), FALSE, FALSE, 0);
 
  g->x = dt_bauhaus_slider_from_params(self, "x");
  gtk_widget_set_tooltip_text(g->x, _("crop from left border"));
  dt_bauhaus_slider_set_soft_max(g->x, 256);
 
  g->y = dt_bauhaus_slider_from_params(self, "y");
  gtk_widget_set_tooltip_text(g->y, _("crop from top"));
  dt_bauhaus_slider_set_soft_max(g->y, 256);
 
  g->width = dt_bauhaus_slider_from_params(self, "width");
  gtk_widget_set_tooltip_text(g->width, _("crop from right border"));
  dt_bauhaus_slider_set_soft_max(g->width, 256);
 
  g->height = dt_bauhaus_slider_from_params(self, "height");
  gtk_widget_set_tooltip_text(g->height, _("crop from bottom"));
  dt_bauhaus_slider_set_soft_max(g->height, 256);
 
  // start building top level widget
  self->gui->widget = gtk_stack_new();
  gtk_stack_set_homogeneous(GTK_STACK(self->gui->widget), FALSE);
 
  GtkWidget *label_non_raw = dt_ui_label_new(_("raw black/white point correction\nonly works for the sensors that need it."));
 
  gtk_stack_add_named(GTK_STACK(self->gui->widget), label_non_raw, "non_raw");
  gtk_stack_add_named(GTK_STACK(self->gui->widget), box_raw, "raw");
}
 
// 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