sharpen.c

Go to the documentation of this file.

/*
    This file is part of darktable,
    Copyright (C) 2009-2013, 2016 johannes hanika.
    Copyright (C) 2010-2011 Bruce Guenter.
    Copyright (C) 2010-2011 Henrik Andersson.
    Copyright (C) 2010 jan rinze.
    Copyright (C) 2010-2011 Pascal de Bruijn.
    Copyright (C) 2011 Antony Dovgal.
    Copyright (C) 2011 Jérémy Rosen.
    Copyright (C) 2011 Kanstantsin Shautsou.
    Copyright (C) 2011 Olivier Tribout.
    Copyright (C) 2011 Robert Bieber.
    Copyright (C) 2011 Rostyslav Pidgornyi.
    Copyright (C) 2011-2014, 2016, 2019 Tobias Ellinghaus.
    Copyright (C) 2011-2012, 2014, 2016-2017 Ulrich Pegelow.
    Copyright (C) 2012 Richard Wonka.
    Copyright (C) 2013 Simon Spannagel.
    Copyright (C) 2014, 2018, 2020, 2022 Pascal Obry.
    Copyright (C) 2014-2016 Roman Lebedev.
    Copyright (C) 2015 Edouard Gomez.
    Copyright (C) 2015 Pedro Côrte-Real.
    Copyright (C) 2017 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 rawfiner.
    Copyright (C) 2019 Andreas Schneider.
    Copyright (C) 2020 Aldric Renaudin.
    Copyright (C) 2020, 2022 Diederik Ter Rahe.
    Copyright (C) 2020 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 "system/macros.h"
#include "system/mem_alloc.h"
#include "common/module_versioning.h"
#include "common/logging.h"
#include "system/openmp.h"
#include "system/simd.h"
#include "system/target_clones.h"
#include "caches/pixelpipe_cache_alloc.h"
#include "config.h"
#endif
#include "widgets/bauhaus.h"
#include "common/imagebuf.h"
#include "common/opencl.h"
#include "develop/develop.h"
#include "develop/imageop.h"
#include "develop/imageop_math.h"
#include "develop/imageop_gui.h"
#include "develop/tiling.h"
 
#include "gui/presets.h"
#include "iop/iop_api.h"
#include <assert.h>
#include <math.h>
#include <stdlib.h>
#include <string.h>
 
#include <gtk/gtk.h>
#include <inttypes.h>
 
DT_MODULE_INTROSPECTION(1, dt_iop_sharpen_params_t)
 
#define MAXR 12
 
typedef struct dt_iop_sharpen_params_t
{
  float radius;    // $MIN: 0.0 $MAX: 99.0 $DEFAULT: 2.0
  float amount;    // $MIN: 0.0 $MAX: 2.0 $DEFAULT: 0.5
  float threshold; // $MIN: 0.0 $MAX: 100.0 $DEFAULT: 0.5
} dt_iop_sharpen_params_t;
 
typedef struct dt_iop_sharpen_gui_data_t
{
  GtkWidget *radius, *amount, *threshold;
} dt_iop_sharpen_gui_data_t;
 
typedef struct dt_iop_sharpen_data_t
{
  float radius, amount, threshold;
} dt_iop_sharpen_data_t;
 
typedef struct dt_iop_sharpen_global_data_t
{
  int kernel_sharpen_hblur;
  int kernel_sharpen_vblur;
  int kernel_sharpen_mix;
} dt_iop_sharpen_global_data_t;
 
 
const char *name()
{
  return C_("modulename", "sharpen");
}
 
int default_group()
{
  return IOP_GROUP_SHARPNESS;
}
 
int flags()
{
  return IOP_FLAGS_SUPPORTS_BLENDING | IOP_FLAGS_ALLOW_TILING | 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;
}
 
const char **description(struct dt_iop_module_t *self)
{
  return dt_iop_set_description(self, _("sharpen the details in the image using a standard UnSharp Mask (USM)"),
                                      _("corrective"),
                                      _("linear or non-linear, Lab, display or scene-referred"),
                                      _("frequential, Lab"),
                                      _("quasi-linear, Lab, display or scene-referred"));
}
 
void init_presets(dt_iop_module_so_t *self)
{
  dt_iop_sharpen_params_t tmp = (dt_iop_sharpen_params_t){ 2.0, 0.5, 0.5 };
  // add the preset.
  dt_gui_presets_add_generic(_("sharpen"), self->op,
                             self->version(), &tmp, sizeof(dt_iop_sharpen_params_t), 1);
  // restrict to raw images
  dt_gui_presets_update_ldr(_("sharpen"), self->op,
                            self->version(), FOR_RAW);
}
 
__DT_CLONE_TARGETS__
static float *const init_gaussian_kernel(const int rad, const size_t mat_size, const float sigma2)
{
  float weight = 0.0f;
  float *const mat = dt_pixelpipe_cache_alloc_align_float_cache(mat_size, 0);
  if(IS_NULL_PTR(mat)) return NULL;
  memset(mat, 0, sizeof(float) * mat_size);
  for(int l = -rad; l <= rad; l++) weight += mat[l + rad] = expf(-l * l / (2.f * sigma2));
  for(int l = -rad; l <= rad; l++) mat[l + rad] /= weight;
  return mat;
}
 
#ifdef HAVE_OPENCL
int process_cl(struct 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;
  dt_iop_sharpen_data_t *d = (dt_iop_sharpen_data_t *)piece->data;
  dt_iop_sharpen_global_data_t *gd = (dt_iop_sharpen_global_data_t *)self->global_data;
  cl_mem dev_m = NULL;
  cl_mem dev_tmp = NULL;
  cl_int err = -999;
 
  const int devid = pipe->devid;
  const int width = roi_in->width;
  const int height = roi_in->height;
  const int rad = MIN(MAXR, ceilf(d->radius * roi_in->scale));
  const int wd = 2 * rad + 1;
  float *mat = NULL;
 
  if(rad == 0)
  {
    size_t origin[] = { 0, 0, 0 };
    size_t region[] = { width, height, 1 };
    err = dt_opencl_enqueue_copy_image(devid, dev_in, dev_out, origin, origin, region);
    if(err != CL_SUCCESS) goto error;
    return TRUE;
  }
 
  // special case handling: very small image with one or two dimensions below 2*rad+1 => no sharpening,
  // normally not needed for OpenCL but implemented here for identity with CPU code path
  if(width < 2 * rad + 1 || height < 2 * rad + 1)
  {
    size_t origin[] = { 0, 0, 0 };
    size_t region[] = { width, height, 1 };
    err = dt_opencl_enqueue_copy_image(devid, dev_in, dev_out, origin, origin, region);
    if(err != CL_SUCCESS) goto error;
    return TRUE;
  }
 
  const float sigma2 = (1.0f / (2.5 * 2.5)) * (d->radius * roi_in->scale)
                       * (d->radius * roi_in->scale);
  mat = init_gaussian_kernel(rad, wd, sigma2);
  if(IS_NULL_PTR(mat)) goto error;
 
  int hblocksize;
  dt_opencl_local_buffer_t hlocopt
    = (dt_opencl_local_buffer_t){ .xoffset = 2 * rad, .xfactor = 1, .yoffset = 0, .yfactor = 1,
                                  .cellsize = sizeof(float), .overhead = 0,
                                  .sizex = 1 << 16, .sizey = 1 };
 
  if(dt_opencl_local_buffer_opt(devid, gd->kernel_sharpen_hblur, &hlocopt))
    hblocksize = hlocopt.sizex;
  else
    hblocksize = 1;
 
  int vblocksize;
  dt_opencl_local_buffer_t vlocopt
    = (dt_opencl_local_buffer_t){ .xoffset = 1, .xfactor = 1, .yoffset = 2 * rad, .yfactor = 1,
                                  .cellsize = sizeof(float), .overhead = 0,
                                  .sizex = 1, .sizey = 1 << 16 };
 
  if(dt_opencl_local_buffer_opt(devid, gd->kernel_sharpen_vblur, &vlocopt))
    vblocksize = vlocopt.sizey;
  else
    vblocksize = 1;
 
 
  const size_t bwidth = ROUNDUP(width, hblocksize);
  const size_t bheight = ROUNDUP(height, vblocksize);
 
  size_t sizes[3];
  size_t local[3];
 
  dev_tmp = dt_opencl_alloc_device(devid, width, height, sizeof(float) * 4);
  if(IS_NULL_PTR(dev_tmp)) goto error;
 
  dev_m = dt_opencl_copy_host_to_device_constant(devid, sizeof(float) * wd, mat);
  if(IS_NULL_PTR(dev_m)) goto error;
 
  /* horizontal blur */
  sizes[0] = bwidth;
  sizes[1] = ROUNDUPDHT(height, devid);
  sizes[2] = 1;
  local[0] = hblocksize;
  local[1] = 1;
  local[2] = 1;
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_hblur, 0, sizeof(cl_mem), (void *)&dev_in);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_hblur, 1, sizeof(cl_mem), (void *)&dev_out);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_hblur, 2, sizeof(cl_mem), (void *)&dev_m);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_hblur, 3, sizeof(int), (void *)&rad);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_hblur, 4, sizeof(int), (void *)&width);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_hblur, 5, sizeof(int), (void *)&height);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_hblur, 6, sizeof(int), (void *)&hblocksize);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_hblur, 7, (hblocksize + 2 * rad) * sizeof(float), NULL);
  err = dt_opencl_enqueue_kernel_2d_with_local(devid, gd->kernel_sharpen_hblur, sizes, local);
  if(err != CL_SUCCESS) goto error;
 
  /* vertical blur */
  sizes[0] = ROUNDUPDWD(width, devid);
  sizes[1] = bheight;
  sizes[2] = 1;
  local[0] = 1;
  local[1] = vblocksize;
  local[2] = 1;
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_vblur, 0, sizeof(cl_mem), (void *)&dev_out);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_vblur, 1, sizeof(cl_mem), (void *)&dev_tmp);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_vblur, 2, sizeof(cl_mem), (void *)&dev_m);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_vblur, 3, sizeof(int), (void *)&rad);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_vblur, 4, sizeof(int), (void *)&width);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_vblur, 5, sizeof(int), (void *)&height);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_vblur, 6, sizeof(int), (void *)&vblocksize);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_vblur, 7, (vblocksize + 2 * rad) * sizeof(float), NULL);
  err = dt_opencl_enqueue_kernel_2d_with_local(devid, gd->kernel_sharpen_vblur, sizes, local);
  if(err != CL_SUCCESS) goto error;
 
  /* mixing tmp and in -> out */
  sizes[0] = ROUNDUPDWD(width, devid);
  sizes[1] = ROUNDUPDHT(height, devid);
  sizes[2] = 1;
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_mix, 0, sizeof(cl_mem), (void *)&dev_in);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_mix, 1, sizeof(cl_mem), (void *)&dev_tmp);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_mix, 2, sizeof(cl_mem), (void *)&dev_out);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_mix, 3, sizeof(int), (void *)&width);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_mix, 4, sizeof(int), (void *)&height);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_mix, 5, sizeof(float), (void *)&d->amount);
  dt_opencl_set_kernel_arg(devid, gd->kernel_sharpen_mix, 6, sizeof(float), (void *)&d->threshold);
  err = dt_opencl_enqueue_kernel_2d(devid, gd->kernel_sharpen_mix, sizes);
  if(err != CL_SUCCESS) goto error;
 
  dt_opencl_release_mem_object(dev_m);
  dt_opencl_release_mem_object(dev_tmp);
  dt_pixelpipe_cache_free_align(mat);
  return TRUE;
 
error:
  dt_opencl_release_mem_object(dev_m);
  dt_opencl_release_mem_object(dev_tmp);
  dt_pixelpipe_cache_free_align(mat);
  dt_print(DT_DEBUG_OPENCL, "[opencl_sharpen] couldn't enqueue kernel! %d\n", err);
  return FALSE;
}
#endif
 
 
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;
  dt_iop_sharpen_data_t *d = (dt_iop_sharpen_data_t *)piece->data;
  const int rad = MIN(MAXR, ceilf(d->radius * roi_in->scale));
 
  tiling->factor = 2.1f; // in + out + tmprow
  tiling->factor_cl = 3.0f; // in + out + tmp
  tiling->maxbuf = 1.0f;
  tiling->overhead = 0;
  tiling->overlap = rad;
  tiling->xalign = 1;
  tiling->yalign = 1;
  return;
}
 
__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_sharpen_data_t *const data = (dt_iop_sharpen_data_t *)piece->data;
  const int rad = MIN(MAXR, ceilf(data->radius * roi_in->scale));
  // Special case handling: very small image with one or two dimensions below 2*rad+1 treat as no sharpening and just
  // pass through.  This avoids handling of all kinds of border cases below.
  if(rad == 0 ||
     (roi_out->width < 2 * rad + 1 || roi_out->height < 2 * rad + 1))
  {
    dt_iop_image_copy_by_size(ovoid, ivoid, roi_out->width, roi_out->height, 4);
    return 0;
  }
 
  float *restrict tmp;  // one row per thread
  size_t padded_size;
  if (dt_iop_alloc_image_buffers(self, roi_in, roi_out,
                                  1 | DT_IMGSZ_WIDTH | DT_IMGSZ_PERTHREAD, &tmp, &padded_size,
                                  0))
  {
    dt_iop_copy_image_roi(ovoid, ivoid, 4, roi_in, roi_out, TRUE);
    return 1;
  }
 
  const int wd = 2 * rad + 1;
  const int wd4 = (wd & 3) ? (wd >> 2) + 1 : wd >> 2;
 
  const size_t mat_size = (size_t)4 * wd4;
  const float sigma2 = (1.0f / (2.5 * 2.5)) * (data->radius * roi_in->scale)
                       * (data->radius * roi_in->scale);
  float *const mat = init_gaussian_kernel(rad, mat_size, sigma2);
  if(IS_NULL_PTR(mat))
  {
    dt_pixelpipe_cache_free_align(tmp);
    dt_iop_copy_image_roi(ovoid, ivoid, 4, roi_in, roi_out, TRUE);
    return 1;
  }
 
  const float *const restrict in = (float*)ivoid;
  const size_t width = roi_out->width;
  __OMP_PARALLEL_FOR__()
  for(int j = 0; j < roi_out->height; j++)
  {
    // We skip the top and bottom 'rad' rows because the kernel would extend beyond the edge of the image, resulting
    // in an incomplete summation.
    if (j < rad || j >= roi_out->height - rad)
    {
      // fill in the top/bottom border with unchanged luma values from the input image.
      const float *const restrict row_in = in + (size_t)4 * j * width;
      float *const restrict row_out = ((float*)ovoid) + (size_t)4 * j * width;
      memcpy(row_out, row_in, 4 * sizeof(float) * width);
      continue;
    }
    // Get a thread-local temporary buffer for processing the current row of the image.
    float *const restrict temp_buf = dt_get_perthread(tmp, padded_size);
    // vertically blur the pixels of the current row into the temp buffer
    const size_t start_row = j-rad;
    const size_t end_row = j+rad;
    // do the bulk of the row four at a time
    for(int i = 0; i < width; i += 4)
    {
      dt_aligned_pixel_t sum = { 0.0f };
      for(int k = start_row; k <= end_row; k++)
      {
        const int k_adj = k - (j-rad);
        for_four_channels(c,aligned(in))
          sum[c] += mat[k_adj] * in[4*(k*width+i+c)];
      }
      float *const vblurred = temp_buf + i;
      for_four_channels(c,aligned(vblurred))
        vblurred[c] = sum[c];
    }
    // do the leftover 0-3 pixels of the row
    for(int i = width & ~3; i < width; i++)
    {
      float sum = 0.0f;
      for(int k = start_row; k <= end_row; k++)
      {
        const int k_adj = k - (j-rad);
        sum += mat[k_adj] * in[4*(k*width+i)];
      }
      temp_buf[i] = sum;
    }
 
    // now horizontally blur the already vertically-blurred pixels from the temp buffer to the final output buffer
    // we can skip the left-most and right-most pixels for the same reason as we skipped the top and bottom borders.
    float *const restrict row_out = ((float*)ovoid) + (size_t)4 * j * width;
    for(int i = 0; i < rad; i++)
      copy_pixel(row_out + 4*i, in + 4*(j*width+i));  //copy unsharpened border pixel
    const float threshold = data->threshold;
    const float amount = data->amount;
    for(int i = rad; i < roi_out->width - rad; i++)
    {
      float sum = 0.0f;
      for(int k = i-rad; k <= i+rad; k++)
      {
        const int k_adj = k - (i-rad);
        sum += mat[k_adj] * temp_buf[k];
      }
      // subtract the blurred pixel's luma from the original input pixel's luma
      const size_t index = 4 * (j * width + i);
      const float diff = in[index] - sum;
      const float absdiff = fabs(diff);
      const float detail = (absdiff > threshold) ? copysignf(MAX(absdiff - threshold, 0.0f), diff) : 0.0f;
      row_out[4*i] = in[index] + detail * amount;
      row_out[4*i + 1] = in[index + 1];
      row_out[4*i + 2] = in[index + 2];
    }
    for(int i = roi_out->width - rad; i < roi_out->width; i++)
      copy_pixel(row_out + 4*i, in + 4*(j*width+i));  //copy unsharpened border pixel
  }
 
  dt_pixelpipe_cache_free_align(mat);
  dt_pixelpipe_cache_free_align(tmp);
 
  if(pipe->mask_display & DT_DEV_PIXELPIPE_DISPLAY_MASK)
    dt_iop_alpha_copy(ivoid, ovoid, roi_out->width, roi_out->height);
  return 0;
}
 
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_sharpen_params_t *p = (dt_iop_sharpen_params_t *)p1;
  dt_iop_sharpen_data_t *d = (dt_iop_sharpen_data_t *)piece->data;
 
  // actually need to increase the mask to fit 2.5 sigma inside
  d->radius = 2.5f * p->radius;
  d->amount = p->amount;
  d->threshold = p->threshold;
}
 
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_sharpen_data_t));
  piece->data_size = sizeof(dt_iop_sharpen_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 init_global(dt_iop_module_so_t *module)
{
  const int program = 7; // sharpen.cl, from programs.conf
  dt_iop_sharpen_global_data_t *gd
      = (dt_iop_sharpen_global_data_t *)calloc(1, sizeof(dt_iop_sharpen_global_data_t));
  if(IS_NULL_PTR(gd)) return;
  module->data = gd;
  gd->kernel_sharpen_hblur = dt_opencl_create_kernel(program, "sharpen_hblur");
  gd->kernel_sharpen_vblur = dt_opencl_create_kernel(program, "sharpen_vblur");
  gd->kernel_sharpen_mix = dt_opencl_create_kernel(program, "sharpen_mix");
}
 
void cleanup_global(dt_iop_module_so_t *module)
{
  dt_iop_sharpen_global_data_t *gd = (dt_iop_sharpen_global_data_t *)module->data;
  dt_opencl_free_kernel(gd->kernel_sharpen_hblur);
  dt_opencl_free_kernel(gd->kernel_sharpen_vblur);
  dt_opencl_free_kernel(gd->kernel_sharpen_mix);
  dt_free(module->data);
}
 
void gui_init(struct dt_iop_module_t *self)
{
  dt_iop_sharpen_gui_data_t *g = IOP_GUI_ALLOC(sharpen);
 
  g->radius = dt_bauhaus_slider_from_params(self, N_("radius"));
  dt_bauhaus_slider_set_soft_max(g->radius, 8.0);
  dt_bauhaus_slider_set_digits(g->radius, 3);
  gtk_widget_set_tooltip_text(g->radius, _("spatial extent of the unblurring"));
 
  g->amount = dt_bauhaus_slider_from_params(self, N_("amount"));
  dt_bauhaus_slider_set_digits(g->amount, 3);
  gtk_widget_set_tooltip_text(g->amount, _("strength of the sharpen"));
 
  g->threshold = dt_bauhaus_slider_from_params(self, N_("threshold"));
  dt_bauhaus_slider_set_digits(g->threshold, 3);
  gtk_widget_set_tooltip_text(g->threshold, _("threshold to activate sharpen"));
}
 
#undef MAXR
 
// 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