diff --git a/kmeans.c b/kmeans.c
index f596009..7ee273d 100644
--- a/kmeans.c
+++ b/kmeans.c
@@ -1,20 +1,6 @@
/*
-© 2011-2016 by Kornel Lesiński.
-
-This file is part of libimagequant.
-
-libimagequant 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.
-
-libimagequant 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 libimagequant. If not, see .
+** © 2011-2016 by Kornel Lesiński.
+** See COPYRIGHT file for license.
*/
#include "libimagequant.h"
@@ -80,14 +66,14 @@ LIQ_PRIVATE void kmeans_finalize(colormap *map, const unsigned int max_threads,
LIQ_PRIVATE double kmeans_do_iteration(histogram *hist, colormap *const map, kmeans_callback callback)
{
const unsigned int max_threads = omp_get_max_threads();
- kmeans_state average_color[(KMEANS_CACHE_LINE_GAP+map->colors) * max_threads];
+ LIQ_ARRAY(kmeans_state, average_color, (KMEANS_CACHE_LINE_GAP+map->colors) * max_threads);
kmeans_init(map, max_threads, average_color);
struct nearest_map *const n = nearest_init(map);
hist_item *const achv = hist->achv;
const int hist_size = hist->size;
double total_diff=0;
- #pragma omp parallel for if (hist_size > 3000) \
+ #pragma omp parallel for if (hist_size > 2000) \
schedule(static) default(none) shared(average_color,callback) reduction(+:total_diff)
for(int j=0; j < hist_size; j++) {
float diff;
diff --git a/libimagequant.c b/libimagequant.c
index 08f6962..3506564 100644
--- a/libimagequant.c
+++ b/libimagequant.c
@@ -1,31 +1,9 @@
/*
-** © 2009-2016 by Kornel Lesiński.
+** © 2009-2018 by Kornel Lesiński.
+** © 1989, 1991 by Jef Poskanzer.
+** © 1997, 2000, 2002 by Greg Roelofs; based on an idea by Stefan Schneider.
**
-** This file is part of libimagequant.
-**
-** libimagequant 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.
-**
-** libimagequant 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 libimagequant. If not, see .
-*/
-/* Copyright (C) 1989, 1991 by Jef Poskanzer.
-** Copyright (C) 1997, 2000, 2002 by Greg Roelofs; based on an idea by
-** Stefan Schneider.
-**
-** Permission to use, copy, modify, and distribute this software and its
-** documentation for any purpose and without fee is hereby granted, provided
-** that the above copyright notice appear in all copies and that both that
-** copyright notice and this permission notice appear in supporting
-** documentation. This software is provided "as is" without express or
-** implied warranty.
+** See COPYRIGHT file for license.
*/
#include
@@ -43,7 +21,9 @@
#ifdef _OPENMP
#include
+#define LIQ_TEMP_ROW_WIDTH(img_width) (((img_width) | 15) + 1) /* keep alignment & leave space between rows to avoid cache line contention */
#else
+#define LIQ_TEMP_ROW_WIDTH(img_width) (img_width)
#define omp_get_max_threads() 1
#define omp_get_thread_num() 0
#endif
@@ -78,7 +58,8 @@ struct liq_attr {
unsigned int max_colors, max_histogram_entries;
unsigned int min_posterization_output /* user setting */, min_posterization_input /* speed setting */;
unsigned int kmeans_iterations, feedback_loop_trials;
- bool last_index_transparent, use_contrast_maps, use_dither_map;
+ bool last_index_transparent, use_contrast_maps;
+ unsigned char use_dither_map;
unsigned char speed;
unsigned char progress_stage1, progress_stage2, progress_stage3;
@@ -100,11 +81,12 @@ struct liq_image {
rgba_pixel **rows;
double gamma;
unsigned int width, height;
- unsigned char *noise, *edges, *dither_map;
+ unsigned char *importance_map, *edges, *dither_map;
rgba_pixel *pixels, *temp_row;
f_pixel *temp_f_row;
liq_image_get_rgba_row_callback *row_callback;
void *row_callback_user_info;
+ liq_image *background;
float min_opaque_val;
f_pixel fixed_colors[256];
unsigned short fixed_colors_count;
@@ -124,7 +106,8 @@ typedef struct liq_remapping_result {
liq_palette int_palette;
double gamma, palette_error;
float dither_level;
- bool use_dither_map; unsigned char progress_stage1;
+ unsigned char use_dither_map;
+ unsigned char progress_stage1;
} liq_remapping_result;
struct liq_result {
@@ -141,7 +124,7 @@ struct liq_result {
float dither_level;
double gamma, palette_error;
int min_posterization_output;
- bool use_dither_map;
+ unsigned char use_dither_map;
};
struct liq_histogram {
@@ -161,6 +144,7 @@ static void modify_alpha(liq_image *input_image, rgba_pixel *const row_pixels) L
static void contrast_maps(liq_image *image) LIQ_NONNULL;
static liq_error finalize_histogram(liq_histogram *input_hist, liq_attr *options, histogram **hist_output) LIQ_NONNULL;
static const rgba_pixel *liq_image_get_row_rgba(liq_image *input_image, unsigned int row) LIQ_NONNULL;
+static bool liq_image_get_row_f_init(liq_image *img) LIQ_NONNULL;
static const f_pixel *liq_image_get_row_f(liq_image *input_image, unsigned int row) LIQ_NONNULL;
static void liq_remapping_result_destroy(liq_remapping_result *result) LIQ_NONNULL;
static liq_error pngquant_quantize(histogram *hist, const liq_attr *options, const int fixed_colors_count, const f_pixel fixed_colors[], const double gamma, bool fixed_result_colors, liq_result **) LIQ_NONNULL;
@@ -174,7 +158,7 @@ LIQ_NONNULL static void liq_verbose_printf(const liq_attr *context, const char *
int required_space = vsnprintf(NULL, 0, fmt, va)+1; // +\0
va_end(va);
- char buf[required_space];
+ LIQ_ARRAY(char, buf, required_space);
va_start(va, fmt);
vsnprintf(buf, required_space, fmt, va);
va_end(va);
@@ -210,8 +194,13 @@ LIQ_NONNULL static bool liq_remap_progress(const liq_remapping_result *quant, co
#if USE_SSE
inline static bool is_sse_available()
{
-#if (defined(__x86_64__) || defined(__amd64))
+#if (defined(__x86_64__) || defined(__amd64) || defined(_WIN64))
return true;
+#elif _MSC_VER
+ int info[4];
+ __cpuid(info, 1);
+ /* bool is implemented as a built-in type of size 1 in MSVC */
+ return info[3] & (1<<26) ? true : false;
#else
int a,b,c,d;
cpuid(1, a, b, c, d);
@@ -348,7 +337,8 @@ LIQ_EXPORT LIQ_NONNULL liq_error liq_set_speed(liq_attr* attr, int speed)
if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return LIQ_INVALID_POINTER;
if (speed < 1 || speed > 10) return LIQ_VALUE_OUT_OF_RANGE;
- unsigned int iterations = MAX(8-speed, 0); iterations += iterations * iterations/2;
+ unsigned int iterations = MAX(8-speed, 0);
+ iterations += iterations * iterations/2;
attr->kmeans_iterations = iterations;
attr->kmeans_iteration_limit = 1.0/(double)(1<<(23-speed));
attr->feedback_loop_trials = MAX(56-9*speed, 0);
@@ -356,11 +346,16 @@ LIQ_EXPORT LIQ_NONNULL liq_error liq_set_speed(liq_attr* attr, int speed)
attr->max_histogram_entries = (1<<17) + (1<<18)*(10-speed);
attr->min_posterization_input = (speed >= 8) ? 1 : 0;
attr->use_dither_map = (speed <= (omp_get_max_threads() > 1 ? 7 : 5)); // parallelized dither map might speed up floyd remapping
+ if (attr->use_dither_map && speed < 3) {
+ attr->use_dither_map = 2; // always
+ }
attr->use_contrast_maps = (speed <= 7) || attr->use_dither_map;
attr->speed = speed;
attr->progress_stage1 = attr->use_contrast_maps ? 20 : 8;
- if (attr->feedback_loop_trials < 2) attr->progress_stage1 += 30;
+ if (attr->feedback_loop_trials < 2) {
+ attr->progress_stage1 += 30;
+ }
attr->progress_stage3 = 50 / (1+speed);
attr->progress_stage2 = 100 - attr->progress_stage1 - attr->progress_stage3;
return LIQ_OK;
@@ -400,7 +395,7 @@ LIQ_EXPORT LIQ_NONNULL int liq_get_min_opacity(const liq_attr *attr)
{
if (!CHECK_STRUCT_TYPE(attr, liq_attr)) return -1;
- return MIN(255, 256.0 * attr->min_opaque_val);
+ return MIN(255.f, 256.f * attr->min_opaque_val);
}
LIQ_EXPORT LIQ_NONNULL void liq_set_last_index_transparent(liq_attr* attr, int is_last)
@@ -520,7 +515,7 @@ LIQ_EXPORT liq_attr* liq_attr_create_with_allocator(void* (*custom_malloc)(size_
.target_mse = 0,
.max_mse = MAX_DIFF,
};
- liq_set_speed(attr, 3);
+ liq_set_speed(attr, 4);
return attr;
}
@@ -540,7 +535,7 @@ LIQ_EXPORT LIQ_NONNULL liq_error liq_image_add_fixed_color(liq_image *img, liq_c
return LIQ_OK;
}
-LIQ_NONNULL static liq_error liq_histogram_add_fixed_color_internal(liq_histogram *hist, f_pixel color)
+LIQ_NONNULL static liq_error liq_histogram_add_fixed_color_f(liq_histogram *hist, f_pixel color)
{
if (hist->fixed_colors_count > 255) return LIQ_UNSUPPORTED;
@@ -548,9 +543,24 @@ LIQ_NONNULL static liq_error liq_histogram_add_fixed_color_internal(liq_histogra
return LIQ_OK;
}
+LIQ_EXPORT LIQ_NONNULL liq_error liq_histogram_add_fixed_color(liq_histogram *hist, liq_color color, double gamma)
+{
+ if (!CHECK_STRUCT_TYPE(hist, liq_histogram)) return LIQ_INVALID_POINTER;
+
+ float gamma_lut[256];
+ to_f_set_gamma(gamma_lut, gamma ? gamma : 0.45455);
+ const f_pixel px = rgba_to_f(gamma_lut, (rgba_pixel){
+ .r = color.r,
+ .g = color.g,
+ .b = color.b,
+ .a = color.a,
+ });
+ return liq_histogram_add_fixed_color_f(hist, px);
+}
+
LIQ_NONNULL static bool liq_image_use_low_memory(liq_image *img)
{
- img->temp_f_row = img->malloc(sizeof(img->f_pixels[0]) * img->width * omp_get_max_threads());
+ img->temp_f_row = img->malloc(sizeof(img->f_pixels[0]) * LIQ_TEMP_ROW_WIDTH(img->width) * omp_get_max_threads());
return img->temp_f_row != NULL;
}
@@ -586,7 +596,7 @@ static liq_image *liq_image_create_internal(const liq_attr *attr, rgba_pixel* ro
};
if (!rows || attr->min_opaque_val < 1.f) {
- img->temp_row = attr->malloc(sizeof(img->temp_row[0]) * width * omp_get_max_threads());
+ img->temp_row = attr->malloc(sizeof(img->temp_row[0]) * LIQ_TEMP_ROW_WIDTH(width) * omp_get_max_threads());
if (!img->temp_row) return NULL;
}
@@ -630,6 +640,57 @@ LIQ_EXPORT LIQ_NONNULL liq_error liq_image_set_memory_ownership(liq_image *img,
return LIQ_OK;
}
+LIQ_NONNULL static void liq_image_free_maps(liq_image *input_image);
+LIQ_NONNULL static void liq_image_free_importance_map(liq_image *input_image);
+
+LIQ_EXPORT LIQ_NONNULL liq_error liq_image_set_importance_map(liq_image *img, unsigned char importance_map[], size_t buffer_size, enum liq_ownership ownership) {
+ if (!CHECK_STRUCT_TYPE(img, liq_image)) return LIQ_INVALID_POINTER;
+ if (!CHECK_USER_POINTER(importance_map)) return LIQ_INVALID_POINTER;
+
+ const size_t required_size = img->width * img->height;
+ if (buffer_size < required_size) {
+ return LIQ_BUFFER_TOO_SMALL;
+ }
+
+ if (ownership == LIQ_COPY_PIXELS) {
+ unsigned char *tmp = img->malloc(required_size);
+ if (!tmp) {
+ return LIQ_OUT_OF_MEMORY;
+ }
+ memcpy(tmp, importance_map, required_size);
+ importance_map = tmp;
+ } else if (ownership != LIQ_OWN_PIXELS) {
+ return LIQ_UNSUPPORTED;
+ }
+
+ liq_image_free_importance_map(img);
+ img->importance_map = importance_map;
+
+ return LIQ_OK;
+}
+
+LIQ_EXPORT LIQ_NONNULL liq_error liq_image_set_background(liq_image *img, liq_image *background)
+{
+ if (!CHECK_STRUCT_TYPE(img, liq_image)) return LIQ_INVALID_POINTER;
+ if (!CHECK_STRUCT_TYPE(background, liq_image)) return LIQ_INVALID_POINTER;
+
+ if (background->background) {
+ return LIQ_UNSUPPORTED;
+ }
+ if (img->width != background->width || img->height != background->height) {
+ return LIQ_BUFFER_TOO_SMALL;
+ }
+
+ if (img->background) {
+ liq_image_destroy(img->background);
+ }
+
+ img->background = background;
+ liq_image_free_maps(img); // Force them to be re-analyzed with the background
+
+ return LIQ_OK;
+}
+
LIQ_NONNULL static bool check_image_size(const liq_attr *attr, const int width, const int height)
{
if (!CHECK_STRUCT_TYPE(attr, liq_attr)) {
@@ -730,7 +791,7 @@ LIQ_NONNULL static const rgba_pixel *liq_image_get_row_rgba(liq_image *img, unsi
}
assert(img->temp_row);
- rgba_pixel *temp_row = img->temp_row + img->width * omp_get_thread_num();
+ rgba_pixel *temp_row = img->temp_row + LIQ_TEMP_ROW_WIDTH(img->width) * omp_get_thread_num();
if (img->rows) {
memcpy(temp_row, img->rows[row], img->width * sizeof(temp_row[0]));
} else {
@@ -753,31 +814,40 @@ LIQ_NONNULL static void convert_row_to_f(liq_image *img, f_pixel *row_f_pixels,
}
}
+LIQ_NONNULL static bool liq_image_get_row_f_init(liq_image *img)
+{
+ assert(omp_get_thread_num() == 0);
+ if (img->f_pixels) {
+ return true;
+ }
+ if (!liq_image_should_use_low_memory(img, false)) {
+ img->f_pixels = img->malloc(sizeof(img->f_pixels[0]) * img->width * img->height);
+ }
+ if (!img->f_pixels) {
+ return liq_image_use_low_memory(img);
+ }
+
+ if (!liq_image_has_rgba_pixels(img)) {
+ return false;
+ }
+
+ float gamma_lut[256];
+ to_f_set_gamma(gamma_lut, img->gamma);
+ for(unsigned int i=0; i < img->height; i++) {
+ convert_row_to_f(img, &img->f_pixels[i*img->width], i, gamma_lut);
+ }
+ return true;
+}
+
LIQ_NONNULL static const f_pixel *liq_image_get_row_f(liq_image *img, unsigned int row)
{
if (!img->f_pixels) {
- if (img->temp_f_row) {
- float gamma_lut[256];
- to_f_set_gamma(gamma_lut, img->gamma);
- f_pixel *row_for_thread = img->temp_f_row + img->width * omp_get_thread_num();
- convert_row_to_f(img, row_for_thread, row, gamma_lut);
- return row_for_thread;
- }
-
- assert(omp_get_thread_num() == 0);
- if (!liq_image_should_use_low_memory(img, false)) {
- img->f_pixels = img->malloc(sizeof(img->f_pixels[0]) * img->width * img->height);
- }
- if (!img->f_pixels) {
- if (!liq_image_use_low_memory(img)) return NULL;
- return liq_image_get_row_f(img, row);
- }
-
+ assert(img->temp_f_row); // init should have done that
float gamma_lut[256];
to_f_set_gamma(gamma_lut, img->gamma);
- for(unsigned int i=0; i < img->height; i++) {
- convert_row_to_f(img, &img->f_pixels[i*img->width], i, gamma_lut);
- }
+ f_pixel *row_for_thread = img->temp_f_row + LIQ_TEMP_ROW_WIDTH(img->width) * omp_get_thread_num();
+ convert_row_to_f(img, row_for_thread, row, gamma_lut);
+ return row_for_thread;
}
return img->f_pixels + img->width * row;
}
@@ -818,23 +888,34 @@ LIQ_NONNULL static void liq_image_free_rgba_source(liq_image *input_image)
}
}
+LIQ_NONNULL static void liq_image_free_importance_map(liq_image *input_image) {
+ if (input_image->importance_map) {
+ input_image->free(input_image->importance_map);
+ input_image->importance_map = NULL;
+ }
+}
+
+LIQ_NONNULL static void liq_image_free_maps(liq_image *input_image) {
+ liq_image_free_importance_map(input_image);
+
+ if (input_image->edges) {
+ input_image->free(input_image->edges);
+ input_image->edges = NULL;
+ }
+
+ if (input_image->dither_map) {
+ input_image->free(input_image->dither_map);
+ input_image->dither_map = NULL;
+ }
+}
+
LIQ_EXPORT LIQ_NONNULL void liq_image_destroy(liq_image *input_image)
{
if (!CHECK_STRUCT_TYPE(input_image, liq_image)) return;
liq_image_free_rgba_source(input_image);
- if (input_image->noise) {
- input_image->free(input_image->noise);
- }
-
- if (input_image->edges) {
- input_image->free(input_image->edges);
- }
-
- if (input_image->dither_map) {
- input_image->free(input_image->dither_map);
- }
+ liq_image_free_maps(input_image);
if (input_image->f_pixels) {
input_image->free(input_image->f_pixels);
@@ -848,6 +929,10 @@ LIQ_EXPORT LIQ_NONNULL void liq_image_destroy(liq_image *input_image)
input_image->free(input_image->temp_f_row);
}
+ if (input_image->background) {
+ liq_image_destroy(input_image->background);
+ }
+
input_image->magic_header = liq_freed_magic;
input_image->free(input_image);
}
@@ -1077,7 +1162,7 @@ LIQ_NONNULL static void sort_palette(colormap *map, const liq_attr *options)
*/
if (options->last_index_transparent) {
for(unsigned int i=0; i < map->colors; i++) {
- if (map->palette[i].acolor.a < 1.0/256.0) {
+ if (map->palette[i].acolor.a < 1.f/256.f) {
const unsigned int old = i, transparent_dest = map->colors-1;
SWAP_PALETTE(map, transparent_dest, old);
@@ -1100,7 +1185,7 @@ LIQ_NONNULL static void sort_palette(colormap *map, const liq_attr *options)
/* move transparent colors to the beginning to shrink trns chunk */
unsigned int num_transparent = 0;
for(unsigned int i = 0; i < non_fixed_colors; i++) {
- if (map->palette[i].acolor.a < 255.0/256.0) {
+ if (map->palette[i].acolor.a < 255.f/256.f) {
// current transparent color is swapped with earlier opaque one
if (i != num_transparent) {
SWAP_PALETTE(map, num_transparent, i);
@@ -1174,24 +1259,37 @@ LIQ_NONNULL static float remap_to_palette(liq_image *const input_image, unsigned
const unsigned int cols = input_image->width;
double remapping_error=0;
- if (!liq_image_get_row_f(input_image, 0)) { // trigger lazy conversion
+ if (!liq_image_get_row_f_init(input_image)) {
+ return -1;
+ }
+ if (input_image->background && !liq_image_get_row_f_init(input_image->background)) {
return -1;
}
+ const colormap_item *acolormap = map->palette;
+
struct nearest_map *const n = nearest_init(map);
+ const int transparent_index = input_image->background ? nearest_search(n, &(f_pixel){0,0,0,0}, 0, NULL) : 0;
+
const unsigned int max_threads = omp_get_max_threads();
- kmeans_state average_color[(KMEANS_CACHE_LINE_GAP+map->colors) * max_threads];
+ LIQ_ARRAY(kmeans_state, average_color, (KMEANS_CACHE_LINE_GAP+map->colors) * max_threads);
kmeans_init(map, max_threads, average_color);
#pragma omp parallel for if (rows*cols > 3000) \
- schedule(static) default(none) shared(average_color) reduction(+:remapping_error)
+ schedule(static) default(none) shared(acolormap) shared(average_color) reduction(+:remapping_error)
for(int row = 0; row < rows; ++row) {
const f_pixel *const row_pixels = liq_image_get_row_f(input_image, row);
+ const f_pixel *const bg_pixels = input_image->background && acolormap[transparent_index].acolor.a < 1.f/256.f ? liq_image_get_row_f(input_image->background, row) : NULL;
+
unsigned int last_match=0;
for(unsigned int col = 0; col < cols; ++col) {
float diff;
- output_pixels[row][col] = last_match = nearest_search(n, &row_pixels[col], last_match, &diff);
+ last_match = nearest_search(n, &row_pixels[col], last_match, &diff);
+ if (bg_pixels && colordifference(bg_pixels[col], acolormap[last_match].acolor) <= diff) {
+ last_match = transparent_index;
+ }
+ output_pixels[row][col] = last_match;
remapping_error += diff;
kmeans_update_color(row_pixels[col], 1.0, map, last_match, omp_get_thread_num(), average_color);
@@ -1226,25 +1324,25 @@ inline static f_pixel get_dithered_pixel(const float dither_level, const float m
else { if (px.b + sb < max_underflow) ratio = MIN(ratio, (max_underflow-px.b)/sb); }
float a = px.a + sa;
- if (a > 1.0) { a = 1.0; }
+ if (a > 1.f) { a = 1.f; }
else if (a < 0) { a = 0; }
// If dithering error is crazy high, don't propagate it that much
// This prevents crazy geen pixels popping out of the blue (or red or black! ;)
const float dither_error = sr*sr + sg*sg + sb*sb + sa*sa;
if (dither_error > max_dither_error) {
- ratio *= 0.8;
+ ratio *= 0.8f;
} else if (dither_error < 2.f/256.f/256.f) {
// don't dither areas that don't have noticeable error — makes file smaller
return px;
- }
+ }
- return (f_pixel){
- .r=px.r + sr * ratio,
- .g=px.g + sg * ratio,
- .b=px.b + sb * ratio,
- .a=a,
- };
+ return (f_pixel) {
+ .r=px.r + sr * ratio,
+ .g=px.g + sg * ratio,
+ .b=px.b + sb * ratio,
+ .a=a,
+ };
}
/**
@@ -1260,25 +1358,32 @@ LIQ_NONNULL static bool remap_to_palette_floyd(liq_image *input_image, unsigned
const colormap *map = quant->palette;
const colormap_item *acolormap = map->palette;
+ if (!liq_image_get_row_f_init(input_image)) {
+ return false;
+ }
+ if (input_image->background && !liq_image_get_row_f_init(input_image->background)) {
+ return false;
+ }
+
/* Initialize Floyd-Steinberg error vectors. */
- f_pixel *restrict thiserr, *restrict nexterr;
- const size_t errsize = (cols + 2) * sizeof(*thiserr) * 2;
- thiserr = input_image->malloc(errsize); // +2 saves from checking out of bounds access
+ const size_t errwidth = cols+2;
+ f_pixel *restrict thiserr = input_image->malloc(errwidth * sizeof(thiserr[0]) * 2); // +2 saves from checking out of bounds access
if (!thiserr) return false;
- memset(thiserr, 0, errsize);
- nexterr = thiserr + (cols + 2);
+ f_pixel *restrict nexterr = thiserr + errwidth;
+ memset(thiserr, 0, errwidth * sizeof(thiserr[0]));
bool ok = true;
struct nearest_map *const n = nearest_init(map);
+ const int transparent_index = input_image->background ? nearest_search(n, &(f_pixel){0,0,0,0}, 0, NULL) : 0;
// response to this value is non-linear and without it any value < 0.8 would give almost no dithering
float base_dithering_level = quant->dither_level;
- base_dithering_level = 1.0 - (1.0-base_dithering_level)*(1.0-base_dithering_level);
+ base_dithering_level = 1.f - (1.f-base_dithering_level)*(1.f-base_dithering_level);
if (dither_map) {
- base_dithering_level *= 1.0/255.0; // convert byte to float
+ base_dithering_level *= 1.f/255.f; // convert byte to float
}
- base_dithering_level *= 15.0/16.0; // prevent small errors from accumulating
+ base_dithering_level *= 15.f/16.f; // prevent small errors from accumulating
int fs_direction = 1;
unsigned int last_match=0;
@@ -1288,10 +1393,11 @@ LIQ_NONNULL static bool remap_to_palette_floyd(liq_image *input_image, unsigned
break;
}
- memset(nexterr, 0, (cols + 2) * sizeof(*nexterr));
+ memset(nexterr, 0, errwidth * sizeof(nexterr[0]));
int col = (fs_direction > 0) ? 0 : (cols - 1);
const f_pixel *const row_pixels = liq_image_get_row_f(input_image, row);
+ const f_pixel *const bg_pixels = input_image->background && acolormap[transparent_index].acolor.a < 1.f/256.f ? liq_image_get_row_f(input_image->background, row) : NULL;
do {
float dither_level = base_dithering_level;
@@ -1302,9 +1408,16 @@ LIQ_NONNULL static bool remap_to_palette_floyd(liq_image *input_image, unsigned
const f_pixel spx = get_dithered_pixel(dither_level, max_dither_error, thiserr[col + 1], row_pixels[col]);
const unsigned int guessed_match = output_image_is_remapped ? output_pixels[row][col] : last_match;
- output_pixels[row][col] = last_match = nearest_search(n, &spx, guessed_match, NULL);
+ float diff;
+ last_match = nearest_search(n, &spx, guessed_match, &diff);
+ f_pixel output_px = acolormap[last_match].acolor;
+ if (bg_pixels && colordifference(bg_pixels[col], output_px) <= diff) {
+ output_px = bg_pixels[col];
+ output_pixels[row][col] = transparent_index;
+ } else {
+ output_pixels[row][col] = last_match;
+ }
- const f_pixel output_px = acolormap[last_match].acolor;
f_pixel err = {
.r = (spx.r - output_px.r),
.g = (spx.g - output_px.g),
@@ -1315,10 +1428,10 @@ LIQ_NONNULL static bool remap_to_palette_floyd(liq_image *input_image, unsigned
// If dithering error is crazy high, don't propagate it that much
// This prevents crazy geen pixels popping out of the blue (or red or black! ;)
if (err.r*err.r + err.g*err.g + err.b*err.b + err.a*err.a > max_dither_error) {
- err.r *= 0.75;
- err.g *= 0.75;
- err.b *= 0.75;
- err.a *= 0.75;
+ err.r *= 0.75f;
+ err.g *= 0.75f;
+ err.b *= 0.75f;
+ err.a *= 0.75f;
}
/* Propagate Floyd-Steinberg error terms. */
@@ -1370,7 +1483,7 @@ LIQ_NONNULL static bool remap_to_palette_floyd(liq_image *input_image, unsigned
if (fs_direction > 0) {
if (col >= cols) break;
} else {
- if (col <= 0) break;
+ if (col < 0) break;
}
} while(1);
@@ -1389,7 +1502,7 @@ LIQ_NONNULL static bool remap_to_palette_floyd(liq_image *input_image, unsigned
/* fixed colors are always included in the palette, so it would be wasteful to duplicate them in palette from histogram */
LIQ_NONNULL static void remove_fixed_colors_from_histogram(histogram *hist, const int fixed_colors_count, const f_pixel fixed_colors[], const float target_mse)
{
- const float max_difference = MAX(target_mse/2.0, 2.0/256.0/256.0);
+ const float max_difference = MAX(target_mse/2.f, 2.f/256.f/256.f);
if (fixed_colors_count) {
for(int j=0; j < hist->size; j++) {
for(unsigned int i=0; i < fixed_colors_count; i++) {
@@ -1461,26 +1574,28 @@ LIQ_EXPORT LIQ_NONNULL liq_error liq_histogram_add_image(liq_histogram *input_hi
const unsigned int cols = input_image->width, rows = input_image->height;
- if (!input_image->noise && options->use_contrast_maps) {
+ if (!input_image->importance_map && options->use_contrast_maps) {
contrast_maps(input_image);
}
input_hist->gamma = input_image->gamma;
for(int i = 0; i < input_image->fixed_colors_count; i++) {
- liq_error res = liq_histogram_add_fixed_color_internal(input_hist, input_image->fixed_colors[i]);
+ liq_error res = liq_histogram_add_fixed_color_f(input_hist, input_image->fixed_colors[i]);
if (res != LIQ_OK) {
return res;
}
}
- /*
- ** Step 2: attempt to make a histogram of the colors, unclustered.
- ** If at first we don't succeed, increase ignorebits to increase color
- ** coherence and try again.
- */
+ /*
+ ** Step 2: attempt to make a histogram of the colors, unclustered.
+ ** If at first we don't succeed, increase ignorebits to increase color
+ ** coherence and try again.
+ */
- if (liq_progress(options, options->progress_stage1 * 0.4f)) return LIQ_ABORTED;
+ if (liq_progress(options, options->progress_stage1 * 0.4f)) {
+ return LIQ_ABORTED;
+ }
const bool all_rows_at_once = liq_image_can_use_rgba_rows(input_image);
@@ -1498,11 +1613,11 @@ LIQ_EXPORT LIQ_NONNULL liq_error liq_histogram_add_image(liq_histogram *input_hi
for(unsigned int row=0; row < rows; row++) {
bool added_ok;
if (all_rows_at_once) {
- added_ok = pam_computeacolorhash(input_hist->acht, (const rgba_pixel *const *)input_image->rows, cols, rows, input_image->noise);
+ added_ok = pam_computeacolorhash(input_hist->acht, (const rgba_pixel *const *)input_image->rows, cols, rows, input_image->importance_map);
if (added_ok) break;
} else {
const rgba_pixel* rows_p[1] = { liq_image_get_row_rgba(input_image, row) };
- added_ok = pam_computeacolorhash(input_hist->acht, rows_p, cols, 1, input_image->noise ? &input_image->noise[row * cols] : NULL);
+ added_ok = pam_computeacolorhash(input_hist->acht, rows_p, cols, 1, input_image->importance_map ? &input_image->importance_map[row * cols] : NULL);
}
if (!added_ok) {
input_hist->ignorebits++;
@@ -1517,10 +1632,7 @@ LIQ_EXPORT LIQ_NONNULL liq_error liq_histogram_add_image(liq_histogram *input_hi
input_hist->had_image_added = true;
- if (input_image->noise) {
- input_image->free(input_image->noise);
- input_image->noise = NULL;
- }
+ liq_image_free_importance_map(input_image);
if (input_image->free_pixels && input_image->f_pixels) {
liq_image_free_rgba_source(input_image); // bow can free the RGBA source if copy has been made in f_pixels
@@ -1578,7 +1690,7 @@ LIQ_NONNULL static void modify_alpha(liq_image *input_image, rgba_pixel *const r
/**
Builds two maps:
- noise - approximation of areas with high-frequency noise, except straight edges. 1=flat, 0=noisy.
+ importance_map - approximation of areas with high-frequency noise, except straight edges. 1=flat, 0=noisy.
edges - noise map including all edges
*/
LIQ_NONNULL static void contrast_maps(liq_image *image)
@@ -1588,14 +1700,14 @@ LIQ_NONNULL static void contrast_maps(liq_image *image)
return;
}
- unsigned char *restrict noise = image->noise ? image->noise : image->malloc(cols*rows);
- image->noise = NULL;
+ unsigned char *restrict noise = image->importance_map ? image->importance_map : image->malloc(cols*rows);
+ image->importance_map = NULL;
unsigned char *restrict edges = image->edges ? image->edges : image->malloc(cols*rows);
image->edges = NULL;
unsigned char *restrict tmp = image->malloc(cols*rows);
- if (!noise || !edges || !tmp) {
+ if (!noise || !edges || !tmp || !liq_image_get_row_f_init(image)) {
image->free(noise);
image->free(edges);
image->free(tmp);
@@ -1637,11 +1749,11 @@ LIQ_NONNULL static void contrast_maps(liq_image *image)
z = 1.f - MAX(z,MIN(horiz,vert));
z *= z; // noise is amplified
z *= z;
-
- z *= 256.f;
- noise[j*cols+i] = z < 256 ? z : 255;
- z = (1.f-edge)*256.f;
- edges[j*cols+i] = z > 0 ? (z < 256 ? z : 255) : 0;
+ // 85 is about 1/3rd of weight (not 0, because noisy pixels still need to be included, just not as precisely).
+ const unsigned int z_int = 85 + (unsigned int)(z * 171.f);
+ noise[j*cols+i] = MIN(z_int, 255);
+ const int e_int = 255 - (int)(edge * 256.f);
+ edges[j*cols+i] = e_int > 0 ? MIN(e_int, 255) : 0;
}
}
@@ -1663,7 +1775,7 @@ LIQ_NONNULL static void contrast_maps(liq_image *image)
image->free(tmp);
- image->noise = noise;
+ image->importance_map = noise;
image->edges = edges;
}
@@ -1674,7 +1786,7 @@ LIQ_NONNULL static void contrast_maps(liq_image *image)
* and peeks 1 pixel above/below. Full 2d algorithm doesn't improve it significantly.
* Correct flood fill doesn't have visually good properties.
*/
-LIQ_NONNULL static void update_dither_map(unsigned char *const *const row_pointers, liq_image *input_image)
+LIQ_NONNULL static void update_dither_map(liq_image *input_image, unsigned char *const *const row_pointers, colormap *map)
{
const unsigned int width = input_image->width;
const unsigned int height = input_image->height;
@@ -1686,6 +1798,10 @@ LIQ_NONNULL static void update_dither_map(unsigned char *const *const row_pointe
for(unsigned int col=1; col < width; col++) {
const unsigned char px = row_pointers[row][col];
+ if (input_image->background && map->palette[px].acolor.a < 1.f/256.f) {
+ // Transparency may or may not create an edge. When there's an explicit background set, assume no edge.
+ continue;
+ }
if (px != lastpixel || col == width-1) {
int neighbor_count = 10 * (col-lastcol);
@@ -1758,6 +1874,10 @@ static colormap *find_best_palette(histogram *hist, const liq_attr *options, con
// at this point actual gamma is not set, so very conservative posterization estimate is used
const double target_mse = MIN(max_mse, MAX(options->target_mse, pow((1<min_posterization_output)/1024.0, 2)));
int feedback_loop_trials = options->feedback_loop_trials;
+ if (hist->size > 5000) {feedback_loop_trials = (feedback_loop_trials*3 + 3)/4;}
+ if (hist->size > 25000) {feedback_loop_trials = (feedback_loop_trials*3 + 3)/4;}
+ if (hist->size > 50000) {feedback_loop_trials = (feedback_loop_trials*3 + 3)/4;}
+ if (hist->size > 100000) {feedback_loop_trials = (feedback_loop_trials*3 + 3)/4;}
colormap *acolormap = NULL;
double least_error = MAX_DIFF;
double target_mse_overshoot = feedback_loop_trials>0 ? 1.05 : 1.0;
@@ -1767,7 +1887,7 @@ static colormap *find_best_palette(histogram *hist, const liq_attr *options, con
colormap *newmap;
if (hist->size && fixed_colors_count < max_colors) {
newmap = mediancut(hist, max_colors-fixed_colors_count, target_mse * target_mse_overshoot, MAX(MAX(45.0/65536.0, target_mse), least_error)*1.2,
- options->malloc, options->free);
+ options->malloc, options->free);
} else {
feedback_loop_trials = 0;
newmap = NULL;
@@ -1864,7 +1984,7 @@ LIQ_NONNULL static liq_error pngquant_quantize(histogram *hist, const liq_attr *
}
// K-Means iteration approaches local minimum for the palette
- const double iteration_limit = options->kmeans_iteration_limit;
+ double iteration_limit = options->kmeans_iteration_limit;
unsigned int iterations = options->kmeans_iterations;
if (!iterations && palette_error < 0 && max_mse < MAX_DIFF) iterations = 1; // otherwise total error is never calculated and MSE limit won't work
@@ -1877,6 +1997,11 @@ LIQ_NONNULL static liq_error pngquant_quantize(histogram *hist, const liq_attr *
}
}
+ if (hist->size > 5000) {iterations = (iterations*3 + 3)/4;}
+ if (hist->size > 25000) {iterations = (iterations*3 + 3)/4;}
+ if (hist->size > 50000) {iterations = (iterations*3 + 3)/4;}
+ if (hist->size > 100000) {iterations = (iterations*3 + 3)/4; iteration_limit *= 2;}
+
verbose_print(options, " moving colormap towards local minimum");
double previous_palette_error = MAX_DIFF;
@@ -1958,7 +2083,7 @@ LIQ_EXPORT LIQ_NONNULL liq_error liq_write_remapped_image(liq_result *result, li
return LIQ_BUFFER_TOO_SMALL;
}
- unsigned char *rows[input_image->height];
+ LIQ_ARRAY(unsigned char *, rows, input_image->height);
unsigned char *buffer_bytes = buffer;
for(unsigned int i=0; i < input_image->height; i++) {
rows[i] = &buffer_bytes[input_image->width * i];
@@ -1998,11 +2123,13 @@ LIQ_EXPORT LIQ_NONNULL liq_error liq_write_remapped_image_rows(liq_result *quant
set_rounded_palette(&result->int_palette, result->palette, result->gamma, quant->min_posterization_output);
remapping_error = remap_to_palette(input_image, row_pointers, result->palette);
} else {
- const bool generate_dither_map = result->use_dither_map && (input_image->edges && !input_image->dither_map);
+ const bool is_image_huge = (input_image->width * input_image->height) > 2000 * 2000;
+ const bool allow_dither_map = result->use_dither_map == 2 || (!is_image_huge && result->use_dither_map);
+ const bool generate_dither_map = allow_dither_map && (input_image->edges && !input_image->dither_map);
if (generate_dither_map) {
// If dithering (with dither map) is required, this image is used to find areas that require dithering
remapping_error = remap_to_palette(input_image, row_pointers, result->palette);
- update_dither_map(row_pointers, input_image);
+ update_dither_map(input_image, row_pointers, result->palette);
}
if (liq_remap_progress(result, result->progress_stage1 * 0.5f)) {
diff --git a/libimagequant.h b/libimagequant.h
index c86f5a5..c8f84b5 100644
--- a/libimagequant.h
+++ b/libimagequant.h
@@ -13,8 +13,8 @@
#define LIQ_EXPORT extern
#endif
-#define LIQ_VERSION 20900
-#define LIQ_VERSION_STRING "2.9.0"
+#define LIQ_VERSION 21200
+#define LIQ_VERSION_STRING "2.12.2"
#ifndef LIQ_PRIVATE
#if defined(__GNUC__) || defined (__llvm__)
@@ -60,7 +60,11 @@ typedef enum liq_error {
LIQ_UNSUPPORTED,
} liq_error;
-enum liq_ownership {LIQ_OWN_ROWS=4, LIQ_OWN_PIXELS=8};
+enum liq_ownership {
+ LIQ_OWN_ROWS=4,
+ LIQ_OWN_PIXELS=8,
+ LIQ_COPY_PIXELS=16,
+};
typedef struct liq_histogram_entry {
liq_color color;
@@ -75,6 +79,7 @@ LIQ_EXPORT void liq_attr_destroy(liq_attr *attr) LIQ_NONNULL;
LIQ_EXPORT LIQ_USERESULT liq_histogram* liq_histogram_create(const liq_attr* attr);
LIQ_EXPORT liq_error liq_histogram_add_image(liq_histogram *hist, const liq_attr *attr, liq_image* image) LIQ_NONNULL;
LIQ_EXPORT liq_error liq_histogram_add_colors(liq_histogram *hist, const liq_attr *attr, const liq_histogram_entry entries[], int num_entries, double gamma) LIQ_NONNULL;
+LIQ_EXPORT liq_error liq_histogram_add_fixed_color(liq_histogram *hist, liq_color color, double gamma) LIQ_NONNULL;
LIQ_EXPORT void liq_histogram_destroy(liq_histogram *hist) LIQ_NONNULL;
LIQ_EXPORT liq_error liq_set_max_colors(liq_attr* attr, int colors) LIQ_NONNULL;
@@ -107,6 +112,8 @@ typedef void liq_image_get_rgba_row_callback(liq_color row_out[], int row, int w
LIQ_EXPORT LIQ_USERESULT liq_image *liq_image_create_custom(const liq_attr *attr, liq_image_get_rgba_row_callback *row_callback, void* user_info, int width, int height, double gamma);
LIQ_EXPORT liq_error liq_image_set_memory_ownership(liq_image *image, int ownership_flags) LIQ_NONNULL;
+LIQ_EXPORT liq_error liq_image_set_background(liq_image *img, liq_image *background_image) LIQ_NONNULL;
+LIQ_EXPORT liq_error liq_image_set_importance_map(liq_image *img, unsigned char buffer[], size_t buffer_size, enum liq_ownership memory_handling) LIQ_NONNULL;
LIQ_EXPORT liq_error liq_image_add_fixed_color(liq_image *img, liq_color color) LIQ_NONNULL;
LIQ_EXPORT LIQ_USERESULT int liq_image_get_width(const liq_image *img) LIQ_NONNULL;
LIQ_EXPORT LIQ_USERESULT int liq_image_get_height(const liq_image *img) LIQ_NONNULL;
diff --git a/mediancut.c b/mediancut.c
index 6e3e590..447a4af 100644
--- a/mediancut.c
+++ b/mediancut.c
@@ -1,32 +1,9 @@
/*
-** © 2009-2015 by Kornel Lesiński.
+** © 2009-2018 by Kornel Lesiński.
+** © 1989, 1991 by Jef Poskanzer.
+** © 1997, 2000, 2002 by Greg Roelofs; based on an idea by Stefan Schneider.
**
-** This file is part of libimagequant.
-**
-** libimagequant 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.
-**
-** libimagequant 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 libimagequant. If not, see .
-*/
-/*
-** Copyright (C) 1989, 1991 by Jef Poskanzer.
-** Copyright (C) 1997, 2000, 2002 by Greg Roelofs; based on an idea by
-** Stefan Schneider.
-**
-** Permission to use, copy, modify, and distribute this software and its
-** documentation for any purpose and without fee is hereby granted, provided
-** that the above copyright notice appear in all copies and that both that
-** copyright notice and this permission notice appear in supporting
-** documentation. This software is provided "as is" without express or
-** implied warranty.
+** See COPYRIGHT file for license.
*/
#include
@@ -141,7 +118,7 @@ inline static unsigned int qsort_partition(hist_item *const base, const unsigned
}
/** quick select algorithm */
-static void hist_item_sort_range(hist_item *base, unsigned int len, unsigned int sort_start)
+static void hist_item_sort_range(hist_item base[], unsigned int len, unsigned int sort_start)
{
for(;;) {
const unsigned int l = qsort_partition(base, len), r = l+1;
@@ -157,7 +134,7 @@ static void hist_item_sort_range(hist_item *base, unsigned int len, unsigned int
}
/** sorts array to make sum of weights lower than halfvar one side, returns edge between halfvar parts of the set */
-static hist_item *hist_item_sort_halfvar(hist_item *base, unsigned int len, double *const lowervar, const double halfvar)
+static hist_item *hist_item_sort_halfvar(hist_item base[], unsigned int len, double *const lowervar, const double halfvar)
{
do {
const unsigned int l = qsort_partition(base, len), r = l+1;
@@ -198,7 +175,7 @@ typedef struct {
static int comparevariance(const void *ch1, const void *ch2)
{
return ((const channelvariance*)ch1)->variance > ((const channelvariance*)ch2)->variance ? -1 :
- (((const channelvariance*)ch1)->variance < ((const channelvariance*)ch2)->variance ? 1 : 0);
+ (((const channelvariance*)ch1)->variance < ((const channelvariance*)ch2)->variance ? 1 : 0);
}
/** Finds which channels need to be sorted first and preproceses achv for fast sort */
@@ -216,12 +193,16 @@ static double prepare_sort(struct box *b, hist_item achv[])
qsort(channels, 4, sizeof(channels[0]), comparevariance);
- for(unsigned int i=0; i < b->colors; i++) {
- const float *chans = (const float *)&achv[b->ind + i].acolor;
+ const unsigned int ind1 = b->ind;
+ const unsigned int colors = b->colors;
+ #pragma omp parallel for if (colors > 25000) \
+ schedule(static) default(none) shared(achv, channels)
+ for(unsigned int i=0; i < colors; i++) {
+ const float *chans = (const float *)&achv[ind1 + i].acolor;
// Only the first channel really matters. When trying median cut many times
// with different histogram weights, I don't want sort randomness to influence outcome.
- achv[b->ind + i].tmp.sort_value = ((unsigned int)(chans[channels[0].chan]*65535.0)<<16) |
- (unsigned int)((chans[channels[2].chan] + chans[channels[1].chan]/2.0 + chans[channels[3].chan]/4.0)*65535.0);
+ achv[ind1 + i].tmp.sort_value = ((unsigned int)(chans[channels[0].chan]*65535.0)<<16) |
+ (unsigned int)((chans[channels[2].chan] + chans[channels[1].chan]/2.0 + chans[channels[3].chan]/4.0)*65535.0);
}
const f_pixel median = get_median(b, achv);
@@ -229,6 +210,8 @@ static double prepare_sort(struct box *b, hist_item achv[])
// box will be split to make color_weight of each side even
const unsigned int ind = b->ind, end = ind+b->colors;
double totalvar = 0;
+ #pragma omp parallel for if (end - ind > 15000) \
+ schedule(static) default(shared) reduction(+:totalvar)
for(unsigned int j=ind; j < end; j++) totalvar += (achv[j].color_weight = color_weight(median, achv[j]));
return totalvar / 2.0;
}
@@ -251,7 +234,7 @@ static f_pixel get_median(const struct box *b, hist_item achv[])
/*
** Find the best splittable box. -1 if no boxes are splittable.
*/
-static int best_splittable_box(struct box* bv, unsigned int boxes, const double max_mse)
+static int best_splittable_box(struct box bv[], unsigned int boxes, const double max_mse)
{
int bi=-1; double maxsum=0;
for(unsigned int i=0; i < boxes; i++) {
@@ -281,8 +264,8 @@ inline static double color_weight(f_pixel median, hist_item h)
return sqrt(diff) * (sqrt(1.0+h.adjusted_weight)-1.0);
}
-static void set_colormap_from_boxes(colormap *map, struct box* bv, unsigned int boxes, hist_item *achv);
-static void adjust_histogram(hist_item *achv, const struct box* bv, unsigned int boxes);
+static void set_colormap_from_boxes(colormap *map, struct box bv[], unsigned int boxes, hist_item *achv);
+static void adjust_histogram(hist_item *achv, const struct box bv[], unsigned int boxes);
static double box_error(const struct box *box, const hist_item achv[])
{
@@ -326,8 +309,11 @@ static void box_init(struct box *box, const hist_item *achv, const unsigned int
box->colors = colors;
box->sum = sum;
box->total_error = -1;
+
box->color = averagepixels(colors, &achv[ind]);
+ #pragma omp task if (colors > 5000)
box->variance = box_variance(achv, box);
+ #pragma omp task if (colors > 8000)
box->max_error = box_max_error(achv, box);
}
@@ -339,67 +325,78 @@ static void box_init(struct box *box, const hist_item *achv, const unsigned int
LIQ_PRIVATE colormap *mediancut(histogram *hist, unsigned int newcolors, const double target_mse, const double max_mse, void* (*malloc)(size_t), void (*free)(void*))
{
hist_item *achv = hist->achv;
- struct box bv[newcolors];
+ LIQ_ARRAY(struct box, bv, newcolors);
+ unsigned int boxes = 1;
/*
** Set up the initial box.
*/
+ #pragma omp parallel
+ #pragma omp single
+ {
double sum = 0;
- for(unsigned int i=0; i < hist->size; i++) {
+ for(unsigned int i=0; i < hist->size; i++) {
sum += achv[i].adjusted_weight;
}
- box_init(&bv[0], achv, 0, hist->size, sum);
+ #pragma omp taskgroup
+ {
+ box_init(&bv[0], achv, 0, hist->size, sum);
+ }
- unsigned int boxes = 1;
-
- /*
- ** Main loop: split boxes until we have enough.
- */
- while (boxes < newcolors) {
-
- // first splits boxes that exceed quality limit (to have colors for things like odd green pixel),
- // later raises the limit to allow large smooth areas/gradients get colors.
- const double current_max_mse = max_mse + (boxes/(double)newcolors)*16.0*max_mse;
- const int bi = best_splittable_box(bv, boxes, current_max_mse);
- if (bi < 0)
- break; /* ran out of colors! */
-
- unsigned int indx = bv[bi].ind;
- unsigned int clrs = bv[bi].colors;
/*
- Classic implementation tries to get even number of colors or pixels in each subdivision.
-
- Here, instead of popularity I use (sqrt(popularity)*variance) metric.
- Each subdivision balances number of pixels (popular colors) and low variance -
- boxes can be large if they have similar colors. Later boxes with high variance
- will be more likely to be split.
-
- Median used as expected value gives much better results than mean.
+ ** Main loop: split boxes until we have enough.
*/
+ while (boxes < newcolors) {
- const double halfvar = prepare_sort(&bv[bi], achv);
- double lowervar=0;
+ // first splits boxes that exceed quality limit (to have colors for things like odd green pixel),
+ // later raises the limit to allow large smooth areas/gradients get colors.
+ const double current_max_mse = max_mse + (boxes/(double)newcolors)*16.0*max_mse;
+ const int bi = best_splittable_box(bv, boxes, current_max_mse);
+ if (bi < 0) {
+ break; /* ran out of colors! */
+ }
- // hist_item_sort_halfvar sorts and sums lowervar at the same time
- // returns item to break at …minus one, which does smell like an off-by-one error.
- hist_item *break_p = hist_item_sort_halfvar(&achv[indx], clrs, &lowervar, halfvar);
- unsigned int break_at = MIN(clrs-1, break_p - &achv[indx] + 1);
+ unsigned int indx = bv[bi].ind;
+ unsigned int clrs = bv[bi].colors;
- /*
- ** Split the box.
- */
- double sm = bv[bi].sum;
- double lowersum = 0;
- for(unsigned int i=0; i < break_at; i++) lowersum += achv[indx + i].adjusted_weight;
+ /*
+ Classic implementation tries to get even number of colors or pixels in each subdivision.
- box_init(&bv[bi], achv, bv[bi].ind, break_at, lowersum);
- box_init(&bv[boxes], achv, indx + break_at, clrs - break_at, sm - lowersum);
+ Here, instead of popularity I use (sqrt(popularity)*variance) metric.
+ Each subdivision balances number of pixels (popular colors) and low variance -
+ boxes can be large if they have similar colors. Later boxes with high variance
+ will be more likely to be split.
- ++boxes;
+ Median used as expected value gives much better results than mean.
+ */
- if (total_box_error_below_target(target_mse, bv, boxes, hist)) {
- break;
+ const double halfvar = prepare_sort(&bv[bi], achv);
+ double lowervar=0;
+
+ // hist_item_sort_halfvar sorts and sums lowervar at the same time
+ // returns item to break at …minus one, which does smell like an off-by-one error.
+ hist_item *break_p = hist_item_sort_halfvar(&achv[indx], clrs, &lowervar, halfvar);
+ unsigned int break_at = MIN(clrs-1, break_p - &achv[indx] + 1);
+
+ /*
+ ** Split the box.
+ */
+ double sm = bv[bi].sum;
+ double lowersum = 0;
+ for(unsigned int i=0; i < break_at; i++) lowersum += achv[indx + i].adjusted_weight;
+
+ #pragma omp taskgroup
+ {
+ box_init(&bv[bi], achv, indx, break_at, lowersum);
+ box_init(&bv[boxes], achv, indx + break_at, clrs - break_at, sm - lowersum);
+ }
+
+ ++boxes;
+
+ if (total_box_error_below_target(target_mse, bv, boxes, hist)) {
+ break;
+ }
}
}
@@ -446,6 +443,8 @@ static f_pixel averagepixels(unsigned int clrs, const hist_item achv[])
{
double r = 0, g = 0, b = 0, a = 0, sum = 0;
+ #pragma omp parallel for if (clrs > 25000) \
+ schedule(static) default(shared) reduction(+:a) reduction(+:r) reduction(+:g) reduction(+:b) reduction(+:sum)
for(unsigned int i = 0; i < clrs; i++) {
const f_pixel px = achv[i].acolor;
const double weight = achv[i].adjusted_weight;
diff --git a/mempool.c b/mempool.c
index cd345a7..cd49f59 100644
--- a/mempool.c
+++ b/mempool.c
@@ -1,20 +1,9 @@
/*
-© 2011-2016 by Kornel Lesiński.
-
-This file is part of libimagequant.
-
-libimagequant 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.
-
-libimagequant 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 libimagequant. If not, see .
+** © 2009-2017 by Kornel Lesiński.
+** © 1989, 1991 by Jef Poskanzer.
+** © 1997, 2000, 2002 by Greg Roelofs; based on an idea by Stefan Schneider.
+**
+** See COPYRIGHT file for license.
*/
#include "libimagequant.h"
@@ -32,7 +21,7 @@ struct mempool {
void (*free)(void*);
struct mempool *next;
};
-LIQ_PRIVATE void* mempool_create(mempool *mptr, const unsigned int size, unsigned int max_size, void* (*malloc)(size_t), void (*free)(void*))
+LIQ_PRIVATE void* mempool_create(mempoolptr *mptr, const unsigned int size, unsigned int max_size, void* (*malloc)(size_t), void (*free)(void*))
{
if (*mptr && ((*mptr)->used+size) <= (*mptr)->size) {
unsigned int prevused = (*mptr)->used;
@@ -40,7 +29,7 @@ LIQ_PRIVATE void* mempool_create(mempool *mptr, const unsigned int size, unsigne
return ((char*)(*mptr)) + prevused;
}
- mempool old = *mptr;
+ mempoolptr old = *mptr;
if (!max_size) max_size = (1<<17);
max_size = size+ALIGN_MASK > max_size ? size+ALIGN_MASK : max_size;
@@ -60,7 +49,7 @@ LIQ_PRIVATE void* mempool_create(mempool *mptr, const unsigned int size, unsigne
return mempool_alloc(mptr, size, size);
}
-LIQ_PRIVATE void* mempool_alloc(mempool *mptr, const unsigned int size, const unsigned int max_size)
+LIQ_PRIVATE void* mempool_alloc(mempoolptr *mptr, const unsigned int size, const unsigned int max_size)
{
if (((*mptr)->used+size) <= (*mptr)->size) {
unsigned int prevused = (*mptr)->used;
@@ -71,10 +60,10 @@ LIQ_PRIVATE void* mempool_alloc(mempool *mptr, const unsigned int size, const un
return mempool_create(mptr, size, max_size, (*mptr)->malloc, (*mptr)->free);
}
-LIQ_PRIVATE void mempool_destroy(mempool m)
+LIQ_PRIVATE void mempool_destroy(mempoolptr m)
{
while (m) {
- mempool next = m->next;
+ mempoolptr next = m->next;
m->free(m);
m = next;
}
diff --git a/mempool.h b/mempool.h
index 0797e7c..9b7333b 100644
--- a/mempool.h
+++ b/mempool.h
@@ -4,10 +4,10 @@
#include
struct mempool;
-typedef struct mempool *mempool;
+typedef struct mempool *mempoolptr;
-LIQ_PRIVATE void* mempool_create(mempool *mptr, const unsigned int size, unsigned int capacity, void* (*malloc)(size_t), void (*free)(void*));
-LIQ_PRIVATE void* mempool_alloc(mempool *mptr, const unsigned int size, const unsigned int capacity);
-LIQ_PRIVATE void mempool_destroy(mempool m);
+LIQ_PRIVATE void* mempool_create(mempoolptr *mptr, const unsigned int size, unsigned int capacity, void* (*malloc)(size_t), void (*free)(void*));
+LIQ_PRIVATE void* mempool_alloc(mempoolptr *mptr, const unsigned int size, const unsigned int capacity);
+LIQ_PRIVATE void mempool_destroy(mempoolptr m);
#endif
diff --git a/nearest.c b/nearest.c
index 2f1e926..aeb4dc1 100644
--- a/nearest.c
+++ b/nearest.c
@@ -1,20 +1,9 @@
/*
-© 2011-2015 by Kornel Lesiński.
-
-This file is part of libimagequant.
-
-libimagequant 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.
-
-libimagequant 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 libimagequant. If not, see .
+** © 2009-2015 by Kornel Lesiński.
+** © 1989, 1991 by Jef Poskanzer.
+** © 1997, 2000, 2002 by Greg Roelofs; based on an idea by Stefan Schneider.
+**
+** See COPYRIGHT file for license.
*/
#include "libimagequant.h"
@@ -45,7 +34,7 @@ struct nearest_map {
vp_node *root;
const colormap_item *palette;
float nearest_other_color_dist[256];
- mempool mempool;
+ mempoolptr mempool;
};
static void vp_search_node(const vp_node *node, const f_pixel *const needle, vp_search_tmp *const best_candidate);
@@ -56,7 +45,7 @@ static int vp_compare_distance(const void *ap, const void *bp) {
return a > b ? 1 : -1;
}
-static void vp_sort_indexes_by_distance(const f_pixel vantage_point, vp_sort_tmp *indexes, int num_indexes, const colormap_item items[]) {
+static void vp_sort_indexes_by_distance(const f_pixel vantage_point, vp_sort_tmp indexes[], int num_indexes, const colormap_item items[]) {
for(int i=0; i < num_indexes; i++) {
indexes[i].distance_squared = colordifference(vantage_point, items[indexes[i].idx].acolor);
}
@@ -66,7 +55,7 @@ static void vp_sort_indexes_by_distance(const f_pixel vantage_point, vp_sort_tmp
/*
* Usually it should pick farthest point, but picking most popular point seems to make search quicker anyway
*/
-static int vp_find_best_vantage_point_index(vp_sort_tmp *indexes, int num_indexes, const colormap_item items[]) {
+static int vp_find_best_vantage_point_index(vp_sort_tmp indexes[], int num_indexes, const colormap_item items[]) {
int best = 0;
float best_popularity = items[indexes[0].idx].popularity;
for(int i = 1; i < num_indexes; i++) {
@@ -78,7 +67,7 @@ static int vp_find_best_vantage_point_index(vp_sort_tmp *indexes, int num_indexe
return best;
}
-static vp_node *vp_create_node(mempool *m, vp_sort_tmp *indexes, int num_indexes, const colormap_item items[]) {
+static vp_node *vp_create_node(mempoolptr *m, vp_sort_tmp indexes[], int num_indexes, const colormap_item items[]) {
if (num_indexes <= 0) {
return NULL;
}
@@ -118,10 +107,10 @@ static vp_node *vp_create_node(mempool *m, vp_sort_tmp *indexes, int num_indexes
}
LIQ_PRIVATE struct nearest_map *nearest_init(const colormap *map) {
- mempool m = NULL;
+ mempoolptr m = NULL;
struct nearest_map *handle = mempool_create(&m, sizeof(handle[0]), sizeof(handle[0]) + sizeof(vp_node)*map->colors+16, map->malloc, map->free);
- vp_sort_tmp indexes[map->colors];
+ LIQ_ARRAY(vp_sort_tmp, indexes, map->colors);
for(unsigned int i=0; i < map->colors; i++) {
indexes[i].idx = i;
diff --git a/pam.c b/pam.c
index 74165e1..660f829 100644
--- a/pam.c
+++ b/pam.c
@@ -1,16 +1,10 @@
/* pam.c - pam (portable alpha map) utility library
**
-** Copyright (C) 1989, 1991 by Jef Poskanzer.
-** Copyright (C) 1997, 2000, 2002 by Greg Roelofs; based on an idea by
-** Stefan Schneider.
-** © 2009-2016 by Kornel Lesinski.
+** © 2009-2017 by Kornel Lesiński.
+** © 1989, 1991 by Jef Poskanzer.
+** © 1997, 2000, 2002 by Greg Roelofs; based on an idea by Stefan Schneider.
**
-** Permission to use, copy, modify, and distribute this software and its
-** documentation for any purpose and without fee is hereby granted, provided
-** that the above copyright notice appear in all copies and that both that
-** copyright notice and this permission notice appear in supporting
-** documentation. This software is provided "as is" without express or
-** implied warranty.
+** See COPYRIGHT file for license.
*/
#include
@@ -33,11 +27,8 @@ LIQ_PRIVATE bool pam_computeacolorhash(struct acolorhash_table *acht, const rgba
/* Go through the entire image, building a hash table of colors. */
for(unsigned int row = 0; row < rows; ++row) {
- float boost=1.0;
for(unsigned int col = 0; col < cols; ++col) {
- if (importance_map) {
- boost = 0.5f+ (double)*importance_map++/255.f;
- }
+ unsigned int boost;
// RGBA color is casted to long for easier hasing/comparisons
union rgba_as_int px = {pixels[row][col]};
@@ -45,12 +36,22 @@ LIQ_PRIVATE bool pam_computeacolorhash(struct acolorhash_table *acht, const rgba
if (!px.rgba.a) {
// "dirty alpha" has different RGBA values that end up being the same fully transparent color
px.l=0; hash=0;
- boost = 10;
+
+ boost = 2000;
+ if (importance_map) {
+ importance_map++;
+ }
} else {
// mask posterizes all 4 channels in one go
px.l = (px.l & posterize_mask) | ((px.l & posterize_high_mask) >> (8-ignorebits));
// fancier hashing algorithms didn't improve much
hash = px.l % hash_size;
+
+ if (importance_map) {
+ boost = *importance_map++;
+ } else {
+ boost = 255;
+ }
}
if (!pam_add_to_hash(acht, hash, boost, px, row, rows)) {
@@ -64,7 +65,7 @@ LIQ_PRIVATE bool pam_computeacolorhash(struct acolorhash_table *acht, const rgba
return true;
}
-LIQ_PRIVATE bool pam_add_to_hash(struct acolorhash_table *acht, unsigned int hash, float boost, union rgba_as_int px, unsigned int row, unsigned int rows)
+LIQ_PRIVATE bool pam_add_to_hash(struct acolorhash_table *acht, unsigned int hash, unsigned int boost, union rgba_as_int px, unsigned int row, unsigned int rows)
{
/* head of the hash function stores first 2 colors inline (achl->used = 1..2),
to reduce number of allocations of achl->other_items.
@@ -160,7 +161,7 @@ LIQ_PRIVATE struct acolorhash_table *pam_allocacolorhash(unsigned int maxcolors,
const size_t estimated_colors = MIN(maxcolors, surface/(ignorebits + (surface > 512*512 ? 6 : 5)));
const size_t hash_size = estimated_colors < 66000 ? 6673 : (estimated_colors < 200000 ? 12011 : 24019);
- mempool m = NULL;
+ mempoolptr m = NULL;
const size_t buckets_size = hash_size * sizeof(struct acolorhist_arr_head);
const size_t mempool_size = sizeof(struct acolorhash_table) + buckets_size + estimated_colors * sizeof(struct acolorhist_arr_item);
struct acolorhash_table *t = mempool_create(&m, sizeof(*t) + buckets_size, mempool_size, malloc, free);
@@ -175,11 +176,15 @@ LIQ_PRIVATE struct acolorhash_table *pam_allocacolorhash(unsigned int maxcolors,
return t;
}
-ALWAYS_INLINE static float pam_add_to_hist(const float *gamma_lut, hist_item *achv, unsigned int j, const struct acolorhist_arr_item *entry, const float max_perceptual_weight)
+ALWAYS_INLINE static float pam_add_to_hist(const float *gamma_lut, hist_item *achv, unsigned int *j, const struct acolorhist_arr_item *entry, const float max_perceptual_weight)
{
- achv[j].acolor = rgba_to_f(gamma_lut, entry->color.rgba);
- const float w = MIN(entry->perceptual_weight, max_perceptual_weight);
- achv[j].adjusted_weight = achv[j].perceptual_weight = w;
+ if (entry->perceptual_weight == 0) {
+ return 0;
+ }
+ const float w = MIN(entry->perceptual_weight/128.f, max_perceptual_weight);
+ achv[*j].adjusted_weight = achv[*j].perceptual_weight = w;
+ achv[*j].acolor = rgba_to_f(gamma_lut, entry->color.rgba);
+ *j += 1;
return w;
}
@@ -203,22 +208,27 @@ LIQ_PRIVATE histogram *pam_acolorhashtoacolorhist(const struct acolorhash_table
float max_perceptual_weight = 0.1f * acht->cols * acht->rows;
double total_weight = 0;
- for(unsigned int j=0, i=0; i < acht->hash_size; ++i) {
+ unsigned int j=0;
+ for(unsigned int i=0; i < acht->hash_size; ++i) {
const struct acolorhist_arr_head *const achl = &acht->buckets[i];
if (achl->used) {
- total_weight += pam_add_to_hist(gamma_lut, hist->achv, j++, &achl->inline1, max_perceptual_weight);
+ total_weight += pam_add_to_hist(gamma_lut, hist->achv, &j, &achl->inline1, max_perceptual_weight);
if (achl->used > 1) {
- total_weight += pam_add_to_hist(gamma_lut, hist->achv, j++, &achl->inline2, max_perceptual_weight);
+ total_weight += pam_add_to_hist(gamma_lut, hist->achv, &j, &achl->inline2, max_perceptual_weight);
for(unsigned int k=0; k < achl->used-2; k++) {
- total_weight += pam_add_to_hist(gamma_lut, hist->achv, j++, &achl->other_items[k], max_perceptual_weight);
+ total_weight += pam_add_to_hist(gamma_lut, hist->achv, &j, &achl->other_items[k], max_perceptual_weight);
}
}
}
}
-
+ hist->size = j;
hist->total_perceptual_weight = total_weight;
+ if (!j) {
+ pam_freeacolorhist(hist);
+ return NULL;
+ }
return hist;
}
diff --git a/pam.h b/pam.h
index 6657a14..2ca4327 100644
--- a/pam.h
+++ b/pam.h
@@ -62,6 +62,12 @@
# define SSE_ALIGN
#endif
+#ifndef _MSC_VER
+#define LIQ_ARRAY(type, var, count) type var[count]
+#else
+#define LIQ_ARRAY(type, var, count) type* var = (type*)_alloca(sizeof(type)*(count))
+#endif
+
#if defined(__GNUC__) || defined (__llvm__)
#define ALWAYS_INLINE __attribute__((always_inline)) inline
#define NEVER_INLINE __attribute__ ((noinline))
@@ -85,7 +91,7 @@ typedef struct {
float a, r, g, b;
} SSE_ALIGN f_pixel;
-static const double internal_gamma = 0.5499;
+static const float internal_gamma = 0.5499f;
LIQ_PRIVATE void to_f_set_gamma(float gamma_lut[], const double gamma);
@@ -238,7 +244,7 @@ typedef struct colormap {
struct acolorhist_arr_item {
union rgba_as_int color;
- float perceptual_weight;
+ unsigned int perceptual_weight;
};
struct acolorhist_arr_head {
@@ -260,7 +266,7 @@ LIQ_PRIVATE void pam_freeacolorhash(struct acolorhash_table *acht);
LIQ_PRIVATE struct acolorhash_table *pam_allocacolorhash(unsigned int maxcolors, unsigned int surface, unsigned int ignorebits, void* (*malloc)(size_t), void (*free)(void*));
LIQ_PRIVATE histogram *pam_acolorhashtoacolorhist(const struct acolorhash_table *acht, const double gamma, void* (*malloc)(size_t), void (*free)(void*));
LIQ_PRIVATE bool pam_computeacolorhash(struct acolorhash_table *acht, const rgba_pixel *const pixels[], unsigned int cols, unsigned int rows, const unsigned char *importance_map);
-LIQ_PRIVATE bool pam_add_to_hash(struct acolorhash_table *acht, unsigned int hash, float boost, union rgba_as_int px, unsigned int row, unsigned int rows);
+LIQ_PRIVATE bool pam_add_to_hash(struct acolorhash_table *acht, unsigned int hash, unsigned int boost, union rgba_as_int px, unsigned int row, unsigned int rows);
LIQ_PRIVATE void pam_freeacolorhist(histogram *h);