doc: update README for v2.12.6 release

.gitignore

@@ -0,0 +1,2 @@
+cmd/gopngquant/gopngquant
+cmd/gopngquant/gopngquant.exe

.hgignore

@@ -1,4 +0,0 @@
-mode:regex
-
-^_dist/
-\.exe$

README.md

@@ -0,0 +1,70 @@
+# imagequant
+
+Go bindings for libimagequant
+
+`libimagequant` is a library for lossy recompression of PNG images to reduce their filesize. It is used by the `pngquant` tool. This `go-imagequant` project is a set of bindings for libimagequant to enable its use from the Go programming language.
+
+This binding was written by hand. The result is somewhat more idiomatic than an automated conversion, but some `defer foo.Release()` calls are required for memory management.
+
+## Usage
+
+Usage example is provided by a sample utility `cmd/gopngquant` which mimics some functionality of the upstream `pngquant`.
+
+The sample utility has the following options:
+
+```
+Usage of gopngquant:
+  -In string
+        Input filename
+  -Out string
+        Output filename
+  -Speed int
+        Speed (1 slowest, 10 fastest) (default 3)
+  -Version
+```
+
+## Building
+
+This package can be installed via go get: `go get code.ivysaur.me/imagequant`
+
+Import the package as `code.ivysaur.me/imagequant/v2` and compile via `go build`.
+
+This is a CGO package and requires a C compiler installed. However, if you use `go install` then future invocations of `go build` do not require the C compiler to be present.
+
+The `imagequant.go` file also declares a number of `CFLAGS` for GCC that allow the included libimagequant to build in a reasonably optimal way without using the upstream configure/make scripts.
+
+## License
+
+I am releasing this binding under the ISC license, however, `libimagequant` itself is released under GPLv3-or-later and/or commercial licenses. You must comply with the terms of such a license when using this binding in a Go project.
+
+## Changelog
+
+2020-07-25 v2.12.6
+- Update bundled libimagequant from 2.12.2 to 2.12.6
+- Fix missing v2 in go.mod file causing failure to import in Modules mode
+
+2018-12-31 v2.12.2-go1.2
+- go-imagequant: Update bundled libimagequant from 2.9.0 to 2.12.2
+- build: Switch to Go Modules
+- build: Update bundled CFLAGS for new CGo whitelist (reduces performance)
+- build: Remove nonportable Cygwin makefile
+
+2017-03-03 v2.9.0-go1.1
+- *Previously tagged as 2.9go1.1*
+- go-imagequant: Update bundled libimagequant from 2.8.0 to 2.9.0
+- go-imagequant: Separate `CGO_LDFLAGS` for Linux and Windows targets
+- gopngquant: Fix an issue with non-square images
+- [⬇️ go-imagequant-2.9go1.1-src.zip](https://git.ivysaur.me/attachments/381bbeac-c50f-4235-b61a-54785cf49b11) *(56.13 KiB)*
+- [⬇️ gopngquant-2.9go1.1-win64.7z](https://git.ivysaur.me/attachments/ced44b97-4666-43cc-a578-44f4272be686) *(497.54 KiB)*
+
+2016-11-24 v2.8.0-go1.0
+- *Previously tagged as 2.8go1.0*
+- Initial public release
+- [⬇️ go-imagequant-2.8go1.0-src.zip](https://git.ivysaur.me/attachments/0f45207f-aeb8-4dd7-a62c-499a3f378f1a) *(55.40 KiB)*
+- [⬇️ gopngquant-2.8go1.0-win64.7z](https://git.ivysaur.me/attachments/46229661-fa5d-418d-92ff-d6c4f72a5204) *(497.98 KiB)*
+
+## See also
+
+- Pngquant homepage https://pngquant.org/
+- Pngquant source code https://github.com/pornel/pngquant
+- Libimagequant source code https://github.com/ImageOptim/libimagequant

_dist/README.txt

@@ -1,45 +0,0 @@
-Go bindings for libimagequant
-
-`libimagequant` is a library for lossy recompression of PNG images to reduce their filesize. It is used by the `pngquant` tool. This `go-imagequant` project is a set of bindings for libimagequant to enable its use from the Go programming language.
-
-This binding was written by hand. The result is somewhat more idiomatic than an automated conversion, but some `defer foo.Release()` calls are required for memory management.
-
-Written in Golang
-
-=USAGE=
-
-Usage example is provided by a sample utility `cmd/gopngquant` which mimics some functionality of the upstream `pngquant`.
-
-The sample utility has the following options:
-
-`Usage of gopngquant:
-  -In string
-        Input filename
-  -Out string
-        Output filename
-  -Speed int
-        Speed (1 slowest, 10 fastest) (default 3)
-  -Version`
-
-=BUILDING=
-
-The expected package path is `code.ivysaur.me/imagequant`. Build via `go build`.
-
-This is a CGO package and requires a C compiler installed. However, if you use `go install` then future invocations of `go build` do not require the C compiler to be present.
-
-The `imagequant.go` file also declares a number of `CFLAGS` for GCC that allow the included libimagequant (2.8 git-a425e83) to build in an optimal way without using the upstream configure/make scripts.
-
-=LICENSE=
-
-I am releasing this binding under the ISC license, however, `libimagequant` itself is released under GPLv3-or-later and/or commercial licenses. You must comply with the terms of such a license when using this binding in a Go project.
-
-=CHANGELOG=
-
-2016-11-24 2.8go1.0
-- Initial public release
-
-=SEE ALSO=
-
-- Pngquant homepage https://pngquant.org/
-- Pngquant source code https://github.com/pornel/pngquant
-- Libimagequant source code https://github.com/ImageOptim/libimagequant

blur.h

@@ -1,4 +1,8 @@
+#ifndef BLUR_H
+#define BLUR_H
 
 LIQ_PRIVATE void liq_blur(unsigned char *src, unsigned char *tmp, unsigned char *dst, unsigned int width, unsigned int height, unsigned int size);
 LIQ_PRIVATE void liq_max3(unsigned char *src, unsigned char *dst, unsigned int width, unsigned int height);
 LIQ_PRIVATE void liq_min3(unsigned char *src, unsigned char *dst, unsigned int width, unsigned int height);
+
+#endif

cflags_linux.go

@@ -0,0 +1,9 @@
+//+build !windows
+
+package imagequant
+
+/*
+#cgo CFLAGS: -O3 -fopenmp -fomit-frame-pointer -Wall -Wno-attributes -std=c99 -DNDEBUG -DUSE_SSE=1 -msse
+#cgo LDFLAGS: -lm -fopenmp -ldl
+*/
+import "C"

cflags_windows.go

@@ -0,0 +1,9 @@
+//+build windows
+
+package imagequant
+
+/*
+#cgo CFLAGS: -O3 -fno-math-errno -fopenmp -funroll-loops -fomit-frame-pointer -Wall -Wno-attributes -std=c99 -DNDEBUG -DUSE_SSE=1 -msse -fexcess-precision=fast
+#cgo LDFLAGS: -fopenmp -static
+*/
+import "C"

cmd/gopngquant/Makefile.cygwin

@@ -1,10 +0,0 @@
-
-.PHONY: all clean
-
-all: gopngquant.exe
-
-clean:
-	rm ./gopngquant.exe || true
-	
-gopngquant.exe: main.go
-	( source ~/.cygwin-toolchain-switcher/use-toolchain-x86_64-w64-mingw32.sh && go build -a -ldflags '-s -w' )

cmd/gopngquant/main.go

@@ -24,7 +24,7 @@ import (
 	"image/png"
 	"os"
 
-	"code.ivysaur.me/imagequant"
+	"code.ivysaur.me/imagequant/v2"
 )
 
 func GoImageToRgba32(im image.Image) []byte {
@@ -61,7 +61,7 @@ func Rgb8PaletteToGoImage(w, h int, rgb8data []byte, pal color.Palette) image.Im
 
 	for y := 0; y < h; y += 1 {
 		for x := 0; x < w; x += 1 {
-			ret.SetColorIndex(x, y, rgb8data[y*h+x])
+			ret.SetColorIndex(x, y, rgb8data[y*w+x])
 		}
 	}
 

go.mod

@@ -0,0 +1,3 @@
+module code.ivysaur.me/imagequant/v2
+
+go 1.13

imagequant.go

@@ -21,8 +21,6 @@ import (
 )
 
 /*
-#cgo CFLAGS: -O3 -fno-math-errno -fopenmp -funroll-loops -fomit-frame-pointer -Wall -Wno-attributes -std=c99 -DNDEBUG -DUSE_SSE=1 -msse -fexcess-precision=fast
-#cgo LDFLAGS: -fopenmp -static
 #include "libimagequant.h"
 
 const char* liqVersionString() {

kmeans.c

@@ -0,0 +1,98 @@
+/*
+** © 2011-2016 by Kornel Lesiński.
+** See COPYRIGHT file for license.
+*/
+
+#include "libimagequant.h"
+#include "pam.h"
+#include "kmeans.h"
+#include "nearest.h"
+#include <stdlib.h>
+#include <string.h>
+
+#ifdef _OPENMP
+#include <omp.h>
+#else
+#define omp_get_max_threads() 1
+#define omp_get_thread_num() 0
+#endif
+
+/*
+ * K-Means iteration: new palette color is computed from weighted average of colors that map to that palette entry.
+ */
+LIQ_PRIVATE void kmeans_init(const colormap *map, const unsigned int max_threads, kmeans_state average_color[])
+{
+    memset(average_color, 0, sizeof(average_color[0])*(KMEANS_CACHE_LINE_GAP+map->colors)*max_threads);
+}
+
+LIQ_PRIVATE void kmeans_update_color(const f_pixel acolor, const float value, const colormap *map, unsigned int match, const unsigned int thread, kmeans_state average_color[])
+{
+    match += thread * (KMEANS_CACHE_LINE_GAP+map->colors);
+    average_color[match].a += acolor.a * value;
+    average_color[match].r += acolor.r * value;
+    average_color[match].g += acolor.g * value;
+    average_color[match].b += acolor.b * value;
+    average_color[match].total += value;
+}
+
+LIQ_PRIVATE void kmeans_finalize(colormap *map, const unsigned int max_threads, const kmeans_state average_color[])
+{
+    for (unsigned int i=0; i < map->colors; i++) {
+        double a=0, r=0, g=0, b=0, total=0;
+
+        // Aggregate results from all threads
+        for(unsigned int t=0; t < max_threads; t++) {
+            const unsigned int offset = (KMEANS_CACHE_LINE_GAP+map->colors) * t + i;
+
+            a += average_color[offset].a;
+            r += average_color[offset].r;
+            g += average_color[offset].g;
+            b += average_color[offset].b;
+            total += average_color[offset].total;
+        }
+
+        if (total && !map->palette[i].fixed) {
+            map->palette[i].acolor = (f_pixel){
+                .a = a / total,
+                .r = r / total,
+                .g = g / total,
+                .b = b / total,
+            };
+            map->palette[i].popularity = total;
+        }
+    }
+}
+
+LIQ_PRIVATE double kmeans_do_iteration(histogram *hist, colormap *const map, kmeans_callback callback)
+{
+    const unsigned int max_threads = omp_get_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;
+#if __GNUC__ >= 9
+    #pragma omp parallel for if (hist_size > 2000) \
+        schedule(static) default(none) shared(achv,average_color,callback,hist_size,map,n) reduction(+:total_diff)
+#else
+    #pragma omp parallel for if (hist_size > 2000) \
+        schedule(static) default(none) shared(average_color,callback) reduction(+:total_diff)
+#endif
+    for(int j=0; j < hist_size; j++) {
+        float diff;
+        unsigned int match = nearest_search(n, &achv[j].acolor, achv[j].tmp.likely_colormap_index, &diff);
+        achv[j].tmp.likely_colormap_index = match;
+        total_diff += diff * achv[j].perceptual_weight;
+
+        kmeans_update_color(achv[j].acolor, achv[j].perceptual_weight, map, match, omp_get_thread_num(), average_color);
+
+        if (callback) callback(&achv[j], diff);
+    }
+
+    nearest_free(n);
+    kmeans_finalize(map, max_threads, average_color);
+
+    return total_diff / hist->total_perceptual_weight;
+}

kmeans.h

@@ -0,0 +1,19 @@
+
+#ifndef KMEANS_H
+#define KMEANS_H
+
+// Spread memory touched by different threads at least 64B apart which I assume is the cache line size. This should avoid memory write contention.
+#define KMEANS_CACHE_LINE_GAP ((64+sizeof(kmeans_state)-1)/sizeof(kmeans_state))
+
+typedef struct {
+    double a, r, g, b, total;
+} kmeans_state;
+
+typedef void (*kmeans_callback)(hist_item *item, float diff);
+
+LIQ_PRIVATE void kmeans_init(const colormap *map, const unsigned int max_threads, kmeans_state state[]);
+LIQ_PRIVATE void kmeans_update_color(const f_pixel acolor, const float value, const colormap *map, unsigned int match, const unsigned int thread, kmeans_state average_color[]);
+LIQ_PRIVATE void kmeans_finalize(colormap *map, const unsigned int max_threads, const kmeans_state state[]);
+LIQ_PRIVATE double kmeans_do_iteration(histogram *hist, colormap *const map, kmeans_callback callback);
+
+#endif

libimagequant.h

@@ -13,8 +13,8 @@
 #define LIQ_EXPORT extern
 #endif
 
-#define LIQ_VERSION 20800
-#define LIQ_VERSION_STRING "2.8.0"
+#define LIQ_VERSION 21206
+#define LIQ_VERSION_STRING "2.12.6"
 
 #ifndef LIQ_PRIVATE
 #if defined(__GNUC__) || defined (__llvm__)
@@ -57,17 +57,29 @@ typedef enum liq_error {
     LIQ_BITMAP_NOT_AVAILABLE,
     LIQ_BUFFER_TOO_SMALL,
     LIQ_INVALID_POINTER,
+    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;
+    unsigned int count;
+} liq_histogram_entry;
 
 LIQ_EXPORT LIQ_USERESULT liq_attr* liq_attr_create(void);
 LIQ_EXPORT LIQ_USERESULT liq_attr* liq_attr_create_with_allocator(void* (*malloc)(size_t), void (*free)(void*));
-LIQ_EXPORT LIQ_USERESULT liq_attr* liq_attr_copy(liq_attr *orig) LIQ_NONNULL;
+LIQ_EXPORT LIQ_USERESULT liq_attr* liq_attr_copy(const liq_attr *orig) LIQ_NONNULL;
 LIQ_EXPORT void liq_attr_destroy(liq_attr *attr) LIQ_NONNULL;
 
-LIQ_EXPORT LIQ_USERESULT liq_histogram* liq_histogram_create(liq_attr* attr);
-LIQ_EXPORT LIQ_USERESULT liq_error liq_histogram_add_image(liq_histogram *hist, liq_attr *attr, liq_image* image);
+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;
@@ -100,6 +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;
@@ -117,10 +131,10 @@ LIQ_EXPORT LIQ_USERESULT const liq_palette *liq_get_palette(liq_result *result)
 LIQ_EXPORT liq_error liq_write_remapped_image(liq_result *result, liq_image *input_image, void *buffer, size_t buffer_size) LIQ_NONNULL;
 LIQ_EXPORT liq_error liq_write_remapped_image_rows(liq_result *result, liq_image *input_image, unsigned char **row_pointers) LIQ_NONNULL;
 
-LIQ_EXPORT double liq_get_quantization_error(liq_result *result) LIQ_NONNULL;
-LIQ_EXPORT int liq_get_quantization_quality(liq_result *result) LIQ_NONNULL;
-LIQ_EXPORT double liq_get_remapping_error(liq_result *result) LIQ_NONNULL;
-LIQ_EXPORT int liq_get_remapping_quality(liq_result *result) LIQ_NONNULL;
+LIQ_EXPORT double liq_get_quantization_error(const liq_result *result) LIQ_NONNULL;
+LIQ_EXPORT int liq_get_quantization_quality(const liq_result *result) LIQ_NONNULL;
+LIQ_EXPORT double liq_get_remapping_error(const liq_result *result) LIQ_NONNULL;
+LIQ_EXPORT int liq_get_remapping_quality(const liq_result *result) LIQ_NONNULL;
 
 LIQ_EXPORT void liq_result_destroy(liq_result *) LIQ_NONNULL;
 

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 <http://www.gnu.org/licenses/>.
-*/
-/*
-** 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 <stdlib.h>
@@ -38,7 +15,7 @@
 
 #define index_of_channel(ch) (offsetof(f_pixel,ch)/sizeof(float))
 
-static f_pixel averagepixels(unsigned int clrs, const hist_item achv[], const f_pixel center);
+static f_pixel averagepixels(unsigned int clrs, const hist_item achv[]);
 
 struct box {
     f_pixel color;
@@ -63,7 +40,7 @@ static f_pixel box_variance(const hist_item achv[], const struct box *box)
     double variancea=0, variancer=0, varianceg=0, varianceb=0;
 
     for(unsigned int i = 0; i < box->colors; ++i) {
-        f_pixel px = achv[box->ind + i].acolor;
+        const f_pixel px = achv[box->ind + i].acolor;
         double weight = achv[box->ind + i].adjusted_weight;
         variancea += variance_diff(mean.a - px.a, 2.0/256.0)*weight;
         variancer += variance_diff(mean.r - px.r, 1.0/256.0)*weight;
@@ -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 and >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 */
@@ -208,20 +185,29 @@ static double prepare_sort(struct box *b, hist_item achv[])
      ** Sort dimensions by their variance, and then sort colors first by dimension with highest variance
      */
     channelvariance channels[4] = {
+        {index_of_channel(a), b->variance.a},
         {index_of_channel(r), b->variance.r},
         {index_of_channel(g), b->variance.g},
         {index_of_channel(b), b->variance.b},
-        {index_of_channel(a), b->variance.a},
     };
 
     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;
+#if __GNUC__ >= 9
+    #pragma omp parallel for if (colors > 25000) \
+        schedule(static) default(none) shared(achv, channels, colors, ind1)
+#else
+    #pragma omp parallel for if (colors > 25000) \
+        schedule(static) default(none) shared(achv, channels)
+#endif
+    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 +215,8 @@ 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;
 }
@@ -245,13 +233,13 @@ static f_pixel get_median(const struct box *b, hist_item achv[])
 
     // technically the second color is not guaranteed to be sorted correctly
     // but most of the time it is good enough to be useful
-    return averagepixels(2, &achv[b->ind + median_start], (f_pixel){0.5,0.5,0.5,0.5});
+    return averagepixels(2, &achv[b->ind + median_start]);
 }
 
 /*
  ** 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++) {
@@ -278,13 +266,11 @@ static int best_splittable_box(struct box* bv, unsigned int boxes, const double
 inline static double color_weight(f_pixel median, hist_item h)
 {
     float diff = colordifference(median, h.acolor);
-    // if color is "good enough", don't split further
-    if (diff < 1.f/256.f/256.f) diff /= 2.f;
     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 colormap *map, 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[])
 {
@@ -323,6 +309,17 @@ static bool total_box_error_below_target(double target_mse, struct box bv[], uns
     return true;
 }
 
+static void box_init(struct box *box, const hist_item *achv, const unsigned int ind, const unsigned int colors, const double sum) {
+    box->ind = ind;
+    box->colors = colors;
+    box->sum = sum;
+    box->total_error = -1;
+
+    box->color = averagepixels(colors, &achv[ind]);
+    box->variance = box_variance(achv, box);
+    box->max_error = box_max_error(achv, box);
+}
+
 /*
  ** Here is the fun part, the median-cut colormap generator.  This is based
  ** on Paul Heckbert's paper, "Color Image Quantization for Frame Buffer
@@ -331,89 +328,77 @@ static bool total_box_error_below_target(double target_mse, struct box bv[], uns
 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.
      */
-    bv[0].ind = 0;
-    bv[0].colors = hist->size;
-    bv[0].color = averagepixels(bv[0].colors, &achv[bv[0].ind], (f_pixel){0.5,0.5,0.5,0.5});
-    bv[0].variance = box_variance(achv, &bv[0]);
-    bv[0].max_error = box_max_error(achv, &bv[0]);
-    bv[0].sum = 0;
-    bv[0].total_error = -1;
-    for(unsigned int i=0; i < bv[0].colors; i++) bv[0].sum += achv[i].adjusted_weight;
+    {
+        double sum = 0;
+        for(unsigned int i=0; i < hist->size; i++) {
+            sum += achv[i].adjusted_weight;
+        }
+        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.
 
-        const f_pixel previous_center = bv[bi].color;
-        bv[bi].colors = break_at;
-        bv[bi].sum = lowersum;
-        bv[bi].color = averagepixels(bv[bi].colors, &achv[bv[bi].ind], previous_center);
-        bv[bi].total_error = -1;
-        bv[bi].variance = box_variance(achv, &bv[bi]);
-        bv[bi].max_error = box_max_error(achv, &bv[bi]);
-        bv[boxes].ind = indx + break_at;
-        bv[boxes].colors = clrs - break_at;
-        bv[boxes].sum = sm - lowersum;
-        bv[boxes].color = averagepixels(bv[boxes].colors, &achv[bv[boxes].ind], previous_center);
-        bv[boxes].total_error = -1;
-        bv[boxes].variance = box_variance(achv, &bv[boxes]);
-        bv[boxes].max_error = box_max_error(achv, &bv[boxes]);
+             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;
+
+            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;
+            }
         }
     }
 
     colormap *map = pam_colormap(boxes, malloc, free);
     set_colormap_from_boxes(map, bv, boxes, achv);
 
-    adjust_histogram(achv, map, bv, boxes);
+    adjust_histogram(achv, bv, boxes);
 
     return map;
 }
@@ -440,29 +425,30 @@ static void set_colormap_from_boxes(colormap *map, struct box* bv, unsigned int
 }
 
 /* increase histogram popularity by difference from the final color (this is used as part of feedback loop) */
-static void adjust_histogram(hist_item *achv, const colormap *map, const struct box* bv, unsigned int boxes)
+static void adjust_histogram(hist_item *achv, const struct box* bv, unsigned int boxes)
 {
     for(unsigned int bi = 0; bi < boxes; ++bi) {
         for(unsigned int i=bv[bi].ind; i < bv[bi].ind+bv[bi].colors; i++) {
-            achv[i].adjusted_weight *= sqrt(1.0 +colordifference(map->palette[bi].acolor, achv[i].acolor)/2.0);
             achv[i].tmp.likely_colormap_index = bi;
         }
     }
 }
 
-static f_pixel averagepixels(unsigned int clrs, const hist_item achv[], f_pixel center)
+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;
 
         sum += weight;
+        a += px.a * weight;
         r += px.r * weight;
         g += px.g * weight;
         b += px.b * weight;
-        a += px.a * weight;
     }
 
     if (sum) {

mediancut.h

@@ -1,2 +1,6 @@
+#ifndef MEDIANCUT_H
+#define MEDIANCUT_H
 
 LIQ_PRIVATE colormap *mediancut(histogram *hist, unsigned int newcolors, const double target_mse, const double max_mse, void* (*malloc)(size_t), void (*free)(void*));
+
+#endif

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 <http://www.gnu.org/licenses/>.
+** © 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;
     }

mempool.h

@@ -4,10 +4,10 @@
 #include <stddef.h>
 
 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

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 <http://www.gnu.org/licenses/>.
+** © 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;
     }
@@ -117,11 +106,11 @@ static vp_node *vp_create_node(mempool *m, vp_sort_tmp *indexes, int num_indexes
     return node;
 }
 
-LIQ_PRIVATE struct nearest_map *nearest_init(const colormap *map, const bool fast) {
-    mempool m = NULL;
+LIQ_PRIVATE struct nearest_map *nearest_init(const colormap *map) {
+    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;

nearest.h

@@ -2,7 +2,13 @@
 //  nearest.h
 //  pngquant
 //
+
+#ifndef NEAREST_H
+#define NEAREST_H
+
 struct nearest_map;
-LIQ_PRIVATE struct nearest_map *nearest_init(const colormap *palette, const bool fast);
+LIQ_PRIVATE struct nearest_map *nearest_init(const colormap *palette);
 LIQ_PRIVATE unsigned int nearest_search(const struct nearest_map *map, const f_pixel *px, const int palette_index_guess, float *diff);
 LIQ_PRIVATE void nearest_free(struct nearest_map *map);
+
+#endif

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 <stdlib.h>
@@ -22,28 +16,19 @@
 
 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)
 {
-    const unsigned int maxacolors = acht->maxcolors, ignorebits = acht->ignorebits;
+    const unsigned int ignorebits = acht->ignorebits;
     const unsigned int channel_mask = 255U>>ignorebits<<ignorebits;
     const unsigned int channel_hmask = (255U>>ignorebits) ^ 0xFFU;
     const unsigned int posterize_mask = channel_mask << 24 | channel_mask << 16 | channel_mask << 8 | channel_mask;
     const unsigned int posterize_high_mask = channel_hmask << 24 | channel_hmask << 16 | channel_hmask << 8 | channel_hmask;
-    struct acolorhist_arr_head *const buckets = acht->buckets;
 
-    unsigned int colors = acht->colors;
     const unsigned int hash_size = acht->hash_size;
 
-    const unsigned int stacksize = sizeof(acht->freestack)/sizeof(acht->freestack[0]);
-    struct acolorhist_arr_item **freestack = acht->freestack;
-    unsigned int freestackp=acht->freestackp;
-
     /* 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]};
@@ -51,27 +36,50 @@ 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)) {
+                return false;
+            }
+        }
+
+    }
+    acht->cols = cols;
+    acht->rows += rows;
+    return true;
+}
+
+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.
              */
-            struct acolorhist_arr_head *achl = &buckets[hash];
+            struct acolorhist_arr_head *achl = &acht->buckets[hash];
             if (achl->inline1.color.l == px.l && achl->used) {
                 achl->inline1.perceptual_weight += boost;
-                continue;
+                return true;
             }
             if (achl->used) {
                 if (achl->used > 1) {
                     if (achl->inline2.color.l == px.l) {
                         achl->inline2.perceptual_weight += boost;
-                        continue;
+                        return true;
                     }
                     // other items are stored as an array (which gets reallocated if needed)
                     struct acolorhist_arr_item *other_items = achl->other_items;
@@ -79,7 +87,7 @@ LIQ_PRIVATE bool pam_computeacolorhash(struct acolorhash_table *acht, const rgba
                     for (; i < achl->used-2; i++) {
                         if (other_items[i].color.l == px.l) {
                             other_items[i].perceptual_weight += boost;
-                            goto continue_outer_loop;
+                            return true;
                         }
                     }
 
@@ -90,13 +98,11 @@ LIQ_PRIVATE bool pam_computeacolorhash(struct acolorhash_table *acht, const rgba
                             .perceptual_weight = boost,
                         };
                         achl->used++;
-                        ++colors;
-                        continue;
+                        ++acht->colors;
+                        return true;
                     }
 
-                    if (++colors > maxacolors) {
-                        acht->colors = colors;
-                        acht->freestackp = freestackp;
+                    if (++acht->colors > acht->maxcolors) {
                         return false;
                     }
 
@@ -104,22 +110,24 @@ LIQ_PRIVATE bool pam_computeacolorhash(struct acolorhash_table *acht, const rgba
                     unsigned int capacity;
                     if (!other_items) { // there was no array previously, alloc "small" array
                         capacity = 8;
-                        if (freestackp <= 0) {
+                        if (acht->freestackp <= 0) {
                             // estimate how many colors are going to be + headroom
-                            const size_t mempool_size = ((acht->rows + rows-row) * 2 * colors / (acht->rows + row + 1) + 1024) * sizeof(struct acolorhist_arr_item);
+                            const size_t mempool_size = ((acht->rows + rows-row) * 2 * acht->colors / (acht->rows + row + 1) + 1024) * sizeof(struct acolorhist_arr_item);
                             new_items = mempool_alloc(&acht->mempool, sizeof(struct acolorhist_arr_item)*capacity, mempool_size);
                         } else {
                             // freestack stores previously freed (reallocated) arrays that can be reused
                             // (all pesimistically assumed to be capacity = 8)
-                            new_items = freestack[--freestackp];
+                            new_items = acht->freestack[--acht->freestackp];
                         }
                     } else {
+                        const unsigned int stacksize = sizeof(acht->freestack)/sizeof(acht->freestack[0]);
+
                         // simply reallocs and copies array to larger capacity
                         capacity = achl->capacity*2 + 16;
-                        if (freestackp < stacksize-1) {
-                            freestack[freestackp++] = other_items;
+                        if (acht->freestackp < stacksize-1) {
+                            acht->freestack[acht->freestackp++] = other_items;
                         }
-                        const size_t mempool_size = ((acht->rows + rows-row) * 2 * colors / (acht->rows + row + 1) + 32*capacity) * sizeof(struct acolorhist_arr_item);
+                        const size_t mempool_size = ((acht->rows + rows-row) * 2 * acht->colors / (acht->rows + row + 1) + 32*capacity) * sizeof(struct acolorhist_arr_item);
                         new_items = mempool_alloc(&acht->mempool, sizeof(struct acolorhist_arr_item)*capacity, mempool_size);
                         if (!new_items) return false;
                         memcpy(new_items, other_items, sizeof(other_items[0])*achl->capacity);
@@ -137,23 +145,14 @@ LIQ_PRIVATE bool pam_computeacolorhash(struct acolorhash_table *acht, const rgba
                     achl->inline2.color.l = px.l;
                     achl->inline2.perceptual_weight = boost;
                     achl->used = 2;
-                    ++colors;
+                    ++acht->colors;
                 }
             } else {
                 achl->inline1.color.l = px.l;
                 achl->inline1.perceptual_weight = boost;
                 achl->used = 1;
-                ++colors;
+                ++acht->colors;
             }
-
-            continue_outer_loop:;
-        }
-
-    }
-    acht->colors = colors;
-    acht->cols = cols;
-    acht->rows += rows;
-    acht->freestackp = freestackp;
     return true;
 }
 
@@ -162,7 +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);
@@ -177,10 +176,16 @@ LIQ_PRIVATE struct acolorhash_table *pam_allocacolorhash(unsigned int maxcolors,
     return t;
 }
 
-#define PAM_ADD_TO_HIST(entry) { \
-    hist->achv[j].acolor = to_f(gamma_lut, entry.color.rgba); \
-    total_weight += hist->achv[j].adjusted_weight = hist->achv[j].perceptual_weight = MIN(entry.perceptual_weight, max_perceptual_weight); \
-    ++j; \
+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)
+{
+    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;
 }
 
 LIQ_PRIVATE histogram *pam_acolorhashtoacolorhist(const struct acolorhash_table *acht, const double gamma, void* (*malloc)(size_t), void (*free)(void*))
@@ -203,22 +208,30 @@ 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) {
-            PAM_ADD_TO_HIST(achl->inline1);
+            total_weight += pam_add_to_hist(gamma_lut, hist->achv, &j, &achl->inline1, max_perceptual_weight);
 
             if (achl->used > 1) {
-                PAM_ADD_TO_HIST(achl->inline2);
+                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++) {
-                    PAM_ADD_TO_HIST(achl->other_items[k]);
+                    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;
+    for(unsigned int k=0; k < hist->size; k++) {
+        hist->achv[k].tmp.likely_colormap_index = 0;
+    }
+    if (!j) {
+        pam_freeacolorhist(hist);
+        return NULL;
+    }
     return hist;
 }
 

pam.h

@@ -16,6 +16,12 @@
 #ifndef PAM_H
 #define PAM_H
 
+// accidental debug assertions make color search much slower,
+// so force assertions off if there's no explicit setting
+#if !defined(NDEBUG) && !defined(DEBUG)
+#define NDEBUG
+#endif
+
 #include <math.h>
 #include <assert.h>
 #include <stdlib.h>
@@ -62,6 +68,12 @@
 #  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 +97,7 @@ 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);
 
@@ -93,8 +105,8 @@ LIQ_PRIVATE void to_f_set_gamma(float gamma_lut[], const double gamma);
  Converts 8-bit color to internal gamma and premultiplied alpha.
  (premultiplied color space is much better for blending of semitransparent colors)
  */
-ALWAYS_INLINE static f_pixel to_f(const float gamma_lut[], const rgba_pixel px);
-inline static f_pixel to_f(const float gamma_lut[], const rgba_pixel px)
+ALWAYS_INLINE static f_pixel rgba_to_f(const float gamma_lut[], const rgba_pixel px);
+inline static f_pixel rgba_to_f(const float gamma_lut[], const rgba_pixel px)
 {
     float a = px.a/255.f;
 
@@ -106,7 +118,7 @@ inline static f_pixel to_f(const float gamma_lut[], const rgba_pixel px)
     };
 }
 
-inline static rgba_pixel to_rgb(const float gamma, const f_pixel px)
+inline static rgba_pixel f_to_rgb(const float gamma, const f_pixel px)
 {
     if (px.a < 1.f/256.f) {
         return (rgba_pixel){0,0,0,0};
@@ -226,7 +238,7 @@ typedef struct {
 typedef struct {
     f_pixel acolor;
     float popularity;
-    bool fixed; // if true it's user-supplied and must not be changed (e.g in voronoi iteration)
+    bool fixed; // if true it's user-supplied and must not be changed (e.g in K-Means iteration)
 } colormap_item;
 
 typedef struct colormap {
@@ -238,7 +250,7 @@ typedef struct colormap {
 
 struct acolorhist_arr_item {
     union rgba_as_int color;
-    float perceptual_weight;
+    unsigned int perceptual_weight;
 };
 
 struct acolorhist_arr_head {
@@ -260,6 +272,7 @@ 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, unsigned int boost, union rgba_as_int px, unsigned int row, unsigned int rows);
 
 LIQ_PRIVATE void pam_freeacolorhist(histogram *h);
 

viter.c

@@ -1,107 +0,0 @@
-/*
-© 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 <http://www.gnu.org/licenses/>.
-*/
-
-#include "libimagequant.h"
-#include "pam.h"
-#include "viter.h"
-#include "nearest.h"
-#include <stdlib.h>
-#include <string.h>
-
-#ifdef _OPENMP
-#include <omp.h>
-#else
-#define omp_get_max_threads() 1
-#define omp_get_thread_num() 0
-#endif
-
-/*
- * Voronoi iteration: new palette color is computed from weighted average of colors that map to that palette entry.
- */
-LIQ_PRIVATE void viter_init(const colormap *map, const unsigned int max_threads, viter_state average_color[])
-{
-    memset(average_color, 0, sizeof(average_color[0])*(VITER_CACHE_LINE_GAP+map->colors)*max_threads);
-}
-
-LIQ_PRIVATE void viter_update_color(const f_pixel acolor, const float value, const colormap *map, unsigned int match, const unsigned int thread, viter_state average_color[])
-{
-    match += thread * (VITER_CACHE_LINE_GAP+map->colors);
-    average_color[match].a += acolor.a * value;
-    average_color[match].r += acolor.r * value;
-    average_color[match].g += acolor.g * value;
-    average_color[match].b += acolor.b * value;
-    average_color[match].total += value;
-}
-
-LIQ_PRIVATE void viter_finalize(colormap *map, const unsigned int max_threads, const viter_state average_color[])
-{
-    for (unsigned int i=0; i < map->colors; i++) {
-        double a=0, r=0, g=0, b=0, total=0;
-
-        // Aggregate results from all threads
-        for(unsigned int t=0; t < max_threads; t++) {
-            const unsigned int offset = (VITER_CACHE_LINE_GAP+map->colors) * t + i;
-
-            a += average_color[offset].a;
-            r += average_color[offset].r;
-            g += average_color[offset].g;
-            b += average_color[offset].b;
-            total += average_color[offset].total;
-        }
-
-        if (total && !map->palette[i].fixed) {
-            map->palette[i].acolor = (f_pixel){
-                .a = a / total,
-                .r = r / total,
-                .g = g / total,
-                .b = b / total,
-            };
-            map->palette[i].popularity = total;
-        }
-    }
-}
-
-LIQ_PRIVATE double viter_do_iteration(histogram *hist, colormap *const map, viter_callback callback, const bool fast_palette)
-{
-    const unsigned int max_threads = omp_get_max_threads();
-    viter_state average_color[(VITER_CACHE_LINE_GAP+map->colors) * max_threads];
-    viter_init(map, max_threads, average_color);
-    struct nearest_map *const n = nearest_init(map, fast_palette);
-    hist_item *const achv = hist->achv;
-    const int hist_size = hist->size;
-
-    double total_diff=0;
-    #pragma omp parallel for if (hist_size > 3000) \
-        schedule(static) default(none) shared(average_color,callback) reduction(+:total_diff)
-    for(int j=0; j < hist_size; j++) {
-        float diff;
-        unsigned int match = nearest_search(n, &achv[j].acolor, achv[j].tmp.likely_colormap_index, &diff);
-        achv[j].tmp.likely_colormap_index = match;
-        total_diff += diff * achv[j].perceptual_weight;
-
-        viter_update_color(achv[j].acolor, achv[j].perceptual_weight, map, match, omp_get_thread_num(), average_color);
-
-        if (callback) callback(&achv[j], diff);
-    }
-
-    nearest_free(n);
-    viter_finalize(map, max_threads, average_color);
-
-    return total_diff / hist->total_perceptual_weight;
-}

viter.h

@@ -1,19 +0,0 @@
-
-#ifndef VITER_H
-#define VITER_H
-
-// Spread memory touched by different threads at least 64B apart which I assume is the cache line size. This should avoid memory write contention.
-#define VITER_CACHE_LINE_GAP ((64+sizeof(viter_state)-1)/sizeof(viter_state))
-
-typedef struct {
-    double a, r, g, b, total;
-} viter_state;
-
-typedef void (*viter_callback)(hist_item *item, float diff);
-
-LIQ_PRIVATE void viter_init(const colormap *map, const unsigned int max_threads, viter_state state[]);
-LIQ_PRIVATE void viter_update_color(const f_pixel acolor, const float value, const colormap *map, unsigned int match, const unsigned int thread, viter_state average_color[]);
-LIQ_PRIVATE void viter_finalize(colormap *map, const unsigned int max_threads, const viter_state state[]);
-LIQ_PRIVATE double viter_do_iteration(histogram *hist, colormap *const map, viter_callback callback, const bool fast_palette);
-
-#endif