xref: /third_party/ffmpeg/libavcodec/rpzaenc.c (revision cabdff1a)
1/*
2 * QuickTime RPZA Video Encoder
3 *
4 * This file is part of FFmpeg.
5 *
6 * FFmpeg is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
10 *
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14 * Lesser General Public License for more details.
15 *
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19 */
20
21/**
22 * @file rpzaenc.c
23 * QT RPZA Video Encoder by Todd Kirby <doubleshot@pacbell.net> and David Adler
24 */
25
26#include "libavutil/avassert.h"
27#include "libavutil/common.h"
28#include "libavutil/opt.h"
29
30#include "avcodec.h"
31#include "codec_internal.h"
32#include "encode.h"
33#include "put_bits.h"
34
35typedef struct RpzaContext {
36    AVClass *avclass;
37
38    int skip_frame_thresh;
39    int start_one_color_thresh;
40    int continue_one_color_thresh;
41    int sixteen_color_thresh;
42
43    AVFrame *prev_frame;    // buffer for previous source frame
44    PutBitContext pb;       // buffer for encoded frame data.
45
46    int frame_width;        // width in pixels of source frame
47    int frame_height;       // height in pixesl of source frame
48
49    int first_frame;        // flag set to one when the first frame is being processed
50                            // so that comparisons with previous frame data in not attempted
51} RpzaContext;
52
53typedef enum channel_offset {
54    RED = 2,
55    GREEN = 1,
56    BLUE = 0,
57} channel_offset;
58
59typedef struct rgb {
60    uint8_t r;
61    uint8_t g;
62    uint8_t b;
63} rgb;
64
65#define SQR(x) ((x) * (x))
66
67/* 15 bit components */
68#define GET_CHAN(color, chan) (((color) >> ((chan) * 5) & 0x1F) * 8)
69#define R(color) GET_CHAN(color, RED)
70#define G(color) GET_CHAN(color, GREEN)
71#define B(color) GET_CHAN(color, BLUE)
72
73typedef struct BlockInfo {
74    int row;
75    int col;
76    int block_width;
77    int block_height;
78    int image_width;
79    int image_height;
80    int block_index;
81    uint16_t start;
82    int rowstride;
83    int blocks_per_row;
84    int total_blocks;
85} BlockInfo;
86
87static void get_colors(uint8_t *min, uint8_t *max, uint8_t color4[4][3])
88{
89    uint8_t step;
90
91    color4[0][0] = min[0];
92    color4[0][1] = min[1];
93    color4[0][2] = min[2];
94
95    color4[3][0] = max[0];
96    color4[3][1] = max[1];
97    color4[3][2] = max[2];
98
99    // red components
100    step = (color4[3][0] - color4[0][0] + 1) / 3;
101    color4[1][0] = color4[0][0] + step;
102    color4[2][0] = color4[3][0] - step;
103
104    // green components
105    step = (color4[3][1] - color4[0][1] + 1) / 3;
106    color4[1][1] = color4[0][1] + step;
107    color4[2][1] = color4[3][1] - step;
108
109    // blue components
110    step = (color4[3][2] - color4[0][2] + 1) / 3;
111    color4[1][2] = color4[0][2] + step;
112    color4[2][2] = color4[3][2] - step;
113}
114
115/* Fill BlockInfo struct with information about a 4x4 block of the image */
116static int get_block_info(BlockInfo *bi, int block)
117{
118    bi->row = block / bi->blocks_per_row;
119    bi->col = block % bi->blocks_per_row;
120
121    // test for right edge block
122    if (bi->col == bi->blocks_per_row - 1 && (bi->image_width % 4) != 0) {
123        bi->block_width = bi->image_width % 4;
124    } else {
125        bi->block_width = 4;
126    }
127
128    // test for bottom edge block
129    if (bi->row == (bi->image_height / 4) && (bi->image_height % 4) != 0) {
130        bi->block_height = bi->image_height % 4;
131    } else {
132        bi->block_height = 4;
133    }
134
135    return block ? (bi->col * 4) + (bi->row * bi->rowstride * 4) : 0;
136}
137
138static uint16_t rgb24_to_rgb555(uint8_t *rgb24)
139{
140    uint16_t rgb555 = 0;
141    uint32_t r, g, b;
142
143    r = rgb24[0] >> 3;
144    g = rgb24[1] >> 3;
145    b = rgb24[2] >> 3;
146
147    rgb555 |= (r << 10);
148    rgb555 |= (g << 5);
149    rgb555 |= (b << 0);
150
151    return rgb555;
152}
153
154/*
155 * Returns the total difference between two 24 bit color values
156 */
157static int diff_colors(uint8_t *colorA, uint8_t *colorB)
158{
159    int tot;
160
161    tot  = SQR(colorA[0] - colorB[0]);
162    tot += SQR(colorA[1] - colorB[1]);
163    tot += SQR(colorA[2] - colorB[2]);
164
165    return tot;
166}
167
168/*
169 * Returns the maximum channel difference
170 */
171static int max_component_diff(uint16_t *colorA, uint16_t *colorB)
172{
173    int diff, max = 0;
174
175    diff = FFABS(R(colorA[0]) - R(colorB[0]));
176    if (diff > max) {
177        max = diff;
178    }
179    diff = FFABS(G(colorA[0]) - G(colorB[0]));
180    if (diff > max) {
181        max = diff;
182    }
183    diff = FFABS(B(colorA[0]) - B(colorB[0]));
184    if (diff > max) {
185        max = diff;
186    }
187    return max * 8;
188}
189
190/*
191 * Find the channel that has the largest difference between minimum and maximum
192 * color values. Put the minimum value in min, maximum in max and the channel
193 * in chan.
194 */
195static void get_max_component_diff(BlockInfo *bi, uint16_t *block_ptr,
196                                   uint8_t *min, uint8_t *max, channel_offset *chan)
197{
198    int x, y;
199    uint8_t min_r, max_r, min_g, max_g, min_b, max_b;
200    uint8_t r, g, b;
201
202    // fix warning about uninitialized vars
203    min_r = min_g = min_b = UINT8_MAX;
204    max_r = max_g = max_b = 0;
205
206    // loop thru and compare pixels
207    for (y = 0; y < bi->block_height; y++) {
208        for (x = 0; x < bi->block_width; x++) {
209            // TODO:  optimize
210            min_r = FFMIN(R(block_ptr[x]), min_r);
211            min_g = FFMIN(G(block_ptr[x]), min_g);
212            min_b = FFMIN(B(block_ptr[x]), min_b);
213
214            max_r = FFMAX(R(block_ptr[x]), max_r);
215            max_g = FFMAX(G(block_ptr[x]), max_g);
216            max_b = FFMAX(B(block_ptr[x]), max_b);
217        }
218        block_ptr += bi->rowstride;
219    }
220
221    r = max_r - min_r;
222    g = max_g - min_g;
223    b = max_b - min_b;
224
225    if (r > g && r > b) {
226        *max = max_r;
227        *min = min_r;
228        *chan = RED;
229    } else if (g > b && g >= r) {
230        *max = max_g;
231        *min = min_g;
232        *chan = GREEN;
233    } else {
234        *max = max_b;
235        *min = min_b;
236        *chan = BLUE;
237    }
238}
239
240/*
241 * Compare two 4x4 blocks to determine if the total difference between the
242 * blocks is greater than the thresh parameter. Returns -1 if difference
243 * exceeds threshold or zero otherwise.
244 */
245static int compare_blocks(uint16_t *block1, uint16_t *block2, BlockInfo *bi, int thresh)
246{
247    int x, y, diff = 0;
248    for (y = 0; y < bi->block_height; y++) {
249        for (x = 0; x < bi->block_width; x++) {
250            diff = max_component_diff(&block1[x], &block2[x]);
251            if (diff >= thresh) {
252                return -1;
253            }
254        }
255        block1 += bi->rowstride;
256        block2 += bi->rowstride;
257    }
258    return 0;
259}
260
261/*
262 * Determine the fit of one channel to another within a 4x4 block. This
263 * is used to determine the best palette choices for 4-color encoding.
264 */
265static int leastsquares(uint16_t *block_ptr, BlockInfo *bi,
266                        channel_offset xchannel, channel_offset ychannel,
267                        double *slope, double *y_intercept, double *correlation_coef)
268{
269    double sumx = 0, sumy = 0, sumx2 = 0, sumy2 = 0, sumxy = 0,
270           sumx_sq = 0, sumy_sq = 0, tmp, tmp2;
271    int i, j, count;
272    uint8_t x, y;
273
274    count = bi->block_height * bi->block_width;
275
276    if (count < 2)
277        return -1;
278
279    for (i = 0; i < bi->block_height; i++) {
280        for (j = 0; j < bi->block_width; j++) {
281            x = GET_CHAN(block_ptr[j], xchannel);
282            y = GET_CHAN(block_ptr[j], ychannel);
283            sumx += x;
284            sumy += y;
285            sumx2 += x * x;
286            sumy2 += y * y;
287            sumxy += x * y;
288        }
289        block_ptr += bi->rowstride;
290    }
291
292    sumx_sq = sumx * sumx;
293    tmp = (count * sumx2 - sumx_sq);
294
295    // guard against div/0
296    if (tmp == 0)
297        return -2;
298
299    sumy_sq = sumy * sumy;
300
301    *slope = (sumx * sumy - sumxy) / tmp;
302    *y_intercept = (sumy - (*slope) * sumx) / count;
303
304    tmp2 = count * sumy2 - sumy_sq;
305    if (tmp2 == 0) {
306        *correlation_coef = 0.0;
307    } else {
308        *correlation_coef = (count * sumxy - sumx * sumy) /
309            sqrt(tmp * tmp2);
310    }
311
312    return 0; // success
313}
314
315/*
316 * Determine the amount of error in the leastsquares fit.
317 */
318static int calc_lsq_max_fit_error(uint16_t *block_ptr, BlockInfo *bi,
319                                  int min, int max, int tmp_min, int tmp_max,
320                                  channel_offset xchannel, channel_offset ychannel)
321{
322    int i, j, x, y;
323    int err;
324    int max_err = 0;
325
326    for (i = 0; i < bi->block_height; i++) {
327        for (j = 0; j < bi->block_width; j++) {
328            int x_inc, lin_y, lin_x;
329            x = GET_CHAN(block_ptr[j], xchannel);
330            y = GET_CHAN(block_ptr[j], ychannel);
331
332            /* calculate x_inc as the 4-color index (0..3) */
333            x_inc = floor( (x - min) * 3.0 / (max - min) + 0.5);
334            x_inc = FFMAX(FFMIN(3, x_inc), 0);
335
336            /* calculate lin_y corresponding to x_inc */
337            lin_y = (int)(tmp_min + (tmp_max - tmp_min) * x_inc / 3.0 + 0.5);
338
339            err = FFABS(lin_y - y);
340            if (err > max_err)
341                max_err = err;
342
343            /* calculate lin_x corresponding to x_inc */
344            lin_x = (int)(min + (max - min) * x_inc / 3.0 + 0.5);
345
346            err = FFABS(lin_x - x);
347            if (err > max_err)
348                max_err += err;
349        }
350        block_ptr += bi->rowstride;
351    }
352
353    return max_err;
354}
355
356/*
357 * Find the closest match to a color within the 4-color palette
358 */
359static int match_color(uint16_t *color, uint8_t colors[4][3])
360{
361    int ret = 0;
362    int smallest_variance = INT_MAX;
363    uint8_t dithered_color[3];
364
365    for (int channel = 0; channel < 3; channel++) {
366        dithered_color[channel] = GET_CHAN(color[0], channel);
367    }
368
369    for (int palette_entry = 0; palette_entry < 4; palette_entry++) {
370        int variance = diff_colors(dithered_color, colors[palette_entry]);
371
372        if (variance < smallest_variance) {
373            smallest_variance = variance;
374            ret = palette_entry;
375        }
376    }
377
378    return ret;
379}
380
381/*
382 * Encode a block using the 4-color opcode and palette. return number of
383 * blocks encoded (until we implement multi-block 4 color runs this will
384 * always be 1)
385 */
386static int encode_four_color_block(uint8_t *min_color, uint8_t *max_color,
387                                   PutBitContext *pb, uint16_t *block_ptr, BlockInfo *bi)
388{
389    int x, y, idx;
390    uint8_t color4[4][3];
391    uint16_t rounded_max, rounded_min;
392
393    // round min and max wider
394    rounded_min = rgb24_to_rgb555(min_color);
395    rounded_max = rgb24_to_rgb555(max_color);
396
397    // put a and b colors
398    // encode 4 colors = first 16 bit color with MSB zeroed and...
399    put_bits(pb, 16, rounded_max & ~0x8000);
400    // ...second 16 bit color with MSB on.
401    put_bits(pb, 16, rounded_min | 0x8000);
402
403    get_colors(min_color, max_color, color4);
404
405    for (y = 0; y < 4; y++) {
406        for (x = 0; x < 4; x++) {
407            idx = match_color(&block_ptr[x], color4);
408            put_bits(pb, 2, idx);
409        }
410        block_ptr += bi->rowstride;
411    }
412    return 1; // num blocks encoded
413}
414
415/*
416 * Copy a 4x4 block from the current frame buffer to the previous frame buffer.
417 */
418static void update_block_in_prev_frame(const uint16_t *src_pixels,
419                                       uint16_t *dest_pixels,
420                                       const BlockInfo *bi, int block_counter)
421{
422    const int y_size = FFMIN(4, bi->image_height - bi->row * 4);
423
424    for (int y = 0; y < y_size; y++) {
425        memcpy(dest_pixels, src_pixels, 8);
426        dest_pixels += bi->rowstride;
427        src_pixels += bi->rowstride;
428    }
429}
430
431/*
432 * update statistics for the specified block. If first_block,
433 * it initializes the statistics.  Otherwise it updates the statistics IF THIS
434 * BLOCK IS SUITABLE TO CONTINUE A 1-COLOR RUN. That is, it checks whether
435 * the range of colors (since the routine was called first_block != 0) are
436 * all close enough intensities to be represented by a single color.
437
438 * The routine returns 0 if this block is too different to be part of
439 * the same run of 1-color blocks. The routine returns 1 if this
440 * block can be part of the same 1-color block run.
441
442 * If the routine returns 1, it also updates its arguments to include
443 * the statistics of this block. Otherwise, the stats are unchanged
444 * and don't include the current block.
445 */
446static int update_block_stats(RpzaContext *s, BlockInfo *bi, uint16_t *block,
447                              uint8_t min_color[3], uint8_t max_color[3],
448                              int *total_rgb, int *total_pixels,
449                              uint8_t avg_color[3], int first_block)
450{
451    int x, y;
452    int is_in_range;
453    int total_pixels_blk;
454    int threshold;
455
456    uint8_t min_color_blk[3], max_color_blk[3];
457    int total_rgb_blk[3];
458    uint8_t avg_color_blk[3];
459
460    if (first_block) {
461        min_color[0] = UINT8_MAX;
462        min_color[1] = UINT8_MAX;
463        min_color[2] = UINT8_MAX;
464        max_color[0] = 0;
465        max_color[1] = 0;
466        max_color[2] = 0;
467        total_rgb[0] = 0;
468        total_rgb[1] = 0;
469        total_rgb[2] = 0;
470        *total_pixels = 0;
471        threshold = s->start_one_color_thresh;
472    } else {
473        threshold = s->continue_one_color_thresh;
474    }
475
476    /*
477       The *_blk variables will include the current block.
478       Initialize them based on the blocks so far.
479     */
480    min_color_blk[0] = min_color[0];
481    min_color_blk[1] = min_color[1];
482    min_color_blk[2] = min_color[2];
483    max_color_blk[0] = max_color[0];
484    max_color_blk[1] = max_color[1];
485    max_color_blk[2] = max_color[2];
486    total_rgb_blk[0] = total_rgb[0];
487    total_rgb_blk[1] = total_rgb[1];
488    total_rgb_blk[2] = total_rgb[2];
489    total_pixels_blk = *total_pixels + bi->block_height * bi->block_width;
490
491    /*
492       Update stats for this block's pixels
493     */
494    for (y = 0; y < bi->block_height; y++) {
495        for (x = 0; x < bi->block_width; x++) {
496            total_rgb_blk[0] += R(block[x]);
497            total_rgb_blk[1] += G(block[x]);
498            total_rgb_blk[2] += B(block[x]);
499
500            min_color_blk[0] = FFMIN(R(block[x]), min_color_blk[0]);
501            min_color_blk[1] = FFMIN(G(block[x]), min_color_blk[1]);
502            min_color_blk[2] = FFMIN(B(block[x]), min_color_blk[2]);
503
504            max_color_blk[0] = FFMAX(R(block[x]), max_color_blk[0]);
505            max_color_blk[1] = FFMAX(G(block[x]), max_color_blk[1]);
506            max_color_blk[2] = FFMAX(B(block[x]), max_color_blk[2]);
507        }
508        block += bi->rowstride;
509    }
510
511    /*
512       Calculate average color including current block.
513     */
514    avg_color_blk[0] = total_rgb_blk[0] / total_pixels_blk;
515    avg_color_blk[1] = total_rgb_blk[1] / total_pixels_blk;
516    avg_color_blk[2] = total_rgb_blk[2] / total_pixels_blk;
517
518    /*
519       Are all the pixels within threshold of the average color?
520     */
521    is_in_range = (max_color_blk[0] - avg_color_blk[0] <= threshold &&
522                   max_color_blk[1] - avg_color_blk[1] <= threshold &&
523                   max_color_blk[2] - avg_color_blk[2] <= threshold &&
524                   avg_color_blk[0] - min_color_blk[0] <= threshold &&
525                   avg_color_blk[1] - min_color_blk[1] <= threshold &&
526                   avg_color_blk[2] - min_color_blk[2] <= threshold);
527
528    if (is_in_range) {
529        /*
530           Set the output variables to include this block.
531         */
532        min_color[0] = min_color_blk[0];
533        min_color[1] = min_color_blk[1];
534        min_color[2] = min_color_blk[2];
535        max_color[0] = max_color_blk[0];
536        max_color[1] = max_color_blk[1];
537        max_color[2] = max_color_blk[2];
538        total_rgb[0] = total_rgb_blk[0];
539        total_rgb[1] = total_rgb_blk[1];
540        total_rgb[2] = total_rgb_blk[2];
541        *total_pixels = total_pixels_blk;
542        avg_color[0] = avg_color_blk[0];
543        avg_color[1] = avg_color_blk[1];
544        avg_color[2] = avg_color_blk[2];
545    }
546
547    return is_in_range;
548}
549
550static void rpza_encode_stream(RpzaContext *s, const AVFrame *pict)
551{
552    BlockInfo bi;
553    int block_counter = 0;
554    int n_blocks;
555    int total_blocks;
556    int prev_block_offset;
557    int block_offset = 0;
558    uint8_t min = 0, max = 0;
559    channel_offset chan;
560    int i;
561    int tmp_min, tmp_max;
562    int total_rgb[3];
563    uint8_t avg_color[3];
564    int pixel_count;
565    uint8_t min_color[3], max_color[3];
566    double slope, y_intercept, correlation_coef;
567    uint16_t *src_pixels = (uint16_t *)pict->data[0];
568    uint16_t *prev_pixels = (uint16_t *)s->prev_frame->data[0];
569
570    /* Number of 4x4 blocks in frame. */
571    total_blocks = ((s->frame_width + 3) / 4) * ((s->frame_height + 3) / 4);
572
573    bi.image_width = s->frame_width;
574    bi.image_height = s->frame_height;
575    bi.rowstride = pict->linesize[0] / 2;
576
577    bi.blocks_per_row = (s->frame_width + 3) / 4;
578
579    while (block_counter < total_blocks) {
580        // SKIP CHECK
581        // make sure we have a valid previous frame and we're not writing
582        // a key frame
583        if (!s->first_frame) {
584            n_blocks = 0;
585            prev_block_offset = 0;
586
587            while (n_blocks < 32 && block_counter + n_blocks < total_blocks) {
588
589                block_offset = get_block_info(&bi, block_counter + n_blocks);
590
591                // multi-block opcodes cannot span multiple rows.
592                // If we're starting a new row, break out and write the opcode
593                /* TODO: Should eventually use bi.row here to determine when a
594                   row break occurs, but that is currently breaking the
595                   quicktime player. This is probably due to a bug in the
596                   way I'm calculating the current row.
597                 */
598                if (prev_block_offset && block_offset - prev_block_offset > 12) {
599                    break;
600                }
601
602                prev_block_offset = block_offset;
603
604                if (compare_blocks(&prev_pixels[block_offset],
605                                   &src_pixels[block_offset], &bi, s->skip_frame_thresh) != 0) {
606                    // write out skipable blocks
607                    if (n_blocks) {
608
609                        // write skip opcode
610                        put_bits(&s->pb, 8, 0x80 | (n_blocks - 1));
611                        block_counter += n_blocks;
612
613                        goto post_skip;
614                    }
615                    break;
616                }
617
618                /*
619                 * NOTE: we don't update skipped blocks in the previous frame buffer
620                 * since skipped needs always to be compared against the first skipped
621                 * block to avoid artifacts during gradual fade in/outs.
622                 */
623
624                // update_block_in_prev_frame(&src_pixels[block_offset],
625                //   &prev_pixels[block_offset], &bi, block_counter + n_blocks);
626
627                n_blocks++;
628            }
629
630            // we're either at the end of the frame or we've reached the maximum
631            // of 32 blocks in a run. Write out the run.
632            if (n_blocks) {
633                // write skip opcode
634                put_bits(&s->pb, 8, 0x80 | (n_blocks - 1));
635                block_counter += n_blocks;
636
637                continue;
638            }
639
640        } else {
641            block_offset = get_block_info(&bi, block_counter);
642        }
643post_skip :
644
645        // ONE COLOR CHECK
646        if (update_block_stats(s, &bi, &src_pixels[block_offset],
647                               min_color, max_color,
648                               total_rgb, &pixel_count, avg_color, 1)) {
649            prev_block_offset = block_offset;
650
651            n_blocks = 1;
652
653            /* update this block in the previous frame buffer */
654            update_block_in_prev_frame(&src_pixels[block_offset],
655                                       &prev_pixels[block_offset], &bi, block_counter + n_blocks);
656
657            // check for subsequent blocks with the same color
658            while (n_blocks < 32 && block_counter + n_blocks < total_blocks) {
659                block_offset = get_block_info(&bi, block_counter + n_blocks);
660
661                // multi-block opcodes cannot span multiple rows.
662                // If we've hit end of a row, break out and write the opcode
663                if (block_offset - prev_block_offset > 12) {
664                    break;
665                }
666
667                if (!update_block_stats(s, &bi, &src_pixels[block_offset],
668                                        min_color, max_color,
669                                        total_rgb, &pixel_count, avg_color, 0)) {
670                    break;
671                }
672
673                prev_block_offset = block_offset;
674
675                /* update this block in the previous frame buffer */
676                update_block_in_prev_frame(&src_pixels[block_offset],
677                                           &prev_pixels[block_offset], &bi, block_counter + n_blocks);
678
679                n_blocks++;
680            }
681
682            // write one color opcode.
683            put_bits(&s->pb, 8, 0xa0 | (n_blocks - 1));
684            // write color to encode.
685            put_bits(&s->pb, 16, rgb24_to_rgb555(avg_color));
686            // skip past the blocks we've just encoded.
687            block_counter += n_blocks;
688        } else { // FOUR COLOR CHECK
689            int err = 0;
690
691            // get max component diff for block
692            get_max_component_diff(&bi, &src_pixels[block_offset], &min, &max, &chan);
693
694            min_color[0] = 0;
695            max_color[0] = 0;
696            min_color[1] = 0;
697            max_color[1] = 0;
698            min_color[2] = 0;
699            max_color[2] = 0;
700
701            // run least squares against other two components
702            for (i = 0; i < 3; i++) {
703                if (i == chan) {
704                    min_color[i] = min;
705                    max_color[i] = max;
706                    continue;
707                }
708
709                slope = y_intercept = correlation_coef = 0;
710
711                if (leastsquares(&src_pixels[block_offset], &bi, chan, i,
712                                 &slope, &y_intercept, &correlation_coef)) {
713                    min_color[i] = GET_CHAN(src_pixels[block_offset], i);
714                    max_color[i] = GET_CHAN(src_pixels[block_offset], i);
715                } else {
716                    tmp_min = (int)(0.5 + min * slope + y_intercept);
717                    tmp_max = (int)(0.5 + max * slope + y_intercept);
718
719                    av_assert0(tmp_min <= tmp_max);
720                    // clamp min and max color values
721                    tmp_min = av_clip_uint8(tmp_min);
722                    tmp_max = av_clip_uint8(tmp_max);
723
724                    err = FFMAX(calc_lsq_max_fit_error(&src_pixels[block_offset], &bi,
725                                                       min, max, tmp_min, tmp_max, chan, i), err);
726
727                    min_color[i] = tmp_min;
728                    max_color[i] = tmp_max;
729                }
730            }
731
732            if (err > s->sixteen_color_thresh) { // DO SIXTEEN COLOR BLOCK
733                uint16_t *row_ptr;
734                int y_size, x_size, rgb555;
735
736                block_offset = get_block_info(&bi, block_counter);
737
738                row_ptr = &src_pixels[block_offset];
739                y_size = FFMIN(4, bi.image_height - bi.row * 4);
740                x_size = FFMIN(4, bi.image_width  - bi.col * 4);
741
742                for (int y = 0; y < y_size; y++) {
743                    for (int x = 0; x < x_size; x++) {
744                        rgb555 = row_ptr[x] & ~0x8000;
745
746                        put_bits(&s->pb, 16, rgb555);
747                    }
748                    for (int x = x_size; x < 4; x++)
749                        put_bits(&s->pb, 16, 0);
750                    row_ptr += bi.rowstride;
751                }
752
753                for (int y = y_size; y < 4; y++) {
754                    for (int x = 0; x < 4; x++)
755                        put_bits(&s->pb, 16, 0);
756                }
757
758                block_counter++;
759            } else { // FOUR COLOR BLOCK
760                block_counter += encode_four_color_block(min_color, max_color,
761                                                         &s->pb, &src_pixels[block_offset], &bi);
762            }
763
764            /* update this block in the previous frame buffer */
765            update_block_in_prev_frame(&src_pixels[block_offset],
766                                       &prev_pixels[block_offset], &bi, block_counter);
767        }
768    }
769}
770
771static int rpza_encode_init(AVCodecContext *avctx)
772{
773    RpzaContext *s = avctx->priv_data;
774
775    s->frame_width = avctx->width;
776    s->frame_height = avctx->height;
777
778    s->prev_frame = av_frame_alloc();
779    if (!s->prev_frame)
780        return AVERROR(ENOMEM);
781
782    return 0;
783}
784
785static int rpza_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
786                                const AVFrame *frame, int *got_packet)
787{
788    RpzaContext *s = avctx->priv_data;
789    const AVFrame *pict = frame;
790    uint8_t *buf;
791    int ret = ff_alloc_packet(avctx, pkt, 6LL * avctx->height * avctx->width);
792
793    if (ret < 0)
794        return ret;
795
796    init_put_bits(&s->pb, pkt->data, pkt->size);
797
798    // skip 4 byte header, write it later once the size of the chunk is known
799    put_bits32(&s->pb, 0x00);
800
801    if (!s->prev_frame->data[0]) {
802        s->first_frame = 1;
803        s->prev_frame->format = pict->format;
804        s->prev_frame->width = pict->width;
805        s->prev_frame->height = pict->height;
806        ret = av_frame_get_buffer(s->prev_frame, 0);
807        if (ret < 0)
808            return ret;
809    } else {
810        s->first_frame = 0;
811    }
812
813    rpza_encode_stream(s, pict);
814
815    flush_put_bits(&s->pb);
816
817    av_shrink_packet(pkt, put_bytes_output(&s->pb));
818    buf = pkt->data;
819
820    // write header opcode
821    buf[0] = 0xe1; // chunk opcode
822
823    // write chunk length
824    AV_WB24(buf + 1, pkt->size);
825
826    *got_packet = 1;
827
828    return 0;
829}
830
831static int rpza_encode_end(AVCodecContext *avctx)
832{
833    RpzaContext *s = (RpzaContext *)avctx->priv_data;
834
835    av_frame_free(&s->prev_frame);
836
837    return 0;
838}
839
840#define OFFSET(x) offsetof(RpzaContext, x)
841#define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
842static const AVOption options[] = {
843    { "skip_frame_thresh", NULL, OFFSET(skip_frame_thresh), AV_OPT_TYPE_INT, {.i64=1}, 0, 24, VE},
844    { "start_one_color_thresh", NULL, OFFSET(start_one_color_thresh), AV_OPT_TYPE_INT, {.i64=1}, 0, 24, VE},
845    { "continue_one_color_thresh", NULL, OFFSET(continue_one_color_thresh), AV_OPT_TYPE_INT, {.i64=0}, 0, 24, VE},
846    { "sixteen_color_thresh", NULL, OFFSET(sixteen_color_thresh), AV_OPT_TYPE_INT, {.i64=1}, 0, 24, VE},
847    { NULL },
848};
849
850static const AVClass rpza_class = {
851    .class_name = "rpza",
852    .item_name  = av_default_item_name,
853    .option     = options,
854    .version    = LIBAVUTIL_VERSION_INT,
855};
856
857const FFCodec ff_rpza_encoder = {
858    .p.name         = "rpza",
859    .p.long_name    = NULL_IF_CONFIG_SMALL("QuickTime video (RPZA)"),
860    .p.type         = AVMEDIA_TYPE_VIDEO,
861    .p.id           = AV_CODEC_ID_RPZA,
862    .priv_data_size = sizeof(RpzaContext),
863    .p.priv_class   = &rpza_class,
864    .init           = rpza_encode_init,
865    FF_CODEC_ENCODE_CB(rpza_encode_frame),
866    .close          = rpza_encode_end,
867    .caps_internal  = FF_CODEC_CAP_INIT_THREADSAFE,
868    .p.pix_fmts     = (const enum AVPixelFormat[]) { AV_PIX_FMT_RGB555,
869                                                     AV_PIX_FMT_NONE},
870};
871