| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /* | ||
| 2 | * Cinepak encoder (c) 2011 Tomas Härdin | ||
| 3 | * http://titan.codemill.se/~tomhar/cinepakenc.patch | ||
| 4 | * | ||
| 5 | * Fixes and improvements, vintage decoders compatibility | ||
| 6 | * (c) 2013, 2014 Rl, Aetey Global Technologies AB | ||
| 7 | * | ||
| 8 | * Permission is hereby granted, free of charge, to any person obtaining a | ||
| 9 | * copy of this software and associated documentation files (the "Software"), | ||
| 10 | * to deal in the Software without restriction, including without limitation | ||
| 11 | * the rights to use, copy, modify, merge, publish, distribute, sublicense, | ||
| 12 | * and/or sell copies of the Software, and to permit persons to whom the | ||
| 13 | * Software is furnished to do so, subject to the following conditions: | ||
| 14 | * | ||
| 15 | * The above copyright notice and this permission notice shall be included | ||
| 16 | * in all copies or substantial portions of the Software. | ||
| 17 | * | ||
| 18 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | ||
| 19 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | ||
| 20 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL | ||
| 21 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR | ||
| 22 | * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, | ||
| 23 | * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR | ||
| 24 | * OTHER DEALINGS IN THE SOFTWARE. | ||
| 25 | */ | ||
| 26 | |||
| 27 | /* | ||
| 28 | * TODO: | ||
| 29 | * - optimize: color space conversion (move conversion to libswscale), ... | ||
| 30 | * MAYBE: | ||
| 31 | * - "optimally" split the frame into several non-regular areas | ||
| 32 | * using a separate codebook pair for each area and approximating | ||
| 33 | * the area by several rectangular strips (generally not full width ones) | ||
| 34 | * (use quadtree splitting? a simple fixed-granularity grid?) | ||
| 35 | */ | ||
| 36 | |||
| 37 | #include <string.h> | ||
| 38 | |||
| 39 | #include "libavutil/avassert.h" | ||
| 40 | #include "libavutil/intreadwrite.h" | ||
| 41 | #include "libavutil/lfg.h" | ||
| 42 | #include "libavutil/mem.h" | ||
| 43 | #include "libavutil/opt.h" | ||
| 44 | |||
| 45 | #include "avcodec.h" | ||
| 46 | #include "codec_internal.h" | ||
| 47 | #include "elbg.h" | ||
| 48 | #include "encode.h" | ||
| 49 | |||
| 50 | #define CVID_HEADER_SIZE 10 | ||
| 51 | #define STRIP_HEADER_SIZE 12 | ||
| 52 | #define CHUNK_HEADER_SIZE 4 | ||
| 53 | |||
| 54 | #define MB_SIZE 4 //4x4 MBs | ||
| 55 | #define MB_AREA (MB_SIZE * MB_SIZE) | ||
| 56 | |||
| 57 | #define VECTOR_MAX 6 // six or four entries per vector depending on format | ||
| 58 | #define CODEBOOK_MAX 256 // size of a codebook | ||
| 59 | |||
| 60 | #define MAX_STRIPS 32 // Note: having fewer choices regarding the number of strips speeds up encoding (obviously) | ||
| 61 | #define MIN_STRIPS 1 // Note: having more strips speeds up encoding the frame (this is less obvious) | ||
| 62 | // MAX_STRIPS limits the maximum quality you can reach | ||
| 63 | // when you want high quality on high resolutions, | ||
| 64 | // MIN_STRIPS limits the minimum efficiently encodable bit rate | ||
| 65 | // on low resolutions | ||
| 66 | // the numbers are only used for brute force optimization for the first frame, | ||
| 67 | // for the following frames they are adaptively readjusted | ||
| 68 | // NOTE the decoder in ffmpeg has its own arbitrary limitation on the number | ||
| 69 | // of strips, currently 32 | ||
| 70 | |||
| 71 | typedef enum CinepakMode { | ||
| 72 | MODE_V1_ONLY = 0, | ||
| 73 | MODE_V1_V4, | ||
| 74 | MODE_MC, | ||
| 75 | |||
| 76 | MODE_COUNT, | ||
| 77 | } CinepakMode; | ||
| 78 | |||
| 79 | typedef enum mb_encoding { | ||
| 80 | ENC_V1, | ||
| 81 | ENC_V4, | ||
| 82 | ENC_SKIP, | ||
| 83 | |||
| 84 | ENC_UNCERTAIN | ||
| 85 | } mb_encoding; | ||
| 86 | |||
| 87 | typedef struct mb_info { | ||
| 88 | int v1_vector; // index into v1 codebook | ||
| 89 | int v1_error; // error when using V1 encoding | ||
| 90 | int v4_vector[4]; // indices into v4 codebook | ||
| 91 | int v4_error; // error when using V4 encoding | ||
| 92 | int skip_error; // error when block is skipped (aka copied from last frame) | ||
| 93 | mb_encoding best_encoding; // last result from calculate_mode_score() | ||
| 94 | } mb_info; | ||
| 95 | |||
| 96 | typedef struct strip_info { | ||
| 97 | int v1_codebook[CODEBOOK_MAX * VECTOR_MAX]; | ||
| 98 | int v4_codebook[CODEBOOK_MAX * VECTOR_MAX]; | ||
| 99 | int v1_size; | ||
| 100 | int v4_size; | ||
| 101 | CinepakMode mode; | ||
| 102 | } strip_info; | ||
| 103 | |||
| 104 | typedef struct CinepakEncContext { | ||
| 105 | const AVClass *class; | ||
| 106 | AVCodecContext *avctx; | ||
| 107 | unsigned char *pict_bufs[4], *strip_buf, *frame_buf; | ||
| 108 | AVFrame *last_frame; | ||
| 109 | AVFrame *best_frame; | ||
| 110 | AVFrame *scratch_frame; | ||
| 111 | AVFrame *input_frame; | ||
| 112 | enum AVPixelFormat pix_fmt; | ||
| 113 | int w, h; | ||
| 114 | int frame_buf_size; | ||
| 115 | int curframe; | ||
| 116 | AVLFG randctx; | ||
| 117 | uint64_t lambda; | ||
| 118 | int *codebook_input; | ||
| 119 | int *codebook_closest; | ||
| 120 | mb_info *mb; // MB RD state | ||
| 121 | int min_strips; // the current limit | ||
| 122 | int max_strips; // the current limit | ||
| 123 | // options | ||
| 124 | int max_extra_cb_iterations; | ||
| 125 | int skip_empty_cb; | ||
| 126 | int min_min_strips; | ||
| 127 | int max_max_strips; | ||
| 128 | int strip_number_delta_range; | ||
| 129 | struct ELBGContext *elbg; | ||
| 130 | } CinepakEncContext; | ||
| 131 | |||
| 132 | #define OFFSET(x) offsetof(CinepakEncContext, x) | ||
| 133 | #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM | ||
| 134 | static const AVOption options[] = { | ||
| 135 | { "max_extra_cb_iterations", "Max extra codebook recalculation passes, more is better and slower", | ||
| 136 | OFFSET(max_extra_cb_iterations), AV_OPT_TYPE_INT, { .i64 = 2 }, 0, INT_MAX, VE }, | ||
| 137 | { "skip_empty_cb", "Avoid wasting bytes, ignore vintage MacOS decoder", | ||
| 138 | OFFSET(skip_empty_cb), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, VE }, | ||
| 139 | { "max_strips", "Limit strips/frame, vintage compatible is 1..3, otherwise the more the better", | ||
| 140 | OFFSET(max_max_strips), AV_OPT_TYPE_INT, { .i64 = 3 }, MIN_STRIPS, MAX_STRIPS, VE }, | ||
| 141 | { "min_strips", "Enforce min strips/frame, more is worse and faster, must be <= max_strips", | ||
| 142 | OFFSET(min_min_strips), AV_OPT_TYPE_INT, { .i64 = MIN_STRIPS }, MIN_STRIPS, MAX_STRIPS, VE }, | ||
| 143 | { "strip_number_adaptivity", "How fast the strip number adapts, more is slightly better, much slower", | ||
| 144 | OFFSET(strip_number_delta_range), AV_OPT_TYPE_INT, { .i64 = 0 }, 0, MAX_STRIPS - MIN_STRIPS, VE }, | ||
| 145 | { NULL }, | ||
| 146 | }; | ||
| 147 | |||
| 148 | static const AVClass cinepak_class = { | ||
| 149 | .class_name = "cinepak", | ||
| 150 | .item_name = av_default_item_name, | ||
| 151 | .option = options, | ||
| 152 | .version = LIBAVUTIL_VERSION_INT, | ||
| 153 | }; | ||
| 154 | |||
| 155 | 3 | static av_cold int cinepak_encode_init(AVCodecContext *avctx) | |
| 156 | { | ||
| 157 | 3 | CinepakEncContext *s = avctx->priv_data; | |
| 158 | int x, mb_count, strip_buf_size, frame_buf_size; | ||
| 159 | |||
| 160 |
2/4✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 3 times.
|
3 | if (avctx->width & 3 || avctx->height & 3) { |
| 161 | ✗ | av_log(avctx, AV_LOG_ERROR, "width and height must be multiples of four (got %ix%i)\n", | |
| 162 | avctx->width, avctx->height); | ||
| 163 | ✗ | return AVERROR(EINVAL); | |
| 164 | } | ||
| 165 | |||
| 166 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 3 times.
|
3 | if (s->min_min_strips > s->max_max_strips) { |
| 167 | ✗ | av_log(avctx, AV_LOG_ERROR, "minimum number of strips must not exceed maximum (got %i and %i)\n", | |
| 168 | s->min_min_strips, s->max_max_strips); | ||
| 169 | ✗ | return AVERROR(EINVAL); | |
| 170 | } | ||
| 171 | |||
| 172 |
1/2✗ Branch 1 not taken.
✓ Branch 2 taken 3 times.
|
3 | if (!(s->last_frame = av_frame_alloc())) |
| 173 | ✗ | return AVERROR(ENOMEM); | |
| 174 |
1/2✗ Branch 1 not taken.
✓ Branch 2 taken 3 times.
|
3 | if (!(s->best_frame = av_frame_alloc())) |
| 175 | ✗ | return AVERROR(ENOMEM); | |
| 176 |
1/2✗ Branch 1 not taken.
✓ Branch 2 taken 3 times.
|
3 | if (!(s->scratch_frame = av_frame_alloc())) |
| 177 | ✗ | return AVERROR(ENOMEM); | |
| 178 |
1/2✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
|
3 | if (avctx->pix_fmt == AV_PIX_FMT_RGB24) |
| 179 |
1/2✗ Branch 1 not taken.
✓ Branch 2 taken 3 times.
|
3 | if (!(s->input_frame = av_frame_alloc())) |
| 180 | ✗ | return AVERROR(ENOMEM); | |
| 181 | |||
| 182 |
2/4✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
✗ Branch 3 not taken.
✓ Branch 4 taken 3 times.
|
3 | if (!(s->codebook_input = av_malloc_array((avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4) * (avctx->width * avctx->height) >> 2, sizeof(*s->codebook_input)))) |
| 183 | ✗ | return AVERROR(ENOMEM); | |
| 184 | |||
| 185 |
1/2✗ Branch 1 not taken.
✓ Branch 2 taken 3 times.
|
3 | if (!(s->codebook_closest = av_malloc_array((avctx->width * avctx->height) >> 2, sizeof(*s->codebook_closest)))) |
| 186 | ✗ | return AVERROR(ENOMEM); | |
| 187 | |||
| 188 |
3/4✓ Branch 0 taken 15 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 12 times.
✓ Branch 3 taken 3 times.
|
15 | for (x = 0; x < (avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 4 : 3); x++) |
| 189 |
2/4✓ Branch 0 taken 12 times.
✗ Branch 1 not taken.
✗ Branch 3 not taken.
✓ Branch 4 taken 12 times.
|
12 | if (!(s->pict_bufs[x] = av_malloc((avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4) * (avctx->width * avctx->height) >> 2))) |
| 190 | ✗ | return AVERROR(ENOMEM); | |
| 191 | |||
| 192 | 3 | mb_count = avctx->width * avctx->height / MB_AREA; | |
| 193 | |||
| 194 | // the largest possible chunk is 0x31 with all MBs encoded in V4 mode | ||
| 195 | // and full codebooks being replaced in INTER mode, | ||
| 196 | // which is 34 bits per MB | ||
| 197 | // and 2*256 extra flag bits per strip | ||
| 198 | 3 | strip_buf_size = STRIP_HEADER_SIZE + 3 * CHUNK_HEADER_SIZE + 2 * VECTOR_MAX * CODEBOOK_MAX + 4 * (mb_count + (mb_count + 15) / 16) + (2 * CODEBOOK_MAX) / 8; | |
| 199 | |||
| 200 | 3 | frame_buf_size = CVID_HEADER_SIZE + s->max_max_strips * strip_buf_size; | |
| 201 | |||
| 202 |
1/2✗ Branch 1 not taken.
✓ Branch 2 taken 3 times.
|
3 | if (!(s->strip_buf = av_malloc(strip_buf_size))) |
| 203 | ✗ | return AVERROR(ENOMEM); | |
| 204 | |||
| 205 |
1/2✗ Branch 1 not taken.
✓ Branch 2 taken 3 times.
|
3 | if (!(s->frame_buf = av_malloc(frame_buf_size))) |
| 206 | ✗ | return AVERROR(ENOMEM); | |
| 207 | |||
| 208 |
1/2✗ Branch 1 not taken.
✓ Branch 2 taken 3 times.
|
3 | if (!(s->mb = av_malloc_array(mb_count, sizeof(mb_info)))) |
| 209 | ✗ | return AVERROR(ENOMEM); | |
| 210 | |||
| 211 | 3 | av_lfg_init(&s->randctx, 1); | |
| 212 | 3 | s->avctx = avctx; | |
| 213 | 3 | s->w = avctx->width; | |
| 214 | 3 | s->h = avctx->height; | |
| 215 | 3 | s->frame_buf_size = frame_buf_size; | |
| 216 | 3 | s->curframe = 0; | |
| 217 | 3 | s->pix_fmt = avctx->pix_fmt; | |
| 218 | |||
| 219 | // set up AVFrames | ||
| 220 | 3 | s->last_frame->data[0] = s->pict_bufs[0]; | |
| 221 | 3 | s->last_frame->linesize[0] = s->w; | |
| 222 | 3 | s->best_frame->data[0] = s->pict_bufs[1]; | |
| 223 | 3 | s->best_frame->linesize[0] = s->w; | |
| 224 | 3 | s->scratch_frame->data[0] = s->pict_bufs[2]; | |
| 225 | 3 | s->scratch_frame->linesize[0] = s->w; | |
| 226 | |||
| 227 |
1/2✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
|
3 | if (s->pix_fmt == AV_PIX_FMT_RGB24) { |
| 228 | 3 | s->last_frame->data[1] = s->last_frame->data[0] + s->w * s->h; | |
| 229 | 3 | s->last_frame->data[2] = s->last_frame->data[1] + ((s->w * s->h) >> 2); | |
| 230 | 3 | s->last_frame->linesize[1] = | |
| 231 | 3 | s->last_frame->linesize[2] = s->w >> 1; | |
| 232 | |||
| 233 | 3 | s->best_frame->data[1] = s->best_frame->data[0] + s->w * s->h; | |
| 234 | 3 | s->best_frame->data[2] = s->best_frame->data[1] + ((s->w * s->h) >> 2); | |
| 235 | 3 | s->best_frame->linesize[1] = | |
| 236 | 3 | s->best_frame->linesize[2] = s->w >> 1; | |
| 237 | |||
| 238 | 3 | s->scratch_frame->data[1] = s->scratch_frame->data[0] + s->w * s->h; | |
| 239 | 3 | s->scratch_frame->data[2] = s->scratch_frame->data[1] + ((s->w * s->h) >> 2); | |
| 240 | 3 | s->scratch_frame->linesize[1] = | |
| 241 | 3 | s->scratch_frame->linesize[2] = s->w >> 1; | |
| 242 | |||
| 243 | 3 | s->input_frame->data[0] = s->pict_bufs[3]; | |
| 244 | 3 | s->input_frame->linesize[0] = s->w; | |
| 245 | 3 | s->input_frame->data[1] = s->input_frame->data[0] + s->w * s->h; | |
| 246 | 3 | s->input_frame->data[2] = s->input_frame->data[1] + ((s->w * s->h) >> 2); | |
| 247 | 3 | s->input_frame->linesize[1] = | |
| 248 | 3 | s->input_frame->linesize[2] = s->w >> 1; | |
| 249 | } | ||
| 250 | |||
| 251 | 3 | s->min_strips = s->min_min_strips; | |
| 252 | 3 | s->max_strips = s->max_max_strips; | |
| 253 | |||
| 254 | 3 | return 0; | |
| 255 | } | ||
| 256 | |||
| 257 | 47049 | static int64_t calculate_mode_score(CinepakEncContext *s, int h, | |
| 258 | strip_info *info, int report, | ||
| 259 | int *training_set_v1_shrunk, | ||
| 260 | int *training_set_v4_shrunk) | ||
| 261 | { | ||
| 262 | // score = FF_LAMBDA_SCALE * error + lambda * bits | ||
| 263 | int x; | ||
| 264 |
1/2✓ Branch 0 taken 47049 times.
✗ Branch 1 not taken.
|
47049 | int entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4; |
| 265 | 47049 | int mb_count = s->w * h / MB_AREA; | |
| 266 | mb_info *mb; | ||
| 267 | int64_t score1, score2, score3; | ||
| 268 |
2/2✓ Branch 0 taken 47019 times.
✓ Branch 1 taken 30 times.
|
47049 | int64_t ret = s->lambda * ((info->v1_size ? CHUNK_HEADER_SIZE + info->v1_size * entry_size : 0) + |
| 269 |
2/2✓ Branch 0 taken 36472 times.
✓ Branch 1 taken 10577 times.
|
47049 | (info->v4_size ? CHUNK_HEADER_SIZE + info->v4_size * entry_size : 0) + |
| 270 | 47049 | CHUNK_HEADER_SIZE) << 3; | |
| 271 | |||
| 272 |
3/4✓ Branch 0 taken 10542 times.
✓ Branch 1 taken 13554 times.
✓ Branch 2 taken 22953 times.
✗ Branch 3 not taken.
|
47049 | switch (info->mode) { |
| 273 | 10542 | case MODE_V1_ONLY: | |
| 274 | // one byte per MB | ||
| 275 | 10542 | ret += s->lambda * 8 * mb_count; | |
| 276 | |||
| 277 | // while calculating we assume all blocks are ENC_V1 | ||
| 278 |
2/2✓ Branch 0 taken 3257856 times.
✓ Branch 1 taken 10542 times.
|
3268398 | for (x = 0; x < mb_count; x++) { |
| 279 | 3257856 | mb = &s->mb[x]; | |
| 280 | 3257856 | ret += FF_LAMBDA_SCALE * mb->v1_error; | |
| 281 | // this function is never called for report in MODE_V1_ONLY | ||
| 282 | // if (!report) | ||
| 283 | 3257856 | mb->best_encoding = ENC_V1; | |
| 284 | } | ||
| 285 | |||
| 286 | 10542 | break; | |
| 287 | 13554 | case MODE_V1_V4: | |
| 288 | // 9 or 33 bits per MB | ||
| 289 |
2/2✓ Branch 0 taken 6777 times.
✓ Branch 1 taken 6777 times.
|
13554 | if (report) { |
| 290 | // no moves between the corresponding training sets are allowed | ||
| 291 | 6777 | *training_set_v1_shrunk = *training_set_v4_shrunk = 0; | |
| 292 |
2/2✓ Branch 0 taken 2094336 times.
✓ Branch 1 taken 6777 times.
|
2101113 | for (x = 0; x < mb_count; x++) { |
| 293 | int mberr; | ||
| 294 | 2094336 | mb = &s->mb[x]; | |
| 295 |
2/2✓ Branch 0 taken 666470 times.
✓ Branch 1 taken 1427866 times.
|
2094336 | if (mb->best_encoding == ENC_V1) |
| 296 | 666470 | score1 = s->lambda * 9 + FF_LAMBDA_SCALE * (mberr = mb->v1_error); | |
| 297 | else | ||
| 298 | 1427866 | score1 = s->lambda * 33 + FF_LAMBDA_SCALE * (mberr = mb->v4_error); | |
| 299 | 2094336 | ret += score1; | |
| 300 | } | ||
| 301 | } else { // find best mode per block | ||
| 302 |
2/2✓ Branch 0 taken 2094336 times.
✓ Branch 1 taken 6777 times.
|
2101113 | for (x = 0; x < mb_count; x++) { |
| 303 | 2094336 | mb = &s->mb[x]; | |
| 304 | 2094336 | score1 = s->lambda * 9 + FF_LAMBDA_SCALE * mb->v1_error; | |
| 305 | 2094336 | score2 = s->lambda * 33 + FF_LAMBDA_SCALE * mb->v4_error; | |
| 306 | |||
| 307 |
2/2✓ Branch 0 taken 666470 times.
✓ Branch 1 taken 1427866 times.
|
2094336 | if (score1 <= score2) { |
| 308 | 666470 | ret += score1; | |
| 309 | 666470 | mb->best_encoding = ENC_V1; | |
| 310 | } else { | ||
| 311 | 1427866 | ret += score2; | |
| 312 | 1427866 | mb->best_encoding = ENC_V4; | |
| 313 | } | ||
| 314 | } | ||
| 315 | } | ||
| 316 | |||
| 317 | 13554 | break; | |
| 318 | 22953 | case MODE_MC: | |
| 319 | // 1, 10 or 34 bits per MB | ||
| 320 |
2/2✓ Branch 0 taken 16788 times.
✓ Branch 1 taken 6165 times.
|
22953 | if (report) { |
| 321 | 16788 | int v1_shrunk = 0, v4_shrunk = 0; | |
| 322 |
2/2✓ Branch 0 taken 5184512 times.
✓ Branch 1 taken 16788 times.
|
5201300 | for (x = 0; x < mb_count; x++) { |
| 323 | 5184512 | mb = &s->mb[x]; | |
| 324 | // it is OK to move blocks to ENC_SKIP here | ||
| 325 | // but not to any codebook encoding! | ||
| 326 | 5184512 | score1 = s->lambda * 1 + FF_LAMBDA_SCALE * mb->skip_error; | |
| 327 |
2/2✓ Branch 0 taken 2396023 times.
✓ Branch 1 taken 2788489 times.
|
5184512 | if (mb->best_encoding == ENC_SKIP) { |
| 328 | 2396023 | ret += score1; | |
| 329 |
2/2✓ Branch 0 taken 786928 times.
✓ Branch 1 taken 2001561 times.
|
2788489 | } else if (mb->best_encoding == ENC_V1) { |
| 330 |
2/2✓ Branch 0 taken 23055 times.
✓ Branch 1 taken 763873 times.
|
786928 | if ((score2 = s->lambda * 10 + FF_LAMBDA_SCALE * mb->v1_error) >= score1) { |
| 331 | 23055 | mb->best_encoding = ENC_SKIP; | |
| 332 | 23055 | ++v1_shrunk; | |
| 333 | 23055 | ret += score1; | |
| 334 | } else { | ||
| 335 | 763873 | ret += score2; | |
| 336 | } | ||
| 337 | } else { | ||
| 338 |
2/2✓ Branch 0 taken 71655 times.
✓ Branch 1 taken 1929906 times.
|
2001561 | if ((score3 = s->lambda * 34 + FF_LAMBDA_SCALE * mb->v4_error) >= score1) { |
| 339 | 71655 | mb->best_encoding = ENC_SKIP; | |
| 340 | 71655 | ++v4_shrunk; | |
| 341 | 71655 | ret += score1; | |
| 342 | } else { | ||
| 343 | 1929906 | ret += score3; | |
| 344 | } | ||
| 345 | } | ||
| 346 | } | ||
| 347 | 16788 | *training_set_v1_shrunk = v1_shrunk; | |
| 348 | 16788 | *training_set_v4_shrunk = v4_shrunk; | |
| 349 | } else { // find best mode per block | ||
| 350 |
2/2✓ Branch 0 taken 1893888 times.
✓ Branch 1 taken 6165 times.
|
1900053 | for (x = 0; x < mb_count; x++) { |
| 351 | 1893888 | mb = &s->mb[x]; | |
| 352 | 1893888 | score1 = s->lambda * 1 + FF_LAMBDA_SCALE * mb->skip_error; | |
| 353 | 1893888 | score2 = s->lambda * 10 + FF_LAMBDA_SCALE * mb->v1_error; | |
| 354 | 1893888 | score3 = s->lambda * 34 + FF_LAMBDA_SCALE * mb->v4_error; | |
| 355 | |||
| 356 |
4/4✓ Branch 0 taken 1039787 times.
✓ Branch 1 taken 854101 times.
✓ Branch 2 taken 803498 times.
✓ Branch 3 taken 236289 times.
|
1893888 | if (score1 <= score2 && score1 <= score3) { |
| 357 | 803498 | ret += score1; | |
| 358 | 803498 | mb->best_encoding = ENC_SKIP; | |
| 359 |
2/2✓ Branch 0 taken 298201 times.
✓ Branch 1 taken 792189 times.
|
1090390 | } else if (score2 <= score3) { |
| 360 | 298201 | ret += score2; | |
| 361 | 298201 | mb->best_encoding = ENC_V1; | |
| 362 | } else { | ||
| 363 | 792189 | ret += score3; | |
| 364 | 792189 | mb->best_encoding = ENC_V4; | |
| 365 | } | ||
| 366 | } | ||
| 367 | } | ||
| 368 | |||
| 369 | 22953 | break; | |
| 370 | } | ||
| 371 | |||
| 372 | 47049 | return ret; | |
| 373 | } | ||
| 374 | |||
| 375 | 30375 | static int write_chunk_header(unsigned char *buf, int chunk_type, int chunk_size) | |
| 376 | { | ||
| 377 | 30375 | buf[0] = chunk_type; | |
| 378 | 30375 | AV_WB24(&buf[1], chunk_size + CHUNK_HEADER_SIZE); | |
| 379 | 30375 | return CHUNK_HEADER_SIZE; | |
| 380 | } | ||
| 381 | |||
| 382 | 20250 | static int encode_codebook(CinepakEncContext *s, int *codebook, int size, | |
| 383 | int chunk_type_yuv, int chunk_type_gray, | ||
| 384 | unsigned char *buf) | ||
| 385 | { | ||
| 386 |
1/2✓ Branch 0 taken 20250 times.
✗ Branch 1 not taken.
|
20250 | int x, y, ret, entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4; |
| 387 | 20250 | int incremental_codebook_replacement_mode = 0; // hardcoded here, | |
| 388 | // the compiler should notice that this is a constant -- rl | ||
| 389 | |||
| 390 | 20250 | ret = write_chunk_header(buf, | |
| 391 |
1/2✓ Branch 0 taken 20250 times.
✗ Branch 1 not taken.
|
20250 | s->pix_fmt == AV_PIX_FMT_RGB24 ? |
| 392 | 20250 | chunk_type_yuv + (incremental_codebook_replacement_mode ? 1 : 0) : | |
| 393 | ✗ | chunk_type_gray + (incremental_codebook_replacement_mode ? 1 : 0), | |
| 394 | 20250 | entry_size * size + | |
| 395 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 20250 times.
|
20250 | (incremental_codebook_replacement_mode ? (size + 31) / 32 * 4 : 0)); |
| 396 | |||
| 397 | // we do codebook encoding according to the "intra" mode | ||
| 398 | // but we keep the "dead" code for reference in case we will want | ||
| 399 | // to use incremental codebook updates (which actually would give us | ||
| 400 | // "kind of" motion compensation, especially in 1 strip/frame case) -- rl | ||
| 401 | // (of course, the code will be not useful as-is) | ||
| 402 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 20250 times.
|
20250 | if (incremental_codebook_replacement_mode) { |
| 403 | ✗ | int flags = 0; | |
| 404 | int flagsind; | ||
| 405 | ✗ | for (x = 0; x < size; x++) { | |
| 406 | ✗ | if (flags == 0) { | |
| 407 | ✗ | flagsind = ret; | |
| 408 | ✗ | ret += 4; | |
| 409 | ✗ | flags = 0x80000000; | |
| 410 | } else | ||
| 411 | ✗ | flags = ((flags >> 1) | 0x80000000); | |
| 412 | ✗ | for (y = 0; y < entry_size; y++) | |
| 413 | ✗ | buf[ret++] = codebook[y + x * entry_size] ^ (y >= 4 ? 0x80 : 0); | |
| 414 | ✗ | if ((flags & 0xffffffff) == 0xffffffff) { | |
| 415 | ✗ | AV_WB32(&buf[flagsind], flags); | |
| 416 | ✗ | flags = 0; | |
| 417 | } | ||
| 418 | } | ||
| 419 | ✗ | if (flags) | |
| 420 | ✗ | AV_WB32(&buf[flagsind], flags); | |
| 421 | } else | ||
| 422 |
2/2✓ Branch 0 taken 719124 times.
✓ Branch 1 taken 20250 times.
|
739374 | for (x = 0; x < size; x++) |
| 423 |
2/2✓ Branch 0 taken 4314744 times.
✓ Branch 1 taken 719124 times.
|
5033868 | for (y = 0; y < entry_size; y++) |
| 424 |
2/2✓ Branch 0 taken 1438248 times.
✓ Branch 1 taken 2876496 times.
|
4314744 | buf[ret++] = codebook[y + x * entry_size] ^ (y >= 4 ? 0x80 : 0); |
| 425 | |||
| 426 | 20250 | return ret; | |
| 427 | } | ||
| 428 | |||
| 429 | // sets out to the sub picture starting at (x,y) in in | ||
| 430 | 14757134 | static void get_sub_picture(CinepakEncContext *s, int x, int y, | |
| 431 | uint8_t *const in_data[4], const int in_linesize[4], | ||
| 432 | uint8_t *out_data[4], int out_linesize[4]) | ||
| 433 | { | ||
| 434 | 14757134 | out_data[0] = in_data[0] + x + y * in_linesize[0]; | |
| 435 | 14757134 | out_linesize[0] = in_linesize[0]; | |
| 436 | |||
| 437 |
1/2✓ Branch 0 taken 14757134 times.
✗ Branch 1 not taken.
|
14757134 | if (s->pix_fmt == AV_PIX_FMT_RGB24) { |
| 438 | 14757134 | out_data[1] = in_data[1] + (x >> 1) + (y >> 1) * in_linesize[1]; | |
| 439 | 14757134 | out_linesize[1] = in_linesize[1]; | |
| 440 | |||
| 441 | 14757134 | out_data[2] = in_data[2] + (x >> 1) + (y >> 1) * in_linesize[2]; | |
| 442 | 14757134 | out_linesize[2] = in_linesize[2]; | |
| 443 | } | ||
| 444 | 14757134 | } | |
| 445 | |||
| 446 | // decodes the V1 vector in mb into the 4x4 MB pointed to by data | ||
| 447 | 5305363 | static void decode_v1_vector(CinepakEncContext *s, uint8_t *data[4], | |
| 448 | int linesize[4], int v1_vector, strip_info *info) | ||
| 449 | { | ||
| 450 |
1/2✓ Branch 0 taken 5305363 times.
✗ Branch 1 not taken.
|
5305363 | int entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4; |
| 451 | |||
| 452 | 5305363 | data[0][0] = | |
| 453 | 5305363 | data[0][1] = | |
| 454 | 5305363 | data[0][ linesize[0]] = | |
| 455 | 5305363 | data[0][1 + linesize[0]] = info->v1_codebook[v1_vector * entry_size]; | |
| 456 | |||
| 457 | 5305363 | data[0][2] = | |
| 458 | 5305363 | data[0][3] = | |
| 459 | 5305363 | data[0][2 + linesize[0]] = | |
| 460 | 5305363 | data[0][3 + linesize[0]] = info->v1_codebook[v1_vector * entry_size + 1]; | |
| 461 | |||
| 462 | 5305363 | data[0][ 2 * linesize[0]] = | |
| 463 | 5305363 | data[0][1 + 2 * linesize[0]] = | |
| 464 | 5305363 | data[0][ 3 * linesize[0]] = | |
| 465 | 5305363 | data[0][1 + 3 * linesize[0]] = info->v1_codebook[v1_vector * entry_size + 2]; | |
| 466 | |||
| 467 | 5305363 | data[0][2 + 2 * linesize[0]] = | |
| 468 | 5305363 | data[0][3 + 2 * linesize[0]] = | |
| 469 | 5305363 | data[0][2 + 3 * linesize[0]] = | |
| 470 | 5305363 | data[0][3 + 3 * linesize[0]] = info->v1_codebook[v1_vector * entry_size + 3]; | |
| 471 | |||
| 472 |
1/2✓ Branch 0 taken 5305363 times.
✗ Branch 1 not taken.
|
5305363 | if (s->pix_fmt == AV_PIX_FMT_RGB24) { |
| 473 | 5305363 | data[1][0] = | |
| 474 | 5305363 | data[1][1] = | |
| 475 | 5305363 | data[1][ linesize[1]] = | |
| 476 | 5305363 | data[1][1 + linesize[1]] = info->v1_codebook[v1_vector * entry_size + 4]; | |
| 477 | |||
| 478 | 5305363 | data[2][0] = | |
| 479 | 5305363 | data[2][1] = | |
| 480 | 5305363 | data[2][ linesize[2]] = | |
| 481 | 5305363 | data[2][1 + linesize[2]] = info->v1_codebook[v1_vector * entry_size + 5]; | |
| 482 | } | ||
| 483 | 5305363 | } | |
| 484 | |||
| 485 | // decodes the V4 vectors in mb into the 4x4 MB pointed to by data | ||
| 486 | 6880234 | static void decode_v4_vector(CinepakEncContext *s, uint8_t *data[4], | |
| 487 | int linesize[4], int *v4_vector, strip_info *info) | ||
| 488 | { | ||
| 489 |
1/2✓ Branch 0 taken 6880234 times.
✗ Branch 1 not taken.
|
6880234 | int i, x, y, entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4; |
| 490 | |||
| 491 |
2/2✓ Branch 0 taken 13760468 times.
✓ Branch 1 taken 6880234 times.
|
20640702 | for (i = y = 0; y < 4; y += 2) { |
| 492 |
2/2✓ Branch 0 taken 27520936 times.
✓ Branch 1 taken 13760468 times.
|
41281404 | for (x = 0; x < 4; x += 2, i++) { |
| 493 | 27520936 | data[0][x + y * linesize[0]] = info->v4_codebook[v4_vector[i] * entry_size]; | |
| 494 | 27520936 | data[0][x + 1 + y * linesize[0]] = info->v4_codebook[v4_vector[i] * entry_size + 1]; | |
| 495 | 27520936 | data[0][x + (y + 1) * linesize[0]] = info->v4_codebook[v4_vector[i] * entry_size + 2]; | |
| 496 | 27520936 | data[0][x + 1 + (y + 1) * linesize[0]] = info->v4_codebook[v4_vector[i] * entry_size + 3]; | |
| 497 | |||
| 498 |
1/2✓ Branch 0 taken 27520936 times.
✗ Branch 1 not taken.
|
27520936 | if (s->pix_fmt == AV_PIX_FMT_RGB24) { |
| 499 | 27520936 | data[1][(x >> 1) + (y >> 1) * linesize[1]] = info->v4_codebook[v4_vector[i] * entry_size + 4]; | |
| 500 | 27520936 | data[2][(x >> 1) + (y >> 1) * linesize[2]] = info->v4_codebook[v4_vector[i] * entry_size + 5]; | |
| 501 | } | ||
| 502 | } | ||
| 503 | } | ||
| 504 | 6880234 | } | |
| 505 | |||
| 506 | 843807 | static void copy_mb(CinepakEncContext *s, | |
| 507 | uint8_t *a_data[4], int a_linesize[4], | ||
| 508 | uint8_t *b_data[4], int b_linesize[4]) | ||
| 509 | { | ||
| 510 | int y, p; | ||
| 511 | |||
| 512 |
2/2✓ Branch 0 taken 3375228 times.
✓ Branch 1 taken 843807 times.
|
4219035 | for (y = 0; y < MB_SIZE; y++) |
| 513 | 3375228 | memcpy(a_data[0] + y * a_linesize[0], b_data[0] + y * b_linesize[0], | |
| 514 | MB_SIZE); | ||
| 515 | |||
| 516 |
1/2✓ Branch 0 taken 843807 times.
✗ Branch 1 not taken.
|
843807 | if (s->pix_fmt == AV_PIX_FMT_RGB24) { |
| 517 |
2/2✓ Branch 0 taken 1687614 times.
✓ Branch 1 taken 843807 times.
|
2531421 | for (p = 1; p <= 2; p++) |
| 518 |
2/2✓ Branch 0 taken 3375228 times.
✓ Branch 1 taken 1687614 times.
|
5062842 | for (y = 0; y < MB_SIZE / 2; y++) |
| 519 | 3375228 | memcpy(a_data[p] + y * a_linesize[p], | |
| 520 | 3375228 | b_data[p] + y * b_linesize[p], | |
| 521 | MB_SIZE / 2); | ||
| 522 | } | ||
| 523 | 843807 | } | |
| 524 | |||
| 525 | 10125 | static int encode_mode(CinepakEncContext *s, int h, | |
| 526 | uint8_t *scratch_data[4], int scratch_linesize[4], | ||
| 527 | uint8_t *last_data[4], int last_linesize[4], | ||
| 528 | strip_info *info, unsigned char *buf) | ||
| 529 | { | ||
| 530 | 10125 | int x, y, z, bits, temp_size, header_ofs, ret = 0, mb_count = s->w * h / MB_AREA; | |
| 531 | int needs_extra_bit, should_write_temp; | ||
| 532 | uint32_t flags; | ||
| 533 | unsigned char temp[64]; // 32/2 = 16 V4 blocks at 4 B each -> 64 B | ||
| 534 | mb_info *mb; | ||
| 535 | 10125 | uint8_t *sub_scratch_data[4] = { 0 }, *sub_last_data[4] = { 0 }; | |
| 536 | 10125 | int sub_scratch_linesize[4] = { 0 }, sub_last_linesize[4] = { 0 }; | |
| 537 | |||
| 538 | // encode codebooks | ||
| 539 | ////// MacOS vintage decoder compatibility dictates the presence of | ||
| 540 | ////// the codebook chunk even when the codebook is empty - pretty dumb... | ||
| 541 | ////// and also the certain order of the codebook chunks -- rl | ||
| 542 |
3/4✓ Branch 0 taken 1061 times.
✓ Branch 1 taken 9064 times.
✓ Branch 2 taken 1061 times.
✗ Branch 3 not taken.
|
10125 | if (info->v4_size || !s->skip_empty_cb) |
| 543 | 10125 | ret += encode_codebook(s, info->v4_codebook, info->v4_size, 0x20, 0x24, buf + ret); | |
| 544 | |||
| 545 |
3/4✓ Branch 0 taken 13 times.
✓ Branch 1 taken 10112 times.
✓ Branch 2 taken 13 times.
✗ Branch 3 not taken.
|
10125 | if (info->v1_size || !s->skip_empty_cb) |
| 546 | 10125 | ret += encode_codebook(s, info->v1_codebook, info->v1_size, 0x22, 0x26, buf + ret); | |
| 547 | |||
| 548 | // update scratch picture | ||
| 549 |
2/2✓ Branch 0 taken 97896 times.
✓ Branch 1 taken 10125 times.
|
108021 | for (z = y = 0; y < h; y += MB_SIZE) |
| 550 |
2/2✓ Branch 0 taken 3132672 times.
✓ Branch 1 taken 97896 times.
|
3230568 | for (x = 0; x < s->w; x += MB_SIZE, z++) { |
| 551 | 3132672 | mb = &s->mb[z]; | |
| 552 | |||
| 553 | 3132672 | get_sub_picture(s, x, y, scratch_data, scratch_linesize, | |
| 554 | sub_scratch_data, sub_scratch_linesize); | ||
| 555 | |||
| 556 |
4/4✓ Branch 0 taken 1630592 times.
✓ Branch 1 taken 1502080 times.
✓ Branch 2 taken 843807 times.
✓ Branch 3 taken 786785 times.
|
3132672 | if (info->mode == MODE_MC && mb->best_encoding == ENC_SKIP) { |
| 557 | 843807 | get_sub_picture(s, x, y, last_data, last_linesize, | |
| 558 | sub_last_data, sub_last_linesize); | ||
| 559 | 843807 | copy_mb(s, sub_scratch_data, sub_scratch_linesize, | |
| 560 | sub_last_data, sub_last_linesize); | ||
| 561 |
4/4✓ Branch 0 taken 1964385 times.
✓ Branch 1 taken 324480 times.
✓ Branch 2 taken 513483 times.
✓ Branch 3 taken 1450902 times.
|
2288865 | } else if (info->mode == MODE_V1_ONLY || mb->best_encoding == ENC_V1) |
| 562 | 837963 | decode_v1_vector(s, sub_scratch_data, sub_scratch_linesize, | |
| 563 | mb->v1_vector, info); | ||
| 564 | else | ||
| 565 | 1450902 | decode_v4_vector(s, sub_scratch_data, sub_scratch_linesize, | |
| 566 | 1450902 | mb->v4_vector, info); | |
| 567 | } | ||
| 568 | |||
| 569 |
3/4✓ Branch 0 taken 1044 times.
✓ Branch 1 taken 3784 times.
✓ Branch 2 taken 5297 times.
✗ Branch 3 not taken.
|
10125 | switch (info->mode) { |
| 570 | 1044 | case MODE_V1_ONLY: | |
| 571 | 1044 | ret += write_chunk_header(buf + ret, 0x32, mb_count); | |
| 572 | |||
| 573 |
2/2✓ Branch 0 taken 324480 times.
✓ Branch 1 taken 1044 times.
|
325524 | for (x = 0; x < mb_count; x++) |
| 574 | 324480 | buf[ret++] = s->mb[x].v1_vector; | |
| 575 | |||
| 576 | 1044 | break; | |
| 577 | 3784 | case MODE_V1_V4: | |
| 578 | // remember header position | ||
| 579 | 3784 | header_ofs = ret; | |
| 580 | 3784 | ret += CHUNK_HEADER_SIZE; | |
| 581 | |||
| 582 |
2/2✓ Branch 0 taken 36800 times.
✓ Branch 1 taken 3784 times.
|
40584 | for (x = 0; x < mb_count; x += 32) { |
| 583 | 36800 | flags = 0; | |
| 584 |
3/4✓ Branch 0 taken 1214400 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 1177600 times.
✓ Branch 3 taken 36800 times.
|
1214400 | for (y = x; y < FFMIN(x + 32, mb_count); y++) |
| 585 |
2/2✓ Branch 0 taken 883984 times.
✓ Branch 1 taken 293616 times.
|
1177600 | if (s->mb[y].best_encoding == ENC_V4) |
| 586 | 883984 | flags |= 1U << (31 - y + x); | |
| 587 | |||
| 588 | 36800 | AV_WB32(&buf[ret], flags); | |
| 589 | 36800 | ret += 4; | |
| 590 | |||
| 591 |
3/4✓ Branch 0 taken 1214400 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 1177600 times.
✓ Branch 3 taken 36800 times.
|
1214400 | for (y = x; y < FFMIN(x + 32, mb_count); y++) { |
| 592 | 1177600 | mb = &s->mb[y]; | |
| 593 | |||
| 594 |
2/2✓ Branch 0 taken 293616 times.
✓ Branch 1 taken 883984 times.
|
1177600 | if (mb->best_encoding == ENC_V1) |
| 595 | 293616 | buf[ret++] = mb->v1_vector; | |
| 596 | else | ||
| 597 |
2/2✓ Branch 0 taken 3535936 times.
✓ Branch 1 taken 883984 times.
|
4419920 | for (z = 0; z < 4; z++) |
| 598 | 3535936 | buf[ret++] = mb->v4_vector[z]; | |
| 599 | } | ||
| 600 | } | ||
| 601 | |||
| 602 | 3784 | write_chunk_header(buf + header_ofs, 0x30, ret - header_ofs - CHUNK_HEADER_SIZE); | |
| 603 | |||
| 604 | 3784 | break; | |
| 605 | 5297 | case MODE_MC: | |
| 606 | // remember header position | ||
| 607 | 5297 | header_ofs = ret; | |
| 608 | 5297 | ret += CHUNK_HEADER_SIZE; | |
| 609 | 5297 | flags = bits = temp_size = 0; | |
| 610 | |||
| 611 |
2/2✓ Branch 0 taken 1630592 times.
✓ Branch 1 taken 5297 times.
|
1635889 | for (x = 0; x < mb_count; x++) { |
| 612 | 1630592 | mb = &s->mb[x]; | |
| 613 | 1630592 | flags |= (uint32_t)(mb->best_encoding != ENC_SKIP) << (31 - bits++); | |
| 614 | 1630592 | needs_extra_bit = 0; | |
| 615 | 1630592 | should_write_temp = 0; | |
| 616 | |||
| 617 |
2/2✓ Branch 0 taken 786785 times.
✓ Branch 1 taken 843807 times.
|
1630592 | if (mb->best_encoding != ENC_SKIP) { |
| 618 |
2/2✓ Branch 0 taken 764622 times.
✓ Branch 1 taken 22163 times.
|
786785 | if (bits < 32) |
| 619 | 764622 | flags |= (uint32_t)(mb->best_encoding == ENC_V4) << (31 - bits++); | |
| 620 | else | ||
| 621 | 22163 | needs_extra_bit = 1; | |
| 622 | } | ||
| 623 | |||
| 624 |
2/2✓ Branch 0 taken 72909 times.
✓ Branch 1 taken 1557683 times.
|
1630592 | if (bits == 32) { |
| 625 | 72909 | AV_WB32(&buf[ret], flags); | |
| 626 | 72909 | ret += 4; | |
| 627 | 72909 | flags = bits = 0; | |
| 628 | |||
| 629 |
4/4✓ Branch 0 taken 47192 times.
✓ Branch 1 taken 25717 times.
✓ Branch 2 taken 22163 times.
✓ Branch 3 taken 25029 times.
|
72909 | if (mb->best_encoding == ENC_SKIP || needs_extra_bit) { |
| 630 | 47880 | memcpy(&buf[ret], temp, temp_size); | |
| 631 | 47880 | ret += temp_size; | |
| 632 | 47880 | temp_size = 0; | |
| 633 | } else | ||
| 634 | 25029 | should_write_temp = 1; | |
| 635 | } | ||
| 636 | |||
| 637 |
2/2✓ Branch 0 taken 22163 times.
✓ Branch 1 taken 1608429 times.
|
1630592 | if (needs_extra_bit) { |
| 638 |
2/2✓ Branch 0 taken 15892 times.
✓ Branch 1 taken 6271 times.
|
22163 | flags = (uint32_t)(mb->best_encoding == ENC_V4) << 31; |
| 639 | 22163 | bits = 1; | |
| 640 | } | ||
| 641 | |||
| 642 |
2/2✓ Branch 0 taken 219867 times.
✓ Branch 1 taken 1410725 times.
|
1630592 | if (mb->best_encoding == ENC_V1) |
| 643 | 219867 | temp[temp_size++] = mb->v1_vector; | |
| 644 |
2/2✓ Branch 0 taken 566918 times.
✓ Branch 1 taken 843807 times.
|
1410725 | else if (mb->best_encoding == ENC_V4) |
| 645 |
2/2✓ Branch 0 taken 2267672 times.
✓ Branch 1 taken 566918 times.
|
2834590 | for (z = 0; z < 4; z++) |
| 646 | 2267672 | temp[temp_size++] = mb->v4_vector[z]; | |
| 647 | |||
| 648 |
2/2✓ Branch 0 taken 25029 times.
✓ Branch 1 taken 1605563 times.
|
1630592 | if (should_write_temp) { |
| 649 | 25029 | memcpy(&buf[ret], temp, temp_size); | |
| 650 | 25029 | ret += temp_size; | |
| 651 | 25029 | temp_size = 0; | |
| 652 | } | ||
| 653 | } | ||
| 654 | |||
| 655 |
2/2✓ Branch 0 taken 5127 times.
✓ Branch 1 taken 170 times.
|
5297 | if (bits > 0) { |
| 656 | 5127 | AV_WB32(&buf[ret], flags); | |
| 657 | 5127 | ret += 4; | |
| 658 | 5127 | memcpy(&buf[ret], temp, temp_size); | |
| 659 | 5127 | ret += temp_size; | |
| 660 | } | ||
| 661 | |||
| 662 | 5297 | write_chunk_header(buf + header_ofs, 0x31, ret - header_ofs - CHUNK_HEADER_SIZE); | |
| 663 | |||
| 664 | 5297 | break; | |
| 665 | } | ||
| 666 | |||
| 667 | 10125 | return ret; | |
| 668 | } | ||
| 669 | |||
| 670 | // computes distortion of 4x4 MB in b compared to a | ||
| 671 | 10107164 | static int compute_mb_distortion(CinepakEncContext *s, | |
| 672 | uint8_t *a_data[4], int a_linesize[4], | ||
| 673 | uint8_t *b_data[4], int b_linesize[4]) | ||
| 674 | { | ||
| 675 | 10107164 | int x, y, p, d, ret = 0; | |
| 676 | |||
| 677 |
2/2✓ Branch 0 taken 40428656 times.
✓ Branch 1 taken 10107164 times.
|
50535820 | for (y = 0; y < MB_SIZE; y++) |
| 678 |
2/2✓ Branch 0 taken 161714624 times.
✓ Branch 1 taken 40428656 times.
|
202143280 | for (x = 0; x < MB_SIZE; x++) { |
| 679 | 161714624 | d = a_data[0][x + y * a_linesize[0]] - b_data[0][x + y * b_linesize[0]]; | |
| 680 | 161714624 | ret += d * d; | |
| 681 | } | ||
| 682 | |||
| 683 |
1/2✓ Branch 0 taken 10107164 times.
✗ Branch 1 not taken.
|
10107164 | if (s->pix_fmt == AV_PIX_FMT_RGB24) { |
| 684 |
2/2✓ Branch 0 taken 20214328 times.
✓ Branch 1 taken 10107164 times.
|
30321492 | for (p = 1; p <= 2; p++) { |
| 685 |
2/2✓ Branch 0 taken 40428656 times.
✓ Branch 1 taken 20214328 times.
|
60642984 | for (y = 0; y < MB_SIZE / 2; y++) |
| 686 |
2/2✓ Branch 0 taken 80857312 times.
✓ Branch 1 taken 40428656 times.
|
121285968 | for (x = 0; x < MB_SIZE / 2; x++) { |
| 687 | 80857312 | d = a_data[p][x + y * a_linesize[p]] - b_data[p][x + y * b_linesize[p]]; | |
| 688 | 80857312 | ret += d * d; | |
| 689 | } | ||
| 690 | } | ||
| 691 | } | ||
| 692 | |||
| 693 | 10107164 | return ret; | |
| 694 | } | ||
| 695 | |||
| 696 | // return the possibly adjusted size of the codebook | ||
| 697 | #define CERTAIN(x) ((x) != ENC_UNCERTAIN) | ||
| 698 | 57111 | static int quantize(CinepakEncContext *s, int h, uint8_t *data[4], | |
| 699 | int linesize[4], int v1mode, strip_info *info, | ||
| 700 | mb_encoding encoding) | ||
| 701 | { | ||
| 702 | int x, y, i, j, k, x2, y2, x3, y3, plane, shift, mbn; | ||
| 703 |
1/2✓ Branch 0 taken 57111 times.
✗ Branch 1 not taken.
|
57111 | int entry_size = s->pix_fmt == AV_PIX_FMT_RGB24 ? 6 : 4; |
| 704 |
2/2✓ Branch 0 taken 28215 times.
✓ Branch 1 taken 28896 times.
|
57111 | int *codebook = v1mode ? info->v1_codebook : info->v4_codebook; |
| 705 |
2/2✓ Branch 0 taken 28215 times.
✓ Branch 1 taken 28896 times.
|
57111 | int size = v1mode ? info->v1_size : info->v4_size; |
| 706 | uint8_t vq_pict_buf[(MB_AREA * 3) / 2]; | ||
| 707 | uint8_t *sub_data[4], *vq_data[4]; | ||
| 708 | int sub_linesize[4], vq_linesize[4]; | ||
| 709 | int ret; | ||
| 710 | |||
| 711 |
2/2✓ Branch 0 taken 552032 times.
✓ Branch 1 taken 57111 times.
|
609143 | for (mbn = i = y = 0; y < h; y += MB_SIZE) { |
| 712 |
2/2✓ Branch 0 taken 17665024 times.
✓ Branch 1 taken 552032 times.
|
18217056 | for (x = 0; x < s->w; x += MB_SIZE, ++mbn) { |
| 713 | int *base; | ||
| 714 | |||
| 715 |
2/2✓ Branch 0 taken 12312832 times.
✓ Branch 1 taken 5352192 times.
|
17665024 | if (CERTAIN(encoding)) { |
| 716 | // use for the training only the blocks known to be to be encoded [sic:-] | ||
| 717 |
2/2✓ Branch 0 taken 7768292 times.
✓ Branch 1 taken 4544540 times.
|
12312832 | if (s->mb[mbn].best_encoding != encoding) |
| 718 | 7768292 | continue; | |
| 719 | } | ||
| 720 | |||
| 721 | 9896732 | base = s->codebook_input + i * entry_size; | |
| 722 |
2/2✓ Branch 0 taken 4467400 times.
✓ Branch 1 taken 5429332 times.
|
9896732 | if (v1mode) { |
| 723 | // subsample | ||
| 724 |
2/2✓ Branch 0 taken 13402200 times.
✓ Branch 1 taken 4467400 times.
|
17869600 | for (j = y2 = 0; y2 < entry_size; y2 += 2) |
| 725 |
2/2✓ Branch 0 taken 26804400 times.
✓ Branch 1 taken 13402200 times.
|
40206600 | for (x2 = 0; x2 < 4; x2 += 2, j++) { |
| 726 |
2/2✓ Branch 0 taken 8934800 times.
✓ Branch 1 taken 17869600 times.
|
26804400 | plane = y2 < 4 ? 0 : 1 + (x2 >> 1); |
| 727 | 26804400 | shift = y2 < 4 ? 0 : 1; | |
| 728 |
2/2✓ Branch 0 taken 17869600 times.
✓ Branch 1 taken 8934800 times.
|
26804400 | x3 = shift ? 0 : x2; |
| 729 |
2/2✓ Branch 0 taken 17869600 times.
✓ Branch 1 taken 8934800 times.
|
26804400 | y3 = shift ? 0 : y2; |
| 730 | 26804400 | base[j] = (data[plane][((x + x3) >> shift) + ((y + y3) >> shift) * linesize[plane]] + | |
| 731 | 26804400 | data[plane][((x + x3) >> shift) + 1 + ((y + y3) >> shift) * linesize[plane]] + | |
| 732 | 26804400 | data[plane][((x + x3) >> shift) + (((y + y3) >> shift) + 1) * linesize[plane]] + | |
| 733 | 26804400 | data[plane][((x + x3) >> shift) + 1 + (((y + y3) >> shift) + 1) * linesize[plane]]) >> 2; | |
| 734 | } | ||
| 735 | } else { | ||
| 736 | // copy | ||
| 737 |
2/2✓ Branch 0 taken 10858664 times.
✓ Branch 1 taken 5429332 times.
|
16287996 | for (j = y2 = 0; y2 < MB_SIZE; y2 += 2) { |
| 738 |
2/2✓ Branch 0 taken 21717328 times.
✓ Branch 1 taken 10858664 times.
|
32575992 | for (x2 = 0; x2 < MB_SIZE; x2 += 2) |
| 739 |
2/2✓ Branch 0 taken 130303968 times.
✓ Branch 1 taken 21717328 times.
|
152021296 | for (k = 0; k < entry_size; k++, j++) { |
| 740 | 130303968 | plane = k >= 4 ? k - 3 : 0; | |
| 741 | |||
| 742 |
2/2✓ Branch 0 taken 43434656 times.
✓ Branch 1 taken 86869312 times.
|
130303968 | if (k >= 4) { |
| 743 | 43434656 | x3 = (x + x2) >> 1; | |
| 744 | 43434656 | y3 = (y + y2) >> 1; | |
| 745 | } else { | ||
| 746 | 86869312 | x3 = x + x2 + (k & 1); | |
| 747 | 86869312 | y3 = y + y2 + (k >> 1); | |
| 748 | } | ||
| 749 | |||
| 750 | 130303968 | base[j] = data[plane][x3 + y3 * linesize[plane]]; | |
| 751 | } | ||
| 752 | } | ||
| 753 | } | ||
| 754 |
2/2✓ Branch 0 taken 4467400 times.
✓ Branch 1 taken 5429332 times.
|
9896732 | i += v1mode ? 1 : 4; |
| 755 | } | ||
| 756 | } | ||
| 757 | |||
| 758 |
2/2✓ Branch 0 taken 36 times.
✓ Branch 1 taken 57075 times.
|
57111 | if (i == 0) // empty training set, nothing to do |
| 759 | 36 | return 0; | |
| 760 |
2/2✓ Branch 0 taken 4889 times.
✓ Branch 1 taken 52186 times.
|
57075 | if (i < size) |
| 761 | 4889 | size = i; | |
| 762 | |||
| 763 | 57075 | ret = avpriv_elbg_do(&s->elbg, s->codebook_input, entry_size, i, codebook, | |
| 764 | size, 1, s->codebook_closest, &s->randctx, 0); | ||
| 765 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 57075 times.
|
57075 | if (ret < 0) |
| 766 | ✗ | return ret; | |
| 767 | |||
| 768 | // set up vq_data, which contains a single MB | ||
| 769 | 57075 | vq_data[0] = vq_pict_buf; | |
| 770 | 57075 | vq_linesize[0] = MB_SIZE; | |
| 771 | 57075 | vq_data[1] = &vq_pict_buf[MB_AREA]; | |
| 772 | 57075 | vq_data[2] = vq_data[1] + (MB_AREA >> 2); | |
| 773 | 57075 | vq_linesize[1] = | |
| 774 | 57075 | vq_linesize[2] = MB_SIZE >> 1; | |
| 775 | |||
| 776 | // copy indices | ||
| 777 |
2/2✓ Branch 0 taken 551724 times.
✓ Branch 1 taken 57075 times.
|
608799 | for (i = j = y = 0; y < h; y += MB_SIZE) |
| 778 |
2/2✓ Branch 0 taken 17655168 times.
✓ Branch 1 taken 551724 times.
|
18206892 | for (x = 0; x < s->w; x += MB_SIZE, j++) { |
| 779 | 17655168 | mb_info *mb = &s->mb[j]; | |
| 780 | // skip uninteresting blocks if we know their preferred encoding | ||
| 781 |
4/4✓ Branch 0 taken 12302976 times.
✓ Branch 1 taken 5352192 times.
✓ Branch 2 taken 7758436 times.
✓ Branch 3 taken 4544540 times.
|
17655168 | if (CERTAIN(encoding) && mb->best_encoding != encoding) |
| 782 | 7758436 | continue; | |
| 783 | |||
| 784 | // point sub_data to current MB | ||
| 785 | 9896732 | get_sub_picture(s, x, y, data, linesize, sub_data, sub_linesize); | |
| 786 | |||
| 787 |
2/2✓ Branch 0 taken 4467400 times.
✓ Branch 1 taken 5429332 times.
|
9896732 | if (v1mode) { |
| 788 | 4467400 | mb->v1_vector = s->codebook_closest[i]; | |
| 789 | |||
| 790 | // fill in vq_data with V1 data | ||
| 791 | 4467400 | decode_v1_vector(s, vq_data, vq_linesize, mb->v1_vector, info); | |
| 792 | |||
| 793 | 4467400 | mb->v1_error = compute_mb_distortion(s, sub_data, sub_linesize, | |
| 794 | vq_data, vq_linesize); | ||
| 795 | } else { | ||
| 796 |
2/2✓ Branch 0 taken 21717328 times.
✓ Branch 1 taken 5429332 times.
|
27146660 | for (k = 0; k < 4; k++) |
| 797 | 21717328 | mb->v4_vector[k] = s->codebook_closest[i + k]; | |
| 798 | |||
| 799 | // fill in vq_data with V4 data | ||
| 800 | 5429332 | decode_v4_vector(s, vq_data, vq_linesize, mb->v4_vector, info); | |
| 801 | |||
| 802 | 5429332 | mb->v4_error = compute_mb_distortion(s, sub_data, sub_linesize, | |
| 803 | vq_data, vq_linesize); | ||
| 804 | } | ||
| 805 |
2/2✓ Branch 0 taken 4467400 times.
✓ Branch 1 taken 5429332 times.
|
9896732 | i += v1mode ? 1 : 4; |
| 806 | } | ||
| 807 | // check that we did it right in the beginning of the function | ||
| 808 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 57075 times.
|
57075 | av_assert0(i >= size); // training set is no smaller than the codebook |
| 809 | |||
| 810 | 57075 | return size; | |
| 811 | } | ||
| 812 | |||
| 813 | 685 | static void calculate_skip_errors(CinepakEncContext *s, int h, | |
| 814 | uint8_t *last_data[4], int last_linesize[4], | ||
| 815 | uint8_t *data[4], int linesize[4], | ||
| 816 | strip_info *info) | ||
| 817 | { | ||
| 818 | int x, y, i; | ||
| 819 | uint8_t *sub_last_data [4], *sub_pict_data [4]; | ||
| 820 | int sub_last_linesize[4], sub_pict_linesize[4]; | ||
| 821 | |||
| 822 |
2/2✓ Branch 0 taken 6576 times.
✓ Branch 1 taken 685 times.
|
7261 | for (i = y = 0; y < h; y += MB_SIZE) |
| 823 |
2/2✓ Branch 0 taken 210432 times.
✓ Branch 1 taken 6576 times.
|
217008 | for (x = 0; x < s->w; x += MB_SIZE, i++) { |
| 824 | 210432 | get_sub_picture(s, x, y, last_data, last_linesize, | |
| 825 | sub_last_data, sub_last_linesize); | ||
| 826 | 210432 | get_sub_picture(s, x, y, data, linesize, | |
| 827 | sub_pict_data, sub_pict_linesize); | ||
| 828 | |||
| 829 | 210432 | s->mb[i].skip_error = | |
| 830 | 210432 | compute_mb_distortion(s, | |
| 831 | sub_last_data, sub_last_linesize, | ||
| 832 | sub_pict_data, sub_pict_linesize); | ||
| 833 | } | ||
| 834 | 685 | } | |
| 835 | |||
| 836 | 10575 | static void write_strip_keyframe(unsigned char *buf, int keyframe) | |
| 837 | { | ||
| 838 | // actually we are exclusively using intra strip coding (how much can we win | ||
| 839 | // otherwise? how to choose which part of a codebook to update?), | ||
| 840 | // keyframes are different only because we disallow ENC_SKIP on them -- rl | ||
| 841 | // (besides, the logic here used to be inverted: ) | ||
| 842 | // buf[0] = keyframe ? 0x11: 0x10; | ||
| 843 |
2/2✓ Branch 0 taken 782 times.
✓ Branch 1 taken 9793 times.
|
10575 | buf[0] = keyframe ? 0x10 : 0x11; |
| 844 | 10575 | } | |
| 845 | |||
| 846 | 10125 | static void write_strip_header(CinepakEncContext *s, int y, int h, int keyframe, | |
| 847 | unsigned char *buf, int strip_size) | ||
| 848 | { | ||
| 849 | 10125 | write_strip_keyframe(buf, keyframe); | |
| 850 | 10125 | AV_WB24(&buf[1], strip_size + STRIP_HEADER_SIZE); | |
| 851 | // AV_WB16(&buf[4], y); /* using absolute y values works -- rl */ | ||
| 852 | 10125 | AV_WB16(&buf[4], 0); /* using relative values works as well -- rl */ | |
| 853 | 10125 | AV_WB16(&buf[6], 0); | |
| 854 | // AV_WB16(&buf[8], y + h); /* using absolute y values works -- rl */ | ||
| 855 | 10125 | AV_WB16(&buf[8], h); /* using relative values works as well -- rl */ | |
| 856 | 10125 | AV_WB16(&buf[10], s->w); | |
| 857 | 10125 | } | |
| 858 | |||
| 859 | 753 | static int rd_strip(CinepakEncContext *s, int y, int h, int keyframe, | |
| 860 | uint8_t *last_data[4], int last_linesize[4], | ||
| 861 | uint8_t *data[4], int linesize[4], | ||
| 862 | uint8_t *scratch_data[4], int scratch_linesize[4], | ||
| 863 | unsigned char *buf, int64_t *best_score, int *no_skip) | ||
| 864 | { | ||
| 865 | 753 | int64_t score = 0; | |
| 866 | 753 | int best_size = 0; | |
| 867 | strip_info info; | ||
| 868 | // for codebook optimization: | ||
| 869 | int v1enough, v1_size, v4enough, v4_size; | ||
| 870 | int new_v1_size, new_v4_size; | ||
| 871 | int v1shrunk, v4shrunk; | ||
| 872 | |||
| 873 |
2/2✓ Branch 0 taken 685 times.
✓ Branch 1 taken 68 times.
|
753 | if (!keyframe) |
| 874 | 685 | calculate_skip_errors(s, h, last_data, last_linesize, data, linesize, | |
| 875 | &info); | ||
| 876 | |||
| 877 | // try some powers of 4 for the size of the codebooks | ||
| 878 | // constraint the v4 codebook to be no bigger than v1 one, | ||
| 879 | // (and no less than v1_size/4) | ||
| 880 | // thus making v1 preferable and possibly losing small details? should be ok | ||
| 881 | #define SMALLEST_CODEBOOK 1 | ||
| 882 |
3/4✓ Branch 0 taken 3765 times.
✓ Branch 1 taken 753 times.
✓ Branch 2 taken 3765 times.
✗ Branch 3 not taken.
|
4518 | for (v1enough = 0, v1_size = SMALLEST_CODEBOOK; v1_size <= CODEBOOK_MAX && !v1enough; v1_size <<= 2) { |
| 883 |
7/8✓ Branch 0 taken 6777 times.
✓ Branch 1 taken 3765 times.
✓ Branch 2 taken 3012 times.
✓ Branch 3 taken 753 times.
✓ Branch 4 taken 10542 times.
✓ Branch 5 taken 3765 times.
✓ Branch 6 taken 10542 times.
✗ Branch 7 not taken.
|
14307 | for (v4enough = 0, v4_size = 0; v4_size <= v1_size && !v4enough; v4_size = v4_size ? v4_size << 2 : v1_size >= SMALLEST_CODEBOOK << 2 ? v1_size >> 2 : SMALLEST_CODEBOOK) { |
| 884 | CinepakMode mode; | ||
| 885 | // try all modes | ||
| 886 |
2/2✓ Branch 0 taken 31626 times.
✓ Branch 1 taken 10542 times.
|
42168 | for (mode = 0; mode < MODE_COUNT; mode++) { |
| 887 | // don't allow MODE_MC in intra frames | ||
| 888 |
4/4✓ Branch 0 taken 2856 times.
✓ Branch 1 taken 28770 times.
✓ Branch 2 taken 952 times.
✓ Branch 3 taken 1904 times.
|
31626 | if (keyframe && mode == MODE_MC) |
| 889 | 952 | continue; | |
| 890 | |||
| 891 |
2/2✓ Branch 0 taken 10542 times.
✓ Branch 1 taken 20132 times.
|
30674 | if (mode == MODE_V1_ONLY) { |
| 892 | 10542 | info.v1_size = v1_size; | |
| 893 | // the size may shrink even before optimizations if the input is short: | ||
| 894 |
1/2✗ Branch 1 not taken.
✓ Branch 2 taken 10542 times.
|
10542 | if ((new_v1_size = quantize(s, h, data, linesize, 1, |
| 895 | &info, ENC_UNCERTAIN)) < 0) | ||
| 896 | ✗ | return new_v1_size; | |
| 897 | 10542 | info.v1_size = new_v1_size; | |
| 898 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 10542 times.
|
10542 | if (info.v1_size < v1_size) |
| 899 | // too few eligible blocks, no sense in trying bigger sizes | ||
| 900 | ✗ | v1enough = 1; | |
| 901 | |||
| 902 | 10542 | info.v4_size = 0; | |
| 903 | } else { // mode != MODE_V1_ONLY | ||
| 904 | // if v4 codebook is empty then only allow V1-only mode | ||
| 905 |
2/2✓ Branch 0 taken 7190 times.
✓ Branch 1 taken 12942 times.
|
20132 | if (!v4_size) |
| 906 | 7190 | continue; | |
| 907 | |||
| 908 |
2/2✓ Branch 0 taken 6777 times.
✓ Branch 1 taken 6165 times.
|
12942 | if (mode == MODE_V1_V4) { |
| 909 | 6777 | info.v4_size = v4_size; | |
| 910 | 6777 | new_v4_size = quantize(s, h, data, linesize, 0, | |
| 911 | &info, ENC_UNCERTAIN); | ||
| 912 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 6777 times.
|
6777 | if (new_v4_size < 0) |
| 913 | ✗ | return new_v4_size; | |
| 914 | 6777 | info.v4_size = new_v4_size; | |
| 915 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 6777 times.
|
6777 | if (info.v4_size < v4_size) |
| 916 | // too few eligible blocks, no sense in trying bigger sizes | ||
| 917 | ✗ | v4enough = 1; | |
| 918 | } | ||
| 919 | } | ||
| 920 | |||
| 921 | 23484 | info.mode = mode; | |
| 922 | // choose the best encoding per block, based on current experience | ||
| 923 | 23484 | score = calculate_mode_score(s, h, &info, 0, | |
| 924 | &v1shrunk, &v4shrunk); | ||
| 925 | |||
| 926 |
2/2✓ Branch 0 taken 12942 times.
✓ Branch 1 taken 10542 times.
|
23484 | if (mode != MODE_V1_ONLY) { |
| 927 | 12942 | int extra_iterations_limit = s->max_extra_cb_iterations; | |
| 928 | // recompute the codebooks, omitting the extra blocks | ||
| 929 | // we assume we _may_ come here with more blocks to encode than before | ||
| 930 | 12942 | info.v1_size = v1_size; | |
| 931 | 12942 | new_v1_size = quantize(s, h, data, linesize, 1, &info, ENC_V1); | |
| 932 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 12942 times.
|
12942 | if (new_v1_size < 0) |
| 933 | ✗ | return new_v1_size; | |
| 934 |
2/2✓ Branch 0 taken 4899 times.
✓ Branch 1 taken 8043 times.
|
12942 | if (new_v1_size < info.v1_size) |
| 935 | 4899 | info.v1_size = new_v1_size; | |
| 936 | // we assume we _may_ come here with more blocks to encode than before | ||
| 937 | 12942 | info.v4_size = v4_size; | |
| 938 | 12942 | new_v4_size = quantize(s, h, data, linesize, 0, &info, ENC_V4); | |
| 939 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 12942 times.
|
12942 | if (new_v4_size < 0) |
| 940 | ✗ | return new_v4_size; | |
| 941 |
2/2✓ Branch 0 taken 17 times.
✓ Branch 1 taken 12925 times.
|
12942 | if (new_v4_size < info.v4_size) |
| 942 | 17 | info.v4_size = new_v4_size; | |
| 943 | // calculate the resulting score | ||
| 944 | // (do not move blocks to codebook encodings now, as some blocks may have | ||
| 945 | // got bigger errors despite a smaller training set - but we do not | ||
| 946 | // ever grow the training sets back) | ||
| 947 | for (;;) { | ||
| 948 | 23565 | score = calculate_mode_score(s, h, &info, 1, | |
| 949 | &v1shrunk, &v4shrunk); | ||
| 950 | // do we have a reason to reiterate? if so, have we reached the limit? | ||
| 951 |
6/6✓ Branch 0 taken 17781 times.
✓ Branch 1 taken 5784 times.
✓ Branch 2 taken 8848 times.
✓ Branch 3 taken 8933 times.
✓ Branch 4 taken 10623 times.
✓ Branch 5 taken 4009 times.
|
23565 | if ((!v1shrunk && !v4shrunk) || !extra_iterations_limit--) |
| 952 | break; | ||
| 953 | // recompute the codebooks, omitting the extra blocks | ||
| 954 |
2/2✓ Branch 0 taken 4731 times.
✓ Branch 1 taken 5892 times.
|
10623 | if (v1shrunk) { |
| 955 | 4731 | info.v1_size = v1_size; | |
| 956 | 4731 | new_v1_size = quantize(s, h, data, linesize, 1, &info, ENC_V1); | |
| 957 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 4731 times.
|
4731 | if (new_v1_size < 0) |
| 958 | ✗ | return new_v1_size; | |
| 959 |
2/2✓ Branch 0 taken 6 times.
✓ Branch 1 taken 4725 times.
|
4731 | if (new_v1_size < info.v1_size) |
| 960 | 6 | info.v1_size = new_v1_size; | |
| 961 | } | ||
| 962 |
2/2✓ Branch 0 taken 9177 times.
✓ Branch 1 taken 1446 times.
|
10623 | if (v4shrunk) { |
| 963 | 9177 | info.v4_size = v4_size; | |
| 964 | 9177 | new_v4_size = quantize(s, h, data, linesize, 0, &info, ENC_V4); | |
| 965 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 9177 times.
|
9177 | if (new_v4_size < 0) |
| 966 | ✗ | return new_v4_size; | |
| 967 |
2/2✓ Branch 0 taken 3 times.
✓ Branch 1 taken 9174 times.
|
9177 | if (new_v4_size < info.v4_size) |
| 968 | 3 | info.v4_size = new_v4_size; | |
| 969 | } | ||
| 970 | } | ||
| 971 | } | ||
| 972 | |||
| 973 |
4/4✓ Branch 0 taken 22731 times.
✓ Branch 1 taken 753 times.
✓ Branch 2 taken 9372 times.
✓ Branch 3 taken 13359 times.
|
23484 | if (best_size == 0 || score < *best_score) { |
| 974 | 10125 | *best_score = score; | |
| 975 | 10125 | best_size = encode_mode(s, h, | |
| 976 | scratch_data, scratch_linesize, | ||
| 977 | last_data, last_linesize, &info, | ||
| 978 | 10125 | s->strip_buf + STRIP_HEADER_SIZE); | |
| 979 | // in theory we could have MODE_MC without ENC_SKIP, | ||
| 980 | // but MODE_V1_V4 will always be more efficient | ||
| 981 | 10125 | *no_skip = info.mode != MODE_MC; | |
| 982 | |||
| 983 | 10125 | write_strip_header(s, y, h, keyframe, s->strip_buf, best_size); | |
| 984 | } | ||
| 985 | } | ||
| 986 | } | ||
| 987 | } | ||
| 988 | |||
| 989 | 753 | best_size += STRIP_HEADER_SIZE; | |
| 990 | 753 | memcpy(buf, s->strip_buf, best_size); | |
| 991 | |||
| 992 | 753 | return best_size; | |
| 993 | } | ||
| 994 | |||
| 995 | 303 | static int write_cvid_header(CinepakEncContext *s, unsigned char *buf, | |
| 996 | int num_strips, int data_size, int isakeyframe) | ||
| 997 | { | ||
| 998 | 303 | buf[0] = isakeyframe ? 0 : 1; | |
| 999 | 303 | AV_WB24(&buf[1], data_size + CVID_HEADER_SIZE); | |
| 1000 | 303 | AV_WB16(&buf[4], s->w); | |
| 1001 | 303 | AV_WB16(&buf[6], s->h); | |
| 1002 | 303 | AV_WB16(&buf[8], num_strips); | |
| 1003 | |||
| 1004 | 303 | return CVID_HEADER_SIZE; | |
| 1005 | } | ||
| 1006 | |||
| 1007 | 150 | static int rd_frame(CinepakEncContext *s, const AVFrame *frame, | |
| 1008 | int isakeyframe, unsigned char *buf, int buf_size, int *got_keyframe) | ||
| 1009 | { | ||
| 1010 | int num_strips, strip, i, y, nexty, size, temp_size, best_size; | ||
| 1011 | uint8_t *last_data [4], *data [4], *scratch_data [4]; | ||
| 1012 | int last_linesize[4], linesize[4], scratch_linesize[4]; | ||
| 1013 | 150 | int64_t best_score = 0, score, score_temp; | |
| 1014 | int best_nstrips, best_strip_offsets[MAX_STRIPS]; | ||
| 1015 | |||
| 1016 |
1/2✓ Branch 0 taken 150 times.
✗ Branch 1 not taken.
|
150 | if (s->pix_fmt == AV_PIX_FMT_RGB24) { |
| 1017 | int x; | ||
| 1018 | // build a copy of the given frame in the correct colorspace | ||
| 1019 |
2/2✓ Branch 0 taken 7200 times.
✓ Branch 1 taken 150 times.
|
7350 | for (y = 0; y < s->h; y += 2) |
| 1020 |
2/2✓ Branch 0 taken 460800 times.
✓ Branch 1 taken 7200 times.
|
468000 | for (x = 0; x < s->w; x += 2) { |
| 1021 | const uint8_t *ir[2]; | ||
| 1022 | int32_t r, g, b, rr, gg, bb; | ||
| 1023 | 460800 | ir[0] = frame->data[0] + x * 3 + y * frame->linesize[0]; | |
| 1024 | 460800 | ir[1] = ir[0] + frame->linesize[0]; | |
| 1025 | 460800 | get_sub_picture(s, x, y, | |
| 1026 | 460800 | s->input_frame->data, s->input_frame->linesize, | |
| 1027 | scratch_data, scratch_linesize); | ||
| 1028 | 460800 | r = g = b = 0; | |
| 1029 |
2/2✓ Branch 0 taken 1843200 times.
✓ Branch 1 taken 460800 times.
|
2304000 | for (i = 0; i < 4; ++i) { |
| 1030 | int i1, i2; | ||
| 1031 | 1843200 | i1 = (i & 1); | |
| 1032 | 1843200 | i2 = (i >= 2); | |
| 1033 | 1843200 | rr = ir[i2][i1 * 3 + 0]; | |
| 1034 | 1843200 | gg = ir[i2][i1 * 3 + 1]; | |
| 1035 | 1843200 | bb = ir[i2][i1 * 3 + 2]; | |
| 1036 | 1843200 | r += rr; | |
| 1037 | 1843200 | g += gg; | |
| 1038 | 1843200 | b += bb; | |
| 1039 | // using fixed point arithmetic for portable repeatability, scaling by 2^23 | ||
| 1040 | // "Y" | ||
| 1041 | // rr = 0.2857 * rr + 0.5714 * gg + 0.1429 * bb; | ||
| 1042 | 1843200 | rr = (2396625 * rr + 4793251 * gg + 1198732 * bb) >> 23; | |
| 1043 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 1843200 times.
|
1843200 | if (rr < 0) |
| 1044 | ✗ | rr = 0; | |
| 1045 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 1843200 times.
|
1843200 | else if (rr > 255) |
| 1046 | ✗ | rr = 255; | |
| 1047 | 1843200 | scratch_data[0][i1 + i2 * scratch_linesize[0]] = rr; | |
| 1048 | } | ||
| 1049 | // let us scale down as late as possible | ||
| 1050 | // r /= 4; g /= 4; b /= 4; | ||
| 1051 | // "U" | ||
| 1052 | // rr = -0.1429 * r - 0.2857 * g + 0.4286 * b; | ||
| 1053 | 460800 | rr = (-299683 * r - 599156 * g + 898839 * b) >> 23; | |
| 1054 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 460800 times.
|
460800 | if (rr < -128) |
| 1055 | ✗ | rr = -128; | |
| 1056 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 460800 times.
|
460800 | else if (rr > 127) |
| 1057 | ✗ | rr = 127; | |
| 1058 | 460800 | scratch_data[1][0] = rr + 128; // quantize needs unsigned | |
| 1059 | // "V" | ||
| 1060 | // rr = 0.3571 * r - 0.2857 * g - 0.0714 * b; | ||
| 1061 | 460800 | rr = (748893 * r - 599156 * g - 149737 * b) >> 23; | |
| 1062 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 460800 times.
|
460800 | if (rr < -128) |
| 1063 | ✗ | rr = -128; | |
| 1064 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 460800 times.
|
460800 | else if (rr > 127) |
| 1065 | ✗ | rr = 127; | |
| 1066 | 460800 | scratch_data[2][0] = rr + 128; // quantize needs unsigned | |
| 1067 | } | ||
| 1068 | } | ||
| 1069 | |||
| 1070 | // would be nice but quite certainly incompatible with vintage players: | ||
| 1071 | // support encoding zero strips (meaning skip the whole frame) | ||
| 1072 |
3/4✓ Branch 0 taken 303 times.
✓ Branch 1 taken 150 times.
✓ Branch 2 taken 303 times.
✗ Branch 3 not taken.
|
453 | for (num_strips = s->min_strips; num_strips <= s->max_strips && num_strips <= s->h / MB_SIZE; num_strips++) { |
| 1073 | int strip_offsets[MAX_STRIPS]; | ||
| 1074 | 303 | int all_no_skip = 1; | |
| 1075 | 303 | score = 0; | |
| 1076 | 303 | size = 0; | |
| 1077 | |||
| 1078 |
2/2✓ Branch 0 taken 753 times.
✓ Branch 1 taken 303 times.
|
1056 | for (y = 0, strip = 1; y < s->h; strip++, y = nexty) { |
| 1079 | int strip_height, no_skip; | ||
| 1080 | |||
| 1081 | 753 | strip_offsets[strip-1] = size + CVID_HEADER_SIZE; | |
| 1082 | 753 | nexty = strip * s->h / num_strips; // <= s->h | |
| 1083 | // make nexty the next multiple of 4 if not already there | ||
| 1084 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 753 times.
|
753 | if (nexty & 3) |
| 1085 | ✗ | nexty += 4 - (nexty & 3); | |
| 1086 | |||
| 1087 | 753 | strip_height = nexty - y; | |
| 1088 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 753 times.
|
753 | if (strip_height <= 0) { // can this ever happen? |
| 1089 | ✗ | av_log(s->avctx, AV_LOG_INFO, "skipping zero height strip %i of %i\n", strip, num_strips); | |
| 1090 | ✗ | continue; | |
| 1091 | } | ||
| 1092 | |||
| 1093 |
1/2✓ Branch 0 taken 753 times.
✗ Branch 1 not taken.
|
753 | if (s->pix_fmt == AV_PIX_FMT_RGB24) |
| 1094 | 753 | get_sub_picture(s, 0, y, | |
| 1095 | 753 | s->input_frame->data, s->input_frame->linesize, | |
| 1096 | data, linesize); | ||
| 1097 | else | ||
| 1098 | ✗ | get_sub_picture(s, 0, y, | |
| 1099 | ✗ | frame->data, frame->linesize, | |
| 1100 | data, linesize); | ||
| 1101 | 753 | get_sub_picture(s, 0, y, | |
| 1102 | 753 | s->last_frame->data, s->last_frame->linesize, | |
| 1103 | last_data, last_linesize); | ||
| 1104 | 753 | get_sub_picture(s, 0, y, | |
| 1105 | 753 | s->scratch_frame->data, s->scratch_frame->linesize, | |
| 1106 | scratch_data, scratch_linesize); | ||
| 1107 | |||
| 1108 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 753 times.
|
753 | if ((temp_size = rd_strip(s, y, strip_height, isakeyframe, |
| 1109 | last_data, last_linesize, data, linesize, | ||
| 1110 | scratch_data, scratch_linesize, | ||
| 1111 | 753 | s->frame_buf + strip_offsets[strip-1], | |
| 1112 | &score_temp, &no_skip)) < 0) | ||
| 1113 | ✗ | return temp_size; | |
| 1114 | |||
| 1115 | 753 | score += score_temp; | |
| 1116 | 753 | size += temp_size; | |
| 1117 | 753 | all_no_skip &= no_skip; | |
| 1118 | } | ||
| 1119 | |||
| 1120 |
3/4✓ Branch 0 taken 153 times.
✓ Branch 1 taken 150 times.
✓ Branch 2 taken 153 times.
✗ Branch 3 not taken.
|
303 | if (best_score == 0 || score < best_score) { |
| 1121 | 303 | best_score = score; | |
| 1122 | 303 | best_size = size + write_cvid_header(s, s->frame_buf, num_strips, size, all_no_skip); | |
| 1123 | |||
| 1124 | 303 | FFSWAP(AVFrame *, s->best_frame, s->scratch_frame); | |
| 1125 | 303 | memcpy(buf, s->frame_buf, best_size); | |
| 1126 | 303 | best_nstrips = num_strips; | |
| 1127 | 303 | *got_keyframe = all_no_skip; // no skip MBs in any strip -> keyframe | |
| 1128 | 303 | memcpy(best_strip_offsets, strip_offsets, sizeof(strip_offsets)); | |
| 1129 | } | ||
| 1130 | // avoid trying too many strip numbers without a real reason | ||
| 1131 | // (this makes the processing of the very first frame faster) | ||
| 1132 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 303 times.
|
303 | if (num_strips - best_nstrips > 4) |
| 1133 | ✗ | break; | |
| 1134 | } | ||
| 1135 | |||
| 1136 | // update strip headers | ||
| 1137 |
2/2✓ Branch 0 taken 450 times.
✓ Branch 1 taken 150 times.
|
600 | for (i = 0; i < best_nstrips; i++) { |
| 1138 | 450 | write_strip_keyframe(s->frame_buf + best_strip_offsets[i], *got_keyframe); | |
| 1139 | } | ||
| 1140 | |||
| 1141 | // let the number of strips slowly adapt to the changes in the contents, | ||
| 1142 | // compared to full bruteforcing every time this will occasionally lead | ||
| 1143 | // to some r/d performance loss but makes encoding up to several times faster | ||
| 1144 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 150 times.
|
150 | if (!s->strip_number_delta_range) { |
| 1145 | ✗ | if (best_nstrips == s->max_strips) { // let us try to step up | |
| 1146 | ✗ | s->max_strips = best_nstrips + 1; | |
| 1147 | ✗ | if (s->max_strips >= s->max_max_strips) | |
| 1148 | ✗ | s->max_strips = s->max_max_strips; | |
| 1149 | } else { // try to step down | ||
| 1150 | ✗ | s->max_strips = best_nstrips; | |
| 1151 | } | ||
| 1152 | ✗ | s->min_strips = s->max_strips - 1; | |
| 1153 | ✗ | if (s->min_strips < s->min_min_strips) | |
| 1154 | ✗ | s->min_strips = s->min_min_strips; | |
| 1155 | } else { | ||
| 1156 | 150 | s->max_strips = best_nstrips + s->strip_number_delta_range; | |
| 1157 |
1/2✓ Branch 0 taken 150 times.
✗ Branch 1 not taken.
|
150 | if (s->max_strips >= s->max_max_strips) |
| 1158 | 150 | s->max_strips = s->max_max_strips; | |
| 1159 | 150 | s->min_strips = best_nstrips - s->strip_number_delta_range; | |
| 1160 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 150 times.
|
150 | if (s->min_strips < s->min_min_strips) |
| 1161 | ✗ | s->min_strips = s->min_min_strips; | |
| 1162 | } | ||
| 1163 | |||
| 1164 | 150 | return best_size; | |
| 1165 | } | ||
| 1166 | |||
| 1167 | 150 | static int cinepak_encode_frame(AVCodecContext *avctx, AVPacket *pkt, | |
| 1168 | const AVFrame *frame, int *got_packet) | ||
| 1169 | { | ||
| 1170 | 150 | CinepakEncContext *s = avctx->priv_data; | |
| 1171 | int ret, got_keyframe; | ||
| 1172 | |||
| 1173 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 150 times.
|
150 | s->lambda = frame->quality ? frame->quality - 1 : 2 * FF_LAMBDA_SCALE; |
| 1174 | |||
| 1175 |
1/2✗ Branch 1 not taken.
✓ Branch 2 taken 150 times.
|
150 | if ((ret = ff_alloc_packet(avctx, pkt, s->frame_buf_size)) < 0) |
| 1176 | ✗ | return ret; | |
| 1177 | 150 | ret = rd_frame(s, frame, (s->curframe == 0), pkt->data, s->frame_buf_size, &got_keyframe); | |
| 1178 | 150 | pkt->size = ret; | |
| 1179 |
2/2✓ Branch 0 taken 15 times.
✓ Branch 1 taken 135 times.
|
150 | if (got_keyframe) { |
| 1180 | 15 | pkt->flags |= AV_PKT_FLAG_KEY; | |
| 1181 | 15 | s->curframe = 0; | |
| 1182 | } | ||
| 1183 | 150 | *got_packet = 1; | |
| 1184 | |||
| 1185 | 150 | FFSWAP(AVFrame *, s->last_frame, s->best_frame); | |
| 1186 | |||
| 1187 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 140 times.
|
150 | if (++s->curframe >= avctx->gop_size) |
| 1188 | 10 | s->curframe = 0; | |
| 1189 | |||
| 1190 | 150 | return 0; | |
| 1191 | } | ||
| 1192 | |||
| 1193 | 3 | static av_cold int cinepak_encode_end(AVCodecContext *avctx) | |
| 1194 | { | ||
| 1195 | 3 | CinepakEncContext *s = avctx->priv_data; | |
| 1196 | int x; | ||
| 1197 | |||
| 1198 | 3 | avpriv_elbg_free(&s->elbg); | |
| 1199 | 3 | av_frame_free(&s->last_frame); | |
| 1200 | 3 | av_frame_free(&s->best_frame); | |
| 1201 | 3 | av_frame_free(&s->scratch_frame); | |
| 1202 |
1/2✓ Branch 0 taken 3 times.
✗ Branch 1 not taken.
|
3 | if (avctx->pix_fmt == AV_PIX_FMT_RGB24) |
| 1203 | 3 | av_frame_free(&s->input_frame); | |
| 1204 | 3 | av_freep(&s->codebook_input); | |
| 1205 | 3 | av_freep(&s->codebook_closest); | |
| 1206 | 3 | av_freep(&s->strip_buf); | |
| 1207 | 3 | av_freep(&s->frame_buf); | |
| 1208 | 3 | av_freep(&s->mb); | |
| 1209 | |||
| 1210 |
3/4✓ Branch 0 taken 15 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 12 times.
✓ Branch 3 taken 3 times.
|
15 | for (x = 0; x < (avctx->pix_fmt == AV_PIX_FMT_RGB24 ? 4 : 3); x++) |
| 1211 | 12 | av_freep(&s->pict_bufs[x]); | |
| 1212 | |||
| 1213 | 3 | return 0; | |
| 1214 | } | ||
| 1215 | |||
| 1216 | const FFCodec ff_cinepak_encoder = { | ||
| 1217 | .p.name = "cinepak", | ||
| 1218 | CODEC_LONG_NAME("Cinepak"), | ||
| 1219 | .p.type = AVMEDIA_TYPE_VIDEO, | ||
| 1220 | .p.id = AV_CODEC_ID_CINEPAK, | ||
| 1221 | .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_ENCODER_REORDERED_OPAQUE, | ||
| 1222 | .priv_data_size = sizeof(CinepakEncContext), | ||
| 1223 | .init = cinepak_encode_init, | ||
| 1224 | FF_CODEC_ENCODE_CB(cinepak_encode_frame), | ||
| 1225 | .close = cinepak_encode_end, | ||
| 1226 | CODEC_PIXFMTS(AV_PIX_FMT_RGB24, AV_PIX_FMT_GRAY8), | ||
| 1227 | .p.priv_class = &cinepak_class, | ||
| 1228 | .caps_internal = FF_CODEC_CAP_INIT_CLEANUP, | ||
| 1229 | }; | ||
| 1230 |