| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /* | ||
| 2 | * Apple ProRes encoder | ||
| 3 | * | ||
| 4 | * Copyright (c) 2011 Anatoliy Wasserman | ||
| 5 | * Copyright (c) 2012 Konstantin Shishkov | ||
| 6 | * | ||
| 7 | * This file is part of FFmpeg. | ||
| 8 | * | ||
| 9 | * FFmpeg is free software; you can redistribute it and/or | ||
| 10 | * modify it under the terms of the GNU Lesser General Public | ||
| 11 | * License as published by the Free Software Foundation; either | ||
| 12 | * version 2.1 of the License, or (at your option) any later version. | ||
| 13 | * | ||
| 14 | * FFmpeg is distributed in the hope that it will be useful, | ||
| 15 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
| 16 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | ||
| 17 | * Lesser General Public License for more details. | ||
| 18 | * | ||
| 19 | * You should have received a copy of the GNU Lesser General Public | ||
| 20 | * License along with FFmpeg; if not, write to the Free Software | ||
| 21 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | ||
| 22 | */ | ||
| 23 | |||
| 24 | #include "libavutil/mem.h" | ||
| 25 | #include "libavutil/mem_internal.h" | ||
| 26 | #include "libavutil/opt.h" | ||
| 27 | #include "libavutil/pixdesc.h" | ||
| 28 | #include "avcodec.h" | ||
| 29 | #include "codec_internal.h" | ||
| 30 | #include "encode.h" | ||
| 31 | #include "fdctdsp.h" | ||
| 32 | #include "put_bits.h" | ||
| 33 | #include "profiles.h" | ||
| 34 | #include "bytestream.h" | ||
| 35 | #include "proresdata.h" | ||
| 36 | #include "proresenc_kostya_common.h" | ||
| 37 | |||
| 38 | #define TRELLIS_WIDTH 16 | ||
| 39 | #define SCORE_LIMIT INT_MAX / 2 | ||
| 40 | |||
| 41 | struct TrellisNode { | ||
| 42 | int prev_node; | ||
| 43 | int quant; | ||
| 44 | int bits; | ||
| 45 | int score; | ||
| 46 | }; | ||
| 47 | |||
| 48 | typedef struct ProresThreadData { | ||
| 49 | DECLARE_ALIGNED(16, int16_t, blocks)[MAX_PLANES][64 * 4 * MAX_MBS_PER_SLICE]; | ||
| 50 | DECLARE_ALIGNED(16, uint16_t, emu_buf)[16 * 16]; | ||
| 51 | int16_t custom_q[64]; | ||
| 52 | int16_t custom_chroma_q[64]; | ||
| 53 | struct TrellisNode *nodes; | ||
| 54 | } ProresThreadData; | ||
| 55 | |||
| 56 | 66600 | static void get_slice_data(ProresContext *ctx, const uint16_t *src, | |
| 57 | ptrdiff_t linesize, int x, int y, int w, int h, | ||
| 58 | int16_t *blocks, uint16_t *emu_buf, | ||
| 59 | int mbs_per_slice, int blocks_per_mb, int is_chroma) | ||
| 60 | { | ||
| 61 | const uint16_t *esrc; | ||
| 62 | 66600 | const int mb_width = 4 * blocks_per_mb; | |
| 63 | ptrdiff_t elinesize; | ||
| 64 | int i, j, k; | ||
| 65 | |||
| 66 |
2/2✓ Branch 0 taken 359100 times.
✓ Branch 1 taken 66600 times.
|
425700 | for (i = 0; i < mbs_per_slice; i++, src += mb_width) { |
| 67 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 359100 times.
|
359100 | if (x >= w) { |
| 68 | ✗ | memset(blocks, 0, 64 * (mbs_per_slice - i) * blocks_per_mb | |
| 69 | * sizeof(*blocks)); | ||
| 70 | ✗ | return; | |
| 71 | } | ||
| 72 |
4/4✓ Branch 0 taken 358200 times.
✓ Branch 1 taken 900 times.
✓ Branch 2 taken 357600 times.
✓ Branch 3 taken 600 times.
|
359100 | if (x + mb_width <= w && y + 16 <= h) { |
| 73 | 357600 | esrc = src; | |
| 74 | 357600 | elinesize = linesize; | |
| 75 | } else { | ||
| 76 | int bw, bh, pix; | ||
| 77 | |||
| 78 | 1500 | esrc = emu_buf; | |
| 79 | 1500 | elinesize = 16 * sizeof(*emu_buf); | |
| 80 | |||
| 81 | 1500 | bw = FFMIN(w - x, mb_width); | |
| 82 | 1500 | bh = FFMIN(h - y, 16); | |
| 83 | |||
| 84 |
2/2✓ Branch 0 taken 11400 times.
✓ Branch 1 taken 1500 times.
|
12900 | for (j = 0; j < bh; j++) { |
| 85 | 11400 | memcpy(emu_buf + j * 16, | |
| 86 | 11400 | (const uint8_t*)src + j * linesize, | |
| 87 | bw * sizeof(*src)); | ||
| 88 | 11400 | pix = emu_buf[j * 16 + bw - 1]; | |
| 89 |
2/2✓ Branch 0 taken 95200 times.
✓ Branch 1 taken 11400 times.
|
106600 | for (k = bw; k < mb_width; k++) |
| 90 | 95200 | emu_buf[j * 16 + k] = pix; | |
| 91 | } | ||
| 92 |
2/2✓ Branch 0 taken 12600 times.
✓ Branch 1 taken 1500 times.
|
14100 | for (; j < 16; j++) |
| 93 | 12600 | memcpy(emu_buf + j * 16, | |
| 94 | 12600 | emu_buf + (bh - 1) * 16, | |
| 95 | mb_width * sizeof(*emu_buf)); | ||
| 96 | } | ||
| 97 |
2/2✓ Branch 0 taken 119700 times.
✓ Branch 1 taken 239400 times.
|
359100 | if (!is_chroma) { |
| 98 | 119700 | ctx->fdct(&ctx->fdsp, esrc, elinesize, blocks); | |
| 99 | 119700 | blocks += 64; | |
| 100 |
1/2✓ Branch 0 taken 119700 times.
✗ Branch 1 not taken.
|
119700 | if (blocks_per_mb > 2) { |
| 101 | 119700 | ctx->fdct(&ctx->fdsp, esrc + 8, elinesize, blocks); | |
| 102 | 119700 | blocks += 64; | |
| 103 | } | ||
| 104 | 119700 | ctx->fdct(&ctx->fdsp, esrc + elinesize * 4, elinesize, blocks); | |
| 105 | 119700 | blocks += 64; | |
| 106 |
1/2✓ Branch 0 taken 119700 times.
✗ Branch 1 not taken.
|
119700 | if (blocks_per_mb > 2) { |
| 107 | 119700 | ctx->fdct(&ctx->fdsp, esrc + elinesize * 4 + 8, elinesize, blocks); | |
| 108 | 119700 | blocks += 64; | |
| 109 | } | ||
| 110 | } else { | ||
| 111 | 239400 | ctx->fdct(&ctx->fdsp, esrc, elinesize, blocks); | |
| 112 | 239400 | blocks += 64; | |
| 113 | 239400 | ctx->fdct(&ctx->fdsp, esrc + elinesize * 4, elinesize, blocks); | |
| 114 | 239400 | blocks += 64; | |
| 115 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 239400 times.
|
239400 | if (blocks_per_mb > 2) { |
| 116 | ✗ | ctx->fdct(&ctx->fdsp, esrc + 8, elinesize, blocks); | |
| 117 | ✗ | blocks += 64; | |
| 118 | ✗ | ctx->fdct(&ctx->fdsp, esrc + elinesize * 4 + 8, elinesize, blocks); | |
| 119 | ✗ | blocks += 64; | |
| 120 | } | ||
| 121 | } | ||
| 122 | |||
| 123 | 359100 | x += mb_width; | |
| 124 | } | ||
| 125 | } | ||
| 126 | |||
| 127 | ✗ | static void get_alpha_data(ProresContext *ctx, const uint16_t *src, | |
| 128 | ptrdiff_t linesize, int x, int y, int w, int h, | ||
| 129 | uint16_t *blocks, int mbs_per_slice, int abits) | ||
| 130 | { | ||
| 131 | ✗ | const int slice_width = 16 * mbs_per_slice; | |
| 132 | int i, j, copy_w, copy_h; | ||
| 133 | |||
| 134 | ✗ | copy_w = FFMIN(w - x, slice_width); | |
| 135 | ✗ | copy_h = FFMIN(h - y, 16); | |
| 136 | ✗ | for (i = 0; i < copy_h; i++) { | |
| 137 | ✗ | memcpy(blocks, src, copy_w * sizeof(*src)); | |
| 138 | ✗ | if (abits == 8) | |
| 139 | ✗ | for (j = 0; j < copy_w; j++) | |
| 140 | ✗ | blocks[j] >>= 2; | |
| 141 | else | ||
| 142 | ✗ | for (j = 0; j < copy_w; j++) | |
| 143 | ✗ | blocks[j] = (blocks[j] << 6) | (blocks[j] >> 4); | |
| 144 | ✗ | for (j = copy_w; j < slice_width; j++) | |
| 145 | ✗ | blocks[j] = blocks[copy_w - 1]; | |
| 146 | ✗ | blocks += slice_width; | |
| 147 | ✗ | src += linesize >> 1; | |
| 148 | } | ||
| 149 | ✗ | for (; i < 16; i++) { | |
| 150 | ✗ | memcpy(blocks, blocks - slice_width, slice_width * sizeof(*blocks)); | |
| 151 | ✗ | blocks += slice_width; | |
| 152 | } | ||
| 153 | ✗ | } | |
| 154 | |||
| 155 | /** | ||
| 156 | * Write an unsigned rice/exp golomb codeword. | ||
| 157 | */ | ||
| 158 | 26885174 | static inline void encode_vlc_codeword(PutBitContext *pb, unsigned codebook, int val) | |
| 159 | { | ||
| 160 | unsigned int rice_order, exp_order, switch_bits, switch_val; | ||
| 161 | int exponent; | ||
| 162 | |||
| 163 | /* number of prefix bits to switch between Rice and expGolomb */ | ||
| 164 | 26885174 | switch_bits = (codebook & 3) + 1; | |
| 165 | 26885174 | rice_order = codebook >> 5; /* rice code order */ | |
| 166 | 26885174 | exp_order = (codebook >> 2) & 7; /* exp golomb code order */ | |
| 167 | |||
| 168 | 26885174 | switch_val = switch_bits << rice_order; | |
| 169 | |||
| 170 |
2/2✓ Branch 0 taken 8074309 times.
✓ Branch 1 taken 18810865 times.
|
26885174 | if (val >= switch_val) { |
| 171 | 8074309 | val -= switch_val - (1 << exp_order); | |
| 172 | 8074309 | exponent = av_log2(val); | |
| 173 | |||
| 174 | 8074309 | put_bits(pb, exponent - exp_order + switch_bits, 0); | |
| 175 | 8074309 | put_bits(pb, exponent + 1, val); | |
| 176 | } else { | ||
| 177 | 18810865 | exponent = val >> rice_order; | |
| 178 | |||
| 179 |
2/2✓ Branch 0 taken 4838901 times.
✓ Branch 1 taken 13971964 times.
|
18810865 | if (exponent) |
| 180 | 4838901 | put_bits(pb, exponent, 0); | |
| 181 | 18810865 | put_bits(pb, 1, 1); | |
| 182 |
2/2✓ Branch 0 taken 1904515 times.
✓ Branch 1 taken 16906350 times.
|
18810865 | if (rice_order) |
| 183 | 1904515 | put_sbits(pb, rice_order, val); | |
| 184 | } | ||
| 185 | 26885174 | } | |
| 186 | |||
| 187 | #define GET_SIGN(x) ((x) >> 31) | ||
| 188 | #define MAKE_CODE(x) (((x) * 2) ^ GET_SIGN(x)) | ||
| 189 | |||
| 190 | 33300 | static void encode_dcs(PutBitContext *pb, int16_t *blocks, | |
| 191 | int blocks_per_slice, int scale) | ||
| 192 | { | ||
| 193 | int i; | ||
| 194 | 33300 | int codebook = 5, code, dc, prev_dc, delta, sign, new_sign; | |
| 195 | |||
| 196 | 33300 | prev_dc = (blocks[0] - 0x4000) / scale; | |
| 197 | 33300 | encode_vlc_codeword(pb, FIRST_DC_CB, MAKE_CODE(prev_dc)); | |
| 198 | 33300 | sign = 0; | |
| 199 | 33300 | blocks += 64; | |
| 200 | |||
| 201 |
2/2✓ Branch 0 taken 445500 times.
✓ Branch 1 taken 33300 times.
|
478800 | for (i = 1; i < blocks_per_slice; i++, blocks += 64) { |
| 202 | 445500 | dc = (blocks[0] - 0x4000) / scale; | |
| 203 | 445500 | delta = dc - prev_dc; | |
| 204 | 445500 | new_sign = GET_SIGN(delta); | |
| 205 | 445500 | delta = (delta ^ sign) - sign; | |
| 206 | 445500 | code = MAKE_CODE(delta); | |
| 207 | 445500 | encode_vlc_codeword(pb, ff_prores_dc_codebook[codebook], code); | |
| 208 | 445500 | codebook = FFMIN(code, 6); | |
| 209 | 445500 | sign = new_sign; | |
| 210 | 445500 | prev_dc = dc; | |
| 211 | } | ||
| 212 | 33300 | } | |
| 213 | |||
| 214 | 33300 | static void encode_acs(PutBitContext *pb, int16_t *blocks, | |
| 215 | int blocks_per_slice, | ||
| 216 | const uint8_t *scan, const int16_t *qmat) | ||
| 217 | { | ||
| 218 | int idx, i; | ||
| 219 | 33300 | int prev_run = 4; | |
| 220 | 33300 | int prev_level = 2; | |
| 221 | 33300 | int run = 0, level; | |
| 222 | int max_coeffs, abs_level; | ||
| 223 | 33300 | max_coeffs = blocks_per_slice << 6; | |
| 224 | |||
| 225 |
2/2✓ Branch 0 taken 2097900 times.
✓ Branch 1 taken 33300 times.
|
2131200 | for (i = 1; i < 64; i++) { |
| 226 |
2/2✓ Branch 0 taken 30164400 times.
✓ Branch 1 taken 2097900 times.
|
32262300 | for (idx = scan[i]; idx < max_coeffs; idx += 64) { |
| 227 | 30164400 | level = blocks[idx] / qmat[scan[i]]; | |
| 228 |
2/2✓ Branch 0 taken 13203187 times.
✓ Branch 1 taken 16961213 times.
|
30164400 | if (level) { |
| 229 | 13203187 | abs_level = FFABS(level); | |
| 230 | 13203187 | encode_vlc_codeword(pb, ff_prores_run_to_cb[prev_run], run); | |
| 231 | 13203187 | encode_vlc_codeword(pb, ff_prores_level_to_cb[prev_level], abs_level - 1); | |
| 232 | 13203187 | put_sbits(pb, 1, GET_SIGN(level)); | |
| 233 | |||
| 234 | 13203187 | prev_run = FFMIN(run, 15); | |
| 235 | 13203187 | prev_level = FFMIN(abs_level, 9); | |
| 236 | 13203187 | run = 0; | |
| 237 | } else { | ||
| 238 | 16961213 | run++; | |
| 239 | } | ||
| 240 | } | ||
| 241 | } | ||
| 242 | 33300 | } | |
| 243 | |||
| 244 | 33300 | static void encode_slice_plane(ProresContext *ctx, PutBitContext *pb, | |
| 245 | const uint16_t *src, ptrdiff_t linesize, | ||
| 246 | int mbs_per_slice, int16_t *blocks, | ||
| 247 | int blocks_per_mb, | ||
| 248 | const int16_t *qmat) | ||
| 249 | { | ||
| 250 | 33300 | int blocks_per_slice = mbs_per_slice * blocks_per_mb; | |
| 251 | |||
| 252 | 33300 | encode_dcs(pb, blocks, blocks_per_slice, qmat[0]); | |
| 253 | 33300 | encode_acs(pb, blocks, blocks_per_slice, ctx->scantable, qmat); | |
| 254 | 33300 | } | |
| 255 | |||
| 256 | ✗ | static void put_alpha_diff(PutBitContext *pb, int cur, int prev, int abits) | |
| 257 | { | ||
| 258 | ✗ | const int dbits = (abits == 8) ? 4 : 7; | |
| 259 | ✗ | const int dsize = 1 << dbits - 1; | |
| 260 | ✗ | int diff = cur - prev; | |
| 261 | |||
| 262 | ✗ | diff = av_zero_extend(diff, abits); | |
| 263 | ✗ | if (diff >= (1 << abits) - dsize) | |
| 264 | ✗ | diff -= 1 << abits; | |
| 265 | ✗ | if (diff < -dsize || diff > dsize || !diff) { | |
| 266 | ✗ | put_bits(pb, 1, 1); | |
| 267 | ✗ | put_bits(pb, abits, diff); | |
| 268 | } else { | ||
| 269 | ✗ | put_bits(pb, 1, 0); | |
| 270 | ✗ | put_bits(pb, dbits - 1, FFABS(diff) - 1); | |
| 271 | ✗ | put_bits(pb, 1, diff < 0); | |
| 272 | } | ||
| 273 | ✗ | } | |
| 274 | |||
| 275 | ✗ | static void put_alpha_run(PutBitContext *pb, int run) | |
| 276 | { | ||
| 277 | ✗ | if (run) { | |
| 278 | ✗ | put_bits(pb, 1, 0); | |
| 279 | ✗ | if (run < 0x10) | |
| 280 | ✗ | put_bits(pb, 4, run); | |
| 281 | else | ||
| 282 | ✗ | put_bits(pb, 15, run); | |
| 283 | } else { | ||
| 284 | ✗ | put_bits(pb, 1, 1); | |
| 285 | } | ||
| 286 | ✗ | } | |
| 287 | |||
| 288 | // todo alpha quantisation for high quants | ||
| 289 | ✗ | static void encode_alpha_plane(ProresContext *ctx, PutBitContext *pb, | |
| 290 | int mbs_per_slice, uint16_t *blocks, | ||
| 291 | int quant) | ||
| 292 | { | ||
| 293 | ✗ | const int abits = ctx->alpha_bits; | |
| 294 | ✗ | const int mask = (1 << abits) - 1; | |
| 295 | ✗ | const int num_coeffs = mbs_per_slice * 256; | |
| 296 | ✗ | int prev = mask, cur; | |
| 297 | ✗ | int idx = 0; | |
| 298 | ✗ | int run = 0; | |
| 299 | |||
| 300 | ✗ | cur = blocks[idx++]; | |
| 301 | ✗ | put_alpha_diff(pb, cur, prev, abits); | |
| 302 | ✗ | prev = cur; | |
| 303 | do { | ||
| 304 | ✗ | cur = blocks[idx++]; | |
| 305 | ✗ | if (cur != prev) { | |
| 306 | ✗ | put_alpha_run (pb, run); | |
| 307 | ✗ | put_alpha_diff(pb, cur, prev, abits); | |
| 308 | ✗ | prev = cur; | |
| 309 | ✗ | run = 0; | |
| 310 | } else { | ||
| 311 | ✗ | run++; | |
| 312 | } | ||
| 313 | ✗ | } while (idx < num_coeffs); | |
| 314 | ✗ | put_alpha_run(pb, run); | |
| 315 | ✗ | } | |
| 316 | |||
| 317 | 11100 | static int encode_slice(AVCodecContext *avctx, const AVFrame *pic, | |
| 318 | PutBitContext *pb, | ||
| 319 | int sizes[4], int x, int y, int quant, | ||
| 320 | int mbs_per_slice) | ||
| 321 | { | ||
| 322 | 11100 | ProresContext *ctx = avctx->priv_data; | |
| 323 | int i, xp, yp; | ||
| 324 | 11100 | int total_size = 0; | |
| 325 | const uint16_t *src; | ||
| 326 | int num_cblocks, pwidth, line_add, picture_height; | ||
| 327 | ptrdiff_t linesize; | ||
| 328 | int is_chroma; | ||
| 329 | uint16_t *qmat; | ||
| 330 | uint16_t *qmat_chroma; | ||
| 331 | |||
| 332 |
1/2✓ Branch 0 taken 11100 times.
✗ Branch 1 not taken.
|
11100 | if (ctx->pictures_per_frame == 1) |
| 333 | 11100 | line_add = 0; | |
| 334 | else | ||
| 335 | ✗ | line_add = ctx->cur_picture_idx ^ !(pic->flags & AV_FRAME_FLAG_TOP_FIELD_FIRST); | |
| 336 | |||
| 337 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 11100 times.
|
11100 | if ((y << 4) * ctx->pictures_per_frame + line_add >= avctx->height) |
| 338 | ✗ | line_add = 0; | |
| 339 | 11100 | picture_height = (avctx->height - line_add + ctx->pictures_per_frame - 1) / ctx->pictures_per_frame; | |
| 340 | |||
| 341 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 11100 times.
|
11100 | if (ctx->force_quant) { |
| 342 | ✗ | qmat = ctx->quants[0]; | |
| 343 | ✗ | qmat_chroma = ctx->quants_chroma[0]; | |
| 344 |
2/2✓ Branch 0 taken 8353 times.
✓ Branch 1 taken 2747 times.
|
11100 | } else if (quant < MAX_STORED_Q) { |
| 345 | 8353 | qmat = ctx->quants[quant]; | |
| 346 | 8353 | qmat_chroma = ctx->quants_chroma[quant]; | |
| 347 | } else { | ||
| 348 | 2747 | qmat = ctx->custom_q; | |
| 349 | 2747 | qmat_chroma = ctx->custom_chroma_q; | |
| 350 |
2/2✓ Branch 0 taken 175808 times.
✓ Branch 1 taken 2747 times.
|
178555 | for (i = 0; i < 64; i++) { |
| 351 | 175808 | qmat[i] = ctx->quant_mat[i] * quant; | |
| 352 | 175808 | qmat_chroma[i] = ctx->quant_chroma_mat[i] * quant; | |
| 353 | } | ||
| 354 | } | ||
| 355 | |||
| 356 |
2/2✓ Branch 0 taken 33300 times.
✓ Branch 1 taken 11100 times.
|
44400 | for (i = 0; i < ctx->num_planes; i++) { |
| 357 |
4/4✓ Branch 0 taken 22200 times.
✓ Branch 1 taken 11100 times.
✓ Branch 2 taken 11100 times.
✓ Branch 3 taken 11100 times.
|
33300 | is_chroma = (i == 1 || i == 2); |
| 358 |
3/4✓ Branch 0 taken 22200 times.
✓ Branch 1 taken 11100 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 22200 times.
|
33300 | if (!is_chroma || ctx->chroma_factor == CFACTOR_Y444) { |
| 359 | 11100 | xp = x << 4; | |
| 360 | 11100 | yp = y << 4; | |
| 361 | 11100 | num_cblocks = 4; | |
| 362 | 11100 | pwidth = avctx->width; | |
| 363 | } else { | ||
| 364 | 22200 | xp = x << 3; | |
| 365 | 22200 | yp = y << 4; | |
| 366 | 22200 | num_cblocks = 2; | |
| 367 | 22200 | pwidth = avctx->width >> 1; | |
| 368 | } | ||
| 369 | |||
| 370 | 33300 | linesize = pic->linesize[i] * ctx->pictures_per_frame; | |
| 371 | 33300 | src = (const uint16_t*)(pic->data[i] + yp * linesize + | |
| 372 | 33300 | line_add * pic->linesize[i]) + xp; | |
| 373 | |||
| 374 |
1/2✓ Branch 0 taken 33300 times.
✗ Branch 1 not taken.
|
33300 | if (i < 3) { |
| 375 | 33300 | get_slice_data(ctx, src, linesize, xp, yp, | |
| 376 | pwidth, picture_height, | ||
| 377 | 33300 | ctx->blocks[0], ctx->emu_buf, | |
| 378 | mbs_per_slice, num_cblocks, is_chroma); | ||
| 379 |
2/2✓ Branch 0 taken 11100 times.
✓ Branch 1 taken 22200 times.
|
33300 | if (!is_chroma) {/* luma quant */ |
| 380 | 11100 | encode_slice_plane(ctx, pb, src, linesize, | |
| 381 | 11100 | mbs_per_slice, ctx->blocks[0], | |
| 382 | num_cblocks, qmat); | ||
| 383 | } else { /* chroma plane */ | ||
| 384 | 22200 | encode_slice_plane(ctx, pb, src, linesize, | |
| 385 | 22200 | mbs_per_slice, ctx->blocks[0], | |
| 386 | num_cblocks, qmat_chroma); | ||
| 387 | } | ||
| 388 | } else { | ||
| 389 | ✗ | get_alpha_data(ctx, src, linesize, xp, yp, | |
| 390 | pwidth, picture_height, | ||
| 391 | ✗ | ctx->blocks[0], mbs_per_slice, ctx->alpha_bits); | |
| 392 | ✗ | encode_alpha_plane(ctx, pb, mbs_per_slice, ctx->blocks[0], quant); | |
| 393 | } | ||
| 394 | 33300 | flush_put_bits(pb); | |
| 395 | 33300 | sizes[i] = put_bytes_output(pb) - total_size; | |
| 396 | 33300 | total_size = put_bytes_output(pb); | |
| 397 | } | ||
| 398 | 11100 | return total_size; | |
| 399 | } | ||
| 400 | |||
| 401 | 390320122 | static inline int estimate_vlc(unsigned codebook, int val) | |
| 402 | { | ||
| 403 | unsigned int rice_order, exp_order, switch_bits, switch_val; | ||
| 404 | int exponent; | ||
| 405 | |||
| 406 | /* number of prefix bits to switch between Rice and expGolomb */ | ||
| 407 | 390320122 | switch_bits = (codebook & 3) + 1; | |
| 408 | 390320122 | rice_order = codebook >> 5; /* rice code order */ | |
| 409 | 390320122 | exp_order = (codebook >> 2) & 7; /* exp golomb code order */ | |
| 410 | |||
| 411 | 390320122 | switch_val = switch_bits << rice_order; | |
| 412 | |||
| 413 |
2/2✓ Branch 0 taken 135366530 times.
✓ Branch 1 taken 254953592 times.
|
390320122 | if (val >= switch_val) { |
| 414 | 135366530 | val -= switch_val - (1 << exp_order); | |
| 415 | 135366530 | exponent = av_log2(val); | |
| 416 | |||
| 417 | 135366530 | return exponent * 2 - exp_order + switch_bits + 1; | |
| 418 | } else { | ||
| 419 | 254953592 | return (val >> rice_order) + rice_order + 1; | |
| 420 | } | ||
| 421 | } | ||
| 422 | |||
| 423 | 379686 | static int estimate_dcs(int *error, int16_t *blocks, int blocks_per_slice, | |
| 424 | int scale) | ||
| 425 | { | ||
| 426 | int i; | ||
| 427 | 379686 | int codebook = 5, code, dc, prev_dc, delta, sign, new_sign; | |
| 428 | int bits; | ||
| 429 | |||
| 430 | 379686 | prev_dc = (blocks[0] - 0x4000) / scale; | |
| 431 | 379686 | bits = estimate_vlc(FIRST_DC_CB, MAKE_CODE(prev_dc)); | |
| 432 | 379686 | sign = 0; | |
| 433 | 379686 | blocks += 64; | |
| 434 |
2/2✓ Branch 0 taken 199846 times.
✓ Branch 1 taken 179840 times.
|
379686 | *error += FFABS(blocks[0] - 0x4000) % scale; |
| 435 | |||
| 436 |
2/2✓ Branch 0 taken 5069986 times.
✓ Branch 1 taken 379686 times.
|
5449672 | for (i = 1; i < blocks_per_slice; i++, blocks += 64) { |
| 437 | 5069986 | dc = (blocks[0] - 0x4000) / scale; | |
| 438 |
2/2✓ Branch 0 taken 2618934 times.
✓ Branch 1 taken 2451052 times.
|
5069986 | *error += FFABS(blocks[0] - 0x4000) % scale; |
| 439 | 5069986 | delta = dc - prev_dc; | |
| 440 | 5069986 | new_sign = GET_SIGN(delta); | |
| 441 | 5069986 | delta = (delta ^ sign) - sign; | |
| 442 | 5069986 | code = MAKE_CODE(delta); | |
| 443 | 5069986 | bits += estimate_vlc(ff_prores_dc_codebook[codebook], code); | |
| 444 | 5069986 | codebook = FFMIN(code, 6); | |
| 445 | 5069986 | sign = new_sign; | |
| 446 | 5069986 | prev_dc = dc; | |
| 447 | } | ||
| 448 | |||
| 449 | 379686 | return bits; | |
| 450 | } | ||
| 451 | |||
| 452 | 379686 | static int estimate_acs(int *error, int16_t *blocks, int blocks_per_slice, | |
| 453 | const uint8_t *scan, const int16_t *qmat) | ||
| 454 | { | ||
| 455 | int idx, i; | ||
| 456 | 379686 | int prev_run = 4; | |
| 457 | 379686 | int prev_level = 2; | |
| 458 | int run, level; | ||
| 459 | int max_coeffs, abs_level; | ||
| 460 | 379686 | int bits = 0; | |
| 461 | |||
| 462 | 379686 | max_coeffs = blocks_per_slice << 6; | |
| 463 | 379686 | run = 0; | |
| 464 | |||
| 465 |
2/2✓ Branch 0 taken 23920218 times.
✓ Branch 1 taken 379686 times.
|
24299904 | for (i = 1; i < 64; i++) { |
| 466 |
2/2✓ Branch 0 taken 343329336 times.
✓ Branch 1 taken 23920218 times.
|
367249554 | for (idx = scan[i]; idx < max_coeffs; idx += 64) { |
| 467 | 343329336 | level = blocks[idx] / qmat[scan[i]]; | |
| 468 | 343329336 | *error += FFABS(blocks[idx]) % qmat[scan[i]]; | |
| 469 |
2/2✓ Branch 0 taken 192435225 times.
✓ Branch 1 taken 150894111 times.
|
343329336 | if (level) { |
| 470 | 192435225 | abs_level = FFABS(level); | |
| 471 | 192435225 | bits += estimate_vlc(ff_prores_run_to_cb[prev_run], run); | |
| 472 | 192435225 | bits += estimate_vlc(ff_prores_level_to_cb[prev_level], | |
| 473 | 192435225 | abs_level - 1) + 1; | |
| 474 | 192435225 | prev_run = FFMIN(run, 15); | |
| 475 | 192435225 | prev_level = FFMIN(abs_level, 9); | |
| 476 | 192435225 | run = 0; | |
| 477 | } else { | ||
| 478 | 150894111 | run++; | |
| 479 | } | ||
| 480 | } | ||
| 481 | } | ||
| 482 | |||
| 483 | 379686 | return bits; | |
| 484 | } | ||
| 485 | |||
| 486 | 379686 | static int estimate_slice_plane(ProresContext *ctx, int *error, int plane, | |
| 487 | const uint16_t *src, ptrdiff_t linesize, | ||
| 488 | int mbs_per_slice, | ||
| 489 | int blocks_per_mb, | ||
| 490 | const int16_t *qmat, ProresThreadData *td) | ||
| 491 | { | ||
| 492 | int blocks_per_slice; | ||
| 493 | int bits; | ||
| 494 | |||
| 495 | 379686 | blocks_per_slice = mbs_per_slice * blocks_per_mb; | |
| 496 | |||
| 497 | 379686 | bits = estimate_dcs(error, td->blocks[plane], blocks_per_slice, qmat[0]); | |
| 498 | 379686 | bits += estimate_acs(error, td->blocks[plane], blocks_per_slice, ctx->scantable, qmat); | |
| 499 | |||
| 500 | 379686 | return FFALIGN(bits, 8); | |
| 501 | } | ||
| 502 | |||
| 503 | ✗ | static int est_alpha_diff(int cur, int prev, int abits) | |
| 504 | { | ||
| 505 | ✗ | const int dbits = (abits == 8) ? 4 : 7; | |
| 506 | ✗ | const int dsize = 1 << dbits - 1; | |
| 507 | ✗ | int diff = cur - prev; | |
| 508 | |||
| 509 | ✗ | diff = av_zero_extend(diff, abits); | |
| 510 | ✗ | if (diff >= (1 << abits) - dsize) | |
| 511 | ✗ | diff -= 1 << abits; | |
| 512 | ✗ | if (diff < -dsize || diff > dsize || !diff) | |
| 513 | ✗ | return abits + 1; | |
| 514 | else | ||
| 515 | ✗ | return dbits + 1; | |
| 516 | } | ||
| 517 | |||
| 518 | ✗ | static int estimate_alpha_plane(ProresContext *ctx, | |
| 519 | const uint16_t *src, ptrdiff_t linesize, | ||
| 520 | int mbs_per_slice, int16_t *blocks) | ||
| 521 | { | ||
| 522 | ✗ | const int abits = ctx->alpha_bits; | |
| 523 | ✗ | const int mask = (1 << abits) - 1; | |
| 524 | ✗ | const int num_coeffs = mbs_per_slice * 256; | |
| 525 | ✗ | int prev = mask, cur; | |
| 526 | ✗ | int idx = 0; | |
| 527 | ✗ | int run = 0; | |
| 528 | int bits; | ||
| 529 | |||
| 530 | ✗ | cur = blocks[idx++]; | |
| 531 | ✗ | bits = est_alpha_diff(cur, prev, abits); | |
| 532 | ✗ | prev = cur; | |
| 533 | do { | ||
| 534 | ✗ | cur = blocks[idx++]; | |
| 535 | ✗ | if (cur != prev) { | |
| 536 | ✗ | if (!run) | |
| 537 | ✗ | bits++; | |
| 538 | ✗ | else if (run < 0x10) | |
| 539 | ✗ | bits += 4; | |
| 540 | else | ||
| 541 | ✗ | bits += 15; | |
| 542 | ✗ | bits += est_alpha_diff(cur, prev, abits); | |
| 543 | ✗ | prev = cur; | |
| 544 | ✗ | run = 0; | |
| 545 | } else { | ||
| 546 | ✗ | run++; | |
| 547 | } | ||
| 548 | ✗ | } while (idx < num_coeffs); | |
| 549 | |||
| 550 | ✗ | if (run) { | |
| 551 | ✗ | if (run < 0x10) | |
| 552 | ✗ | bits += 4; | |
| 553 | else | ||
| 554 | ✗ | bits += 15; | |
| 555 | } | ||
| 556 | |||
| 557 | ✗ | return bits; | |
| 558 | } | ||
| 559 | |||
| 560 | 11100 | static int find_slice_quant(AVCodecContext *avctx, | |
| 561 | int trellis_node, int x, int y, int mbs_per_slice, | ||
| 562 | ProresThreadData *td) | ||
| 563 | { | ||
| 564 | 11100 | ProresContext *ctx = avctx->priv_data; | |
| 565 | int i, q, pq, xp, yp; | ||
| 566 | const uint16_t *src; | ||
| 567 | int num_cblocks[MAX_PLANES], pwidth; | ||
| 568 | int is_chroma[MAX_PLANES]; | ||
| 569 | 11100 | const int min_quant = ctx->profile_info->min_quant; | |
| 570 | 11100 | const int max_quant = ctx->profile_info->max_quant; | |
| 571 | int error, bits, bits_limit; | ||
| 572 | int mbs, prev, cur, new_score; | ||
| 573 | int slice_bits[TRELLIS_WIDTH], slice_score[TRELLIS_WIDTH]; | ||
| 574 | int overquant; | ||
| 575 | uint16_t *qmat; | ||
| 576 | uint16_t *qmat_chroma; | ||
| 577 | int linesize[4], line_add, picture_height; | ||
| 578 | 11100 | int alpha_bits = 0; | |
| 579 | |||
| 580 |
1/2✓ Branch 0 taken 11100 times.
✗ Branch 1 not taken.
|
11100 | if (ctx->pictures_per_frame == 1) |
| 581 | 11100 | line_add = 0; | |
| 582 | else | ||
| 583 | ✗ | line_add = ctx->cur_picture_idx ^ !(ctx->pic->flags & AV_FRAME_FLAG_TOP_FIELD_FIRST); | |
| 584 | |||
| 585 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 11100 times.
|
11100 | if ((y << 4) * ctx->pictures_per_frame + line_add >= avctx->height) |
| 586 | ✗ | line_add = 0; | |
| 587 | 11100 | picture_height = (avctx->height - line_add + ctx->pictures_per_frame - 1) / ctx->pictures_per_frame; | |
| 588 | 11100 | mbs = x + mbs_per_slice; | |
| 589 | |||
| 590 |
2/2✓ Branch 0 taken 33300 times.
✓ Branch 1 taken 11100 times.
|
44400 | for (i = 0; i < ctx->num_planes; i++) { |
| 591 |
4/4✓ Branch 0 taken 22200 times.
✓ Branch 1 taken 11100 times.
✓ Branch 2 taken 11100 times.
✓ Branch 3 taken 11100 times.
|
33300 | is_chroma[i] = (i == 1 || i == 2); |
| 592 |
3/4✓ Branch 0 taken 22200 times.
✓ Branch 1 taken 11100 times.
✗ Branch 2 not taken.
✓ Branch 3 taken 22200 times.
|
33300 | if (!is_chroma[i] || ctx->chroma_factor == CFACTOR_Y444) { |
| 593 | 11100 | xp = x << 4; | |
| 594 | 11100 | yp = y << 4; | |
| 595 | 11100 | num_cblocks[i] = 4; | |
| 596 | 11100 | pwidth = avctx->width; | |
| 597 | } else { | ||
| 598 | 22200 | xp = x << 3; | |
| 599 | 22200 | yp = y << 4; | |
| 600 | 22200 | num_cblocks[i] = 2; | |
| 601 | 22200 | pwidth = avctx->width >> 1; | |
| 602 | } | ||
| 603 | |||
| 604 | 33300 | linesize[i] = ctx->pic->linesize[i] * ctx->pictures_per_frame; | |
| 605 | 33300 | src = (const uint16_t *)(ctx->pic->data[i] + yp * linesize[i] + | |
| 606 | 33300 | line_add * ctx->pic->linesize[i]) + xp; | |
| 607 | |||
| 608 |
1/2✓ Branch 0 taken 33300 times.
✗ Branch 1 not taken.
|
33300 | if (i < 3) { |
| 609 | 33300 | get_slice_data(ctx, src, linesize[i], xp, yp, | |
| 610 | pwidth, picture_height, | ||
| 611 | 33300 | td->blocks[i], td->emu_buf, | |
| 612 | mbs_per_slice, num_cblocks[i], is_chroma[i]); | ||
| 613 | } else { | ||
| 614 | ✗ | get_alpha_data(ctx, src, linesize[i], xp, yp, | |
| 615 | pwidth, picture_height, | ||
| 616 | ✗ | td->blocks[i], mbs_per_slice, ctx->alpha_bits); | |
| 617 | } | ||
| 618 | } | ||
| 619 | |||
| 620 |
2/2✓ Branch 0 taken 77700 times.
✓ Branch 1 taken 11100 times.
|
88800 | for (q = min_quant; q < max_quant + 2; q++) { |
| 621 | 77700 | td->nodes[trellis_node + q].prev_node = -1; | |
| 622 | 77700 | td->nodes[trellis_node + q].quant = q; | |
| 623 | } | ||
| 624 | |||
| 625 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 11100 times.
|
11100 | if (ctx->alpha_bits) |
| 626 | ✗ | alpha_bits = estimate_alpha_plane(ctx, src, linesize[3], | |
| 627 | ✗ | mbs_per_slice, td->blocks[3]); | |
| 628 | // todo: maybe perform coarser quantising to fit into frame size when needed | ||
| 629 |
2/2✓ Branch 0 taken 66600 times.
✓ Branch 1 taken 11100 times.
|
77700 | for (q = min_quant; q <= max_quant; q++) { |
| 630 | 66600 | bits = alpha_bits; | |
| 631 | 66600 | error = 0; | |
| 632 | 133200 | bits += estimate_slice_plane(ctx, &error, 0, | |
| 633 | 66600 | src, linesize[0], | |
| 634 | mbs_per_slice, | ||
| 635 | num_cblocks[0], | ||
| 636 | 66600 | ctx->quants[q], td); /* estimate luma plane */ | |
| 637 |
2/2✓ Branch 0 taken 133200 times.
✓ Branch 1 taken 66600 times.
|
199800 | for (i = 1; i < ctx->num_planes - !!ctx->alpha_bits; i++) { /* estimate chroma plane */ |
| 638 | 133200 | bits += estimate_slice_plane(ctx, &error, i, | |
| 639 | 133200 | src, linesize[i], | |
| 640 | mbs_per_slice, | ||
| 641 | num_cblocks[i], | ||
| 642 | 133200 | ctx->quants_chroma[q], td); | |
| 643 | } | ||
| 644 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 66600 times.
|
66600 | if (bits > 65000 * 8) |
| 645 | ✗ | error = SCORE_LIMIT; | |
| 646 | |||
| 647 | 66600 | slice_bits[q] = bits; | |
| 648 | 66600 | slice_score[q] = error; | |
| 649 | } | ||
| 650 |
2/2✓ Branch 0 taken 3643 times.
✓ Branch 1 taken 7457 times.
|
11100 | if (slice_bits[max_quant] <= ctx->bits_per_mb * mbs_per_slice) { |
| 651 | 3643 | slice_bits[max_quant + 1] = slice_bits[max_quant]; | |
| 652 | 3643 | slice_score[max_quant + 1] = slice_score[max_quant] + 1; | |
| 653 | 3643 | overquant = max_quant; | |
| 654 | } else { | ||
| 655 |
1/2✓ Branch 0 taken 59962 times.
✗ Branch 1 not taken.
|
59962 | for (q = max_quant + 1; q < 128; q++) { |
| 656 | 59962 | bits = alpha_bits; | |
| 657 | 59962 | error = 0; | |
| 658 |
2/2✓ Branch 0 taken 44939 times.
✓ Branch 1 taken 15023 times.
|
59962 | if (q < MAX_STORED_Q) { |
| 659 | 44939 | qmat = ctx->quants[q]; | |
| 660 | 44939 | qmat_chroma = ctx->quants_chroma[q]; | |
| 661 | } else { | ||
| 662 | 15023 | qmat = td->custom_q; | |
| 663 | 15023 | qmat_chroma = td->custom_chroma_q; | |
| 664 |
2/2✓ Branch 0 taken 961472 times.
✓ Branch 1 taken 15023 times.
|
976495 | for (i = 0; i < 64; i++) { |
| 665 | 961472 | qmat[i] = ctx->quant_mat[i] * q; | |
| 666 | 961472 | qmat_chroma[i] = ctx->quant_chroma_mat[i] * q; | |
| 667 | } | ||
| 668 | } | ||
| 669 | 119924 | bits += estimate_slice_plane(ctx, &error, 0, | |
| 670 | 59962 | src, linesize[0], | |
| 671 | mbs_per_slice, | ||
| 672 | num_cblocks[0], | ||
| 673 | qmat, td);/* estimate luma plane */ | ||
| 674 |
2/2✓ Branch 0 taken 119924 times.
✓ Branch 1 taken 59962 times.
|
179886 | for (i = 1; i < ctx->num_planes - !!ctx->alpha_bits; i++) { /* estimate chroma plane */ |
| 675 | 119924 | bits += estimate_slice_plane(ctx, &error, i, | |
| 676 | 119924 | src, linesize[i], | |
| 677 | mbs_per_slice, | ||
| 678 | num_cblocks[i], | ||
| 679 | qmat_chroma, td); | ||
| 680 | } | ||
| 681 |
2/2✓ Branch 0 taken 7457 times.
✓ Branch 1 taken 52505 times.
|
59962 | if (bits <= ctx->bits_per_mb * mbs_per_slice) |
| 682 | 7457 | break; | |
| 683 | } | ||
| 684 | |||
| 685 | 7457 | slice_bits[max_quant + 1] = bits; | |
| 686 | 7457 | slice_score[max_quant + 1] = error; | |
| 687 | 7457 | overquant = q; | |
| 688 | } | ||
| 689 | 11100 | td->nodes[trellis_node + max_quant + 1].quant = overquant; | |
| 690 | |||
| 691 | 11100 | bits_limit = mbs * ctx->bits_per_mb; | |
| 692 |
2/2✓ Branch 0 taken 77700 times.
✓ Branch 1 taken 11100 times.
|
88800 | for (pq = min_quant; pq < max_quant + 2; pq++) { |
| 693 | 77700 | prev = trellis_node - TRELLIS_WIDTH + pq; | |
| 694 | |||
| 695 |
2/2✓ Branch 0 taken 543900 times.
✓ Branch 1 taken 77700 times.
|
621600 | for (q = min_quant; q < max_quant + 2; q++) { |
| 696 | 543900 | cur = trellis_node + q; | |
| 697 | 543900 | bits = td->nodes[prev].bits + slice_bits[q]; | |
| 698 | 543900 | error = slice_score[q]; | |
| 699 |
2/2✓ Branch 0 taken 438632 times.
✓ Branch 1 taken 105268 times.
|
543900 | if (bits > bits_limit) |
| 700 | 438632 | error = SCORE_LIMIT; | |
| 701 | |||
| 702 |
4/4✓ Branch 0 taken 268352 times.
✓ Branch 1 taken 275548 times.
✓ Branch 2 taken 98480 times.
✓ Branch 3 taken 169872 times.
|
543900 | if (td->nodes[prev].score < SCORE_LIMIT && error < SCORE_LIMIT) |
| 703 | 98480 | new_score = td->nodes[prev].score + error; | |
| 704 | else | ||
| 705 | 445420 | new_score = SCORE_LIMIT; | |
| 706 |
2/2✓ Branch 0 taken 466200 times.
✓ Branch 1 taken 77700 times.
|
543900 | if (td->nodes[cur].prev_node == -1 || |
| 707 |
2/2✓ Branch 0 taken 427864 times.
✓ Branch 1 taken 38336 times.
|
466200 | td->nodes[cur].score >= new_score) { |
| 708 | |||
| 709 | 505564 | td->nodes[cur].bits = bits; | |
| 710 | 505564 | td->nodes[cur].score = new_score; | |
| 711 | 505564 | td->nodes[cur].prev_node = prev; | |
| 712 | } | ||
| 713 | } | ||
| 714 | } | ||
| 715 | |||
| 716 | 11100 | error = td->nodes[trellis_node + min_quant].score; | |
| 717 | 11100 | pq = trellis_node + min_quant; | |
| 718 |
2/2✓ Branch 0 taken 66600 times.
✓ Branch 1 taken 11100 times.
|
77700 | for (q = min_quant + 1; q < max_quant + 2; q++) { |
| 719 |
2/2✓ Branch 0 taken 52033 times.
✓ Branch 1 taken 14567 times.
|
66600 | if (td->nodes[trellis_node + q].score <= error) { |
| 720 | 52033 | error = td->nodes[trellis_node + q].score; | |
| 721 | 52033 | pq = trellis_node + q; | |
| 722 | } | ||
| 723 | } | ||
| 724 | |||
| 725 | 11100 | return pq; | |
| 726 | } | ||
| 727 | |||
| 728 | 2850 | static int find_quant_thread(AVCodecContext *avctx, void *arg, | |
| 729 | int jobnr, int threadnr) | ||
| 730 | { | ||
| 731 | 2850 | ProresContext *ctx = avctx->priv_data; | |
| 732 | 2850 | ProresThreadData *td = ctx->tdata + threadnr; | |
| 733 | 2850 | int mbs_per_slice = ctx->mbs_per_slice; | |
| 734 | 2850 | int x, y = jobnr, mb, q = 0; | |
| 735 | |||
| 736 |
2/2✓ Branch 0 taken 11100 times.
✓ Branch 1 taken 2850 times.
|
13950 | for (x = mb = 0; x < ctx->mb_width; x += mbs_per_slice, mb++) { |
| 737 |
2/2✓ Branch 0 taken 5850 times.
✓ Branch 1 taken 11100 times.
|
16950 | while (ctx->mb_width - x < mbs_per_slice) |
| 738 | 5850 | mbs_per_slice >>= 1; | |
| 739 | 11100 | q = find_slice_quant(avctx, | |
| 740 | 11100 | (mb + 1) * TRELLIS_WIDTH, x, y, | |
| 741 | mbs_per_slice, td); | ||
| 742 | } | ||
| 743 | |||
| 744 |
2/2✓ Branch 0 taken 11100 times.
✓ Branch 1 taken 2850 times.
|
13950 | for (x = ctx->slices_width - 1; x >= 0; x--) { |
| 745 | 11100 | ctx->slice_q[x + y * ctx->slices_width] = td->nodes[q].quant; | |
| 746 | 11100 | q = td->nodes[q].prev_node; | |
| 747 | } | ||
| 748 | |||
| 749 | 2850 | return 0; | |
| 750 | } | ||
| 751 | |||
| 752 | 200 | static int encode_frame(AVCodecContext *avctx, AVPacket *pkt, | |
| 753 | const AVFrame *pic, int *got_packet) | ||
| 754 | { | ||
| 755 | 200 | ProresContext *ctx = avctx->priv_data; | |
| 756 | uint8_t *orig_buf, *buf, *slice_hdr, *slice_sizes, *tmp; | ||
| 757 | uint8_t *picture_size_pos; | ||
| 758 | PutBitContext pb; | ||
| 759 | 200 | int x, y, i, mb, q = 0; | |
| 760 | 200 | int sizes[4] = { 0 }; | |
| 761 | 200 | int slice_hdr_size = 2 * ctx->num_planes; | |
| 762 | int frame_size, picture_size, slice_size; | ||
| 763 | int pkt_size, ret; | ||
| 764 | 200 | int max_slice_size = (ctx->frame_size_upper_bound - 200) / (ctx->pictures_per_frame * ctx->slices_per_picture + 1); | |
| 765 | uint8_t frame_flags; | ||
| 766 | |||
| 767 | 200 | ctx->pic = pic; | |
| 768 | 200 | pkt_size = ctx->frame_size_upper_bound; | |
| 769 | |||
| 770 |
1/2✗ Branch 1 not taken.
✓ Branch 2 taken 200 times.
|
200 | if ((ret = ff_alloc_packet(avctx, pkt, pkt_size + FF_INPUT_BUFFER_MIN_SIZE)) < 0) |
| 771 | ✗ | return ret; | |
| 772 | |||
| 773 | 200 | orig_buf = pkt->data; | |
| 774 | |||
| 775 | // frame atom | ||
| 776 | 200 | orig_buf += 4; // frame size | |
| 777 | 200 | bytestream_put_be32 (&orig_buf, FRAME_ID); // frame container ID | |
| 778 | 200 | buf = orig_buf; | |
| 779 | |||
| 780 | // frame header | ||
| 781 | 200 | tmp = buf; | |
| 782 | 200 | buf += 2; // frame header size will be stored here | |
| 783 |
2/4✓ Branch 0 taken 200 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 200 times.
|
200 | bytestream_put_be16 (&buf, ctx->chroma_factor != CFACTOR_Y422 || ctx->alpha_bits ? 1 : 0); |
| 784 | 200 | bytestream_put_buffer(&buf, ctx->vendor, 4); | |
| 785 | 200 | bytestream_put_be16 (&buf, avctx->width); | |
| 786 | 200 | bytestream_put_be16 (&buf, avctx->height); | |
| 787 | |||
| 788 | 200 | frame_flags = ctx->chroma_factor << 6; | |
| 789 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 200 times.
|
200 | if (avctx->flags & AV_CODEC_FLAG_INTERLACED_DCT) |
| 790 | ✗ | frame_flags |= (pic->flags & AV_FRAME_FLAG_TOP_FIELD_FIRST) ? 0x04 : 0x08; | |
| 791 | 200 | bytestream_put_byte (&buf, frame_flags); | |
| 792 | |||
| 793 | 200 | bytestream_put_byte (&buf, 0); // reserved | |
| 794 | 200 | bytestream_put_byte (&buf, pic->color_primaries); | |
| 795 | 200 | bytestream_put_byte (&buf, pic->color_trc); | |
| 796 | 200 | bytestream_put_byte (&buf, pic->colorspace); | |
| 797 | 200 | bytestream_put_byte (&buf, ctx->alpha_bits >> 3); | |
| 798 | 200 | bytestream_put_byte (&buf, 0); // reserved | |
| 799 |
1/2✓ Branch 0 taken 200 times.
✗ Branch 1 not taken.
|
200 | if (ctx->quant_sel != QUANT_MAT_DEFAULT) { |
| 800 | 200 | bytestream_put_byte (&buf, 0x03); // matrix flags - both matrices are present | |
| 801 | 200 | bytestream_put_buffer(&buf, ctx->quant_mat, 64); // luma quantisation matrix | |
| 802 | 200 | bytestream_put_buffer(&buf, ctx->quant_chroma_mat, 64); // chroma quantisation matrix | |
| 803 | } else { | ||
| 804 | ✗ | bytestream_put_byte (&buf, 0x00); // matrix flags - default matrices are used | |
| 805 | } | ||
| 806 | 200 | bytestream_put_be16 (&tmp, buf - orig_buf); // write back frame header size | |
| 807 | |||
| 808 | 200 | for (ctx->cur_picture_idx = 0; | |
| 809 |
2/2✓ Branch 0 taken 200 times.
✓ Branch 1 taken 200 times.
|
400 | ctx->cur_picture_idx < ctx->pictures_per_frame; |
| 810 | 200 | ctx->cur_picture_idx++) { | |
| 811 | // picture header | ||
| 812 | 200 | picture_size_pos = buf + 1; | |
| 813 | 200 | bytestream_put_byte (&buf, 0x40); // picture header size (in bits) | |
| 814 | 200 | buf += 4; // picture data size will be stored here | |
| 815 | 200 | bytestream_put_be16 (&buf, ctx->slices_per_picture); | |
| 816 | 200 | bytestream_put_byte (&buf, av_log2(ctx->mbs_per_slice) << 4); // slice width and height in MBs | |
| 817 | |||
| 818 | // seek table - will be filled during slice encoding | ||
| 819 | 200 | slice_sizes = buf; | |
| 820 | 200 | buf += ctx->slices_per_picture * 2; | |
| 821 | |||
| 822 | // slices | ||
| 823 |
1/2✓ Branch 0 taken 200 times.
✗ Branch 1 not taken.
|
200 | if (!ctx->force_quant) { |
| 824 | 200 | ret = avctx->execute2(avctx, find_quant_thread, NULL, NULL, | |
| 825 | ctx->mb_height); | ||
| 826 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 200 times.
|
200 | if (ret) |
| 827 | ✗ | return ret; | |
| 828 | } | ||
| 829 | |||
| 830 |
2/2✓ Branch 0 taken 2850 times.
✓ Branch 1 taken 200 times.
|
3050 | for (y = 0; y < ctx->mb_height; y++) { |
| 831 | 2850 | int mbs_per_slice = ctx->mbs_per_slice; | |
| 832 |
2/2✓ Branch 0 taken 11100 times.
✓ Branch 1 taken 2850 times.
|
13950 | for (x = mb = 0; x < ctx->mb_width; x += mbs_per_slice, mb++) { |
| 833 | 22200 | q = ctx->force_quant ? ctx->force_quant | |
| 834 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 11100 times.
|
11100 | : ctx->slice_q[mb + y * ctx->slices_width]; |
| 835 | |||
| 836 |
2/2✓ Branch 0 taken 5850 times.
✓ Branch 1 taken 11100 times.
|
16950 | while (ctx->mb_width - x < mbs_per_slice) |
| 837 | 5850 | mbs_per_slice >>= 1; | |
| 838 | |||
| 839 | 11100 | bytestream_put_byte(&buf, slice_hdr_size << 3); | |
| 840 | 11100 | slice_hdr = buf; | |
| 841 | 11100 | buf += slice_hdr_size - 1; | |
| 842 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 11100 times.
|
11100 | if (pkt_size <= buf - orig_buf + 2 * max_slice_size) { |
| 843 | ✗ | uint8_t *start = pkt->data; | |
| 844 | // Recompute new size according to max_slice_size | ||
| 845 | // and deduce delta | ||
| 846 | ✗ | int delta = 200 + (ctx->pictures_per_frame * | |
| 847 | ✗ | ctx->slices_per_picture + 1) * | |
| 848 | max_slice_size - pkt_size; | ||
| 849 | |||
| 850 | ✗ | delta = FFMAX(delta, 2 * max_slice_size); | |
| 851 | ✗ | ctx->frame_size_upper_bound += delta; | |
| 852 | |||
| 853 | ✗ | if (!ctx->warn) { | |
| 854 | ✗ | avpriv_request_sample(avctx, | |
| 855 | "Packet too small: is %i," | ||
| 856 | " needs %i (slice: %i). " | ||
| 857 | "Correct allocation", | ||
| 858 | pkt_size, delta, max_slice_size); | ||
| 859 | ✗ | ctx->warn = 1; | |
| 860 | } | ||
| 861 | |||
| 862 | ✗ | ret = av_grow_packet(pkt, delta); | |
| 863 | ✗ | if (ret < 0) | |
| 864 | ✗ | return ret; | |
| 865 | |||
| 866 | ✗ | pkt_size += delta; | |
| 867 | ✗ | orig_buf = pkt->data + (orig_buf - start); | |
| 868 | ✗ | buf = pkt->data + (buf - start); | |
| 869 | ✗ | picture_size_pos = pkt->data + (picture_size_pos - start); | |
| 870 | ✗ | slice_sizes = pkt->data + (slice_sizes - start); | |
| 871 | ✗ | slice_hdr = pkt->data + (slice_hdr - start); | |
| 872 | ✗ | tmp = pkt->data + (tmp - start); | |
| 873 | } | ||
| 874 | 11100 | init_put_bits(&pb, buf, (pkt_size - (buf - orig_buf))); | |
| 875 | 11100 | ret = encode_slice(avctx, pic, &pb, sizes, x, y, q, | |
| 876 | mbs_per_slice); | ||
| 877 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 11100 times.
|
11100 | if (ret < 0) |
| 878 | ✗ | return ret; | |
| 879 | |||
| 880 | 11100 | bytestream_put_byte(&slice_hdr, q); | |
| 881 | 11100 | slice_size = slice_hdr_size + sizes[ctx->num_planes - 1]; | |
| 882 |
2/2✓ Branch 0 taken 22200 times.
✓ Branch 1 taken 11100 times.
|
33300 | for (i = 0; i < ctx->num_planes - 1; i++) { |
| 883 | 22200 | bytestream_put_be16(&slice_hdr, sizes[i]); | |
| 884 | 22200 | slice_size += sizes[i]; | |
| 885 | } | ||
| 886 | 11100 | bytestream_put_be16(&slice_sizes, slice_size); | |
| 887 | 11100 | buf += slice_size - slice_hdr_size; | |
| 888 |
2/2✓ Branch 0 taken 216 times.
✓ Branch 1 taken 10884 times.
|
11100 | if (max_slice_size < slice_size) |
| 889 | 216 | max_slice_size = slice_size; | |
| 890 | } | ||
| 891 | } | ||
| 892 | |||
| 893 | 200 | picture_size = buf - (picture_size_pos - 1); | |
| 894 | 200 | bytestream_put_be32(&picture_size_pos, picture_size); | |
| 895 | } | ||
| 896 | |||
| 897 | 200 | orig_buf -= 8; | |
| 898 | 200 | frame_size = buf - orig_buf; | |
| 899 | 200 | bytestream_put_be32(&orig_buf, frame_size); | |
| 900 | |||
| 901 | 200 | pkt->size = frame_size; | |
| 902 | 200 | *got_packet = 1; | |
| 903 | |||
| 904 | 200 | return 0; | |
| 905 | } | ||
| 906 | |||
| 907 | 4 | static av_cold int encode_close(AVCodecContext *avctx) | |
| 908 | { | ||
| 909 | 4 | ProresContext *ctx = avctx->priv_data; | |
| 910 | int i; | ||
| 911 | |||
| 912 |
1/2✓ Branch 0 taken 4 times.
✗ Branch 1 not taken.
|
4 | if (ctx->tdata) { |
| 913 |
2/2✓ Branch 0 taken 4 times.
✓ Branch 1 taken 4 times.
|
8 | for (i = 0; i < avctx->thread_count; i++) |
| 914 | 4 | av_freep(&ctx->tdata[i].nodes); | |
| 915 | } | ||
| 916 | 4 | av_freep(&ctx->tdata); | |
| 917 | 4 | av_freep(&ctx->slice_q); | |
| 918 | |||
| 919 | 4 | return 0; | |
| 920 | } | ||
| 921 | |||
| 922 | 957600 | static void prores_fdct(FDCTDSPContext *fdsp, const uint16_t *src, | |
| 923 | ptrdiff_t linesize, int16_t *block) | ||
| 924 | { | ||
| 925 | int x, y; | ||
| 926 | 957600 | const uint16_t *tsrc = src; | |
| 927 | |||
| 928 |
2/2✓ Branch 0 taken 7660800 times.
✓ Branch 1 taken 957600 times.
|
8618400 | for (y = 0; y < 8; y++) { |
| 929 |
2/2✓ Branch 0 taken 61286400 times.
✓ Branch 1 taken 7660800 times.
|
68947200 | for (x = 0; x < 8; x++) |
| 930 | 61286400 | block[y * 8 + x] = tsrc[x]; | |
| 931 | 7660800 | tsrc += linesize >> 1; | |
| 932 | } | ||
| 933 | 957600 | fdsp->fdct(block); | |
| 934 | 957600 | } | |
| 935 | |||
| 936 | 4 | static av_cold int encode_init(AVCodecContext *avctx) | |
| 937 | { | ||
| 938 | 4 | ProresContext *ctx = avctx->priv_data; | |
| 939 | 4 | int err = 0, i, j, min_quant, max_quant; | |
| 940 | |||
| 941 | 4 | err = ff_prores_kostya_encode_init(avctx, ctx, avctx->pix_fmt); | |
| 942 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 4 times.
|
4 | if (err < 0) |
| 943 | ✗ | return err; | |
| 944 | |||
| 945 | 4 | ctx->fdct = prores_fdct; | |
| 946 | 4 | ff_fdctdsp_init(&ctx->fdsp, avctx); | |
| 947 | |||
| 948 |
1/2✓ Branch 0 taken 4 times.
✗ Branch 1 not taken.
|
4 | if (!ctx->force_quant) { |
| 949 | 4 | min_quant = ctx->profile_info->min_quant; | |
| 950 | 4 | max_quant = ctx->profile_info->max_quant; | |
| 951 | |||
| 952 | 4 | ctx->slice_q = av_malloc_array(ctx->slices_per_picture, sizeof(*ctx->slice_q)); | |
| 953 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 4 times.
|
4 | if (!ctx->slice_q) |
| 954 | ✗ | return AVERROR(ENOMEM); | |
| 955 | |||
| 956 | 4 | ctx->tdata = av_calloc(avctx->thread_count, sizeof(*ctx->tdata)); | |
| 957 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 4 times.
|
4 | if (!ctx->tdata) |
| 958 | ✗ | return AVERROR(ENOMEM); | |
| 959 | |||
| 960 |
2/2✓ Branch 0 taken 4 times.
✓ Branch 1 taken 4 times.
|
8 | for (j = 0; j < avctx->thread_count; j++) { |
| 961 | 4 | ctx->tdata[j].nodes = av_malloc_array(ctx->slices_width + 1, | |
| 962 | TRELLIS_WIDTH | ||
| 963 | * sizeof(*ctx->tdata->nodes)); | ||
| 964 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 4 times.
|
4 | if (!ctx->tdata[j].nodes) |
| 965 | ✗ | return AVERROR(ENOMEM); | |
| 966 |
2/2✓ Branch 0 taken 28 times.
✓ Branch 1 taken 4 times.
|
32 | for (i = min_quant; i < max_quant + 2; i++) { |
| 967 | 28 | ctx->tdata[j].nodes[i].prev_node = -1; | |
| 968 | 28 | ctx->tdata[j].nodes[i].bits = 0; | |
| 969 | 28 | ctx->tdata[j].nodes[i].score = 0; | |
| 970 | } | ||
| 971 | } | ||
| 972 | } | ||
| 973 | |||
| 974 | 4 | return 0; | |
| 975 | } | ||
| 976 | |||
| 977 | #define OFFSET(x) offsetof(ProresContext, x) | ||
| 978 | #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM | ||
| 979 | |||
| 980 | static const AVOption options[] = { | ||
| 981 | { "mbs_per_slice", "macroblocks per slice", OFFSET(mbs_per_slice), | ||
| 982 | AV_OPT_TYPE_INT, { .i64 = 8 }, 1, MAX_MBS_PER_SLICE, VE }, | ||
| 983 | { "profile", NULL, OFFSET(profile), AV_OPT_TYPE_INT, | ||
| 984 | { .i64 = PRORES_PROFILE_AUTO }, | ||
| 985 | PRORES_PROFILE_AUTO, PRORES_PROFILE_4444XQ, VE, .unit = "profile" }, | ||
| 986 | { "auto", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = PRORES_PROFILE_AUTO }, | ||
| 987 | 0, 0, VE, .unit = "profile" }, | ||
| 988 | { "proxy", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = PRORES_PROFILE_PROXY }, | ||
| 989 | 0, 0, VE, .unit = "profile" }, | ||
| 990 | { "lt", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = PRORES_PROFILE_LT }, | ||
| 991 | 0, 0, VE, .unit = "profile" }, | ||
| 992 | { "standard", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = PRORES_PROFILE_STANDARD }, | ||
| 993 | 0, 0, VE, .unit = "profile" }, | ||
| 994 | { "hq", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = PRORES_PROFILE_HQ }, | ||
| 995 | 0, 0, VE, .unit = "profile" }, | ||
| 996 | { "4444", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = PRORES_PROFILE_4444 }, | ||
| 997 | 0, 0, VE, .unit = "profile" }, | ||
| 998 | { "4444xq", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = PRORES_PROFILE_4444XQ }, | ||
| 999 | 0, 0, VE, .unit = "profile" }, | ||
| 1000 | { "vendor", "vendor ID", OFFSET(vendor), | ||
| 1001 | AV_OPT_TYPE_STRING, { .str = "Lavc" }, 0, 0, VE }, | ||
| 1002 | { "bits_per_mb", "desired bits per macroblock", OFFSET(bits_per_mb), | ||
| 1003 | AV_OPT_TYPE_INT, { .i64 = 0 }, 0, 8192, VE }, | ||
| 1004 | { "quant_mat", "quantiser matrix", OFFSET(quant_sel), AV_OPT_TYPE_INT, | ||
| 1005 | { .i64 = -1 }, -1, QUANT_MAT_DEFAULT, VE, .unit = "quant_mat" }, | ||
| 1006 | { "auto", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = -1 }, | ||
| 1007 | 0, 0, VE, .unit = "quant_mat" }, | ||
| 1008 | { "proxy", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = QUANT_MAT_PROXY }, | ||
| 1009 | 0, 0, VE, .unit = "quant_mat" }, | ||
| 1010 | { "lt", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = QUANT_MAT_LT }, | ||
| 1011 | 0, 0, VE, .unit = "quant_mat" }, | ||
| 1012 | { "standard", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = QUANT_MAT_STANDARD }, | ||
| 1013 | 0, 0, VE, .unit = "quant_mat" }, | ||
| 1014 | { "hq", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = QUANT_MAT_HQ }, | ||
| 1015 | 0, 0, VE, .unit = "quant_mat" }, | ||
| 1016 | { "default", NULL, 0, AV_OPT_TYPE_CONST, { .i64 = QUANT_MAT_DEFAULT }, | ||
| 1017 | 0, 0, VE, .unit = "quant_mat" }, | ||
| 1018 | { "alpha_bits", "bits for alpha plane", OFFSET(alpha_bits), AV_OPT_TYPE_INT, | ||
| 1019 | { .i64 = 16 }, 0, 16, VE }, | ||
| 1020 | { NULL } | ||
| 1021 | }; | ||
| 1022 | |||
| 1023 | static const AVClass proresenc_class = { | ||
| 1024 | .class_name = "ProRes encoder", | ||
| 1025 | .item_name = av_default_item_name, | ||
| 1026 | .option = options, | ||
| 1027 | .version = LIBAVUTIL_VERSION_INT, | ||
| 1028 | }; | ||
| 1029 | |||
| 1030 | const FFCodec ff_prores_ks_encoder = { | ||
| 1031 | .p.name = "prores_ks", | ||
| 1032 | CODEC_LONG_NAME("Apple ProRes (iCodec Pro)"), | ||
| 1033 | .p.type = AVMEDIA_TYPE_VIDEO, | ||
| 1034 | .p.id = AV_CODEC_ID_PRORES, | ||
| 1035 | .priv_data_size = sizeof(ProresContext), | ||
| 1036 | .init = encode_init, | ||
| 1037 | .close = encode_close, | ||
| 1038 | FF_CODEC_ENCODE_CB(encode_frame), | ||
| 1039 | .p.capabilities = AV_CODEC_CAP_SLICE_THREADS | AV_CODEC_CAP_FRAME_THREADS | | ||
| 1040 | AV_CODEC_CAP_ENCODER_REORDERED_OPAQUE, | ||
| 1041 | CODEC_PIXFMTS(AV_PIX_FMT_YUV422P10, AV_PIX_FMT_YUV444P10, AV_PIX_FMT_YUVA444P10), | ||
| 1042 | .color_ranges = AVCOL_RANGE_MPEG, | ||
| 1043 | .p.priv_class = &proresenc_class, | ||
| 1044 | .p.profiles = NULL_IF_CONFIG_SMALL(ff_prores_profiles), | ||
| 1045 | .caps_internal = FF_CODEC_CAP_INIT_CLEANUP, | ||
| 1046 | }; | ||
| 1047 |