You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

923 lines
36KB

  1. /*
  2. * Copyright (c) 2020 Paul B Mahol
  3. *
  4. * This file is part of FFmpeg.
  5. *
  6. * FFmpeg is free software; you can redistribute it and/or
  7. * modify it under the terms of the GNU Lesser General Public
  8. * License as published by the Free Software Foundation; either
  9. * version 2.1 of the License, or (at your option) any later version.
  10. *
  11. * FFmpeg is distributed in the hope that it will be useful,
  12. * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  14. * Lesser General Public License for more details.
  15. *
  16. * You should have received a copy of the GNU Lesser General Public
  17. * License along with FFmpeg; if not, write to the Free Software
  18. * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  19. */
  20. /**
  21. * @file
  22. * Cineform HD video encoder
  23. */
  24. #include <stdlib.h>
  25. #include <string.h>
  26. #include "libavutil/avassert.h"
  27. #include "libavutil/imgutils.h"
  28. #include "libavutil/opt.h"
  29. #include "avcodec.h"
  30. #include "bytestream.h"
  31. #include "cfhd.h"
  32. #include "put_bits.h"
  33. #include "internal.h"
  34. #include "thread.h"
  35. /* Derived from existing tables from decoder */
  36. static const unsigned codebook[256][2] = {
  37. { 1, 0x00000000 }, { 2, 0x00000002 }, { 3, 0x00000007 }, { 5, 0x00000019 }, { 6, 0x00000030 },
  38. { 6, 0x00000036 }, { 7, 0x00000063 }, { 7, 0x0000006B }, { 7, 0x0000006F }, { 8, 0x000000D4 },
  39. { 8, 0x000000DC }, { 9, 0x00000189 }, { 9, 0x000001A0 }, { 9, 0x000001AB }, {10, 0x00000310 },
  40. {10, 0x00000316 }, {10, 0x00000354 }, {10, 0x00000375 }, {10, 0x00000377 }, {11, 0x00000623 },
  41. {11, 0x00000684 }, {11, 0x000006AB }, {11, 0x000006EC }, {12, 0x00000C44 }, {12, 0x00000C5C },
  42. {12, 0x00000C5E }, {12, 0x00000D55 }, {12, 0x00000DD1 }, {12, 0x00000DD3 }, {12, 0x00000DDB },
  43. {13, 0x0000188B }, {13, 0x000018BB }, {13, 0x00001AA8 }, {13, 0x00001BA0 }, {13, 0x00001BA4 },
  44. {13, 0x00001BB5 }, {14, 0x00003115 }, {14, 0x00003175 }, {14, 0x0000317D }, {14, 0x00003553 },
  45. {14, 0x00003768 }, {15, 0x00006228 }, {15, 0x000062E8 }, {15, 0x000062F8 }, {15, 0x00006AA4 },
  46. {15, 0x00006E85 }, {15, 0x00006E87 }, {15, 0x00006ED3 }, {16, 0x0000C453 }, {16, 0x0000C5D3 },
  47. {16, 0x0000C5F3 }, {16, 0x0000DD08 }, {16, 0x0000DD0C }, {16, 0x0000DDA4 }, {17, 0x000188A4 },
  48. {17, 0x00018BA5 }, {17, 0x00018BE5 }, {17, 0x0001AA95 }, {17, 0x0001AA97 }, {17, 0x0001BA13 },
  49. {17, 0x0001BB4A }, {17, 0x0001BB4B }, {18, 0x00031748 }, {18, 0x000317C8 }, {18, 0x00035528 },
  50. {18, 0x0003552C }, {18, 0x00037424 }, {18, 0x00037434 }, {18, 0x00037436 }, {19, 0x00062294 },
  51. {19, 0x00062E92 }, {19, 0x00062F92 }, {19, 0x0006AA52 }, {19, 0x0006AA5A }, {19, 0x0006E84A },
  52. {19, 0x0006E86A }, {19, 0x0006E86E }, {20, 0x000C452A }, {20, 0x000C5D27 }, {20, 0x000C5F26 },
  53. {20, 0x000D54A6 }, {20, 0x000D54B6 }, {20, 0x000DD096 }, {20, 0x000DD0D6 }, {20, 0x000DD0DE },
  54. {21, 0x00188A56 }, {21, 0x0018BA4D }, {21, 0x0018BE4E }, {21, 0x0018BE4F }, {21, 0x001AA96E },
  55. {21, 0x001BA12E }, {21, 0x001BA12F }, {21, 0x001BA1AF }, {21, 0x001BA1BF }, {22, 0x00317498 },
  56. {22, 0x0035529C }, {22, 0x0035529D }, {22, 0x003552DE }, {22, 0x003552DF }, {22, 0x0037435D },
  57. {22, 0x0037437D }, {23, 0x0062295D }, {23, 0x0062E933 }, {23, 0x006AA53D }, {23, 0x006AA53E },
  58. {23, 0x006AA53F }, {23, 0x006E86B9 }, {23, 0x006E86F8 }, {24, 0x00C452B8 }, {24, 0x00C5D265 },
  59. {24, 0x00D54A78 }, {24, 0x00D54A79 }, {24, 0x00DD0D70 }, {24, 0x00DD0D71 }, {24, 0x00DD0DF2 },
  60. {24, 0x00DD0DF3 }, {26, 0x03114BA2 }, {25, 0x0188A5B1 }, {25, 0x0188A58B }, {25, 0x0188A595 },
  61. {25, 0x0188A5D6 }, {25, 0x0188A5D7 }, {25, 0x0188A5A8 }, {25, 0x0188A5AE }, {25, 0x0188A5AF },
  62. {25, 0x0188A5C4 }, {25, 0x0188A5C5 }, {25, 0x0188A587 }, {25, 0x0188A584 }, {25, 0x0188A585 },
  63. {25, 0x0188A5C6 }, {25, 0x0188A5C7 }, {25, 0x0188A5CC }, {25, 0x0188A5CD }, {25, 0x0188A581 },
  64. {25, 0x0188A582 }, {25, 0x0188A583 }, {25, 0x0188A5CE }, {25, 0x0188A5CF }, {25, 0x0188A5C2 },
  65. {25, 0x0188A5C3 }, {25, 0x0188A5C1 }, {25, 0x0188A5B4 }, {25, 0x0188A5B5 }, {25, 0x0188A5E6 },
  66. {25, 0x0188A5E7 }, {25, 0x0188A5E4 }, {25, 0x0188A5E5 }, {25, 0x0188A5AB }, {25, 0x0188A5E0 },
  67. {25, 0x0188A5E1 }, {25, 0x0188A5E2 }, {25, 0x0188A5E3 }, {25, 0x0188A5B6 }, {25, 0x0188A5B7 },
  68. {25, 0x0188A5FD }, {25, 0x0188A57E }, {25, 0x0188A57F }, {25, 0x0188A5EC }, {25, 0x0188A5ED },
  69. {25, 0x0188A5FE }, {25, 0x0188A5FF }, {25, 0x0188A57D }, {25, 0x0188A59C }, {25, 0x0188A59D },
  70. {25, 0x0188A5E8 }, {25, 0x0188A5E9 }, {25, 0x0188A5EA }, {25, 0x0188A5EB }, {25, 0x0188A5EF },
  71. {25, 0x0188A57A }, {25, 0x0188A57B }, {25, 0x0188A578 }, {25, 0x0188A579 }, {25, 0x0188A5BA },
  72. {25, 0x0188A5BB }, {25, 0x0188A5B8 }, {25, 0x0188A5B9 }, {25, 0x0188A588 }, {25, 0x0188A589 },
  73. {25, 0x018BA4C8 }, {25, 0x018BA4C9 }, {25, 0x0188A5FA }, {25, 0x0188A5FB }, {25, 0x0188A5BC },
  74. {25, 0x0188A5BD }, {25, 0x0188A598 }, {25, 0x0188A599 }, {25, 0x0188A5F4 }, {25, 0x0188A5F5 },
  75. {25, 0x0188A59B }, {25, 0x0188A5DE }, {25, 0x0188A5DF }, {25, 0x0188A596 }, {25, 0x0188A597 },
  76. {25, 0x0188A5F8 }, {25, 0x0188A5F9 }, {25, 0x0188A5F1 }, {25, 0x0188A58E }, {25, 0x0188A58F },
  77. {25, 0x0188A5DC }, {25, 0x0188A5DD }, {25, 0x0188A5F2 }, {25, 0x0188A5F3 }, {25, 0x0188A58C },
  78. {25, 0x0188A58D }, {25, 0x0188A5A4 }, {25, 0x0188A5F0 }, {25, 0x0188A5A5 }, {25, 0x0188A5A6 },
  79. {25, 0x0188A5A7 }, {25, 0x0188A59A }, {25, 0x0188A5A2 }, {25, 0x0188A5A3 }, {25, 0x0188A58A },
  80. {25, 0x0188A5B0 }, {25, 0x0188A5A0 }, {25, 0x0188A5A1 }, {25, 0x0188A5DA }, {25, 0x0188A5DB },
  81. {25, 0x0188A59E }, {25, 0x0188A59F }, {25, 0x0188A5D8 }, {25, 0x0188A5EE }, {25, 0x0188A5D9 },
  82. {25, 0x0188A5F6 }, {25, 0x0188A5F7 }, {25, 0x0188A57C }, {25, 0x0188A5C8 }, {25, 0x0188A5C9 },
  83. {25, 0x0188A594 }, {25, 0x0188A5FC }, {25, 0x0188A5CA }, {25, 0x0188A5CB }, {25, 0x0188A5B2 },
  84. {25, 0x0188A5AA }, {25, 0x0188A5B3 }, {25, 0x0188A572 }, {25, 0x0188A573 }, {25, 0x0188A5C0 },
  85. {25, 0x0188A5BE }, {25, 0x0188A5BF }, {25, 0x0188A592 }, {25, 0x0188A580 }, {25, 0x0188A593 },
  86. {25, 0x0188A590 }, {25, 0x0188A591 }, {25, 0x0188A586 }, {25, 0x0188A5A9 }, {25, 0x0188A5D2 },
  87. {25, 0x0188A5D3 }, {25, 0x0188A5D4 }, {25, 0x0188A5D5 }, {25, 0x0188A5AC }, {25, 0x0188A5AD },
  88. {25, 0x0188A5D0 },
  89. };
  90. /* Derived by extracting runcodes from existing tables from decoder */
  91. static const uint16_t runbook[18][3] = {
  92. {1, 0x0000, 1}, {2, 0x0000, 2}, {3, 0x0000, 3}, {4, 0x0000, 4},
  93. {5, 0x0000, 5}, {6, 0x0000, 6}, {7, 0x0000, 7}, {8, 0x0000, 8},
  94. {9, 0x0000, 9}, {10, 0x0000, 10}, {11, 0x0000, 11},
  95. {7, 0x0069, 12}, {8, 0x00D1, 20}, {9, 0x018A, 32},
  96. {10, 0x0343, 60}, {11, 0x0685, 100}, {13, 0x18BF, 180}, {13, 0x1BA5, 320},
  97. };
  98. /*
  99. * Derived by inspecting various quality encodes
  100. * and adding some more from scratch.
  101. */
  102. static const uint16_t quantization_per_subband[2][3][13][9] = {
  103. {{
  104. { 16, 16, 8, 4, 4, 2, 6, 6, 9, }, // film3+
  105. { 16, 16, 8, 4, 4, 2, 6, 6, 9, }, // film3
  106. { 16, 16, 8, 4, 4, 2, 7, 7, 10, }, // film2+
  107. { 16, 16, 8, 4, 4, 2, 8, 8, 12, }, // film2
  108. { 16, 16, 8, 4, 4, 2, 16, 16, 26, }, // film1++
  109. { 24, 24, 12, 6, 6, 3, 24, 24, 36, }, // film1+
  110. { 24, 24, 12, 6, 6, 3, 24, 24, 36, }, // film1
  111. { 32, 32, 24, 8, 8, 6, 32, 32, 48, }, // high+
  112. { 32, 32, 24, 8, 8, 6, 32, 32, 48, }, // high
  113. { 48, 48, 32, 12, 12, 8, 64, 64, 96, }, // medium+
  114. { 48, 48, 32, 12, 12, 8, 64, 64, 96, }, // medium
  115. { 64, 64, 48, 16, 16, 12, 96, 96, 144, }, // low+
  116. { 64, 64, 48, 16, 16, 12, 128, 128, 192, }, // low
  117. },
  118. {
  119. { 16, 16, 8, 4, 4, 2, 6, 6, 9, }, // film3+
  120. { 16, 16, 8, 4, 4, 2, 6, 6, 12, }, // film3
  121. { 16, 16, 8, 4, 4, 2, 7, 7, 14, }, // film2+
  122. { 16, 16, 8, 4, 4, 2, 8, 8, 16, }, // film2
  123. { 16, 16, 8, 4, 4, 2, 16, 16, 26, }, // film1++
  124. { 24, 24, 12, 6, 6, 3, 24, 24, 36, }, // film1+
  125. { 24, 24, 12, 6, 6, 3, 24, 24, 48, }, // film1
  126. { 32, 32, 24, 8, 8, 6, 32, 32, 48, }, // high+
  127. { 48, 48, 32, 12, 12, 8, 32, 32, 64, }, // high
  128. { 48, 48, 32, 12, 12, 8, 64, 64, 96, }, // medium+
  129. { 48, 48, 32, 12, 12, 8, 64, 64, 128, }, // medium
  130. { 64, 64, 48, 16, 16, 12, 96, 96, 160, }, // low+
  131. { 64, 64, 48, 16, 16, 12, 128, 128, 192, }, // low
  132. },
  133. {
  134. { 16, 16, 8, 4, 4, 2, 6, 6, 9, }, // film3+
  135. { 16, 16, 8, 4, 4, 2, 6, 6, 12, }, // film3
  136. { 16, 16, 8, 4, 4, 2, 7, 7, 14, }, // film2+
  137. { 16, 16, 8, 4, 4, 2, 8, 8, 16, }, // film2
  138. { 16, 16, 8, 4, 4, 2, 16, 16, 26, }, // film1++
  139. { 24, 24, 12, 6, 6, 3, 24, 24, 36, }, // film1+
  140. { 24, 24, 12, 6, 6, 3, 24, 24, 48, }, // film1
  141. { 32, 32, 24, 8, 8, 6, 32, 32, 48, }, // high+
  142. { 48, 48, 32, 12, 12, 8, 32, 32, 64, }, // high
  143. { 48, 48, 32, 12, 12, 8, 64, 64, 96, }, // medium+
  144. { 48, 48, 32, 12, 12, 8, 64, 64, 128, }, // medium
  145. { 64, 64, 48, 16, 16, 12, 96, 96, 160, }, // low+
  146. { 64, 64, 48, 16, 16, 12, 128, 128, 192, }, // low
  147. }},
  148. {{
  149. { 16, 16, 8, 16, 16, 8, 24, 24, 36, }, // film3+
  150. { 16, 16, 8, 16, 16, 8, 24, 24, 36, }, // film3
  151. { 16, 16, 8, 16, 16, 8, 32, 32, 48, }, // film2+
  152. { 16, 16, 8, 16, 16, 8, 32, 32, 48, }, // film2
  153. { 16, 16, 8, 20, 20, 10, 80, 80, 128, }, // film1++
  154. { 24, 24, 12, 24, 24, 12, 96, 96, 144, }, // film1+
  155. { 24, 24, 12, 24, 24, 12, 96, 96, 144, }, // film1
  156. { 32, 32, 24, 32, 32, 24, 128, 128, 192, }, // high+
  157. { 32, 32, 24, 32, 32, 24, 128, 128, 192, }, // high
  158. { 48, 48, 32, 48, 48, 32, 256, 256, 384, }, // medium+
  159. { 48, 48, 32, 48, 48, 32, 256, 256, 384, }, // medium
  160. { 56, 56, 40, 56, 56, 40, 512, 512, 768, }, // low+
  161. { 64, 64, 48, 64, 64, 48, 512, 512, 768, }, // low
  162. },
  163. {
  164. { 16, 16, 8, 16, 16, 8, 24, 24, 36, }, // film3+
  165. { 16, 16, 8, 16, 16, 8, 48, 48, 72, }, // film3
  166. { 16, 16, 8, 16, 16, 8, 48, 48, 72, }, // film2+
  167. { 16, 16, 8, 16, 16, 8, 64, 64, 96, }, // film2
  168. { 16, 16, 8, 20, 20, 10, 80, 80, 128, }, // film1++
  169. { 24, 24, 12, 24, 24, 12, 96, 96, 144, }, // film1+
  170. { 24, 24, 12, 24, 24, 12, 192, 192, 288, }, // film1
  171. { 32, 32, 24, 32, 32, 24, 128, 128, 192, }, // high+
  172. { 32, 32, 24, 32, 32, 24, 256, 256, 384, }, // high
  173. { 48, 48, 32, 48, 48, 32, 256, 256, 384, }, // medium+
  174. { 48, 48, 32, 48, 48, 32, 512, 512, 768, }, // medium
  175. { 56, 56, 40, 56, 56, 40, 512, 512, 768, }, // low+
  176. { 64, 64, 48, 64, 64, 48,1024,1024,1536, }, // low
  177. },
  178. {
  179. { 16, 16, 8, 16, 16, 8, 24, 24, 36, }, // film3+
  180. { 16, 16, 8, 16, 16, 8, 48, 48, 72, }, // film3
  181. { 16, 16, 8, 16, 16, 8, 48, 48, 72, }, // film2+
  182. { 16, 16, 8, 16, 16, 8, 64, 64, 96, }, // film2
  183. { 16, 16, 10, 20, 20, 10, 80, 80, 128, }, // film1++
  184. { 24, 24, 12, 24, 24, 12, 96, 96, 144, }, // film1+
  185. { 24, 24, 12, 24, 24, 12, 192, 192, 288, }, // film1
  186. { 32, 32, 24, 32, 32, 24, 128, 128, 192, }, // high+
  187. { 32, 32, 24, 32, 32, 24, 256, 256, 384, }, // high
  188. { 48, 48, 32, 48, 48, 32, 256, 256, 384, }, // medium+
  189. { 48, 48, 32, 48, 48, 32, 512, 512, 768, }, // medium
  190. { 56, 56, 40, 56, 56, 40, 512, 512, 768, }, // low+
  191. { 64, 64, 48, 64, 64, 48,1024,1024,1536, }, // low
  192. }},
  193. };
  194. typedef struct Codebook {
  195. unsigned bits;
  196. unsigned size;
  197. } Codebook;
  198. typedef struct Runbook {
  199. unsigned size;
  200. unsigned bits;
  201. unsigned run;
  202. } Runbook;
  203. typedef struct PlaneEnc {
  204. unsigned size;
  205. int16_t *dwt_buf;
  206. int16_t *dwt_tmp;
  207. unsigned quantization[SUBBAND_COUNT];
  208. int16_t *subband[SUBBAND_COUNT];
  209. int16_t *l_h[8];
  210. SubBand band[DWT_LEVELS][4];
  211. } PlaneEnc;
  212. typedef struct CFHDEncContext {
  213. const AVClass *class;
  214. PutBitContext pb;
  215. PutByteContext pby;
  216. int quality;
  217. int planes;
  218. int chroma_h_shift;
  219. int chroma_v_shift;
  220. PlaneEnc plane[4];
  221. uint16_t lut[1024];
  222. Runbook rb[321];
  223. Codebook cb[513];
  224. int16_t *alpha;
  225. } CFHDEncContext;
  226. static av_cold int cfhd_encode_init(AVCodecContext *avctx)
  227. {
  228. CFHDEncContext *s = avctx->priv_data;
  229. const int sign_mask = 256;
  230. const int twos_complement = -sign_mask;
  231. const int mag_mask = sign_mask - 1;
  232. int ret, last = 0;
  233. ret = av_pix_fmt_get_chroma_sub_sample(avctx->pix_fmt,
  234. &s->chroma_h_shift,
  235. &s->chroma_v_shift);
  236. if (ret < 0)
  237. return ret;
  238. if (avctx->width & 15) {
  239. av_log(avctx, AV_LOG_ERROR, "Width must be multiple of 16.\n");
  240. return AVERROR_INVALIDDATA;
  241. }
  242. s->planes = av_pix_fmt_count_planes(avctx->pix_fmt);
  243. for (int i = 0; i < s->planes; i++) {
  244. int w8, h8, w4, h4, w2, h2;
  245. int width = i ? avctx->width >> s->chroma_h_shift : avctx->width;
  246. int height = i ? FFALIGN(avctx->height >> s->chroma_v_shift, 8) :
  247. FFALIGN(avctx->height >> s->chroma_v_shift, 8);
  248. ptrdiff_t stride = FFALIGN(width / 8, 8) * 8;
  249. w8 = FFALIGN(width / 8, 8);
  250. h8 = height / 8;
  251. w4 = w8 * 2;
  252. h4 = h8 * 2;
  253. w2 = w4 * 2;
  254. h2 = h4 * 2;
  255. s->plane[i].dwt_buf =
  256. av_mallocz_array(height * stride, sizeof(*s->plane[i].dwt_buf));
  257. s->plane[i].dwt_tmp =
  258. av_malloc_array(height * stride, sizeof(*s->plane[i].dwt_tmp));
  259. if (!s->plane[i].dwt_buf || !s->plane[i].dwt_tmp)
  260. return AVERROR(ENOMEM);
  261. s->plane[i].subband[0] = s->plane[i].dwt_buf;
  262. s->plane[i].subband[1] = s->plane[i].dwt_buf + 2 * w8 * h8;
  263. s->plane[i].subband[2] = s->plane[i].dwt_buf + 1 * w8 * h8;
  264. s->plane[i].subband[3] = s->plane[i].dwt_buf + 3 * w8 * h8;
  265. s->plane[i].subband[4] = s->plane[i].dwt_buf + 2 * w4 * h4;
  266. s->plane[i].subband[5] = s->plane[i].dwt_buf + 1 * w4 * h4;
  267. s->plane[i].subband[6] = s->plane[i].dwt_buf + 3 * w4 * h4;
  268. s->plane[i].subband[7] = s->plane[i].dwt_buf + 2 * w2 * h2;
  269. s->plane[i].subband[8] = s->plane[i].dwt_buf + 1 * w2 * h2;
  270. s->plane[i].subband[9] = s->plane[i].dwt_buf + 3 * w2 * h2;
  271. for (int j = 0; j < DWT_LEVELS; j++) {
  272. for (int k = 0; k < FF_ARRAY_ELEMS(s->plane[i].band[j]); k++) {
  273. s->plane[i].band[j][k].width = (width / 8) << j;
  274. s->plane[i].band[j][k].height = (height / 8) << j;
  275. s->plane[i].band[j][k].a_width = w8 << j;
  276. s->plane[i].band[j][k].a_height = h8 << j;
  277. }
  278. }
  279. /* ll2 and ll1 commented out because they are done in-place */
  280. s->plane[i].l_h[0] = s->plane[i].dwt_tmp;
  281. s->plane[i].l_h[1] = s->plane[i].dwt_tmp + 2 * w8 * h8;
  282. // s->plane[i].l_h[2] = ll2;
  283. s->plane[i].l_h[3] = s->plane[i].dwt_tmp;
  284. s->plane[i].l_h[4] = s->plane[i].dwt_tmp + 2 * w4 * h4;
  285. // s->plane[i].l_h[5] = ll1;
  286. s->plane[i].l_h[6] = s->plane[i].dwt_tmp;
  287. s->plane[i].l_h[7] = s->plane[i].dwt_tmp + 2 * w2 * h2;
  288. }
  289. for (int i = 0; i < 512; i++) {
  290. int value = (i & sign_mask) ? twos_complement + (i & mag_mask): i;
  291. int mag = FFMIN(FFABS(value), 255);
  292. if (mag) {
  293. s->cb[i].bits = (codebook[mag][1] << 1) | (value > 0 ? 0 : 1);
  294. s->cb[i].size = codebook[mag][0] + 1;
  295. } else {
  296. s->cb[i].bits = codebook[mag][1];
  297. s->cb[i].size = codebook[mag][0];
  298. }
  299. }
  300. s->cb[512].bits = 0x3114ba3;
  301. s->cb[512].size = 26;
  302. s->rb[0].run = 0;
  303. for (int i = 1, j = 0; i < 320 && j < 17; j++) {
  304. int run = runbook[j][2];
  305. int end = runbook[j+1][2];
  306. while (i < end) {
  307. s->rb[i].run = run;
  308. s->rb[i].bits = runbook[j][1];
  309. s->rb[i++].size = runbook[j][0];
  310. }
  311. }
  312. s->rb[320].bits = runbook[17][1];
  313. s->rb[320].size = runbook[17][0];
  314. s->rb[320].run = 320;
  315. for (int i = 0; i < 256; i++) {
  316. int idx = i + ((768LL * i * i * i) / (256 * 256 * 256));
  317. s->lut[idx] = i;
  318. }
  319. for (int i = 0; i < 1024; i++) {
  320. if (s->lut[i])
  321. last = s->lut[i];
  322. else
  323. s->lut[i] = last;
  324. }
  325. if (s->planes != 4)
  326. return 0;
  327. s->alpha = av_calloc(avctx->width * avctx->height, sizeof(*s->alpha));
  328. if (!s->alpha)
  329. return AVERROR(ENOMEM);
  330. return 0;
  331. }
  332. static av_always_inline void filter(int16_t *input, ptrdiff_t in_stride,
  333. int16_t *low, ptrdiff_t low_stride,
  334. int16_t *high, ptrdiff_t high_stride,
  335. int len)
  336. {
  337. low[(0>>1) * low_stride] = av_clip_int16(input[0*in_stride] + input[1*in_stride]);
  338. high[(0>>1) * high_stride] = av_clip_int16((5 * input[0*in_stride] - 11 * input[1*in_stride] +
  339. 4 * input[2*in_stride] + 4 * input[3*in_stride] -
  340. 1 * input[4*in_stride] - 1 * input[5*in_stride] + 4) >> 3);
  341. for (int i = 2; i < len - 2; i += 2) {
  342. low[(i>>1) * low_stride] = av_clip_int16(input[i*in_stride] + input[(i+1)*in_stride]);
  343. high[(i>>1) * high_stride] = av_clip_int16(((-input[(i-2)*in_stride] - input[(i-1)*in_stride] +
  344. input[(i+2)*in_stride] + input[(i+3)*in_stride] + 4) >> 3) +
  345. input[(i+0)*in_stride] - input[(i+1)*in_stride]);
  346. }
  347. low[((len-2)>>1) * low_stride] = av_clip_int16(input[((len-2)+0)*in_stride] + input[((len-2)+1)*in_stride]);
  348. high[((len-2)>>1) * high_stride] = av_clip_int16((11* input[((len-2)+0)*in_stride] - 5 * input[((len-2)+1)*in_stride] -
  349. 4 * input[((len-2)-1)*in_stride] - 4 * input[((len-2)-2)*in_stride] +
  350. 1 * input[((len-2)-3)*in_stride] + 1 * input[((len-2)-4)*in_stride] + 4) >> 3);
  351. }
  352. static void horiz_filter(int16_t *input, int16_t *low, int16_t *high,
  353. int width)
  354. {
  355. filter(input, 1, low, 1, high, 1, width);
  356. }
  357. static void vert_filter(int16_t *input, ptrdiff_t in_stride,
  358. int16_t *low, ptrdiff_t low_stride,
  359. int16_t *high, ptrdiff_t high_stride, int len)
  360. {
  361. filter(input, in_stride, low, low_stride, high, high_stride, len);
  362. }
  363. static void quantize_band(int16_t *input, int width, int a_width,
  364. int height, unsigned quantization)
  365. {
  366. const int16_t factor = (uint32_t)(1U << 15) / quantization;
  367. for (int i = 0; i < height; i++) {
  368. for (int j = 0; j < width; j++)
  369. input[j] = av_clip_intp2(((input[j] * factor + 16384 * FFSIGN(input[j])) / 32768), 10);
  370. input += a_width;
  371. }
  372. }
  373. static int put_runcode(PutBitContext *pb, int count, const Runbook *const rb)
  374. {
  375. while (count > 0) {
  376. const int index = FFMIN(320, count);
  377. put_bits(pb, rb[index].size, rb[index].bits);
  378. count -= rb[index].run;
  379. }
  380. return 0;
  381. }
  382. static void process_alpha(const int16_t *src, int width, int height, ptrdiff_t stride, int16_t *dst)
  383. {
  384. for (int i = 0; i < height; i++) {
  385. for (int j = 0; j < width; j++) {
  386. int alpha = src[j];
  387. if (alpha > 0 && alpha < 4080) {
  388. alpha *= 223;
  389. alpha += 128;
  390. alpha >>= 8;
  391. alpha += 256;
  392. }
  393. dst[j] = av_clip_uintp2(alpha, 12);
  394. }
  395. src += stride;
  396. dst += width;
  397. }
  398. }
  399. static int cfhd_encode_frame(AVCodecContext *avctx, AVPacket *pkt,
  400. const AVFrame *frame, int *got_packet)
  401. {
  402. CFHDEncContext *s = avctx->priv_data;
  403. PutByteContext *pby = &s->pby;
  404. PutBitContext *pb = &s->pb;
  405. const Codebook *const cb = s->cb;
  406. const Runbook *const rb = s->rb;
  407. const uint16_t *lut = s->lut;
  408. unsigned pos;
  409. int ret;
  410. for (int plane = 0; plane < s->planes; plane++) {
  411. int width = s->plane[plane].band[2][0].width;
  412. int a_width = s->plane[plane].band[2][0].a_width;
  413. int height = s->plane[plane].band[2][0].height;
  414. int act_plane = plane == 1 ? 2 : plane == 2 ? 1 : plane;
  415. int16_t *input = (int16_t *)frame->data[act_plane];
  416. int16_t *low = s->plane[plane].l_h[6];
  417. int16_t *high = s->plane[plane].l_h[7];
  418. ptrdiff_t in_stride = frame->linesize[act_plane] / 2;
  419. int low_stride, high_stride;
  420. if (plane == 3) {
  421. process_alpha(input, avctx->width, avctx->height,
  422. in_stride, s->alpha);
  423. input = s->alpha;
  424. in_stride = avctx->width;
  425. }
  426. for (int i = 0; i < height * 2; i++) {
  427. horiz_filter(input, low, high, width * 2);
  428. input += in_stride;
  429. low += a_width;
  430. high += a_width;
  431. }
  432. input = s->plane[plane].l_h[7];
  433. low = s->plane[plane].subband[7];
  434. low_stride = s->plane[plane].band[2][0].a_width;
  435. high = s->plane[plane].subband[9];
  436. high_stride = s->plane[plane].band[2][0].a_width;
  437. for (int i = 0; i < width; i++) {
  438. vert_filter(input, a_width, low, low_stride, high, high_stride, height * 2);
  439. input++;
  440. low++;
  441. high++;
  442. }
  443. input = s->plane[plane].l_h[6];
  444. low = s->plane[plane].l_h[7];
  445. high = s->plane[plane].subband[8];
  446. for (int i = 0; i < width; i++) {
  447. vert_filter(input, a_width, low, low_stride, high, high_stride, height * 2);
  448. input++;
  449. low++;
  450. high++;
  451. }
  452. a_width = s->plane[plane].band[1][0].a_width;
  453. width = s->plane[plane].band[1][0].width;
  454. height = s->plane[plane].band[1][0].height;
  455. input = s->plane[plane].l_h[7];
  456. low = s->plane[plane].l_h[3];
  457. low_stride = s->plane[plane].band[1][0].a_width;
  458. high = s->plane[plane].l_h[4];
  459. high_stride = s->plane[plane].band[1][0].a_width;
  460. for (int i = 0; i < height * 2; i++) {
  461. for (int j = 0; j < width * 2; j++)
  462. input[j] /= 4;
  463. input += a_width * 2;
  464. }
  465. input = s->plane[plane].l_h[7];
  466. for (int i = 0; i < height * 2; i++) {
  467. horiz_filter(input, low, high, width * 2);
  468. input += a_width * 2;
  469. low += low_stride;
  470. high += high_stride;
  471. }
  472. input = s->plane[plane].l_h[4];
  473. low = s->plane[plane].subband[4];
  474. high = s->plane[plane].subband[6];
  475. for (int i = 0; i < width; i++) {
  476. vert_filter(input, a_width, low, low_stride, high, high_stride, height * 2);
  477. input++;
  478. low++;
  479. high++;
  480. }
  481. input = s->plane[plane].l_h[3];
  482. low = s->plane[plane].l_h[4];
  483. high = s->plane[plane].subband[5];
  484. for (int i = 0; i < width; i++) {
  485. vert_filter(input, a_width, low, low_stride, high, high_stride, height * 2);
  486. input++;
  487. low++;
  488. high++;
  489. }
  490. a_width = s->plane[plane].band[0][0].a_width;
  491. width = s->plane[plane].band[0][0].width;
  492. height = s->plane[plane].band[0][0].height;
  493. input = s->plane[plane].l_h[4];
  494. low = s->plane[plane].l_h[0];
  495. low_stride = s->plane[plane].band[0][0].a_width;
  496. high = s->plane[plane].l_h[1];
  497. high_stride = s->plane[plane].band[0][0].a_width;
  498. if (avctx->pix_fmt != AV_PIX_FMT_YUV422P10) {
  499. for (int i = 0; i < height * 2; i++) {
  500. for (int j = 0; j < width * 2; j++)
  501. input[j] /= 4;
  502. input += a_width * 2;
  503. }
  504. }
  505. input = s->plane[plane].l_h[4];
  506. for (int i = 0; i < height * 2; i++) {
  507. horiz_filter(input, low, high, width * 2);
  508. input += a_width * 2;
  509. low += low_stride;
  510. high += high_stride;
  511. }
  512. low = s->plane[plane].subband[1];
  513. high = s->plane[plane].subband[3];
  514. input = s->plane[plane].l_h[1];
  515. for (int i = 0; i < width; i++) {
  516. vert_filter(input, a_width, low, low_stride, high, high_stride, height * 2);
  517. input++;
  518. low++;
  519. high++;
  520. }
  521. low = s->plane[plane].subband[0];
  522. high = s->plane[plane].subband[2];
  523. input = s->plane[plane].l_h[0];
  524. for (int i = 0; i < width; i++) {
  525. vert_filter(input, a_width, low, low_stride, high, high_stride, height * 2);
  526. input++;
  527. low++;
  528. high++;
  529. }
  530. }
  531. ret = ff_alloc_packet2(avctx, pkt, 64LL + s->planes * (2LL * avctx->width * avctx->height + 1000LL), 0);
  532. if (ret < 0)
  533. return ret;
  534. bytestream2_init_writer(pby, pkt->data, pkt->size);
  535. bytestream2_put_be16(pby, SampleType);
  536. bytestream2_put_be16(pby, 9);
  537. bytestream2_put_be16(pby, SampleIndexTable);
  538. bytestream2_put_be16(pby, s->planes);
  539. for (int i = 0; i < s->planes; i++)
  540. bytestream2_put_be32(pby, 0);
  541. bytestream2_put_be16(pby, TransformType);
  542. bytestream2_put_be16(pby, 0);
  543. bytestream2_put_be16(pby, NumFrames);
  544. bytestream2_put_be16(pby, 1);
  545. bytestream2_put_be16(pby, ChannelCount);
  546. bytestream2_put_be16(pby, s->planes);
  547. bytestream2_put_be16(pby, EncodedFormat);
  548. bytestream2_put_be16(pby, avctx->pix_fmt == AV_PIX_FMT_YUV422P10 ? 1 : 3 + (s->planes == 4));
  549. bytestream2_put_be16(pby, WaveletCount);
  550. bytestream2_put_be16(pby, 3);
  551. bytestream2_put_be16(pby, SubbandCount);
  552. bytestream2_put_be16(pby, SUBBAND_COUNT);
  553. bytestream2_put_be16(pby, NumSpatial);
  554. bytestream2_put_be16(pby, 2);
  555. bytestream2_put_be16(pby, FirstWavelet);
  556. bytestream2_put_be16(pby, 3);
  557. bytestream2_put_be16(pby, ImageWidth);
  558. bytestream2_put_be16(pby, avctx->width);
  559. bytestream2_put_be16(pby, ImageHeight);
  560. bytestream2_put_be16(pby, avctx->height);
  561. bytestream2_put_be16(pby, -FrameNumber);
  562. bytestream2_put_be16(pby, avctx->frame_number);
  563. bytestream2_put_be16(pby, Precision);
  564. bytestream2_put_be16(pby, avctx->pix_fmt == AV_PIX_FMT_YUV422P10 ? 10 : 12);
  565. bytestream2_put_be16(pby, PrescaleTable);
  566. bytestream2_put_be16(pby, avctx->pix_fmt == AV_PIX_FMT_YUV422P10 ? 0x2000 : 0x2800);
  567. bytestream2_put_be16(pby, SampleFlags);
  568. bytestream2_put_be16(pby, 1);
  569. for (int p = 0; p < s->planes; p++) {
  570. int width = s->plane[p].band[0][0].width;
  571. int a_width = s->plane[p].band[0][0].a_width;
  572. int height = s->plane[p].band[0][0].height;
  573. int16_t *data = s->plane[p].subband[0];
  574. if (p) {
  575. bytestream2_put_be16(pby, SampleType);
  576. bytestream2_put_be16(pby, 3);
  577. bytestream2_put_be16(pby, ChannelNumber);
  578. bytestream2_put_be16(pby, p);
  579. }
  580. bytestream2_put_be16(pby, BitstreamMarker);
  581. bytestream2_put_be16(pby, 0x1a4a);
  582. pos = bytestream2_tell_p(pby);
  583. bytestream2_put_be16(pby, LowpassSubband);
  584. bytestream2_put_be16(pby, 0);
  585. bytestream2_put_be16(pby, NumLevels);
  586. bytestream2_put_be16(pby, 3);
  587. bytestream2_put_be16(pby, LowpassWidth);
  588. bytestream2_put_be16(pby, width);
  589. bytestream2_put_be16(pby, LowpassHeight);
  590. bytestream2_put_be16(pby, height);
  591. bytestream2_put_be16(pby, PixelOffset);
  592. bytestream2_put_be16(pby, 0);
  593. bytestream2_put_be16(pby, LowpassQuantization);
  594. bytestream2_put_be16(pby, 1);
  595. bytestream2_put_be16(pby, LowpassPrecision);
  596. bytestream2_put_be16(pby, 16);
  597. bytestream2_put_be16(pby, BitstreamMarker);
  598. bytestream2_put_be16(pby, 0x0f0f);
  599. for (int i = 0; i < height; i++) {
  600. for (int j = 0; j < width; j++)
  601. bytestream2_put_be16(pby, data[j]);
  602. data += a_width;
  603. }
  604. bytestream2_put_be16(pby, BitstreamMarker);
  605. bytestream2_put_be16(pby, 0x1b4b);
  606. for (int l = 0; l < 3; l++) {
  607. for (int i = 0; i < 3; i++) {
  608. s->plane[p].quantization[1 + l * 3 + i] = quantization_per_subband[avctx->pix_fmt != AV_PIX_FMT_YUV422P10][p >= 3 ? 0 : p][s->quality][l * 3 + i];
  609. }
  610. }
  611. for (int l = 0; l < 3; l++) {
  612. int a_width = s->plane[p].band[l][0].a_width;
  613. int width = s->plane[p].band[l][0].width;
  614. int stride = FFALIGN(width, 8);
  615. int height = s->plane[p].band[l][0].height;
  616. bytestream2_put_be16(pby, BitstreamMarker);
  617. bytestream2_put_be16(pby, 0x0d0d);
  618. bytestream2_put_be16(pby, WaveletType);
  619. bytestream2_put_be16(pby, 3 + 2 * (l == 2));
  620. bytestream2_put_be16(pby, WaveletNumber);
  621. bytestream2_put_be16(pby, 3 - l);
  622. bytestream2_put_be16(pby, WaveletLevel);
  623. bytestream2_put_be16(pby, 3 - l);
  624. bytestream2_put_be16(pby, NumBands);
  625. bytestream2_put_be16(pby, 4);
  626. bytestream2_put_be16(pby, HighpassWidth);
  627. bytestream2_put_be16(pby, width);
  628. bytestream2_put_be16(pby, HighpassHeight);
  629. bytestream2_put_be16(pby, height);
  630. bytestream2_put_be16(pby, LowpassBorder);
  631. bytestream2_put_be16(pby, 0);
  632. bytestream2_put_be16(pby, HighpassBorder);
  633. bytestream2_put_be16(pby, 0);
  634. bytestream2_put_be16(pby, LowpassScale);
  635. bytestream2_put_be16(pby, 1);
  636. bytestream2_put_be16(pby, LowpassDivisor);
  637. bytestream2_put_be16(pby, 1);
  638. for (int i = 0; i < 3; i++) {
  639. int16_t *data = s->plane[p].subband[1 + l * 3 + i];
  640. int count = 0, padd = 0;
  641. bytestream2_put_be16(pby, BitstreamMarker);
  642. bytestream2_put_be16(pby, 0x0e0e);
  643. bytestream2_put_be16(pby, SubbandNumber);
  644. bytestream2_put_be16(pby, i + 1);
  645. bytestream2_put_be16(pby, BandCodingFlags);
  646. bytestream2_put_be16(pby, 1);
  647. bytestream2_put_be16(pby, BandWidth);
  648. bytestream2_put_be16(pby, width);
  649. bytestream2_put_be16(pby, BandHeight);
  650. bytestream2_put_be16(pby, height);
  651. bytestream2_put_be16(pby, SubbandBand);
  652. bytestream2_put_be16(pby, 1 + l * 3 + i);
  653. bytestream2_put_be16(pby, BandEncoding);
  654. bytestream2_put_be16(pby, 3);
  655. bytestream2_put_be16(pby, Quantization);
  656. bytestream2_put_be16(pby, s->plane[p].quantization[1 + l * 3 + i]);
  657. bytestream2_put_be16(pby, BandScale);
  658. bytestream2_put_be16(pby, 1);
  659. bytestream2_put_be16(pby, BandHeader);
  660. bytestream2_put_be16(pby, 0);
  661. quantize_band(data, width, a_width, height,
  662. s->plane[p].quantization[1 + l * 3 + i]);
  663. init_put_bits(pb, pkt->data + bytestream2_tell_p(pby), bytestream2_get_bytes_left_p(pby));
  664. for (int m = 0; m < height; m++) {
  665. for (int j = 0; j < stride; j++) {
  666. int16_t index = FFSIGN(data[j]) * lut[FFABS(data[j])];
  667. if (index < 0)
  668. index += 512;
  669. if (index == 0) {
  670. count++;
  671. continue;
  672. } else if (count > 0) {
  673. count = put_runcode(pb, count, rb);
  674. }
  675. put_bits(pb, cb[index].size, cb[index].bits);
  676. }
  677. data += a_width;
  678. }
  679. if (count > 0) {
  680. count = put_runcode(pb, count, rb);
  681. }
  682. put_bits(pb, cb[512].size, cb[512].bits);
  683. flush_put_bits(pb);
  684. bytestream2_skip_p(pby, put_bits_count(pb) >> 3);
  685. padd = (4 - (bytestream2_tell_p(pby) & 3)) & 3;
  686. while (padd--)
  687. bytestream2_put_byte(pby, 0);
  688. bytestream2_put_be16(pby, BandTrailer);
  689. bytestream2_put_be16(pby, 0);
  690. }
  691. bytestream2_put_be16(pby, BitstreamMarker);
  692. bytestream2_put_be16(pby, 0x0c0c);
  693. }
  694. s->plane[p].size = bytestream2_tell_p(pby) - pos;
  695. }
  696. bytestream2_put_be16(pby, GroupTrailer);
  697. bytestream2_put_be16(pby, 0);
  698. av_shrink_packet(pkt, bytestream2_tell_p(pby));
  699. pkt->flags |= AV_PKT_FLAG_KEY;
  700. bytestream2_seek_p(pby, 8, SEEK_SET);
  701. for (int i = 0; i < s->planes; i++)
  702. bytestream2_put_be32(pby, s->plane[i].size);
  703. *got_packet = 1;
  704. return 0;
  705. }
  706. static av_cold int cfhd_encode_close(AVCodecContext *avctx)
  707. {
  708. CFHDEncContext *s = avctx->priv_data;
  709. for (int i = 0; i < s->planes; i++) {
  710. av_freep(&s->plane[i].dwt_buf);
  711. av_freep(&s->plane[i].dwt_tmp);
  712. for (int j = 0; j < SUBBAND_COUNT; j++)
  713. s->plane[i].subband[j] = NULL;
  714. for (int j = 0; j < 8; j++)
  715. s->plane[i].l_h[j] = NULL;
  716. }
  717. av_freep(&s->alpha);
  718. return 0;
  719. }
  720. #define OFFSET(x) offsetof(CFHDEncContext, x)
  721. #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
  722. static const AVOption options[] = {
  723. { "quality", "set quality", OFFSET(quality), AV_OPT_TYPE_INT, {.i64= 0}, 0, 12, VE, "q" },
  724. { "film3+", NULL, 0, AV_OPT_TYPE_CONST, {.i64= 0}, 0, 0, VE, "q" },
  725. { "film3", NULL, 0, AV_OPT_TYPE_CONST, {.i64= 1}, 0, 0, VE, "q" },
  726. { "film2+", NULL, 0, AV_OPT_TYPE_CONST, {.i64= 2}, 0, 0, VE, "q" },
  727. { "film2", NULL, 0, AV_OPT_TYPE_CONST, {.i64= 3}, 0, 0, VE, "q" },
  728. { "film1.5", NULL, 0, AV_OPT_TYPE_CONST, {.i64= 4}, 0, 0, VE, "q" },
  729. { "film1+", NULL, 0, AV_OPT_TYPE_CONST, {.i64= 5}, 0, 0, VE, "q" },
  730. { "film1", NULL, 0, AV_OPT_TYPE_CONST, {.i64= 6}, 0, 0, VE, "q" },
  731. { "high+", NULL, 0, AV_OPT_TYPE_CONST, {.i64= 7}, 0, 0, VE, "q" },
  732. { "high", NULL, 0, AV_OPT_TYPE_CONST, {.i64= 8}, 0, 0, VE, "q" },
  733. { "medium+", NULL, 0, AV_OPT_TYPE_CONST, {.i64= 9}, 0, 0, VE, "q" },
  734. { "medium", NULL, 0, AV_OPT_TYPE_CONST, {.i64=10}, 0, 0, VE, "q" },
  735. { "low+", NULL, 0, AV_OPT_TYPE_CONST, {.i64=11}, 0, 0, VE, "q" },
  736. { "low", NULL, 0, AV_OPT_TYPE_CONST, {.i64=12}, 0, 0, VE, "q" },
  737. { NULL},
  738. };
  739. static const AVClass cfhd_class = {
  740. .class_name = "cfhd",
  741. .item_name = av_default_item_name,
  742. .option = options,
  743. .version = LIBAVUTIL_VERSION_INT,
  744. };
  745. AVCodec ff_cfhd_encoder = {
  746. .name = "cfhd",
  747. .long_name = NULL_IF_CONFIG_SMALL("GoPro CineForm HD"),
  748. .type = AVMEDIA_TYPE_VIDEO,
  749. .id = AV_CODEC_ID_CFHD,
  750. .priv_data_size = sizeof(CFHDEncContext),
  751. .priv_class = &cfhd_class,
  752. .init = cfhd_encode_init,
  753. .close = cfhd_encode_close,
  754. .encode2 = cfhd_encode_frame,
  755. .capabilities = AV_CODEC_CAP_FRAME_THREADS,
  756. .pix_fmts = (const enum AVPixelFormat[]) {
  757. AV_PIX_FMT_YUV422P10,
  758. AV_PIX_FMT_GBRP12,
  759. AV_PIX_FMT_GBRAP12,
  760. AV_PIX_FMT_NONE
  761. },
  762. };