Go to most recent revision | Details | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line |
|---|---|---|---|
| 4349 | Serge | 1 | /* |
| 2 | * VC3/DNxHD encoder |
||
| 3 | * Copyright (c) 2007 Baptiste Coudurier |
||
| 4 | * Copyright (c) 2011 MirriAd Ltd |
||
| 5 | * |
||
| 6 | * VC-3 encoder funded by the British Broadcasting Corporation |
||
| 7 | * 10 bit support added by MirriAd Ltd, Joseph Artsimovich |
||
| 8 | * |
||
| 9 | * This file is part of FFmpeg. |
||
| 10 | * |
||
| 11 | * FFmpeg is free software; you can redistribute it and/or |
||
| 12 | * modify it under the terms of the GNU Lesser General Public |
||
| 13 | * License as published by the Free Software Foundation; either |
||
| 14 | * version 2.1 of the License, or (at your option) any later version. |
||
| 15 | * |
||
| 16 | * FFmpeg is distributed in the hope that it will be useful, |
||
| 17 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
||
| 18 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
||
| 19 | * Lesser General Public License for more details. |
||
| 20 | * |
||
| 21 | * You should have received a copy of the GNU Lesser General Public |
||
| 22 | * License along with FFmpeg; if not, write to the Free Software |
||
| 23 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA |
||
| 24 | */ |
||
| 25 | |||
| 26 | #define RC_VARIANCE 1 // use variance or ssd for fast rc |
||
| 27 | |||
| 28 | #include "libavutil/attributes.h" |
||
| 29 | #include "libavutil/internal.h" |
||
| 30 | #include "libavutil/opt.h" |
||
| 31 | #include "avcodec.h" |
||
| 32 | #include "dsputil.h" |
||
| 33 | #include "internal.h" |
||
| 34 | #include "mpegvideo.h" |
||
| 35 | #include "dnxhdenc.h" |
||
| 36 | |||
| 37 | #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM |
||
| 38 | #define DNX10BIT_QMAT_SHIFT 18 // The largest value that will not lead to overflow for 10bit samples. |
||
| 39 | |||
| 40 | static const AVOption options[]={ |
||
| 41 | {"nitris_compat", "encode with Avid Nitris compatibility", offsetof(DNXHDEncContext, nitris_compat), AV_OPT_TYPE_INT, {.i64 = 0}, 0, 1, VE}, |
||
| 42 | {NULL} |
||
| 43 | }; |
||
| 44 | |||
| 45 | static const AVClass dnxhd_class = { |
||
| 46 | .class_name = "dnxhd", |
||
| 47 | .item_name = av_default_item_name, |
||
| 48 | .option = options, |
||
| 49 | .version = LIBAVUTIL_VERSION_INT, |
||
| 50 | }; |
||
| 51 | |||
| 52 | #define LAMBDA_FRAC_BITS 10 |
||
| 53 | |||
| 54 | static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *av_restrict block, const uint8_t *pixels, int line_size) |
||
| 55 | { |
||
| 56 | int i; |
||
| 57 | for (i = 0; i < 4; i++) { |
||
| 58 | block[0] = pixels[0]; block[1] = pixels[1]; |
||
| 59 | block[2] = pixels[2]; block[3] = pixels[3]; |
||
| 60 | block[4] = pixels[4]; block[5] = pixels[5]; |
||
| 61 | block[6] = pixels[6]; block[7] = pixels[7]; |
||
| 62 | pixels += line_size; |
||
| 63 | block += 8; |
||
| 64 | } |
||
| 65 | memcpy(block, block - 8, sizeof(*block) * 8); |
||
| 66 | memcpy(block + 8, block - 16, sizeof(*block) * 8); |
||
| 67 | memcpy(block + 16, block - 24, sizeof(*block) * 8); |
||
| 68 | memcpy(block + 24, block - 32, sizeof(*block) * 8); |
||
| 69 | } |
||
| 70 | |||
| 71 | static av_always_inline void dnxhd_10bit_get_pixels_8x4_sym(int16_t *av_restrict block, const uint8_t *pixels, int line_size) |
||
| 72 | { |
||
| 73 | int i; |
||
| 74 | const uint16_t* pixels16 = (const uint16_t*)pixels; |
||
| 75 | line_size >>= 1; |
||
| 76 | |||
| 77 | for (i = 0; i < 4; i++) { |
||
| 78 | block[0] = pixels16[0]; block[1] = pixels16[1]; |
||
| 79 | block[2] = pixels16[2]; block[3] = pixels16[3]; |
||
| 80 | block[4] = pixels16[4]; block[5] = pixels16[5]; |
||
| 81 | block[6] = pixels16[6]; block[7] = pixels16[7]; |
||
| 82 | pixels16 += line_size; |
||
| 83 | block += 8; |
||
| 84 | } |
||
| 85 | memcpy(block, block - 8, sizeof(*block) * 8); |
||
| 86 | memcpy(block + 8, block - 16, sizeof(*block) * 8); |
||
| 87 | memcpy(block + 16, block - 24, sizeof(*block) * 8); |
||
| 88 | memcpy(block + 24, block - 32, sizeof(*block) * 8); |
||
| 89 | } |
||
| 90 | |||
| 91 | static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *block, |
||
| 92 | int n, int qscale, int *overflow) |
||
| 93 | { |
||
| 94 | const uint8_t *scantable= ctx->intra_scantable.scantable; |
||
| 95 | const int *qmat = n<4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale]; |
||
| 96 | int last_non_zero = 0; |
||
| 97 | int i; |
||
| 98 | |||
| 99 | ctx->dsp.fdct(block); |
||
| 100 | |||
| 101 | // Divide by 4 with rounding, to compensate scaling of DCT coefficients |
||
| 102 | block[0] = (block[0] + 2) >> 2; |
||
| 103 | |||
| 104 | for (i = 1; i < 64; ++i) { |
||
| 105 | int j = scantable[i]; |
||
| 106 | int sign = block[j] >> 31; |
||
| 107 | int level = (block[j] ^ sign) - sign; |
||
| 108 | level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT; |
||
| 109 | block[j] = (level ^ sign) - sign; |
||
| 110 | if (level) |
||
| 111 | last_non_zero = i; |
||
| 112 | } |
||
| 113 | |||
| 114 | return last_non_zero; |
||
| 115 | } |
||
| 116 | |||
| 117 | static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx) |
||
| 118 | { |
||
| 119 | int i, j, level, run; |
||
| 120 | int max_level = 1<<(ctx->cid_table->bit_depth+2); |
||
| 121 | |||
| 122 | FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_codes, max_level*4*sizeof(*ctx->vlc_codes), fail); |
||
| 123 | FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_bits, max_level*4*sizeof(*ctx->vlc_bits) , fail); |
||
| 124 | FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes, 63*2, fail); |
||
| 125 | FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits, 63, fail); |
||
| 126 | |||
| 127 | ctx->vlc_codes += max_level*2; |
||
| 128 | ctx->vlc_bits += max_level*2; |
||
| 129 | for (level = -max_level; level < max_level; level++) { |
||
| 130 | for (run = 0; run < 2; run++) { |
||
| 131 | int index = (level<<1)|run; |
||
| 132 | int sign, offset = 0, alevel = level; |
||
| 133 | |||
| 134 | MASK_ABS(sign, alevel); |
||
| 135 | if (alevel > 64) { |
||
| 136 | offset = (alevel-1)>>6; |
||
| 137 | alevel -= offset<<6; |
||
| 138 | } |
||
| 139 | for (j = 0; j < 257; j++) { |
||
| 140 | if (ctx->cid_table->ac_level[j] >> 1 == alevel && |
||
| 141 | (!offset || (ctx->cid_table->ac_flags[j] & 1) && offset) && |
||
| 142 | (!run || (ctx->cid_table->ac_flags[j] & 2) && run)) { |
||
| 143 | av_assert1(!ctx->vlc_codes[index]); |
||
| 144 | if (alevel) { |
||
| 145 | ctx->vlc_codes[index] = (ctx->cid_table->ac_codes[j]<<1)|(sign&1); |
||
| 146 | ctx->vlc_bits [index] = ctx->cid_table->ac_bits[j]+1; |
||
| 147 | } else { |
||
| 148 | ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j]; |
||
| 149 | ctx->vlc_bits [index] = ctx->cid_table->ac_bits [j]; |
||
| 150 | } |
||
| 151 | break; |
||
| 152 | } |
||
| 153 | } |
||
| 154 | av_assert0(!alevel || j < 257); |
||
| 155 | if (offset) { |
||
| 156 | ctx->vlc_codes[index] = (ctx->vlc_codes[index]< |
||
| 157 | ctx->vlc_bits [index]+= ctx->cid_table->index_bits; |
||
| 158 | } |
||
| 159 | } |
||
| 160 | } |
||
| 161 | for (i = 0; i < 62; i++) { |
||
| 162 | int run = ctx->cid_table->run[i]; |
||
| 163 | av_assert0(run < 63); |
||
| 164 | ctx->run_codes[run] = ctx->cid_table->run_codes[i]; |
||
| 165 | ctx->run_bits [run] = ctx->cid_table->run_bits[i]; |
||
| 166 | } |
||
| 167 | return 0; |
||
| 168 | fail: |
||
| 169 | return -1; |
||
| 170 | } |
||
| 171 | |||
| 172 | static av_cold int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias) |
||
| 173 | { |
||
| 174 | // init first elem to 1 to avoid div by 0 in convert_matrix |
||
| 175 | uint16_t weight_matrix[64] = {1,}; // convert_matrix needs uint16_t* |
||
| 176 | int qscale, i; |
||
| 177 | const uint8_t *luma_weight_table = ctx->cid_table->luma_weight; |
||
| 178 | const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight; |
||
| 179 | |||
| 180 | FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l, (ctx->m.avctx->qmax+1) * 64 * sizeof(int), fail); |
||
| 181 | FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c, (ctx->m.avctx->qmax+1) * 64 * sizeof(int), fail); |
||
| 182 | FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t), fail); |
||
| 183 | FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t), fail); |
||
| 184 | |||
| 185 | if (ctx->cid_table->bit_depth == 8) { |
||
| 186 | for (i = 1; i < 64; i++) { |
||
| 187 | int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]]; |
||
| 188 | weight_matrix[j] = ctx->cid_table->luma_weight[i]; |
||
| 189 | } |
||
| 190 | ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_l, ctx->qmatrix_l16, weight_matrix, |
||
| 191 | ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1); |
||
| 192 | for (i = 1; i < 64; i++) { |
||
| 193 | int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]]; |
||
| 194 | weight_matrix[j] = ctx->cid_table->chroma_weight[i]; |
||
| 195 | } |
||
| 196 | ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_c, ctx->qmatrix_c16, weight_matrix, |
||
| 197 | ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1); |
||
| 198 | |||
| 199 | for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) { |
||
| 200 | for (i = 0; i < 64; i++) { |
||
| 201 | ctx->qmatrix_l [qscale] [i] <<= 2; ctx->qmatrix_c [qscale] [i] <<= 2; |
||
| 202 | ctx->qmatrix_l16[qscale][0][i] <<= 2; ctx->qmatrix_l16[qscale][1][i] <<= 2; |
||
| 203 | ctx->qmatrix_c16[qscale][0][i] <<= 2; ctx->qmatrix_c16[qscale][1][i] <<= 2; |
||
| 204 | } |
||
| 205 | } |
||
| 206 | } else { |
||
| 207 | // 10-bit |
||
| 208 | for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) { |
||
| 209 | for (i = 1; i < 64; i++) { |
||
| 210 | int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]]; |
||
| 211 | |||
| 212 | // The quantization formula from the VC-3 standard is: |
||
| 213 | // quantized = sign(block[i]) * floor(abs(block[i]/s) * p / (qscale * weight_table[i])) |
||
| 214 | // Where p is 32 for 8-bit samples and 8 for 10-bit ones. |
||
| 215 | // The s factor compensates scaling of DCT coefficients done by the DCT routines, |
||
| 216 | // and therefore is not present in standard. It's 8 for 8-bit samples and 4 for 10-bit ones. |
||
| 217 | // We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be: |
||
| 218 | // ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) / (qscale * weight_table[i]) |
||
| 219 | // For 10-bit samples, p / s == 2 |
||
| 220 | ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) / (qscale * luma_weight_table[i]); |
||
| 221 | ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) / (qscale * chroma_weight_table[i]); |
||
| 222 | } |
||
| 223 | } |
||
| 224 | } |
||
| 225 | |||
| 226 | ctx->m.q_chroma_intra_matrix16 = ctx->qmatrix_c16; |
||
| 227 | ctx->m.q_chroma_intra_matrix = ctx->qmatrix_c; |
||
| 228 | ctx->m.q_intra_matrix16 = ctx->qmatrix_l16; |
||
| 229 | ctx->m.q_intra_matrix = ctx->qmatrix_l; |
||
| 230 | |||
| 231 | return 0; |
||
| 232 | fail: |
||
| 233 | return -1; |
||
| 234 | } |
||
| 235 | |||
| 236 | static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx) |
||
| 237 | { |
||
| 238 | FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_rc, 8160*ctx->m.avctx->qmax*sizeof(RCEntry), fail); |
||
| 239 | if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD) |
||
| 240 | FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_cmp, ctx->m.mb_num*sizeof(RCCMPEntry), fail); |
||
| 241 | |||
| 242 | ctx->frame_bits = (ctx->cid_table->coding_unit_size - 640 - 4 - ctx->min_padding) * 8; |
||
| 243 | ctx->qscale = 1; |
||
| 244 | ctx->lambda = 2< |
||
| 245 | return 0; |
||
| 246 | fail: |
||
| 247 | return -1; |
||
| 248 | } |
||
| 249 | |||
| 250 | static av_cold int dnxhd_encode_init(AVCodecContext *avctx) |
||
| 251 | { |
||
| 252 | DNXHDEncContext *ctx = avctx->priv_data; |
||
| 253 | int i, index, bit_depth; |
||
| 254 | |||
| 255 | switch (avctx->pix_fmt) { |
||
| 256 | case AV_PIX_FMT_YUV422P: |
||
| 257 | bit_depth = 8; |
||
| 258 | break; |
||
| 259 | case AV_PIX_FMT_YUV422P10: |
||
| 260 | bit_depth = 10; |
||
| 261 | break; |
||
| 262 | default: |
||
| 263 | av_log(avctx, AV_LOG_ERROR, "pixel format is incompatible with DNxHD\n"); |
||
| 264 | return -1; |
||
| 265 | } |
||
| 266 | |||
| 267 | ctx->cid = ff_dnxhd_find_cid(avctx, bit_depth); |
||
| 268 | if (!ctx->cid) { |
||
| 269 | av_log(avctx, AV_LOG_ERROR, "video parameters incompatible with DNxHD. Valid DNxHD profiles:\n"); |
||
| 270 | ff_dnxhd_print_profiles(avctx, AV_LOG_ERROR); |
||
| 271 | return -1; |
||
| 272 | } |
||
| 273 | av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid); |
||
| 274 | |||
| 275 | index = ff_dnxhd_get_cid_table(ctx->cid); |
||
| 276 | av_assert0(index >= 0); |
||
| 277 | ctx->cid_table = &ff_dnxhd_cid_table[index]; |
||
| 278 | |||
| 279 | ctx->m.avctx = avctx; |
||
| 280 | ctx->m.mb_intra = 1; |
||
| 281 | ctx->m.h263_aic = 1; |
||
| 282 | |||
| 283 | avctx->bits_per_raw_sample = ctx->cid_table->bit_depth; |
||
| 284 | |||
| 285 | ff_dct_common_init(&ctx->m); |
||
| 286 | ff_dct_encode_init(&ctx->m); |
||
| 287 | |||
| 288 | if (!ctx->m.dct_quantize) |
||
| 289 | ctx->m.dct_quantize = ff_dct_quantize_c; |
||
| 290 | |||
| 291 | if (ctx->cid_table->bit_depth == 10) { |
||
| 292 | ctx->m.dct_quantize = dnxhd_10bit_dct_quantize; |
||
| 293 | ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym; |
||
| 294 | ctx->block_width_l2 = 4; |
||
| 295 | } else { |
||
| 296 | ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym; |
||
| 297 | ctx->block_width_l2 = 3; |
||
| 298 | } |
||
| 299 | |||
| 300 | if (ARCH_X86) |
||
| 301 | ff_dnxhdenc_init_x86(ctx); |
||
| 302 | |||
| 303 | ctx->m.mb_height = (avctx->height + 15) / 16; |
||
| 304 | ctx->m.mb_width = (avctx->width + 15) / 16; |
||
| 305 | |||
| 306 | if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) { |
||
| 307 | ctx->interlaced = 1; |
||
| 308 | ctx->m.mb_height /= 2; |
||
| 309 | } |
||
| 310 | |||
| 311 | ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width; |
||
| 312 | |||
| 313 | if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS) |
||
| 314 | ctx->m.intra_quant_bias = avctx->intra_quant_bias; |
||
| 315 | if (dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0) < 0) // XXX tune lbias/cbias |
||
| 316 | return -1; |
||
| 317 | |||
| 318 | // Avid Nitris hardware decoder requires a minimum amount of padding in the coding unit payload |
||
| 319 | if (ctx->nitris_compat) |
||
| 320 | ctx->min_padding = 1600; |
||
| 321 | |||
| 322 | if (dnxhd_init_vlc(ctx) < 0) |
||
| 323 | return -1; |
||
| 324 | if (dnxhd_init_rc(ctx) < 0) |
||
| 325 | return -1; |
||
| 326 | |||
| 327 | FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t), fail); |
||
| 328 | FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs, ctx->m.mb_height*sizeof(uint32_t), fail); |
||
| 329 | FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits, ctx->m.mb_num *sizeof(uint16_t), fail); |
||
| 330 | FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale, ctx->m.mb_num *sizeof(uint8_t), fail); |
||
| 331 | |||
| 332 | ctx->frame.key_frame = 1; |
||
| 333 | ctx->frame.pict_type = AV_PICTURE_TYPE_I; |
||
| 334 | ctx->m.avctx->coded_frame = &ctx->frame; |
||
| 335 | |||
| 336 | if (avctx->thread_count > MAX_THREADS) { |
||
| 337 | av_log(avctx, AV_LOG_ERROR, "too many threads\n"); |
||
| 338 | return -1; |
||
| 339 | } |
||
| 340 | |||
| 341 | ctx->thread[0] = ctx; |
||
| 342 | for (i = 1; i < avctx->thread_count; i++) { |
||
| 343 | ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext)); |
||
| 344 | memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext)); |
||
| 345 | } |
||
| 346 | |||
| 347 | return 0; |
||
| 348 | fail: //for FF_ALLOCZ_OR_GOTO |
||
| 349 | return -1; |
||
| 350 | } |
||
| 351 | |||
| 352 | static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf) |
||
| 353 | { |
||
| 354 | DNXHDEncContext *ctx = avctx->priv_data; |
||
| 355 | static const uint8_t header_prefix[5] = { 0x00,0x00,0x02,0x80,0x01 }; |
||
| 356 | |||
| 357 | memset(buf, 0, 640); |
||
| 358 | |||
| 359 | memcpy(buf, header_prefix, 5); |
||
| 360 | buf[5] = ctx->interlaced ? ctx->cur_field+2 : 0x01; |
||
| 361 | buf[6] = 0x80; // crc flag off |
||
| 362 | buf[7] = 0xa0; // reserved |
||
| 363 | AV_WB16(buf + 0x18, avctx->height>>ctx->interlaced); // ALPF |
||
| 364 | AV_WB16(buf + 0x1a, avctx->width); // SPL |
||
| 365 | AV_WB16(buf + 0x1d, avctx->height>>ctx->interlaced); // NAL |
||
| 366 | |||
| 367 | buf[0x21] = ctx->cid_table->bit_depth == 10 ? 0x58 : 0x38; |
||
| 368 | buf[0x22] = 0x88 + (ctx->interlaced<<2); |
||
| 369 | AV_WB32(buf + 0x28, ctx->cid); // CID |
||
| 370 | buf[0x2c] = ctx->interlaced ? 0 : 0x80; |
||
| 371 | |||
| 372 | buf[0x5f] = 0x01; // UDL |
||
| 373 | |||
| 374 | buf[0x167] = 0x02; // reserved |
||
| 375 | AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4); // MSIPS |
||
| 376 | buf[0x16d] = ctx->m.mb_height; // Ns |
||
| 377 | buf[0x16f] = 0x10; // reserved |
||
| 378 | |||
| 379 | ctx->msip = buf + 0x170; |
||
| 380 | return 0; |
||
| 381 | } |
||
| 382 | |||
| 383 | static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff) |
||
| 384 | { |
||
| 385 | int nbits; |
||
| 386 | if (diff < 0) { |
||
| 387 | nbits = av_log2_16bit(-2*diff); |
||
| 388 | diff--; |
||
| 389 | } else { |
||
| 390 | nbits = av_log2_16bit(2*diff); |
||
| 391 | } |
||
| 392 | put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits, |
||
| 393 | (ctx->cid_table->dc_codes[nbits]< |
||
| 394 | } |
||
| 395 | |||
| 396 | static av_always_inline void dnxhd_encode_block(DNXHDEncContext *ctx, int16_t *block, int last_index, int n) |
||
| 397 | { |
||
| 398 | int last_non_zero = 0; |
||
| 399 | int slevel, i, j; |
||
| 400 | |||
| 401 | dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]); |
||
| 402 | ctx->m.last_dc[n] = block[0]; |
||
| 403 | |||
| 404 | for (i = 1; i <= last_index; i++) { |
||
| 405 | j = ctx->m.intra_scantable.permutated[i]; |
||
| 406 | slevel = block[j]; |
||
| 407 | if (slevel) { |
||
| 408 | int run_level = i - last_non_zero - 1; |
||
| 409 | int rlevel = (slevel<<1)|!!run_level; |
||
| 410 | put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]); |
||
| 411 | if (run_level) |
||
| 412 | put_bits(&ctx->m.pb, ctx->run_bits[run_level], ctx->run_codes[run_level]); |
||
| 413 | last_non_zero = i; |
||
| 414 | } |
||
| 415 | } |
||
| 416 | put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB |
||
| 417 | } |
||
| 418 | |||
| 419 | static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n, int qscale, int last_index) |
||
| 420 | { |
||
| 421 | const uint8_t *weight_matrix; |
||
| 422 | int level; |
||
| 423 | int i; |
||
| 424 | |||
| 425 | weight_matrix = (n&2) ? ctx->cid_table->chroma_weight : ctx->cid_table->luma_weight; |
||
| 426 | |||
| 427 | for (i = 1; i <= last_index; i++) { |
||
| 428 | int j = ctx->m.intra_scantable.permutated[i]; |
||
| 429 | level = block[j]; |
||
| 430 | if (level) { |
||
| 431 | if (level < 0) { |
||
| 432 | level = (1-2*level) * qscale * weight_matrix[i]; |
||
| 433 | if (ctx->cid_table->bit_depth == 10) { |
||
| 434 | if (weight_matrix[i] != 8) |
||
| 435 | level += 8; |
||
| 436 | level >>= 4; |
||
| 437 | } else { |
||
| 438 | if (weight_matrix[i] != 32) |
||
| 439 | level += 32; |
||
| 440 | level >>= 6; |
||
| 441 | } |
||
| 442 | level = -level; |
||
| 443 | } else { |
||
| 444 | level = (2*level+1) * qscale * weight_matrix[i]; |
||
| 445 | if (ctx->cid_table->bit_depth == 10) { |
||
| 446 | if (weight_matrix[i] != 8) |
||
| 447 | level += 8; |
||
| 448 | level >>= 4; |
||
| 449 | } else { |
||
| 450 | if (weight_matrix[i] != 32) |
||
| 451 | level += 32; |
||
| 452 | level >>= 6; |
||
| 453 | } |
||
| 454 | } |
||
| 455 | block[j] = level; |
||
| 456 | } |
||
| 457 | } |
||
| 458 | } |
||
| 459 | |||
| 460 | static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *block) |
||
| 461 | { |
||
| 462 | int score = 0; |
||
| 463 | int i; |
||
| 464 | for (i = 0; i < 64; i++) |
||
| 465 | score += (block[i] - qblock[i]) * (block[i] - qblock[i]); |
||
| 466 | return score; |
||
| 467 | } |
||
| 468 | |||
| 469 | static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index) |
||
| 470 | { |
||
| 471 | int last_non_zero = 0; |
||
| 472 | int bits = 0; |
||
| 473 | int i, j, level; |
||
| 474 | for (i = 1; i <= last_index; i++) { |
||
| 475 | j = ctx->m.intra_scantable.permutated[i]; |
||
| 476 | level = block[j]; |
||
| 477 | if (level) { |
||
| 478 | int run_level = i - last_non_zero - 1; |
||
| 479 | bits += ctx->vlc_bits[(level<<1)|!!run_level]+ctx->run_bits[run_level]; |
||
| 480 | last_non_zero = i; |
||
| 481 | } |
||
| 482 | } |
||
| 483 | return bits; |
||
| 484 | } |
||
| 485 | |||
| 486 | static av_always_inline void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y) |
||
| 487 | { |
||
| 488 | const int bs = ctx->block_width_l2; |
||
| 489 | const int bw = 1 << bs; |
||
| 490 | const uint8_t *ptr_y = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs+1); |
||
| 491 | const uint8_t *ptr_u = ctx->thread[0]->src[1] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs); |
||
| 492 | const uint8_t *ptr_v = ctx->thread[0]->src[2] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs); |
||
| 493 | DSPContext *dsp = &ctx->m.dsp; |
||
| 494 | |||
| 495 | dsp->get_pixels(ctx->blocks[0], ptr_y, ctx->m.linesize); |
||
| 496 | dsp->get_pixels(ctx->blocks[1], ptr_y + bw, ctx->m.linesize); |
||
| 497 | dsp->get_pixels(ctx->blocks[2], ptr_u, ctx->m.uvlinesize); |
||
| 498 | dsp->get_pixels(ctx->blocks[3], ptr_v, ctx->m.uvlinesize); |
||
| 499 | |||
| 500 | if (mb_y+1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) { |
||
| 501 | if (ctx->interlaced) { |
||
| 502 | ctx->get_pixels_8x4_sym(ctx->blocks[4], ptr_y + ctx->dct_y_offset, ctx->m.linesize); |
||
| 503 | ctx->get_pixels_8x4_sym(ctx->blocks[5], ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize); |
||
| 504 | ctx->get_pixels_8x4_sym(ctx->blocks[6], ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize); |
||
| 505 | ctx->get_pixels_8x4_sym(ctx->blocks[7], ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize); |
||
| 506 | } else { |
||
| 507 | dsp->clear_block(ctx->blocks[4]); |
||
| 508 | dsp->clear_block(ctx->blocks[5]); |
||
| 509 | dsp->clear_block(ctx->blocks[6]); |
||
| 510 | dsp->clear_block(ctx->blocks[7]); |
||
| 511 | } |
||
| 512 | } else { |
||
| 513 | dsp->get_pixels(ctx->blocks[4], ptr_y + ctx->dct_y_offset, ctx->m.linesize); |
||
| 514 | dsp->get_pixels(ctx->blocks[5], ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize); |
||
| 515 | dsp->get_pixels(ctx->blocks[6], ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize); |
||
| 516 | dsp->get_pixels(ctx->blocks[7], ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize); |
||
| 517 | } |
||
| 518 | } |
||
| 519 | |||
| 520 | static av_always_inline int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i) |
||
| 521 | { |
||
| 522 | const static uint8_t component[8]={0,0,1,2,0,0,1,2}; |
||
| 523 | return component[i]; |
||
| 524 | } |
||
| 525 | |||
| 526 | static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr) |
||
| 527 | { |
||
| 528 | DNXHDEncContext *ctx = avctx->priv_data; |
||
| 529 | int mb_y = jobnr, mb_x; |
||
| 530 | int qscale = ctx->qscale; |
||
| 531 | LOCAL_ALIGNED_16(int16_t, block, [64]); |
||
| 532 | ctx = ctx->thread[threadnr]; |
||
| 533 | |||
| 534 | ctx->m.last_dc[0] = |
||
| 535 | ctx->m.last_dc[1] = |
||
| 536 | ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2); |
||
| 537 | |||
| 538 | for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) { |
||
| 539 | unsigned mb = mb_y * ctx->m.mb_width + mb_x; |
||
| 540 | int ssd = 0; |
||
| 541 | int ac_bits = 0; |
||
| 542 | int dc_bits = 0; |
||
| 543 | int i; |
||
| 544 | |||
| 545 | dnxhd_get_blocks(ctx, mb_x, mb_y); |
||
| 546 | |||
| 547 | for (i = 0; i < 8; i++) { |
||
| 548 | int16_t *src_block = ctx->blocks[i]; |
||
| 549 | int overflow, nbits, diff, last_index; |
||
| 550 | int n = dnxhd_switch_matrix(ctx, i); |
||
| 551 | |||
| 552 | memcpy(block, src_block, 64*sizeof(*block)); |
||
| 553 | last_index = ctx->m.dct_quantize(&ctx->m, block, 4&(2*i), qscale, &overflow); |
||
| 554 | ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index); |
||
| 555 | |||
| 556 | diff = block[0] - ctx->m.last_dc[n]; |
||
| 557 | if (diff < 0) nbits = av_log2_16bit(-2*diff); |
||
| 558 | else nbits = av_log2_16bit( 2*diff); |
||
| 559 | |||
| 560 | av_assert1(nbits < ctx->cid_table->bit_depth + 4); |
||
| 561 | dc_bits += ctx->cid_table->dc_bits[nbits] + nbits; |
||
| 562 | |||
| 563 | ctx->m.last_dc[n] = block[0]; |
||
| 564 | |||
| 565 | if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) { |
||
| 566 | dnxhd_unquantize_c(ctx, block, i, qscale, last_index); |
||
| 567 | ctx->m.dsp.idct(block); |
||
| 568 | ssd += dnxhd_ssd_block(block, src_block); |
||
| 569 | } |
||
| 570 | } |
||
| 571 | ctx->mb_rc[qscale][mb].ssd = ssd; |
||
| 572 | ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->vlc_bits[0]; |
||
| 573 | } |
||
| 574 | return 0; |
||
| 575 | } |
||
| 576 | |||
| 577 | static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr) |
||
| 578 | { |
||
| 579 | DNXHDEncContext *ctx = avctx->priv_data; |
||
| 580 | int mb_y = jobnr, mb_x; |
||
| 581 | ctx = ctx->thread[threadnr]; |
||
| 582 | init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr], ctx->slice_size[jobnr]); |
||
| 583 | |||
| 584 | ctx->m.last_dc[0] = |
||
| 585 | ctx->m.last_dc[1] = |
||
| 586 | ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2); |
||
| 587 | for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) { |
||
| 588 | unsigned mb = mb_y * ctx->m.mb_width + mb_x; |
||
| 589 | int qscale = ctx->mb_qscale[mb]; |
||
| 590 | int i; |
||
| 591 | |||
| 592 | put_bits(&ctx->m.pb, 12, qscale<<1); |
||
| 593 | |||
| 594 | dnxhd_get_blocks(ctx, mb_x, mb_y); |
||
| 595 | |||
| 596 | for (i = 0; i < 8; i++) { |
||
| 597 | int16_t *block = ctx->blocks[i]; |
||
| 598 | int overflow, n = dnxhd_switch_matrix(ctx, i); |
||
| 599 | int last_index = ctx->m.dct_quantize(&ctx->m, block, 4&(2*i), qscale, &overflow); |
||
| 600 | //START_TIMER; |
||
| 601 | dnxhd_encode_block(ctx, block, last_index, n); |
||
| 602 | //STOP_TIMER("encode_block"); |
||
| 603 | } |
||
| 604 | } |
||
| 605 | if (put_bits_count(&ctx->m.pb)&31) |
||
| 606 | put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0); |
||
| 607 | flush_put_bits(&ctx->m.pb); |
||
| 608 | return 0; |
||
| 609 | } |
||
| 610 | |||
| 611 | static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx) |
||
| 612 | { |
||
| 613 | int mb_y, mb_x; |
||
| 614 | int offset = 0; |
||
| 615 | for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) { |
||
| 616 | int thread_size; |
||
| 617 | ctx->slice_offs[mb_y] = offset; |
||
| 618 | ctx->slice_size[mb_y] = 0; |
||
| 619 | for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) { |
||
| 620 | unsigned mb = mb_y * ctx->m.mb_width + mb_x; |
||
| 621 | ctx->slice_size[mb_y] += ctx->mb_bits[mb]; |
||
| 622 | } |
||
| 623 | ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31; |
||
| 624 | ctx->slice_size[mb_y] >>= 3; |
||
| 625 | thread_size = ctx->slice_size[mb_y]; |
||
| 626 | offset += thread_size; |
||
| 627 | } |
||
| 628 | } |
||
| 629 | |||
| 630 | static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr) |
||
| 631 | { |
||
| 632 | DNXHDEncContext *ctx = avctx->priv_data; |
||
| 633 | int mb_y = jobnr, mb_x, x, y; |
||
| 634 | int partial_last_row = (mb_y == ctx->m.mb_height - 1) && |
||
| 635 | ((avctx->height >> ctx->interlaced) & 0xF); |
||
| 636 | |||
| 637 | ctx = ctx->thread[threadnr]; |
||
| 638 | if (ctx->cid_table->bit_depth == 8) { |
||
| 639 | uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y<<4) * ctx->m.linesize); |
||
| 640 | for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) { |
||
| 641 | unsigned mb = mb_y * ctx->m.mb_width + mb_x; |
||
| 642 | int sum; |
||
| 643 | int varc; |
||
| 644 | |||
| 645 | if (!partial_last_row && mb_x * 16 <= avctx->width - 16) { |
||
| 646 | sum = ctx->m.dsp.pix_sum(pix, ctx->m.linesize); |
||
| 647 | varc = ctx->m.dsp.pix_norm1(pix, ctx->m.linesize); |
||
| 648 | } else { |
||
| 649 | int bw = FFMIN(avctx->width - 16 * mb_x, 16); |
||
| 650 | int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16); |
||
| 651 | sum = varc = 0; |
||
| 652 | for (y = 0; y < bh; y++) { |
||
| 653 | for (x = 0; x < bw; x++) { |
||
| 654 | uint8_t val = pix[x + y * ctx->m.linesize]; |
||
| 655 | sum += val; |
||
| 656 | varc += val * val; |
||
| 657 | } |
||
| 658 | } |
||
| 659 | } |
||
| 660 | varc = (varc - (((unsigned)sum * sum) >> 8) + 128) >> 8; |
||
| 661 | |||
| 662 | ctx->mb_cmp[mb].value = varc; |
||
| 663 | ctx->mb_cmp[mb].mb = mb; |
||
| 664 | } |
||
| 665 | } else { // 10-bit |
||
| 666 | int const linesize = ctx->m.linesize >> 1; |
||
| 667 | for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) { |
||
| 668 | uint16_t *pix = (uint16_t*)ctx->thread[0]->src[0] + ((mb_y << 4) * linesize) + (mb_x << 4); |
||
| 669 | unsigned mb = mb_y * ctx->m.mb_width + mb_x; |
||
| 670 | int sum = 0; |
||
| 671 | int sqsum = 0; |
||
| 672 | int mean, sqmean; |
||
| 673 | int i, j; |
||
| 674 | // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8. |
||
| 675 | for (i = 0; i < 16; ++i) { |
||
| 676 | for (j = 0; j < 16; ++j) { |
||
| 677 | // Turn 16-bit pixels into 10-bit ones. |
||
| 678 | int const sample = (unsigned)pix[j] >> 6; |
||
| 679 | sum += sample; |
||
| 680 | sqsum += sample * sample; |
||
| 681 | // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX |
||
| 682 | } |
||
| 683 | pix += linesize; |
||
| 684 | } |
||
| 685 | mean = sum >> 8; // 16*16 == 2^8 |
||
| 686 | sqmean = sqsum >> 8; |
||
| 687 | ctx->mb_cmp[mb].value = sqmean - mean * mean; |
||
| 688 | ctx->mb_cmp[mb].mb = mb; |
||
| 689 | } |
||
| 690 | } |
||
| 691 | return 0; |
||
| 692 | } |
||
| 693 | |||
| 694 | static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx) |
||
| 695 | { |
||
| 696 | int lambda, up_step, down_step; |
||
| 697 | int last_lower = INT_MAX, last_higher = 0; |
||
| 698 | int x, y, q; |
||
| 699 | |||
| 700 | for (q = 1; q < avctx->qmax; q++) { |
||
| 701 | ctx->qscale = q; |
||
| 702 | avctx->execute2(avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height); |
||
| 703 | } |
||
| 704 | up_step = down_step = 2< |
||
| 705 | lambda = ctx->lambda; |
||
| 706 | |||
| 707 | for (;;) { |
||
| 708 | int bits = 0; |
||
| 709 | int end = 0; |
||
| 710 | if (lambda == last_higher) { |
||
| 711 | lambda++; |
||
| 712 | end = 1; // need to set final qscales/bits |
||
| 713 | } |
||
| 714 | for (y = 0; y < ctx->m.mb_height; y++) { |
||
| 715 | for (x = 0; x < ctx->m.mb_width; x++) { |
||
| 716 | unsigned min = UINT_MAX; |
||
| 717 | int qscale = 1; |
||
| 718 | int mb = y*ctx->m.mb_width+x; |
||
| 719 | for (q = 1; q < avctx->qmax; q++) { |
||
| 720 | unsigned score = ctx->mb_rc[q][mb].bits*lambda+ |
||
| 721 | ((unsigned)ctx->mb_rc[q][mb].ssd< |
||
| 722 | if (score < min) { |
||
| 723 | min = score; |
||
| 724 | qscale = q; |
||
| 725 | } |
||
| 726 | } |
||
| 727 | bits += ctx->mb_rc[qscale][mb].bits; |
||
| 728 | ctx->mb_qscale[mb] = qscale; |
||
| 729 | ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits; |
||
| 730 | } |
||
| 731 | bits = (bits+31)&~31; // padding |
||
| 732 | if (bits > ctx->frame_bits) |
||
| 733 | break; |
||
| 734 | } |
||
| 735 | //av_dlog(ctx->m.avctx, "lambda %d, up %u, down %u, bits %d, frame %d\n", |
||
| 736 | // lambda, last_higher, last_lower, bits, ctx->frame_bits); |
||
| 737 | if (end) { |
||
| 738 | if (bits > ctx->frame_bits) |
||
| 739 | return -1; |
||
| 740 | break; |
||
| 741 | } |
||
| 742 | if (bits < ctx->frame_bits) { |
||
| 743 | last_lower = FFMIN(lambda, last_lower); |
||
| 744 | if (last_higher != 0) |
||
| 745 | lambda = (lambda+last_higher)>>1; |
||
| 746 | else |
||
| 747 | lambda -= down_step; |
||
| 748 | down_step = FFMIN((int64_t)down_step*5, INT_MAX); |
||
| 749 | up_step = 1< |
||
| 750 | lambda = FFMAX(1, lambda); |
||
| 751 | if (lambda == last_lower) |
||
| 752 | break; |
||
| 753 | } else { |
||
| 754 | last_higher = FFMAX(lambda, last_higher); |
||
| 755 | if (last_lower != INT_MAX) |
||
| 756 | lambda = (lambda+last_lower)>>1; |
||
| 757 | else if ((int64_t)lambda + up_step > INT_MAX) |
||
| 758 | return -1; |
||
| 759 | else |
||
| 760 | lambda += up_step; |
||
| 761 | up_step = FFMIN((int64_t)up_step*5, INT_MAX); |
||
| 762 | down_step = 1< |
||
| 763 | } |
||
| 764 | } |
||
| 765 | //av_dlog(ctx->m.avctx, "out lambda %d\n", lambda); |
||
| 766 | ctx->lambda = lambda; |
||
| 767 | return 0; |
||
| 768 | } |
||
| 769 | |||
| 770 | static int dnxhd_find_qscale(DNXHDEncContext *ctx) |
||
| 771 | { |
||
| 772 | int bits = 0; |
||
| 773 | int up_step = 1; |
||
| 774 | int down_step = 1; |
||
| 775 | int last_higher = 0; |
||
| 776 | int last_lower = INT_MAX; |
||
| 777 | int qscale; |
||
| 778 | int x, y; |
||
| 779 | |||
| 780 | qscale = ctx->qscale; |
||
| 781 | for (;;) { |
||
| 782 | bits = 0; |
||
| 783 | ctx->qscale = qscale; |
||
| 784 | // XXX avoid recalculating bits |
||
| 785 | ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height); |
||
| 786 | for (y = 0; y < ctx->m.mb_height; y++) { |
||
| 787 | for (x = 0; x < ctx->m.mb_width; x++) |
||
| 788 | bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits; |
||
| 789 | bits = (bits+31)&~31; // padding |
||
| 790 | if (bits > ctx->frame_bits) |
||
| 791 | break; |
||
| 792 | } |
||
| 793 | //av_dlog(ctx->m.avctx, "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n", |
||
| 794 | // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits, last_higher, last_lower); |
||
| 795 | if (bits < ctx->frame_bits) { |
||
| 796 | if (qscale == 1) |
||
| 797 | return 1; |
||
| 798 | if (last_higher == qscale - 1) { |
||
| 799 | qscale = last_higher; |
||
| 800 | break; |
||
| 801 | } |
||
| 802 | last_lower = FFMIN(qscale, last_lower); |
||
| 803 | if (last_higher != 0) |
||
| 804 | qscale = (qscale+last_higher)>>1; |
||
| 805 | else |
||
| 806 | qscale -= down_step++; |
||
| 807 | if (qscale < 1) |
||
| 808 | qscale = 1; |
||
| 809 | up_step = 1; |
||
| 810 | } else { |
||
| 811 | if (last_lower == qscale + 1) |
||
| 812 | break; |
||
| 813 | last_higher = FFMAX(qscale, last_higher); |
||
| 814 | if (last_lower != INT_MAX) |
||
| 815 | qscale = (qscale+last_lower)>>1; |
||
| 816 | else |
||
| 817 | qscale += up_step++; |
||
| 818 | down_step = 1; |
||
| 819 | if (qscale >= ctx->m.avctx->qmax) |
||
| 820 | return -1; |
||
| 821 | } |
||
| 822 | } |
||
| 823 | //av_dlog(ctx->m.avctx, "out qscale %d\n", qscale); |
||
| 824 | ctx->qscale = qscale; |
||
| 825 | return 0; |
||
| 826 | } |
||
| 827 | |||
| 828 | #define BUCKET_BITS 8 |
||
| 829 | #define RADIX_PASSES 4 |
||
| 830 | #define NBUCKETS (1 << BUCKET_BITS) |
||
| 831 | |||
| 832 | static inline int get_bucket(int value, int shift) |
||
| 833 | { |
||
| 834 | value >>= shift; |
||
| 835 | value &= NBUCKETS - 1; |
||
| 836 | return NBUCKETS - 1 - value; |
||
| 837 | } |
||
| 838 | |||
| 839 | static void radix_count(const RCCMPEntry *data, int size, int buckets[RADIX_PASSES][NBUCKETS]) |
||
| 840 | { |
||
| 841 | int i, j; |
||
| 842 | memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS); |
||
| 843 | for (i = 0; i < size; i++) { |
||
| 844 | int v = data[i].value; |
||
| 845 | for (j = 0; j < RADIX_PASSES; j++) { |
||
| 846 | buckets[j][get_bucket(v, 0)]++; |
||
| 847 | v >>= BUCKET_BITS; |
||
| 848 | } |
||
| 849 | av_assert1(!v); |
||
| 850 | } |
||
| 851 | for (j = 0; j < RADIX_PASSES; j++) { |
||
| 852 | int offset = size; |
||
| 853 | for (i = NBUCKETS - 1; i >= 0; i--) |
||
| 854 | buckets[j][i] = offset -= buckets[j][i]; |
||
| 855 | av_assert1(!buckets[j][0]); |
||
| 856 | } |
||
| 857 | } |
||
| 858 | |||
| 859 | static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data, int size, int buckets[NBUCKETS], int pass) |
||
| 860 | { |
||
| 861 | int shift = pass * BUCKET_BITS; |
||
| 862 | int i; |
||
| 863 | for (i = 0; i < size; i++) { |
||
| 864 | int v = get_bucket(data[i].value, shift); |
||
| 865 | int pos = buckets[v]++; |
||
| 866 | dst[pos] = data[i]; |
||
| 867 | } |
||
| 868 | } |
||
| 869 | |||
| 870 | static void radix_sort(RCCMPEntry *data, int size) |
||
| 871 | { |
||
| 872 | int buckets[RADIX_PASSES][NBUCKETS]; |
||
| 873 | RCCMPEntry *tmp = av_malloc(sizeof(*tmp) * size); |
||
| 874 | radix_count(data, size, buckets); |
||
| 875 | radix_sort_pass(tmp, data, size, buckets[0], 0); |
||
| 876 | radix_sort_pass(data, tmp, size, buckets[1], 1); |
||
| 877 | if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) { |
||
| 878 | radix_sort_pass(tmp, data, size, buckets[2], 2); |
||
| 879 | radix_sort_pass(data, tmp, size, buckets[3], 3); |
||
| 880 | } |
||
| 881 | av_free(tmp); |
||
| 882 | } |
||
| 883 | |||
| 884 | static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx) |
||
| 885 | { |
||
| 886 | int max_bits = 0; |
||
| 887 | int ret, x, y; |
||
| 888 | if ((ret = dnxhd_find_qscale(ctx)) < 0) |
||
| 889 | return -1; |
||
| 890 | for (y = 0; y < ctx->m.mb_height; y++) { |
||
| 891 | for (x = 0; x < ctx->m.mb_width; x++) { |
||
| 892 | int mb = y*ctx->m.mb_width+x; |
||
| 893 | int delta_bits; |
||
| 894 | ctx->mb_qscale[mb] = ctx->qscale; |
||
| 895 | ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits; |
||
| 896 | max_bits += ctx->mb_rc[ctx->qscale][mb].bits; |
||
| 897 | if (!RC_VARIANCE) { |
||
| 898 | delta_bits = ctx->mb_rc[ctx->qscale][mb].bits-ctx->mb_rc[ctx->qscale+1][mb].bits; |
||
| 899 | ctx->mb_cmp[mb].mb = mb; |
||
| 900 | ctx->mb_cmp[mb].value = delta_bits ? |
||
| 901 | ((ctx->mb_rc[ctx->qscale][mb].ssd-ctx->mb_rc[ctx->qscale+1][mb].ssd)*100)/delta_bits |
||
| 902 | : INT_MIN; //avoid increasing qscale |
||
| 903 | } |
||
| 904 | } |
||
| 905 | max_bits += 31; //worst padding |
||
| 906 | } |
||
| 907 | if (!ret) { |
||
| 908 | if (RC_VARIANCE) |
||
| 909 | avctx->execute2(avctx, dnxhd_mb_var_thread, NULL, NULL, ctx->m.mb_height); |
||
| 910 | radix_sort(ctx->mb_cmp, ctx->m.mb_num); |
||
| 911 | for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) { |
||
| 912 | int mb = ctx->mb_cmp[x].mb; |
||
| 913 | max_bits -= ctx->mb_rc[ctx->qscale][mb].bits - ctx->mb_rc[ctx->qscale+1][mb].bits; |
||
| 914 | ctx->mb_qscale[mb] = ctx->qscale+1; |
||
| 915 | ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale+1][mb].bits; |
||
| 916 | } |
||
| 917 | } |
||
| 918 | return 0; |
||
| 919 | } |
||
| 920 | |||
| 921 | static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame) |
||
| 922 | { |
||
| 923 | int i; |
||
| 924 | |||
| 925 | for (i = 0; i < 3; i++) { |
||
| 926 | ctx->frame.data[i] = frame->data[i]; |
||
| 927 | ctx->frame.linesize[i] = frame->linesize[i]; |
||
| 928 | } |
||
| 929 | |||
| 930 | for (i = 0; i < ctx->m.avctx->thread_count; i++) { |
||
| 931 | ctx->thread[i]->m.linesize = ctx->frame.linesize[0]< |
||
| 932 | ctx->thread[i]->m.uvlinesize = ctx->frame.linesize[1]< |
||
| 933 | ctx->thread[i]->dct_y_offset = ctx->m.linesize *8; |
||
| 934 | ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8; |
||
| 935 | } |
||
| 936 | |||
| 937 | ctx->frame.interlaced_frame = frame->interlaced_frame; |
||
| 938 | ctx->cur_field = frame->interlaced_frame && !frame->top_field_first; |
||
| 939 | } |
||
| 940 | |||
| 941 | static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt, |
||
| 942 | const AVFrame *frame, int *got_packet) |
||
| 943 | { |
||
| 944 | DNXHDEncContext *ctx = avctx->priv_data; |
||
| 945 | int first_field = 1; |
||
| 946 | int offset, i, ret; |
||
| 947 | uint8_t *buf; |
||
| 948 | |||
| 949 | if ((ret = ff_alloc_packet2(avctx, pkt, ctx->cid_table->frame_size)) < 0) |
||
| 950 | return ret; |
||
| 951 | buf = pkt->data; |
||
| 952 | |||
| 953 | dnxhd_load_picture(ctx, frame); |
||
| 954 | |||
| 955 | encode_coding_unit: |
||
| 956 | for (i = 0; i < 3; i++) { |
||
| 957 | ctx->src[i] = ctx->frame.data[i]; |
||
| 958 | if (ctx->interlaced && ctx->cur_field) |
||
| 959 | ctx->src[i] += ctx->frame.linesize[i]; |
||
| 960 | } |
||
| 961 | |||
| 962 | dnxhd_write_header(avctx, buf); |
||
| 963 | |||
| 964 | if (avctx->mb_decision == FF_MB_DECISION_RD) |
||
| 965 | ret = dnxhd_encode_rdo(avctx, ctx); |
||
| 966 | else |
||
| 967 | ret = dnxhd_encode_fast(avctx, ctx); |
||
| 968 | if (ret < 0) { |
||
| 969 | av_log(avctx, AV_LOG_ERROR, |
||
| 970 | "picture could not fit ratecontrol constraints, increase qmax\n"); |
||
| 971 | return -1; |
||
| 972 | } |
||
| 973 | |||
| 974 | dnxhd_setup_threads_slices(ctx); |
||
| 975 | |||
| 976 | offset = 0; |
||
| 977 | for (i = 0; i < ctx->m.mb_height; i++) { |
||
| 978 | AV_WB32(ctx->msip + i * 4, offset); |
||
| 979 | offset += ctx->slice_size[i]; |
||
| 980 | av_assert1(!(ctx->slice_size[i] & 3)); |
||
| 981 | } |
||
| 982 | |||
| 983 | avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height); |
||
| 984 | |||
| 985 | av_assert1(640 + offset + 4 <= ctx->cid_table->coding_unit_size); |
||
| 986 | memset(buf + 640 + offset, 0, ctx->cid_table->coding_unit_size - 4 - offset - 640); |
||
| 987 | |||
| 988 | AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF |
||
| 989 | |||
| 990 | if (ctx->interlaced && first_field) { |
||
| 991 | first_field = 0; |
||
| 992 | ctx->cur_field ^= 1; |
||
| 993 | buf += ctx->cid_table->coding_unit_size; |
||
| 994 | goto encode_coding_unit; |
||
| 995 | } |
||
| 996 | |||
| 997 | ctx->frame.quality = ctx->qscale*FF_QP2LAMBDA; |
||
| 998 | |||
| 999 | pkt->flags |= AV_PKT_FLAG_KEY; |
||
| 1000 | *got_packet = 1; |
||
| 1001 | return 0; |
||
| 1002 | } |
||
| 1003 | |||
| 1004 | static av_cold int dnxhd_encode_end(AVCodecContext *avctx) |
||
| 1005 | { |
||
| 1006 | DNXHDEncContext *ctx = avctx->priv_data; |
||
| 1007 | int max_level = 1<<(ctx->cid_table->bit_depth+2); |
||
| 1008 | int i; |
||
| 1009 | |||
| 1010 | av_free(ctx->vlc_codes-max_level*2); |
||
| 1011 | av_free(ctx->vlc_bits -max_level*2); |
||
| 1012 | av_freep(&ctx->run_codes); |
||
| 1013 | av_freep(&ctx->run_bits); |
||
| 1014 | |||
| 1015 | av_freep(&ctx->mb_bits); |
||
| 1016 | av_freep(&ctx->mb_qscale); |
||
| 1017 | av_freep(&ctx->mb_rc); |
||
| 1018 | av_freep(&ctx->mb_cmp); |
||
| 1019 | av_freep(&ctx->slice_size); |
||
| 1020 | av_freep(&ctx->slice_offs); |
||
| 1021 | |||
| 1022 | av_freep(&ctx->qmatrix_c); |
||
| 1023 | av_freep(&ctx->qmatrix_l); |
||
| 1024 | av_freep(&ctx->qmatrix_c16); |
||
| 1025 | av_freep(&ctx->qmatrix_l16); |
||
| 1026 | |||
| 1027 | for (i = 1; i < avctx->thread_count; i++) |
||
| 1028 | av_freep(&ctx->thread[i]); |
||
| 1029 | |||
| 1030 | return 0; |
||
| 1031 | } |
||
| 1032 | |||
| 1033 | static const AVCodecDefault dnxhd_defaults[] = { |
||
| 1034 | { "qmax", "1024" }, /* Maximum quantization scale factor allowed for VC-3 */ |
||
| 1035 | { NULL }, |
||
| 1036 | }; |
||
| 1037 | |||
| 1038 | AVCodec ff_dnxhd_encoder = { |
||
| 1039 | .name = "dnxhd", |
||
| 1040 | .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"), |
||
| 1041 | .type = AVMEDIA_TYPE_VIDEO, |
||
| 1042 | .id = AV_CODEC_ID_DNXHD, |
||
| 1043 | .priv_data_size = sizeof(DNXHDEncContext), |
||
| 1044 | .init = dnxhd_encode_init, |
||
| 1045 | .encode2 = dnxhd_encode_picture, |
||
| 1046 | .close = dnxhd_encode_end, |
||
| 1047 | .capabilities = CODEC_CAP_SLICE_THREADS, |
||
| 1048 | .pix_fmts = (const enum AVPixelFormat[]){ AV_PIX_FMT_YUV422P, |
||
| 1049 | AV_PIX_FMT_YUV422P10, |
||
| 1050 | AV_PIX_FMT_NONE }, |
||
| 1051 | .priv_class = &dnxhd_class, |
||
| 1052 | .defaults = dnxhd_defaults, |
||
| 1053 | }; |