00001
00002
00003
00004
00005
00006
00007
00008
00009
00010
00011
00012
00013
00014
00015
00016
00017
00018
00019
00020
00021
00022
00023
00024
00025
00026
00027 #define RC_VARIANCE 1 // use variance or ssd for fast rc
00028
00029 #include "libavutil/opt.h"
00030 #include "avcodec.h"
00031 #include "dsputil.h"
00032 #include "internal.h"
00033 #include "mpegvideo.h"
00034 #include "dnxhdenc.h"
00035 #include "internal.h"
00036
00037 #define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM
00038 #define DNX10BIT_QMAT_SHIFT 18 // The largest value that will not lead to overflow for 10bit samples.
00039
00040 static const AVOption options[]={
00041 {"nitris_compat", "encode with Avid Nitris compatibility", offsetof(DNXHDEncContext, nitris_compat), AV_OPT_TYPE_INT, {.i64 = 0}, 0, 1, VE},
00042 {NULL}
00043 };
00044
00045 static const AVClass class = {
00046 .class_name = "dnxhd",
00047 .item_name = av_default_item_name,
00048 .option = options,
00049 .version = LIBAVUTIL_VERSION_INT,
00050 };
00051
00052 #define LAMBDA_FRAC_BITS 10
00053
00054 static void dnxhd_8bit_get_pixels_8x4_sym(DCTELEM *av_restrict block, const uint8_t *pixels, int line_size)
00055 {
00056 int i;
00057 for (i = 0; i < 4; i++) {
00058 block[0] = pixels[0]; block[1] = pixels[1];
00059 block[2] = pixels[2]; block[3] = pixels[3];
00060 block[4] = pixels[4]; block[5] = pixels[5];
00061 block[6] = pixels[6]; block[7] = pixels[7];
00062 pixels += line_size;
00063 block += 8;
00064 }
00065 memcpy(block, block - 8, sizeof(*block) * 8);
00066 memcpy(block + 8, block - 16, sizeof(*block) * 8);
00067 memcpy(block + 16, block - 24, sizeof(*block) * 8);
00068 memcpy(block + 24, block - 32, sizeof(*block) * 8);
00069 }
00070
00071 static av_always_inline void dnxhd_10bit_get_pixels_8x4_sym(DCTELEM *av_restrict block, const uint8_t *pixels, int line_size)
00072 {
00073 int i;
00074 const uint16_t* pixels16 = (const uint16_t*)pixels;
00075 line_size >>= 1;
00076
00077 for (i = 0; i < 4; i++) {
00078 block[0] = pixels16[0]; block[1] = pixels16[1];
00079 block[2] = pixels16[2]; block[3] = pixels16[3];
00080 block[4] = pixels16[4]; block[5] = pixels16[5];
00081 block[6] = pixels16[6]; block[7] = pixels16[7];
00082 pixels16 += line_size;
00083 block += 8;
00084 }
00085 memcpy(block, block - 8, sizeof(*block) * 8);
00086 memcpy(block + 8, block - 16, sizeof(*block) * 8);
00087 memcpy(block + 16, block - 24, sizeof(*block) * 8);
00088 memcpy(block + 24, block - 32, sizeof(*block) * 8);
00089 }
00090
00091 static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, DCTELEM *block,
00092 int n, int qscale, int *overflow)
00093 {
00094 const uint8_t *scantable= ctx->intra_scantable.scantable;
00095 const int *qmat = n<4 ? ctx->q_intra_matrix[qscale] : ctx->q_chroma_intra_matrix[qscale];
00096 int last_non_zero = 0;
00097 int i;
00098
00099 ctx->dsp.fdct(block);
00100
00101
00102 block[0] = (block[0] + 2) >> 2;
00103
00104 for (i = 1; i < 64; ++i) {
00105 int j = scantable[i];
00106 int sign = block[j] >> 31;
00107 int level = (block[j] ^ sign) - sign;
00108 level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT;
00109 block[j] = (level ^ sign) - sign;
00110 if (level)
00111 last_non_zero = i;
00112 }
00113
00114 return last_non_zero;
00115 }
00116
00117 static int dnxhd_init_vlc(DNXHDEncContext *ctx)
00118 {
00119 int i, j, level, run;
00120 int max_level = 1<<(ctx->cid_table->bit_depth+2);
00121
00122 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_codes, max_level*4*sizeof(*ctx->vlc_codes), fail);
00123 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_bits, max_level*4*sizeof(*ctx->vlc_bits) , fail);
00124 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes, 63*2, fail);
00125 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits, 63, fail);
00126
00127 ctx->vlc_codes += max_level*2;
00128 ctx->vlc_bits += max_level*2;
00129 for (level = -max_level; level < max_level; level++) {
00130 for (run = 0; run < 2; run++) {
00131 int index = (level<<1)|run;
00132 int sign, offset = 0, alevel = level;
00133
00134 MASK_ABS(sign, alevel);
00135 if (alevel > 64) {
00136 offset = (alevel-1)>>6;
00137 alevel -= offset<<6;
00138 }
00139 for (j = 0; j < 257; j++) {
00140 if (ctx->cid_table->ac_level[j] >> 1 == alevel &&
00141 (!offset || (ctx->cid_table->ac_flags[j] & 1) && offset) &&
00142 (!run || (ctx->cid_table->ac_flags[j] & 2) && run)) {
00143 av_assert1(!ctx->vlc_codes[index]);
00144 if (alevel) {
00145 ctx->vlc_codes[index] = (ctx->cid_table->ac_codes[j]<<1)|(sign&1);
00146 ctx->vlc_bits [index] = ctx->cid_table->ac_bits[j]+1;
00147 } else {
00148 ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j];
00149 ctx->vlc_bits [index] = ctx->cid_table->ac_bits [j];
00150 }
00151 break;
00152 }
00153 }
00154 av_assert0(!alevel || j < 257);
00155 if (offset) {
00156 ctx->vlc_codes[index] = (ctx->vlc_codes[index]<<ctx->cid_table->index_bits)|offset;
00157 ctx->vlc_bits [index]+= ctx->cid_table->index_bits;
00158 }
00159 }
00160 }
00161 for (i = 0; i < 62; i++) {
00162 int run = ctx->cid_table->run[i];
00163 av_assert0(run < 63);
00164 ctx->run_codes[run] = ctx->cid_table->run_codes[i];
00165 ctx->run_bits [run] = ctx->cid_table->run_bits[i];
00166 }
00167 return 0;
00168 fail:
00169 return -1;
00170 }
00171
00172 static int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias)
00173 {
00174
00175 uint16_t weight_matrix[64] = {1,};
00176 int qscale, i;
00177 const uint8_t *luma_weight_table = ctx->cid_table->luma_weight;
00178 const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight;
00179
00180 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l, (ctx->m.avctx->qmax+1) * 64 * sizeof(int), fail);
00181 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c, (ctx->m.avctx->qmax+1) * 64 * sizeof(int), fail);
00182 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t), fail);
00183 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t), fail);
00184
00185 if (ctx->cid_table->bit_depth == 8) {
00186 for (i = 1; i < 64; i++) {
00187 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
00188 weight_matrix[j] = ctx->cid_table->luma_weight[i];
00189 }
00190 ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_l, ctx->qmatrix_l16, weight_matrix,
00191 ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
00192 for (i = 1; i < 64; i++) {
00193 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
00194 weight_matrix[j] = ctx->cid_table->chroma_weight[i];
00195 }
00196 ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_c, ctx->qmatrix_c16, weight_matrix,
00197 ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1);
00198
00199 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
00200 for (i = 0; i < 64; i++) {
00201 ctx->qmatrix_l [qscale] [i] <<= 2; ctx->qmatrix_c [qscale] [i] <<= 2;
00202 ctx->qmatrix_l16[qscale][0][i] <<= 2; ctx->qmatrix_l16[qscale][1][i] <<= 2;
00203 ctx->qmatrix_c16[qscale][0][i] <<= 2; ctx->qmatrix_c16[qscale][1][i] <<= 2;
00204 }
00205 }
00206 } else {
00207
00208 for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) {
00209 for (i = 1; i < 64; i++) {
00210 int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]];
00211
00212
00213
00214
00215
00216
00217
00218
00219
00220 ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) / (qscale * luma_weight_table[i]);
00221 ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) / (qscale * chroma_weight_table[i]);
00222 }
00223 }
00224 }
00225
00226 ctx->m.q_chroma_intra_matrix16 = ctx->qmatrix_c16;
00227 ctx->m.q_chroma_intra_matrix = ctx->qmatrix_c;
00228 ctx->m.q_intra_matrix16 = ctx->qmatrix_l16;
00229 ctx->m.q_intra_matrix = ctx->qmatrix_l;
00230
00231 return 0;
00232 fail:
00233 return -1;
00234 }
00235
00236 static int dnxhd_init_rc(DNXHDEncContext *ctx)
00237 {
00238 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_rc, 8160*ctx->m.avctx->qmax*sizeof(RCEntry), fail);
00239 if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD)
00240 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_cmp, ctx->m.mb_num*sizeof(RCCMPEntry), fail);
00241
00242 ctx->frame_bits = (ctx->cid_table->coding_unit_size - 640 - 4 - ctx->min_padding) * 8;
00243 ctx->qscale = 1;
00244 ctx->lambda = 2<<LAMBDA_FRAC_BITS;
00245 return 0;
00246 fail:
00247 return -1;
00248 }
00249
00250 static int dnxhd_encode_init(AVCodecContext *avctx)
00251 {
00252 DNXHDEncContext *ctx = avctx->priv_data;
00253 int i, index, bit_depth;
00254
00255 switch (avctx->pix_fmt) {
00256 case AV_PIX_FMT_YUV422P:
00257 bit_depth = 8;
00258 break;
00259 case AV_PIX_FMT_YUV422P10:
00260 bit_depth = 10;
00261 break;
00262 default:
00263 av_log(avctx, AV_LOG_ERROR, "pixel format is incompatible with DNxHD\n");
00264 return -1;
00265 }
00266
00267 ctx->cid = ff_dnxhd_find_cid(avctx, bit_depth);
00268 if (!ctx->cid) {
00269 av_log(avctx, AV_LOG_ERROR, "video parameters incompatible with DNxHD\n");
00270 return -1;
00271 }
00272 av_log(avctx, AV_LOG_DEBUG, "cid %d\n", ctx->cid);
00273
00274 index = ff_dnxhd_get_cid_table(ctx->cid);
00275 av_assert0(index >= 0);
00276 ctx->cid_table = &ff_dnxhd_cid_table[index];
00277
00278 ctx->m.avctx = avctx;
00279 ctx->m.mb_intra = 1;
00280 ctx->m.h263_aic = 1;
00281
00282 avctx->bits_per_raw_sample = ctx->cid_table->bit_depth;
00283
00284 ff_dct_common_init(&ctx->m);
00285 ff_dct_encode_init(&ctx->m);
00286
00287 if (!ctx->m.dct_quantize)
00288 ctx->m.dct_quantize = ff_dct_quantize_c;
00289
00290 if (ctx->cid_table->bit_depth == 10) {
00291 ctx->m.dct_quantize = dnxhd_10bit_dct_quantize;
00292 ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym;
00293 ctx->block_width_l2 = 4;
00294 } else {
00295 ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym;
00296 ctx->block_width_l2 = 3;
00297 }
00298
00299 if (ARCH_X86)
00300 ff_dnxhdenc_init_x86(ctx);
00301
00302 ctx->m.mb_height = (avctx->height + 15) / 16;
00303 ctx->m.mb_width = (avctx->width + 15) / 16;
00304
00305 if (avctx->flags & CODEC_FLAG_INTERLACED_DCT) {
00306 ctx->interlaced = 1;
00307 ctx->m.mb_height /= 2;
00308 }
00309
00310 ctx->m.mb_num = ctx->m.mb_height * ctx->m.mb_width;
00311
00312 if (avctx->intra_quant_bias != FF_DEFAULT_QUANT_BIAS)
00313 ctx->m.intra_quant_bias = avctx->intra_quant_bias;
00314 if (dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0) < 0)
00315 return -1;
00316
00317
00318 if (ctx->nitris_compat)
00319 ctx->min_padding = 1600;
00320
00321 if (dnxhd_init_vlc(ctx) < 0)
00322 return -1;
00323 if (dnxhd_init_rc(ctx) < 0)
00324 return -1;
00325
00326 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t), fail);
00327 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs, ctx->m.mb_height*sizeof(uint32_t), fail);
00328 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits, ctx->m.mb_num *sizeof(uint16_t), fail);
00329 FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale, ctx->m.mb_num *sizeof(uint8_t), fail);
00330
00331 ctx->frame.key_frame = 1;
00332 ctx->frame.pict_type = AV_PICTURE_TYPE_I;
00333 ctx->m.avctx->coded_frame = &ctx->frame;
00334
00335 if (avctx->thread_count > MAX_THREADS) {
00336 av_log(avctx, AV_LOG_ERROR, "too many threads\n");
00337 return -1;
00338 }
00339
00340 ctx->thread[0] = ctx;
00341 for (i = 1; i < avctx->thread_count; i++) {
00342 ctx->thread[i] = av_malloc(sizeof(DNXHDEncContext));
00343 memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext));
00344 }
00345
00346 return 0;
00347 fail:
00348 return -1;
00349 }
00350
00351 static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf)
00352 {
00353 DNXHDEncContext *ctx = avctx->priv_data;
00354 const uint8_t header_prefix[5] = { 0x00,0x00,0x02,0x80,0x01 };
00355
00356 memset(buf, 0, 640);
00357
00358 memcpy(buf, header_prefix, 5);
00359 buf[5] = ctx->interlaced ? ctx->cur_field+2 : 0x01;
00360 buf[6] = 0x80;
00361 buf[7] = 0xa0;
00362 AV_WB16(buf + 0x18, avctx->height>>ctx->interlaced);
00363 AV_WB16(buf + 0x1a, avctx->width);
00364 AV_WB16(buf + 0x1d, avctx->height>>ctx->interlaced);
00365
00366 buf[0x21] = ctx->cid_table->bit_depth == 10 ? 0x58 : 0x38;
00367 buf[0x22] = 0x88 + (ctx->interlaced<<2);
00368 AV_WB32(buf + 0x28, ctx->cid);
00369 buf[0x2c] = ctx->interlaced ? 0 : 0x80;
00370
00371 buf[0x5f] = 0x01;
00372
00373 buf[0x167] = 0x02;
00374 AV_WB16(buf + 0x16a, ctx->m.mb_height * 4 + 4);
00375 buf[0x16d] = ctx->m.mb_height;
00376 buf[0x16f] = 0x10;
00377
00378 ctx->msip = buf + 0x170;
00379 return 0;
00380 }
00381
00382 static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff)
00383 {
00384 int nbits;
00385 if (diff < 0) {
00386 nbits = av_log2_16bit(-2*diff);
00387 diff--;
00388 } else {
00389 nbits = av_log2_16bit(2*diff);
00390 }
00391 put_bits(&ctx->m.pb, ctx->cid_table->dc_bits[nbits] + nbits,
00392 (ctx->cid_table->dc_codes[nbits]<<nbits) + (diff & ((1 << nbits) - 1)));
00393 }
00394
00395 static av_always_inline void dnxhd_encode_block(DNXHDEncContext *ctx, DCTELEM *block, int last_index, int n)
00396 {
00397 int last_non_zero = 0;
00398 int slevel, i, j;
00399
00400 dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]);
00401 ctx->m.last_dc[n] = block[0];
00402
00403 for (i = 1; i <= last_index; i++) {
00404 j = ctx->m.intra_scantable.permutated[i];
00405 slevel = block[j];
00406 if (slevel) {
00407 int run_level = i - last_non_zero - 1;
00408 int rlevel = (slevel<<1)|!!run_level;
00409 put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]);
00410 if (run_level)
00411 put_bits(&ctx->m.pb, ctx->run_bits[run_level], ctx->run_codes[run_level]);
00412 last_non_zero = i;
00413 }
00414 }
00415 put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]);
00416 }
00417
00418 static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, DCTELEM *block, int n, int qscale, int last_index)
00419 {
00420 const uint8_t *weight_matrix;
00421 int level;
00422 int i;
00423
00424 weight_matrix = (n&2) ? ctx->cid_table->chroma_weight : ctx->cid_table->luma_weight;
00425
00426 for (i = 1; i <= last_index; i++) {
00427 int j = ctx->m.intra_scantable.permutated[i];
00428 level = block[j];
00429 if (level) {
00430 if (level < 0) {
00431 level = (1-2*level) * qscale * weight_matrix[i];
00432 if (ctx->cid_table->bit_depth == 10) {
00433 if (weight_matrix[i] != 8)
00434 level += 8;
00435 level >>= 4;
00436 } else {
00437 if (weight_matrix[i] != 32)
00438 level += 32;
00439 level >>= 6;
00440 }
00441 level = -level;
00442 } else {
00443 level = (2*level+1) * qscale * weight_matrix[i];
00444 if (ctx->cid_table->bit_depth == 10) {
00445 if (weight_matrix[i] != 8)
00446 level += 8;
00447 level >>= 4;
00448 } else {
00449 if (weight_matrix[i] != 32)
00450 level += 32;
00451 level >>= 6;
00452 }
00453 }
00454 block[j] = level;
00455 }
00456 }
00457 }
00458
00459 static av_always_inline int dnxhd_ssd_block(DCTELEM *qblock, DCTELEM *block)
00460 {
00461 int score = 0;
00462 int i;
00463 for (i = 0; i < 64; i++)
00464 score += (block[i] - qblock[i]) * (block[i] - qblock[i]);
00465 return score;
00466 }
00467
00468 static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, DCTELEM *block, int last_index)
00469 {
00470 int last_non_zero = 0;
00471 int bits = 0;
00472 int i, j, level;
00473 for (i = 1; i <= last_index; i++) {
00474 j = ctx->m.intra_scantable.permutated[i];
00475 level = block[j];
00476 if (level) {
00477 int run_level = i - last_non_zero - 1;
00478 bits += ctx->vlc_bits[(level<<1)|!!run_level]+ctx->run_bits[run_level];
00479 last_non_zero = i;
00480 }
00481 }
00482 return bits;
00483 }
00484
00485 static av_always_inline void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y)
00486 {
00487 const int bs = ctx->block_width_l2;
00488 const int bw = 1 << bs;
00489 const uint8_t *ptr_y = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs+1);
00490 const uint8_t *ptr_u = ctx->thread[0]->src[1] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
00491 const uint8_t *ptr_v = ctx->thread[0]->src[2] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs);
00492 DSPContext *dsp = &ctx->m.dsp;
00493
00494 dsp->get_pixels(ctx->blocks[0], ptr_y, ctx->m.linesize);
00495 dsp->get_pixels(ctx->blocks[1], ptr_y + bw, ctx->m.linesize);
00496 dsp->get_pixels(ctx->blocks[2], ptr_u, ctx->m.uvlinesize);
00497 dsp->get_pixels(ctx->blocks[3], ptr_v, ctx->m.uvlinesize);
00498
00499 if (mb_y+1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) {
00500 if (ctx->interlaced) {
00501 ctx->get_pixels_8x4_sym(ctx->blocks[4], ptr_y + ctx->dct_y_offset, ctx->m.linesize);
00502 ctx->get_pixels_8x4_sym(ctx->blocks[5], ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
00503 ctx->get_pixels_8x4_sym(ctx->blocks[6], ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
00504 ctx->get_pixels_8x4_sym(ctx->blocks[7], ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
00505 } else {
00506 dsp->clear_block(ctx->blocks[4]);
00507 dsp->clear_block(ctx->blocks[5]);
00508 dsp->clear_block(ctx->blocks[6]);
00509 dsp->clear_block(ctx->blocks[7]);
00510 }
00511 } else {
00512 dsp->get_pixels(ctx->blocks[4], ptr_y + ctx->dct_y_offset, ctx->m.linesize);
00513 dsp->get_pixels(ctx->blocks[5], ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize);
00514 dsp->get_pixels(ctx->blocks[6], ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize);
00515 dsp->get_pixels(ctx->blocks[7], ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize);
00516 }
00517 }
00518
00519 static av_always_inline int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i)
00520 {
00521 const static uint8_t component[8]={0,0,1,2,0,0,1,2};
00522 return component[i];
00523 }
00524
00525 static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
00526 {
00527 DNXHDEncContext *ctx = avctx->priv_data;
00528 int mb_y = jobnr, mb_x;
00529 int qscale = ctx->qscale;
00530 LOCAL_ALIGNED_16(DCTELEM, block, [64]);
00531 ctx = ctx->thread[threadnr];
00532
00533 ctx->m.last_dc[0] =
00534 ctx->m.last_dc[1] =
00535 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
00536
00537 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
00538 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
00539 int ssd = 0;
00540 int ac_bits = 0;
00541 int dc_bits = 0;
00542 int i;
00543
00544 dnxhd_get_blocks(ctx, mb_x, mb_y);
00545
00546 for (i = 0; i < 8; i++) {
00547 DCTELEM *src_block = ctx->blocks[i];
00548 int overflow, nbits, diff, last_index;
00549 int n = dnxhd_switch_matrix(ctx, i);
00550
00551 memcpy(block, src_block, 64*sizeof(*block));
00552 last_index = ctx->m.dct_quantize(&ctx->m, block, 4&(2*i), qscale, &overflow);
00553 ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index);
00554
00555 diff = block[0] - ctx->m.last_dc[n];
00556 if (diff < 0) nbits = av_log2_16bit(-2*diff);
00557 else nbits = av_log2_16bit( 2*diff);
00558
00559 av_assert1(nbits < ctx->cid_table->bit_depth + 4);
00560 dc_bits += ctx->cid_table->dc_bits[nbits] + nbits;
00561
00562 ctx->m.last_dc[n] = block[0];
00563
00564 if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) {
00565 dnxhd_unquantize_c(ctx, block, i, qscale, last_index);
00566 ctx->m.dsp.idct(block);
00567 ssd += dnxhd_ssd_block(block, src_block);
00568 }
00569 }
00570 ctx->mb_rc[qscale][mb].ssd = ssd;
00571 ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->vlc_bits[0];
00572 }
00573 return 0;
00574 }
00575
00576 static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
00577 {
00578 DNXHDEncContext *ctx = avctx->priv_data;
00579 int mb_y = jobnr, mb_x;
00580 ctx = ctx->thread[threadnr];
00581 init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr], ctx->slice_size[jobnr]);
00582
00583 ctx->m.last_dc[0] =
00584 ctx->m.last_dc[1] =
00585 ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2);
00586 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
00587 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
00588 int qscale = ctx->mb_qscale[mb];
00589 int i;
00590
00591 put_bits(&ctx->m.pb, 12, qscale<<1);
00592
00593 dnxhd_get_blocks(ctx, mb_x, mb_y);
00594
00595 for (i = 0; i < 8; i++) {
00596 DCTELEM *block = ctx->blocks[i];
00597 int overflow, n = dnxhd_switch_matrix(ctx, i);
00598 int last_index = ctx->m.dct_quantize(&ctx->m, block, 4&(2*i), qscale, &overflow);
00599
00600 dnxhd_encode_block(ctx, block, last_index, n);
00601
00602 }
00603 }
00604 if (put_bits_count(&ctx->m.pb)&31)
00605 put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0);
00606 flush_put_bits(&ctx->m.pb);
00607 return 0;
00608 }
00609
00610 static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx)
00611 {
00612 int mb_y, mb_x;
00613 int offset = 0;
00614 for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) {
00615 int thread_size;
00616 ctx->slice_offs[mb_y] = offset;
00617 ctx->slice_size[mb_y] = 0;
00618 for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) {
00619 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
00620 ctx->slice_size[mb_y] += ctx->mb_bits[mb];
00621 }
00622 ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31;
00623 ctx->slice_size[mb_y] >>= 3;
00624 thread_size = ctx->slice_size[mb_y];
00625 offset += thread_size;
00626 }
00627 }
00628
00629 static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr)
00630 {
00631 DNXHDEncContext *ctx = avctx->priv_data;
00632 int mb_y = jobnr, mb_x;
00633 ctx = ctx->thread[threadnr];
00634 if (ctx->cid_table->bit_depth == 8) {
00635 uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y<<4) * ctx->m.linesize);
00636 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) {
00637 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
00638 int sum = ctx->m.dsp.pix_sum(pix, ctx->m.linesize);
00639 int varc = (ctx->m.dsp.pix_norm1(pix, ctx->m.linesize) - (((unsigned)sum*sum)>>8)+128)>>8;
00640 ctx->mb_cmp[mb].value = varc;
00641 ctx->mb_cmp[mb].mb = mb;
00642 }
00643 } else {
00644 int const linesize = ctx->m.linesize >> 1;
00645 for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) {
00646 uint16_t *pix = (uint16_t*)ctx->thread[0]->src[0] + ((mb_y << 4) * linesize) + (mb_x << 4);
00647 unsigned mb = mb_y * ctx->m.mb_width + mb_x;
00648 int sum = 0;
00649 int sqsum = 0;
00650 int mean, sqmean;
00651 int i, j;
00652
00653 for (i = 0; i < 16; ++i) {
00654 for (j = 0; j < 16; ++j) {
00655
00656 int const sample = (unsigned)pix[j] >> 6;
00657 sum += sample;
00658 sqsum += sample * sample;
00659
00660 }
00661 pix += linesize;
00662 }
00663 mean = sum >> 8;
00664 sqmean = sqsum >> 8;
00665 ctx->mb_cmp[mb].value = sqmean - mean * mean;
00666 ctx->mb_cmp[mb].mb = mb;
00667 }
00668 }
00669 return 0;
00670 }
00671
00672 static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx)
00673 {
00674 int lambda, up_step, down_step;
00675 int last_lower = INT_MAX, last_higher = 0;
00676 int x, y, q;
00677
00678 for (q = 1; q < avctx->qmax; q++) {
00679 ctx->qscale = q;
00680 avctx->execute2(avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height);
00681 }
00682 up_step = down_step = 2<<LAMBDA_FRAC_BITS;
00683 lambda = ctx->lambda;
00684
00685 for (;;) {
00686 int bits = 0;
00687 int end = 0;
00688 if (lambda == last_higher) {
00689 lambda++;
00690 end = 1;
00691 }
00692 for (y = 0; y < ctx->m.mb_height; y++) {
00693 for (x = 0; x < ctx->m.mb_width; x++) {
00694 unsigned min = UINT_MAX;
00695 int qscale = 1;
00696 int mb = y*ctx->m.mb_width+x;
00697 for (q = 1; q < avctx->qmax; q++) {
00698 unsigned score = ctx->mb_rc[q][mb].bits*lambda+
00699 ((unsigned)ctx->mb_rc[q][mb].ssd<<LAMBDA_FRAC_BITS);
00700 if (score < min) {
00701 min = score;
00702 qscale = q;
00703 }
00704 }
00705 bits += ctx->mb_rc[qscale][mb].bits;
00706 ctx->mb_qscale[mb] = qscale;
00707 ctx->mb_bits[mb] = ctx->mb_rc[qscale][mb].bits;
00708 }
00709 bits = (bits+31)&~31;
00710 if (bits > ctx->frame_bits)
00711 break;
00712 }
00713
00714
00715 if (end) {
00716 if (bits > ctx->frame_bits)
00717 return -1;
00718 break;
00719 }
00720 if (bits < ctx->frame_bits) {
00721 last_lower = FFMIN(lambda, last_lower);
00722 if (last_higher != 0)
00723 lambda = (lambda+last_higher)>>1;
00724 else
00725 lambda -= down_step;
00726 down_step = FFMIN((int64_t)down_step*5, INT_MAX);
00727 up_step = 1<<LAMBDA_FRAC_BITS;
00728 lambda = FFMAX(1, lambda);
00729 if (lambda == last_lower)
00730 break;
00731 } else {
00732 last_higher = FFMAX(lambda, last_higher);
00733 if (last_lower != INT_MAX)
00734 lambda = (lambda+last_lower)>>1;
00735 else if ((int64_t)lambda + up_step > INT_MAX)
00736 return -1;
00737 else
00738 lambda += up_step;
00739 up_step = FFMIN((int64_t)up_step*5, INT_MAX);
00740 down_step = 1<<LAMBDA_FRAC_BITS;
00741 }
00742 }
00743
00744 ctx->lambda = lambda;
00745 return 0;
00746 }
00747
00748 static int dnxhd_find_qscale(DNXHDEncContext *ctx)
00749 {
00750 int bits = 0;
00751 int up_step = 1;
00752 int down_step = 1;
00753 int last_higher = 0;
00754 int last_lower = INT_MAX;
00755 int qscale;
00756 int x, y;
00757
00758 qscale = ctx->qscale;
00759 for (;;) {
00760 bits = 0;
00761 ctx->qscale = qscale;
00762
00763 ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height);
00764 for (y = 0; y < ctx->m.mb_height; y++) {
00765 for (x = 0; x < ctx->m.mb_width; x++)
00766 bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits;
00767 bits = (bits+31)&~31;
00768 if (bits > ctx->frame_bits)
00769 break;
00770 }
00771
00772
00773 if (bits < ctx->frame_bits) {
00774 if (qscale == 1)
00775 return 1;
00776 if (last_higher == qscale - 1) {
00777 qscale = last_higher;
00778 break;
00779 }
00780 last_lower = FFMIN(qscale, last_lower);
00781 if (last_higher != 0)
00782 qscale = (qscale+last_higher)>>1;
00783 else
00784 qscale -= down_step++;
00785 if (qscale < 1)
00786 qscale = 1;
00787 up_step = 1;
00788 } else {
00789 if (last_lower == qscale + 1)
00790 break;
00791 last_higher = FFMAX(qscale, last_higher);
00792 if (last_lower != INT_MAX)
00793 qscale = (qscale+last_lower)>>1;
00794 else
00795 qscale += up_step++;
00796 down_step = 1;
00797 if (qscale >= ctx->m.avctx->qmax)
00798 return -1;
00799 }
00800 }
00801
00802 ctx->qscale = qscale;
00803 return 0;
00804 }
00805
00806 #define BUCKET_BITS 8
00807 #define RADIX_PASSES 4
00808 #define NBUCKETS (1 << BUCKET_BITS)
00809
00810 static inline int get_bucket(int value, int shift)
00811 {
00812 value >>= shift;
00813 value &= NBUCKETS - 1;
00814 return NBUCKETS - 1 - value;
00815 }
00816
00817 static void radix_count(const RCCMPEntry *data, int size, int buckets[RADIX_PASSES][NBUCKETS])
00818 {
00819 int i, j;
00820 memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS);
00821 for (i = 0; i < size; i++) {
00822 int v = data[i].value;
00823 for (j = 0; j < RADIX_PASSES; j++) {
00824 buckets[j][get_bucket(v, 0)]++;
00825 v >>= BUCKET_BITS;
00826 }
00827 av_assert1(!v);
00828 }
00829 for (j = 0; j < RADIX_PASSES; j++) {
00830 int offset = size;
00831 for (i = NBUCKETS - 1; i >= 0; i--)
00832 buckets[j][i] = offset -= buckets[j][i];
00833 av_assert1(!buckets[j][0]);
00834 }
00835 }
00836
00837 static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data, int size, int buckets[NBUCKETS], int pass)
00838 {
00839 int shift = pass * BUCKET_BITS;
00840 int i;
00841 for (i = 0; i < size; i++) {
00842 int v = get_bucket(data[i].value, shift);
00843 int pos = buckets[v]++;
00844 dst[pos] = data[i];
00845 }
00846 }
00847
00848 static void radix_sort(RCCMPEntry *data, int size)
00849 {
00850 int buckets[RADIX_PASSES][NBUCKETS];
00851 RCCMPEntry *tmp = av_malloc(sizeof(*tmp) * size);
00852 radix_count(data, size, buckets);
00853 radix_sort_pass(tmp, data, size, buckets[0], 0);
00854 radix_sort_pass(data, tmp, size, buckets[1], 1);
00855 if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) {
00856 radix_sort_pass(tmp, data, size, buckets[2], 2);
00857 radix_sort_pass(data, tmp, size, buckets[3], 3);
00858 }
00859 av_free(tmp);
00860 }
00861
00862 static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx)
00863 {
00864 int max_bits = 0;
00865 int ret, x, y;
00866 if ((ret = dnxhd_find_qscale(ctx)) < 0)
00867 return -1;
00868 for (y = 0; y < ctx->m.mb_height; y++) {
00869 for (x = 0; x < ctx->m.mb_width; x++) {
00870 int mb = y*ctx->m.mb_width+x;
00871 int delta_bits;
00872 ctx->mb_qscale[mb] = ctx->qscale;
00873 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale][mb].bits;
00874 max_bits += ctx->mb_rc[ctx->qscale][mb].bits;
00875 if (!RC_VARIANCE) {
00876 delta_bits = ctx->mb_rc[ctx->qscale][mb].bits-ctx->mb_rc[ctx->qscale+1][mb].bits;
00877 ctx->mb_cmp[mb].mb = mb;
00878 ctx->mb_cmp[mb].value = delta_bits ?
00879 ((ctx->mb_rc[ctx->qscale][mb].ssd-ctx->mb_rc[ctx->qscale+1][mb].ssd)*100)/delta_bits
00880 : INT_MIN;
00881 }
00882 }
00883 max_bits += 31;
00884 }
00885 if (!ret) {
00886 if (RC_VARIANCE)
00887 avctx->execute2(avctx, dnxhd_mb_var_thread, NULL, NULL, ctx->m.mb_height);
00888 radix_sort(ctx->mb_cmp, ctx->m.mb_num);
00889 for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) {
00890 int mb = ctx->mb_cmp[x].mb;
00891 max_bits -= ctx->mb_rc[ctx->qscale][mb].bits - ctx->mb_rc[ctx->qscale+1][mb].bits;
00892 ctx->mb_qscale[mb] = ctx->qscale+1;
00893 ctx->mb_bits[mb] = ctx->mb_rc[ctx->qscale+1][mb].bits;
00894 }
00895 }
00896 return 0;
00897 }
00898
00899 static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame)
00900 {
00901 int i;
00902
00903 for (i = 0; i < 3; i++) {
00904 ctx->frame.data[i] = frame->data[i];
00905 ctx->frame.linesize[i] = frame->linesize[i];
00906 }
00907
00908 for (i = 0; i < ctx->m.avctx->thread_count; i++) {
00909 ctx->thread[i]->m.linesize = ctx->frame.linesize[0]<<ctx->interlaced;
00910 ctx->thread[i]->m.uvlinesize = ctx->frame.linesize[1]<<ctx->interlaced;
00911 ctx->thread[i]->dct_y_offset = ctx->m.linesize *8;
00912 ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8;
00913 }
00914
00915 ctx->frame.interlaced_frame = frame->interlaced_frame;
00916 ctx->cur_field = frame->interlaced_frame && !frame->top_field_first;
00917 }
00918
00919 static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt,
00920 const AVFrame *frame, int *got_packet)
00921 {
00922 DNXHDEncContext *ctx = avctx->priv_data;
00923 int first_field = 1;
00924 int offset, i, ret;
00925 uint8_t *buf;
00926
00927 if ((ret = ff_alloc_packet2(avctx, pkt, ctx->cid_table->frame_size)) < 0)
00928 return ret;
00929 buf = pkt->data;
00930
00931 dnxhd_load_picture(ctx, frame);
00932
00933 encode_coding_unit:
00934 for (i = 0; i < 3; i++) {
00935 ctx->src[i] = ctx->frame.data[i];
00936 if (ctx->interlaced && ctx->cur_field)
00937 ctx->src[i] += ctx->frame.linesize[i];
00938 }
00939
00940 dnxhd_write_header(avctx, buf);
00941
00942 if (avctx->mb_decision == FF_MB_DECISION_RD)
00943 ret = dnxhd_encode_rdo(avctx, ctx);
00944 else
00945 ret = dnxhd_encode_fast(avctx, ctx);
00946 if (ret < 0) {
00947 av_log(avctx, AV_LOG_ERROR,
00948 "picture could not fit ratecontrol constraints, increase qmax\n");
00949 return -1;
00950 }
00951
00952 dnxhd_setup_threads_slices(ctx);
00953
00954 offset = 0;
00955 for (i = 0; i < ctx->m.mb_height; i++) {
00956 AV_WB32(ctx->msip + i * 4, offset);
00957 offset += ctx->slice_size[i];
00958 av_assert1(!(ctx->slice_size[i] & 3));
00959 }
00960
00961 avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height);
00962
00963 av_assert1(640 + offset + 4 <= ctx->cid_table->coding_unit_size);
00964 memset(buf + 640 + offset, 0, ctx->cid_table->coding_unit_size - 4 - offset - 640);
00965
00966 AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE);
00967
00968 if (ctx->interlaced && first_field) {
00969 first_field = 0;
00970 ctx->cur_field ^= 1;
00971 buf += ctx->cid_table->coding_unit_size;
00972 goto encode_coding_unit;
00973 }
00974
00975 ctx->frame.quality = ctx->qscale*FF_QP2LAMBDA;
00976
00977 pkt->flags |= AV_PKT_FLAG_KEY;
00978 *got_packet = 1;
00979 return 0;
00980 }
00981
00982 static int dnxhd_encode_end(AVCodecContext *avctx)
00983 {
00984 DNXHDEncContext *ctx = avctx->priv_data;
00985 int max_level = 1<<(ctx->cid_table->bit_depth+2);
00986 int i;
00987
00988 av_free(ctx->vlc_codes-max_level*2);
00989 av_free(ctx->vlc_bits -max_level*2);
00990 av_freep(&ctx->run_codes);
00991 av_freep(&ctx->run_bits);
00992
00993 av_freep(&ctx->mb_bits);
00994 av_freep(&ctx->mb_qscale);
00995 av_freep(&ctx->mb_rc);
00996 av_freep(&ctx->mb_cmp);
00997 av_freep(&ctx->slice_size);
00998 av_freep(&ctx->slice_offs);
00999
01000 av_freep(&ctx->qmatrix_c);
01001 av_freep(&ctx->qmatrix_l);
01002 av_freep(&ctx->qmatrix_c16);
01003 av_freep(&ctx->qmatrix_l16);
01004
01005 for (i = 1; i < avctx->thread_count; i++)
01006 av_freep(&ctx->thread[i]);
01007
01008 return 0;
01009 }
01010
01011 static const AVCodecDefault dnxhd_defaults[] = {
01012 { "qmax", "1024" },
01013 { NULL },
01014 };
01015
01016 AVCodec ff_dnxhd_encoder = {
01017 .name = "dnxhd",
01018 .type = AVMEDIA_TYPE_VIDEO,
01019 .id = AV_CODEC_ID_DNXHD,
01020 .priv_data_size = sizeof(DNXHDEncContext),
01021 .init = dnxhd_encode_init,
01022 .encode2 = dnxhd_encode_picture,
01023 .close = dnxhd_encode_end,
01024 .capabilities = CODEC_CAP_SLICE_THREADS,
01025 .pix_fmts = (const enum AVPixelFormat[]){ AV_PIX_FMT_YUV422P,
01026 AV_PIX_FMT_YUV422P10,
01027 AV_PIX_FMT_NONE },
01028 .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"),
01029 .priv_class = &class,
01030 .defaults = dnxhd_defaults,
01031 };