X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libavcodec%2Fdnxhdenc.c;h=8531fe0ea5d57002ab1bdb81567b1c36d95dc363;hb=90fcac0e95b7d266c148a86506f301a2072d9de3;hp=9ca842e1d74f9e1ed53bc4da1ecfc2e6d6666db6;hpb=0c39c38b1456d4a59c4982db74facaa1722edc18;p=ffmpeg diff --git a/libavcodec/dnxhdenc.c b/libavcodec/dnxhdenc.c index 9ca842e1d74..8531fe0ea5d 100644 --- a/libavcodec/dnxhdenc.c +++ b/libavcodec/dnxhdenc.c @@ -1,116 +1,153 @@ /* * VC3/DNxHD encoder * Copyright (c) 2007 Baptiste Coudurier + * Copyright (c) 2011 MirriAd Ltd * * VC-3 encoder funded by the British Broadcasting Corporation + * 10 bit support added by MirriAd Ltd, Joseph Artsimovich * - * This file is part of FFmpeg. + * This file is part of Libav. * - * FFmpeg is free software; you can redistribute it and/or + * Libav is free software; you can redistribute it and/or * modify it under the terms of the GNU Lesser General Public * License as published by the Free Software Foundation; either * version 2.1 of the License, or (at your option) any later version. * - * FFmpeg is distributed in the hope that it will be useful, + * Libav is distributed in the hope that it will be useful, * but WITHOUT ANY WARRANTY; without even the implied warranty of * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU * Lesser General Public License for more details. * * You should have received a copy of the GNU Lesser General Public - * License along with FFmpeg; if not, write to the Free Software + * License along with Libav; if not, write to the Free Software * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA */ //#define DEBUG #define RC_VARIANCE 1 // use variance or ssd for fast rc +#include "libavutil/opt.h" #include "avcodec.h" #include "dsputil.h" +#include "internal.h" #include "mpegvideo.h" -#include "dnxhddata.h" +#include "dnxhdenc.h" -typedef struct { - uint16_t mb; - int value; -} RCCMPEntry; +#define VE AV_OPT_FLAG_VIDEO_PARAM | AV_OPT_FLAG_ENCODING_PARAM +#define DNX10BIT_QMAT_SHIFT 18 // The largest value that will not lead to overflow for 10bit samples. -typedef struct { - int ssd; - int bits; -} RCEntry; - -int dct_quantize_c(MpegEncContext *s, DCTELEM *block, int n, int qscale, int *overflow); - -typedef struct DNXHDEncContext { - MpegEncContext m; ///< Used for quantization dsp functions - - AVFrame frame; - int cid; - const CIDEntry *cid_table; - uint8_t *msip; ///< Macroblock Scan Indices Payload - uint32_t *slice_size; - - struct DNXHDEncContext *thread[MAX_THREADS]; +static const AVOption options[]={ + {"nitris_compat", "encode with Avid Nitris compatibility", offsetof(DNXHDEncContext, nitris_compat), AV_OPT_TYPE_INT, {.i64 = 0}, 0, 1, VE}, +{NULL} +}; +static const AVClass class = { "dnxhd", av_default_item_name, options, LIBAVUTIL_VERSION_INT }; - unsigned dct_y_offset; - unsigned dct_uv_offset; - int interlaced; - int cur_field; +#define LAMBDA_FRAC_BITS 10 - DECLARE_ALIGNED_16(DCTELEM, blocks[8][64]); +static void dnxhd_8bit_get_pixels_8x4_sym(DCTELEM *restrict block, const uint8_t *pixels, int line_size) +{ + int i; + for (i = 0; i < 4; i++) { + block[0] = pixels[0]; block[1] = pixels[1]; + block[2] = pixels[2]; block[3] = pixels[3]; + block[4] = pixels[4]; block[5] = pixels[5]; + block[6] = pixels[6]; block[7] = pixels[7]; + pixels += line_size; + block += 8; + } + memcpy(block, block - 8, sizeof(*block) * 8); + memcpy(block + 8, block - 16, sizeof(*block) * 8); + memcpy(block + 16, block - 24, sizeof(*block) * 8); + memcpy(block + 24, block - 32, sizeof(*block) * 8); +} - int (*qmatrix_c) [64]; - int (*qmatrix_l) [64]; - uint16_t (*qmatrix_l16)[2][64]; - uint16_t (*qmatrix_c16)[2][64]; +static av_always_inline void dnxhd_10bit_get_pixels_8x4_sym(DCTELEM *restrict block, const uint8_t *pixels, int line_size) +{ + int i; - unsigned frame_bits; - uint8_t *src[3]; + block += 32; - uint16_t *table_vlc_codes; - uint8_t *table_vlc_bits; - uint16_t *table_run_codes; - uint8_t *table_run_bits; + for (i = 0; i < 4; i++) { + memcpy(block + i * 8, pixels + i * line_size, 8 * sizeof(*block)); + memcpy(block - (i+1) * 8, pixels + i * line_size, 8 * sizeof(*block)); + } +} - /** Rate control */ - unsigned slice_bits; - unsigned qscale; - unsigned lambda; +static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, DCTELEM *block, + int n, int qscale, int *overflow) +{ + const uint8_t *scantable= ctx->intra_scantable.scantable; + const int *qmat = ctx->q_intra_matrix[qscale]; + int last_non_zero = 0; + int i; - unsigned thread_size; + ctx->dsp.fdct(block); - uint16_t *mb_bits; - uint8_t *mb_qscale; + // Divide by 4 with rounding, to compensate scaling of DCT coefficients + block[0] = (block[0] + 2) >> 2; - RCCMPEntry *mb_cmp; - RCEntry (*mb_rc)[8160]; -} DNXHDEncContext; + for (i = 1; i < 64; ++i) { + int j = scantable[i]; + int sign = block[j] >> 31; + int level = (block[j] ^ sign) - sign; + level = level * qmat[j] >> DNX10BIT_QMAT_SHIFT; + block[j] = (level ^ sign) - sign; + if (level) + last_non_zero = i; + } -#define LAMBDA_FRAC_BITS 10 + return last_non_zero; +} static int dnxhd_init_vlc(DNXHDEncContext *ctx) { - int i; - - CHECKED_ALLOCZ(ctx->table_vlc_codes, 449*2); - CHECKED_ALLOCZ(ctx->table_vlc_bits, 449); - CHECKED_ALLOCZ(ctx->table_run_codes, 63*2); - CHECKED_ALLOCZ(ctx->table_run_bits, 63); - - for (i = 0; i < 257; i++) { - int level = ctx->cid_table->ac_level[i] + - (ctx->cid_table->ac_run_flag[i] << 7) + (ctx->cid_table->ac_index_flag[i] << 8); - assert(level < 449); - if (ctx->cid_table->ac_level[i] == 64 && ctx->cid_table->ac_index_flag[i]) - level -= 64; // use 0+(1<<8) level - ctx->table_vlc_codes[level] = ctx->cid_table->ac_codes[i]; - ctx->table_vlc_bits [level] = ctx->cid_table->ac_bits[i]; + int i, j, level, run; + int max_level = 1<<(ctx->cid_table->bit_depth+2); + + FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_codes, max_level*4*sizeof(*ctx->vlc_codes), fail); + FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->vlc_bits, max_level*4*sizeof(*ctx->vlc_bits) , fail); + FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_codes, 63*2, fail); + FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->run_bits, 63, fail); + + ctx->vlc_codes += max_level*2; + ctx->vlc_bits += max_level*2; + for (level = -max_level; level < max_level; level++) { + for (run = 0; run < 2; run++) { + int index = (level<<1)|run; + int sign, offset = 0, alevel = level; + + MASK_ABS(sign, alevel); + if (alevel > 64) { + offset = (alevel-1)>>6; + alevel -= offset<<6; + } + for (j = 0; j < 257; j++) { + if (ctx->cid_table->ac_level[j] == alevel && + (!offset || (ctx->cid_table->ac_index_flag[j] && offset)) && + (!run || (ctx->cid_table->ac_run_flag [j] && run))) { + assert(!ctx->vlc_codes[index]); + if (alevel) { + ctx->vlc_codes[index] = (ctx->cid_table->ac_codes[j]<<1)|(sign&1); + ctx->vlc_bits [index] = ctx->cid_table->ac_bits[j]+1; + } else { + ctx->vlc_codes[index] = ctx->cid_table->ac_codes[j]; + ctx->vlc_bits [index] = ctx->cid_table->ac_bits [j]; + } + break; + } + } + assert(!alevel || j < 257); + if (offset) { + ctx->vlc_codes[index] = (ctx->vlc_codes[index]<cid_table->index_bits)|offset; + ctx->vlc_bits [index]+= ctx->cid_table->index_bits; + } + } } for (i = 0; i < 62; i++) { int run = ctx->cid_table->run[i]; assert(run < 63); - ctx->table_run_codes[run] = ctx->cid_table->run_codes[i]; - ctx->table_run_bits [run] = ctx->cid_table->run_bits[i]; + ctx->run_codes[run] = ctx->cid_table->run_codes[i]; + ctx->run_bits [run] = ctx->cid_table->run_bits[i]; } return 0; fail: @@ -122,31 +159,55 @@ static int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias) // init first elem to 1 to avoid div by 0 in convert_matrix uint16_t weight_matrix[64] = {1,}; // convert_matrix needs uint16_t* int qscale, i; - - CHECKED_ALLOCZ(ctx->qmatrix_l, (ctx->m.avctx->qmax+1) * 64 * sizeof(int)); - CHECKED_ALLOCZ(ctx->qmatrix_c, (ctx->m.avctx->qmax+1) * 64 * sizeof(int)); - CHECKED_ALLOCZ(ctx->qmatrix_l16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t)); - CHECKED_ALLOCZ(ctx->qmatrix_c16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t)); - - for (i = 1; i < 64; i++) { - int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]]; - weight_matrix[j] = ctx->cid_table->luma_weight[i]; - } - ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_l, ctx->qmatrix_l16, weight_matrix, - ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1); - for (i = 1; i < 64; i++) { - int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]]; - weight_matrix[j] = ctx->cid_table->chroma_weight[i]; - } - ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_c, ctx->qmatrix_c16, weight_matrix, - ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1); - for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) { - for (i = 0; i < 64; i++) { - ctx->qmatrix_l [qscale] [i] <<= 2; ctx->qmatrix_c [qscale] [i] <<= 2; - ctx->qmatrix_l16[qscale][0][i] <<= 2; ctx->qmatrix_l16[qscale][1][i] <<= 2; - ctx->qmatrix_c16[qscale][0][i] <<= 2; ctx->qmatrix_c16[qscale][1][i] <<= 2; + const uint8_t *luma_weight_table = ctx->cid_table->luma_weight; + const uint8_t *chroma_weight_table = ctx->cid_table->chroma_weight; + + FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l, (ctx->m.avctx->qmax+1) * 64 * sizeof(int), fail); + FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c, (ctx->m.avctx->qmax+1) * 64 * sizeof(int), fail); + FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_l16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t), fail); + FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->qmatrix_c16, (ctx->m.avctx->qmax+1) * 64 * 2 * sizeof(uint16_t), fail); + + if (ctx->cid_table->bit_depth == 8) { + for (i = 1; i < 64; i++) { + int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]]; + weight_matrix[j] = ctx->cid_table->luma_weight[i]; + } + ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_l, ctx->qmatrix_l16, weight_matrix, + ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1); + for (i = 1; i < 64; i++) { + int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]]; + weight_matrix[j] = ctx->cid_table->chroma_weight[i]; + } + ff_convert_matrix(&ctx->m.dsp, ctx->qmatrix_c, ctx->qmatrix_c16, weight_matrix, + ctx->m.intra_quant_bias, 1, ctx->m.avctx->qmax, 1); + + for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) { + for (i = 0; i < 64; i++) { + ctx->qmatrix_l [qscale] [i] <<= 2; ctx->qmatrix_c [qscale] [i] <<= 2; + ctx->qmatrix_l16[qscale][0][i] <<= 2; ctx->qmatrix_l16[qscale][1][i] <<= 2; + ctx->qmatrix_c16[qscale][0][i] <<= 2; ctx->qmatrix_c16[qscale][1][i] <<= 2; + } + } + } else { + // 10-bit + for (qscale = 1; qscale <= ctx->m.avctx->qmax; qscale++) { + for (i = 1; i < 64; i++) { + int j = ctx->m.dsp.idct_permutation[ff_zigzag_direct[i]]; + + // The quantization formula from the VC-3 standard is: + // quantized = sign(block[i]) * floor(abs(block[i]/s) * p / (qscale * weight_table[i])) + // Where p is 32 for 8-bit samples and 8 for 10-bit ones. + // The s factor compensates scaling of DCT coefficients done by the DCT routines, + // and therefore is not present in standard. It's 8 for 8-bit samples and 4 for 10-bit ones. + // We want values of ctx->qtmatrix_l and ctx->qtmatrix_r to be: + // ((1 << DNX10BIT_QMAT_SHIFT) * (p / s)) / (qscale * weight_table[i]) + // For 10-bit samples, p / s == 2 + ctx->qmatrix_l[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) / (qscale * luma_weight_table[i]); + ctx->qmatrix_c[qscale][j] = (1 << (DNX10BIT_QMAT_SHIFT + 1)) / (qscale * chroma_weight_table[i]); + } } } + return 0; fail: return -1; @@ -154,11 +215,11 @@ static int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias) static int dnxhd_init_rc(DNXHDEncContext *ctx) { - CHECKED_ALLOCZ(ctx->mb_rc, 8160*ctx->m.avctx->qmax*sizeof(RCEntry)); + FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_rc, 8160*ctx->m.avctx->qmax*sizeof(RCEntry), fail); if (ctx->m.avctx->mb_decision != FF_MB_DECISION_RD) - CHECKED_ALLOCZ(ctx->mb_cmp, ctx->m.mb_num*sizeof(RCCMPEntry)); + FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_cmp, ctx->m.mb_num*sizeof(RCCMPEntry), fail); - ctx->frame_bits = (ctx->cid_table->coding_unit_size - 640 - 4) * 8; + ctx->frame_bits = (ctx->cid_table->coding_unit_size - 640 - 4 - ctx->min_padding) * 8; ctx->qscale = 1; ctx->lambda = 2<priv_data; - int i, index; + int i, index, bit_depth; + + switch (avctx->pix_fmt) { + case AV_PIX_FMT_YUV422P: + bit_depth = 8; + break; + case AV_PIX_FMT_YUV422P10: + bit_depth = 10; + break; + default: + av_log(avctx, AV_LOG_ERROR, "pixel format is incompatible with DNxHD\n"); + return -1; + } - ctx->cid = ff_dnxhd_find_cid(avctx); - if (!ctx->cid || avctx->pix_fmt != PIX_FMT_YUV422P) { + ctx->cid = ff_dnxhd_find_cid(avctx, bit_depth); + if (!ctx->cid) { av_log(avctx, AV_LOG_ERROR, "video parameters incompatible with DNxHD\n"); return -1; } @@ -185,10 +258,24 @@ static int dnxhd_encode_init(AVCodecContext *avctx) ctx->m.mb_intra = 1; ctx->m.h263_aic = 1; - dsputil_init(&ctx->m.dsp, avctx); + avctx->bits_per_raw_sample = ctx->cid_table->bit_depth; + + ff_dsputil_init(&ctx->m.dsp, avctx); ff_dct_common_init(&ctx->m); if (!ctx->m.dct_quantize) - ctx->m.dct_quantize = dct_quantize_c; + ctx->m.dct_quantize = ff_dct_quantize_c; + + if (ctx->cid_table->bit_depth == 10) { + ctx->m.dct_quantize = dnxhd_10bit_dct_quantize; + ctx->get_pixels_8x4_sym = dnxhd_10bit_get_pixels_8x4_sym; + ctx->block_width_l2 = 4; + } else { + ctx->get_pixels_8x4_sym = dnxhd_8bit_get_pixels_8x4_sym; + ctx->block_width_l2 = 3; + } + + if (ARCH_X86) + ff_dnxhdenc_init_x86(ctx); ctx->m.mb_height = (avctx->height + 15) / 16; ctx->m.mb_width = (avctx->width + 15) / 16; @@ -205,20 +292,25 @@ static int dnxhd_encode_init(AVCodecContext *avctx) if (dnxhd_init_qmat(ctx, ctx->m.intra_quant_bias, 0) < 0) // XXX tune lbias/cbias return -1; + // Avid Nitris hardware decoder requires a minimum amount of padding in the coding unit payload + if (ctx->nitris_compat) + ctx->min_padding = 1600; + if (dnxhd_init_vlc(ctx) < 0) return -1; if (dnxhd_init_rc(ctx) < 0) return -1; - CHECKED_ALLOCZ(ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t)); - CHECKED_ALLOCZ(ctx->mb_bits, ctx->m.mb_num *sizeof(uint16_t)); - CHECKED_ALLOCZ(ctx->mb_qscale, ctx->m.mb_num *sizeof(uint8_t)); + FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_size, ctx->m.mb_height*sizeof(uint32_t), fail); + FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->slice_offs, ctx->m.mb_height*sizeof(uint32_t), fail); + FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_bits, ctx->m.mb_num *sizeof(uint16_t), fail); + FF_ALLOCZ_OR_GOTO(ctx->m.avctx, ctx->mb_qscale, ctx->m.mb_num *sizeof(uint8_t), fail); ctx->frame.key_frame = 1; - ctx->frame.pict_type = FF_I_TYPE; + ctx->frame.pict_type = AV_PICTURE_TYPE_I; ctx->m.avctx->coded_frame = &ctx->frame; - if (avctx->thread_count > MAX_THREADS || (avctx->thread_count > ctx->m.mb_height)) { + if (avctx->thread_count > MAX_THREADS) { av_log(avctx, AV_LOG_ERROR, "too many threads\n"); return -1; } @@ -229,13 +321,8 @@ static int dnxhd_encode_init(AVCodecContext *avctx) memcpy(ctx->thread[i], ctx, sizeof(DNXHDEncContext)); } - for (i = 0; i < avctx->thread_count; i++) { - ctx->thread[i]->m.start_mb_y = (ctx->m.mb_height*(i ) + avctx->thread_count/2) / avctx->thread_count; - ctx->thread[i]->m.end_mb_y = (ctx->m.mb_height*(i+1) + avctx->thread_count/2) / avctx->thread_count; - } - return 0; - fail: //for CHECKED_ALLOCZ + fail: //for FF_ALLOCZ_OR_GOTO return -1; } @@ -244,16 +331,18 @@ static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf) DNXHDEncContext *ctx = avctx->priv_data; const uint8_t header_prefix[5] = { 0x00,0x00,0x02,0x80,0x01 }; + memset(buf, 0, 640); + memcpy(buf, header_prefix, 5); buf[5] = ctx->interlaced ? ctx->cur_field+2 : 0x01; buf[6] = 0x80; // crc flag off buf[7] = 0xa0; // reserved - AV_WB16(buf + 0x18, avctx->height); // ALPF + AV_WB16(buf + 0x18, avctx->height>>ctx->interlaced); // ALPF AV_WB16(buf + 0x1a, avctx->width); // SPL - AV_WB16(buf + 0x1d, avctx->height); // NAL + AV_WB16(buf + 0x1d, avctx->height>>ctx->interlaced); // NAL - buf[0x21] = 0x38; // FIXME 8 bit per comp - buf[0x22] = 0x88 + (ctx->frame.interlaced_frame<<2); + buf[0x21] = ctx->cid_table->bit_depth == 10 ? 0x58 : 0x38; + buf[0x22] = 0x88 + (ctx->interlaced<<2); AV_WB32(buf + 0x28, ctx->cid); // CID buf[0x2c] = ctx->interlaced ? 0 : 0x80; @@ -284,7 +373,6 @@ static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff) static av_always_inline void dnxhd_encode_block(DNXHDEncContext *ctx, DCTELEM *block, int last_index, int n) { int last_non_zero = 0; - int offset = 0; int slevel, i, j; dnxhd_encode_dc(ctx, block[0] - ctx->m.last_dc[n]); @@ -295,25 +383,14 @@ static av_always_inline void dnxhd_encode_block(DNXHDEncContext *ctx, DCTELEM *b slevel = block[j]; if (slevel) { int run_level = i - last_non_zero - 1; - int sign; - MASK_ABS(sign, slevel); - if (slevel > 64) { - offset = (slevel-1) >> 6; - slevel = 256 | (slevel & 63); // level 64 is treated as 0 - } + int rlevel = (slevel<<1)|!!run_level; + put_bits(&ctx->m.pb, ctx->vlc_bits[rlevel], ctx->vlc_codes[rlevel]); if (run_level) - slevel |= 128; - put_bits(&ctx->m.pb, ctx->table_vlc_bits[slevel]+1, (ctx->table_vlc_codes[slevel]<<1)|(sign&1)); - if (offset) { - put_bits(&ctx->m.pb, 4, offset); - offset = 0; - } - if (run_level) - put_bits(&ctx->m.pb, ctx->table_run_bits[run_level], ctx->table_run_codes[run_level]); + put_bits(&ctx->m.pb, ctx->run_bits[run_level], ctx->run_codes[run_level]); last_non_zero = i; } } - put_bits(&ctx->m.pb, ctx->table_vlc_bits[0], ctx->table_vlc_codes[0]); // EOB + put_bits(&ctx->m.pb, ctx->vlc_bits[0], ctx->vlc_codes[0]); // EOB } static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, DCTELEM *block, int n, int qscale, int last_index) @@ -330,15 +407,27 @@ static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, DCTELEM *b if (level) { if (level < 0) { level = (1-2*level) * qscale * weight_matrix[i]; - if (weight_matrix[i] != 32) - level += 32; - level >>= 6; + if (ctx->cid_table->bit_depth == 10) { + if (weight_matrix[i] != 8) + level += 8; + level >>= 4; + } else { + if (weight_matrix[i] != 32) + level += 32; + level >>= 6; + } level = -level; } else { level = (2*level+1) * qscale * weight_matrix[i]; - if (weight_matrix[i] != 32) - level += 32; - level >>= 6; + if (ctx->cid_table->bit_depth == 10) { + if (weight_matrix[i] != 8) + level += 8; + level >>= 4; + } else { + if (weight_matrix[i] != 32) + level += 32; + level >>= 6; + } } block[j] = level; } @@ -350,7 +439,7 @@ static av_always_inline int dnxhd_ssd_block(DCTELEM *qblock, DCTELEM *block) int score = 0; int i; for (i = 0; i < 64; i++) - score += (block[i]-qblock[i])*(block[i]-qblock[i]); + score += (block[i] - qblock[i]) * (block[i] - qblock[i]); return score; } @@ -364,65 +453,44 @@ static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, DCTELEM *bl level = block[j]; if (level) { int run_level = i - last_non_zero - 1; - level = FFABS(level); - if (level > 64) { - level = 256 | (level & 63); // level 64 is treated as 0 - bits += 4; - } - level |= (!!run_level)<<7; - bits += ctx->table_vlc_bits[level]+1 + ctx->table_run_bits[run_level]; + bits += ctx->vlc_bits[(level<<1)|!!run_level]+ctx->run_bits[run_level]; last_non_zero = i; } } return bits; } -static av_always_inline void dnxhd_get_pixels_4x8(DCTELEM *restrict block, const uint8_t *pixels, int line_size) -{ - int i; - for (i = 0; i < 4; i++) { - block[0] = pixels[0]; - block[1] = pixels[1]; - block[2] = pixels[2]; - block[3] = pixels[3]; - block[4] = pixels[4]; - block[5] = pixels[5]; - block[6] = pixels[6]; - block[7] = pixels[7]; - pixels += line_size; - block += 8; - } - memcpy(block , block- 8, sizeof(*block)*8); - memcpy(block+ 8, block-16, sizeof(*block)*8); - memcpy(block+16, block-24, sizeof(*block)*8); - memcpy(block+24, block-32, sizeof(*block)*8); -} - static av_always_inline void dnxhd_get_blocks(DNXHDEncContext *ctx, int mb_x, int mb_y) { - const uint8_t *ptr_y = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize) + (mb_x << 4); - const uint8_t *ptr_u = ctx->thread[0]->src[1] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3); - const uint8_t *ptr_v = ctx->thread[0]->src[2] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << 3); + const int bs = ctx->block_width_l2; + const int bw = 1 << bs; + const uint8_t *ptr_y = ctx->thread[0]->src[0] + ((mb_y << 4) * ctx->m.linesize) + (mb_x << bs+1); + const uint8_t *ptr_u = ctx->thread[0]->src[1] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs); + const uint8_t *ptr_v = ctx->thread[0]->src[2] + ((mb_y << 4) * ctx->m.uvlinesize) + (mb_x << bs); DSPContext *dsp = &ctx->m.dsp; - dsp->get_pixels(ctx->blocks[0], ptr_y , ctx->m.linesize); - dsp->get_pixels(ctx->blocks[1], ptr_y + 8, ctx->m.linesize); - dsp->get_pixels(ctx->blocks[2], ptr_u , ctx->m.uvlinesize); - dsp->get_pixels(ctx->blocks[3], ptr_v , ctx->m.uvlinesize); + dsp->get_pixels(ctx->blocks[0], ptr_y, ctx->m.linesize); + dsp->get_pixels(ctx->blocks[1], ptr_y + bw, ctx->m.linesize); + dsp->get_pixels(ctx->blocks[2], ptr_u, ctx->m.uvlinesize); + dsp->get_pixels(ctx->blocks[3], ptr_v, ctx->m.uvlinesize); if (mb_y+1 == ctx->m.mb_height && ctx->m.avctx->height == 1080) { if (ctx->interlaced) { - dnxhd_get_pixels_4x8(ctx->blocks[4], ptr_y + ctx->dct_y_offset , ctx->m.linesize); - dnxhd_get_pixels_4x8(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize); - dnxhd_get_pixels_4x8(ctx->blocks[6], ptr_u + ctx->dct_uv_offset , ctx->m.uvlinesize); - dnxhd_get_pixels_4x8(ctx->blocks[7], ptr_v + ctx->dct_uv_offset , ctx->m.uvlinesize); - } else - memset(ctx->blocks[4], 0, 4*64*sizeof(DCTELEM)); + ctx->get_pixels_8x4_sym(ctx->blocks[4], ptr_y + ctx->dct_y_offset, ctx->m.linesize); + ctx->get_pixels_8x4_sym(ctx->blocks[5], ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize); + ctx->get_pixels_8x4_sym(ctx->blocks[6], ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize); + ctx->get_pixels_8x4_sym(ctx->blocks[7], ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize); + } else { + dsp->clear_block(ctx->blocks[4]); + dsp->clear_block(ctx->blocks[5]); + dsp->clear_block(ctx->blocks[6]); + dsp->clear_block(ctx->blocks[7]); + } } else { - dsp->get_pixels(ctx->blocks[4], ptr_y + ctx->dct_y_offset , ctx->m.linesize); - dsp->get_pixels(ctx->blocks[5], ptr_y + ctx->dct_y_offset + 8, ctx->m.linesize); - dsp->get_pixels(ctx->blocks[6], ptr_u + ctx->dct_uv_offset , ctx->m.uvlinesize); - dsp->get_pixels(ctx->blocks[7], ptr_v + ctx->dct_uv_offset , ctx->m.uvlinesize); + dsp->get_pixels(ctx->blocks[4], ptr_y + ctx->dct_y_offset, ctx->m.linesize); + dsp->get_pixels(ctx->blocks[5], ptr_y + ctx->dct_y_offset + bw, ctx->m.linesize); + dsp->get_pixels(ctx->blocks[6], ptr_u + ctx->dct_uv_offset, ctx->m.uvlinesize); + dsp->get_pixels(ctx->blocks[7], ptr_v + ctx->dct_uv_offset, ctx->m.uvlinesize); } } @@ -439,123 +507,150 @@ static av_always_inline int dnxhd_switch_matrix(DNXHDEncContext *ctx, int i) } } -static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg) +static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr) { - DNXHDEncContext *ctx = arg; - int mb_y, mb_x; - int qscale = ctx->thread[0]->qscale; - - for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) { - ctx->m.last_dc[0] = - ctx->m.last_dc[1] = - ctx->m.last_dc[2] = 1024; - - for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) { - unsigned mb = mb_y * ctx->m.mb_width + mb_x; - int ssd = 0; - int ac_bits = 0; - int dc_bits = 0; - int i; - - dnxhd_get_blocks(ctx, mb_x, mb_y); - - for (i = 0; i < 8; i++) { - DECLARE_ALIGNED_16(DCTELEM, block[64]); - DCTELEM *src_block = ctx->blocks[i]; - int overflow, nbits, diff, last_index; - int n = dnxhd_switch_matrix(ctx, i); - - memcpy(block, src_block, sizeof(block)); - last_index = ctx->m.dct_quantize((MpegEncContext*)ctx, block, i, qscale, &overflow); - ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index); - - diff = block[0] - ctx->m.last_dc[n]; - if (diff < 0) nbits = av_log2_16bit(-2*diff); - else nbits = av_log2_16bit( 2*diff); - dc_bits += ctx->cid_table->dc_bits[nbits] + nbits; - - ctx->m.last_dc[n] = block[0]; - - if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) { - dnxhd_unquantize_c(ctx, block, i, qscale, last_index); - ctx->m.dsp.idct(block); - ssd += dnxhd_ssd_block(block, src_block); - } + DNXHDEncContext *ctx = avctx->priv_data; + int mb_y = jobnr, mb_x; + int qscale = ctx->qscale; + LOCAL_ALIGNED_16(DCTELEM, block, [64]); + ctx = ctx->thread[threadnr]; + + ctx->m.last_dc[0] = + ctx->m.last_dc[1] = + ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2); + + for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) { + unsigned mb = mb_y * ctx->m.mb_width + mb_x; + int ssd = 0; + int ac_bits = 0; + int dc_bits = 0; + int i; + + dnxhd_get_blocks(ctx, mb_x, mb_y); + + for (i = 0; i < 8; i++) { + DCTELEM *src_block = ctx->blocks[i]; + int overflow, nbits, diff, last_index; + int n = dnxhd_switch_matrix(ctx, i); + + memcpy(block, src_block, 64*sizeof(*block)); + last_index = ctx->m.dct_quantize(&ctx->m, block, i, qscale, &overflow); + ac_bits += dnxhd_calc_ac_bits(ctx, block, last_index); + + diff = block[0] - ctx->m.last_dc[n]; + if (diff < 0) nbits = av_log2_16bit(-2*diff); + else nbits = av_log2_16bit( 2*diff); + + assert(nbits < ctx->cid_table->bit_depth + 4); + dc_bits += ctx->cid_table->dc_bits[nbits] + nbits; + + ctx->m.last_dc[n] = block[0]; + + if (avctx->mb_decision == FF_MB_DECISION_RD || !RC_VARIANCE) { + dnxhd_unquantize_c(ctx, block, i, qscale, last_index); + ctx->m.dsp.idct(block); + ssd += dnxhd_ssd_block(block, src_block); } - ctx->mb_rc[qscale][mb].ssd = ssd; - ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->table_vlc_bits[0]; } + ctx->mb_rc[qscale][mb].ssd = ssd; + ctx->mb_rc[qscale][mb].bits = ac_bits+dc_bits+12+8*ctx->vlc_bits[0]; } return 0; } -static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg) +static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr) { - DNXHDEncContext *ctx = arg; - int mb_y, mb_x; - - for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) { - ctx->m.last_dc[0] = - ctx->m.last_dc[1] = - ctx->m.last_dc[2] = 1024; - for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) { - unsigned mb = mb_y * ctx->m.mb_width + mb_x; - int qscale = ctx->mb_qscale[mb]; - int i; - - put_bits(&ctx->m.pb, 12, qscale<<1); - - dnxhd_get_blocks(ctx, mb_x, mb_y); - - for (i = 0; i < 8; i++) { - DCTELEM *block = ctx->blocks[i]; - int last_index, overflow; - int n = dnxhd_switch_matrix(ctx, i); - last_index = ctx->m.dct_quantize((MpegEncContext*)ctx, block, i, qscale, &overflow); - dnxhd_encode_block(ctx, block, last_index, n); - } + DNXHDEncContext *ctx = avctx->priv_data; + int mb_y = jobnr, mb_x; + ctx = ctx->thread[threadnr]; + init_put_bits(&ctx->m.pb, (uint8_t *)arg + 640 + ctx->slice_offs[jobnr], ctx->slice_size[jobnr]); + + ctx->m.last_dc[0] = + ctx->m.last_dc[1] = + ctx->m.last_dc[2] = 1 << (ctx->cid_table->bit_depth + 2); + for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) { + unsigned mb = mb_y * ctx->m.mb_width + mb_x; + int qscale = ctx->mb_qscale[mb]; + int i; + + put_bits(&ctx->m.pb, 12, qscale<<1); + + dnxhd_get_blocks(ctx, mb_x, mb_y); + + for (i = 0; i < 8; i++) { + DCTELEM *block = ctx->blocks[i]; + int overflow, n = dnxhd_switch_matrix(ctx, i); + int last_index = ctx->m.dct_quantize(&ctx->m, block, i, + qscale, &overflow); + //START_TIMER; + dnxhd_encode_block(ctx, block, last_index, n); + //STOP_TIMER("encode_block"); } - if (put_bits_count(&ctx->m.pb)&31) - put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0); } + if (put_bits_count(&ctx->m.pb)&31) + put_bits(&ctx->m.pb, 32-(put_bits_count(&ctx->m.pb)&31), 0); flush_put_bits(&ctx->m.pb); return 0; } -static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx, uint8_t *buf) +static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx) { int mb_y, mb_x; - int i, offset = 0; - for (i = 0; i < ctx->m.avctx->thread_count; i++) { - int thread_size = 0; - for (mb_y = ctx->thread[i]->m.start_mb_y; mb_y < ctx->thread[i]->m.end_mb_y; mb_y++) { - ctx->slice_size[mb_y] = 0; - for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) { - unsigned mb = mb_y * ctx->m.mb_width + mb_x; - ctx->slice_size[mb_y] += ctx->mb_bits[mb]; - } - ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31; - ctx->slice_size[mb_y] >>= 3; - thread_size += ctx->slice_size[mb_y]; + int offset = 0; + for (mb_y = 0; mb_y < ctx->m.mb_height; mb_y++) { + int thread_size; + ctx->slice_offs[mb_y] = offset; + ctx->slice_size[mb_y] = 0; + for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) { + unsigned mb = mb_y * ctx->m.mb_width + mb_x; + ctx->slice_size[mb_y] += ctx->mb_bits[mb]; } - init_put_bits(&ctx->thread[i]->m.pb, buf + 640 + offset, thread_size); + ctx->slice_size[mb_y] = (ctx->slice_size[mb_y]+31)&~31; + ctx->slice_size[mb_y] >>= 3; + thread_size = ctx->slice_size[mb_y]; offset += thread_size; } } -static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg) +static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr) { - DNXHDEncContext *ctx = arg; - int mb_y, mb_x; - for (mb_y = ctx->m.start_mb_y; mb_y < ctx->m.end_mb_y; mb_y++) { - for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) { + DNXHDEncContext *ctx = avctx->priv_data; + int mb_y = jobnr, mb_x; + ctx = ctx->thread[threadnr]; + if (ctx->cid_table->bit_depth == 8) { + uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y<<4) * ctx->m.linesize); + for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x, pix += 16) { unsigned mb = mb_y * ctx->m.mb_width + mb_x; - uint8_t *pix = ctx->thread[0]->src[0] + ((mb_y<<4) * ctx->m.linesize) + (mb_x<<4); - int sum = ctx->m.dsp.pix_sum(pix, ctx->m.linesize); - int varc = (ctx->m.dsp.pix_norm1(pix, ctx->m.linesize) - (((unsigned)(sum*sum))>>8)+128)>>8; + int sum = ctx->m.dsp.pix_sum(pix, ctx->m.linesize); + int varc = (ctx->m.dsp.pix_norm1(pix, ctx->m.linesize) - (((unsigned)sum*sum)>>8)+128)>>8; ctx->mb_cmp[mb].value = varc; ctx->mb_cmp[mb].mb = mb; } + } else { // 10-bit + int const linesize = ctx->m.linesize >> 1; + for (mb_x = 0; mb_x < ctx->m.mb_width; ++mb_x) { + uint16_t *pix = (uint16_t*)ctx->thread[0]->src[0] + ((mb_y << 4) * linesize) + (mb_x << 4); + unsigned mb = mb_y * ctx->m.mb_width + mb_x; + int sum = 0; + int sqsum = 0; + int mean, sqmean; + int i, j; + // Macroblocks are 16x16 pixels, unlike DCT blocks which are 8x8. + for (i = 0; i < 16; ++i) { + for (j = 0; j < 16; ++j) { + // Turn 16-bit pixels into 10-bit ones. + int const sample = (unsigned)pix[j] >> 6; + sum += sample; + sqsum += sample * sample; + // 2^10 * 2^10 * 16 * 16 = 2^28, which is less than INT_MAX + } + pix += linesize; + } + mean = sum >> 8; // 16*16 == 2^8 + sqmean = sqsum >> 8; + ctx->mb_cmp[mb].value = sqmean - mean * mean; + ctx->mb_cmp[mb].mb = mb; + } } return 0; } @@ -568,7 +663,7 @@ static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx) for (q = 1; q < avctx->qmax; q++) { ctx->qscale = q; - avctx->execute(avctx, dnxhd_calc_bits_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count); + avctx->execute2(avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height); } up_step = down_step = 2<lambda; @@ -586,7 +681,8 @@ static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx) int qscale = 1; int mb = y*ctx->m.mb_width+x; for (q = 1; q < avctx->qmax; q++) { - unsigned score = ctx->mb_rc[q][mb].bits*lambda+(ctx->mb_rc[q][mb].ssd<mb_rc[q][mb].bits*lambda+ + ((unsigned)ctx->mb_rc[q][mb].ssd< ctx->frame_bits) break; } - //dprintf(ctx->m.avctx, "lambda %d, up %u, down %u, bits %d, frame %d\n", + //av_dlog(ctx->m.avctx, "lambda %d, up %u, down %u, bits %d, frame %d\n", // lambda, last_higher, last_lower, bits, ctx->frame_bits); if (end) { if (bits > ctx->frame_bits) @@ -613,7 +709,7 @@ static int dnxhd_encode_rdo(AVCodecContext *avctx, DNXHDEncContext *ctx) lambda = (lambda+last_higher)>>1; else lambda -= down_step; - down_step *= 5; // XXX tune ? + down_step = FFMIN((int64_t)down_step*5, INT_MAX); up_step = 1<>1; + else if ((int64_t)lambda + up_step > INT_MAX) + return -1; else lambda += up_step; - up_step *= 5; + up_step = FFMIN((int64_t)up_step*5, INT_MAX); down_step = 1<m.avctx, "out lambda %d\n", lambda); + //av_dlog(ctx->m.avctx, "out lambda %d\n", lambda); ctx->lambda = lambda; return 0; } @@ -648,7 +746,7 @@ static int dnxhd_find_qscale(DNXHDEncContext *ctx) bits = 0; ctx->qscale = qscale; // XXX avoid recalculating bits - ctx->m.avctx->execute(ctx->m.avctx, dnxhd_calc_bits_thread, (void**)&ctx->thread[0], NULL, ctx->m.avctx->thread_count); + ctx->m.avctx->execute2(ctx->m.avctx, dnxhd_calc_bits_thread, NULL, NULL, ctx->m.mb_height); for (y = 0; y < ctx->m.mb_height; y++) { for (x = 0; x < ctx->m.mb_width; x++) bits += ctx->mb_rc[qscale][y*ctx->m.mb_width+x].bits; @@ -656,7 +754,7 @@ static int dnxhd_find_qscale(DNXHDEncContext *ctx) if (bits > ctx->frame_bits) break; } - //dprintf(ctx->m.avctx, "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n", + //av_dlog(ctx->m.avctx, "%d, qscale %d, bits %d, frame %d, higher %d, lower %d\n", // ctx->m.avctx->frame_number, qscale, bits, ctx->frame_bits, last_higher, last_lower); if (bits < ctx->frame_bits) { if (qscale == 1) @@ -686,14 +784,65 @@ static int dnxhd_find_qscale(DNXHDEncContext *ctx) return -1; } } - //dprintf(ctx->m.avctx, "out qscale %d\n", qscale); + //av_dlog(ctx->m.avctx, "out qscale %d\n", qscale); ctx->qscale = qscale; return 0; } -static int dnxhd_rc_cmp(const void *a, const void *b) +#define BUCKET_BITS 8 +#define RADIX_PASSES 4 +#define NBUCKETS (1 << BUCKET_BITS) + +static inline int get_bucket(int value, int shift) +{ + value >>= shift; + value &= NBUCKETS - 1; + return NBUCKETS - 1 - value; +} + +static void radix_count(const RCCMPEntry *data, int size, int buckets[RADIX_PASSES][NBUCKETS]) { - return ((RCCMPEntry *)b)->value - ((RCCMPEntry *)a)->value; + int i, j; + memset(buckets, 0, sizeof(buckets[0][0]) * RADIX_PASSES * NBUCKETS); + for (i = 0; i < size; i++) { + int v = data[i].value; + for (j = 0; j < RADIX_PASSES; j++) { + buckets[j][get_bucket(v, 0)]++; + v >>= BUCKET_BITS; + } + assert(!v); + } + for (j = 0; j < RADIX_PASSES; j++) { + int offset = size; + for (i = NBUCKETS - 1; i >= 0; i--) + buckets[j][i] = offset -= buckets[j][i]; + assert(!buckets[j][0]); + } +} + +static void radix_sort_pass(RCCMPEntry *dst, const RCCMPEntry *data, int size, int buckets[NBUCKETS], int pass) +{ + int shift = pass * BUCKET_BITS; + int i; + for (i = 0; i < size; i++) { + int v = get_bucket(data[i].value, shift); + int pos = buckets[v]++; + dst[pos] = data[i]; + } +} + +static void radix_sort(RCCMPEntry *data, int size) +{ + int buckets[RADIX_PASSES][NBUCKETS]; + RCCMPEntry *tmp = av_malloc(sizeof(*tmp) * size); + radix_count(data, size, buckets); + radix_sort_pass(tmp, data, size, buckets[0], 0); + radix_sort_pass(data, tmp, size, buckets[1], 1); + if (buckets[2][NBUCKETS - 1] || buckets[3][NBUCKETS - 1]) { + radix_sort_pass(tmp, data, size, buckets[2], 2); + radix_sort_pass(data, tmp, size, buckets[3], 3); + } + av_free(tmp); } static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx) @@ -721,8 +870,8 @@ static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx) } if (!ret) { if (RC_VARIANCE) - avctx->execute(avctx, dnxhd_mb_var_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count); - qsort(ctx->mb_cmp, ctx->m.mb_num, sizeof(RCEntry), dnxhd_rc_cmp); + avctx->execute2(avctx, dnxhd_mb_var_thread, NULL, NULL, ctx->m.mb_height); + radix_sort(ctx->mb_cmp, ctx->m.mb_num); for (x = 0; x < ctx->m.mb_num && max_bits > ctx->frame_bits; x++) { int mb = ctx->mb_cmp[x].mb; max_bits -= ctx->mb_rc[ctx->qscale][mb].bits - ctx->mb_rc[ctx->qscale+1][mb].bits; @@ -733,7 +882,7 @@ static int dnxhd_encode_fast(AVCodecContext *avctx, DNXHDEncContext *ctx) return 0; } -static void dnxhd_load_picture(DNXHDEncContext *ctx, AVFrame *frame) +static void dnxhd_load_picture(DNXHDEncContext *ctx, const AVFrame *frame) { int i; @@ -753,18 +902,21 @@ static void dnxhd_load_picture(DNXHDEncContext *ctx, AVFrame *frame) ctx->cur_field = frame->interlaced_frame && !frame->top_field_first; } -static int dnxhd_encode_picture(AVCodecContext *avctx, unsigned char *buf, int buf_size, void *data) +static int dnxhd_encode_picture(AVCodecContext *avctx, AVPacket *pkt, + const AVFrame *frame, int *got_packet) { DNXHDEncContext *ctx = avctx->priv_data; int first_field = 1; int offset, i, ret; + uint8_t *buf; - if (buf_size < ctx->cid_table->frame_size) { + if ((ret = ff_alloc_packet(pkt, ctx->cid_table->frame_size)) < 0) { av_log(avctx, AV_LOG_ERROR, "output buffer is too small to compress picture\n"); - return -1; + return ret; } + buf = pkt->data; - dnxhd_load_picture(ctx, data); + dnxhd_load_picture(ctx, frame); encode_coding_unit: for (i = 0; i < 3; i++) { @@ -780,11 +932,12 @@ static int dnxhd_encode_picture(AVCodecContext *avctx, unsigned char *buf, int b else ret = dnxhd_encode_fast(avctx, ctx); if (ret < 0) { - av_log(avctx, AV_LOG_ERROR, "picture could not fit ratecontrol constraints\n"); + av_log(avctx, AV_LOG_ERROR, + "picture could not fit ratecontrol constraints, increase qmax\n"); return -1; } - dnxhd_setup_threads_slices(ctx, buf); + dnxhd_setup_threads_slices(ctx); offset = 0; for (i = 0; i < ctx->m.mb_height; i++) { @@ -793,7 +946,10 @@ static int dnxhd_encode_picture(AVCodecContext *avctx, unsigned char *buf, int b assert(!(ctx->slice_size[i] & 3)); } - avctx->execute(avctx, dnxhd_encode_thread, (void**)&ctx->thread[0], NULL, avctx->thread_count); + avctx->execute2(avctx, dnxhd_encode_thread, buf, NULL, ctx->m.mb_height); + + assert(640 + offset + 4 <= ctx->cid_table->coding_unit_size); + memset(buf + 640 + offset, 0, ctx->cid_table->coding_unit_size - 4 - offset - 640); AV_WB32(buf + ctx->cid_table->coding_unit_size - 4, 0x600DC0DE); // EOF @@ -801,28 +957,33 @@ static int dnxhd_encode_picture(AVCodecContext *avctx, unsigned char *buf, int b first_field = 0; ctx->cur_field ^= 1; buf += ctx->cid_table->coding_unit_size; - buf_size -= ctx->cid_table->coding_unit_size; goto encode_coding_unit; } - return ctx->cid_table->frame_size; + ctx->frame.quality = ctx->qscale*FF_QP2LAMBDA; + + pkt->flags |= AV_PKT_FLAG_KEY; + *got_packet = 1; + return 0; } static int dnxhd_encode_end(AVCodecContext *avctx) { DNXHDEncContext *ctx = avctx->priv_data; + int max_level = 1<<(ctx->cid_table->bit_depth+2); int i; - av_freep(&ctx->table_vlc_codes); - av_freep(&ctx->table_vlc_bits); - av_freep(&ctx->table_run_codes); - av_freep(&ctx->table_run_bits); + av_free(ctx->vlc_codes-max_level*2); + av_free(ctx->vlc_bits -max_level*2); + av_freep(&ctx->run_codes); + av_freep(&ctx->run_bits); av_freep(&ctx->mb_bits); av_freep(&ctx->mb_qscale); av_freep(&ctx->mb_rc); av_freep(&ctx->mb_cmp); av_freep(&ctx->slice_size); + av_freep(&ctx->slice_offs); av_freep(&ctx->qmatrix_c); av_freep(&ctx->qmatrix_l); @@ -835,13 +996,18 @@ static int dnxhd_encode_end(AVCodecContext *avctx) return 0; } -AVCodec dnxhd_encoder = { - "dnxhd", - CODEC_TYPE_VIDEO, - CODEC_ID_DNXHD, - sizeof(DNXHDEncContext), - dnxhd_encode_init, - dnxhd_encode_picture, - dnxhd_encode_end, - .pix_fmts = (enum PixelFormat[]){PIX_FMT_YUV422P, -1}, +AVCodec ff_dnxhd_encoder = { + .name = "dnxhd", + .type = AVMEDIA_TYPE_VIDEO, + .id = AV_CODEC_ID_DNXHD, + .priv_data_size = sizeof(DNXHDEncContext), + .init = dnxhd_encode_init, + .encode2 = dnxhd_encode_picture, + .close = dnxhd_encode_end, + .capabilities = CODEC_CAP_SLICE_THREADS, + .pix_fmts = (const enum AVPixelFormat[]){ AV_PIX_FMT_YUV422P, + AV_PIX_FMT_YUV422P10, + AV_PIX_FMT_NONE }, + .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"), + .priv_class = &class, };