X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libavcodec%2Fdnxhdenc.c;h=990c3b21b4a49f1ba6e6d422f4df386d0aa31f66;hb=30517a9f056c8d01b3790871db4aee95d9c176e6;hp=958f9d4081b76ff7aafc502e1beb621d5b0082cc;hpb=2912e87a6c9264d556734e2bf94a99c64cf9b102;p=ffmpeg diff --git a/libavcodec/dnxhdenc.c b/libavcodec/dnxhdenc.c index 958f9d4081b..990c3b21b4a 100644 --- a/libavcodec/dnxhdenc.c +++ b/libavcodec/dnxhdenc.c @@ -1,8 +1,10 @@ /* * 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 Libav. * @@ -21,28 +23,30 @@ * 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/attributes.h" +#include "libavutil/internal.h" #include "libavutil/opt.h" +#include "libavutil/timer.h" #include "avcodec.h" #include "dsputil.h" +#include "internal.h" #include "mpegvideo.h" #include "dnxhdenc.h" #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. static const AVOption options[]={ - {"nitris_compat", "encode with Avid Nitris compatibility", offsetof(DNXHDEncContext, nitris_compat), FF_OPT_TYPE_INT, 0, 0, 1, VE}, + {"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 }; -int dct_quantize_c(MpegEncContext *s, DCTELEM *block, int n, int qscale, int *overflow); - #define LAMBDA_FRAC_BITS 10 -static av_always_inline void dnxhd_get_pixels_8x4(DCTELEM *restrict block, const uint8_t *pixels, int line_size) +static void dnxhd_8bit_get_pixels_8x4_sym(int16_t *restrict block, const uint8_t *pixels, int line_size) { int i; for (i = 0; i < 4; i++) { @@ -53,21 +57,59 @@ static av_always_inline void dnxhd_get_pixels_8x4(DCTELEM *restrict block, const 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); + 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 int dnxhd_init_vlc(DNXHDEncContext *ctx) +static av_always_inline void dnxhd_10bit_get_pixels_8x4_sym(int16_t *restrict block, const uint8_t *pixels, int line_size) +{ + int i; + + block += 32; + + 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)); + } +} + +static int dnxhd_10bit_dct_quantize(MpegEncContext *ctx, int16_t *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; + + ctx->dsp.fdct(block); + + // Divide by 4 with rounding, to compensate scaling of DCT coefficients + block[0] = (block[0] + 2) >> 2; + + 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; + } + + return last_non_zero; +} + +static av_cold int dnxhd_init_vlc(DNXHDEncContext *ctx) { 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); + 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; @@ -114,42 +156,66 @@ static int dnxhd_init_vlc(DNXHDEncContext *ctx) return -1; } -static int dnxhd_init_qmat(DNXHDEncContext *ctx, int lbias, int cbias) +static av_cold 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; + 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_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); - 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; + 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; } -static int dnxhd_init_rc(DNXHDEncContext *ctx) +static av_cold int dnxhd_init_rc(DNXHDEncContext *ctx) { 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) @@ -163,13 +229,25 @@ static int dnxhd_init_rc(DNXHDEncContext *ctx) return -1; } -static int dnxhd_encode_init(AVCodecContext *avctx) +static av_cold int dnxhd_encode_init(AVCodecContext *avctx) { DNXHDEncContext *ctx = avctx->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; } @@ -182,15 +260,24 @@ static int dnxhd_encode_init(AVCodecContext *avctx) ctx->m.mb_intra = 1; ctx->m.h263_aic = 1; - ctx->get_pixels_8x4_sym = dnxhd_get_pixels_8x4; + avctx->bits_per_raw_sample = ctx->cid_table->bit_depth; - dsputil_init(&ctx->m.dsp, avctx); + ff_dsputil_init(&ctx->m.dsp, avctx); ff_dct_common_init(&ctx->m); -#if HAVE_MMX - ff_dnxhd_init_mmx(ctx); -#endif 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; @@ -219,11 +306,14 @@ static int dnxhd_encode_init(AVCodecContext *avctx) 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); + 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->m.avctx->coded_frame = &ctx->frame; + avctx->coded_frame = av_frame_alloc(); + if (!avctx->coded_frame) + return AVERROR(ENOMEM); + + avctx->coded_frame->key_frame = 1; + avctx->coded_frame->pict_type = AV_PICTURE_TYPE_I; if (avctx->thread_count > MAX_THREADS) { av_log(avctx, AV_LOG_ERROR, "too many threads\n"); @@ -256,7 +346,7 @@ static int dnxhd_write_header(AVCodecContext *avctx, uint8_t *buf) AV_WB16(buf + 0x1a, avctx->width); // SPL AV_WB16(buf + 0x1d, avctx->height>>ctx->interlaced); // NAL - buf[0x21] = 0x38; // FIXME 8 bit per comp + 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; @@ -285,7 +375,7 @@ static av_always_inline void dnxhd_encode_dc(DNXHDEncContext *ctx, int diff) (ctx->cid_table->dc_codes[nbits]<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) +static av_always_inline void dnxhd_unquantize_c(DNXHDEncContext *ctx, int16_t *block, int n, int qscale, int last_index) { const uint8_t *weight_matrix; int level; @@ -322,31 +412,43 @@ 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; } } } -static av_always_inline int dnxhd_ssd_block(DCTELEM *qblock, DCTELEM *block) +static av_always_inline int dnxhd_ssd_block(int16_t *qblock, int16_t *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; } -static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, DCTELEM *block, int last_index) +static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, int16_t *block, int last_index) { int last_non_zero = 0; int bits = 0; @@ -365,31 +467,35 @@ static av_always_inline int dnxhd_calc_ac_bits(DNXHDEncContext *ctx, DCTELEM *bl 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) { - 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 + 8, 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); + 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]); + 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); } } @@ -411,12 +517,12 @@ static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg, int jobnr, i DNXHDEncContext *ctx = avctx->priv_data; int mb_y = jobnr, mb_x; int qscale = ctx->qscale; - LOCAL_ALIGNED_16(DCTELEM, block, [64]); + LOCAL_ALIGNED_16(int16_t, block, [64]); ctx = ctx->thread[threadnr]; ctx->m.last_dc[0] = ctx->m.last_dc[1] = - ctx->m.last_dc[2] = 1024; + 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; @@ -428,17 +534,19 @@ static int dnxhd_calc_bits_thread(AVCodecContext *avctx, void *arg, int jobnr, i dnxhd_get_blocks(ctx, mb_x, mb_y); for (i = 0; i < 8; i++) { - DCTELEM *src_block = ctx->blocks[i]; + int16_t *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((MpegEncContext*)ctx, block, i, qscale, &overflow); + 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]; @@ -464,7 +572,7 @@ static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg, int jobnr, int ctx->m.last_dc[0] = ctx->m.last_dc[1] = - ctx->m.last_dc[2] = 1024; + 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]; @@ -475,10 +583,10 @@ static int dnxhd_encode_thread(AVCodecContext *avctx, void *arg, int jobnr, int 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); + int16_t *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"); @@ -497,14 +605,14 @@ static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx) 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]; - } - 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]; + 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]; offset += thread_size; } } @@ -512,15 +620,63 @@ static void dnxhd_setup_threads_slices(DNXHDEncContext *ctx) static int dnxhd_mb_var_thread(AVCodecContext *avctx, void *arg, int jobnr, int threadnr) { DNXHDEncContext *ctx = avctx->priv_data; - int mb_y = jobnr, mb_x; + int mb_y = jobnr, mb_x, x, y; + int partial_last_row = (mb_y == ctx->m.mb_height - 1) && + ((avctx->height >> ctx->interlaced) & 0xF); + ctx = ctx->thread[threadnr]; - for (mb_x = 0; mb_x < ctx->m.mb_width; mb_x++) { - 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; - ctx->mb_cmp[mb].value = varc; - ctx->mb_cmp[mb].mb = mb; + 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; + int sum; + int varc; + + if (!partial_last_row && mb_x * 16 <= avctx->width - 16) { + sum = ctx->m.dsp.pix_sum(pix, ctx->m.linesize); + varc = ctx->m.dsp.pix_norm1(pix, ctx->m.linesize); + } else { + int bw = FFMIN(avctx->width - 16 * mb_x, 16); + int bh = FFMIN((avctx->height >> ctx->interlaced) - 16 * mb_y, 16); + sum = varc = 0; + for (y = 0; y < bh; y++) { + for (x = 0; x < bw; x++) { + uint8_t val = pix[x + y * ctx->m.linesize]; + sum += val; + varc += val * val; + } + } + } + varc = (varc - (((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; } @@ -551,7 +707,8 @@ 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<>1; else lambda -= down_step; - down_step *= 5; // XXX tune ? + down_step = FFMIN((int64_t)down_step*5, INT_MAX); up_step = 1<frame.data[i] = frame->data[i]; - ctx->frame.linesize[i] = frame->linesize[i]; - } - for (i = 0; i < ctx->m.avctx->thread_count; i++) { - ctx->thread[i]->m.linesize = ctx->frame.linesize[0]<interlaced; - ctx->thread[i]->m.uvlinesize = ctx->frame.linesize[1]<interlaced; + ctx->thread[i]->m.linesize = frame->linesize[0] << ctx->interlaced; + ctx->thread[i]->m.uvlinesize = frame->linesize[1] << ctx->interlaced; ctx->thread[i]->dct_y_offset = ctx->m.linesize *8; ctx->thread[i]->dct_uv_offset = ctx->m.uvlinesize*8; } - ctx->frame.interlaced_frame = frame->interlaced_frame; + ctx->m.avctx->coded_frame->interlaced_frame = frame->interlaced_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++) { - ctx->src[i] = ctx->frame.data[i]; + ctx->src[i] = frame->data[i]; if (ctx->interlaced && ctx->cur_field) - ctx->src[i] += ctx->frame.linesize[i]; + ctx->src[i] += frame->linesize[i]; } dnxhd_write_header(avctx, buf); @@ -823,16 +978,17 @@ 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; } - ctx->frame.quality = ctx->qscale*FF_QP2LAMBDA; + avctx->coded_frame->quality = ctx->qscale * FF_QP2LAMBDA; - return ctx->cid_table->frame_size; + pkt->flags |= AV_PKT_FLAG_KEY; + *got_packet = 1; + return 0; } -static int dnxhd_encode_end(AVCodecContext *avctx) +static av_cold int dnxhd_encode_end(AVCodecContext *avctx) { DNXHDEncContext *ctx = avctx->priv_data; int max_level = 1<<(ctx->cid_table->bit_depth+2); @@ -858,18 +1014,23 @@ static int dnxhd_encode_end(AVCodecContext *avctx) for (i = 1; i < avctx->thread_count; i++) av_freep(&ctx->thread[i]); + av_frame_free(&avctx->coded_frame); + return 0; } AVCodec ff_dnxhd_encoder = { - "dnxhd", - AVMEDIA_TYPE_VIDEO, - CODEC_ID_DNXHD, - sizeof(DNXHDEncContext), - dnxhd_encode_init, - dnxhd_encode_picture, - dnxhd_encode_end, - .pix_fmts = (const enum PixelFormat[]){PIX_FMT_YUV422P, PIX_FMT_NONE}, - .long_name = NULL_IF_CONFIG_SMALL("VC3/DNxHD"), - .priv_class = &class, + .name = "dnxhd", + .long_name = NULL_IF_CONFIG_SMALL("VC3/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 }, + .priv_class = &class, };