X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=libavcodec%2Fvp8.c;h=833819890de4a167db0d2454c8d5cd68416ca7ec;hb=c18365402bbb6fbfa7854b47c3288f9e31f39f44;hp=2a0f93bb1671ad9b02b7ba0b945491523c8167b9;hpb=b0d58795138c738c45c5b738418765939021b49f;p=ffmpeg diff --git a/libavcodec/vp8.c b/libavcodec/vp8.c index 2a0f93bb167..833819890de 100644 --- a/libavcodec/vp8.c +++ b/libavcodec/vp8.c @@ -5,249 +5,130 @@ * Copyright (C) 2010 Ronald S. Bultje * Copyright (C) 2010 Jason Garrett-Glaser * - * 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 */ +#include "libavutil/imgutils.h" #include "avcodec.h" -#include "vp56.h" +#include "internal.h" +#include "vp8.h" #include "vp8data.h" -#include "vp8dsp.h" -#include "h264pred.h" #include "rectangle.h" +#include "thread.h" -typedef struct { - uint8_t filter_level; - uint8_t inner_limit; - uint8_t inner_filter; -} VP8FilterStrength; - -typedef struct { - uint8_t skip; - // todo: make it possible to check for at least (i4x4 or split_mv) - // in one op. are others needed? - uint8_t mode; - uint8_t ref_frame; - uint8_t partitioning; - VP56mv mv; - VP56mv bmv[16]; -} VP8Macroblock; - -typedef struct { - AVCodecContext *avctx; - DSPContext dsp; - VP8DSPContext vp8dsp; - H264PredContext hpc; - vp8_mc_func put_pixels_tab[3][3][3]; - AVFrame frames[4]; - AVFrame *framep[4]; - uint8_t *edge_emu_buffer; - VP56RangeCoder c; ///< header context, includes mb modes and motion vectors - int profile; - - int mb_width; /* number of horizontal MB */ - int mb_height; /* number of vertical MB */ - int linesize; - int uvlinesize; - - int keyframe; - int invisible; - int update_last; ///< update VP56_FRAME_PREVIOUS with the current one - int update_golden; ///< VP56_FRAME_NONE if not updated, or which frame to copy if so - int update_altref; - int deblock_filter; - - /** - * If this flag is not set, all the probability updates - * are discarded after this frame is decoded. - */ - int update_probabilities; - - /** - * All coefficients are contained in separate arith coding contexts. - * There can be 1, 2, 4, or 8 of these after the header context. - */ - int num_coeff_partitions; - VP56RangeCoder coeff_partition[8]; - - VP8Macroblock *macroblocks; - VP8Macroblock *macroblocks_base; - VP8FilterStrength *filter_strength; - int mb_stride; - - uint8_t *intra4x4_pred_mode_top; - uint8_t intra4x4_pred_mode_left[4]; - uint8_t *segmentation_map; - int b4_stride; - - /** - * Cache of the top row needed for intra prediction - * 16 for luma, 8 for each chroma plane - */ - uint8_t (*top_border)[16+8+8]; - - /** - * For coeff decode, we need to know whether the above block had non-zero - * coefficients. This means for each macroblock, we need data for 4 luma - * blocks, 2 u blocks, 2 v blocks, and the luma dc block, for a total of 9 - * per macroblock. We keep the last row in top_nnz. - */ - uint8_t (*top_nnz)[9]; - DECLARE_ALIGNED(8, uint8_t, left_nnz)[9]; - - /** - * This is the index plus one of the last non-zero coeff - * for each of the blocks in the current macroblock. - * So, 0 -> no coeffs - * 1 -> dc-only (special transform) - * 2+-> full transform - */ - DECLARE_ALIGNED(16, uint8_t, non_zero_count_cache)[6][4]; - DECLARE_ALIGNED(16, DCTELEM, block)[6][4][16]; - DECLARE_ALIGNED(16, DCTELEM, block_dc)[16]; - uint8_t intra4x4_pred_mode_mb[16]; - - int chroma_pred_mode; ///< 8x8c pred mode of the current macroblock - int segment; ///< segment of the current macroblock - - int mbskip_enabled; - int sign_bias[4]; ///< one state [0, 1] per ref frame type - int ref_count[3]; - - /** - * Base parameters for segmentation, i.e. per-macroblock parameters. - * These must be kept unchanged even if segmentation is not used for - * a frame, since the values persist between interframes. - */ - struct { - int enabled; - int absolute_vals; - int update_map; - int8_t base_quant[4]; - int8_t filter_level[4]; ///< base loop filter level - } segmentation; - - /** - * Macroblocks can have one of 4 different quants in a frame when - * segmentation is enabled. - * If segmentation is disabled, only the first segment's values are used. - */ - struct { - // [0] - DC qmul [1] - AC qmul - int16_t luma_qmul[2]; - int16_t luma_dc_qmul[2]; ///< luma dc-only block quant - int16_t chroma_qmul[2]; - } qmat[4]; - - struct { - int simple; - int level; - int sharpness; - } filter; - - struct { - int enabled; ///< whether each mb can have a different strength based on mode/ref - - /** - * filter strength adjustment for the following macroblock modes: - * [0] - i4x4 - * [1] - zero mv - * [2] - inter modes except for zero or split mv - * [3] - split mv - * i16x16 modes never have any adjustment - */ - int8_t mode[4]; - - /** - * filter strength adjustment for macroblocks that reference: - * [0] - intra / VP56_FRAME_CURRENT - * [1] - VP56_FRAME_PREVIOUS - * [2] - VP56_FRAME_GOLDEN - * [3] - altref / VP56_FRAME_GOLDEN2 - */ - int8_t ref[4]; - } lf_delta; - - /** - * These are all of the updatable probabilities for binary decisions. - * They are only implictly reset on keyframes, making it quite likely - * for an interframe to desync if a prior frame's header was corrupt - * or missing outright! - */ - struct { - uint8_t segmentid[3]; - uint8_t mbskip; - uint8_t intra; - uint8_t last; - uint8_t golden; - uint8_t pred16x16[4]; - uint8_t pred8x8c[3]; - /* Padded to allow overreads */ - uint8_t token[4][17][3][NUM_DCT_TOKENS-1]; - uint8_t mvc[2][19]; - } prob[2]; -} VP8Context; +#if ARCH_ARM +# include "arm/vp8.h" +#endif -static void vp8_decode_flush(AVCodecContext *avctx) +static void free_buffers(VP8Context *s) { - VP8Context *s = avctx->priv_data; - int i; - - for (i = 0; i < 4; i++) - if (s->frames[i].data[0]) - avctx->release_buffer(avctx, &s->frames[i]); - memset(s->framep, 0, sizeof(s->framep)); - av_freep(&s->macroblocks_base); av_freep(&s->filter_strength); av_freep(&s->intra4x4_pred_mode_top); av_freep(&s->top_nnz); av_freep(&s->edge_emu_buffer); av_freep(&s->top_border); - av_freep(&s->segmentation_map); - s->macroblocks = NULL; + s->macroblocks = NULL; +} + +static int vp8_alloc_frame(VP8Context *s, AVFrame *f) +{ + int ret; + if ((ret = ff_thread_get_buffer(s->avctx, f)) < 0) + return ret; + if (s->num_maps_to_be_freed && !s->maps_are_invalid) { + f->ref_index[0] = s->segmentation_maps[--s->num_maps_to_be_freed]; + } else if (!(f->ref_index[0] = av_mallocz(s->mb_width * s->mb_height))) { + ff_thread_release_buffer(s->avctx, f); + return AVERROR(ENOMEM); + } + return 0; +} + +static void vp8_release_frame(VP8Context *s, AVFrame *f, int prefer_delayed_free, int can_direct_free) +{ + if (f->ref_index[0]) { + if (prefer_delayed_free) { + /* Upon a size change, we want to free the maps but other threads may still + * be using them, so queue them. Upon a seek, all threads are inactive so + * we want to cache one to prevent re-allocation in the next decoding + * iteration, but the rest we can free directly. */ + int max_queued_maps = can_direct_free ? 1 : FF_ARRAY_ELEMS(s->segmentation_maps); + if (s->num_maps_to_be_freed < max_queued_maps) { + s->segmentation_maps[s->num_maps_to_be_freed++] = f->ref_index[0]; + } else if (can_direct_free) /* vp8_decode_flush(), but our queue is full */ { + av_free(f->ref_index[0]); + } /* else: MEMLEAK (should never happen, but better that than crash) */ + f->ref_index[0] = NULL; + } else /* vp8_decode_free() */ { + av_free(f->ref_index[0]); + } + } + ff_thread_release_buffer(s->avctx, f); +} + +static void vp8_decode_flush_impl(AVCodecContext *avctx, + int prefer_delayed_free, int can_direct_free, int free_mem) +{ + VP8Context *s = avctx->priv_data; + int i; + + if (!avctx->internal->is_copy) { + for (i = 0; i < 5; i++) + if (s->frames[i].data[0]) + vp8_release_frame(s, &s->frames[i], prefer_delayed_free, can_direct_free); + } + memset(s->framep, 0, sizeof(s->framep)); + + if (free_mem) { + free_buffers(s); + s->maps_are_invalid = 1; + } +} + +static void vp8_decode_flush(AVCodecContext *avctx) +{ + vp8_decode_flush_impl(avctx, 1, 1, 0); } static int update_dimensions(VP8Context *s, int width, int height) { - if (avcodec_check_dimensions(s->avctx, width, height)) - return AVERROR_INVALIDDATA; + if (width != s->avctx->width || + height != s->avctx->height) { + if (av_image_check_size(width, height, 0, s->avctx)) + return AVERROR_INVALIDDATA; - vp8_decode_flush(s->avctx); + vp8_decode_flush_impl(s->avctx, 1, 0, 1); - avcodec_set_dimensions(s->avctx, width, height); + avcodec_set_dimensions(s->avctx, width, height); + } s->mb_width = (s->avctx->coded_width +15) / 16; s->mb_height = (s->avctx->coded_height+15) / 16; - // we allocate a border around the top/left of intra4x4 modes - // this is 4 blocks for intra4x4 to keep 4-byte alignment for fill_rectangle - s->mb_stride = s->mb_width+1; - s->b4_stride = 4*s->mb_stride; - - s->macroblocks_base = av_mallocz((s->mb_stride+s->mb_height*2+2)*sizeof(*s->macroblocks)); - s->filter_strength = av_mallocz(s->mb_stride*sizeof(*s->filter_strength)); - s->intra4x4_pred_mode_top = av_mallocz(s->b4_stride*4); + s->macroblocks_base = av_mallocz((s->mb_width+s->mb_height*2+1)*sizeof(*s->macroblocks)); + s->filter_strength = av_mallocz(s->mb_width*sizeof(*s->filter_strength)); + s->intra4x4_pred_mode_top = av_mallocz(s->mb_width*4); s->top_nnz = av_mallocz(s->mb_width*sizeof(*s->top_nnz)); s->top_border = av_mallocz((s->mb_width+1)*sizeof(*s->top_border)); - s->segmentation_map = av_mallocz(s->mb_stride*s->mb_height); if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_top || - !s->top_nnz || !s->top_border || !s->segmentation_map) + !s->top_nnz || !s->top_border) return AVERROR(ENOMEM); s->macroblocks = s->macroblocks_base + 1; @@ -284,7 +165,7 @@ static void update_lf_deltas(VP8Context *s) for (i = 0; i < 4; i++) s->lf_delta.ref[i] = vp8_rac_get_sint(c, 6); - for (i = 0; i < 4; i++) + for (i = MODE_I4x4; i <= VP8_MVMODE_SPLIT; i++) s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6); } @@ -305,11 +186,11 @@ static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size) if (buf_size - size < 0) return -1; - vp56_init_range_decoder(&s->coeff_partition[i], buf, size); + ff_vp56_init_range_decoder(&s->coeff_partition[i], buf, size); buf += size; buf_size -= size; } - vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size); + ff_vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size); return 0; } @@ -334,12 +215,12 @@ static void get_quants(VP8Context *s) } else base_qi = yac_qi; - s->qmat[i].luma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + ydc_delta , 0, 127)]; - s->qmat[i].luma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi , 0, 127)]; - s->qmat[i].luma_dc_qmul[0] = 2 * vp8_dc_qlookup[av_clip(base_qi + y2dc_delta, 0, 127)]; - s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip(base_qi + y2ac_delta, 0, 127)] / 100; - s->qmat[i].chroma_qmul[0] = vp8_dc_qlookup[av_clip(base_qi + uvdc_delta, 0, 127)]; - s->qmat[i].chroma_qmul[1] = vp8_ac_qlookup[av_clip(base_qi + uvac_delta, 0, 127)]; + s->qmat[i].luma_qmul[0] = vp8_dc_qlookup[av_clip_uintp2(base_qi + ydc_delta , 7)]; + s->qmat[i].luma_qmul[1] = vp8_ac_qlookup[av_clip_uintp2(base_qi , 7)]; + s->qmat[i].luma_dc_qmul[0] = 2 * vp8_dc_qlookup[av_clip_uintp2(base_qi + y2dc_delta, 7)]; + s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip_uintp2(base_qi + y2ac_delta, 7)] / 100; + s->qmat[i].chroma_qmul[0] = vp8_dc_qlookup[av_clip_uintp2(base_qi + uvdc_delta, 7)]; + s->qmat[i].chroma_qmul[1] = vp8_ac_qlookup[av_clip_uintp2(base_qi + uvac_delta, 7)]; s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8); s->qmat[i].chroma_qmul[0] = FFMIN(s->qmat[i].chroma_qmul[0], 132); @@ -441,11 +322,11 @@ static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size) if (!s->macroblocks_base || /* first frame */ width != s->avctx->width || height != s->avctx->height) { - if ((ret = update_dimensions(s, width, height) < 0)) + if ((ret = update_dimensions(s, width, height)) < 0) return ret; } - vp56_init_range_decoder(c, buf, header_size); + ff_vp56_init_range_decoder(c, buf, header_size); buf += header_size; buf_size -= header_size; @@ -523,76 +404,10 @@ static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size) return 0; } -static av_always_inline -void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src, int mb_x, int mb_y) -{ -#define MARGIN (16 << 2) - dst->x = av_clip(src->x, -((mb_x << 6) + MARGIN), - ((s->mb_width - 1 - mb_x) << 6) + MARGIN); - dst->y = av_clip(src->y, -((mb_y << 6) + MARGIN), - ((s->mb_height - 1 - mb_y) << 6) + MARGIN); -} - -static av_always_inline -void find_near_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, - VP56mv near[2], VP56mv *best, uint8_t cnt[4]) +static av_always_inline void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src) { - VP8Macroblock *mb_edge[3] = { mb + 2 /* top */, - mb - 1 /* left */, - mb + 1 /* top-left */ }; - enum { EDGE_TOP, EDGE_LEFT, EDGE_TOPLEFT }; - VP56mv near_mv[4] = {{ 0 }}; - enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV }; - int idx = CNT_ZERO; - int best_idx = CNT_ZERO; - int cur_sign_bias = s->sign_bias[mb->ref_frame]; - int *sign_bias = s->sign_bias; - - /* Process MB on top, left and top-left */ - #define MV_EDGE_CHECK(n)\ - {\ - VP8Macroblock *edge = mb_edge[n];\ - int edge_ref = edge->ref_frame;\ - if (edge_ref != VP56_FRAME_CURRENT) {\ - uint32_t mv = AV_RN32A(&edge->mv);\ - if (mv) {\ - if (cur_sign_bias != sign_bias[edge_ref]) {\ - /* SWAR negate of the values in mv. */\ - mv = ~mv;\ - mv = ((mv&0x7fff7fff) + 0x00010001) ^ (mv&0x80008000);\ - }\ - if (!n || mv != AV_RN32A(&near_mv[idx]))\ - AV_WN32A(&near_mv[++idx], mv);\ - cnt[idx] += 1 + (n != 2);\ - } else\ - cnt[CNT_ZERO] += 1 + (n != 2);\ - }\ - } - MV_EDGE_CHECK(0) - MV_EDGE_CHECK(1) - MV_EDGE_CHECK(2) - - /* If we have three distinct MVs, merge first and last if they're the same */ - if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1+EDGE_TOP]) == AV_RN32A(&near_mv[1+EDGE_TOPLEFT])) - cnt[CNT_NEAREST] += 1; - - cnt[CNT_SPLITMV] = ((mb_edge[EDGE_LEFT]->mode == VP8_MVMODE_SPLIT) + - (mb_edge[EDGE_TOP]->mode == VP8_MVMODE_SPLIT)) * 2 + - (mb_edge[EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT); - - /* Swap near and nearest if necessary */ - if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) { - FFSWAP(uint8_t, cnt[CNT_NEAREST], cnt[CNT_NEAR]); - FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]); - } - - /* Choose the best mv out of 0,0 and the nearest mv */ - if (cnt[CNT_NEAREST] >= cnt[CNT_ZERO]) - best_idx = CNT_NEAREST; - - mb->mv = near_mv[best_idx]; - near[0] = near_mv[CNT_NEAREST]; - near[1] = near_mv[CNT_NEAR]; + dst->x = av_clip(src->x, s->mv_min.x, s->mv_max.x); + dst->y = av_clip(src->y, s->mv_min.y, s->mv_max.y); } /** @@ -704,6 +519,96 @@ int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb) return num; } +static av_always_inline +void decode_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y) +{ + VP8Macroblock *mb_edge[3] = { mb + 2 /* top */, + mb - 1 /* left */, + mb + 1 /* top-left */ }; + enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV }; + enum { VP8_EDGE_TOP, VP8_EDGE_LEFT, VP8_EDGE_TOPLEFT }; + int idx = CNT_ZERO; + int cur_sign_bias = s->sign_bias[mb->ref_frame]; + int8_t *sign_bias = s->sign_bias; + VP56mv near_mv[4]; + uint8_t cnt[4] = { 0 }; + VP56RangeCoder *c = &s->c; + + AV_ZERO32(&near_mv[0]); + AV_ZERO32(&near_mv[1]); + AV_ZERO32(&near_mv[2]); + + /* Process MB on top, left and top-left */ + #define MV_EDGE_CHECK(n)\ + {\ + VP8Macroblock *edge = mb_edge[n];\ + int edge_ref = edge->ref_frame;\ + if (edge_ref != VP56_FRAME_CURRENT) {\ + uint32_t mv = AV_RN32A(&edge->mv);\ + if (mv) {\ + if (cur_sign_bias != sign_bias[edge_ref]) {\ + /* SWAR negate of the values in mv. */\ + mv = ~mv;\ + mv = ((mv&0x7fff7fff) + 0x00010001) ^ (mv&0x80008000);\ + }\ + if (!n || mv != AV_RN32A(&near_mv[idx]))\ + AV_WN32A(&near_mv[++idx], mv);\ + cnt[idx] += 1 + (n != 2);\ + } else\ + cnt[CNT_ZERO] += 1 + (n != 2);\ + }\ + } + + MV_EDGE_CHECK(0) + MV_EDGE_CHECK(1) + MV_EDGE_CHECK(2) + + mb->partitioning = VP8_SPLITMVMODE_NONE; + if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[CNT_ZERO]][0])) { + mb->mode = VP8_MVMODE_MV; + + /* If we have three distinct MVs, merge first and last if they're the same */ + if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1 + VP8_EDGE_TOP]) == AV_RN32A(&near_mv[1 + VP8_EDGE_TOPLEFT])) + cnt[CNT_NEAREST] += 1; + + /* Swap near and nearest if necessary */ + if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) { + FFSWAP(uint8_t, cnt[CNT_NEAREST], cnt[CNT_NEAR]); + FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]); + } + + if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[CNT_NEAREST]][1])) { + if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[CNT_NEAR]][2])) { + + /* Choose the best mv out of 0,0 and the nearest mv */ + clamp_mv(s, &mb->mv, &near_mv[CNT_ZERO + (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])]); + cnt[CNT_SPLITMV] = ((mb_edge[VP8_EDGE_LEFT]->mode == VP8_MVMODE_SPLIT) + + (mb_edge[VP8_EDGE_TOP]->mode == VP8_MVMODE_SPLIT)) * 2 + + (mb_edge[VP8_EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT); + + if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[CNT_SPLITMV]][3])) { + mb->mode = VP8_MVMODE_SPLIT; + mb->mv = mb->bmv[decode_splitmvs(s, c, mb) - 1]; + } else { + mb->mv.y += read_mv_component(c, s->prob->mvc[0]); + mb->mv.x += read_mv_component(c, s->prob->mvc[1]); + mb->bmv[0] = mb->mv; + } + } else { + clamp_mv(s, &mb->mv, &near_mv[CNT_NEAR]); + mb->bmv[0] = mb->mv; + } + } else { + clamp_mv(s, &mb->mv, &near_mv[CNT_NEAREST]); + mb->bmv[0] = mb->mv; + } + } else { + mb->mode = VP8_MVMODE_ZERO; + AV_ZERO32(&mb->mv); + mb->bmv[0] = mb->mv; + } +} + static av_always_inline void decode_intra4x4_modes(VP8Context *s, VP56RangeCoder *c, int mb_x, int keyframe) @@ -730,12 +635,14 @@ void decode_intra4x4_modes(VP8Context *s, VP56RangeCoder *c, } static av_always_inline -void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, uint8_t *segment) +void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, uint8_t *segment, uint8_t *ref) { VP56RangeCoder *c = &s->c; if (s->segmentation.update_map) *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid); + else + *segment = ref ? *ref : *segment; s->segment = *segment; mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0; @@ -754,9 +661,6 @@ void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, uint8_ s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra); mb->ref_frame = VP56_FRAME_CURRENT; } else if (vp56_rac_get_prob_branchy(c, s->prob->intra)) { - VP56mv near[2], best; - uint8_t cnt[4] = { 0 }; - // inter MB, 16.2 if (vp56_rac_get_prob_branchy(c, s->prob->last)) mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ? @@ -766,36 +670,7 @@ void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, uint8_ s->ref_count[mb->ref_frame-1]++; // motion vectors, 16.3 - find_near_mvs(s, mb, mb_x, mb_y, near, &best, cnt); - if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[0]][0])) { - if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[1]][1])) { - if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[2]][2])) { - if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[3]][3])) { - mb->mode = VP8_MVMODE_SPLIT; - clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y); - mb->mv = mb->bmv[decode_splitmvs(s, c, mb) - 1]; - } else { - mb->mode = VP8_MVMODE_NEW; - clamp_mv(s, &mb->mv, &mb->mv, mb_x, mb_y); - mb->mv.y += + read_mv_component(c, s->prob->mvc[0]); - mb->mv.x += + read_mv_component(c, s->prob->mvc[1]); - } - } else { - mb->mode = VP8_MVMODE_NEAR; - clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y); - } - } else { - mb->mode = VP8_MVMODE_NEAREST; - clamp_mv(s, &mb->mv, &near[0], mb_x, mb_y); - } - } else { - mb->mode = VP8_MVMODE_ZERO; - AV_ZERO32(&mb->mv); - } - if (mb->mode != VP8_MVMODE_SPLIT) { - mb->partitioning = VP8_SPLITMVMODE_NONE; - mb->bmv[0] = mb->mv; - } + decode_mvs(s, mb, mb_x, mb_y); } else { // intra MB, 16.1 mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16); @@ -810,33 +685,30 @@ void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, uint8_ } } +#ifndef decode_block_coeffs_internal /** * @param c arithmetic bitstream reader context * @param block destination for block coefficients * @param probs probabilities to use when reading trees from the bitstream * @param i initial coeff index, 0 unless a separate DC block is coded - * @param zero_nhood the initial prediction context for number of surrounding - * all-zero blocks (only left/top, so 0-2) * @param qmul array holding the dc/ac dequant factor at position 0/1 * @return 0 if no coeffs were decoded * otherwise, the index of the last coeff decoded plus one */ -static int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16], - uint8_t probs[8][3][NUM_DCT_TOKENS-1], - int i, int zero_nhood, int16_t qmul[2]) +static int decode_block_coeffs_internal(VP56RangeCoder *c, DCTELEM block[16], + uint8_t probs[16][3][NUM_DCT_TOKENS-1], + int i, uint8_t *token_prob, int16_t qmul[2]) { - uint8_t *token_prob = probs[i][zero_nhood]; - int nonzero = 0; - int coeff; - + goto skip_eob; do { + int coeff; if (!vp56_rac_get_prob_branchy(c, token_prob[0])) // DCT_EOB - return nonzero; + return i; skip_eob: if (!vp56_rac_get_prob_branchy(c, token_prob[1])) { // DCT_0 if (++i == 16) - return nonzero; // invalid input; blocks should end with EOB + return i; // invalid input; blocks should end with EOB token_prob = probs[i][0]; goto skip_eob; } @@ -865,18 +737,38 @@ skip_eob: int b = vp56_rac_get_prob(c, token_prob[9+a]); int cat = (a<<1) + b; coeff = 3 + (8<avctx; + int x, y, mode, nnz; + uint32_t tr; // for the first row, we need to run xchg_mb_border to init the top edge to 127 // otherwise, skip it if we aren't going to deblock - if (s->deblock_filter || !mb_y) + if (!(avctx->flags & CODEC_FLAG_EMU_EDGE && !mb_y) && (s->deblock_filter || !mb_y)) xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width, s->filter.simple, 1); if (mb->mode < MODE_I4x4) { - mode = check_intra_pred_mode(mb->mode, mb_x, mb_y); + if (avctx->flags & CODEC_FLAG_EMU_EDGE) { // tested + mode = check_intra_pred8x8_mode_emuedge(mb->mode, mb_x, mb_y); + } else { + mode = check_intra_pred8x8_mode(mb->mode, mb_x, mb_y); + } s->hpc.pred16x16[mode](dst[0], s->linesize); } else { uint8_t *ptr = dst[0]; uint8_t *intra4x4 = s->intra4x4_pred_mode_mb; + uint8_t tr_top[4] = { 127, 127, 127, 127 }; // all blocks on the right edge of the macroblock use bottom edge // the top macroblock for their topright edge @@ -1021,8 +997,9 @@ void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, // if we're on the right edge of the frame, said edge is extended // from the top macroblock - if (mb_x == s->mb_width-1) { - tr = tr_right[-1]*0x01010101; + if (!(!mb_y && avctx->flags & CODEC_FLAG_EMU_EDGE) && + mb_x == s->mb_width-1) { + tr = tr_right[-1]*0x01010101u; tr_right = (uint8_t *)&tr; } @@ -1032,10 +1009,53 @@ void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, for (y = 0; y < 4; y++) { uint8_t *topright = ptr + 4 - s->linesize; for (x = 0; x < 4; x++) { - if (x == 3) + int copy = 0, linesize = s->linesize; + uint8_t *dst = ptr+4*x; + DECLARE_ALIGNED(4, uint8_t, copy_dst)[5*8]; + + if ((y == 0 || x == 3) && mb_y == 0 && avctx->flags & CODEC_FLAG_EMU_EDGE) { + topright = tr_top; + } else if (x == 3) topright = tr_right; - s->hpc.pred4x4[intra4x4[x]](ptr+4*x, topright, s->linesize); + if (avctx->flags & CODEC_FLAG_EMU_EDGE) { // mb_x+x or mb_y+y is a hack but works + mode = check_intra_pred4x4_mode_emuedge(intra4x4[x], mb_x + x, mb_y + y, ©); + if (copy) { + dst = copy_dst + 12; + linesize = 8; + if (!(mb_y + y)) { + copy_dst[3] = 127U; + AV_WN32A(copy_dst+4, 127U * 0x01010101U); + } else { + AV_COPY32(copy_dst+4, ptr+4*x-s->linesize); + if (!(mb_x + x)) { + copy_dst[3] = 129U; + } else { + copy_dst[3] = ptr[4*x-s->linesize-1]; + } + } + if (!(mb_x + x)) { + copy_dst[11] = + copy_dst[19] = + copy_dst[27] = + copy_dst[35] = 129U; + } else { + copy_dst[11] = ptr[4*x -1]; + copy_dst[19] = ptr[4*x+s->linesize -1]; + copy_dst[27] = ptr[4*x+s->linesize*2-1]; + copy_dst[35] = ptr[4*x+s->linesize*3-1]; + } + } + } else { + mode = intra4x4[x]; + } + s->hpc.pred4x4[mode](dst, topright, linesize); + if (copy) { + AV_COPY32(ptr+4*x , copy_dst+12); + AV_COPY32(ptr+4*x+s->linesize , copy_dst+20); + AV_COPY32(ptr+4*x+s->linesize*2, copy_dst+28); + AV_COPY32(ptr+4*x+s->linesize*3, copy_dst+36); + } nnz = s->non_zero_count_cache[y][x]; if (nnz) { @@ -1052,23 +1072,33 @@ void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, } } - mode = check_intra_pred_mode(s->chroma_pred_mode, mb_x, mb_y); + if (avctx->flags & CODEC_FLAG_EMU_EDGE) { + mode = check_intra_pred8x8_mode_emuedge(s->chroma_pred_mode, mb_x, mb_y); + } else { + mode = check_intra_pred8x8_mode(s->chroma_pred_mode, mb_x, mb_y); + } s->hpc.pred8x8[mode](dst[1], s->uvlinesize); s->hpc.pred8x8[mode](dst[2], s->uvlinesize); - if (s->deblock_filter || !mb_y) + if (!(avctx->flags & CODEC_FLAG_EMU_EDGE && !mb_y) && (s->deblock_filter || !mb_y)) xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width, s->filter.simple, 0); } +static const uint8_t subpel_idx[3][8] = { + { 0, 1, 2, 1, 2, 1, 2, 1 }, // nr. of left extra pixels, + // also function pointer index + { 0, 3, 5, 3, 5, 3, 5, 3 }, // nr. of extra pixels required + { 0, 2, 3, 2, 3, 2, 3, 2 }, // nr. of right extra pixels +}; + /** - * Generic MC function. + * luma MC function * * @param s VP8 decoding context - * @param luma 1 for luma (Y) planes, 0 for chroma (Cb/Cr) planes * @param dst target buffer for block data at block position - * @param src reference picture buffer at origin (0, 0) + * @param ref reference picture buffer at origin (0, 0) * @param mv motion vector (relative to block position) to get pixel data from * @param x_off horizontal position of block from origin (0, 0) * @param y_off vertical position of block from origin (0, 0) @@ -1080,32 +1110,96 @@ void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, * @param mc_func motion compensation function pointers (bilinear or sixtap MC) */ static av_always_inline -void vp8_mc(VP8Context *s, int luma, - uint8_t *dst, uint8_t *src, const VP56mv *mv, - int x_off, int y_off, int block_w, int block_h, - int width, int height, int linesize, - vp8_mc_func mc_func[3][3]) +void vp8_mc_luma(VP8Context *s, uint8_t *dst, AVFrame *ref, const VP56mv *mv, + int x_off, int y_off, int block_w, int block_h, + int width, int height, int linesize, + vp8_mc_func mc_func[3][3]) { + uint8_t *src = ref->data[0]; + if (AV_RN32A(mv)) { - static const uint8_t idx[8] = { 0, 1, 2, 1, 2, 1, 2, 1 }; - int mx = (mv->x << luma)&7, mx_idx = idx[mx]; - int my = (mv->y << luma)&7, my_idx = idx[my]; - x_off += mv->x >> (3 - luma); - y_off += mv->y >> (3 - luma); + int mx = (mv->x << 1)&7, mx_idx = subpel_idx[0][mx]; + int my = (mv->y << 1)&7, my_idx = subpel_idx[0][my]; + + x_off += mv->x >> 2; + y_off += mv->y >> 2; // edge emulation + ff_thread_await_progress(ref, (3 + y_off + block_h + subpel_idx[2][my]) >> 4, 0); src += y_off * linesize + x_off; - if (x_off < 2 || x_off >= width - block_w - 3 || - y_off < 2 || y_off >= height - block_h - 3) { - ff_emulated_edge_mc(s->edge_emu_buffer, src - 2 * linesize - 2, linesize, - block_w + 5, block_h + 5, - x_off - 2, y_off - 2, width, height); - src = s->edge_emu_buffer + 2 + linesize * 2; + if (x_off < mx_idx || x_off >= width - block_w - subpel_idx[2][mx] || + y_off < my_idx || y_off >= height - block_h - subpel_idx[2][my]) { + s->dsp.emulated_edge_mc(s->edge_emu_buffer, src - my_idx * linesize - mx_idx, linesize, + block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my], + x_off - mx_idx, y_off - my_idx, width, height); + src = s->edge_emu_buffer + mx_idx + linesize * my_idx; } mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my); - } else + } else { + ff_thread_await_progress(ref, (3 + y_off + block_h) >> 4, 0); mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0); + } +} + +/** + * chroma MC function + * + * @param s VP8 decoding context + * @param dst1 target buffer for block data at block position (U plane) + * @param dst2 target buffer for block data at block position (V plane) + * @param ref reference picture buffer at origin (0, 0) + * @param mv motion vector (relative to block position) to get pixel data from + * @param x_off horizontal position of block from origin (0, 0) + * @param y_off vertical position of block from origin (0, 0) + * @param block_w width of block (16, 8 or 4) + * @param block_h height of block (always same as block_w) + * @param width width of src/dst plane data + * @param height height of src/dst plane data + * @param linesize size of a single line of plane data, including padding + * @param mc_func motion compensation function pointers (bilinear or sixtap MC) + */ +static av_always_inline +void vp8_mc_chroma(VP8Context *s, uint8_t *dst1, uint8_t *dst2, AVFrame *ref, + const VP56mv *mv, int x_off, int y_off, + int block_w, int block_h, int width, int height, int linesize, + vp8_mc_func mc_func[3][3]) +{ + uint8_t *src1 = ref->data[1], *src2 = ref->data[2]; + + if (AV_RN32A(mv)) { + int mx = mv->x&7, mx_idx = subpel_idx[0][mx]; + int my = mv->y&7, my_idx = subpel_idx[0][my]; + + x_off += mv->x >> 3; + y_off += mv->y >> 3; + + // edge emulation + src1 += y_off * linesize + x_off; + src2 += y_off * linesize + x_off; + ff_thread_await_progress(ref, (3 + y_off + block_h + subpel_idx[2][my]) >> 3, 0); + if (x_off < mx_idx || x_off >= width - block_w - subpel_idx[2][mx] || + y_off < my_idx || y_off >= height - block_h - subpel_idx[2][my]) { + s->dsp.emulated_edge_mc(s->edge_emu_buffer, src1 - my_idx * linesize - mx_idx, linesize, + block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my], + x_off - mx_idx, y_off - my_idx, width, height); + src1 = s->edge_emu_buffer + mx_idx + linesize * my_idx; + mc_func[my_idx][mx_idx](dst1, linesize, src1, linesize, block_h, mx, my); + + s->dsp.emulated_edge_mc(s->edge_emu_buffer, src2 - my_idx * linesize - mx_idx, linesize, + block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my], + x_off - mx_idx, y_off - my_idx, width, height); + src2 = s->edge_emu_buffer + mx_idx + linesize * my_idx; + mc_func[my_idx][mx_idx](dst2, linesize, src2, linesize, block_h, mx, my); + } else { + mc_func[my_idx][mx_idx](dst1, linesize, src1, linesize, block_h, mx, my); + mc_func[my_idx][mx_idx](dst2, linesize, src2, linesize, block_h, mx, my); + } + } else { + ff_thread_await_progress(ref, (3 + y_off + block_h) >> 3, 0); + mc_func[0][0](dst1, linesize, src1 + y_off * linesize + x_off, linesize, block_h, 0, 0); + mc_func[0][0](dst2, linesize, src2 + y_off * linesize + x_off, linesize, block_h, 0, 0); + } } static av_always_inline @@ -1118,10 +1212,10 @@ void vp8_mc_part(VP8Context *s, uint8_t *dst[3], VP56mv uvmv = *mv; /* Y */ - vp8_mc(s, 1, dst[0] + by_off * s->linesize + bx_off, - ref_frame->data[0], mv, x_off + bx_off, y_off + by_off, - block_w, block_h, width, height, s->linesize, - s->put_pixels_tab[block_w == 8]); + vp8_mc_luma(s, dst[0] + by_off * s->linesize + bx_off, + ref_frame, mv, x_off + bx_off, y_off + by_off, + block_w, block_h, width, height, s->linesize, + s->put_pixels_tab[block_w == 8]); /* U/V */ if (s->profile == 3) { @@ -1132,14 +1226,11 @@ void vp8_mc_part(VP8Context *s, uint8_t *dst[3], bx_off >>= 1; by_off >>= 1; width >>= 1; height >>= 1; block_w >>= 1; block_h >>= 1; - vp8_mc(s, 0, dst[1] + by_off * s->uvlinesize + bx_off, - ref_frame->data[1], &uvmv, x_off + bx_off, y_off + by_off, - block_w, block_h, width, height, s->uvlinesize, - s->put_pixels_tab[1 + (block_w == 4)]); - vp8_mc(s, 0, dst[2] + by_off * s->uvlinesize + bx_off, - ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off, - block_w, block_h, width, height, s->uvlinesize, - s->put_pixels_tab[1 + (block_w == 4)]); + vp8_mc_chroma(s, dst[1] + by_off * s->uvlinesize + bx_off, + dst[2] + by_off * s->uvlinesize + bx_off, ref_frame, + &uvmv, x_off + bx_off, y_off + by_off, + block_w, block_h, width, height, s->uvlinesize, + s->put_pixels_tab[1 + (block_w == 4)]); } /* Fetch pixels for estimated mv 4 macroblocks ahead. @@ -1149,10 +1240,13 @@ static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, i /* Don't prefetch refs that haven't been used very often this frame. */ if (s->ref_count[ref-1] > (mb_xy >> 5)) { int x_off = mb_x << 4, y_off = mb_y << 4; - int mx = mb->mv.x + x_off + 8; - int my = mb->mv.y + y_off; + int mx = (mb->mv.x>>2) + x_off + 8; + int my = (mb->mv.y>>2) + y_off; uint8_t **src= s->framep[ref]->data; int off= mx + (my + (mb_x&3)*4)*s->linesize + 64; + /* For threading, a ff_thread_await_progress here might be useful, but + * it actually slows down the decoder. Since a bad prefetch doesn't + * generate bad decoder output, we don't run it here. */ s->dsp.prefetch(src[0]+off, s->linesize, 4); off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64; s->dsp.prefetch(src[1]+off, src[2]-src[1], 2); @@ -1171,10 +1265,11 @@ void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, AVFrame *ref = s->framep[mb->ref_frame]; VP56mv *bmv = mb->bmv; - if (mb->mode < VP8_MVMODE_SPLIT) { + switch (mb->partitioning) { + case VP8_SPLITMVMODE_NONE: vp8_mc_part(s, dst, ref, x_off, y_off, 0, 0, 16, 16, width, height, &mb->mv); - } else switch (mb->partitioning) { + break; case VP8_SPLITMVMODE_4x4: { int x, y; VP56mv uvmv; @@ -1182,11 +1277,11 @@ void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, /* Y */ for (y = 0; y < 4; y++) { for (x = 0; x < 4; x++) { - vp8_mc(s, 1, dst[0] + 4*y*s->linesize + x*4, - ref->data[0], &bmv[4*y + x], - 4*x + x_off, 4*y + y_off, 4, 4, - width, height, s->linesize, - s->put_pixels_tab[2]); + vp8_mc_luma(s, dst[0] + 4*y*s->linesize + x*4, + ref, &bmv[4*y + x], + 4*x + x_off, 4*y + y_off, 4, 4, + width, height, s->linesize, + s->put_pixels_tab[2]); } } @@ -1208,16 +1303,11 @@ void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb, uvmv.x &= ~7; uvmv.y &= ~7; } - vp8_mc(s, 0, dst[1] + 4*y*s->uvlinesize + x*4, - ref->data[1], &uvmv, - 4*x + x_off, 4*y + y_off, 4, 4, - width, height, s->uvlinesize, - s->put_pixels_tab[2]); - vp8_mc(s, 0, dst[2] + 4*y*s->uvlinesize + x*4, - ref->data[2], &uvmv, - 4*x + x_off, 4*y + y_off, 4, 4, - width, height, s->uvlinesize, - s->put_pixels_tab[2]); + vp8_mc_chroma(s, dst[1] + 4*y*s->uvlinesize + x*4, + dst[2] + 4*y*s->uvlinesize + x*4, ref, &uvmv, + 4*x + x_off, 4*y + y_off, 4, 4, + width, height, s->uvlinesize, + s->put_pixels_tab[2]); } } break; @@ -1254,17 +1344,17 @@ static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblo if (mb->mode != MODE_I4x4) { uint8_t *y_dst = dst[0]; for (y = 0; y < 4; y++) { - uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[y]); + uint32_t nnz4 = AV_RL32(s->non_zero_count_cache[y]); if (nnz4) { if (nnz4&~0x01010101) { for (x = 0; x < 4; x++) { - int nnz = s->non_zero_count_cache[y][x]; - if (nnz) { - if (nnz == 1) - s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize); - else - s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize); - } + if ((uint8_t)nnz4 == 1) + s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize); + else if((uint8_t)nnz4 > 1) + s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize); + nnz4 >>= 8; + if (!nnz4) + break; } } else { s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize); @@ -1275,19 +1365,19 @@ static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblo } for (ch = 0; ch < 2; ch++) { - uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[4+ch]); + uint32_t nnz4 = AV_RL32(s->non_zero_count_cache[4+ch]); if (nnz4) { uint8_t *ch_dst = dst[1+ch]; if (nnz4&~0x01010101) { for (y = 0; y < 2; y++) { for (x = 0; x < 2; x++) { - int nnz = s->non_zero_count_cache[4+ch][(y<<1)+x]; - if (nnz) { - if (nnz == 1) - s->vp8dsp.vp8_idct_dc_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize); - else - s->vp8dsp.vp8_idct_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize); - } + if ((uint8_t)nnz4 == 1) + s->vp8dsp.vp8_idct_dc_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize); + else if((uint8_t)nnz4 > 1) + s->vp8dsp.vp8_idct_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize); + nnz4 >>= 8; + if (!nnz4) + goto chroma_idct_end; } ch_dst += 4*s->uvlinesize; } @@ -1295,6 +1385,7 @@ static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblo s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize); } } +chroma_idct_end: ; } } @@ -1311,24 +1402,14 @@ static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *m if (s->lf_delta.enabled) { filter_level += s->lf_delta.ref[mb->ref_frame]; - - if (mb->ref_frame == VP56_FRAME_CURRENT) { - if (mb->mode == MODE_I4x4) - filter_level += s->lf_delta.mode[0]; - } else { - if (mb->mode == VP8_MVMODE_ZERO) - filter_level += s->lf_delta.mode[1]; - else if (mb->mode == VP8_MVMODE_SPLIT) - filter_level += s->lf_delta.mode[3]; - else - filter_level += s->lf_delta.mode[2]; - } + filter_level += s->lf_delta.mode[mb->mode]; } - filter_level = av_clip(filter_level, 0, 63); + + filter_level = av_clip_uintp2(filter_level, 6); interior_limit = filter_level; if (s->filter.sharpness) { - interior_limit >>= s->filter.sharpness > 4 ? 2 : 1; + interior_limit >>= (s->filter.sharpness + 3) >> 2; interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness); } interior_limit = FFMAX(interior_limit, 1); @@ -1346,23 +1427,24 @@ static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8Filter int inner_filter = f->inner_filter; int linesize = s->linesize; int uvlinesize = s->uvlinesize; + static const uint8_t hev_thresh_lut[2][64] = { + { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, + 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, + 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, + 3, 3, 3, 3 }, + { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, + 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, + 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, + 2, 2, 2, 2 } + }; if (!filter_level) return; - mbedge_lim = 2*(filter_level+2) + inner_limit; - bedge_lim = 2* filter_level + inner_limit; - hev_thresh = filter_level >= 15; + bedge_lim = 2*filter_level + inner_limit; + mbedge_lim = bedge_lim + 4; - if (s->keyframe) { - if (filter_level >= 40) - hev_thresh = 2; - } else { - if (filter_level >= 40) - hev_thresh = 3; - else if (filter_level >= 20) - hev_thresh = 2; - } + hev_thresh = hev_thresh_lut[s->keyframe][filter_level]; if (mb_x) { s->vp8dsp.vp8_h_loop_filter16y(dst[0], linesize, @@ -1418,8 +1500,8 @@ static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8Fi if (!filter_level) return; - mbedge_lim = 2*(filter_level+2) + inner_limit; - bedge_lim = 2* filter_level + inner_limit; + bedge_lim = 2*filter_level + inner_limit; + mbedge_lim = bedge_lim + 4; if (mb_x) s->vp8dsp.vp8_h_loop_filter_simple(dst, linesize, mbedge_lim); @@ -1438,13 +1520,13 @@ static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8Fi } } -static void filter_mb_row(VP8Context *s, int mb_y) +static void filter_mb_row(VP8Context *s, AVFrame *curframe, int mb_y) { VP8FilterStrength *f = s->filter_strength; uint8_t *dst[3] = { - s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize, - s->framep[VP56_FRAME_CURRENT]->data[1] + 8*mb_y*s->uvlinesize, - s->framep[VP56_FRAME_CURRENT]->data[2] + 8*mb_y*s->uvlinesize + curframe->data[0] + 16*mb_y*s->linesize, + curframe->data[1] + 8*mb_y*s->uvlinesize, + curframe->data[2] + 8*mb_y*s->uvlinesize }; int mb_x; @@ -1457,10 +1539,10 @@ static void filter_mb_row(VP8Context *s, int mb_y) } } -static void filter_mb_row_simple(VP8Context *s, int mb_y) +static void filter_mb_row_simple(VP8Context *s, AVFrame *curframe, int mb_y) { VP8FilterStrength *f = s->filter_strength; - uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize; + uint8_t *dst = curframe->data[0] + 16*mb_y*s->linesize; int mb_x; for (mb_x = 0; mb_x < s->mb_width; mb_x++) { @@ -1470,17 +1552,29 @@ static void filter_mb_row_simple(VP8Context *s, int mb_y) } } +static void release_queued_segmaps(VP8Context *s, int is_close) +{ + int leave_behind = is_close ? 0 : !s->maps_are_invalid; + while (s->num_maps_to_be_freed > leave_behind) + av_freep(&s->segmentation_maps[--s->num_maps_to_be_freed]); + s->maps_are_invalid = 0; +} + static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size, AVPacket *avpkt) { VP8Context *s = avctx->priv_data; int ret, mb_x, mb_y, i, y, referenced; enum AVDiscard skip_thresh; - AVFrame *av_uninit(curframe); + AVFrame *av_uninit(curframe), *prev_frame; + + release_queued_segmaps(s, 0); if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0) return ret; + prev_frame = s->framep[VP56_FRAME_CURRENT]; + referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT || s->update_altref == VP56_FRAME_CURRENT; @@ -1493,24 +1587,59 @@ static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size, } s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh; - for (i = 0; i < 4; i++) - if (&s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] && + // release no longer referenced frames + for (i = 0; i < 5; i++) + if (s->frames[i].data[0] && + &s->frames[i] != prev_frame && + &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] && + &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] && + &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) + vp8_release_frame(s, &s->frames[i], 1, 0); + + // find a free buffer + for (i = 0; i < 5; i++) + if (&s->frames[i] != prev_frame && + &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] && &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] && &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) { curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i]; break; } + if (i == 5) { + av_log(avctx, AV_LOG_FATAL, "Ran out of free frames!\n"); + abort(); + } if (curframe->data[0]) - avctx->release_buffer(avctx, curframe); + vp8_release_frame(s, curframe, 1, 0); curframe->key_frame = s->keyframe; - curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE; + curframe->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P; curframe->reference = referenced ? 3 : 0; - if ((ret = avctx->get_buffer(avctx, curframe))) { + if ((ret = vp8_alloc_frame(s, curframe))) { av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n"); return ret; } + // check if golden and altref are swapped + if (s->update_altref != VP56_FRAME_NONE) { + s->next_framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref]; + } else { + s->next_framep[VP56_FRAME_GOLDEN2] = s->framep[VP56_FRAME_GOLDEN2]; + } + if (s->update_golden != VP56_FRAME_NONE) { + s->next_framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden]; + } else { + s->next_framep[VP56_FRAME_GOLDEN] = s->framep[VP56_FRAME_GOLDEN]; + } + if (s->update_last) { + s->next_framep[VP56_FRAME_PREVIOUS] = curframe; + } else { + s->next_framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_PREVIOUS]; + } + s->next_framep[VP56_FRAME_CURRENT] = curframe; + + ff_thread_finish_setup(avctx); + // Given that arithmetic probabilities are updated every frame, it's quite likely // that the values we have on a random interframe are complete junk if we didn't // start decode on a keyframe. So just don't display anything rather than junk. @@ -1529,45 +1658,57 @@ static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size, memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz)); - /* Zero macroblock structures for top/left prediction from outside the frame. */ - memset(s->macroblocks, 0, (s->mb_width + s->mb_height*2)*sizeof(*s->macroblocks)); + /* Zero macroblock structures for top/top-left prediction from outside the frame. */ + memset(s->macroblocks + s->mb_height*2 - 1, 0, (s->mb_width+1)*sizeof(*s->macroblocks)); // top edge of 127 for intra prediction - memset(s->top_border, 127, (s->mb_width+1)*sizeof(*s->top_border)); + if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) { + s->top_border[0][15] = s->top_border[0][23] = 127; + memset(s->top_border[1]-1, 127, s->mb_width*sizeof(*s->top_border)+1); + } memset(s->ref_count, 0, sizeof(s->ref_count)); if (s->keyframe) - memset(s->intra4x4_pred_mode_top, DC_PRED, s->b4_stride*4); + memset(s->intra4x4_pred_mode_top, DC_PRED, s->mb_width*4); + +#define MARGIN (16 << 2) + s->mv_min.y = -MARGIN; + s->mv_max.y = ((s->mb_height - 1) << 6) + MARGIN; for (mb_y = 0; mb_y < s->mb_height; mb_y++) { VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)]; VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2; - uint8_t *segment_map = s->segmentation_map + mb_y*s->mb_stride; - int mb_xy = mb_y * s->mb_stride; + int mb_xy = mb_y*s->mb_width; uint8_t *dst[3] = { curframe->data[0] + 16*mb_y*s->linesize, curframe->data[1] + 8*mb_y*s->uvlinesize, curframe->data[2] + 8*mb_y*s->uvlinesize }; + memset(mb - 1, 0, sizeof(*mb)); // zero left macroblock memset(s->left_nnz, 0, sizeof(s->left_nnz)); AV_WN32A(s->intra4x4_pred_mode_left, DC_PRED*0x01010101); // left edge of 129 for intra prediction - if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) + if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) { for (i = 0; i < 3; i++) for (y = 0; y < 16>>!!i; y++) dst[i][y*curframe->linesize[i]-1] = 129; - if (mb_y) - memset(s->top_border, 129, sizeof(*s->top_border)); + if (mb_y == 1) // top left edge is also 129 + s->top_border[0][15] = s->top_border[0][23] = s->top_border[0][31] = 129; + } - for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) { - uint8_t *segment_mb = segment_map+mb_x; + s->mv_min.x = -MARGIN; + s->mv_max.x = ((s->mb_width - 1) << 6) + MARGIN; + if (prev_frame && s->segmentation.enabled && !s->segmentation.update_map) + ff_thread_await_progress(prev_frame, mb_y, 0); + for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) { /* Prefetch the current frame, 4 MBs ahead */ s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4); s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2); - decode_mb_mode(s, mb, mb_x, mb_y, segment_mb); + decode_mb_mode(s, mb, mb_x, mb_y, curframe->ref_index[0] + mb_xy, + prev_frame && prev_frame->ref_index[0] ? prev_frame->ref_index[0] + mb_xy : NULL); prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS); @@ -1602,47 +1743,32 @@ static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size, dst[0] += 16; dst[1] += 8; dst[2] += 8; + s->mv_min.x -= 64; + s->mv_max.x -= 64; } if (s->deblock_filter) { if (s->filter.simple) - filter_mb_row_simple(s, mb_y); + filter_mb_row_simple(s, curframe, mb_y); else - filter_mb_row(s, mb_y); + filter_mb_row(s, curframe, mb_y); } + s->mv_min.y -= 64; + s->mv_max.y -= 64; + + ff_thread_report_progress(curframe, mb_y, 0); } + ff_thread_report_progress(curframe, INT_MAX, 0); skip_decode: // if future frames don't use the updated probabilities, // reset them to the values we saved if (!s->update_probabilities) s->prob[0] = s->prob[1]; - // check if golden and altref are swapped - if (s->update_altref == VP56_FRAME_GOLDEN && - s->update_golden == VP56_FRAME_GOLDEN2) - FFSWAP(AVFrame *, s->framep[VP56_FRAME_GOLDEN], s->framep[VP56_FRAME_GOLDEN2]); - else { - if (s->update_altref != VP56_FRAME_NONE) - s->framep[VP56_FRAME_GOLDEN2] = s->framep[s->update_altref]; - - if (s->update_golden != VP56_FRAME_NONE) - s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden]; - } - - if (s->update_last) // move cur->prev - s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT]; - - // release no longer referenced frames - for (i = 0; i < 4; i++) - if (s->frames[i].data[0] && - &s->frames[i] != s->framep[VP56_FRAME_CURRENT] && - &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] && - &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] && - &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) - avctx->release_buffer(avctx, &s->frames[i]); + memcpy(&s->framep[0], &s->next_framep[0], sizeof(s->framep[0]) * 4); if (!s->invisible) { - *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT]; + *(AVFrame*)data = *curframe; *data_size = sizeof(AVFrame); } @@ -1657,34 +1783,66 @@ static av_cold int vp8_decode_init(AVCodecContext *avctx) avctx->pix_fmt = PIX_FMT_YUV420P; dsputil_init(&s->dsp, avctx); - ff_h264_pred_init(&s->hpc, CODEC_ID_VP8); + ff_h264_pred_init(&s->hpc, CODEC_ID_VP8, 8, 1); ff_vp8dsp_init(&s->vp8dsp); - // intra pred needs edge emulation among other things - if (avctx->flags&CODEC_FLAG_EMU_EDGE) { - av_log(avctx, AV_LOG_ERROR, "Edge emulation not supported\n"); - return AVERROR_PATCHWELCOME; - } - return 0; } static av_cold int vp8_decode_free(AVCodecContext *avctx) { - vp8_decode_flush(avctx); + vp8_decode_flush_impl(avctx, 0, 1, 1); + release_queued_segmaps(avctx->priv_data, 1); + return 0; +} + +static av_cold int vp8_decode_init_thread_copy(AVCodecContext *avctx) +{ + VP8Context *s = avctx->priv_data; + + s->avctx = avctx; + + return 0; +} + +#define REBASE(pic) \ + pic ? pic - &s_src->frames[0] + &s->frames[0] : NULL + +static int vp8_decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src) +{ + VP8Context *s = dst->priv_data, *s_src = src->priv_data; + + if (s->macroblocks_base && + (s_src->mb_width != s->mb_width || s_src->mb_height != s->mb_height)) { + free_buffers(s); + s->maps_are_invalid = 1; + } + + s->prob[0] = s_src->prob[!s_src->update_probabilities]; + s->segmentation = s_src->segmentation; + s->lf_delta = s_src->lf_delta; + memcpy(s->sign_bias, s_src->sign_bias, sizeof(s->sign_bias)); + + memcpy(&s->frames, &s_src->frames, sizeof(s->frames)); + s->framep[0] = REBASE(s_src->next_framep[0]); + s->framep[1] = REBASE(s_src->next_framep[1]); + s->framep[2] = REBASE(s_src->next_framep[2]); + s->framep[3] = REBASE(s_src->next_framep[3]); + return 0; } -AVCodec vp8_decoder = { - "vp8", - AVMEDIA_TYPE_VIDEO, - CODEC_ID_VP8, - sizeof(VP8Context), - vp8_decode_init, - NULL, - vp8_decode_free, - vp8_decode_frame, - CODEC_CAP_DR1, +AVCodec ff_vp8_decoder = { + .name = "vp8", + .type = AVMEDIA_TYPE_VIDEO, + .id = CODEC_ID_VP8, + .priv_data_size = sizeof(VP8Context), + .init = vp8_decode_init, + .close = vp8_decode_free, + .decode = vp8_decode_frame, + .capabilities = CODEC_CAP_DR1 | CODEC_CAP_FRAME_THREADS, .flush = vp8_decode_flush, .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"), + .init_thread_copy = ONLY_IF_THREADS_ENABLED(vp8_decode_init_thread_copy), + .update_thread_context = ONLY_IF_THREADS_ENABLED(vp8_decode_update_thread_context), };