hevc_deblock: Fix compilation with nasm
[ffmpeg.git] / libavcodec / vp8.c
index dce090f..4ea6d56 100644 (file)
-/**
- * VP8 compatible video decoder
+/*
+ * VP7/VP8 compatible video decoder
  *
  * Copyright (C) 2010 David Conrad
  * Copyright (C) 2010 Ronald S. Bultje
- * Copyright (C) 2010 Jason Garrett-Glaser
+ * Copyright (C) 2010 Fiona Glaser
+ * Copyright (C) 2012 Daniel Kang
+ * Copyright (C) 2014 Peter Ross
  *
- * 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 "libavcore/imgutils.h"
+#include "libavutil/imgutils.h"
+
 #include "avcodec.h"
-#include "vp56.h"
-#include "vp8data.h"
-#include "vp8dsp.h"
-#include "h264pred.h"
+#include "internal.h"
 #include "rectangle.h"
+#include "thread.h"
+#include "vp8.h"
+#include "vp8data.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;
-
-    uint8_t *intra4x4_pred_mode_top;
-    uint8_t intra4x4_pred_mode_left[4];
-    uint8_t *segmentation_map;
-
-    /**
-     * 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:
-         * [0-3] - i16x16 (always zero)
-         * [4]   - i4x4
-         * [5]   - zero mv
-         * [6]   - inter modes except for zero or split mv
-         * [7]   - split mv
-         *  i16x16 modes never have any adjustment
-         */
-        int8_t mode[VP8_MVMODE_SPLIT+1];
-
-        /**
-         * 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));
-
+    if (s->thread_data)
+        for (i = 0; i < MAX_THREADS; i++) {
+#if HAVE_THREADS
+            pthread_cond_destroy(&s->thread_data[i].cond);
+            pthread_mutex_destroy(&s->thread_data[i].lock);
+#endif
+            av_freep(&s->thread_data[i].filter_strength);
+        }
+    av_freep(&s->thread_data);
     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 update_dimensions(VP8Context *s, int width, int height)
+static int vp8_alloc_frame(VP8Context *s, VP8Frame *f, int ref)
 {
-    if (av_image_check_size(width, height, 0, s->avctx))
-        return AVERROR_INVALIDDATA;
+    int ret;
+    if ((ret = ff_thread_get_buffer(s->avctx, &f->tf,
+                                    ref ? AV_GET_BUFFER_FLAG_REF : 0)) < 0)
+        return ret;
+    if (!(f->seg_map = av_buffer_allocz(s->mb_width * s->mb_height))) {
+        ff_thread_release_buffer(s->avctx, &f->tf);
+        return AVERROR(ENOMEM);
+    }
+    return 0;
+}
+
+static void vp8_release_frame(VP8Context *s, VP8Frame *f)
+{
+    av_buffer_unref(&f->seg_map);
+    ff_thread_release_buffer(s->avctx, &f->tf);
+}
+
+#if CONFIG_VP8_DECODER
+static int vp8_ref_frame(VP8Context *s, VP8Frame *dst, VP8Frame *src)
+{
+    int ret;
+
+    vp8_release_frame(s, dst);
 
-    vp8_decode_flush(s->avctx);
+    if ((ret = ff_thread_ref_frame(&dst->tf, &src->tf)) < 0)
+        return ret;
+    if (src->seg_map &&
+        !(dst->seg_map = av_buffer_ref(src->seg_map))) {
+        vp8_release_frame(s, dst);
+        return AVERROR(ENOMEM);
+    }
+
+    return 0;
+}
+#endif /* CONFIG_VP8_DECODER */
 
-    avcodec_set_dimensions(s->avctx, width, height);
+static void vp8_decode_flush_impl(AVCodecContext *avctx, int free_mem)
+{
+    VP8Context *s = avctx->priv_data;
+    int i;
 
-    s->mb_width  = (s->avctx->coded_width +15) / 16;
-    s->mb_height = (s->avctx->coded_height+15) / 16;
+    for (i = 0; i < FF_ARRAY_ELEMS(s->frames); i++)
+        vp8_release_frame(s, &s->frames[i]);
+    memset(s->framep, 0, sizeof(s->framep));
 
-    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_width*s->mb_height);
+    if (free_mem)
+        free_buffers(s);
+}
 
-    if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_top ||
-        !s->top_nnz || !s->top_border || !s->segmentation_map)
+static void vp8_decode_flush(AVCodecContext *avctx)
+{
+    vp8_decode_flush_impl(avctx, 0);
+}
+
+static VP8Frame *vp8_find_free_buffer(VP8Context *s)
+{
+    VP8Frame *frame = NULL;
+    int i;
+
+    // find a free buffer
+    for (i = 0; i < 5; i++)
+        if (&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]) {
+            frame = &s->frames[i];
+            break;
+        }
+    if (i == 5) {
+        av_log(s->avctx, AV_LOG_FATAL, "Ran out of free frames!\n");
+        abort();
+    }
+    if (frame->tf.f->data[0])
+        vp8_release_frame(s, frame);
+
+    return frame;
+}
+
+static av_always_inline
+int update_dimensions(VP8Context *s, int width, int height, int is_vp7)
+{
+    AVCodecContext *avctx = s->avctx;
+    int i, ret;
+
+    if (width  != s->avctx->width ||
+        height != s->avctx->height) {
+        vp8_decode_flush_impl(s->avctx, 1);
+
+        ret = ff_set_dimensions(s->avctx, width, height);
+        if (ret < 0)
+            return ret;
+    }
+
+    s->mb_width  = (s->avctx->coded_width  + 15) / 16;
+    s->mb_height = (s->avctx->coded_height + 15) / 16;
+
+    s->mb_layout = is_vp7 || avctx->active_thread_type == FF_THREAD_SLICE &&
+                   FFMIN(s->num_coeff_partitions, avctx->thread_count) > 1;
+    if (!s->mb_layout) { // Frame threading and one thread
+        s->macroblocks_base       = av_mallocz((s->mb_width + s->mb_height * 2 + 1) *
+                                               sizeof(*s->macroblocks));
+        s->intra4x4_pred_mode_top = av_mallocz(s->mb_width * 4);
+    } else // Sliced threading
+        s->macroblocks_base = av_mallocz((s->mb_width + 2) * (s->mb_height + 2) *
+                                         sizeof(*s->macroblocks));
+    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->thread_data = av_mallocz(MAX_THREADS * sizeof(VP8ThreadData));
+
+    if (!s->macroblocks_base || !s->top_nnz || !s->top_border ||
+        !s->thread_data || (!s->intra4x4_pred_mode_top && !s->mb_layout)) {
+        free_buffers(s);
         return AVERROR(ENOMEM);
+    }
 
-    s->macroblocks        = s->macroblocks_base + 1;
+    for (i = 0; i < MAX_THREADS; i++) {
+        s->thread_data[i].filter_strength =
+            av_mallocz(s->mb_width * sizeof(*s->thread_data[0].filter_strength));
+        if (!s->thread_data[i].filter_strength) {
+            free_buffers(s);
+            return AVERROR(ENOMEM);
+        }
+#if HAVE_THREADS
+        pthread_mutex_init(&s->thread_data[i].lock, NULL);
+        pthread_cond_init(&s->thread_data[i].cond, NULL);
+#endif
+    }
+
+    s->macroblocks = s->macroblocks_base + 1;
 
     return 0;
 }
 
+static int vp7_update_dimensions(VP8Context *s, int width, int height)
+{
+    return update_dimensions(s, width, height, IS_VP7);
+}
+
+static int vp8_update_dimensions(VP8Context *s, int width, int height)
+{
+    return update_dimensions(s, width, height, IS_VP8);
+}
+
 static void parse_segment_info(VP8Context *s)
 {
     VP56RangeCoder *c = &s->c;
@@ -276,11 +228,23 @@ static void update_lf_deltas(VP8Context *s)
     VP56RangeCoder *c = &s->c;
     int i;
 
-    for (i = 0; i < 4; i++)
-        s->lf_delta.ref[i]  = vp8_rac_get_sint(c, 6);
+    for (i = 0; i < 4; i++) {
+        if (vp8_rac_get(c)) {
+            s->lf_delta.ref[i] = vp8_rac_get_uint(c, 6);
 
-    for (i = MODE_I4x4; i <= VP8_MVMODE_SPLIT; i++)
-        s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);
+            if (vp8_rac_get(c))
+                s->lf_delta.ref[i] = -s->lf_delta.ref[i];
+        }
+    }
+
+    for (i = MODE_I4x4; i <= VP8_MVMODE_SPLIT; i++) {
+        if (vp8_rac_get(c)) {
+            s->lf_delta.mode[i] = vp8_rac_get_uint(c, 6);
+
+            if (vp8_rac_get(c))
+                s->lf_delta.mode[i] = -s->lf_delta.mode[i];
+        }
+    }
 }
 
 static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
@@ -290,13 +254,13 @@ static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
 
     s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);
 
-    buf      += 3*(s->num_coeff_partitions-1);
-    buf_size -= 3*(s->num_coeff_partitions-1);
+    buf      += 3 * (s->num_coeff_partitions - 1);
+    buf_size -= 3 * (s->num_coeff_partitions - 1);
     if (buf_size < 0)
         return -1;
 
-    for (i = 0; i < s->num_coeff_partitions-1; i++) {
-        int size = AV_RL24(sizes + 3*i);
+    for (i = 0; i < s->num_coeff_partitions - 1; i++) {
+        int size = AV_RL24(sizes + 3 * i);
         if (buf_size - size < 0)
             return -1;
 
@@ -309,6 +273,25 @@ static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)
     return 0;
 }
 
+static void vp7_get_quants(VP8Context *s)
+{
+    VP56RangeCoder *c = &s->c;
+
+    int yac_qi  = vp8_rac_get_uint(c, 7);
+    int ydc_qi  = vp8_rac_get(c) ? vp8_rac_get_uint(c, 7) : yac_qi;
+    int y2dc_qi = vp8_rac_get(c) ? vp8_rac_get_uint(c, 7) : yac_qi;
+    int y2ac_qi = vp8_rac_get(c) ? vp8_rac_get_uint(c, 7) : yac_qi;
+    int uvdc_qi = vp8_rac_get(c) ? vp8_rac_get_uint(c, 7) : yac_qi;
+    int uvac_qi = vp8_rac_get(c) ? vp8_rac_get_uint(c, 7) : yac_qi;
+
+    s->qmat[0].luma_qmul[0]    =       vp7_ydc_qlookup[ydc_qi];
+    s->qmat[0].luma_qmul[1]    =       vp7_yac_qlookup[yac_qi];
+    s->qmat[0].luma_dc_qmul[0] =       vp7_y2dc_qlookup[y2dc_qi];
+    s->qmat[0].luma_dc_qmul[1] =       vp7_y2ac_qlookup[y2ac_qi];
+    s->qmat[0].chroma_qmul[0]  = FFMIN(vp7_ydc_qlookup[uvdc_qi], 132);
+    s->qmat[0].chroma_qmul[1]  =       vp7_yac_qlookup[uvac_qi];
+}
+
 static void get_quants(VP8Context *s)
 {
     VP56RangeCoder *c = &s->c;
@@ -329,12 +312,13 @@ 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] = vp8_dc_qlookup[av_clip_uintp2(base_qi + y2dc_delta, 7)] * 2;
+        /* 101581>>16 is equivalent to 155/100 */
+        s->qmat[i].luma_dc_qmul[1] = vp8_ac_qlookup[av_clip_uintp2(base_qi + y2ac_delta, 7)] * 101581 >> 16;
+        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);
@@ -370,6 +354,54 @@ static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)
     return VP56_FRAME_NONE;
 }
 
+static void vp78_reset_probability_tables(VP8Context *s)
+{
+    int i, j;
+    for (i = 0; i < 4; i++)
+        for (j = 0; j < 16; j++)
+            memcpy(s->prob->token[i][j], vp8_token_default_probs[i][vp8_coeff_band[j]],
+                   sizeof(s->prob->token[i][j]));
+}
+
+static void vp78_update_probability_tables(VP8Context *s)
+{
+    VP56RangeCoder *c = &s->c;
+    int i, j, k, l, m;
+
+    for (i = 0; i < 4; i++)
+        for (j = 0; j < 8; j++)
+            for (k = 0; k < 3; k++)
+                for (l = 0; l < NUM_DCT_TOKENS-1; l++)
+                    if (vp56_rac_get_prob_branchy(c, vp8_token_update_probs[i][j][k][l])) {
+                        int prob = vp8_rac_get_uint(c, 8);
+                        for (m = 0; vp8_coeff_band_indexes[j][m] >= 0; m++)
+                            s->prob->token[i][vp8_coeff_band_indexes[j][m]][k][l] = prob;
+                    }
+}
+
+#define VP7_MVC_SIZE 17
+#define VP8_MVC_SIZE 19
+
+static void vp78_update_pred16x16_pred8x8_mvc_probabilities(VP8Context *s,
+                                                            int mvc_size)
+{
+    VP56RangeCoder *c = &s->c;
+    int i, j;
+
+    if (vp8_rac_get(c))
+        for (i = 0; i < 4; i++)
+            s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);
+    if (vp8_rac_get(c))
+        for (i = 0; i < 3; i++)
+            s->prob->pred8x8c[i]  = vp8_rac_get_uint(c, 8);
+
+    // 17.2 MV probability update
+    for (i = 0; i < 2; i++)
+        for (j = 0; j < mvc_size; j++)
+            if (vp56_rac_get_prob_branchy(c, vp8_mv_update_prob[i][j]))
+                s->prob->mvc[i][j] = vp8_rac_get_nn(c);
+}
+
 static void update_refs(VP8Context *s)
 {
     VP56RangeCoder *c = &s->c;
@@ -381,10 +413,211 @@ static void update_refs(VP8Context *s)
     s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2);
 }
 
-static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
+static void copy_luma(AVFrame *dst, AVFrame *src, int width, int height)
+{
+    int i, j;
+
+    for (j = 1; j < 3; j++) {
+        for (i = 0; i < height / 2; i++)
+            memcpy(dst->data[j] + i * dst->linesize[j],
+                   src->data[j] + i * src->linesize[j], width / 2);
+    }
+}
+
+static void fade(uint8_t *dst, uint8_t *src,
+                 int width, int height, int linesize,
+                 int alpha, int beta)
+{
+    int i, j;
+
+    for (j = 0; j < height; j++) {
+        for (i = 0; i < width; i++) {
+            uint8_t y = src[j * linesize + i];
+            dst[j * linesize + i] = av_clip_uint8(y + ((y * beta) >> 8) + alpha);
+        }
+    }
+}
+
+static int vp7_fade_frame(VP8Context *s, VP56RangeCoder *c)
+{
+    int alpha = (int8_t) vp8_rac_get_uint(c, 8);
+    int beta  = (int8_t) vp8_rac_get_uint(c, 8);
+    int ret;
+
+    if (!s->keyframe && (alpha || beta)) {
+        int width  = s->mb_width * 16;
+        int height = s->mb_height * 16;
+        AVFrame *src, *dst;
+
+        if (!s->framep[VP56_FRAME_PREVIOUS])
+            return AVERROR_INVALIDDATA;
+
+        dst =
+        src = s->framep[VP56_FRAME_PREVIOUS]->tf.f;
+
+        /* preserve the golden frame, write a new previous frame */
+        if (s->framep[VP56_FRAME_GOLDEN] == s->framep[VP56_FRAME_PREVIOUS]) {
+            s->framep[VP56_FRAME_PREVIOUS] = vp8_find_free_buffer(s);
+            if ((ret = vp8_alloc_frame(s, s->framep[VP56_FRAME_PREVIOUS], 1)) < 0)
+               return ret;
+
+            dst = s->framep[VP56_FRAME_PREVIOUS]->tf.f;
+
+            copy_luma(dst, src, width, height);
+        }
+
+        fade(dst->data[0], src->data[0],
+             width, height, dst->linesize[0], alpha, beta);
+    }
+
+    return 0;
+}
+
+static int vp7_decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
 {
     VP56RangeCoder *c = &s->c;
-    int header_size, hscale, vscale, i, j, k, l, m, ret;
+    int part1_size, hscale, vscale, i, j, ret;
+    int width  = s->avctx->width;
+    int height = s->avctx->height;
+
+    s->profile = (buf[0] >> 1) & 7;
+    if (s->profile > 1) {
+        avpriv_request_sample(s->avctx, "Unknown profile %d", s->profile);
+        return AVERROR_INVALIDDATA;
+    }
+
+    s->keyframe  = !(buf[0] & 1);
+    s->invisible = 0;
+    part1_size   = AV_RL24(buf) >> 4;
+
+    buf      += 4 - s->profile;
+    buf_size -= 4 - s->profile;
+
+    memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));
+
+    ff_vp56_init_range_decoder(c, buf, part1_size);
+    buf      += part1_size;
+    buf_size -= part1_size;
+
+    /* A. Dimension information (keyframes only) */
+    if (s->keyframe) {
+        width  = vp8_rac_get_uint(c, 12);
+        height = vp8_rac_get_uint(c, 12);
+        hscale = vp8_rac_get_uint(c, 2);
+        vscale = vp8_rac_get_uint(c, 2);
+        if (hscale || vscale)
+            avpriv_request_sample(s->avctx, "Upscaling");
+
+        s->update_golden = s->update_altref = VP56_FRAME_CURRENT;
+        vp78_reset_probability_tables(s);
+        memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter,
+               sizeof(s->prob->pred16x16));
+        memcpy(s->prob->pred8x8c, vp8_pred8x8c_prob_inter,
+               sizeof(s->prob->pred8x8c));
+        for (i = 0; i < 2; i++)
+            memcpy(s->prob->mvc[i], vp7_mv_default_prob[i],
+                   sizeof(vp7_mv_default_prob[i]));
+        memset(&s->segmentation, 0, sizeof(s->segmentation));
+        memset(&s->lf_delta, 0, sizeof(s->lf_delta));
+        memcpy(s->prob[0].scan, zigzag_scan, sizeof(s->prob[0].scan));
+    }
+
+    if (s->keyframe || s->profile > 0)
+        memset(s->inter_dc_pred, 0 , sizeof(s->inter_dc_pred));
+
+    /* B. Decoding information for all four macroblock-level features */
+    for (i = 0; i < 4; i++) {
+        s->feature_enabled[i] = vp8_rac_get(c);
+        if (s->feature_enabled[i]) {
+             s->feature_present_prob[i] = vp8_rac_get_uint(c, 8);
+
+             for (j = 0; j < 3; j++)
+                 s->feature_index_prob[i][j] =
+                     vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255;
+
+             if (vp7_feature_value_size[s->profile][i])
+                 for (j = 0; j < 4; j++)
+                     s->feature_value[i][j] =
+                         vp8_rac_get(c) ? vp8_rac_get_uint(c, vp7_feature_value_size[s->profile][i]) : 0;
+        }
+    }
+
+    s->segmentation.enabled    = 0;
+    s->segmentation.update_map = 0;
+    s->lf_delta.enabled        = 0;
+
+    s->num_coeff_partitions = 1;
+    ff_vp56_init_range_decoder(&s->coeff_partition[0], buf, buf_size);
+
+    if (!s->macroblocks_base || /* first frame */
+        width != s->avctx->width || height != s->avctx->height ||
+        (width + 15) / 16 != s->mb_width || (height + 15) / 16 != s->mb_height) {
+        if ((ret = vp7_update_dimensions(s, width, height)) < 0)
+            return ret;
+    }
+
+    /* C. Dequantization indices */
+    vp7_get_quants(s);
+
+    /* D. Golden frame update flag (a Flag) for interframes only */
+    if (!s->keyframe) {
+        s->update_golden = vp8_rac_get(c) ? VP56_FRAME_CURRENT : VP56_FRAME_NONE;
+        s->sign_bias[VP56_FRAME_GOLDEN] = 0;
+    }
+
+    s->update_last          = 1;
+    s->update_probabilities = 1;
+    s->fade_present         = 1;
+
+    if (s->profile > 0) {
+        s->update_probabilities = vp8_rac_get(c);
+        if (!s->update_probabilities)
+            s->prob[1] = s->prob[0];
+
+        if (!s->keyframe)
+            s->fade_present = vp8_rac_get(c);
+    }
+
+    /* E. Fading information for previous frame */
+    if (s->fade_present && vp8_rac_get(c)) {
+        if ((ret = vp7_fade_frame(s ,c)) < 0)
+            return ret;
+    }
+
+    /* F. Loop filter type */
+    if (!s->profile)
+        s->filter.simple = vp8_rac_get(c);
+
+    /* G. DCT coefficient ordering specification */
+    if (vp8_rac_get(c))
+        for (i = 1; i < 16; i++)
+            s->prob[0].scan[i] = zigzag_scan[vp8_rac_get_uint(c, 4)];
+
+    /* H. Loop filter levels  */
+    if (s->profile > 0)
+        s->filter.simple = vp8_rac_get(c);
+    s->filter.level     = vp8_rac_get_uint(c, 6);
+    s->filter.sharpness = vp8_rac_get_uint(c, 3);
+
+    /* I. DCT coefficient probability update; 13.3 Token Probability Updates */
+    vp78_update_probability_tables(s);
+
+    s->mbskip_enabled = 0;
+
+    /* J. The remaining frame header data occurs ONLY FOR INTERFRAMES */
+    if (!s->keyframe) {
+        s->prob->intra  = vp8_rac_get_uint(c, 8);
+        s->prob->last   = vp8_rac_get_uint(c, 8);
+        vp78_update_pred16x16_pred8x8_mvc_probabilities(s, VP7_MVC_SIZE);
+    }
+
+    return 0;
+}
+
+static int vp8_decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
+{
+    VP56RangeCoder *c = &s->c;
+    int header_size, hscale, vscale, ret;
     int width  = s->avctx->width;
     int height = s->avctx->height;
 
@@ -399,45 +632,43 @@ static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
         av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);
 
     if (!s->profile)
-        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));
+        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab,
+               sizeof(s->put_pixels_tab));
     else    // profile 1-3 use bilinear, 4+ aren't defined so whatever
-        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab));
+        memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab,
+               sizeof(s->put_pixels_tab));
 
-    if (header_size > buf_size - 7*s->keyframe) {
+    if (header_size > buf_size - 7 * s->keyframe) {
         av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n");
         return AVERROR_INVALIDDATA;
     }
 
     if (s->keyframe) {
         if (AV_RL24(buf) != 0x2a019d) {
-            av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", AV_RL24(buf));
+            av_log(s->avctx, AV_LOG_ERROR,
+                   "Invalid start code 0x%x\n", AV_RL24(buf));
             return AVERROR_INVALIDDATA;
         }
-        width  = AV_RL16(buf+3) & 0x3fff;
-        height = AV_RL16(buf+5) & 0x3fff;
-        hscale = buf[4] >> 6;
-        vscale = buf[6] >> 6;
+        width     = AV_RL16(buf + 3) & 0x3fff;
+        height    = AV_RL16(buf + 5) & 0x3fff;
+        hscale    = buf[4] >> 6;
+        vscale    = buf[6] >> 6;
         buf      += 7;
         buf_size -= 7;
 
         if (hscale || vscale)
-            av_log_missing_feature(s->avctx, "Upscaling", 1);
+            avpriv_request_sample(s->avctx, "Upscaling");
 
         s->update_golden = s->update_altref = VP56_FRAME_CURRENT;
-        for (i = 0; i < 4; i++)
-            for (j = 0; j < 16; j++)
-                memcpy(s->prob->token[i][j], vp8_token_default_probs[i][vp8_coeff_band[j]],
-                       sizeof(s->prob->token[i][j]));
-        memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16));
-        memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c));
-        memcpy(s->prob->mvc      , vp8_mv_default_prob     , sizeof(s->prob->mvc));
+        vp78_reset_probability_tables(s);
+        memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter,
+               sizeof(s->prob->pred16x16));
+        memcpy(s->prob->pred8x8c, vp8_pred8x8c_prob_inter,
+               sizeof(s->prob->pred8x8c));
+        memcpy(s->prob->mvc, vp8_mv_default_prob,
+               sizeof(s->prob->mvc));
         memset(&s->segmentation, 0, sizeof(s->segmentation));
-    }
-
-    if (!s->macroblocks_base || /* first frame */
-        width != s->avctx->width || height != s->avctx->height) {
-        if ((ret = update_dimensions(s, width, height) < 0))
-            return ret;
+        memset(&s->lf_delta, 0, sizeof(s->lf_delta));
     }
 
     ff_vp56_init_range_decoder(c, buf, header_size);
@@ -445,9 +676,10 @@ static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
     buf_size -= header_size;
 
     if (s->keyframe) {
-        if (vp8_rac_get(c))
+        s->colorspace = vp8_rac_get(c);
+        if (s->colorspace)
             av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n");
-        vp8_rac_get(c); // whether we can skip clamping in dsp functions
+        s->fullrange = vp8_rac_get(c);
     }
 
     if ((s->segmentation.enabled = vp8_rac_get(c)))
@@ -468,6 +700,11 @@ static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
         return AVERROR_INVALIDDATA;
     }
 
+    if (!s->macroblocks_base || /* first frame */
+        width != s->avctx->width || height != s->avctx->height)
+        if ((ret = vp8_update_dimensions(s, width, height)) < 0)
+            return ret;
+
     get_quants(s);
 
     if (!s->keyframe) {
@@ -483,15 +720,7 @@ static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
 
     s->update_last = s->keyframe || vp8_rac_get(c);
 
-    for (i = 0; i < 4; i++)
-        for (j = 0; j < 8; j++)
-            for (k = 0; k < 3; k++)
-                for (l = 0; l < NUM_DCT_TOKENS-1; l++)
-                    if (vp56_rac_get_prob_branchy(c, vp8_token_update_probs[i][j][k][l])) {
-                        int prob = vp8_rac_get_uint(c, 8);
-                        for (m = 0; vp8_coeff_band_indexes[j][m] >= 0; m++)
-                            s->prob->token[i][vp8_coeff_band_indexes[j][m]][k][l] = prob;
-                    }
+    vp78_update_probability_tables(s);
 
     if ((s->mbskip_enabled = vp8_rac_get(c)))
         s->prob->mbskip = vp8_rac_get_uint(c, 8);
@@ -500,100 +729,23 @@ static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)
         s->prob->intra  = vp8_rac_get_uint(c, 8);
         s->prob->last   = vp8_rac_get_uint(c, 8);
         s->prob->golden = vp8_rac_get_uint(c, 8);
-
-        if (vp8_rac_get(c))
-            for (i = 0; i < 4; i++)
-                s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);
-        if (vp8_rac_get(c))
-            for (i = 0; i < 3; i++)
-                s->prob->pred8x8c[i]  = vp8_rac_get_uint(c, 8);
-
-        // 17.2 MV probability update
-        for (i = 0; i < 2; i++)
-            for (j = 0; j < 19; j++)
-                if (vp56_rac_get_prob_branchy(c, vp8_mv_update_prob[i][j]))
-                    s->prob->mvc[i][j] = vp8_rac_get_nn(c);
+        vp78_update_pred16x16_pred8x8_mvc_probabilities(s, VP8_MVC_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,
-                   VP56mv near[2], VP56mv *best, uint8_t cnt[4])
+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);
 }
 
 /**
  * Motion vector coding, 17.1.
  */
-static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
+static int read_mv_component(VP56RangeCoder *c, const uint8_t *p, int vp7)
 {
     int bit, x = 0;
 
@@ -602,19 +754,19 @@ static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
 
         for (i = 0; i < 3; i++)
             x += vp56_rac_get_prob(c, p[9 + i]) << i;
-        for (i = 9; i > 3; i--)
+        for (i = (vp7 ? 7 : 9); i > 3; i--)
             x += vp56_rac_get_prob(c, p[9 + i]) << i;
-        if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))
+        if (!(x & (vp7 ? 0xF0 : 0xFFF0)) || vp56_rac_get_prob(c, p[12]))
             x += 8;
     } else {
         // small_mvtree
-        const uint8_t *ps = p+2;
+        const uint8_t *ps = p + 2;
         bit = vp56_rac_get_prob(c, *ps);
-        ps += 1 + 3*bit;
-        x  += 4*bit;
+        ps += 1 + 3 * bit;
+        x  += 4 * bit;
         bit = vp56_rac_get_prob(c, *ps);
         ps += 1 + bit;
-        x  += 2*bit;
+        x  += 2 * bit;
         x  += vp56_rac_get_prob(c, *ps);
     }
 
@@ -622,13 +774,16 @@ static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)
 }
 
 static av_always_inline
-const uint8_t *get_submv_prob(uint32_t left, uint32_t top)
+const uint8_t *get_submv_prob(uint32_t left, uint32_t top, int is_vp7)
 {
+    if (is_vp7)
+        return vp7_submv_prob;
+
     if (left == top)
-        return vp8_submv_prob[4-!!left];
+        return vp8_submv_prob[4 - !!left];
     if (!top)
         return vp8_submv_prob[2];
-    return vp8_submv_prob[1-!!left];
+    return vp8_submv_prob[1 - !!left];
 }
 
 /**
@@ -636,32 +791,38 @@ const uint8_t *get_submv_prob(uint32_t left, uint32_t top)
  * @returns the number of motion vectors parsed (2, 4 or 16)
  */
 static av_always_inline
-int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb)
+int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
+                    int layout, int is_vp7)
 {
     int part_idx;
     int n, num;
-    VP8Macroblock *top_mb  = &mb[2];
+    VP8Macroblock *top_mb;
     VP8Macroblock *left_mb = &mb[-1];
-    const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning],
-                  *mbsplits_top = vp8_mbsplits[top_mb->partitioning],
-                  *mbsplits_cur, *firstidx;
-    VP56mv *top_mv  = top_mb->bmv;
+    const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning];
+    const uint8_t *mbsplits_top, *mbsplits_cur, *firstidx;
+    VP56mv *top_mv;
     VP56mv *left_mv = left_mb->bmv;
     VP56mv *cur_mv  = mb->bmv;
 
+    if (!layout) // layout is inlined, s->mb_layout is not
+        top_mb = &mb[2];
+    else
+        top_mb = &mb[-s->mb_width - 1];
+    mbsplits_top = vp8_mbsplits[top_mb->partitioning];
+    top_mv       = top_mb->bmv;
+
     if (vp56_rac_get_prob_branchy(c, vp8_mbsplit_prob[0])) {
-        if (vp56_rac_get_prob_branchy(c, vp8_mbsplit_prob[1])) {
+        if (vp56_rac_get_prob_branchy(c, vp8_mbsplit_prob[1]))
             part_idx = VP8_SPLITMVMODE_16x8 + vp56_rac_get_prob(c, vp8_mbsplit_prob[2]);
-        } else {
+        else
             part_idx = VP8_SPLITMVMODE_8x8;
-        }
     } else {
         part_idx = VP8_SPLITMVMODE_4x4;
     }
 
-    num = vp8_mbsplit_count[part_idx];
-    mbsplits_cur = vp8_mbsplits[part_idx],
-    firstidx = vp8_mbfirstidx[part_idx];
+    num              = vp8_mbsplit_count[part_idx];
+    mbsplits_cur     = vp8_mbsplits[part_idx],
+    firstidx         = vp8_mbfirstidx[part_idx];
     mb->partitioning = part_idx;
 
     for (n = 0; n < num; n++) {
@@ -672,19 +833,21 @@ int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb)
         if (!(k & 3))
             left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]);
         else
-            left  = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]);
+            left = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]);
         if (k <= 3)
             above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]);
         else
             above = AV_RN32A(&cur_mv[mbsplits_cur[k - 4]]);
 
-        submv_prob = get_submv_prob(left, above);
+        submv_prob = get_submv_prob(left, above, is_vp7);
 
         if (vp56_rac_get_prob_branchy(c, submv_prob[0])) {
             if (vp56_rac_get_prob_branchy(c, submv_prob[1])) {
                 if (vp56_rac_get_prob_branchy(c, submv_prob[2])) {
-                    mb->bmv[n].y = mb->mv.y + read_mv_component(c, s->prob->mvc[0]);
-                    mb->bmv[n].x = mb->mv.x + read_mv_component(c, s->prob->mvc[1]);
+                    mb->bmv[n].y = mb->mv.y +
+                                   read_mv_component(c, s->prob->mvc[0], is_vp7);
+                    mb->bmv[n].x = mb->mv.x +
+                                   read_mv_component(c, s->prob->mvc[1], is_vp7);
                 } else {
                     AV_ZERO32(&mb->bmv[n]);
                 }
@@ -699,219 +862,532 @@ int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb)
     return num;
 }
 
+/**
+ * The vp7 reference decoder uses a padding macroblock column (added to right
+ * edge of the frame) to guard against illegal macroblock offsets. The
+ * algorithm has bugs that permit offsets to straddle the padding column.
+ * This function replicates those bugs.
+ *
+ * @param[out] edge_x macroblock x address
+ * @param[out] edge_y macroblock y address
+ *
+ * @return macroblock offset legal (boolean)
+ */
+static int vp7_calculate_mb_offset(int mb_x, int mb_y, int mb_width,
+                                   int xoffset, int yoffset, int boundary,
+                                   int *edge_x, int *edge_y)
+{
+    int vwidth = mb_width + 1;
+    int new = (mb_y + yoffset) * vwidth + mb_x + xoffset;
+    if (new < boundary || new % vwidth == vwidth - 1)
+        return 0;
+    *edge_y = new / vwidth;
+    *edge_x = new % vwidth;
+    return 1;
+}
+
+static const VP56mv *get_bmv_ptr(const VP8Macroblock *mb, int subblock)
+{
+    return &mb->bmv[mb->mode == VP8_MVMODE_SPLIT ? vp8_mbsplits[mb->partitioning][subblock] : 0];
+}
+
+static av_always_inline
+void vp7_decode_mvs(VP8Context *s, VP8Macroblock *mb,
+                    int mb_x, int mb_y, int layout)
+{
+    VP8Macroblock *mb_edge[12];
+    enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR };
+    enum { VP8_EDGE_TOP, VP8_EDGE_LEFT, VP8_EDGE_TOPLEFT };
+    int idx = CNT_ZERO;
+    VP56mv near_mv[3];
+    uint8_t cnt[3] = { 0 };
+    VP56RangeCoder *c = &s->c;
+    int i;
+
+    AV_ZERO32(&near_mv[0]);
+    AV_ZERO32(&near_mv[1]);
+    AV_ZERO32(&near_mv[2]);
+
+    for (i = 0; i < VP7_MV_PRED_COUNT; i++) {
+        const VP7MVPred * pred = &vp7_mv_pred[i];
+        int edge_x, edge_y;
+
+        if (vp7_calculate_mb_offset(mb_x, mb_y, s->mb_width, pred->xoffset,
+                                    pred->yoffset, !s->profile, &edge_x, &edge_y)) {
+            VP8Macroblock *edge = mb_edge[i] = (s->mb_layout == 1)
+                                             ? s->macroblocks_base + 1 + edge_x +
+                                               (s->mb_width + 1) * (edge_y + 1)
+                                             : s->macroblocks + edge_x +
+                                               (s->mb_height - edge_y - 1) * 2;
+            uint32_t mv = AV_RN32A(get_bmv_ptr(edge, vp7_mv_pred[i].subblock));
+            if (mv) {
+                if (AV_RN32A(&near_mv[CNT_NEAREST])) {
+                    if (mv == AV_RN32A(&near_mv[CNT_NEAREST])) {
+                        idx = CNT_NEAREST;
+                    } else if (AV_RN32A(&near_mv[CNT_NEAR])) {
+                        if (mv != AV_RN32A(&near_mv[CNT_NEAR]))
+                            continue;
+                        idx = CNT_NEAR;
+                    } else {
+                        AV_WN32A(&near_mv[CNT_NEAR], mv);
+                        idx = CNT_NEAR;
+                    }
+                } else {
+                    AV_WN32A(&near_mv[CNT_NEAREST], mv);
+                    idx = CNT_NEAREST;
+                }
+            } else {
+                idx = CNT_ZERO;
+            }
+        } else {
+            idx = CNT_ZERO;
+        }
+        cnt[idx] += vp7_mv_pred[i].score;
+    }
+
+    mb->partitioning = VP8_SPLITMVMODE_NONE;
+
+    if (vp56_rac_get_prob_branchy(c, vp7_mode_contexts[cnt[CNT_ZERO]][0])) {
+        mb->mode = VP8_MVMODE_MV;
+
+        if (vp56_rac_get_prob_branchy(c, vp7_mode_contexts[cnt[CNT_NEAREST]][1])) {
+
+            if (vp56_rac_get_prob_branchy(c, vp7_mode_contexts[cnt[CNT_NEAR]][2])) {
+
+                if (cnt[CNT_NEAREST] > cnt[CNT_NEAR])
+                    AV_WN32A(&mb->mv, cnt[CNT_ZERO] > cnt[CNT_NEAREST] ? 0 : AV_RN32A(&near_mv[CNT_NEAREST]));
+                else
+                    AV_WN32A(&mb->mv, cnt[CNT_ZERO] > cnt[CNT_NEAR]    ? 0 : AV_RN32A(&near_mv[CNT_NEAR]));
+
+                if (vp56_rac_get_prob_branchy(c, vp7_mode_contexts[cnt[CNT_NEAR]][3])) {
+                    mb->mode = VP8_MVMODE_SPLIT;
+                    mb->mv = mb->bmv[decode_splitmvs(s, c, mb, layout, IS_VP7) - 1];
+                } else {
+                    mb->mv.y += read_mv_component(c, s->prob->mvc[0], IS_VP7);
+                    mb->mv.x += read_mv_component(c, s->prob->mvc[1], IS_VP7);
+                    mb->bmv[0] = mb->mv;
+                }
+            } else {
+                mb->mv = near_mv[CNT_NEAR];
+                mb->bmv[0] = mb->mv;
+            }
+        } else {
+            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 vp8_decode_mvs(VP8Context *s, VP8Macroblock *mb,
+                    int mb_x, int mb_y, int layout)
+{
+    VP8Macroblock *mb_edge[3] = { 0      /* top */,
+                                  mb - 1 /* left */,
+                                  0      /* 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;
+
+    if (!layout) { // layout is inlined (s->mb_layout is not)
+        mb_edge[0] = mb + 2;
+        mb_edge[2] = mb + 1;
+    } else {
+        mb_edge[0] = mb - s->mb_width - 1;
+        mb_edge[2] = mb - s->mb_width - 2;
+    }
+
+    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, layout, IS_VP8) - 1];
+                } else {
+                    mb->mv.y  += read_mv_component(c, s->prob->mvc[0], IS_VP8);
+                    mb->mv.x  += read_mv_component(c, s->prob->mvc[1], IS_VP8);
+                    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)
+void decode_intra4x4_modes(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
+                           int mb_x, int keyframe, int layout)
 {
-    uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;
+    uint8_t *intra4x4 = mb->intra4x4_pred_mode_mb;
+
+    if (layout == 1) {
+        VP8Macroblock *mb_top = mb - s->mb_width - 1;
+        memcpy(mb->intra4x4_pred_mode_top, mb_top->intra4x4_pred_mode_top, 4);
+    }
     if (keyframe) {
         int x, y;
-        uint8_t* const top = s->intra4x4_pred_mode_top + 4 * mb_x;
-        uint8_t* const left = s->intra4x4_pred_mode_left;
+        uint8_t *top;
+        uint8_t *const left = s->intra4x4_pred_mode_left;
+        if (layout == 1)
+            top = mb->intra4x4_pred_mode_top;
+        else
+            top = s->intra4x4_pred_mode_top + 4 * mb_x;
         for (y = 0; y < 4; y++) {
             for (x = 0; x < 4; x++) {
                 const uint8_t *ctx;
-                ctx = vp8_pred4x4_prob_intra[top[x]][left[y]];
+                ctx       = vp8_pred4x4_prob_intra[top[x]][left[y]];
                 *intra4x4 = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);
-                left[y] = top[x] = *intra4x4;
+                left[y]   = top[x] = *intra4x4;
                 intra4x4++;
             }
         }
     } else {
         int i;
         for (i = 0; i < 16; i++)
-            intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter);
+            intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree,
+                                           vp8_pred4x4_prob_inter);
     }
 }
 
 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, int layout, int is_vp7)
 {
     VP56RangeCoder *c = &s->c;
-
-    if (s->segmentation.update_map)
+    const char *vp7_feature_name[] = { "q-index",
+                                       "lf-delta",
+                                       "partial-golden-update",
+                                       "blit-pitch" };
+    if (is_vp7) {
+        int i;
+        *segment = 0;
+        for (i = 0; i < 4; i++) {
+            if (s->feature_enabled[i]) {
+                if (vp56_rac_get_prob(c, s->feature_present_prob[i])) {
+                      int index = vp8_rac_get_tree(c, vp7_feature_index_tree,
+                                                   s->feature_index_prob[i]);
+                      av_log(s->avctx, AV_LOG_WARNING,
+                             "Feature %s present in macroblock (value 0x%x)\n",
+                             vp7_feature_name[i], s->feature_value[i][index]);
+                }
+           }
+        }
+    } else if (s->segmentation.update_map)
         *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);
-    s->segment = *segment;
+    else if (s->segmentation.enabled)
+        *segment = ref ? *ref : *segment;
+    mb->segment = *segment;
 
     mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;
 
     if (s->keyframe) {
-        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);
+        mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra,
+                                    vp8_pred16x16_prob_intra);
 
         if (mb->mode == MODE_I4x4) {
-            decode_intra4x4_modes(s, c, mb_x, 1);
+            decode_intra4x4_modes(s, c, mb, mb_x, 1, layout);
         } else {
-            const uint32_t modes = vp8_pred4x4_mode[mb->mode] * 0x01010101u;
-            AV_WN32A(s->intra4x4_pred_mode_top + 4 * mb_x, modes);
+            const uint32_t modes = (is_vp7 ? vp7_pred4x4_mode
+                                           : vp8_pred4x4_mode)[mb->mode] * 0x01010101u;
+            if (s->mb_layout == 1)
+                AV_WN32A(mb->intra4x4_pred_mode_top, modes);
+            else
+                AV_WN32A(s->intra4x4_pred_mode_top + 4 * mb_x, modes);
             AV_WN32A(s->intra4x4_pred_mode_left, modes);
         }
 
-        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);
-        mb->ref_frame = VP56_FRAME_CURRENT;
+        mb->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) ?
-                VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;
+            mb->ref_frame =
+                (!is_vp7 && vp56_rac_get_prob(c, s->prob->golden)) ? VP56_FRAME_GOLDEN2 /* altref */
+                                                                   : VP56_FRAME_GOLDEN;
         else
             mb->ref_frame = VP56_FRAME_PREVIOUS;
-        s->ref_count[mb->ref_frame-1]++;
+        s->ref_count[mb->ref_frame - 1]++;
 
         // motion vectors, 16.3
-        find_near_mvs(s, mb, near, &best, cnt);
-        if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[0]][0])) {
-            mb->mode = VP8_MVMODE_MV;
-            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 {
-                        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
-                    clamp_mv(s, &mb->mv, &near[1], mb_x, mb_y);
-            } else
-                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;
-        }
+        if (is_vp7)
+            vp7_decode_mvs(s, mb, mb_x, mb_y, layout);
+        else
+            vp8_decode_mvs(s, mb, mb_x, mb_y, layout);
     } else {
         // intra MB, 16.1
         mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);
 
         if (mb->mode == MODE_I4x4)
-            decode_intra4x4_modes(s, c, mb_x, 0);
+            decode_intra4x4_modes(s, c, mb, mb_x, 0, layout);
 
-        s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);
-        mb->ref_frame = VP56_FRAME_CURRENT;
-        mb->partitioning = VP8_SPLITMVMODE_NONE;
+        mb->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree,
+                                                s->prob->pred8x8c);
+        mb->ref_frame        = VP56_FRAME_CURRENT;
+        mb->partitioning     = VP8_SPLITMVMODE_NONE;
         AV_ZERO32(&mb->bmv[0]);
     }
 }
 
 /**
- * @param c arithmetic bitstream reader context
+ * @param r     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
+ * @param i     initial coeff index, 0 unless a separate DC block is coded
+ * @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_internal(VP56RangeCoder *c, DCTELEM block[16],
-                                        uint8_t probs[8][3][NUM_DCT_TOKENS-1],
-                                        int i, uint8_t *token_prob, int16_t qmul[2])
+static av_always_inline
+int decode_block_coeffs_internal(VP56RangeCoder *r, int16_t block[16],
+                                 uint8_t probs[16][3][NUM_DCT_TOKENS - 1],
+                                 int i, uint8_t *token_prob, int16_t qmul[2],
+                                 const uint8_t scan[16], int vp7)
 {
+    VP56RangeCoder c = *r;
     goto skip_eob;
     do {
         int coeff;
-        if (!vp56_rac_get_prob_branchy(c, token_prob[0]))   // DCT_EOB
-            return i;
+restart:
+        if (!vp56_rac_get_prob_branchy(&c, token_prob[0]))   // DCT_EOB
+            break;
 
 skip_eob:
-        if (!vp56_rac_get_prob_branchy(c, token_prob[1])) { // DCT_0
+        if (!vp56_rac_get_prob_branchy(&c, token_prob[1])) { // DCT_0
             if (++i == 16)
-                return i; // invalid input; blocks should end with EOB
+                break; // invalid input; blocks should end with EOB
             token_prob = probs[i][0];
+            if (vp7)
+                goto restart;
             goto skip_eob;
         }
 
-        if (!vp56_rac_get_prob_branchy(c, token_prob[2])) { // DCT_1
+        if (!vp56_rac_get_prob_branchy(&c, token_prob[2])) { // DCT_1
             coeff = 1;
-            token_prob = probs[i+1][1];
+            token_prob = probs[i + 1][1];
         } else {
-            if (!vp56_rac_get_prob_branchy(c, token_prob[3])) { // DCT 2,3,4
-                coeff = vp56_rac_get_prob_branchy(c, token_prob[4]);
+            if (!vp56_rac_get_prob_branchy(&c, token_prob[3])) { // DCT 2,3,4
+                coeff = vp56_rac_get_prob_branchy(&c, token_prob[4]);
                 if (coeff)
-                    coeff += vp56_rac_get_prob(c, token_prob[5]);
+                    coeff += vp56_rac_get_prob(&c, token_prob[5]);
                 coeff += 2;
             } else {
                 // DCT_CAT*
-                if (!vp56_rac_get_prob_branchy(c, token_prob[6])) {
-                    if (!vp56_rac_get_prob_branchy(c, token_prob[7])) { // DCT_CAT1
-                        coeff  = 5 + vp56_rac_get_prob(c, vp8_dct_cat1_prob[0]);
+                if (!vp56_rac_get_prob_branchy(&c, token_prob[6])) {
+                    if (!vp56_rac_get_prob_branchy(&c, token_prob[7])) { // DCT_CAT1
+                        coeff = 5 + vp56_rac_get_prob(&c, vp8_dct_cat1_prob[0]);
                     } else {                                    // DCT_CAT2
                         coeff  = 7;
-                        coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[0]) << 1;
-                        coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[1]);
+                        coeff += vp56_rac_get_prob(&c, vp8_dct_cat2_prob[0]) << 1;
+                        coeff += vp56_rac_get_prob(&c, vp8_dct_cat2_prob[1]);
                     }
                 } else {    // DCT_CAT3 and up
-                    int a = vp56_rac_get_prob(c, token_prob[8]);
-                    int b = vp56_rac_get_prob(c, token_prob[9+a]);
-                    int cat = (a<<1) + b;
-                    coeff  = 3 + (8<<cat);
-                    coeff += vp8_rac_get_coeff(c, vp8_dct_cat_prob[cat]);
+                    int a   = vp56_rac_get_prob(&c, token_prob[8]);
+                    int b   = vp56_rac_get_prob(&c, token_prob[9 + a]);
+                    int cat = (a << 1) + b;
+                    coeff  = 3 + (8 << cat);
+                    coeff += vp8_rac_get_coeff(&c, ff_vp8_dct_cat_prob[cat]);
                 }
             }
-            token_prob = probs[i+1][2];
+            token_prob = probs[i + 1][2];
         }
-        block[zigzag_scan[i]] = (vp8_rac_get(c) ? -coeff : coeff) * qmul[!!i];
+        block[scan[i]] = (vp8_rac_get(&c) ? -coeff : coeff) * qmul[!!i];
     } while (++i < 16);
 
+    *r = c;
     return i;
 }
 
 static av_always_inline
-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])
+int inter_predict_dc(int16_t block[16], int16_t pred[2])
+{
+    int16_t dc = block[0];
+    int ret = 0;
+
+    if (pred[1] > 3) {
+        dc += pred[0];
+        ret = 1;
+    }
+
+    if (!pred[0] | !dc | ((int32_t)pred[0] ^ (int32_t)dc) >> 31) {
+        block[0] = pred[0] = dc;
+        pred[1] = 0;
+    } else {
+        if (pred[0] == dc)
+            pred[1]++;
+        block[0] = pred[0] = dc;
+    }
+
+    return ret;
+}
+
+static int vp7_decode_block_coeffs_internal(VP56RangeCoder *r,
+                                            int16_t block[16],
+                                            uint8_t probs[16][3][NUM_DCT_TOKENS - 1],
+                                            int i, uint8_t *token_prob,
+                                            int16_t qmul[2],
+                                            const uint8_t scan[16])
+{
+    return decode_block_coeffs_internal(r, block, probs, i,
+                                        token_prob, qmul, scan, IS_VP7);
+}
+
+#ifndef vp8_decode_block_coeffs_internal
+static int vp8_decode_block_coeffs_internal(VP56RangeCoder *r,
+                                            int16_t block[16],
+                                            uint8_t probs[16][3][NUM_DCT_TOKENS - 1],
+                                            int i, uint8_t *token_prob,
+                                            int16_t qmul[2])
+{
+    return decode_block_coeffs_internal(r, block, probs, i,
+                                        token_prob, qmul, zigzag_scan, IS_VP8);
+}
+#endif
+
+/**
+ * @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 av_always_inline
+int decode_block_coeffs(VP56RangeCoder *c, int16_t block[16],
+                        uint8_t probs[16][3][NUM_DCT_TOKENS - 1],
+                        int i, int zero_nhood, int16_t qmul[2],
+                        const uint8_t scan[16], int vp7)
 {
     uint8_t *token_prob = probs[i][zero_nhood];
     if (!vp56_rac_get_prob_branchy(c, token_prob[0]))   // DCT_EOB
         return 0;
-    return decode_block_coeffs_internal(c, block, probs, i, token_prob, qmul);
+    return vp7 ? vp7_decode_block_coeffs_internal(c, block, probs, i,
+                                                  token_prob, qmul, scan)
+               : vp8_decode_block_coeffs_internal(c, block, probs, i,
+                                                  token_prob, qmul);
 }
 
 static av_always_inline
-void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
-                      uint8_t t_nnz[9], uint8_t l_nnz[9])
+void decode_mb_coeffs(VP8Context *s, VP8ThreadData *td, VP56RangeCoder *c,
+                      VP8Macroblock *mb, uint8_t t_nnz[9], uint8_t l_nnz[9],
+                      int is_vp7)
 {
     int i, x, y, luma_start = 0, luma_ctx = 3;
     int nnz_pred, nnz, nnz_total = 0;
-    int segment = s->segment;
+    int segment = mb->segment;
     int block_dc = 0;
 
-    if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
+    if (mb->mode != MODE_I4x4 && (is_vp7 || mb->mode != VP8_MVMODE_SPLIT)) {
         nnz_pred = t_nnz[8] + l_nnz[8];
 
         // decode DC values and do hadamard
-        nnz = decode_block_coeffs(c, s->block_dc, s->prob->token[1], 0, nnz_pred,
-                                  s->qmat[segment].luma_dc_qmul);
+        nnz = decode_block_coeffs(c, td->block_dc, s->prob->token[1], 0,
+                                  nnz_pred, s->qmat[segment].luma_dc_qmul,
+                                  zigzag_scan, is_vp7);
         l_nnz[8] = t_nnz[8] = !!nnz;
+
+        if (is_vp7 && mb->mode > MODE_I4x4) {
+            nnz |=  inter_predict_dc(td->block_dc,
+                                     s->inter_dc_pred[mb->ref_frame - 1]);
+        }
+
         if (nnz) {
             nnz_total += nnz;
-            block_dc = 1;
+            block_dc   = 1;
             if (nnz == 1)
-                s->vp8dsp.vp8_luma_dc_wht_dc(s->block, s->block_dc);
+                s->vp8dsp.vp8_luma_dc_wht_dc(td->block, td->block_dc);
             else
-                s->vp8dsp.vp8_luma_dc_wht(s->block, s->block_dc);
+                s->vp8dsp.vp8_luma_dc_wht(td->block, td->block_dc);
         }
         luma_start = 1;
-        luma_ctx = 0;
+        luma_ctx   = 0;
     }
 
     // luma blocks
     for (y = 0; y < 4; y++)
         for (x = 0; x < 4; x++) {
             nnz_pred = l_nnz[y] + t_nnz[x];
-            nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,
-                                      nnz_pred, s->qmat[segment].luma_qmul);
-            // nnz+block_dc may be one more than the actual last index, but we don't care
-            s->non_zero_count_cache[y][x] = nnz + block_dc;
+            nnz = decode_block_coeffs(c, td->block[y][x],
+                                      s->prob->token[luma_ctx],
+                                      luma_start, nnz_pred,
+                                      s->qmat[segment].luma_qmul,
+                                      s->prob[0].scan, is_vp7);
+            /* nnz+block_dc may be one more than the actual last index,
+             * but we don't care */
+            td->non_zero_count_cache[y][x] = nnz + block_dc;
             t_nnz[x] = l_nnz[y] = !!nnz;
             nnz_total += nnz;
         }
@@ -922,11 +1398,13 @@ void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
     for (i = 4; i < 6; i++)
         for (y = 0; y < 2; y++)
             for (x = 0; x < 2; x++) {
-                nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];
-                nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,
-                                          nnz_pred, s->qmat[segment].chroma_qmul);
-                s->non_zero_count_cache[i][(y<<1)+x] = nnz;
-                t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;
+                nnz_pred = l_nnz[i + 2 * y] + t_nnz[i + 2 * x];
+                nnz = decode_block_coeffs(c, td->block[i][(y << 1) + x],
+                                          s->prob->token[2], 0, nnz_pred,
+                                          s->qmat[segment].chroma_qmul,
+                                          s->prob[0].scan, is_vp7);
+                td->non_zero_count_cache[i][(y << 1) + x] = nnz;
+                t_nnz[i + 2 * x] = l_nnz[i + 2 * y] = !!nnz;
                 nnz_total += nnz;
             }
 
@@ -938,105 +1416,98 @@ void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,
 }
 
 static av_always_inline
-void backup_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
+void backup_mb_border(uint8_t *top_border, uint8_t *src_y,
+                      uint8_t *src_cb, uint8_t *src_cr,
                       int linesize, int uvlinesize, int simple)
 {
-    AV_COPY128(top_border, src_y + 15*linesize);
+    AV_COPY128(top_border, src_y + 15 * linesize);
     if (!simple) {
-        AV_COPY64(top_border+16, src_cb + 7*uvlinesize);
-        AV_COPY64(top_border+24, src_cr + 7*uvlinesize);
+        AV_COPY64(top_border + 16, src_cb + 7 * uvlinesize);
+        AV_COPY64(top_border + 24, src_cr + 7 * uvlinesize);
     }
 }
 
 static av_always_inline
-void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,
-                    int linesize, int uvlinesize, int mb_x, int mb_y, int mb_width,
-                    int simple, int xchg)
+void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb,
+                    uint8_t *src_cr, int linesize, int uvlinesize, int mb_x,
+                    int mb_y, int mb_width, int simple, int xchg)
 {
-    uint8_t *top_border_m1 = top_border-32;     // for TL prediction
-    src_y  -=   linesize;
+    uint8_t *top_border_m1 = top_border - 32;     // for TL prediction
+    src_y  -= linesize;
     src_cb -= uvlinesize;
     src_cr -= uvlinesize;
 
-#define XCHG(a,b,xchg) do {                     \
-        if (xchg) AV_SWAP64(b,a);               \
-        else      AV_COPY64(b,a);               \
+#define XCHG(a, b, xchg)                                                      \
+    do {                                                                      \
+        if (xchg)                                                             \
+            AV_SWAP64(b, a);                                                  \
+        else                                                                  \
+            AV_COPY64(b, a);                                                  \
     } while (0)
 
-    XCHG(top_border_m1+8, src_y-8, xchg);
-    XCHG(top_border,      src_y,   xchg);
-    XCHG(top_border+8,    src_y+8, 1);
-    if (mb_x < mb_width-1)
-        XCHG(top_border+32, src_y+16, 1);
+    XCHG(top_border_m1 + 8, src_y - 8, xchg);
+    XCHG(top_border, src_y, xchg);
+    XCHG(top_border + 8, src_y + 8, 1);
+    if (mb_x < mb_width - 1)
+        XCHG(top_border + 32, src_y + 16, 1);
 
     // only copy chroma for normal loop filter
     // or to initialize the top row to 127
     if (!simple || !mb_y) {
-        XCHG(top_border_m1+16, src_cb-8, xchg);
-        XCHG(top_border_m1+24, src_cr-8, xchg);
-        XCHG(top_border+16,    src_cb, 1);
-        XCHG(top_border+24,    src_cr, 1);
+        XCHG(top_border_m1 + 16, src_cb - 8, xchg);
+        XCHG(top_border_m1 + 24, src_cr - 8, xchg);
+        XCHG(top_border + 16, src_cb, 1);
+        XCHG(top_border + 24, src_cr, 1);
     }
 }
 
 static av_always_inline
 int check_dc_pred8x8_mode(int mode, int mb_x, int mb_y)
 {
-    if (!mb_x) {
+    if (!mb_x)
         return mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;
-    } else {
+    else
         return mb_y ? mode : LEFT_DC_PRED8x8;
-    }
 }
 
 static av_always_inline
-int check_tm_pred8x8_mode(int mode, int mb_x, int mb_y)
+int check_tm_pred8x8_mode(int mode, int mb_x, int mb_y, int vp7)
 {
-    if (!mb_x) {
-        return mb_y ? VERT_PRED8x8 : DC_129_PRED8x8;
-    } else {
+    if (!mb_x)
+        return mb_y ? VERT_PRED8x8 : (vp7 ? DC_128_PRED8x8 : DC_129_PRED8x8);
+    else
         return mb_y ? mode : HOR_PRED8x8;
-    }
-}
-
-static av_always_inline
-int check_intra_pred8x8_mode(int mode, int mb_x, int mb_y)
-{
-    if (mode == DC_PRED8x8) {
-        return check_dc_pred8x8_mode(mode, mb_x, mb_y);
-    } else {
-        return mode;
-    }
 }
 
 static av_always_inline
-int check_intra_pred8x8_mode_emuedge(int mode, int mb_x, int mb_y)
+int check_intra_pred8x8_mode_emuedge(int mode, int mb_x, int mb_y, int vp7)
 {
     switch (mode) {
     case DC_PRED8x8:
         return check_dc_pred8x8_mode(mode, mb_x, mb_y);
     case VERT_PRED8x8:
-        return !mb_y ? DC_127_PRED8x8 : mode;
+        return !mb_y ? (vp7 ? DC_128_PRED8x8 : DC_127_PRED8x8) : mode;
     case HOR_PRED8x8:
-        return !mb_x ? DC_129_PRED8x8 : mode;
-    case PLANE_PRED8x8 /*TM*/:
-        return check_tm_pred8x8_mode(mode, mb_x, mb_y);
+        return !mb_x ? (vp7 ? DC_128_PRED8x8 : DC_129_PRED8x8) : mode;
+    case PLANE_PRED8x8: /* TM */
+        return check_tm_pred8x8_mode(mode, mb_x, mb_y, vp7);
     }
     return mode;
 }
 
 static av_always_inline
-int check_tm_pred4x4_mode(int mode, int mb_x, int mb_y)
+int check_tm_pred4x4_mode(int mode, int mb_x, int mb_y, int vp7)
 {
     if (!mb_x) {
-        return mb_y ? VERT_VP8_PRED : DC_129_PRED;
+        return mb_y ? VERT_VP8_PRED : (vp7 ? DC_128_PRED : DC_129_PRED);
     } else {
         return mb_y ? mode : HOR_VP8_PRED;
     }
 }
 
 static av_always_inline
-int check_intra_pred4x4_mode_emuedge(int mode, int mb_x, int mb_y, int *copy_buf)
+int check_intra_pred4x4_mode_emuedge(int mode, int mb_x, int mb_y,
+                                     int *copy_buf, int vp7)
 {
     switch (mode) {
     case VERT_PRED:
@@ -1047,7 +1518,7 @@ int check_intra_pred4x4_mode_emuedge(int mode, int mb_x, int mb_y, int *copy_buf
         /* fall-through */
     case DIAG_DOWN_LEFT_PRED:
     case VERT_LEFT_PRED:
-        return !mb_y ? DC_127_PRED : mode;
+        return !mb_y ? (vp7 ? DC_128_PRED : DC_127_PRED) : mode;
     case HOR_PRED:
         if (!mb_y) {
             *copy_buf = 1;
@@ -1055,10 +1526,11 @@ int check_intra_pred4x4_mode_emuedge(int mode, int mb_x, int mb_y, int *copy_buf
         }
         /* fall-through */
     case HOR_UP_PRED:
-        return !mb_x ? DC_129_PRED : mode;
+        return !mb_x ? (vp7 ? DC_128_PRED : DC_129_PRED) : mode;
     case TM_VP8_PRED:
-        return check_tm_pred4x4_mode(mode, mb_x, mb_y);
-    case DC_PRED: // 4x4 DC doesn't use the same "H.264-style" exceptions as 16x16/8x8 DC
+        return check_tm_pred4x4_mode(mode, mb_x, mb_y, vp7);
+    case DC_PRED: /* 4x4 DC doesn't use the same "H.264-style" exceptions
+                   * as 16x16/8x8 DC */
     case DIAG_DOWN_RIGHT_PRED:
     case VERT_RIGHT_PRED:
     case HOR_DOWN_PRED:
@@ -1070,30 +1542,28 @@ int check_intra_pred4x4_mode_emuedge(int mode, int mb_x, int mb_y, int *copy_buf
 }
 
 static av_always_inline
-void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
-                   int mb_x, int mb_y)
+void intra_predict(VP8Context *s, VP8ThreadData *td, uint8_t *dst[3],
+                   VP8Macroblock *mb, int mb_x, int mb_y, int is_vp7)
 {
-    AVCodecContext *avctx = s->avctx;
-    int x, y, mode, nnz, tr;
+    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 (!(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],
+    /* 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 (mb_y && (s->deblock_filter || !mb_y) && td->thread_nr == 0)
+        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) {
-        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);
-        }
+        mode = check_intra_pred8x8_mode_emuedge(mb->mode, mb_x, mb_y, is_vp7);
         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 };
+        uint8_t *intra4x4 = mb->intra4x4_pred_mode_mb;
+        const uint8_t lo = is_vp7 ? 128 : 127;
+        const uint8_t hi = is_vp7 ? 128 : 129;
+        uint8_t tr_top[4] = { lo, lo, lo, lo };
 
         // all blocks on the right edge of the macroblock use bottom edge
         // the top macroblock for their topright edge
@@ -1101,91 +1571,86 @@ 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_y && avctx->flags & CODEC_FLAG_EMU_EDGE) &&
-            mb_x == s->mb_width-1) {
-            tr = tr_right[-1]*0x01010101;
-            tr_right = (uint8_t *)&tr;
+        if (mb_y && mb_x == s->mb_width - 1) {
+            tr       = tr_right[-1] * 0x01010101u;
+            tr_right = (uint8_t *) &tr;
         }
 
         if (mb->skip)
-            AV_ZERO128(s->non_zero_count_cache);
+            AV_ZERO128(td->non_zero_count_cache);
 
         for (y = 0; y < 4; y++) {
             uint8_t *topright = ptr + 4 - s->linesize;
             for (x = 0; x < 4; x++) {
                 int copy = 0, linesize = s->linesize;
-                uint8_t *dst = ptr+4*x;
-                DECLARE_ALIGNED(4, uint8_t, copy_dst)[5*8];
+                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) {
+                if ((y == 0 || x == 3) && mb_y == 0) {
                     topright = tr_top;
                 } else if (x == 3)
                     topright = tr_right;
 
-                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, &copy);
-                    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];
-                            }
-                        }
+                mode = check_intra_pred4x4_mode_emuedge(intra4x4[x], mb_x + x,
+                                                        mb_y + y, &copy, is_vp7);
+                if (copy) {
+                    dst      = copy_dst + 12;
+                    linesize = 8;
+                    if (!(mb_y + y)) {
+                        copy_dst[3] = lo;
+                        AV_WN32A(copy_dst + 4, lo * 0x01010101U);
+                    } else {
+                        AV_COPY32(copy_dst + 4, ptr + 4 * x - s->linesize);
                         if (!(mb_x + x)) {
-                            copy_dst[11] =
-                            copy_dst[19] =
-                            copy_dst[27] =
-                            copy_dst[35] = 129U;
+                            copy_dst[3] = hi;
                         } 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];
+                            copy_dst[3] = ptr[4 * x - s->linesize - 1];
                         }
                     }
-                } else {
-                    mode = intra4x4[x];
+                    if (!(mb_x + x)) {
+                        copy_dst[11] =
+                        copy_dst[19] =
+                        copy_dst[27] =
+                        copy_dst[35] = hi;
+                    } 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];
+                    }
                 }
                 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);
+                    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];
+                nnz = td->non_zero_count_cache[y][x];
                 if (nnz) {
                     if (nnz == 1)
-                        s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);
+                        s->vp8dsp.vp8_idct_dc_add(ptr + 4 * x,
+                                                  td->block[y][x], s->linesize);
                     else
-                        s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);
+                        s->vp8dsp.vp8_idct_add(ptr + 4 * x,
+                                               td->block[y][x], s->linesize);
                 }
                 topright += 4;
             }
 
-            ptr   += 4*s->linesize;
+            ptr      += 4 * s->linesize;
             intra4x4 += 4;
         }
     }
 
-    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);
-    }
+    mode = check_intra_pred8x8_mode_emuedge(mb->chroma_pred_mode,
+                                            mb_x, mb_y, is_vp7);
     s->hpc.pred8x8[mode](dst[1], s->uvlinesize);
     s->hpc.pred8x8[mode](dst[2], s->uvlinesize);
 
-    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],
+    if (mb_y && (s->deblock_filter || !mb_y) && td->thread_nr == 0)
+        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);
 }
@@ -1198,59 +1663,91 @@ static const uint8_t subpel_idx[3][8] = {
 };
 
 /**
- * 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 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 s        VP8 decoding context
+ * @param dst      target buffer for block data at block position
+ * @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)
+ * @param mc_func  motion compensation function pointers (bilinear or sixtap MC)
  */
 static av_always_inline
-void vp8_mc_luma(VP8Context *s, uint8_t *dst, uint8_t *src, const VP56mv *mv,
+void vp8_mc_luma(VP8Context *s, VP8ThreadData *td, uint8_t *dst,
+                 ThreadFrame *ref, const VP56mv *mv,
                  int x_off, int y_off, int block_w, int block_h,
-                 int width, int height, int linesize,
+                 int width, int height, ptrdiff_t linesize,
                  vp8_mc_func mc_func[3][3])
 {
+    uint8_t *src = ref->f->data[0];
+
     if (AV_RN32A(mv)) {
+        int src_linesize = linesize;
 
-        int mx = (mv->x << 1)&7, mx_idx = subpel_idx[0][mx];
-        int my = (mv->y << 1)&7, my_idx = subpel_idx[0][my];
+        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 < 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;
+            s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
+                                     src - my_idx * linesize - mx_idx,
+                                     EDGE_EMU_LINESIZE, linesize,
+                                     block_w + subpel_idx[1][mx],
+                                     block_h + subpel_idx[1][my],
+                                     x_off - mx_idx, y_off - my_idx,
+                                     width, height);
+            src = td->edge_emu_buffer + mx_idx + EDGE_EMU_LINESIZE * my_idx;
+            src_linesize = EDGE_EMU_LINESIZE;
         }
-        mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);
-    } else
-        mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);
+        mc_func[my_idx][mx_idx](dst, linesize, src, src_linesize, block_h, mx, my);
+    } 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, uint8_t *src1,
-                   uint8_t *src2, const VP56mv *mv, int x_off, int y_off,
-                   int block_w, int block_h, int width, int height, int linesize,
+void vp8_mc_chroma(VP8Context *s, VP8ThreadData *td, uint8_t *dst1,
+                   uint8_t *dst2, ThreadFrame *ref, const VP56mv *mv,
+                   int x_off, int y_off, int block_w, int block_h,
+                   int width, int height, ptrdiff_t linesize,
                    vp8_mc_func mc_func[3][3])
 {
+    uint8_t *src1 = ref->f->data[1], *src2 = ref->f->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];
+        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;
@@ -1258,74 +1755,90 @@ void vp8_mc_chroma(VP8Context *s, uint8_t *dst1, uint8_t *dst2, uint8_t *src1,
         // 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);
+            s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
+                                     src1 - my_idx * linesize - mx_idx,
+                                     EDGE_EMU_LINESIZE, linesize,
+                                     block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],
+                                     x_off - mx_idx, y_off - my_idx, width, height);
+            src1 = td->edge_emu_buffer + mx_idx + EDGE_EMU_LINESIZE * my_idx;
+            mc_func[my_idx][mx_idx](dst1, linesize, src1, EDGE_EMU_LINESIZE, block_h, mx, my);
+
+            s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
+                                     src2 - my_idx * linesize - mx_idx,
+                                     EDGE_EMU_LINESIZE, linesize,
+                                     block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],
+                                     x_off - mx_idx, y_off - my_idx, width, height);
+            src2 = td->edge_emu_buffer + mx_idx + EDGE_EMU_LINESIZE * my_idx;
+            mc_func[my_idx][mx_idx](dst2, linesize, src2, EDGE_EMU_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
-void vp8_mc_part(VP8Context *s, uint8_t *dst[3],
-                 AVFrame *ref_frame, int x_off, int y_off,
-                 int bx_off, int by_off,
-                 int block_w, int block_h,
+void vp8_mc_part(VP8Context *s, VP8ThreadData *td, uint8_t *dst[3],
+                 ThreadFrame *ref_frame, int x_off, int y_off,
+                 int bx_off, int by_off, int block_w, int block_h,
                  int width, int height, VP56mv *mv)
 {
     VP56mv uvmv = *mv;
 
     /* Y */
-    vp8_mc_luma(s, dst[0] + by_off * s->linesize + bx_off,
-                ref_frame->data[0], mv, x_off + bx_off, y_off + by_off,
+    vp8_mc_luma(s, td, 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) {
+        /* this block only applies VP8; it is safe to check
+         * only the profile, as VP7 profile <= 1 */
         uvmv.x &= ~7;
         uvmv.y &= ~7;
     }
-    x_off   >>= 1; y_off   >>= 1;
-    bx_off  >>= 1; by_off  >>= 1;
-    width   >>= 1; height  >>= 1;
-    block_w >>= 1; block_h >>= 1;
-    vp8_mc_chroma(s, dst[1] + by_off * s->uvlinesize + bx_off,
-                  dst[2] + by_off * s->uvlinesize + bx_off, ref_frame->data[1],
-                  ref_frame->data[2], &uvmv, x_off + bx_off, y_off + by_off,
+    x_off   >>= 1;
+    y_off   >>= 1;
+    bx_off  >>= 1;
+    by_off  >>= 1;
+    width   >>= 1;
+    height  >>= 1;
+    block_w >>= 1;
+    block_h >>= 1;
+    vp8_mc_chroma(s, td, 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.
- * Optimized for 64-byte cache lines.  Inspired by ffh264 prefetch_motion. */
-static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref)
+ * Optimized for 64-byte cache lines. Inspired by ffh264 prefetch_motion. */
+static av_always_inline
+void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y,
+                     int mb_xy, int ref)
 {
     /* Don't prefetch refs that haven't been used very often this frame. */
-    if (s->ref_count[ref-1] > (mb_xy >> 5)) {
+    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>>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;
-        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);
+        int mx = (mb->mv.x >> 2) + x_off + 8;
+        int my = (mb->mv.y >> 2) + y_off;
+        uint8_t **src = s->framep[ref]->tf.f->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->vdsp.prefetch(src[0] + off, s->linesize, 4);
+        off = (mx >> 1) + ((my >> 1) + (mb_x & 7)) * s->uvlinesize + 64;
+        s->vdsp.prefetch(src[1] + off, src[2] - src[1], 2);
     }
 }
 
@@ -1333,17 +1846,17 @@ static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, i
  * Apply motion vectors to prediction buffer, chapter 18.
  */
 static av_always_inline
-void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
-                   int mb_x, int mb_y)
+void inter_predict(VP8Context *s, VP8ThreadData *td, uint8_t *dst[3],
+                   VP8Macroblock *mb, int mb_x, int mb_y)
 {
     int x_off = mb_x << 4, y_off = mb_y << 4;
-    int width = 16*s->mb_width, height = 16*s->mb_height;
-    AVFrame *ref = s->framep[mb->ref_frame];
+    int width = 16 * s->mb_width, height = 16 * s->mb_height;
+    ThreadFrame *ref = &s->framep[mb->ref_frame]->tf;
     VP56mv *bmv = mb->bmv;
 
     switch (mb->partitioning) {
     case VP8_SPLITMVMODE_NONE:
-        vp8_mc_part(s, dst, ref, x_off, y_off,
+        vp8_mc_part(s, td, dst, ref, x_off, y_off,
                     0, 0, 16, 16, width, height, &mb->mv);
         break;
     case VP8_SPLITMVMODE_4x4: {
@@ -1353,36 +1866,38 @@ 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_luma(s, 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,
+                vp8_mc_luma(s, td, 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]);
             }
         }
 
         /* U/V */
-        x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;
+        x_off  >>= 1;
+        y_off  >>= 1;
+        width  >>= 1;
+        height >>= 1;
         for (y = 0; y < 2; y++) {
             for (x = 0; x < 2; x++) {
-                uvmv.x = mb->bmv[ 2*y    * 4 + 2*x  ].x +
-                         mb->bmv[ 2*y    * 4 + 2*x+1].x +
-                         mb->bmv[(2*y+1) * 4 + 2*x  ].x +
-                         mb->bmv[(2*y+1) * 4 + 2*x+1].x;
-                uvmv.y = mb->bmv[ 2*y    * 4 + 2*x  ].y +
-                         mb->bmv[ 2*y    * 4 + 2*x+1].y +
-                         mb->bmv[(2*y+1) * 4 + 2*x  ].y +
-                         mb->bmv[(2*y+1) * 4 + 2*x+1].y;
-                uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;
-                uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;
+                uvmv.x = mb->bmv[2 * y       * 4 + 2 * x    ].x +
+                         mb->bmv[2 * y       * 4 + 2 * x + 1].x +
+                         mb->bmv[(2 * y + 1) * 4 + 2 * x    ].x +
+                         mb->bmv[(2 * y + 1) * 4 + 2 * x + 1].x;
+                uvmv.y = mb->bmv[2 * y       * 4 + 2 * x    ].y +
+                         mb->bmv[2 * y       * 4 + 2 * x + 1].y +
+                         mb->bmv[(2 * y + 1) * 4 + 2 * x    ].y +
+                         mb->bmv[(2 * y + 1) * 4 + 2 * x + 1].y;
+                uvmv.x = (uvmv.x + 2 + FF_SIGNBIT(uvmv.x)) >> 2;
+                uvmv.y = (uvmv.y + 2 + FF_SIGNBIT(uvmv.y)) >> 2;
                 if (s->profile == 3) {
                     uvmv.x &= ~7;
                     uvmv.y &= ~7;
                 }
-                vp8_mc_chroma(s, dst[1] + 4*y*s->uvlinesize + x*4,
-                              dst[2] + 4*y*s->uvlinesize + x*4,
-                              ref->data[1], ref->data[2], &uvmv,
-                              4*x + x_off, 4*y + y_off, 4, 4,
+                vp8_mc_chroma(s, td, 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]);
             }
@@ -1390,87 +1905,100 @@ void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,
         break;
     }
     case VP8_SPLITMVMODE_16x8:
-        vp8_mc_part(s, dst, ref, x_off, y_off,
+        vp8_mc_part(s, td, dst, ref, x_off, y_off,
                     0, 0, 16, 8, width, height, &bmv[0]);
-        vp8_mc_part(s, dst, ref, x_off, y_off,
+        vp8_mc_part(s, td, dst, ref, x_off, y_off,
                     0, 8, 16, 8, width, height, &bmv[1]);
         break;
     case VP8_SPLITMVMODE_8x16:
-        vp8_mc_part(s, dst, ref, x_off, y_off,
+        vp8_mc_part(s, td, dst, ref, x_off, y_off,
                     0, 0, 8, 16, width, height, &bmv[0]);
-        vp8_mc_part(s, dst, ref, x_off, y_off,
+        vp8_mc_part(s, td, dst, ref, x_off, y_off,
                     8, 0, 8, 16, width, height, &bmv[1]);
         break;
     case VP8_SPLITMVMODE_8x8:
-        vp8_mc_part(s, dst, ref, x_off, y_off,
+        vp8_mc_part(s, td, dst, ref, x_off, y_off,
                     0, 0, 8, 8, width, height, &bmv[0]);
-        vp8_mc_part(s, dst, ref, x_off, y_off,
+        vp8_mc_part(s, td, dst, ref, x_off, y_off,
                     8, 0, 8, 8, width, height, &bmv[1]);
-        vp8_mc_part(s, dst, ref, x_off, y_off,
+        vp8_mc_part(s, td, dst, ref, x_off, y_off,
                     0, 8, 8, 8, width, height, &bmv[2]);
-        vp8_mc_part(s, dst, ref, x_off, y_off,
+        vp8_mc_part(s, td, dst, ref, x_off, y_off,
                     8, 8, 8, 8, width, height, &bmv[3]);
         break;
     }
 }
 
-static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb)
+static av_always_inline
+void idct_mb(VP8Context *s, VP8ThreadData *td, uint8_t *dst[3], VP8Macroblock *mb)
 {
     int x, y, ch;
 
     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(td->non_zero_count_cache[y]);
             if (nnz4) {
-                if (nnz4&~0x01010101) {
+                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,
+                                                      td->block[y][x],
+                                                      s->linesize);
+                        else if ((uint8_t) nnz4 > 1)
+                            s->vp8dsp.vp8_idct_add(y_dst + 4 * x,
+                                                   td->block[y][x],
+                                                   s->linesize);
+                        nnz4 >>= 8;
+                        if (!nnz4)
+                            break;
                     }
                 } else {
-                    s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize);
+                    s->vp8dsp.vp8_idct_dc_add4y(y_dst, td->block[y], s->linesize);
                 }
             }
-            y_dst += 4*s->linesize;
+            y_dst += 4 * s->linesize;
         }
     }
 
     for (ch = 0; ch < 2; ch++) {
-        uint32_t nnz4 = AV_RN32A(s->non_zero_count_cache[4+ch]);
+        uint32_t nnz4 = AV_RL32(td->non_zero_count_cache[4 + ch]);
         if (nnz4) {
-            uint8_t *ch_dst = dst[1+ch];
-            if (nnz4&~0x01010101) {
+            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,
+                                                      td->block[4 + ch][(y << 1) + x],
+                                                      s->uvlinesize);
+                        else if ((uint8_t) nnz4 > 1)
+                            s->vp8dsp.vp8_idct_add(ch_dst + 4 * x,
+                                                   td->block[4 + ch][(y << 1) + x],
+                                                   s->uvlinesize);
+                        nnz4 >>= 8;
+                        if (!nnz4)
+                            goto chroma_idct_end;
                     }
-                    ch_dst += 4*s->uvlinesize;
+                    ch_dst += 4 * s->uvlinesize;
                 }
             } else {
-                s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize);
+                s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, td->block[4 + ch], s->uvlinesize);
             }
         }
+chroma_idct_end:
+        ;
     }
 }
 
-static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )
+static av_always_inline
+void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb,
+                         VP8FilterStrength *f, int is_vp7)
 {
     int interior_limit, filter_level;
 
     if (s->segmentation.enabled) {
-        filter_level = s->segmentation.filter_level[s->segment];
+        filter_level = s->segmentation.filter_level[mb->segment];
         if (!s->segmentation.absolute_vals)
             filter_level += s->filter.level;
     } else
@@ -1481,25 +2009,26 @@ static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *m
         filter_level += s->lf_delta.mode[mb->mode];
     }
 
-/* Like av_clip for inputs 0 and max, where max is equal to (2^n-1) */
-#define POW2CLIP(x,max) (((x) & ~max) ? (-(x))>>31 & max : (x));
-    filter_level = POW2CLIP(filter_level, 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);
 
     f->filter_level = filter_level;
     f->inner_limit = interior_limit;
-    f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;
+    f->inner_filter = is_vp7 || !mb->skip || mb->mode == MODE_I4x4 ||
+                      mb->mode == VP8_MVMODE_SPLIT;
 }
 
-static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)
+static av_always_inline
+void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f,
+               int mb_x, int mb_y, int is_vp7)
 {
-    int mbedge_lim, bedge_lim, hev_thresh;
+    int mbedge_lim, bedge_lim_y, bedge_lim_uv, hev_thresh;
     int filter_level = f->filter_level;
     int inner_limit = f->inner_limit;
     int inner_filter = f->inner_filter;
@@ -1519,259 +2048,551 @@ static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8Filter
     if (!filter_level)
         return;
 
-     bedge_lim = 2*filter_level + inner_limit;
-    mbedge_lim = bedge_lim + 4;
+    if (is_vp7) {
+        bedge_lim_y  = filter_level;
+        bedge_lim_uv = filter_level * 2;
+        mbedge_lim   = filter_level + 2;
+    } else {
+        bedge_lim_y  =
+        bedge_lim_uv = filter_level * 2 + inner_limit;
+        mbedge_lim   = bedge_lim_y + 4;
+    }
 
     hev_thresh = hev_thresh_lut[s->keyframe][filter_level];
 
     if (mb_x) {
-        s->vp8dsp.vp8_h_loop_filter16y(dst[0],     linesize,
+        s->vp8dsp.vp8_h_loop_filter16y(dst[0], linesize,
                                        mbedge_lim, inner_limit, hev_thresh);
-        s->vp8dsp.vp8_h_loop_filter8uv(dst[1],     dst[2],      uvlinesize,
+        s->vp8dsp.vp8_h_loop_filter8uv(dst[1], dst[2], uvlinesize,
                                        mbedge_lim, inner_limit, hev_thresh);
     }
 
-    if (inner_filter) {
-        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, linesize, bedge_lim,
-                                             inner_limit, hev_thresh);
-        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, linesize, bedge_lim,
-                                             inner_limit, hev_thresh);
-        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, linesize, bedge_lim,
-                                             inner_limit, hev_thresh);
-        s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4,
-                                             uvlinesize,  bedge_lim,
-                                             inner_limit, hev_thresh);
+#define H_LOOP_FILTER_16Y_INNER(cond)                                         \
+    if (cond && inner_filter) {                                               \
+        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0] +  4, linesize,           \
+                                             bedge_lim_y, inner_limit,        \
+                                             hev_thresh);                     \
+        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0] +  8, linesize,           \
+                                             bedge_lim_y, inner_limit,        \
+                                             hev_thresh);                     \
+        s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0] + 12, linesize,           \
+                                             bedge_lim_y, inner_limit,        \
+                                             hev_thresh);                     \
+        s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] +  4, dst[2] + 4,         \
+                                             uvlinesize,  bedge_lim_uv,       \
+                                             inner_limit, hev_thresh);        \
     }
 
+    H_LOOP_FILTER_16Y_INNER(!is_vp7)
+
     if (mb_y) {
-        s->vp8dsp.vp8_v_loop_filter16y(dst[0],     linesize,
+        s->vp8dsp.vp8_v_loop_filter16y(dst[0], linesize,
                                        mbedge_lim, inner_limit, hev_thresh);
-        s->vp8dsp.vp8_v_loop_filter8uv(dst[1],     dst[2],      uvlinesize,
+        s->vp8dsp.vp8_v_loop_filter8uv(dst[1], dst[2], uvlinesize,
                                        mbedge_lim, inner_limit, hev_thresh);
     }
 
     if (inner_filter) {
-        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*linesize,
-                                             linesize,    bedge_lim,
+        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0] +  4 * linesize,
+                                             linesize, bedge_lim_y,
                                              inner_limit, hev_thresh);
-        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*linesize,
-                                             linesize,    bedge_lim,
+        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0] +  8 * linesize,
+                                             linesize, bedge_lim_y,
                                              inner_limit, hev_thresh);
-        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*linesize,
-                                             linesize,    bedge_lim,
+        s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0] + 12 * linesize,
+                                             linesize, bedge_lim_y,
                                              inner_limit, hev_thresh);
-        s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * uvlinesize,
-                                             dst[2] + 4 * uvlinesize,
-                                             uvlinesize,  bedge_lim,
+        s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] +  4 * uvlinesize,
+                                             dst[2] +  4 * uvlinesize,
+                                             uvlinesize, bedge_lim_uv,
                                              inner_limit, hev_thresh);
     }
+
+    H_LOOP_FILTER_16Y_INNER(is_vp7)
 }
 
-static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)
+static av_always_inline
+void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f,
+                      int mb_x, int mb_y)
 {
     int mbedge_lim, bedge_lim;
     int filter_level = f->filter_level;
-    int inner_limit = f->inner_limit;
+    int inner_limit  = f->inner_limit;
     int inner_filter = f->inner_filter;
-    int linesize = s->linesize;
+    int linesize     = s->linesize;
 
     if (!filter_level)
         return;
 
-     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);
     if (inner_filter) {
-        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, linesize, bedge_lim);
-        s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, linesize, bedge_lim);
-        s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim);
+        s->vp8dsp.vp8_h_loop_filter_simple(dst +  4, linesize, bedge_lim);
+        s->vp8dsp.vp8_h_loop_filter_simple(dst +  8, linesize, bedge_lim);
+        s->vp8dsp.vp8_h_loop_filter_simple(dst + 12, linesize, bedge_lim);
     }
 
     if (mb_y)
         s->vp8dsp.vp8_v_loop_filter_simple(dst, linesize, mbedge_lim);
     if (inner_filter) {
-        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*linesize, linesize, bedge_lim);
-        s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*linesize, linesize, bedge_lim);
-        s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim);
+        s->vp8dsp.vp8_v_loop_filter_simple(dst +  4 * linesize, linesize, bedge_lim);
+        s->vp8dsp.vp8_v_loop_filter_simple(dst +  8 * linesize, linesize, bedge_lim);
+        s->vp8dsp.vp8_v_loop_filter_simple(dst + 12 * linesize, linesize, bedge_lim);
+    }
+}
+
+#define MARGIN (16 << 2)
+static av_always_inline
+void vp78_decode_mv_mb_modes(AVCodecContext *avctx, VP8Frame *curframe,
+                                    VP8Frame *prev_frame, int is_vp7)
+{
+    VP8Context *s = avctx->priv_data;
+    int mb_x, mb_y;
+
+    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++) {
+        VP8Macroblock *mb = s->macroblocks_base +
+                            ((s->mb_width + 1) * (mb_y + 1) + 1);
+        int mb_xy = mb_y * s->mb_width;
+
+        AV_WN32A(s->intra4x4_pred_mode_left, DC_PRED * 0x01010101);
+
+        s->mv_min.x = -MARGIN;
+        s->mv_max.x = ((s->mb_width - 1) << 6) + MARGIN;
+        for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {
+            if (mb_y == 0)
+                AV_WN32A((mb - s->mb_width - 1)->intra4x4_pred_mode_top,
+                         DC_PRED * 0x01010101);
+            decode_mb_mode(s, mb, mb_x, mb_y, curframe->seg_map->data + mb_xy,
+                           prev_frame && prev_frame->seg_map ?
+                           prev_frame->seg_map->data + mb_xy : NULL, 1, is_vp7);
+            s->mv_min.x -= 64;
+            s->mv_max.x -= 64;
+        }
+        s->mv_min.y -= 64;
+        s->mv_max.y -= 64;
+    }
+}
+
+static void vp7_decode_mv_mb_modes(AVCodecContext *avctx, VP8Frame *cur_frame,
+                                   VP8Frame *prev_frame)
+{
+    vp78_decode_mv_mb_modes(avctx, cur_frame, prev_frame, IS_VP7);
+}
+
+static void vp8_decode_mv_mb_modes(AVCodecContext *avctx, VP8Frame *cur_frame,
+                                   VP8Frame *prev_frame)
+{
+    vp78_decode_mv_mb_modes(avctx, cur_frame, prev_frame, IS_VP8);
+}
+
+#if HAVE_THREADS
+#define check_thread_pos(td, otd, mb_x_check, mb_y_check)                     \
+    do {                                                                      \
+        int tmp = (mb_y_check << 16) | (mb_x_check & 0xFFFF);                 \
+        if (otd->thread_mb_pos < tmp) {                                       \
+            pthread_mutex_lock(&otd->lock);                                   \
+            td->wait_mb_pos = tmp;                                            \
+            do {                                                              \
+                if (otd->thread_mb_pos >= tmp)                                \
+                    break;                                                    \
+                pthread_cond_wait(&otd->cond, &otd->lock);                    \
+            } while (1);                                                      \
+            td->wait_mb_pos = INT_MAX;                                        \
+            pthread_mutex_unlock(&otd->lock);                                 \
+        }                                                                     \
+    } while (0);
+
+#define update_pos(td, mb_y, mb_x)                                            \
+    do {                                                                      \
+        int pos              = (mb_y << 16) | (mb_x & 0xFFFF);                \
+        int sliced_threading = (avctx->active_thread_type == FF_THREAD_SLICE) && \
+                               (num_jobs > 1);                                \
+        int is_null          = !next_td || !prev_td;                          \
+        int pos_check        = (is_null) ? 1                                  \
+                                         : (next_td != td &&                  \
+                                            pos >= next_td->wait_mb_pos) ||   \
+                                           (prev_td != td &&                  \
+                                            pos >= prev_td->wait_mb_pos);     \
+        td->thread_mb_pos = pos;                                              \
+        if (sliced_threading && pos_check) {                                  \
+            pthread_mutex_lock(&td->lock);                                    \
+            pthread_cond_broadcast(&td->cond);                                \
+            pthread_mutex_unlock(&td->lock);                                  \
+        }                                                                     \
+    } while (0);
+#else
+#define check_thread_pos(td, otd, mb_x_check, mb_y_check)
+#define update_pos(td, mb_y, mb_x)
+#endif
+
+static void vp8_decode_mb_row_no_filter(AVCodecContext *avctx, void *tdata,
+                                        int jobnr, int threadnr, int is_vp7)
+{
+    VP8Context *s = avctx->priv_data;
+    VP8ThreadData *prev_td, *next_td, *td = &s->thread_data[threadnr];
+    int mb_y = td->thread_mb_pos >> 16;
+    int mb_x, mb_xy = mb_y * s->mb_width;
+    int num_jobs = s->num_jobs;
+    VP8Frame *curframe = s->curframe, *prev_frame = s->prev_frame;
+    VP56RangeCoder *c  = &s->coeff_partition[mb_y & (s->num_coeff_partitions - 1)];
+    VP8Macroblock *mb;
+    uint8_t *dst[3] = {
+        curframe->tf.f->data[0] + 16 * mb_y * s->linesize,
+        curframe->tf.f->data[1] +  8 * mb_y * s->uvlinesize,
+        curframe->tf.f->data[2] +  8 * mb_y * s->uvlinesize
+    };
+    if (mb_y == 0)
+        prev_td = td;
+    else
+        prev_td = &s->thread_data[(jobnr + num_jobs - 1) % num_jobs];
+    if (mb_y == s->mb_height - 1)
+        next_td = td;
+    else
+        next_td = &s->thread_data[(jobnr + 1) % num_jobs];
+    if (s->mb_layout == 1)
+        mb = s->macroblocks_base + ((s->mb_width + 1) * (mb_y + 1) + 1);
+    else {
+        // Make sure the previous frame has read its segmentation map,
+        // if we re-use the same map.
+        if (prev_frame && s->segmentation.enabled &&
+            !s->segmentation.update_map)
+            ff_thread_await_progress(&prev_frame->tf, mb_y, 0);
+        mb = s->macroblocks + (s->mb_height - mb_y - 1) * 2;
+        memset(mb - 1, 0, sizeof(*mb)); // zero left macroblock
+        AV_WN32A(s->intra4x4_pred_mode_left, DC_PRED * 0x01010101);
+    }
+
+    if (!is_vp7 || mb_y == 0)
+        memset(td->left_nnz, 0, sizeof(td->left_nnz));
+
+    s->mv_min.x = -MARGIN;
+    s->mv_max.x = ((s->mb_width - 1) << 6) + MARGIN;
+
+    for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {
+        // Wait for previous thread to read mb_x+2, and reach mb_y-1.
+        if (prev_td != td) {
+            if (threadnr != 0) {
+                check_thread_pos(td, prev_td,
+                                 mb_x + (is_vp7 ? 2 : 1),
+                                 mb_y - (is_vp7 ? 2 : 1));
+            } else {
+                check_thread_pos(td, prev_td,
+                                 mb_x + (is_vp7 ? 2 : 1) + s->mb_width + 3,
+                                 mb_y - (is_vp7 ? 2 : 1));
+            }
+        }
+
+        s->vdsp.prefetch(dst[0] + (mb_x & 3) * 4 * s->linesize + 64,
+                         s->linesize, 4);
+        s->vdsp.prefetch(dst[1] + (mb_x & 7) * s->uvlinesize + 64,
+                         dst[2] - dst[1], 2);
+
+        if (!s->mb_layout)
+            decode_mb_mode(s, mb, mb_x, mb_y, curframe->seg_map->data + mb_xy,
+                           prev_frame && prev_frame->seg_map ?
+                           prev_frame->seg_map->data + mb_xy : NULL, 0, is_vp7);
+
+        prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS);
+
+        if (!mb->skip)
+            decode_mb_coeffs(s, td, c, mb, s->top_nnz[mb_x], td->left_nnz, is_vp7);
+
+        if (mb->mode <= MODE_I4x4)
+            intra_predict(s, td, dst, mb, mb_x, mb_y, is_vp7);
+        else
+            inter_predict(s, td, dst, mb, mb_x, mb_y);
+
+        prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN);
+
+        if (!mb->skip) {
+            idct_mb(s, td, dst, mb);
+        } else {
+            AV_ZERO64(td->left_nnz);
+            AV_WN64(s->top_nnz[mb_x], 0);   // array of 9, so unaligned
+
+            /* Reset DC block predictors if they would exist
+             * if the mb had coefficients */
+            if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
+                td->left_nnz[8]     = 0;
+                s->top_nnz[mb_x][8] = 0;
+            }
+        }
+
+        if (s->deblock_filter)
+            filter_level_for_mb(s, mb, &td->filter_strength[mb_x], is_vp7);
+
+        if (s->deblock_filter && num_jobs != 1 && threadnr == num_jobs - 1) {
+            if (s->filter.simple)
+                backup_mb_border(s->top_border[mb_x + 1], dst[0],
+                                 NULL, NULL, s->linesize, 0, 1);
+            else
+                backup_mb_border(s->top_border[mb_x + 1], dst[0],
+                                 dst[1], dst[2], s->linesize, s->uvlinesize, 0);
+        }
+
+        prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2);
+
+        dst[0]      += 16;
+        dst[1]      += 8;
+        dst[2]      += 8;
+        s->mv_min.x -= 64;
+        s->mv_max.x -= 64;
+
+        if (mb_x == s->mb_width + 1) {
+            update_pos(td, mb_y, s->mb_width + 3);
+        } else {
+            update_pos(td, mb_y, mb_x);
+        }
     }
 }
 
-static void filter_mb_row(VP8Context *s, int mb_y)
+static void vp8_filter_mb_row(AVCodecContext *avctx, void *tdata,
+                              int jobnr, int threadnr, int is_vp7)
 {
-    VP8FilterStrength *f = s->filter_strength;
+    VP8Context *s = avctx->priv_data;
+    VP8ThreadData *td = &s->thread_data[threadnr];
+    int mb_x, mb_y = td->thread_mb_pos >> 16, num_jobs = s->num_jobs;
+    AVFrame *curframe = s->curframe->tf.f;
+    VP8Macroblock *mb;
+    VP8ThreadData *prev_td, *next_td;
     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;
 
-    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
-        backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);
-        filter_mb(s, dst, f++, mb_x, mb_y);
+    if (s->mb_layout == 1)
+        mb = s->macroblocks_base + ((s->mb_width + 1) * (mb_y + 1) + 1);
+    else
+        mb = s->macroblocks + (s->mb_height - mb_y - 1) * 2;
+
+    if (mb_y == 0)
+        prev_td = td;
+    else
+        prev_td = &s->thread_data[(jobnr + num_jobs - 1) % num_jobs];
+    if (mb_y == s->mb_height - 1)
+        next_td = td;
+    else
+        next_td = &s->thread_data[(jobnr + 1) % num_jobs];
+
+    for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb++) {
+        VP8FilterStrength *f = &td->filter_strength[mb_x];
+        if (prev_td != td)
+            check_thread_pos(td, prev_td,
+                             (mb_x + 1) + (s->mb_width + 3), mb_y - 1);
+        if (next_td != td)
+            if (next_td != &s->thread_data[0])
+                check_thread_pos(td, next_td, mb_x + 1, mb_y + 1);
+
+        if (num_jobs == 1) {
+            if (s->filter.simple)
+                backup_mb_border(s->top_border[mb_x + 1], dst[0],
+                                 NULL, NULL, s->linesize, 0, 1);
+            else
+                backup_mb_border(s->top_border[mb_x + 1], dst[0],
+                                 dst[1], dst[2], s->linesize, s->uvlinesize, 0);
+        }
+
+        if (s->filter.simple)
+            filter_mb_simple(s, dst[0], f, mb_x, mb_y);
+        else
+            filter_mb(s, dst, f, mb_x, mb_y, is_vp7);
         dst[0] += 16;
         dst[1] += 8;
         dst[2] += 8;
+
+        update_pos(td, mb_y, (s->mb_width + 3) + mb_x);
     }
 }
 
-static void filter_mb_row_simple(VP8Context *s, int mb_y)
+static av_always_inline
+int vp78_decode_mb_row_sliced(AVCodecContext *avctx, void *tdata, int jobnr,
+                              int threadnr, int is_vp7)
 {
-    VP8FilterStrength *f = s->filter_strength;
-    uint8_t *dst = s->framep[VP56_FRAME_CURRENT]->data[0] + 16*mb_y*s->linesize;
-    int mb_x;
+    VP8Context *s = avctx->priv_data;
+    VP8ThreadData *td = &s->thread_data[jobnr];
+    VP8ThreadData *next_td = NULL, *prev_td = NULL;
+    VP8Frame *curframe = s->curframe;
+    int mb_y, num_jobs = s->num_jobs;
+
+    td->thread_nr = threadnr;
+    for (mb_y = jobnr; mb_y < s->mb_height; mb_y += num_jobs) {
+        if (mb_y >= s->mb_height)
+            break;
+        td->thread_mb_pos = mb_y << 16;
+        vp8_decode_mb_row_no_filter(avctx, tdata, jobnr, threadnr, is_vp7);
+        if (s->deblock_filter)
+            vp8_filter_mb_row(avctx, tdata, jobnr, threadnr, is_vp7);
+        update_pos(td, mb_y, INT_MAX & 0xFFFF);
 
-    for (mb_x = 0; mb_x < s->mb_width; mb_x++) {
-        backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1);
-        filter_mb_simple(s, dst, f++, mb_x, mb_y);
-        dst += 16;
+        s->mv_min.y -= 64;
+        s->mv_max.y -= 64;
+
+        if (avctx->active_thread_type == FF_THREAD_FRAME)
+            ff_thread_report_progress(&curframe->tf, mb_y, 0);
     }
+
+    return 0;
 }
 
-static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
-                            AVPacket *avpkt)
+static int vp7_decode_mb_row_sliced(AVCodecContext *avctx, void *tdata,
+                                    int jobnr, int threadnr)
+{
+    return vp78_decode_mb_row_sliced(avctx, tdata, jobnr, threadnr, IS_VP7);
+}
+
+static int vp8_decode_mb_row_sliced(AVCodecContext *avctx, void *tdata,
+                                    int jobnr, int threadnr)
+{
+    return vp78_decode_mb_row_sliced(avctx, tdata, jobnr, threadnr, IS_VP8);
+}
+
+
+static av_always_inline
+int vp78_decode_frame(AVCodecContext *avctx, void *data, int *got_frame,
+                      AVPacket *avpkt, int is_vp7)
 {
     VP8Context *s = avctx->priv_data;
-    int ret, mb_x, mb_y, i, y, referenced;
+    int ret, i, referenced, num_jobs;
     enum AVDiscard skip_thresh;
-    AVFrame *av_uninit(curframe);
+    VP8Frame *av_uninit(curframe), *prev_frame;
 
-    if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)
-        return ret;
+    if (is_vp7)
+        ret = vp7_decode_frame_header(s, avpkt->data, avpkt->size);
+    else
+        ret = vp8_decode_frame_header(s, avpkt->data, avpkt->size);
 
-    referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT
-                                || s->update_altref == VP56_FRAME_CURRENT;
+    if (ret < 0)
+        goto err;
 
-    skip_thresh = !referenced ? AVDISCARD_NONREF :
-                    !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;
+    prev_frame = s->framep[VP56_FRAME_CURRENT];
+
+    referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT ||
+                 s->update_altref == VP56_FRAME_CURRENT;
+
+    skip_thresh = !referenced ? AVDISCARD_NONREF
+                              : !s->keyframe ? AVDISCARD_NONKEY
+                                             : AVDISCARD_ALL;
 
     if (avctx->skip_frame >= skip_thresh) {
         s->invisible = 1;
+        memcpy(&s->next_framep[0], &s->framep[0], sizeof(s->framep[0]) * 4);
         goto skip_decode;
     }
     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] &&
-            &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 (curframe->data[0])
-        avctx->release_buffer(avctx, curframe);
+    // release no longer referenced frames
+    for (i = 0; i < 5; i++)
+        if (s->frames[i].tf.f->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]);
 
-    curframe->key_frame = s->keyframe;
-    curframe->pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
-    curframe->reference = referenced ? 3 : 0;
-    if ((ret = avctx->get_buffer(avctx, curframe))) {
-        av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
-        return ret;
-    }
+    curframe = s->framep[VP56_FRAME_CURRENT] = vp8_find_free_buffer(s);
 
-    // 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.
+    if (!s->colorspace)
+        avctx->colorspace = AVCOL_SPC_BT470BG;
+    if (s->fullrange)
+        avctx->color_range = AVCOL_RANGE_JPEG;
+    else
+        avctx->color_range = AVCOL_RANGE_MPEG;
+
+    /* 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. */
     if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||
-                         !s->framep[VP56_FRAME_GOLDEN] ||
+                         !s->framep[VP56_FRAME_GOLDEN]   ||
                          !s->framep[VP56_FRAME_GOLDEN2])) {
-        av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");
-        return AVERROR_INVALIDDATA;
+        av_log(avctx, AV_LOG_WARNING,
+               "Discarding interframe without a prior keyframe!\n");
+        ret = AVERROR_INVALIDDATA;
+        goto err;
     }
 
-    s->linesize   = curframe->linesize[0];
-    s->uvlinesize = curframe->linesize[1];
-
-    if (!s->edge_emu_buffer)
-        s->edge_emu_buffer = av_malloc(21*s->linesize);
-
-    memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));
-
-    /* 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
-    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);
+    curframe->tf.f->key_frame = s->keyframe;
+    curframe->tf.f->pict_type = s->keyframe ? AV_PICTURE_TYPE_I
+                                            : AV_PICTURE_TYPE_P;
+    if ((ret = vp8_alloc_frame(s, curframe, referenced))) {
+        av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");
+        goto err;
     }
-    memset(s->ref_count, 0, sizeof(s->ref_count));
-    if (s->keyframe)
-        memset(s->intra4x4_pred_mode_top, DC_PRED, s->mb_width*4);
-
-    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;
-        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)) {
-            for (i = 0; i < 3; i++)
-                for (y = 0; y < 16>>!!i; y++)
-                    dst[i][y*curframe->linesize[i]-1] = 129;
-            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++) {
-            /* 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, s->segmentation_map + mb_xy);
+    // 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];
 
-            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS);
+    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 (!mb->skip)
-                decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);
+    if (s->update_last)
+        s->next_framep[VP56_FRAME_PREVIOUS] = curframe;
+    else
+        s->next_framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_PREVIOUS];
 
-            if (mb->mode <= MODE_I4x4)
-                intra_predict(s, dst, mb, mb_x, mb_y);
-            else
-                inter_predict(s, dst, mb, mb_x, mb_y);
+    s->next_framep[VP56_FRAME_CURRENT] = curframe;
 
-            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN);
+    ff_thread_finish_setup(avctx);
 
-            if (!mb->skip) {
-                idct_mb(s, dst, mb);
-            } else {
-                AV_ZERO64(s->left_nnz);
-                AV_WN64(s->top_nnz[mb_x], 0);   // array of 9, so unaligned
+    s->linesize   = curframe->tf.f->linesize[0];
+    s->uvlinesize = curframe->tf.f->linesize[1];
 
-                // Reset DC block predictors if they would exist if the mb had coefficients
-                if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {
-                    s->left_nnz[8]      = 0;
-                    s->top_nnz[mb_x][8] = 0;
-                }
-            }
+    memset(s->top_nnz, 0, s->mb_width * sizeof(*s->top_nnz));
+    /* Zero macroblock structures for top/top-left prediction
+     * from outside the frame. */
+    if (!s->mb_layout)
+        memset(s->macroblocks + s->mb_height * 2 - 1, 0,
+               (s->mb_width + 1) * sizeof(*s->macroblocks));
+    if (!s->mb_layout && s->keyframe)
+        memset(s->intra4x4_pred_mode_top, DC_PRED, s->mb_width * 4);
 
-            if (s->deblock_filter)
-                filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);
+    memset(s->ref_count, 0, sizeof(s->ref_count));
 
-            prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2);
+    if (s->mb_layout == 1) {
+        // Make sure the previous frame has read its segmentation map,
+        // if we re-use the same map.
+        if (prev_frame && s->segmentation.enabled &&
+            !s->segmentation.update_map)
+            ff_thread_await_progress(&prev_frame->tf, 1, 0);
+        if (is_vp7)
+            vp7_decode_mv_mb_modes(avctx, curframe, prev_frame);
+        else
+            vp8_decode_mv_mb_modes(avctx, curframe, prev_frame);
+    }
 
-            dst[0] += 16;
-            dst[1] += 8;
-            dst[2] += 8;
-        }
-        if (s->deblock_filter) {
-            if (s->filter.simple)
-                filter_mb_row_simple(s, mb_y);
-            else
-                filter_mb_row(s, mb_y);
-        }
+    if (avctx->active_thread_type == FF_THREAD_FRAME)
+        num_jobs = 1;
+    else
+        num_jobs = FFMIN(s->num_coeff_partitions, avctx->thread_count);
+    s->num_jobs   = num_jobs;
+    s->curframe   = curframe;
+    s->prev_frame = prev_frame;
+    s->mv_min.y   = -MARGIN;
+    s->mv_max.y   = ((s->mb_height - 1) << 6) + MARGIN;
+    for (i = 0; i < MAX_THREADS; i++) {
+        s->thread_data[i].thread_mb_pos = 0;
+        s->thread_data[i].wait_mb_pos   = INT_MAX;
     }
+    if (is_vp7)
+        avctx->execute2(avctx, vp7_decode_mb_row_sliced, s->thread_data, NULL,
+                        num_jobs);
+    else
+        avctx->execute2(avctx, vp8_decode_mb_row_sliced, s->thread_data, NULL,
+                        num_jobs);
+
+    ff_thread_report_progress(&curframe->tf, INT_MAX, 0);
+    memcpy(&s->framep[0], &s->next_framep[0], sizeof(s->framep[0]) * 4);
 
 skip_decode:
     // if future frames don't use the updated probabilities,
@@ -1779,68 +2600,180 @@ skip_decode:
     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->invisible) {
+        if ((ret = av_frame_ref(data, curframe->tf.f)) < 0)
+            return ret;
+        *got_frame = 1;
+    }
+
+    return avpkt->size;
+err:
+    memcpy(&s->next_framep[0], &s->framep[0], sizeof(s->framep[0]) * 4);
+    return ret;
+}
+
+int ff_vp8_decode_frame(AVCodecContext *avctx, void *data, int *got_frame,
+                        AVPacket *avpkt)
+{
+    return vp78_decode_frame(avctx, data, got_frame, avpkt, IS_VP8);
+}
+
+#if CONFIG_VP7_DECODER
+static int vp7_decode_frame(AVCodecContext *avctx, void *data, int *got_frame,
+                            AVPacket *avpkt)
+{
+    return vp78_decode_frame(avctx, data, got_frame, avpkt, IS_VP7);
+}
+#endif /* CONFIG_VP7_DECODER */
+
+av_cold int ff_vp8_decode_free(AVCodecContext *avctx)
+{
+    VP8Context *s = avctx->priv_data;
+    int i;
+
+    vp8_decode_flush_impl(avctx, 1);
+    for (i = 0; i < FF_ARRAY_ELEMS(s->frames); i++)
+        av_frame_free(&s->frames[i].tf.f);
 
-        if (s->update_golden != VP56_FRAME_NONE)
-            s->framep[VP56_FRAME_GOLDEN] = s->framep[s->update_golden];
+    return 0;
+}
+
+static av_cold int vp8_init_frames(VP8Context *s)
+{
+    int i;
+    for (i = 0; i < FF_ARRAY_ELEMS(s->frames); i++) {
+        s->frames[i].tf.f = av_frame_alloc();
+        if (!s->frames[i].tf.f)
+            return AVERROR(ENOMEM);
     }
+    return 0;
+}
+
+static av_always_inline
+int vp78_decode_init(AVCodecContext *avctx, int is_vp7)
+{
+    VP8Context *s = avctx->priv_data;
+    int ret;
 
-    if (s->update_last) // move cur->prev
-        s->framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_CURRENT];
+    s->avctx = avctx;
+    avctx->pix_fmt = AV_PIX_FMT_YUV420P;
+    avctx->internal->allocate_progress = 1;
+
+    ff_videodsp_init(&s->vdsp, 8);
+
+    ff_vp78dsp_init(&s->vp8dsp);
+    if (CONFIG_VP7_DECODER && is_vp7) {
+        ff_h264_pred_init(&s->hpc, AV_CODEC_ID_VP7, 8, 1);
+        ff_vp7dsp_init(&s->vp8dsp);
+    } else if (CONFIG_VP8_DECODER && !is_vp7) {
+        ff_h264_pred_init(&s->hpc, AV_CODEC_ID_VP8, 8, 1);
+        ff_vp8dsp_init(&s->vp8dsp);
+    }
 
-    // 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]);
+    /* does not change for VP8 */
+    memcpy(s->prob[0].scan, zigzag_scan, sizeof(s->prob[0].scan));
 
-    if (!s->invisible) {
-        *(AVFrame*)data = *s->framep[VP56_FRAME_CURRENT];
-        *data_size = sizeof(AVFrame);
+    if ((ret = vp8_init_frames(s)) < 0) {
+        ff_vp8_decode_free(avctx);
+        return ret;
     }
 
-    return avpkt->size;
+    return 0;
+}
+
+#if CONFIG_VP7_DECODER
+static int vp7_decode_init(AVCodecContext *avctx)
+{
+    return vp78_decode_init(avctx, IS_VP7);
 }
+#endif /* CONFIG_VP7_DECODER */
 
-static av_cold int vp8_decode_init(AVCodecContext *avctx)
+av_cold int ff_vp8_decode_init(AVCodecContext *avctx)
+{
+    return vp78_decode_init(avctx, IS_VP8);
+}
+
+#if CONFIG_VP8_DECODER
+static av_cold int vp8_decode_init_thread_copy(AVCodecContext *avctx)
 {
     VP8Context *s = avctx->priv_data;
+    int ret;
 
     s->avctx = avctx;
-    avctx->pix_fmt = PIX_FMT_YUV420P;
 
-    dsputil_init(&s->dsp, avctx);
-    ff_h264_pred_init(&s->hpc, CODEC_ID_VP8);
-    ff_vp8dsp_init(&s->vp8dsp);
+    if ((ret = vp8_init_frames(s)) < 0) {
+        ff_vp8_decode_free(avctx);
+        return ret;
+    }
 
     return 0;
 }
 
-static av_cold int vp8_decode_free(AVCodecContext *avctx)
+#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)
 {
-    vp8_decode_flush(avctx);
+    VP8Context *s = dst->priv_data, *s_src = src->priv_data;
+    int i;
+
+    if (s->macroblocks_base &&
+        (s_src->mb_width != s->mb_width || s_src->mb_height != s->mb_height)) {
+        free_buffers(s);
+        s->mb_width  = s_src->mb_width;
+        s->mb_height = s_src->mb_height;
+    }
+
+    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));
+
+    for (i = 0; i < FF_ARRAY_ELEMS(s_src->frames); i++) {
+        if (s_src->frames[i].tf.f->data[0]) {
+            int ret = vp8_ref_frame(s, &s->frames[i], &s_src->frames[i]);
+            if (ret < 0)
+                return ret;
+        }
+    }
+
+    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;
 }
+#endif /* CONFIG_VP8_DECODER */
+
+#if CONFIG_VP7_DECODER
+AVCodec ff_vp7_decoder = {
+    .name                  = "vp7",
+    .long_name             = NULL_IF_CONFIG_SMALL("On2 VP7"),
+    .type                  = AVMEDIA_TYPE_VIDEO,
+    .id                    = AV_CODEC_ID_VP7,
+    .priv_data_size        = sizeof(VP8Context),
+    .init                  = vp7_decode_init,
+    .close                 = ff_vp8_decode_free,
+    .decode                = vp7_decode_frame,
+    .capabilities          = CODEC_CAP_DR1,
+    .flush                 = vp8_decode_flush,
+};
+#endif /* CONFIG_VP7_DECODER */
 
+#if CONFIG_VP8_DECODER
 AVCodec ff_vp8_decoder = {
-    "vp8",
-    AVMEDIA_TYPE_VIDEO,
-    CODEC_ID_VP8,
-    sizeof(VP8Context),
-    vp8_decode_init,
-    NULL,
-    vp8_decode_free,
-    vp8_decode_frame,
-    CODEC_CAP_DR1,
-    .flush = vp8_decode_flush,
-    .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),
+    .name                  = "vp8",
+    .long_name             = NULL_IF_CONFIG_SMALL("On2 VP8"),
+    .type                  = AVMEDIA_TYPE_VIDEO,
+    .id                    = AV_CODEC_ID_VP8,
+    .priv_data_size        = sizeof(VP8Context),
+    .init                  = ff_vp8_decode_init,
+    .close                 = ff_vp8_decode_free,
+    .decode                = ff_vp8_decode_frame,
+    .capabilities          = CODEC_CAP_DR1 | CODEC_CAP_FRAME_THREADS | CODEC_CAP_SLICE_THREADS,
+    .flush                 = vp8_decode_flush,
+    .init_thread_copy      = ONLY_IF_THREADS_ENABLED(vp8_decode_init_thread_copy),
+    .update_thread_context = ONLY_IF_THREADS_ENABLED(vp8_decode_update_thread_context),
 };
+#endif /* CONFIG_VP7_DECODER */