avcodec/vp9: Add tile threading support

Signed-off-by: Ilia Valiakhmetov <zakne0ne@gmail.com>
Signed-off-by: Ronald S. Bultje <rsbultje@gmail.com>

7 files changed, 1051 insertions, 755 deletions

diff --git a/libavcodec/vp9.c b/libavcodec/vp9.c
index 94430db..a71045e 100644
--- a/libavcodec/vp9.c
+++ b/libavcodec/vp9.c
@@ -36,6 +36,64 @@
 
 #define VP9_SYNCCODE 0x498342
 
+#if HAVE_THREADS
+static void vp9_free_entries(AVCodecContext *avctx) {
+    VP9Context *s = avctx->priv_data;
+
+    if (avctx->active_thread_type & FF_THREAD_SLICE)  {
+        pthread_mutex_destroy(&s->progress_mutex);
+        pthread_cond_destroy(&s->progress_cond);
+        av_freep(&s->entries);
+    }
+}
+
+static int vp9_alloc_entries(AVCodecContext *avctx, int n) {
+    VP9Context *s = avctx->priv_data;
+    int i;
+
+    if (avctx->active_thread_type & FF_THREAD_SLICE)  {
+        if (s->entries)
+            av_freep(&s->entries);
+
+        s->entries = av_malloc_array(n, sizeof(atomic_int));
+
+        if (!s->entries) {
+            av_freep(&s->entries);
+            return AVERROR(ENOMEM);
+        }
+
+        for (i  = 0; i < n; i++)
+            atomic_init(&s->entries[i], 0);
+
+        pthread_mutex_init(&s->progress_mutex, NULL);
+        pthread_cond_init(&s->progress_cond, NULL);
+    }
+    return 0;
+}
+
+static void vp9_report_tile_progress(VP9Context *s, int field, int n) {
+    pthread_mutex_lock(&s->progress_mutex);
+    atomic_fetch_add_explicit(&s->entries[field], n, memory_order_relaxed);
+    pthread_cond_signal(&s->progress_cond);
+    pthread_mutex_unlock(&s->progress_mutex);
+}
+
+static void vp9_await_tile_progress(VP9Context *s, int field, int n) {
+    if (atomic_load_explicit(&s->entries[field], memory_order_acquire) >= n)
+        return;
+
+    pthread_mutex_lock(&s->progress_mutex);
+    while (atomic_load_explicit(&s->entries[field], memory_order_relaxed) != n)
+        pthread_cond_wait(&s->progress_cond, &s->progress_mutex);
+    pthread_mutex_unlock(&s->progress_mutex);
+}
+#else
+static void vp9_free_entries(VP9Context *s) {}
+static int vp9_alloc_entries(AVCodecContext *avctx, int n) { return 0; }
+static void vp9_report_tile_progress(VP9Context *s, int field, int n) {}
+static void vp9_await_tile_progress(VP9Context *s, int field, int n) {}
+#endif
+
 static void vp9_frame_unref(AVCodecContext *avctx, VP9Frame *f)
 {
     ff_thread_release_buffer(avctx, &f->tf);
@@ -118,6 +176,7 @@ static int update_size(AVCodecContext *avctx, int w, int h)
     VP9Context *s = avctx->priv_data;
     uint8_t *p;
     int bytesperpixel = s->bytesperpixel, ret, cols, rows;
+    int lflvl_len, i;
 
     av_assert0(w > 0 && h > 0);
 
@@ -170,13 +229,14 @@ static int update_size(AVCodecContext *avctx, int w, int h)
     s->sb_rows   = (h + 63) >> 6;
     s->cols      = (w + 7) >> 3;
     s->rows      = (h + 7) >> 3;
+    lflvl_len    = avctx->active_thread_type == FF_THREAD_SLICE ? s->sb_rows : 1;
 
 #define assign(var, type, n) var = (type) p; p += s->sb_cols * (n) * sizeof(*var)
     av_freep(&s->intra_pred_data[0]);
     // FIXME we slightly over-allocate here for subsampled chroma, but a little
     // bit of padding shouldn't affect performance...
     p = av_malloc(s->sb_cols * (128 + 192 * bytesperpixel +
-                                sizeof(*s->lflvl) + 16 * sizeof(*s->above_mv_ctx)));
+                                lflvl_len * sizeof(*s->lflvl) + 16 * sizeof(*s->above_mv_ctx)));
     if (!p)
         return AVERROR(ENOMEM);
     assign(s->intra_pred_data[0],  uint8_t *,             64 * bytesperpixel);
@@ -195,12 +255,15 @@ static int update_size(AVCodecContext *avctx, int w, int h)
     assign(s->above_comp_ctx,      uint8_t *,              8);
     assign(s->above_ref_ctx,       uint8_t *,              8);
     assign(s->above_filter_ctx,    uint8_t *,              8);
-    assign(s->lflvl,               VP9Filter *,            1);
+    assign(s->lflvl,               VP9Filter *,            lflvl_len);
 #undef assign
 
-    // these will be re-allocated a little later
-    av_freep(&s->b_base);
-    av_freep(&s->block_base);
+    if (s->td) {
+        for (i = 0; i < s->active_tile_cols; i++) {
+            av_freep(&s->td[i].b_base);
+            av_freep(&s->td[i].block_base);
+        }
+    }
 
     if (s->s.h.bpp != s->last_bpp) {
         ff_vp9dsp_init(&s->dsp, s->s.h.bpp, avctx->flags & AV_CODEC_FLAG_BITEXACT);
@@ -213,40 +276,50 @@ static int update_size(AVCodecContext *avctx, int w, int h)
 
 static int update_block_buffers(AVCodecContext *avctx)
 {
+    int i;
     VP9Context *s = avctx->priv_data;
     int chroma_blocks, chroma_eobs, bytesperpixel = s->bytesperpixel;
+    VP9TileData *td = &s->td[0];
 
-    if (s->b_base && s->block_base && s->block_alloc_using_2pass == s->s.frames[CUR_FRAME].uses_2pass)
+    if (td->b_base && td->block_base && s->block_alloc_using_2pass == s->s.frames[CUR_FRAME].uses_2pass)
         return 0;
 
-    av_free(s->b_base);
-    av_free(s->block_base);
+    av_free(td->b_base);
+    av_free(td->block_base);
     chroma_blocks = 64 * 64 >> (s->ss_h + s->ss_v);
     chroma_eobs   = 16 * 16 >> (s->ss_h + s->ss_v);
     if (s->s.frames[CUR_FRAME].uses_2pass) {
         int sbs = s->sb_cols * s->sb_rows;
 
-        s->b_base = av_malloc_array(s->cols * s->rows, sizeof(VP9Block));
-        s->block_base = av_mallocz(((64 * 64 + 2 * chroma_blocks) * bytesperpixel * sizeof(int16_t) +
+        td->b_base = av_malloc_array(s->cols * s->rows, sizeof(VP9Block));
+        td->block_base = av_mallocz(((64 * 64 + 2 * chroma_blocks) * bytesperpixel * sizeof(int16_t) +
                                     16 * 16 + 2 * chroma_eobs) * sbs);
-        if (!s->b_base || !s->block_base)
+        if (!td->b_base || !td->block_base)
             return AVERROR(ENOMEM);
-        s->uvblock_base[0] = s->block_base + sbs * 64 * 64 * bytesperpixel;
-        s->uvblock_base[1] = s->uvblock_base[0] + sbs * chroma_blocks * bytesperpixel;
-        s->eob_base = (uint8_t *) (s->uvblock_base[1] + sbs * chroma_blocks * bytesperpixel);
-        s->uveob_base[0] = s->eob_base + 16 * 16 * sbs;
-        s->uveob_base[1] = s->uveob_base[0] + chroma_eobs * sbs;
+        td->uvblock_base[0] = td->block_base + sbs * 64 * 64 * bytesperpixel;
+        td->uvblock_base[1] = td->uvblock_base[0] + sbs * chroma_blocks * bytesperpixel;
+        td->eob_base = (uint8_t *) (td->uvblock_base[1] + sbs * chroma_blocks * bytesperpixel);
+        td->uveob_base[0] = td->eob_base + 16 * 16 * sbs;
+        td->uveob_base[1] = td->uveob_base[0] + chroma_eobs * sbs;
     } else {
-        s->b_base = av_malloc(sizeof(VP9Block));
-        s->block_base = av_mallocz((64 * 64 + 2 * chroma_blocks) * bytesperpixel * sizeof(int16_t) +
-                                   16 * 16 + 2 * chroma_eobs);
-        if (!s->b_base || !s->block_base)
-            return AVERROR(ENOMEM);
-        s->uvblock_base[0] = s->block_base + 64 * 64 * bytesperpixel;
-        s->uvblock_base[1] = s->uvblock_base[0] + chroma_blocks * bytesperpixel;
-        s->eob_base = (uint8_t *) (s->uvblock_base[1] + chroma_blocks * bytesperpixel);
-        s->uveob_base[0] = s->eob_base + 16 * 16;
-        s->uveob_base[1] = s->uveob_base[0] + chroma_eobs;
+        for (i = 1; i < s->active_tile_cols; i++) {
+            if (s->td[i].b_base && s->td[i].block_base) {
+                av_free(s->td[i].b_base);
+                av_free(s->td[i].block_base);
+            }
+        }
+        for (i = 0; i < s->active_tile_cols; i++) {
+            s->td[i].b_base = av_malloc(sizeof(VP9Block));
+            s->td[i].block_base = av_mallocz((64 * 64 + 2 * chroma_blocks) * bytesperpixel * sizeof(int16_t) +
+                                       16 * 16 + 2 * chroma_eobs);
+            if (!s->td[i].b_base || !s->td[i].block_base)
+                return AVERROR(ENOMEM);
+            s->td[i].uvblock_base[0] = s->td[i].block_base + 64 * 64 * bytesperpixel;
+            s->td[i].uvblock_base[1] = s->td[i].uvblock_base[0] + chroma_blocks * bytesperpixel;
+            s->td[i].eob_base = (uint8_t *) (s->td[i].uvblock_base[1] + chroma_blocks * bytesperpixel);
+            s->td[i].uveob_base[0] = s->td[i].eob_base + 16 * 16;
+            s->td[i].uveob_base[1] = s->td[i].uveob_base[0] + chroma_eobs;
+        }
     }
     s->block_alloc_using_2pass = s->s.frames[CUR_FRAME].uses_2pass;
 
@@ -543,8 +616,20 @@ static int decode_frame_header(AVCodecContext *avctx,
     sharp = get_bits(&s->gb, 3);
     // if sharpness changed, reinit lim/mblim LUTs. if it didn't change, keep
     // the old cache values since they are still valid
-    if (s->s.h.filter.sharpness != sharp)
-        memset(s->filter_lut.lim_lut, 0, sizeof(s->filter_lut.lim_lut));
+    if (s->s.h.filter.sharpness != sharp) {
+        for (i = 1; i <= 63; i++) {
+            int limit = i;
+
+            if (sharp > 0) {
+                limit >>= (sharp + 3) >> 2;
+                limit = FFMIN(limit, 9 - sharp);
+            }
+            limit = FFMAX(limit, 1);
+
+            s->filter_lut.lim_lut[i] = limit;
+            s->filter_lut.mblim_lut[i] = 2 * (i + 2) + limit;
+        }
+    }
     s->s.h.filter.sharpness = sharp;
     if ((s->s.h.lf_delta.enabled = get_bits1(&s->gb))) {
         if ((s->s.h.lf_delta.updated = get_bits1(&s->gb))) {
@@ -662,12 +747,36 @@ static int decode_frame_header(AVCodecContext *avctx,
     s->s.h.tiling.log2_tile_rows = decode012(&s->gb);
     s->s.h.tiling.tile_rows = 1 << s->s.h.tiling.log2_tile_rows;
     if (s->s.h.tiling.tile_cols != (1 << s->s.h.tiling.log2_tile_cols)) {
+        int n_range_coders;
+        VP56RangeCoder *rc;
+
+        if (s->td) {
+            for (i = 0; i < s->active_tile_cols; i++) {
+                av_free(s->td[i].b_base);
+                av_free(s->td[i].block_base);
+            }
+            av_free(s->td);
+        }
+
         s->s.h.tiling.tile_cols = 1 << s->s.h.tiling.log2_tile_cols;
-        s->c_b = av_fast_realloc(s->c_b, &s->c_b_size,
-                                 sizeof(VP56RangeCoder) * s->s.h.tiling.tile_cols);
-        if (!s->c_b) {
-            av_log(avctx, AV_LOG_ERROR, "Ran out of memory during range coder init\n");
+        vp9_free_entries(avctx);
+        s->active_tile_cols = avctx->active_thread_type == FF_THREAD_SLICE ?
+                              s->s.h.tiling.tile_cols : 1;
+        vp9_alloc_entries(avctx, s->sb_rows);
+        if (avctx->active_thread_type == FF_THREAD_SLICE) {
+            n_range_coders = 4; // max_tile_rows
+        } else {
+            n_range_coders = s->s.h.tiling.tile_cols;
+        }
+        s->td = av_mallocz_array(s->active_tile_cols, sizeof(VP9TileData) +
+                                 n_range_coders * sizeof(VP56RangeCoder));
+        if (!s->td)
             return AVERROR(ENOMEM);
+        rc = (VP56RangeCoder *) &s->td[s->active_tile_cols];
+        for (i = 0; i < s->active_tile_cols; i++) {
+            s->td[i].s = s;
+            s->td[i].c_b = rc;
+            rc += n_range_coders;
         }
     }
 
@@ -735,12 +844,15 @@ static int decode_frame_header(AVCodecContext *avctx,
         return AVERROR_INVALIDDATA;
     }
 
-    if (s->s.h.keyframe || s->s.h.intraonly) {
-        memset(s->counts.coef, 0, sizeof(s->counts.coef));
-        memset(s->counts.eob,  0, sizeof(s->counts.eob));
-    } else {
-        memset(&s->counts, 0, sizeof(s->counts));
+    for (i = 0; i < s->active_tile_cols; i++) {
+        if (s->s.h.keyframe || s->s.h.intraonly) {
+            memset(s->td[i].counts.coef, 0, sizeof(s->td[0].counts.coef));
+            memset(s->td[i].counts.eob,  0, sizeof(s->td[0].counts.eob));
+        } else {
+            memset(&s->td[i].counts, 0, sizeof(s->td[0].counts));
+        }
     }
+
     /* FIXME is it faster to not copy here, but do it down in the fw updates
      * as explicit copies if the fw update is missing (and skip the copy upon
      * fw update)? */
@@ -789,7 +901,7 @@ static int decode_frame_header(AVCodecContext *avctx,
                                 else
                                     p[n] = r[n];
                             }
-                            p[3] = 0;
+                            memcpy(&p[3], ff_vp9_model_pareto8[p[2]], 8);
                         }
         } else {
             for (j = 0; j < 2; j++)
@@ -801,7 +913,7 @@ static int decode_frame_header(AVCodecContext *avctx,
                             if (m > 3 && l == 0) // dc only has 3 pt
                                 break;
                             memcpy(p, r, 3);
-                            p[3] = 0;
+                            memcpy(&p[3], ff_vp9_model_pareto8[p[2]], 8);
                         }
         }
         if (s->s.h.txfmmode == i)
@@ -929,12 +1041,12 @@ static int decode_frame_header(AVCodecContext *avctx,
     return (data2 - data) + size2;
 }
 
-static void decode_sb(AVCodecContext *avctx, int row, int col, VP9Filter *lflvl,
+static void decode_sb(VP9TileData *td, int row, int col, VP9Filter *lflvl,
                       ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)
 {
-    VP9Context *s = avctx->priv_data;
+    const VP9Context *s = td->s;
     int c = ((s->above_partition_ctx[col] >> (3 - bl)) & 1) |
-            (((s->left_partition_ctx[row & 0x7] >> (3 - bl)) & 1) << 1);
+            (((td->left_partition_ctx[row & 0x7] >> (3 - bl)) & 1) << 1);
     const uint8_t *p = s->s.h.keyframe || s->s.h.intraonly ? ff_vp9_default_kf_partition_probs[bl][c] :
                                                      s->prob.p.partition[bl][c];
     enum BlockPartition bp;
@@ -944,75 +1056,75 @@ static void decode_sb(AVCodecContext *avctx, int row, int col, VP9Filter *lflvl,
     int bytesperpixel = s->bytesperpixel;
 
     if (bl == BL_8X8) {
-        bp = vp8_rac_get_tree(&s->c, ff_vp9_partition_tree, p);
-        ff_vp9_decode_block(avctx, row, col, lflvl, yoff, uvoff, bl, bp);
+        bp = vp8_rac_get_tree(td->c, ff_vp9_partition_tree, p);
+        ff_vp9_decode_block(td, row, col, lflvl, yoff, uvoff, bl, bp);
     } else if (col + hbs < s->cols) { // FIXME why not <=?
         if (row + hbs < s->rows) { // FIXME why not <=?
-            bp = vp8_rac_get_tree(&s->c, ff_vp9_partition_tree, p);
+            bp = vp8_rac_get_tree(td->c, ff_vp9_partition_tree, p);
             switch (bp) {
             case PARTITION_NONE:
-                ff_vp9_decode_block(avctx, row, col, lflvl, yoff, uvoff, bl, bp);
+                ff_vp9_decode_block(td, row, col, lflvl, yoff, uvoff, bl, bp);
                 break;
             case PARTITION_H:
-                ff_vp9_decode_block(avctx, row, col, lflvl, yoff, uvoff, bl, bp);
+                ff_vp9_decode_block(td, row, col, lflvl, yoff, uvoff, bl, bp);
                 yoff  += hbs * 8 * y_stride;
                 uvoff += hbs * 8 * uv_stride >> s->ss_v;
-                ff_vp9_decode_block(avctx, row + hbs, col, lflvl, yoff, uvoff, bl, bp);
+                ff_vp9_decode_block(td, row + hbs, col, lflvl, yoff, uvoff, bl, bp);
                 break;
             case PARTITION_V:
-                ff_vp9_decode_block(avctx, row, col, lflvl, yoff, uvoff, bl, bp);
+                ff_vp9_decode_block(td, row, col, lflvl, yoff, uvoff, bl, bp);
                 yoff  += hbs * 8 * bytesperpixel;
                 uvoff += hbs * 8 * bytesperpixel >> s->ss_h;
-                ff_vp9_decode_block(avctx, row, col + hbs, lflvl, yoff, uvoff, bl, bp);
+                ff_vp9_decode_block(td, row, col + hbs, lflvl, yoff, uvoff, bl, bp);
                 break;
             case PARTITION_SPLIT:
-                decode_sb(avctx, row, col, lflvl, yoff, uvoff, bl + 1);
-                decode_sb(avctx, row, col + hbs, lflvl,
+                decode_sb(td, row, col, lflvl, yoff, uvoff, bl + 1);
+                decode_sb(td, row, col + hbs, lflvl,
                           yoff + 8 * hbs * bytesperpixel,
                           uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);
                 yoff  += hbs * 8 * y_stride;
                 uvoff += hbs * 8 * uv_stride >> s->ss_v;
-                decode_sb(avctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);
-                decode_sb(avctx, row + hbs, col + hbs, lflvl,
+                decode_sb(td, row + hbs, col, lflvl, yoff, uvoff, bl + 1);
+                decode_sb(td, row + hbs, col + hbs, lflvl,
                           yoff + 8 * hbs * bytesperpixel,
                           uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);
                 break;
             default:
                 av_assert0(0);
             }
-        } else if (vp56_rac_get_prob_branchy(&s->c, p[1])) {
+        } else if (vp56_rac_get_prob_branchy(td->c, p[1])) {
             bp = PARTITION_SPLIT;
-            decode_sb(avctx, row, col, lflvl, yoff, uvoff, bl + 1);
-            decode_sb(avctx, row, col + hbs, lflvl,
+            decode_sb(td, row, col, lflvl, yoff, uvoff, bl + 1);
+            decode_sb(td, row, col + hbs, lflvl,
                       yoff + 8 * hbs * bytesperpixel,
                       uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);
         } else {
             bp = PARTITION_H;
-            ff_vp9_decode_block(avctx, row, col, lflvl, yoff, uvoff, bl, bp);
+            ff_vp9_decode_block(td, row, col, lflvl, yoff, uvoff, bl, bp);
         }
     } else if (row + hbs < s->rows) { // FIXME why not <=?
-        if (vp56_rac_get_prob_branchy(&s->c, p[2])) {
+        if (vp56_rac_get_prob_branchy(td->c, p[2])) {
             bp = PARTITION_SPLIT;
-            decode_sb(avctx, row, col, lflvl, yoff, uvoff, bl + 1);
+            decode_sb(td, row, col, lflvl, yoff, uvoff, bl + 1);
             yoff  += hbs * 8 * y_stride;
             uvoff += hbs * 8 * uv_stride >> s->ss_v;
-            decode_sb(avctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);
+            decode_sb(td, row + hbs, col, lflvl, yoff, uvoff, bl + 1);
         } else {
             bp = PARTITION_V;
-            ff_vp9_decode_block(avctx, row, col, lflvl, yoff, uvoff, bl, bp);
+            ff_vp9_decode_block(td, row, col, lflvl, yoff, uvoff, bl, bp);
         }
     } else {
         bp = PARTITION_SPLIT;
-        decode_sb(avctx, row, col, lflvl, yoff, uvoff, bl + 1);
+        decode_sb(td, row, col, lflvl, yoff, uvoff, bl + 1);
     }
-    s->counts.partition[bl][c][bp]++;
+    td->counts.partition[bl][c][bp]++;
 }
 
-static void decode_sb_mem(AVCodecContext *avctx, int row, int col, VP9Filter *lflvl,
+static void decode_sb_mem(VP9TileData *td, int row, int col, VP9Filter *lflvl,
                           ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)
 {
-    VP9Context *s = avctx->priv_data;
-    VP9Block *b = s->b;
+    const VP9Context *s = td->s;
+    VP9Block *b = td->b;
     ptrdiff_t hbs = 4 >> bl;
     AVFrame *f = s->s.frames[CUR_FRAME].tf.f;
     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];
@@ -1020,39 +1132,39 @@ static void decode_sb_mem(AVCodecContext *avctx, int row, int col, VP9Filter *lf
 
     if (bl == BL_8X8) {
         av_assert2(b->bl == BL_8X8);
-        ff_vp9_decode_block(avctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);
-    } else if (s->b->bl == bl) {
-        ff_vp9_decode_block(avctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);
+        ff_vp9_decode_block(td, row, col, lflvl, yoff, uvoff, b->bl, b->bp);
+    } else if (td->b->bl == bl) {
+        ff_vp9_decode_block(td, row, col, lflvl, yoff, uvoff, b->bl, b->bp);
         if (b->bp == PARTITION_H && row + hbs < s->rows) {
             yoff  += hbs * 8 * y_stride;
             uvoff += hbs * 8 * uv_stride >> s->ss_v;
-            ff_vp9_decode_block(avctx, row + hbs, col, lflvl, yoff, uvoff, b->bl, b->bp);
+            ff_vp9_decode_block(td, row + hbs, col, lflvl, yoff, uvoff, b->bl, b->bp);
         } else if (b->bp == PARTITION_V && col + hbs < s->cols) {
             yoff  += hbs * 8 * bytesperpixel;
             uvoff += hbs * 8 * bytesperpixel >> s->ss_h;
-            ff_vp9_decode_block(avctx, row, col + hbs, lflvl, yoff, uvoff, b->bl, b->bp);
+            ff_vp9_decode_block(td, row, col + hbs, lflvl, yoff, uvoff, b->bl, b->bp);
         }
     } else {
-        decode_sb_mem(avctx, row, col, lflvl, yoff, uvoff, bl + 1);
+        decode_sb_mem(td, row, col, lflvl, yoff, uvoff, bl + 1);
         if (col + hbs < s->cols) { // FIXME why not <=?
             if (row + hbs < s->rows) {
-                decode_sb_mem(avctx, row, col + hbs, lflvl, yoff + 8 * hbs * bytesperpixel,
+                decode_sb_mem(td, row, col + hbs, lflvl, yoff + 8 * hbs * bytesperpixel,
                               uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);
                 yoff  += hbs * 8 * y_stride;
                 uvoff += hbs * 8 * uv_stride >> s->ss_v;
-                decode_sb_mem(avctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);
-                decode_sb_mem(avctx, row + hbs, col + hbs, lflvl,
+                decode_sb_mem(td, row + hbs, col, lflvl, yoff, uvoff, bl + 1);
+                decode_sb_mem(td, row + hbs, col + hbs, lflvl,
                               yoff + 8 * hbs * bytesperpixel,
                               uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);
             } else {
                 yoff  += hbs * 8 * bytesperpixel;
                 uvoff += hbs * 8 * bytesperpixel >> s->ss_h;
-                decode_sb_mem(avctx, row, col + hbs, lflvl, yoff, uvoff, bl + 1);
+                decode_sb_mem(td, row, col + hbs, lflvl, yoff, uvoff, bl + 1);
             }
         } else if (row + hbs < s->rows) {
             yoff  += hbs * 8 * y_stride;
             uvoff += hbs * 8 * uv_stride >> s->ss_v;
-            decode_sb_mem(avctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);
+            decode_sb_mem(td, row + hbs, col, lflvl, yoff, uvoff, bl + 1);
         }
     }
 }
@@ -1067,9 +1179,13 @@ static void set_tile_offset(int *start, int *end, int idx, int log2_n, int n)
 
 static void free_buffers(VP9Context *s)
 {
+    int i;
+
     av_freep(&s->intra_pred_data[0]);
-    av_freep(&s->b_base);
-    av_freep(&s->block_base);
+    for (i = 0; i < s->active_tile_cols; i++) {
+        av_freep(&s->td[i].b_base);
+        av_freep(&s->td[i].block_base);
+    }
 }
 
 static av_cold int vp9_decode_free(AVCodecContext *avctx)
@@ -1090,10 +1206,249 @@ static av_cold int vp9_decode_free(AVCodecContext *avctx)
             ff_thread_release_buffer(avctx, &s->next_refs[i]);
         av_frame_free(&s->next_refs[i].f);
     }
+
     free_buffers(s);
-    av_freep(&s->c_b);
-    s->c_b_size = 0;
+    vp9_free_entries(avctx);
+    av_freep(&s->td);
+    return 0;
+}
+
+static int decode_tiles(AVCodecContext *avctx,
+                        const uint8_t *data, int size)
+{
+    VP9Context *s = avctx->priv_data;
+    VP9TileData *td = &s->td[0];
+    int row, col, tile_row, tile_col, ret;
+    int bytesperpixel;
+    int tile_row_start, tile_row_end, tile_col_start, tile_col_end;
+    AVFrame *f;
+    ptrdiff_t yoff, uvoff, ls_y, ls_uv;
+
+    f = s->s.frames[CUR_FRAME].tf.f;
+    ls_y = f->linesize[0];
+    ls_uv =f->linesize[1];
+    bytesperpixel = s->bytesperpixel;
+
+    yoff = uvoff = 0;
+    for (tile_row = 0; tile_row < s->s.h.tiling.tile_rows; tile_row++) {
+        set_tile_offset(&tile_row_start, &tile_row_end,
+                        tile_row, s->s.h.tiling.log2_tile_rows, s->sb_rows);
+
+        for (tile_col = 0; tile_col < s->s.h.tiling.tile_cols; tile_col++) {
+            int64_t tile_size;
 
+            if (tile_col == s->s.h.tiling.tile_cols - 1 &&
+                tile_row == s->s.h.tiling.tile_rows - 1) {
+                tile_size = size;
+            } else {
+                tile_size = AV_RB32(data);
+                data += 4;
+                size -= 4;
+            }
+            if (tile_size > size) {
+                ff_thread_report_progress(&s->s.frames[CUR_FRAME].tf, INT_MAX, 0);
+                return AVERROR_INVALIDDATA;
+            }
+            ret = ff_vp56_init_range_decoder(&td->c_b[tile_col], data, tile_size);
+            if (ret < 0)
+                return ret;
+            if (vp56_rac_get_prob_branchy(&td->c_b[tile_col], 128)) { // marker bit
+                ff_thread_report_progress(&s->s.frames[CUR_FRAME].tf, INT_MAX, 0);
+                return AVERROR_INVALIDDATA;
+            }
+            data += tile_size;
+            size -= tile_size;
+        }
+
+        for (row = tile_row_start; row < tile_row_end;
+             row += 8, yoff += ls_y * 64, uvoff += ls_uv * 64 >> s->ss_v) {
+            VP9Filter *lflvl_ptr = s->lflvl;
+            ptrdiff_t yoff2 = yoff, uvoff2 = uvoff;
+
+            for (tile_col = 0; tile_col < s->s.h.tiling.tile_cols; tile_col++) {
+                set_tile_offset(&tile_col_start, &tile_col_end,
+                                tile_col, s->s.h.tiling.log2_tile_cols, s->sb_cols);
+                td->tile_col_start = tile_col_start;
+                if (s->pass != 2) {
+                    memset(td->left_partition_ctx, 0, 8);
+                    memset(td->left_skip_ctx, 0, 8);
+                    if (s->s.h.keyframe || s->s.h.intraonly) {
+                        memset(td->left_mode_ctx, DC_PRED, 16);
+                    } else {
+                        memset(td->left_mode_ctx, NEARESTMV, 8);
+                    }
+                    memset(td->left_y_nnz_ctx, 0, 16);
+                    memset(td->left_uv_nnz_ctx, 0, 32);
+                    memset(td->left_segpred_ctx, 0, 8);
+
+                    td->c = &td->c_b[tile_col];
+                }
+
+                for (col = tile_col_start;
+                     col < tile_col_end;
+                     col += 8, yoff2 += 64 * bytesperpixel,
+                     uvoff2 += 64 * bytesperpixel >> s->ss_h, lflvl_ptr++) {
+                    // FIXME integrate with lf code (i.e. zero after each
+                    // use, similar to invtxfm coefficients, or similar)
+                    if (s->pass != 1) {
+                        memset(lflvl_ptr->mask, 0, sizeof(lflvl_ptr->mask));
+                    }
+
+                    if (s->pass == 2) {
+                        decode_sb_mem(td, row, col, lflvl_ptr,
+                                      yoff2, uvoff2, BL_64X64);
+                    } else {
+                        decode_sb(td, row, col, lflvl_ptr,
+                                  yoff2, uvoff2, BL_64X64);
+                    }
+                }
+            }
+
+            if (s->pass == 1)
+                continue;
+
+            // backup pre-loopfilter reconstruction data for intra
+            // prediction of next row of sb64s
+            if (row + 8 < s->rows) {
+                memcpy(s->intra_pred_data[0],
+                       f->data[0] + yoff + 63 * ls_y,
+                       8 * s->cols * bytesperpixel);
+                memcpy(s->intra_pred_data[1],
+                       f->data[1] + uvoff + ((64 >> s->ss_v) - 1) * ls_uv,
+                       8 * s->cols * bytesperpixel >> s->ss_h);
+                memcpy(s->intra_pred_data[2],
+                       f->data[2] + uvoff + ((64 >> s->ss_v) - 1) * ls_uv,
+                       8 * s->cols * bytesperpixel >> s->ss_h);
+            }
+
+            // loopfilter one row
+            if (s->s.h.filter.level) {
+                yoff2 = yoff;
+                uvoff2 = uvoff;
+                lflvl_ptr = s->lflvl;
+                for (col = 0; col < s->cols;
+                     col += 8, yoff2 += 64 * bytesperpixel,
+                     uvoff2 += 64 * bytesperpixel >> s->ss_h, lflvl_ptr++) {
+                    ff_vp9_loopfilter_sb(avctx, lflvl_ptr, row, col,
+                                         yoff2, uvoff2);
+                }
+            }
+
+            // FIXME maybe we can make this more finegrained by running the
+            // loopfilter per-block instead of after each sbrow
+            // In fact that would also make intra pred left preparation easier?
+            ff_thread_report_progress(&s->s.frames[CUR_FRAME].tf, row >> 3, 0);
+        }
+    }
+    return 0;
+}
+
+
+static av_always_inline
+int decode_tiles_mt(AVCodecContext *avctx, void *tdata, int jobnr,
+                              int threadnr)
+{
+    VP9Context *s = avctx->priv_data;
+    VP9TileData *td = &s->td[jobnr];
+    ptrdiff_t uvoff, yoff, ls_y, ls_uv;
+    int bytesperpixel = s->bytesperpixel, row, col, tile_row;
+    unsigned tile_cols_len;
+    int tile_row_start, tile_row_end, tile_col_start, tile_col_end;
+    VP9Filter *lflvl_ptr_base;
+    AVFrame *f;
+
+    f = s->s.frames[CUR_FRAME].tf.f;
+    ls_y = f->linesize[0];
+    ls_uv =f->linesize[1];
+
+    set_tile_offset(&tile_col_start, &tile_col_end,
+                    jobnr, s->s.h.tiling.log2_tile_cols, s->sb_cols);
+    td->tile_col_start  = tile_col_start;
+    uvoff = (64 * bytesperpixel >> s->ss_h)*(tile_col_start >> 3);
+    yoff = (64 * bytesperpixel)*(tile_col_start >> 3);
+    lflvl_ptr_base = s->lflvl+(tile_col_start >> 3);
+
+    for (tile_row = 0; tile_row < s->s.h.tiling.tile_rows; tile_row++) {
+        set_tile_offset(&tile_row_start, &tile_row_end,
+                        tile_row, s->s.h.tiling.log2_tile_rows, s->sb_rows);
+
+        td->c = &td->c_b[tile_row];
+        for (row = tile_row_start; row < tile_row_end;
+             row += 8, yoff += ls_y * 64, uvoff += ls_uv * 64 >> s->ss_v) {
+            ptrdiff_t yoff2 = yoff, uvoff2 = uvoff;
+            VP9Filter *lflvl_ptr = lflvl_ptr_base+s->sb_cols*(row >> 3);
+
+            memset(td->left_partition_ctx, 0, 8);
+            memset(td->left_skip_ctx, 0, 8);
+            if (s->s.h.keyframe || s->s.h.intraonly) {
+                memset(td->left_mode_ctx, DC_PRED, 16);
+            } else {
+                memset(td->left_mode_ctx, NEARESTMV, 8);
+            }
+            memset(td->left_y_nnz_ctx, 0, 16);
+            memset(td->left_uv_nnz_ctx, 0, 32);
+            memset(td->left_segpred_ctx, 0, 8);
+
+            for (col = tile_col_start;
+                 col < tile_col_end;
+                 col += 8, yoff2 += 64 * bytesperpixel,
+                 uvoff2 += 64 * bytesperpixel >> s->ss_h, lflvl_ptr++) {
+                // FIXME integrate with lf code (i.e. zero after each
+                // use, similar to invtxfm coefficients, or similar)
+                memset(lflvl_ptr->mask, 0, sizeof(lflvl_ptr->mask));
+                decode_sb(td, row, col, lflvl_ptr,
+                            yoff2, uvoff2, BL_64X64);
+            }
+
+            // backup pre-loopfilter reconstruction data for intra
+            // prediction of next row of sb64s
+            tile_cols_len = tile_col_end - tile_col_start;
+            if (row + 8 < s->rows) {
+                memcpy(s->intra_pred_data[0] + (tile_col_start * 8 * bytesperpixel),
+                       f->data[0] + yoff + 63 * ls_y,
+                       8 * tile_cols_len * bytesperpixel);
+                memcpy(s->intra_pred_data[1] + (tile_col_start * 8 * bytesperpixel >> s->ss_h),
+                       f->data[1] + uvoff + ((64 >> s->ss_v) - 1) * ls_uv,
+                       8 * tile_cols_len * bytesperpixel >> s->ss_h);
+                memcpy(s->intra_pred_data[2] + (tile_col_start * 8 * bytesperpixel >> s->ss_h),
+                       f->data[2] + uvoff + ((64 >> s->ss_v) - 1) * ls_uv,
+                       8 * tile_cols_len * bytesperpixel >> s->ss_h);
+            }
+
+            vp9_report_tile_progress(s, row >> 3, 1);
+        }
+    }
+    return 0;
+}
+
+static av_always_inline
+int loopfilter_proc(AVCodecContext *avctx)
+{
+    VP9Context *s = avctx->priv_data;
+    ptrdiff_t uvoff, yoff, ls_y, ls_uv;
+    VP9Filter *lflvl_ptr;
+    int bytesperpixel = s->bytesperpixel, col, i;
+    AVFrame *f;
+
+    f = s->s.frames[CUR_FRAME].tf.f;
+    ls_y = f->linesize[0];
+    ls_uv =f->linesize[1];
+
+    for (i = 0; i < s->sb_rows; i++) {
+        vp9_await_tile_progress(s, i, s->s.h.tiling.tile_cols);
+
+        if (s->s.h.filter.level) {
+            yoff = (ls_y * 64)*i;
+            uvoff =  (ls_uv * 64 >> s->ss_v)*i;
+            lflvl_ptr = s->lflvl+s->sb_cols*i;
+            for (col = 0; col < s->cols;
+                 col += 8, yoff += 64 * bytesperpixel,
+                 uvoff += 64 * bytesperpixel >> s->ss_h, lflvl_ptr++) {
+                ff_vp9_loopfilter_sb(avctx, lflvl_ptr, i << 3, col,
+                                     yoff, uvoff);
+            }
+        }
+    }
     return 0;
 }
 
@@ -1104,12 +1459,10 @@ static int vp9_decode_frame(AVCodecContext *avctx, void *frame,
     const uint8_t *data = pkt->data;
     int size = pkt->size;
     VP9Context *s = avctx->priv_data;
-    int ret, tile_row, tile_col, i, ref, row, col;
+    int ret, i, j, ref;
     int retain_segmap_ref = s->s.frames[REF_FRAME_SEGMAP].segmentation_map &&
                             (!s->s.h.segmentation.enabled || !s->s.h.segmentation.update_map);
-    ptrdiff_t yoff, uvoff, ls_y, ls_uv;
     AVFrame *f;
-    int bytesperpixel;
 
     if ((ret = decode_frame_header(avctx, data, size, &ref)) < 0) {
         return ret;
@@ -1159,8 +1512,6 @@ FF_ENABLE_DEPRECATION_WARNINGS
     f = s->s.frames[CUR_FRAME].tf.f;
     f->key_frame = s->s.h.keyframe;
     f->pict_type = (s->s.h.keyframe || s->s.h.intraonly) ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;
-    ls_y = f->linesize[0];
-    ls_uv =f->linesize[1];
 
     if (s->s.frames[REF_FRAME_SEGMAP].tf.f->buf[0] &&
         (s->s.frames[REF_FRAME_MVPAIR].tf.f->width  != s->s.frames[CUR_FRAME].tf.f->width ||
@@ -1195,7 +1546,6 @@ FF_ENABLE_DEPRECATION_WARNINGS
     }
 
     // main tile decode loop
-    bytesperpixel = s->bytesperpixel;
     memset(s->above_partition_ctx, 0, s->cols);
     memset(s->above_skip_ctx, 0, s->cols);
     if (s->s.h.keyframe || s->s.h.intraonly) {
@@ -1233,20 +1583,28 @@ FF_ENABLE_DEPRECATION_WARNINGS
         ff_thread_finish_setup(avctx);
     }
 
+    if (avctx->active_thread_type & FF_THREAD_SLICE) {
+        for (i = 0; i < s->sb_rows; i++)
+            atomic_store(&s->entries[i], 0);
+    }
+
     do {
-        yoff = uvoff = 0;
-        s->b = s->b_base;
-        s->block = s->block_base;
-        s->uvblock[0] = s->uvblock_base[0];
-        s->uvblock[1] = s->uvblock_base[1];
-        s->eob = s->eob_base;
-        s->uveob[0] = s->uveob_base[0];
-        s->uveob[1] = s->uveob_base[1];
-
-        for (tile_row = 0; tile_row < s->s.h.tiling.tile_rows; tile_row++) {
-            set_tile_offset(&s->tile_row_start, &s->tile_row_end,
-                            tile_row, s->s.h.tiling.log2_tile_rows, s->sb_rows);
-            if (s->pass != 2) {
+        for (i = 0; i < s->active_tile_cols; i++) {
+            s->td[i].b = s->td[i].b_base;
+            s->td[i].block = s->td[i].block_base;
+            s->td[i].uvblock[0] = s->td[i].uvblock_base[0];
+            s->td[i].uvblock[1] = s->td[i].uvblock_base[1];
+            s->td[i].eob = s->td[i].eob_base;
+            s->td[i].uveob[0] = s->td[i].uveob_base[0];
+            s->td[i].uveob[1] = s->td[i].uveob_base[1];
+        }
+
+        if (avctx->active_thread_type == FF_THREAD_SLICE) {
+            int tile_row, tile_col;
+
+            assert(!pass);
+
+            for (tile_row = 0; tile_row < s->s.h.tiling.tile_rows; tile_row++) {
                 for (tile_col = 0; tile_col < s->s.h.tiling.tile_cols; tile_col++) {
                     int64_t tile_size;
 
@@ -1258,105 +1616,31 @@ FF_ENABLE_DEPRECATION_WARNINGS
                         data += 4;
                         size -= 4;
                     }
-                    if (tile_size > size) {
-                        ff_thread_report_progress(&s->s.frames[CUR_FRAME].tf, INT_MAX, 0);
+                    if (tile_size > size)
                         return AVERROR_INVALIDDATA;
-                    }
-                    ret = ff_vp56_init_range_decoder(&s->c_b[tile_col], data, tile_size);
+                    ret = ff_vp56_init_range_decoder(&s->td[tile_col].c_b[tile_row], data, tile_size);
                     if (ret < 0)
                         return ret;
-                    if (vp56_rac_get_prob_branchy(&s->c_b[tile_col], 128)) { // marker bit
-                        ff_thread_report_progress(&s->s.frames[CUR_FRAME].tf, INT_MAX, 0);
+                    if (vp56_rac_get_prob_branchy(&s->td[tile_col].c_b[tile_row], 128)) // marker bit
                         return AVERROR_INVALIDDATA;
-                    }
                     data += tile_size;
                     size -= tile_size;
                 }
             }
 
-            for (row = s->tile_row_start; row < s->tile_row_end;
-                 row += 8, yoff += ls_y * 64, uvoff += ls_uv * 64 >> s->ss_v) {
-                VP9Filter *lflvl_ptr = s->lflvl;
-                ptrdiff_t yoff2 = yoff, uvoff2 = uvoff;
-
-                for (tile_col = 0; tile_col < s->s.h.tiling.tile_cols; tile_col++) {
-                    set_tile_offset(&s->tile_col_start, &s->tile_col_end,
-                                    tile_col, s->s.h.tiling.log2_tile_cols, s->sb_cols);
-
-                    if (s->pass != 2) {
-                        memset(s->left_partition_ctx, 0, 8);
-                        memset(s->left_skip_ctx, 0, 8);
-                        if (s->s.h.keyframe || s->s.h.intraonly) {
-                            memset(s->left_mode_ctx, DC_PRED, 16);
-                        } else {
-                            memset(s->left_mode_ctx, NEARESTMV, 8);
-                        }
-                        memset(s->left_y_nnz_ctx, 0, 16);
-                        memset(s->left_uv_nnz_ctx, 0, 32);
-                        memset(s->left_segpred_ctx, 0, 8);
-
-                        memcpy(&s->c, &s->c_b[tile_col], sizeof(s->c));
-                    }
-
-                    for (col = s->tile_col_start;
-                         col < s->tile_col_end;
-                         col += 8, yoff2 += 64 * bytesperpixel,
-                         uvoff2 += 64 * bytesperpixel >> s->ss_h, lflvl_ptr++) {
-                        // FIXME integrate with lf code (i.e. zero after each
-                        // use, similar to invtxfm coefficients, or similar)
-                        if (s->pass != 1) {
-                            memset(lflvl_ptr->mask, 0, sizeof(lflvl_ptr->mask));
-                        }
-
-                        if (s->pass == 2) {
-                            decode_sb_mem(avctx, row, col, lflvl_ptr,
-                                          yoff2, uvoff2, BL_64X64);
-                        } else {
-                            decode_sb(avctx, row, col, lflvl_ptr,
-                                      yoff2, uvoff2, BL_64X64);
-                        }
-                    }
-                    if (s->pass != 2)
-                        memcpy(&s->c_b[tile_col], &s->c, sizeof(s->c));
-                }
-
-                if (s->pass == 1)
-                    continue;
-
-                // backup pre-loopfilter reconstruction data for intra
-                // prediction of next row of sb64s
-                if (row + 8 < s->rows) {
-                    memcpy(s->intra_pred_data[0],
-                           f->data[0] + yoff + 63 * ls_y,
-                           8 * s->cols * bytesperpixel);
-                    memcpy(s->intra_pred_data[1],
-                           f->data[1] + uvoff + ((64 >> s->ss_v) - 1) * ls_uv,
-                           8 * s->cols * bytesperpixel >> s->ss_h);
-                    memcpy(s->intra_pred_data[2],
-                           f->data[2] + uvoff + ((64 >> s->ss_v) - 1) * ls_uv,
-                           8 * s->cols * bytesperpixel >> s->ss_h);
-                }
-
-                // loopfilter one row
-                if (s->s.h.filter.level) {
-                    yoff2 = yoff;
-                    uvoff2 = uvoff;
-                    lflvl_ptr = s->lflvl;
-                    for (col = 0; col < s->cols;
-                         col += 8, yoff2 += 64 * bytesperpixel,
-                         uvoff2 += 64 * bytesperpixel >> s->ss_h, lflvl_ptr++) {
-                        ff_vp9_loopfilter_sb(avctx, lflvl_ptr, row, col,
-                                             yoff2, uvoff2);
-                    }
-                }
-
-                // FIXME maybe we can make this more finegrained by running the
-                // loopfilter per-block instead of after each sbrow
-                // In fact that would also make intra pred left preparation easier?
-                ff_thread_report_progress(&s->s.frames[CUR_FRAME].tf, row >> 3, 0);
-            }
+            ff_slice_thread_execute_with_mainfunc(avctx, decode_tiles_mt, loopfilter_proc, s->td, NULL, s->s.h.tiling.tile_cols);
+        } else {
+            ret = decode_tiles(avctx, data, size);
+            if (ret < 0)
+                return ret;
         }
 
+        // Sum all counts fields into td[0].counts for tile threading
+        if (avctx->active_thread_type == FF_THREAD_SLICE)
+            for (i = 1; i < s->s.h.tiling.tile_cols; i++)
+                for (j = 0; j < sizeof(s->td[i].counts) / sizeof(unsigned); j++)
+                    ((unsigned *)&s->td[0].counts)[j] += ((unsigned *)&s->td[i].counts)[j];
+
         if (s->pass < 2 && s->s.h.refreshctx && !s->s.h.parallelmode) {
             ff_vp9_adapt_probs(s);
             ff_thread_finish_setup(avctx);
@@ -1492,7 +1776,8 @@ AVCodec ff_vp9_decoder = {
     .init                  = vp9_decode_init,
     .close                 = vp9_decode_free,
     .decode                = vp9_decode_frame,
-    .capabilities          = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_FRAME_THREADS,
+    .capabilities          = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_FRAME_THREADS | AV_CODEC_CAP_SLICE_THREADS,
+    .caps_internal         = FF_CODEC_CAP_SLICE_THREAD_HAS_MF,
     .flush                 = vp9_decode_flush,
     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(vp9_decode_init_thread_copy),
     .update_thread_context = ONLY_IF_THREADS_ENABLED(vp9_decode_update_thread_context),
diff --git a/libavcodec/vp9_mc_template.c b/libavcodec/vp9_mc_template.c
index 8ff654b..31e692f 100644
--- a/libavcodec/vp9_mc_template.c
+++ b/libavcodec/vp9_mc_template.c
@@ -27,19 +27,19 @@
     (VP56mv) { .x = ROUNDED_DIV(a.x + b.x + c.x + d.x, 4), \
                .y = ROUNDED_DIV(a.y + b.y + c.y + d.y, 4) }
 
-static void FN(inter_pred)(AVCodecContext *avctx)
+static void FN(inter_pred)(VP9TileData *td)
 {
     static const uint8_t bwlog_tab[2][N_BS_SIZES] = {
         { 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4 },
         { 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 4, 4 },
     };
-    VP9Context *s = avctx->priv_data;
-    VP9Block *b = s->b;
-    int row = s->row, col = s->col;
+    VP9Context *s = td->s;
+    VP9Block *b = td->b;
+    int row = td->row, col = td->col;
     ThreadFrame *tref1 = &s->s.refs[s->s.h.refidx[b->ref[0]]], *tref2;
     AVFrame *ref1 = tref1->f, *ref2;
     int w1 = ref1->width, h1 = ref1->height, w2, h2;
-    ptrdiff_t ls_y = s->y_stride, ls_uv = s->uv_stride;
+    ptrdiff_t ls_y = td->y_stride, ls_uv = td->uv_stride;
     int bytesperpixel = BYTES_PER_PIXEL;
 
     if (b->comp) {
@@ -55,26 +55,26 @@ static void FN(inter_pred)(AVCodecContext *avctx)
 
 #if SCALED == 0
         if (b->bs == BS_8x4) {
-            mc_luma_dir(s, mc[3][b->filter][0], s->dst[0], ls_y,
+            mc_luma_dir(td, mc[3][b->filter][0], td->dst[0], ls_y,
                         ref1->data[0], ref1->linesize[0], tref1,
                         row << 3, col << 3, &b->mv[0][0],,,,, 8, 4, w1, h1, 0);
-            mc_luma_dir(s, mc[3][b->filter][0],
-                        s->dst[0] + 4 * ls_y, ls_y,
+            mc_luma_dir(td, mc[3][b->filter][0],
+                        td->dst[0] + 4 * ls_y, ls_y,
                         ref1->data[0], ref1->linesize[0], tref1,
                         (row << 3) + 4, col << 3, &b->mv[2][0],,,,, 8, 4, w1, h1, 0);
             w1 = (w1 + s->ss_h) >> s->ss_h;
             if (s->ss_v) {
                 h1 = (h1 + 1) >> 1;
                 uvmv = ROUNDED_DIV_MVx2(b->mv[0][0], b->mv[2][0]);
-                mc_chroma_dir(s, mc[3 + s->ss_h][b->filter][0],
-                              s->dst[1], s->dst[2], ls_uv,
+                mc_chroma_dir(td, mc[3 + s->ss_h][b->filter][0],
+                              td->dst[1], td->dst[2], ls_uv,
                               ref1->data[1], ref1->linesize[1],
                               ref1->data[2], ref1->linesize[2], tref1,
                               row << 2, col << (3 - s->ss_h),
                               &uvmv,,,,, 8 >> s->ss_h, 4, w1, h1, 0);
             } else {
-                mc_chroma_dir(s, mc[3 + s->ss_h][b->filter][0],
-                              s->dst[1], s->dst[2], ls_uv,
+                mc_chroma_dir(td, mc[3 + s->ss_h][b->filter][0],
+                              td->dst[1], td->dst[2], ls_uv,
                               ref1->data[1], ref1->linesize[1],
                               ref1->data[2], ref1->linesize[2], tref1,
                               row << 3, col << (3 - s->ss_h),
@@ -87,8 +87,8 @@ static void FN(inter_pred)(AVCodecContext *avctx)
                 } else {
                     uvmv = ROUNDED_DIV_MVx2(b->mv[0][0], b->mv[2][0]);
                 }
-                mc_chroma_dir(s, mc[3 + s->ss_h][b->filter][0],
-                              s->dst[1] + 4 * ls_uv, s->dst[2] + 4 * ls_uv, ls_uv,
+                mc_chroma_dir(td, mc[3 + s->ss_h][b->filter][0],
+                              td->dst[1] + 4 * ls_uv, td->dst[2] + 4 * ls_uv, ls_uv,
                               ref1->data[1], ref1->linesize[1],
                               ref1->data[2], ref1->linesize[2], tref1,
                               (row << 3) + 4, col << (3 - s->ss_h),
@@ -96,26 +96,26 @@ static void FN(inter_pred)(AVCodecContext *avctx)
             }
 
             if (b->comp) {
-                mc_luma_dir(s, mc[3][b->filter][1], s->dst[0], ls_y,
+                mc_luma_dir(td, mc[3][b->filter][1], td->dst[0], ls_y,
                             ref2->data[0], ref2->linesize[0], tref2,
                             row << 3, col << 3, &b->mv[0][1],,,,, 8, 4, w2, h2, 1);
-                mc_luma_dir(s, mc[3][b->filter][1],
-                            s->dst[0] + 4 * ls_y, ls_y,
+                mc_luma_dir(td, mc[3][b->filter][1],
+                            td->dst[0] + 4 * ls_y, ls_y,
                             ref2->data[0], ref2->linesize[0], tref2,
                             (row << 3) + 4, col << 3, &b->mv[2][1],,,,, 8, 4, w2, h2, 1);
                 w2 = (w2 + s->ss_h) >> s->ss_h;
                 if (s->ss_v) {
                     h2 = (h2 + 1) >> 1;
                     uvmv = ROUNDED_DIV_MVx2(b->mv[0][1], b->mv[2][1]);
-                    mc_chroma_dir(s, mc[3 + s->ss_h][b->filter][1],
-                                  s->dst[1], s->dst[2], ls_uv,
+                    mc_chroma_dir(td, mc[3 + s->ss_h][b->filter][1],
+                                  td->dst[1], td->dst[2], ls_uv,
                                   ref2->data[1], ref2->linesize[1],
                                   ref2->data[2], ref2->linesize[2], tref2,
                                   row << 2, col << (3 - s->ss_h),
                                   &uvmv,,,,, 8 >> s->ss_h, 4, w2, h2, 1);
                 } else {
-                    mc_chroma_dir(s, mc[3 + s->ss_h][b->filter][1],
-                                  s->dst[1], s->dst[2], ls_uv,
+                    mc_chroma_dir(td, mc[3 + s->ss_h][b->filter][1],
+                                  td->dst[1], td->dst[2], ls_uv,
                                   ref2->data[1], ref2->linesize[1],
                                   ref2->data[2], ref2->linesize[2], tref2,
                                   row << 3, col << (3 - s->ss_h),
@@ -128,8 +128,8 @@ static void FN(inter_pred)(AVCodecContext *avctx)
                     } else {
                         uvmv = ROUNDED_DIV_MVx2(b->mv[0][1], b->mv[2][1]);
                     }
-                    mc_chroma_dir(s, mc[3 + s->ss_h][b->filter][1],
-                                  s->dst[1] + 4 * ls_uv, s->dst[2] + 4 * ls_uv, ls_uv,
+                    mc_chroma_dir(td, mc[3 + s->ss_h][b->filter][1],
+                                  td->dst[1] + 4 * ls_uv, td->dst[2] + 4 * ls_uv, ls_uv,
                                   ref2->data[1], ref2->linesize[1],
                                   ref2->data[2], ref2->linesize[2], tref2,
                                   (row << 3) + 4, col << (3 - s->ss_h),
@@ -137,32 +137,32 @@ static void FN(inter_pred)(AVCodecContext *avctx)
                 }
             }
         } else if (b->bs == BS_4x8) {
-            mc_luma_dir(s, mc[4][b->filter][0], s->dst[0], ls_y,
+            mc_luma_dir(td, mc[4][b->filter][0], td->dst[0], ls_y,
                         ref1->data[0], ref1->linesize[0], tref1,
                         row << 3, col << 3, &b->mv[0][0],,,,, 4, 8, w1, h1, 0);
-            mc_luma_dir(s, mc[4][b->filter][0], s->dst[0] + 4 * bytesperpixel, ls_y,
+            mc_luma_dir(td, mc[4][b->filter][0], td->dst[0] + 4 * bytesperpixel, ls_y,
                         ref1->data[0], ref1->linesize[0], tref1,
                         row << 3, (col << 3) + 4, &b->mv[1][0],,,,, 4, 8, w1, h1, 0);
             h1 = (h1 + s->ss_v) >> s->ss_v;
             if (s->ss_h) {
                 w1 = (w1 + 1) >> 1;
                 uvmv = ROUNDED_DIV_MVx2(b->mv[0][0], b->mv[1][0]);
-                mc_chroma_dir(s, mc[4][b->filter][0],
-                              s->dst[1], s->dst[2], ls_uv,
+                mc_chroma_dir(td, mc[4][b->filter][0],
+                              td->dst[1], td->dst[2], ls_uv,
                               ref1->data[1], ref1->linesize[1],
                               ref1->data[2], ref1->linesize[2], tref1,
                               row << (3 - s->ss_v), col << 2,
                               &uvmv,,,,, 4, 8 >> s->ss_v, w1, h1, 0);
             } else {
-                mc_chroma_dir(s, mc[4][b->filter][0],
-                              s->dst[1], s->dst[2], ls_uv,
+                mc_chroma_dir(td, mc[4][b->filter][0],
+                              td->dst[1], td->dst[2], ls_uv,
                               ref1->data[1], ref1->linesize[1],
                               ref1->data[2], ref1->linesize[2], tref1,
                               row << (3 - s->ss_v), col << 3,
                               &b->mv[0][0],,,,, 4, 8 >> s->ss_v, w1, h1, 0);
-                mc_chroma_dir(s, mc[4][b->filter][0],
-                              s->dst[1] + 4 * bytesperpixel,
-                              s->dst[2] + 4 * bytesperpixel, ls_uv,
+                mc_chroma_dir(td, mc[4][b->filter][0],
+                              td->dst[1] + 4 * bytesperpixel,
+                              td->dst[2] + 4 * bytesperpixel, ls_uv,
                               ref1->data[1], ref1->linesize[1],
                               ref1->data[2], ref1->linesize[2], tref1,
                               row << (3 - s->ss_v), (col << 3) + 4,
@@ -170,32 +170,32 @@ static void FN(inter_pred)(AVCodecContext *avctx)
             }
 
             if (b->comp) {
-                mc_luma_dir(s, mc[4][b->filter][1], s->dst[0], ls_y,
+                mc_luma_dir(td, mc[4][b->filter][1], td->dst[0], ls_y,
                             ref2->data[0], ref2->linesize[0], tref2,
                             row << 3, col << 3, &b->mv[0][1],,,,, 4, 8, w2, h2, 1);
-                mc_luma_dir(s, mc[4][b->filter][1], s->dst[0] + 4 * bytesperpixel, ls_y,
+                mc_luma_dir(td, mc[4][b->filter][1], td->dst[0] + 4 * bytesperpixel, ls_y,
                             ref2->data[0], ref2->linesize[0], tref2,
                             row << 3, (col << 3) + 4, &b->mv[1][1],,,,, 4, 8, w2, h2, 1);
                 h2 = (h2 + s->ss_v) >> s->ss_v;
                 if (s->ss_h) {
                     w2 = (w2 + 1) >> 1;
                     uvmv = ROUNDED_DIV_MVx2(b->mv[0][1], b->mv[1][1]);
-                    mc_chroma_dir(s, mc[4][b->filter][1],
-                                  s->dst[1], s->dst[2], ls_uv,
+                    mc_chroma_dir(td, mc[4][b->filter][1],
+                                  td->dst[1], td->dst[2], ls_uv,
                                   ref2->data[1], ref2->linesize[1],
                                   ref2->data[2], ref2->linesize[2], tref2,
                                   row << (3 - s->ss_v), col << 2,
                                   &uvmv,,,,, 4, 8 >> s->ss_v, w2, h2, 1);
                 } else {
-                    mc_chroma_dir(s, mc[4][b->filter][1],
-                                  s->dst[1], s->dst[2], ls_uv,
+                    mc_chroma_dir(td, mc[4][b->filter][1],
+                                  td->dst[1], td->dst[2], ls_uv,
                                   ref2->data[1], ref2->linesize[1],
                                   ref2->data[2], ref2->linesize[2], tref2,
                                   row << (3 - s->ss_v), col << 3,
                                   &b->mv[0][1],,,,, 4, 8 >> s->ss_v, w2, h2, 1);
-                    mc_chroma_dir(s, mc[4][b->filter][1],
-                                  s->dst[1] + 4 * bytesperpixel,
-                                  s->dst[2] + 4 * bytesperpixel, ls_uv,
+                    mc_chroma_dir(td, mc[4][b->filter][1],
+                                  td->dst[1] + 4 * bytesperpixel,
+                                  td->dst[2] + 4 * bytesperpixel, ls_uv,
                                   ref2->data[1], ref2->linesize[1],
                                   ref2->data[2], ref2->linesize[2], tref2,
                                   row << (3 - s->ss_v), (col << 3) + 4,
@@ -211,21 +211,21 @@ static void FN(inter_pred)(AVCodecContext *avctx)
 
             // FIXME if two horizontally adjacent blocks have the same MV,
             // do a w8 instead of a w4 call
-            mc_luma_dir(s, mc[4][b->filter][0], s->dst[0], ls_y,
+            mc_luma_dir(td, mc[4][b->filter][0], td->dst[0], ls_y,
                         ref1->data[0], ref1->linesize[0], tref1,
                         row << 3, col << 3, &b->mv[0][0],
                         0, 0, 8, 8, 4, 4, w1, h1, 0);
-            mc_luma_dir(s, mc[4][b->filter][0], s->dst[0] + 4 * bytesperpixel, ls_y,
+            mc_luma_dir(td, mc[4][b->filter][0], td->dst[0] + 4 * bytesperpixel, ls_y,
                         ref1->data[0], ref1->linesize[0], tref1,
                         row << 3, (col << 3) + 4, &b->mv[1][0],
                         4, 0, 8, 8, 4, 4, w1, h1, 0);
-            mc_luma_dir(s, mc[4][b->filter][0],
-                        s->dst[0] + 4 * ls_y, ls_y,
+            mc_luma_dir(td, mc[4][b->filter][0],
+                        td->dst[0] + 4 * ls_y, ls_y,
                         ref1->data[0], ref1->linesize[0], tref1,
                         (row << 3) + 4, col << 3, &b->mv[2][0],
                         0, 4, 8, 8, 4, 4, w1, h1, 0);
-            mc_luma_dir(s, mc[4][b->filter][0],
-                        s->dst[0] + 4 * ls_y + 4 * bytesperpixel, ls_y,
+            mc_luma_dir(td, mc[4][b->filter][0],
+                        td->dst[0] + 4 * ls_y + 4 * bytesperpixel, ls_y,
                         ref1->data[0], ref1->linesize[0], tref1,
                         (row << 3) + 4, (col << 3) + 4, &b->mv[3][0],
                         4, 4, 8, 8, 4, 4, w1, h1, 0);
@@ -235,24 +235,24 @@ static void FN(inter_pred)(AVCodecContext *avctx)
                     w1 = (w1 + 1) >> 1;
                     uvmv = ROUNDED_DIV_MVx4(b->mv[0][0], b->mv[1][0],
                                             b->mv[2][0], b->mv[3][0]);
-                    mc_chroma_dir(s, mc[4][b->filter][0],
-                                  s->dst[1], s->dst[2], ls_uv,
+                    mc_chroma_dir(td, mc[4][b->filter][0],
+                                  td->dst[1], td->dst[2], ls_uv,
                                   ref1->data[1], ref1->linesize[1],
                                   ref1->data[2], ref1->linesize[2], tref1,
                                   row << 2, col << 2,
                                   &uvmv, 0, 0, 4, 4, 4, 4, w1, h1, 0);
                 } else {
                     uvmv = ROUNDED_DIV_MVx2(b->mv[0][0], b->mv[2][0]);
-                    mc_chroma_dir(s, mc[4][b->filter][0],
-                                  s->dst[1], s->dst[2], ls_uv,
+                    mc_chroma_dir(td, mc[4][b->filter][0],
+                                  td->dst[1], td->dst[2], ls_uv,
                                   ref1->data[1], ref1->linesize[1],
                                   ref1->data[2], ref1->linesize[2], tref1,
                                   row << 2, col << 3,
                                   &uvmv, 0, 0, 8, 4, 4, 4, w1, h1, 0);
                     uvmv = ROUNDED_DIV_MVx2(b->mv[1][0], b->mv[3][0]);
-                    mc_chroma_dir(s, mc[4][b->filter][0],
-                                  s->dst[1] + 4 * bytesperpixel,
-                                  s->dst[2] + 4 * bytesperpixel, ls_uv,
+                    mc_chroma_dir(td, mc[4][b->filter][0],
+                                  td->dst[1] + 4 * bytesperpixel,
+                                  td->dst[2] + 4 * bytesperpixel, ls_uv,
                                   ref1->data[1], ref1->linesize[1],
                                   ref1->data[2], ref1->linesize[2], tref1,
                                   row << 2, (col << 3) + 4,
@@ -262,8 +262,8 @@ static void FN(inter_pred)(AVCodecContext *avctx)
                 if (s->ss_h) {
                     w1 = (w1 + 1) >> 1;
                     uvmv = ROUNDED_DIV_MVx2(b->mv[0][0], b->mv[1][0]);
-                    mc_chroma_dir(s, mc[4][b->filter][0],
-                                  s->dst[1], s->dst[2], ls_uv,
+                    mc_chroma_dir(td, mc[4][b->filter][0],
+                                  td->dst[1], td->dst[2], ls_uv,
                                   ref1->data[1], ref1->linesize[1],
                                   ref1->data[2], ref1->linesize[2], tref1,
                                   row << 3, col << 2,
@@ -272,35 +272,35 @@ static void FN(inter_pred)(AVCodecContext *avctx)
                     // bottom block
                     // https://code.google.com/p/webm/issues/detail?id=993
                     uvmv = ROUNDED_DIV_MVx2(b->mv[1][0], b->mv[2][0]);
-                    mc_chroma_dir(s, mc[4][b->filter][0],
-                                  s->dst[1] + 4 * ls_uv, s->dst[2] + 4 * ls_uv, ls_uv,
+                    mc_chroma_dir(td, mc[4][b->filter][0],
+                                  td->dst[1] + 4 * ls_uv, td->dst[2] + 4 * ls_uv, ls_uv,
                                   ref1->data[1], ref1->linesize[1],
                                   ref1->data[2], ref1->linesize[2], tref1,
                                   (row << 3) + 4, col << 2,
                                   &uvmv, 0, 4, 4, 8, 4, 4, w1, h1, 0);
                 } else {
-                    mc_chroma_dir(s, mc[4][b->filter][0],
-                                  s->dst[1], s->dst[2], ls_uv,
+                    mc_chroma_dir(td, mc[4][b->filter][0],
+                                  td->dst[1], td->dst[2], ls_uv,
                                   ref1->data[1], ref1->linesize[1],
                                   ref1->data[2], ref1->linesize[2], tref1,
                                   row << 3, col << 3,
                                   &b->mv[0][0], 0, 0, 8, 8, 4, 4, w1, h1, 0);
-                    mc_chroma_dir(s, mc[4][b->filter][0],
-                                  s->dst[1] + 4 * bytesperpixel,
-                                  s->dst[2] + 4 * bytesperpixel, ls_uv,
+                    mc_chroma_dir(td, mc[4][b->filter][0],
+                                  td->dst[1] + 4 * bytesperpixel,
+                                  td->dst[2] + 4 * bytesperpixel, ls_uv,
                                   ref1->data[1], ref1->linesize[1],
                                   ref1->data[2], ref1->linesize[2], tref1,
                                   row << 3, (col << 3) + 4,
                                   &b->mv[1][0], 4, 0, 8, 8, 4, 4, w1, h1, 0);
-                    mc_chroma_dir(s, mc[4][b->filter][0],
-                                  s->dst[1] + 4 * ls_uv, s->dst[2] + 4 * ls_uv, ls_uv,
+                    mc_chroma_dir(td, mc[4][b->filter][0],
+                                  td->dst[1] + 4 * ls_uv, td->dst[2] + 4 * ls_uv, ls_uv,
                                   ref1->data[1], ref1->linesize[1],
                                   ref1->data[2], ref1->linesize[2], tref1,
                                   (row << 3) + 4, col << 3,
                                   &b->mv[2][0], 0, 4, 8, 8, 4, 4, w1, h1, 0);
-                    mc_chroma_dir(s, mc[4][b->filter][0],
-                                  s->dst[1] + 4 * ls_uv + 4 * bytesperpixel,
-                                  s->dst[2] + 4 * ls_uv + 4 * bytesperpixel, ls_uv,
+                    mc_chroma_dir(td, mc[4][b->filter][0],
+                                  td->dst[1] + 4 * ls_uv + 4 * bytesperpixel,
+                                  td->dst[2] + 4 * ls_uv + 4 * bytesperpixel, ls_uv,
                                   ref1->data[1], ref1->linesize[1],
                                   ref1->data[2], ref1->linesize[2], tref1,
                                   (row << 3) + 4, (col << 3) + 4,
@@ -309,18 +309,18 @@ static void FN(inter_pred)(AVCodecContext *avctx)
             }
 
             if (b->comp) {
-                mc_luma_dir(s, mc[4][b->filter][1], s->dst[0], ls_y,
+                mc_luma_dir(td, mc[4][b->filter][1], td->dst[0], ls_y,
                             ref2->data[0], ref2->linesize[0], tref2,
                             row << 3, col << 3, &b->mv[0][1], 0, 0, 8, 8, 4, 4, w2, h2, 1);
-                mc_luma_dir(s, mc[4][b->filter][1], s->dst[0] + 4 * bytesperpixel, ls_y,
+                mc_luma_dir(td, mc[4][b->filter][1], td->dst[0] + 4 * bytesperpixel, ls_y,
                             ref2->data[0], ref2->linesize[0], tref2,
                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 0, 8, 8, 4, 4, w2, h2, 1);
-                mc_luma_dir(s, mc[4][b->filter][1],
-                            s->dst[0] + 4 * ls_y, ls_y,
+                mc_luma_dir(td, mc[4][b->filter][1],
+                            td->dst[0] + 4 * ls_y, ls_y,
                             ref2->data[0], ref2->linesize[0], tref2,
                             (row << 3) + 4, col << 3, &b->mv[2][1], 0, 4, 8, 8, 4, 4, w2, h2, 1);
-                mc_luma_dir(s, mc[4][b->filter][1],
-                            s->dst[0] + 4 * ls_y + 4 * bytesperpixel, ls_y,
+                mc_luma_dir(td, mc[4][b->filter][1],
+                            td->dst[0] + 4 * ls_y + 4 * bytesperpixel, ls_y,
                             ref2->data[0], ref2->linesize[0], tref2,
                             (row << 3) + 4, (col << 3) + 4, &b->mv[3][1], 4, 4, 8, 8, 4, 4, w2, h2, 1);
                 if (s->ss_v) {
@@ -329,24 +329,24 @@ static void FN(inter_pred)(AVCodecContext *avctx)
                         w2 = (w2 + 1) >> 1;
                         uvmv = ROUNDED_DIV_MVx4(b->mv[0][1], b->mv[1][1],
                                                 b->mv[2][1], b->mv[3][1]);
-                        mc_chroma_dir(s, mc[4][b->filter][1],
-                                      s->dst[1], s->dst[2], ls_uv,
+                        mc_chroma_dir(td, mc[4][b->filter][1],
+                                      td->dst[1], td->dst[2], ls_uv,
                                       ref2->data[1], ref2->linesize[1],
                                       ref2->data[2], ref2->linesize[2], tref2,
                                       row << 2, col << 2,
                                       &uvmv, 0, 0, 4, 4, 4, 4, w2, h2, 1);
                     } else {
                         uvmv = ROUNDED_DIV_MVx2(b->mv[0][1], b->mv[2][1]);
-                        mc_chroma_dir(s, mc[4][b->filter][1],
-                                      s->dst[1], s->dst[2], ls_uv,
+                        mc_chroma_dir(td, mc[4][b->filter][1],
+                                      td->dst[1], td->dst[2], ls_uv,
                                       ref2->data[1], ref2->linesize[1],
                                       ref2->data[2], ref2->linesize[2], tref2,
                                       row << 2, col << 3,
                                       &uvmv, 0, 0, 8, 4, 4, 4, w2, h2, 1);
                         uvmv = ROUNDED_DIV_MVx2(b->mv[1][1], b->mv[3][1]);
-                        mc_chroma_dir(s, mc[4][b->filter][1],
-                                      s->dst[1] + 4 * bytesperpixel,
-                                      s->dst[2] + 4 * bytesperpixel, ls_uv,
+                        mc_chroma_dir(td, mc[4][b->filter][1],
+                                      td->dst[1] + 4 * bytesperpixel,
+                                      td->dst[2] + 4 * bytesperpixel, ls_uv,
                                       ref2->data[1], ref2->linesize[1],
                                       ref2->data[2], ref2->linesize[2], tref2,
                                       row << 2, (col << 3) + 4,
@@ -356,8 +356,8 @@ static void FN(inter_pred)(AVCodecContext *avctx)
                     if (s->ss_h) {
                         w2 = (w2 + 1) >> 1;
                         uvmv = ROUNDED_DIV_MVx2(b->mv[0][1], b->mv[1][1]);
-                        mc_chroma_dir(s, mc[4][b->filter][1],
-                                      s->dst[1], s->dst[2], ls_uv,
+                        mc_chroma_dir(td, mc[4][b->filter][1],
+                                      td->dst[1], td->dst[2], ls_uv,
                                       ref2->data[1], ref2->linesize[1],
                                       ref2->data[2], ref2->linesize[2], tref2,
                                       row << 3, col << 2,
@@ -366,35 +366,35 @@ static void FN(inter_pred)(AVCodecContext *avctx)
                         // bottom block
                         // https://code.google.com/p/webm/issues/detail?id=993
                         uvmv = ROUNDED_DIV_MVx2(b->mv[1][1], b->mv[2][1]);
-                        mc_chroma_dir(s, mc[4][b->filter][1],
-                                      s->dst[1] + 4 * ls_uv, s->dst[2] + 4 * ls_uv, ls_uv,
+                        mc_chroma_dir(td, mc[4][b->filter][1],
+                                      td->dst[1] + 4 * ls_uv, td->dst[2] + 4 * ls_uv, ls_uv,
                                       ref2->data[1], ref2->linesize[1],
                                       ref2->data[2], ref2->linesize[2], tref2,
                                       (row << 3) + 4, col << 2,
                                       &uvmv, 0, 4, 4, 8, 4, 4, w2, h2, 1);
                     } else {
-                        mc_chroma_dir(s, mc[4][b->filter][1],
-                                      s->dst[1], s->dst[2], ls_uv,
+                        mc_chroma_dir(td, mc[4][b->filter][1],
+                                      td->dst[1], td->dst[2], ls_uv,
                                       ref2->data[1], ref2->linesize[1],
                                       ref2->data[2], ref2->linesize[2], tref2,
                                       row << 3, col << 3,
                                       &b->mv[0][1], 0, 0, 8, 8, 4, 4, w2, h2, 1);
-                        mc_chroma_dir(s, mc[4][b->filter][1],
-                                      s->dst[1] + 4 * bytesperpixel,
-                                      s->dst[2] + 4 * bytesperpixel, ls_uv,
+                        mc_chroma_dir(td, mc[4][b->filter][1],
+                                      td->dst[1] + 4 * bytesperpixel,
+                                      td->dst[2] + 4 * bytesperpixel, ls_uv,
                                       ref2->data[1], ref2->linesize[1],
                                       ref2->data[2], ref2->linesize[2], tref2,
                                       row << 3, (col << 3) + 4,
                                       &b->mv[1][1], 4, 0, 8, 8, 4, 4, w2, h2, 1);
-                        mc_chroma_dir(s, mc[4][b->filter][1],
-                                      s->dst[1] + 4 * ls_uv, s->dst[2] + 4 * ls_uv, ls_uv,
+                        mc_chroma_dir(td, mc[4][b->filter][1],
+                                      td->dst[1] + 4 * ls_uv, td->dst[2] + 4 * ls_uv, ls_uv,
                                       ref2->data[1], ref2->linesize[1],
                                       ref2->data[2], ref2->linesize[2], tref2,
                                       (row << 3) + 4, col << 3,
                                       &b->mv[2][1], 0, 4, 8, 8, 4, 4, w2, h2, 1);
-                        mc_chroma_dir(s, mc[4][b->filter][1],
-                                      s->dst[1] + 4 * ls_uv + 4 * bytesperpixel,
-                                      s->dst[2] + 4 * ls_uv + 4 * bytesperpixel, ls_uv,
+                        mc_chroma_dir(td, mc[4][b->filter][1],
+                                      td->dst[1] + 4 * ls_uv + 4 * bytesperpixel,
+                                      td->dst[2] + 4 * ls_uv + 4 * bytesperpixel, ls_uv,
                                       ref2->data[1], ref2->linesize[1],
                                       ref2->data[2], ref2->linesize[2], tref2,
                                       (row << 3) + 4, (col << 3) + 4,
@@ -410,26 +410,26 @@ static void FN(inter_pred)(AVCodecContext *avctx)
         int uvbw = ff_vp9_bwh_tab[s->ss_h][b->bs][0] * 4;
         int uvbh = ff_vp9_bwh_tab[s->ss_v][b->bs][1] * 4;
 
-        mc_luma_dir(s, mc[bwl][b->filter][0], s->dst[0], ls_y,
+        mc_luma_dir(td, mc[bwl][b->filter][0], td->dst[0], ls_y,
                     ref1->data[0], ref1->linesize[0], tref1,
                     row << 3, col << 3, &b->mv[0][0], 0, 0, bw, bh, bw, bh, w1, h1, 0);
         w1 = (w1 + s->ss_h) >> s->ss_h;
         h1 = (h1 + s->ss_v) >> s->ss_v;
-        mc_chroma_dir(s, mc[bwl + s->ss_h][b->filter][0],
-                      s->dst[1], s->dst[2], ls_uv,
+        mc_chroma_dir(td, mc[bwl + s->ss_h][b->filter][0],
+                      td->dst[1], td->dst[2], ls_uv,
                       ref1->data[1], ref1->linesize[1],
                       ref1->data[2], ref1->linesize[2], tref1,
                       row << (3 - s->ss_v), col << (3 - s->ss_h),
                       &b->mv[0][0], 0, 0, uvbw, uvbh, uvbw, uvbh, w1, h1, 0);
 
         if (b->comp) {
-            mc_luma_dir(s, mc[bwl][b->filter][1], s->dst[0], ls_y,
+            mc_luma_dir(td, mc[bwl][b->filter][1], td->dst[0], ls_y,
                         ref2->data[0], ref2->linesize[0], tref2,
                         row << 3, col << 3, &b->mv[0][1], 0, 0, bw, bh, bw, bh, w2, h2, 1);
             w2 = (w2 + s->ss_h) >> s->ss_h;
             h2 = (h2 + s->ss_v) >> s->ss_v;
-            mc_chroma_dir(s, mc[bwl + s->ss_h][b->filter][1],
-                          s->dst[1], s->dst[2], ls_uv,
+            mc_chroma_dir(td, mc[bwl + s->ss_h][b->filter][1],
+                          td->dst[1], td->dst[2], ls_uv,
                           ref2->data[1], ref2->linesize[1],
                           ref2->data[2], ref2->linesize[2], tref2,
                           row << (3 - s->ss_v), col << (3 - s->ss_h),
diff --git a/libavcodec/vp9block.c b/libavcodec/vp9block.c
index a16ccdc..d4ace14 100644
--- a/libavcodec/vp9block.c
+++ b/libavcodec/vp9block.c
@@ -77,7 +77,7 @@ static av_always_inline void setctx_2d(uint8_t *ptr, int w, int h,
     }
 }
 
-static void decode_mode(AVCodecContext *avctx)
+static void decode_mode(VP9TileData *td)
 {
     static const uint8_t left_ctx[N_BS_SIZES] = {
         0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf
@@ -89,25 +89,25 @@ static void decode_mode(AVCodecContext *avctx)
         TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,
         TX_16X16, TX_8X8,   TX_8X8,   TX_8X8,   TX_4X4,   TX_4X4,  TX_4X4
     };
-    VP9Context *s = avctx->priv_data;
-    VP9Block *b = s->b;
-    int row = s->row, col = s->col, row7 = s->row7;
+    VP9Context *s = td->s;
+    VP9Block *b = td->b;
+    int row = td->row, col = td->col, row7 = td->row7;
     enum TxfmMode max_tx = max_tx_for_bl_bp[b->bs];
     int bw4 = ff_vp9_bwh_tab[1][b->bs][0], w4 = FFMIN(s->cols - col, bw4);
     int bh4 = ff_vp9_bwh_tab[1][b->bs][1], h4 = FFMIN(s->rows - row, bh4), y;
-    int have_a = row > 0, have_l = col > s->tile_col_start;
+    int have_a = row > 0, have_l = col > td->tile_col_start;
     int vref, filter_id;
 
     if (!s->s.h.segmentation.enabled) {
         b->seg_id = 0;
     } else if (s->s.h.keyframe || s->s.h.intraonly) {
         b->seg_id = !s->s.h.segmentation.update_map ? 0 :
-                    vp8_rac_get_tree(&s->c, ff_vp9_segmentation_tree, s->s.h.segmentation.prob);
+                    vp8_rac_get_tree(td->c, ff_vp9_segmentation_tree, s->s.h.segmentation.prob);
     } else if (!s->s.h.segmentation.update_map ||
                (s->s.h.segmentation.temporal &&
-                vp56_rac_get_prob_branchy(&s->c,
+                vp56_rac_get_prob_branchy(td->c,
                     s->s.h.segmentation.pred_prob[s->above_segpred_ctx[col] +
-                                    s->left_segpred_ctx[row7]]))) {
+                                    td->left_segpred_ctx[row7]]))) {
         if (!s->s.h.errorres && s->s.frames[REF_FRAME_SEGMAP].segmentation_map) {
             int pred = 8, x;
             uint8_t *refsegmap = s->s.frames[REF_FRAME_SEGMAP].segmentation_map;
@@ -126,13 +126,13 @@ static void decode_mode(AVCodecContext *avctx)
         }
 
         memset(&s->above_segpred_ctx[col], 1, w4);
-        memset(&s->left_segpred_ctx[row7], 1, h4);
+        memset(&td->left_segpred_ctx[row7], 1, h4);
     } else {
-        b->seg_id = vp8_rac_get_tree(&s->c, ff_vp9_segmentation_tree,
+        b->seg_id = vp8_rac_get_tree(td->c, ff_vp9_segmentation_tree,
                                      s->s.h.segmentation.prob);
 
         memset(&s->above_segpred_ctx[col], 0, w4);
-        memset(&s->left_segpred_ctx[row7], 0, h4);
+        memset(&td->left_segpred_ctx[row7], 0, h4);
     }
     if (s->s.h.segmentation.enabled &&
         (s->s.h.segmentation.update_map || s->s.h.keyframe || s->s.h.intraonly)) {
@@ -143,9 +143,9 @@ static void decode_mode(AVCodecContext *avctx)
     b->skip = s->s.h.segmentation.enabled &&
         s->s.h.segmentation.feat[b->seg_id].skip_enabled;
     if (!b->skip) {
-        int c = s->left_skip_ctx[row7] + s->above_skip_ctx[col];
-        b->skip = vp56_rac_get_prob(&s->c, s->prob.p.skip[c]);
-        s->counts.skip[c][b->skip]++;
+        int c = td->left_skip_ctx[row7] + s->above_skip_ctx[col];
+        b->skip = vp56_rac_get_prob(td->c, s->prob.p.skip[c]);
+        td->counts.skip[c][b->skip]++;
     }
 
     if (s->s.h.keyframe || s->s.h.intraonly) {
@@ -156,14 +156,14 @@ static void decode_mode(AVCodecContext *avctx)
         int c, bit;
 
         if (have_a && have_l) {
-            c = s->above_intra_ctx[col] + s->left_intra_ctx[row7];
+            c = s->above_intra_ctx[col] + td->left_intra_ctx[row7];
             c += (c == 2);
         } else {
             c = have_a ? 2 * s->above_intra_ctx[col] :
-                have_l ? 2 * s->left_intra_ctx[row7] : 0;
+                have_l ? 2 * td->left_intra_ctx[row7] : 0;
         }
-        bit = vp56_rac_get_prob(&s->c, s->prob.p.intra[c]);
-        s->counts.intra[c][bit]++;
+        bit = vp56_rac_get_prob(td->c, s->prob.p.intra[c]);
+        td->counts.intra[c][bit]++;
         b->intra = !bit;
     }
 
@@ -173,37 +173,37 @@ static void decode_mode(AVCodecContext *avctx)
             if (have_l) {
                 c = (s->above_skip_ctx[col] ? max_tx :
                      s->above_txfm_ctx[col]) +
-                    (s->left_skip_ctx[row7] ? max_tx :
-                     s->left_txfm_ctx[row7]) > max_tx;
+                    (td->left_skip_ctx[row7] ? max_tx :
+                     td->left_txfm_ctx[row7]) > max_tx;
             } else {
                 c = s->above_skip_ctx[col] ? 1 :
                     (s->above_txfm_ctx[col] * 2 > max_tx);
             }
         } else if (have_l) {
-            c = s->left_skip_ctx[row7] ? 1 :
-                (s->left_txfm_ctx[row7] * 2 > max_tx);
+            c = td->left_skip_ctx[row7] ? 1 :
+                (td->left_txfm_ctx[row7] * 2 > max_tx);
         } else {
             c = 1;
         }
         switch (max_tx) {
         case TX_32X32:
-            b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][0]);
+            b->tx = vp56_rac_get_prob(td->c, s->prob.p.tx32p[c][0]);
             if (b->tx) {
-                b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][1]);
+                b->tx += vp56_rac_get_prob(td->c, s->prob.p.tx32p[c][1]);
                 if (b->tx == 2)
-                    b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][2]);
+                    b->tx += vp56_rac_get_prob(td->c, s->prob.p.tx32p[c][2]);
             }
-            s->counts.tx32p[c][b->tx]++;
+            td->counts.tx32p[c][b->tx]++;
             break;
         case TX_16X16:
-            b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][0]);
+            b->tx = vp56_rac_get_prob(td->c, s->prob.p.tx16p[c][0]);
             if (b->tx)
-                b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][1]);
-            s->counts.tx16p[c][b->tx]++;
+                b->tx += vp56_rac_get_prob(td->c, s->prob.p.tx16p[c][1]);
+            td->counts.tx16p[c][b->tx]++;
             break;
         case TX_8X8:
-            b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx8p[c]);
-            s->counts.tx8p[c][b->tx]++;
+            b->tx = vp56_rac_get_prob(td->c, s->prob.p.tx8p[c]);
+            td->counts.tx8p[c][b->tx]++;
             break;
         case TX_4X4:
             b->tx = TX_4X4;
@@ -215,7 +215,7 @@ static void decode_mode(AVCodecContext *avctx)
 
     if (s->s.h.keyframe || s->s.h.intraonly) {
         uint8_t *a = &s->above_mode_ctx[col * 2];
-        uint8_t *l = &s->left_mode_ctx[(row7) << 1];
+        uint8_t *l = &td->left_mode_ctx[(row7) << 1];
 
         b->comp = 0;
         if (b->bs > BS_8x8) {
@@ -223,10 +223,10 @@ static void decode_mode(AVCodecContext *avctx)
             // necessary, they're just there to make the code slightly
             // simpler for now
             b->mode[0] =
-            a[0]       = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
+            a[0]       = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
                                           ff_vp9_default_kf_ymode_probs[a[0]][l[0]]);
             if (b->bs != BS_8x4) {
-                b->mode[1] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
+                b->mode[1] = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
                                               ff_vp9_default_kf_ymode_probs[a[1]][b->mode[0]]);
                 l[0]       =
                 a[1]       = b->mode[1];
@@ -237,10 +237,10 @@ static void decode_mode(AVCodecContext *avctx)
             }
             if (b->bs != BS_4x8) {
                 b->mode[2] =
-                a[0]       = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
+                a[0]       = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
                                               ff_vp9_default_kf_ymode_probs[a[0]][l[1]]);
                 if (b->bs != BS_8x4) {
-                    b->mode[3] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
+                    b->mode[3] = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
                                                   ff_vp9_default_kf_ymode_probs[a[1]][b->mode[2]]);
                     l[1]       =
                     a[1]       = b->mode[3];
@@ -256,7 +256,7 @@ static void decode_mode(AVCodecContext *avctx)
                 b->mode[3] = b->mode[1];
             }
         } else {
-            b->mode[0] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
+            b->mode[0] = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
                                           ff_vp9_default_kf_ymode_probs[*a][*l]);
             b->mode[3] =
             b->mode[2] =
@@ -265,29 +265,29 @@ static void decode_mode(AVCodecContext *avctx)
             memset(a, b->mode[0], ff_vp9_bwh_tab[0][b->bs][0]);
             memset(l, b->mode[0], ff_vp9_bwh_tab[0][b->bs][1]);
         }
-        b->uvmode = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
+        b->uvmode = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
                                      ff_vp9_default_kf_uvmode_probs[b->mode[3]]);
     } else if (b->intra) {
         b->comp = 0;
         if (b->bs > BS_8x8) {
-            b->mode[0] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
+            b->mode[0] = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
                                           s->prob.p.y_mode[0]);
-            s->counts.y_mode[0][b->mode[0]]++;
+            td->counts.y_mode[0][b->mode[0]]++;
             if (b->bs != BS_8x4) {
-                b->mode[1] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
+                b->mode[1] = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
                                               s->prob.p.y_mode[0]);
-                s->counts.y_mode[0][b->mode[1]]++;
+                td->counts.y_mode[0][b->mode[1]]++;
             } else {
                 b->mode[1] = b->mode[0];
             }
             if (b->bs != BS_4x8) {
-                b->mode[2] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
+                b->mode[2] = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
                                               s->prob.p.y_mode[0]);
-                s->counts.y_mode[0][b->mode[2]]++;
+                td->counts.y_mode[0][b->mode[2]]++;
                 if (b->bs != BS_8x4) {
-                    b->mode[3] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
+                    b->mode[3] = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
                                                   s->prob.p.y_mode[0]);
-                    s->counts.y_mode[0][b->mode[3]]++;
+                    td->counts.y_mode[0][b->mode[3]]++;
                 } else {
                     b->mode[3] = b->mode[2];
                 }
@@ -301,16 +301,16 @@ static void decode_mode(AVCodecContext *avctx)
             };
             int sz = size_group[b->bs];
 
-            b->mode[0] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
+            b->mode[0] = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
                                           s->prob.p.y_mode[sz]);
             b->mode[1] =
             b->mode[2] =
             b->mode[3] = b->mode[0];
-            s->counts.y_mode[sz][b->mode[3]]++;
+            td->counts.y_mode[sz][b->mode[3]]++;
         }
-        b->uvmode = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,
+        b->uvmode = vp8_rac_get_tree(td->c, ff_vp9_intramode_tree,
                                      s->prob.p.uv_mode[b->mode[3]]);
-        s->counts.uv_mode[b->mode[3]][b->uvmode]++;
+        td->counts.uv_mode[b->mode[3]][b->uvmode]++;
     } else {
         static const uint8_t inter_mode_ctx_lut[14][14] = {
             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },
@@ -343,32 +343,32 @@ static void decode_mode(AVCodecContext *avctx)
                 // FIXME add intra as ref=0xff (or -1) to make these easier?
                 if (have_a) {
                     if (have_l) {
-                        if (s->above_comp_ctx[col] && s->left_comp_ctx[row7]) {
+                        if (s->above_comp_ctx[col] && td->left_comp_ctx[row7]) {
                             c = 4;
                         } else if (s->above_comp_ctx[col]) {
-                            c = 2 + (s->left_intra_ctx[row7] ||
-                                     s->left_ref_ctx[row7] == s->s.h.fixcompref);
-                        } else if (s->left_comp_ctx[row7]) {
+                            c = 2 + (td->left_intra_ctx[row7] ||
+                                     td->left_ref_ctx[row7] == s->s.h.fixcompref);
+                        } else if (td->left_comp_ctx[row7]) {
                             c = 2 + (s->above_intra_ctx[col] ||
                                      s->above_ref_ctx[col] == s->s.h.fixcompref);
                         } else {
                             c = (!s->above_intra_ctx[col] &&
                                  s->above_ref_ctx[col] == s->s.h.fixcompref) ^
-                                (!s->left_intra_ctx[row7] &&
-                                 s->left_ref_ctx[row & 7] == s->s.h.fixcompref);
+                                (!td->left_intra_ctx[row7] &&
+                                 td->left_ref_ctx[row & 7] == s->s.h.fixcompref);
                         }
                     } else {
                         c = s->above_comp_ctx[col] ? 3 :
                         (!s->above_intra_ctx[col] && s->above_ref_ctx[col] == s->s.h.fixcompref);
                     }
                 } else if (have_l) {
-                    c = s->left_comp_ctx[row7] ? 3 :
-                    (!s->left_intra_ctx[row7] && s->left_ref_ctx[row7] == s->s.h.fixcompref);
+                    c = td->left_comp_ctx[row7] ? 3 :
+                    (!td->left_intra_ctx[row7] && td->left_ref_ctx[row7] == s->s.h.fixcompref);
                 } else {
                     c = 1;
                 }
-                b->comp = vp56_rac_get_prob(&s->c, s->prob.p.comp[c]);
-                s->counts.comp[c][b->comp]++;
+                b->comp = vp56_rac_get_prob(td->c, s->prob.p.comp[c]);
+                td->counts.comp[c][b->comp]++;
             }
 
             // read actual references
@@ -382,26 +382,26 @@ static void decode_mode(AVCodecContext *avctx)
                 if (have_a) {
                     if (have_l) {
                         if (s->above_intra_ctx[col]) {
-                            if (s->left_intra_ctx[row7]) {
+                            if (td->left_intra_ctx[row7]) {
                                 c = 2;
                             } else {
-                                c = 1 + 2 * (s->left_ref_ctx[row7] != s->s.h.varcompref[1]);
+                                c = 1 + 2 * (td->left_ref_ctx[row7] != s->s.h.varcompref[1]);
                             }
-                        } else if (s->left_intra_ctx[row7]) {
+                        } else if (td->left_intra_ctx[row7]) {
                             c = 1 + 2 * (s->above_ref_ctx[col] != s->s.h.varcompref[1]);
                         } else {
-                            int refl = s->left_ref_ctx[row7], refa = s->above_ref_ctx[col];
+                            int refl = td->left_ref_ctx[row7], refa = s->above_ref_ctx[col];
 
                             if (refl == refa && refa == s->s.h.varcompref[1]) {
                                 c = 0;
-                            } else if (!s->left_comp_ctx[row7] && !s->above_comp_ctx[col]) {
+                            } else if (!td->left_comp_ctx[row7] && !s->above_comp_ctx[col]) {
                                 if ((refa == s->s.h.fixcompref && refl == s->s.h.varcompref[0]) ||
                                     (refl == s->s.h.fixcompref && refa == s->s.h.varcompref[0])) {
                                     c = 4;
                                 } else {
                                     c = (refa == refl) ? 3 : 1;
                                 }
-                            } else if (!s->left_comp_ctx[row7]) {
+                            } else if (!td->left_comp_ctx[row7]) {
                                 if (refa == s->s.h.varcompref[1] && refl != s->s.h.varcompref[1]) {
                                     c = 1;
                                 } else {
@@ -429,37 +429,37 @@ static void decode_mode(AVCodecContext *avctx)
                         }
                     }
                 } else if (have_l) {
-                    if (s->left_intra_ctx[row7]) {
+                    if (td->left_intra_ctx[row7]) {
                         c = 2;
-                    } else if (s->left_comp_ctx[row7]) {
-                        c = 4 * (s->left_ref_ctx[row7] != s->s.h.varcompref[1]);
+                    } else if (td->left_comp_ctx[row7]) {
+                        c = 4 * (td->left_ref_ctx[row7] != s->s.h.varcompref[1]);
                     } else {
-                        c = 3 * (s->left_ref_ctx[row7] != s->s.h.varcompref[1]);
+                        c = 3 * (td->left_ref_ctx[row7] != s->s.h.varcompref[1]);
                     }
                 } else {
                     c = 2;
                 }
-                bit = vp56_rac_get_prob(&s->c, s->prob.p.comp_ref[c]);
+                bit = vp56_rac_get_prob(td->c, s->prob.p.comp_ref[c]);
                 b->ref[var_idx] = s->s.h.varcompref[bit];
-                s->counts.comp_ref[c][bit]++;
+                td->counts.comp_ref[c][bit]++;
             } else /* single reference */ {
                 int bit, c;
 
                 if (have_a && !s->above_intra_ctx[col]) {
-                    if (have_l && !s->left_intra_ctx[row7]) {
-                        if (s->left_comp_ctx[row7]) {
+                    if (have_l && !td->left_intra_ctx[row7]) {
+                        if (td->left_comp_ctx[row7]) {
                             if (s->above_comp_ctx[col]) {
-                                c = 1 + (!s->s.h.fixcompref || !s->left_ref_ctx[row7] ||
+                                c = 1 + (!s->s.h.fixcompref || !td->left_ref_ctx[row7] ||
                                          !s->above_ref_ctx[col]);
                             } else {
                                 c = (3 * !s->above_ref_ctx[col]) +
-                                    (!s->s.h.fixcompref || !s->left_ref_ctx[row7]);
+                                    (!s->s.h.fixcompref || !td->left_ref_ctx[row7]);
                             }
                         } else if (s->above_comp_ctx[col]) {
-                            c = (3 * !s->left_ref_ctx[row7]) +
+                            c = (3 * !td->left_ref_ctx[row7]) +
                                 (!s->s.h.fixcompref || !s->above_ref_ctx[col]);
                         } else {
-                            c = 2 * !s->left_ref_ctx[row7] + 2 * !s->above_ref_ctx[col];
+                            c = 2 * !td->left_ref_ctx[row7] + 2 * !s->above_ref_ctx[col];
                         }
                     } else if (s->above_intra_ctx[col]) {
                         c = 2;
@@ -468,26 +468,26 @@ static void decode_mode(AVCodecContext *avctx)
                     } else {
                         c = 4 * (!s->above_ref_ctx[col]);
                     }
-                } else if (have_l && !s->left_intra_ctx[row7]) {
-                    if (s->left_intra_ctx[row7]) {
+                } else if (have_l && !td->left_intra_ctx[row7]) {
+                    if (td->left_intra_ctx[row7]) {
                         c = 2;
-                    } else if (s->left_comp_ctx[row7]) {
-                        c = 1 + (!s->s.h.fixcompref || !s->left_ref_ctx[row7]);
+                    } else if (td->left_comp_ctx[row7]) {
+                        c = 1 + (!s->s.h.fixcompref || !td->left_ref_ctx[row7]);
                     } else {
-                        c = 4 * (!s->left_ref_ctx[row7]);
+                        c = 4 * (!td->left_ref_ctx[row7]);
                     }
                 } else {
                     c = 2;
                 }
-                bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][0]);
-                s->counts.single_ref[c][0][bit]++;
+                bit = vp56_rac_get_prob(td->c, s->prob.p.single_ref[c][0]);
+                td->counts.single_ref[c][0][bit]++;
                 if (!bit) {
                     b->ref[0] = 0;
                 } else {
                     // FIXME can this codeblob be replaced by some sort of LUT?
                     if (have_a) {
                         if (have_l) {
-                            if (s->left_intra_ctx[row7]) {
+                            if (td->left_intra_ctx[row7]) {
                                 if (s->above_intra_ctx[col]) {
                                     c = 2;
                                 } else if (s->above_comp_ctx[col]) {
@@ -499,49 +499,49 @@ static void decode_mode(AVCodecContext *avctx)
                                     c = 4 * (s->above_ref_ctx[col] == 1);
                                 }
                             } else if (s->above_intra_ctx[col]) {
-                                if (s->left_intra_ctx[row7]) {
+                                if (td->left_intra_ctx[row7]) {
                                     c = 2;
-                                } else if (s->left_comp_ctx[row7]) {
+                                } else if (td->left_comp_ctx[row7]) {
                                     c = 1 + 2 * (s->s.h.fixcompref == 1 ||
-                                                 s->left_ref_ctx[row7] == 1);
-                                } else if (!s->left_ref_ctx[row7]) {
+                                                 td->left_ref_ctx[row7] == 1);
+                                } else if (!td->left_ref_ctx[row7]) {
                                     c = 3;
                                 } else {
-                                    c = 4 * (s->left_ref_ctx[row7] == 1);
+                                    c = 4 * (td->left_ref_ctx[row7] == 1);
                                 }
                             } else if (s->above_comp_ctx[col]) {
-                                if (s->left_comp_ctx[row7]) {
-                                    if (s->left_ref_ctx[row7] == s->above_ref_ctx[col]) {
+                                if (td->left_comp_ctx[row7]) {
+                                    if (td->left_ref_ctx[row7] == s->above_ref_ctx[col]) {
                                         c = 3 * (s->s.h.fixcompref == 1 ||
-                                                 s->left_ref_ctx[row7] == 1);
+                                                 td->left_ref_ctx[row7] == 1);
                                     } else {
                                         c = 2;
                                     }
-                                } else if (!s->left_ref_ctx[row7]) {
+                                } else if (!td->left_ref_ctx[row7]) {
                                     c = 1 + 2 * (s->s.h.fixcompref == 1 ||
                                                  s->above_ref_ctx[col] == 1);
                                 } else {
-                                    c = 3 * (s->left_ref_ctx[row7] == 1) +
+                                    c = 3 * (td->left_ref_ctx[row7] == 1) +
                                     (s->s.h.fixcompref == 1 || s->above_ref_ctx[col] == 1);
                                 }
-                            } else if (s->left_comp_ctx[row7]) {
+                            } else if (td->left_comp_ctx[row7]) {
                                 if (!s->above_ref_ctx[col]) {
                                     c = 1 + 2 * (s->s.h.fixcompref == 1 ||
-                                                 s->left_ref_ctx[row7] == 1);
+                                                 td->left_ref_ctx[row7] == 1);
                                 } else {
                                     c = 3 * (s->above_ref_ctx[col] == 1) +
-                                    (s->s.h.fixcompref == 1 || s->left_ref_ctx[row7] == 1);
+                                    (s->s.h.fixcompref == 1 || td->left_ref_ctx[row7] == 1);
                                 }
                             } else if (!s->above_ref_ctx[col]) {
-                                if (!s->left_ref_ctx[row7]) {
+                                if (!td->left_ref_ctx[row7]) {
                                     c = 3;
                                 } else {
-                                    c = 4 * (s->left_ref_ctx[row7] == 1);
+                                    c = 4 * (td->left_ref_ctx[row7] == 1);
                                 }
-                            } else if (!s->left_ref_ctx[row7]) {
+                            } else if (!td->left_ref_ctx[row7]) {
                                 c = 4 * (s->above_ref_ctx[col] == 1);
                             } else {
-                                c = 2 * (s->left_ref_ctx[row7] == 1) +
+                                c = 2 * (td->left_ref_ctx[row7] == 1) +
                                     2 * (s->above_ref_ctx[col] == 1);
                             }
                         } else {
@@ -555,19 +555,19 @@ static void decode_mode(AVCodecContext *avctx)
                             }
                         }
                     } else if (have_l) {
-                        if (s->left_intra_ctx[row7] ||
-                            (!s->left_comp_ctx[row7] && !s->left_ref_ctx[row7])) {
+                        if (td->left_intra_ctx[row7] ||
+                            (!td->left_comp_ctx[row7] && !td->left_ref_ctx[row7])) {
                             c = 2;
-                        } else if (s->left_comp_ctx[row7]) {
-                            c = 3 * (s->s.h.fixcompref == 1 || s->left_ref_ctx[row7] == 1);
+                        } else if (td->left_comp_ctx[row7]) {
+                            c = 3 * (s->s.h.fixcompref == 1 || td->left_ref_ctx[row7] == 1);
                         } else {
-                            c = 4 * (s->left_ref_ctx[row7] == 1);
+                            c = 4 * (td->left_ref_ctx[row7] == 1);
                         }
                     } else {
                         c = 2;
                     }
-                    bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][1]);
-                    s->counts.single_ref[c][1][bit]++;
+                    bit = vp56_rac_get_prob(td->c, s->prob.p.single_ref[c][1]);
+                    td->counts.single_ref[c][1][bit]++;
                     b->ref[0] = 1 + bit;
                 }
             }
@@ -587,14 +587,14 @@ static void decode_mode(AVCodecContext *avctx)
                 // FIXME this needs to use the LUT tables from find_ref_mvs
                 // because not all are -1,0/0,-1
                 int c = inter_mode_ctx_lut[s->above_mode_ctx[col + off[b->bs]]]
-                                          [s->left_mode_ctx[row7 + off[b->bs]]];
+                                          [td->left_mode_ctx[row7 + off[b->bs]]];
 
-                b->mode[0] = vp8_rac_get_tree(&s->c, ff_vp9_inter_mode_tree,
+                b->mode[0] = vp8_rac_get_tree(td->c, ff_vp9_inter_mode_tree,
                                               s->prob.p.mv_mode[c]);
                 b->mode[1] =
                 b->mode[2] =
                 b->mode[3] = b->mode[0];
-                s->counts.mv_mode[c][b->mode[0] - 10]++;
+                td->counts.mv_mode[c][b->mode[0] - 10]++;
             }
         }
 
@@ -602,39 +602,39 @@ static void decode_mode(AVCodecContext *avctx)
             int c;
 
             if (have_a && s->above_mode_ctx[col] >= NEARESTMV) {
-                if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {
-                    c = s->above_filter_ctx[col] == s->left_filter_ctx[row7] ?
-                        s->left_filter_ctx[row7] : 3;
+                if (have_l && td->left_mode_ctx[row7] >= NEARESTMV) {
+                    c = s->above_filter_ctx[col] == td->left_filter_ctx[row7] ?
+                        td->left_filter_ctx[row7] : 3;
                 } else {
                     c = s->above_filter_ctx[col];
                 }
-            } else if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {
-                c = s->left_filter_ctx[row7];
+            } else if (have_l && td->left_mode_ctx[row7] >= NEARESTMV) {
+                c = td->left_filter_ctx[row7];
             } else {
                 c = 3;
             }
 
-            filter_id = vp8_rac_get_tree(&s->c, ff_vp9_filter_tree,
+            filter_id = vp8_rac_get_tree(td->c, ff_vp9_filter_tree,
                                          s->prob.p.filter[c]);
-            s->counts.filter[c][filter_id]++;
+            td->counts.filter[c][filter_id]++;
             b->filter = ff_vp9_filter_lut[filter_id];
         } else {
             b->filter = s->s.h.filtermode;
         }
 
         if (b->bs > BS_8x8) {
-            int c = inter_mode_ctx_lut[s->above_mode_ctx[col]][s->left_mode_ctx[row7]];
+            int c = inter_mode_ctx_lut[s->above_mode_ctx[col]][td->left_mode_ctx[row7]];
 
-            b->mode[0] = vp8_rac_get_tree(&s->c, ff_vp9_inter_mode_tree,
+            b->mode[0] = vp8_rac_get_tree(td->c, ff_vp9_inter_mode_tree,
                                           s->prob.p.mv_mode[c]);
-            s->counts.mv_mode[c][b->mode[0] - 10]++;
-            ff_vp9_fill_mv(s, b->mv[0], b->mode[0], 0);
+            td->counts.mv_mode[c][b->mode[0] - 10]++;
+            ff_vp9_fill_mv(td, b->mv[0], b->mode[0], 0);
 
             if (b->bs != BS_8x4) {
-                b->mode[1] = vp8_rac_get_tree(&s->c, ff_vp9_inter_mode_tree,
+                b->mode[1] = vp8_rac_get_tree(td->c, ff_vp9_inter_mode_tree,
                                               s->prob.p.mv_mode[c]);
-                s->counts.mv_mode[c][b->mode[1] - 10]++;
-                ff_vp9_fill_mv(s, b->mv[1], b->mode[1], 1);
+                td->counts.mv_mode[c][b->mode[1] - 10]++;
+                ff_vp9_fill_mv(td, b->mv[1], b->mode[1], 1);
             } else {
                 b->mode[1] = b->mode[0];
                 AV_COPY32(&b->mv[1][0], &b->mv[0][0]);
@@ -642,16 +642,16 @@ static void decode_mode(AVCodecContext *avctx)
             }
 
             if (b->bs != BS_4x8) {
-                b->mode[2] = vp8_rac_get_tree(&s->c, ff_vp9_inter_mode_tree,
+                b->mode[2] = vp8_rac_get_tree(td->c, ff_vp9_inter_mode_tree,
                                               s->prob.p.mv_mode[c]);
-                s->counts.mv_mode[c][b->mode[2] - 10]++;
-                ff_vp9_fill_mv(s, b->mv[2], b->mode[2], 2);
+                td->counts.mv_mode[c][b->mode[2] - 10]++;
+                ff_vp9_fill_mv(td, b->mv[2], b->mode[2], 2);
 
                 if (b->bs != BS_8x4) {
-                    b->mode[3] = vp8_rac_get_tree(&s->c, ff_vp9_inter_mode_tree,
+                    b->mode[3] = vp8_rac_get_tree(td->c, ff_vp9_inter_mode_tree,
                                                   s->prob.p.mv_mode[c]);
-                    s->counts.mv_mode[c][b->mode[3] - 10]++;
-                    ff_vp9_fill_mv(s, b->mv[3], b->mode[3], 3);
+                    td->counts.mv_mode[c][b->mode[3] - 10]++;
+                    ff_vp9_fill_mv(td, b->mv[3], b->mode[3], 3);
                 } else {
                     b->mode[3] = b->mode[2];
                     AV_COPY32(&b->mv[3][0], &b->mv[2][0]);
@@ -666,7 +666,7 @@ static void decode_mode(AVCodecContext *avctx)
                 AV_COPY32(&b->mv[3][1], &b->mv[1][1]);
             }
         } else {
-            ff_vp9_fill_mv(s, b->mv[0], b->mode[0], -1);
+            ff_vp9_fill_mv(td, b->mv[0], b->mode[0], -1);
             AV_COPY32(&b->mv[1][0], &b->mv[0][0]);
             AV_COPY32(&b->mv[2][0], &b->mv[0][0]);
             AV_COPY32(&b->mv[3][0], &b->mv[0][0]);
@@ -716,33 +716,33 @@ static void decode_mode(AVCodecContext *avctx)
 #endif
 
     switch (ff_vp9_bwh_tab[1][b->bs][0]) {
-#define SET_CTXS(dir, off, n) \
+#define SET_CTXS(perf, dir, off, n) \
     do { \
-        SPLAT_CTX(s->dir##_skip_ctx[off],      b->skip,          n); \
-        SPLAT_CTX(s->dir##_txfm_ctx[off],      b->tx,            n); \
-        SPLAT_CTX(s->dir##_partition_ctx[off], dir##_ctx[b->bs], n); \
+        SPLAT_CTX(perf->dir##_skip_ctx[off],      b->skip,          n); \
+        SPLAT_CTX(perf->dir##_txfm_ctx[off],      b->tx,            n); \
+        SPLAT_CTX(perf->dir##_partition_ctx[off], dir##_ctx[b->bs], n); \
         if (!s->s.h.keyframe && !s->s.h.intraonly) { \
-            SPLAT_CTX(s->dir##_intra_ctx[off], b->intra,   n); \
-            SPLAT_CTX(s->dir##_comp_ctx[off],  b->comp,    n); \
-            SPLAT_CTX(s->dir##_mode_ctx[off],  b->mode[3], n); \
+            SPLAT_CTX(perf->dir##_intra_ctx[off], b->intra,   n); \
+            SPLAT_CTX(perf->dir##_comp_ctx[off],  b->comp,    n); \
+            SPLAT_CTX(perf->dir##_mode_ctx[off],  b->mode[3], n); \
             if (!b->intra) { \
-                SPLAT_CTX(s->dir##_ref_ctx[off], vref, n); \
+                SPLAT_CTX(perf->dir##_ref_ctx[off], vref, n); \
                 if (s->s.h.filtermode == FILTER_SWITCHABLE) { \
-                    SPLAT_CTX(s->dir##_filter_ctx[off], filter_id, n); \
+                    SPLAT_CTX(perf->dir##_filter_ctx[off], filter_id, n); \
                 } \
             } \
         } \
     } while (0)
-    case 1: SET_CTXS(above, col, 1); break;
-    case 2: SET_CTXS(above, col, 2); break;
-    case 4: SET_CTXS(above, col, 4); break;
-    case 8: SET_CTXS(above, col, 8); break;
+    case 1: SET_CTXS(s, above, col, 1); break;
+    case 2: SET_CTXS(s, above, col, 2); break;
+    case 4: SET_CTXS(s, above, col, 4); break;
+    case 8: SET_CTXS(s, above, col, 8); break;
     }
     switch (ff_vp9_bwh_tab[1][b->bs][1]) {
-    case 1: SET_CTXS(left, row7, 1); break;
-    case 2: SET_CTXS(left, row7, 2); break;
-    case 4: SET_CTXS(left, row7, 4); break;
-    case 8: SET_CTXS(left, row7, 8); break;
+    case 1: SET_CTXS(td, left, row7, 1); break;
+    case 2: SET_CTXS(td, left, row7, 2); break;
+    case 4: SET_CTXS(td, left, row7, 4); break;
+    case 8: SET_CTXS(td, left, row7, 8); break;
     }
 #undef SPLAT_CTX
 #undef SET_CTXS
@@ -751,10 +751,10 @@ static void decode_mode(AVCodecContext *avctx)
         if (b->bs > BS_8x8) {
             int mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);
 
-            AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][0], &b->mv[1][0]);
-            AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][1], &b->mv[1][1]);
-            AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][0], mv0);
-            AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][1], mv1);
+            AV_COPY32(&td->left_mv_ctx[row7 * 2 + 0][0], &b->mv[1][0]);
+            AV_COPY32(&td->left_mv_ctx[row7 * 2 + 0][1], &b->mv[1][1]);
+            AV_WN32A(&td->left_mv_ctx[row7 * 2 + 1][0], mv0);
+            AV_WN32A(&td->left_mv_ctx[row7 * 2 + 1][1], mv1);
             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][0], &b->mv[2][0]);
             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][1], &b->mv[2][1]);
             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][0], mv0);
@@ -767,8 +767,8 @@ static void decode_mode(AVCodecContext *avctx)
                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][1], mv1);
             }
             for (n = 0; n < h4 * 2; n++) {
-                AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][0], mv0);
-                AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][1], mv1);
+                AV_WN32A(&td->left_mv_ctx[row7 * 2 + n][0], mv0);
+                AV_WN32A(&td->left_mv_ctx[row7 * 2 + n][1], mv1);
             }
         }
     }
@@ -806,10 +806,10 @@ decode_coeffs_b_generic(VP56RangeCoder *c, int16_t *coef, int n_coeffs,
                         int is_tx32x32, int is8bitsperpixel, int bpp, unsigned (*cnt)[6][3],
                         unsigned (*eob)[6][2], uint8_t (*p)[6][11],
                         int nnz, const int16_t *scan, const int16_t (*nb)[2],
-                        const int16_t *band_counts, const int16_t *qmul)
+                        const int16_t *band_counts, int16_t *qmul)
 {
     int i = 0, band = 0, band_left = band_counts[band];
-    uint8_t *tp = p[0][nnz];
+    const uint8_t *tp = p[0][nnz];
     uint8_t cache[1024];
 
     do {
@@ -839,10 +839,6 @@ skip_eob:
             val       = 1;
             cache[rc] = 1;
         } else {
-            // fill in p[3-10] (model fill) - only once per frame for each pos
-            if (!tp[3])
-                memcpy(&tp[3], ff_vp9_model_pareto8[tp[2]], 8);
-
             cnt[band][nnz][2]++;
             if (!vp56_rac_get_prob_branchy(c, tp[3])) { // 2, 3, 4
                 if (!vp56_rac_get_prob_branchy(c, tp[4])) {
@@ -925,54 +921,54 @@ skip_eob:
     return i;
 }
 
-static int decode_coeffs_b_8bpp(VP9Context *s, int16_t *coef, int n_coeffs,
+static int decode_coeffs_b_8bpp(VP9TileData *td, int16_t *coef, int n_coeffs,
                                 unsigned (*cnt)[6][3], unsigned (*eob)[6][2],
                                 uint8_t (*p)[6][11], int nnz, const int16_t *scan,
                                 const int16_t (*nb)[2], const int16_t *band_counts,
-                                const int16_t *qmul)
+                                int16_t *qmul)
 {
-    return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 0, 1, 8, cnt, eob, p,
+    return decode_coeffs_b_generic(td->c, coef, n_coeffs, 0, 1, 8, cnt, eob, p,
                                    nnz, scan, nb, band_counts, qmul);
 }
 
-static int decode_coeffs_b32_8bpp(VP9Context *s, int16_t *coef, int n_coeffs,
+static int decode_coeffs_b32_8bpp(VP9TileData *td, int16_t *coef, int n_coeffs,
                                   unsigned (*cnt)[6][3], unsigned (*eob)[6][2],
                                   uint8_t (*p)[6][11], int nnz, const int16_t *scan,
                                   const int16_t (*nb)[2], const int16_t *band_counts,
-                                  const int16_t *qmul)
+                                  int16_t *qmul)
 {
-    return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 1, 1, 8, cnt, eob, p,
+    return decode_coeffs_b_generic(td->c, coef, n_coeffs, 1, 1, 8, cnt, eob, p,
                                    nnz, scan, nb, band_counts, qmul);
 }
 
-static int decode_coeffs_b_16bpp(VP9Context *s, int16_t *coef, int n_coeffs,
+static int decode_coeffs_b_16bpp(VP9TileData *td, int16_t *coef, int n_coeffs,
                                  unsigned (*cnt)[6][3], unsigned (*eob)[6][2],
                                  uint8_t (*p)[6][11], int nnz, const int16_t *scan,
                                  const int16_t (*nb)[2], const int16_t *band_counts,
-                                 const int16_t *qmul)
+                                 int16_t *qmul)
 {
-    return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 0, 0, s->s.h.bpp, cnt, eob, p,
+    return decode_coeffs_b_generic(td->c, coef, n_coeffs, 0, 0, td->s->s.h.bpp, cnt, eob, p,
                                    nnz, scan, nb, band_counts, qmul);
 }
 
-static int decode_coeffs_b32_16bpp(VP9Context *s, int16_t *coef, int n_coeffs,
+static int decode_coeffs_b32_16bpp(VP9TileData *td, int16_t *coef, int n_coeffs,
                                    unsigned (*cnt)[6][3], unsigned (*eob)[6][2],
                                    uint8_t (*p)[6][11], int nnz, const int16_t *scan,
                                    const int16_t (*nb)[2], const int16_t *band_counts,
-                                   const int16_t *qmul)
+                                   int16_t *qmul)
 {
-    return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 1, 0, s->s.h.bpp, cnt, eob, p,
+    return decode_coeffs_b_generic(td->c, coef, n_coeffs, 1, 0, td->s->s.h.bpp, cnt, eob, p,
                                    nnz, scan, nb, band_counts, qmul);
 }
 
-static av_always_inline int decode_coeffs(AVCodecContext *avctx, int is8bitsperpixel)
+static av_always_inline int decode_coeffs(VP9TileData *td, int is8bitsperpixel)
 {
-    VP9Context *s = avctx->priv_data;
-    VP9Block *b = s->b;
-    int row = s->row, col = s->col;
+    VP9Context *s = td->s;
+    VP9Block *b = td->b;
+    int row = td->row, col = td->col;
     uint8_t (*p)[6][11] = s->prob.coef[b->tx][0 /* y */][!b->intra];
-    unsigned (*c)[6][3] = s->counts.coef[b->tx][0 /* y */][!b->intra];
-    unsigned (*e)[6][2] = s->counts.eob[b->tx][0 /* y */][!b->intra];
+    unsigned (*c)[6][3] = td->counts.coef[b->tx][0 /* y */][!b->intra];
+    unsigned (*e)[6][2] = td->counts.eob[b->tx][0 /* y */][!b->intra];
     int w4 = ff_vp9_bwh_tab[1][b->bs][0] << 1, h4 = ff_vp9_bwh_tab[1][b->bs][1] << 1;
     int end_x = FFMIN(2 * (s->cols - col), w4);
     int end_y = FFMIN(2 * (s->rows - row), h4);
@@ -984,7 +980,7 @@ static av_always_inline int decode_coeffs(AVCodecContext *avctx, int is8bitsperp
     const int16_t *uvscan = ff_vp9_scans[b->uvtx][DCT_DCT];
     const int16_t (*uvnb)[2] = ff_vp9_scans_nb[b->uvtx][DCT_DCT];
     uint8_t *a = &s->above_y_nnz_ctx[col * 2];
-    uint8_t *l = &s->left_y_nnz_ctx[(row & 7) << 1];
+    uint8_t *l = &td->left_y_nnz_ctx[(row & 7) << 1];
     static const int16_t band_counts[4][8] = {
         { 1, 2, 3, 4,  3,   16 - 13 },
         { 1, 2, 3, 4, 11,   64 - 21 },
@@ -1010,15 +1006,15 @@ static av_always_inline int decode_coeffs(AVCodecContext *avctx, int is8bitsperp
         for (x = 0; x < end_x; x += step, n += step * step) { \
             enum TxfmType txtp = ff_vp9_intra_txfm_type[b->mode[mode_index]]; \
             ret = (is8bitsperpixel ? decode_coeffs_b##v##_8bpp : decode_coeffs_b##v##_16bpp) \
-                                    (s, s->block + 16 * n * bytesperpixel, 16 * step * step, \
+                                    (td, td->block + 16 * n * bytesperpixel, 16 * step * step, \
                                      c, e, p, a[x] + l[y], yscans[txtp], \
                                      ynbs[txtp], y_band_counts, qmul[0]); \
             a[x] = l[y] = !!ret; \
             total_coeff |= !!ret; \
             if (step >= 4) { \
-                AV_WN16A(&s->eob[n], ret); \
+                AV_WN16A(&td->eob[n], ret); \
             } else { \
-                s->eob[n] = ret; \
+                td->eob[n] = ret; \
             } \
         } \
     }
@@ -1084,29 +1080,29 @@ static av_always_inline int decode_coeffs(AVCodecContext *avctx, int is8bitsperp
     for (n = 0, y = 0; y < end_y; y += step) { \
         for (x = 0; x < end_x; x += step, n += step * step) { \
             ret = (is8bitsperpixel ? decode_coeffs_b##v##_8bpp : decode_coeffs_b##v##_16bpp) \
-                                    (s, s->uvblock[pl] + 16 * n * bytesperpixel, \
+                                    (td, td->uvblock[pl] + 16 * n * bytesperpixel, \
                                      16 * step * step, c, e, p, a[x] + l[y], \
                                      uvscan, uvnb, uv_band_counts, qmul[1]); \
             a[x] = l[y] = !!ret; \
             total_coeff |= !!ret; \
             if (step >= 4) { \
-                AV_WN16A(&s->uveob[pl][n], ret); \
+                AV_WN16A(&td->uveob[pl][n], ret); \
             } else { \
-                s->uveob[pl][n] = ret; \
+                td->uveob[pl][n] = ret; \
             } \
         } \
     }
 
     p = s->prob.coef[b->uvtx][1 /* uv */][!b->intra];
-    c = s->counts.coef[b->uvtx][1 /* uv */][!b->intra];
-    e = s->counts.eob[b->uvtx][1 /* uv */][!b->intra];
+    c = td->counts.coef[b->uvtx][1 /* uv */][!b->intra];
+    e = td->counts.eob[b->uvtx][1 /* uv */][!b->intra];
     w4 >>= s->ss_h;
     end_x >>= s->ss_h;
     h4 >>= s->ss_v;
     end_y >>= s->ss_v;
     for (pl = 0; pl < 2; pl++) {
         a = &s->above_uv_nnz_ctx[pl][col << !s->ss_h];
-        l = &s->left_uv_nnz_ctx[pl][(row & 7) << !s->ss_v];
+        l = &td->left_uv_nnz_ctx[pl][(row & 7) << !s->ss_v];
         switch (b->uvtx) {
         case TX_4X4:
             DECODE_UV_COEF_LOOP(1,);
@@ -1132,14 +1128,14 @@ static av_always_inline int decode_coeffs(AVCodecContext *avctx, int is8bitsperp
     return total_coeff;
 }
 
-static int decode_coeffs_8bpp(AVCodecContext *avctx)
+static int decode_coeffs_8bpp(VP9TileData *td)
 {
-    return decode_coeffs(avctx, 1);
+    return decode_coeffs(td, 1);
 }
 
-static int decode_coeffs_16bpp(AVCodecContext *avctx)
+static int decode_coeffs_16bpp(VP9TileData *td)
 {
-    return decode_coeffs(avctx, 0);
+    return decode_coeffs(td, 0);
 }
 
 static av_always_inline void mask_edges(uint8_t (*mask)[8][4], int ss_h, int ss_v,
@@ -1264,33 +1260,33 @@ static av_always_inline void mask_edges(uint8_t (*mask)[8][4], int ss_h, int ss_
     }
 }
 
-void ff_vp9_decode_block(AVCodecContext *avctx, int row, int col,
+void ff_vp9_decode_block(VP9TileData *td, int row, int col,
                          VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,
                          enum BlockLevel bl, enum BlockPartition bp)
 {
-    VP9Context *s = avctx->priv_data;
-    VP9Block *b = s->b;
+    VP9Context *s = td->s;
+    VP9Block *b = td->b;
     enum BlockSize bs = bl * 3 + bp;
     int bytesperpixel = s->bytesperpixel;
     int w4 = ff_vp9_bwh_tab[1][bs][0], h4 = ff_vp9_bwh_tab[1][bs][1], lvl;
     int emu[2];
     AVFrame *f = s->s.frames[CUR_FRAME].tf.f;
 
-    s->row = row;
-    s->row7 = row & 7;
-    s->col = col;
-    s->col7 = col & 7;
+    td->row = row;
+    td->row7 = row & 7;
+    td->col = col;
+    td->col7 = col & 7;
 
-    s->min_mv.x = -(128 + col * 64);
-    s->min_mv.y = -(128 + row * 64);
-    s->max_mv.x = 128 + (s->cols - col - w4) * 64;
-    s->max_mv.y = 128 + (s->rows - row - h4) * 64;
+    td->min_mv.x = -(128 + col * 64);
+    td->min_mv.y = -(128 + row * 64);
+    td->max_mv.x = 128 + (s->cols - col - w4) * 64;
+    td->max_mv.y = 128 + (s->rows - row - h4) * 64;
 
     if (s->pass < 2) {
         b->bs = bs;
         b->bl = bl;
         b->bp = bp;
-        decode_mode(avctx);
+        decode_mode(td);
         b->uvtx = b->tx - ((s->ss_h && w4 * 2 == (1 << b->tx)) ||
                            (s->ss_v && h4 * 2 == (1 << b->tx)));
 
@@ -1298,17 +1294,17 @@ void ff_vp9_decode_block(AVCodecContext *avctx, int row, int col,
             int has_coeffs;
 
             if (bytesperpixel == 1) {
-                has_coeffs = decode_coeffs_8bpp(avctx);
+                has_coeffs = decode_coeffs_8bpp(td);
             } else {
-                has_coeffs = decode_coeffs_16bpp(avctx);
+                has_coeffs = decode_coeffs_16bpp(td);
             }
             if (!has_coeffs && b->bs <= BS_8x8 && !b->intra) {
                 b->skip = 1;
                 memset(&s->above_skip_ctx[col], 1, w4);
-                memset(&s->left_skip_ctx[s->row7], 1, h4);
+                memset(&td->left_skip_ctx[td->row7], 1, h4);
             }
         } else {
-            int row7 = s->row7;
+            int row7 = td->row7;
 
 #define SPLAT_ZERO_CTX(v, n) \
     switch (n) { \
@@ -1320,38 +1316,38 @@ void ff_vp9_decode_block(AVCodecContext *avctx, int row, int col,
     }
 #define SPLAT_ZERO_YUV(dir, var, off, n, dir2) \
     do { \
-        SPLAT_ZERO_CTX(s->dir##_y_##var[off * 2], n * 2); \
+        SPLAT_ZERO_CTX(dir##_y_##var[off * 2], n * 2); \
         if (s->ss_##dir2) { \
-            SPLAT_ZERO_CTX(s->dir##_uv_##var[0][off], n); \
-            SPLAT_ZERO_CTX(s->dir##_uv_##var[1][off], n); \
+            SPLAT_ZERO_CTX(dir##_uv_##var[0][off], n); \
+            SPLAT_ZERO_CTX(dir##_uv_##var[1][off], n); \
         } else { \
-            SPLAT_ZERO_CTX(s->dir##_uv_##var[0][off * 2], n * 2); \
-            SPLAT_ZERO_CTX(s->dir##_uv_##var[1][off * 2], n * 2); \
+            SPLAT_ZERO_CTX(dir##_uv_##var[0][off * 2], n * 2); \
+            SPLAT_ZERO_CTX(dir##_uv_##var[1][off * 2], n * 2); \
         } \
     } while (0)
 
             switch (w4) {
-            case 1: SPLAT_ZERO_YUV(above, nnz_ctx, col, 1, h); break;
-            case 2: SPLAT_ZERO_YUV(above, nnz_ctx, col, 2, h); break;
-            case 4: SPLAT_ZERO_YUV(above, nnz_ctx, col, 4, h); break;
-            case 8: SPLAT_ZERO_YUV(above, nnz_ctx, col, 8, h); break;
+            case 1: SPLAT_ZERO_YUV(s->above, nnz_ctx, col, 1, h); break;
+            case 2: SPLAT_ZERO_YUV(s->above, nnz_ctx, col, 2, h); break;
+            case 4: SPLAT_ZERO_YUV(s->above, nnz_ctx, col, 4, h); break;
+            case 8: SPLAT_ZERO_YUV(s->above, nnz_ctx, col, 8, h); break;
             }
             switch (h4) {
-            case 1: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 1, v); break;
-            case 2: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 2, v); break;
-            case 4: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 4, v); break;
-            case 8: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 8, v); break;
+            case 1: SPLAT_ZERO_YUV(td->left, nnz_ctx, row7, 1, v); break;
+            case 2: SPLAT_ZERO_YUV(td->left, nnz_ctx, row7, 2, v); break;
+            case 4: SPLAT_ZERO_YUV(td->left, nnz_ctx, row7, 4, v); break;
+            case 8: SPLAT_ZERO_YUV(td->left, nnz_ctx, row7, 8, v); break;
             }
         }
 
         if (s->pass == 1) {
-            s->b++;
-            s->block += w4 * h4 * 64 * bytesperpixel;
-            s->uvblock[0] += w4 * h4 * 64 * bytesperpixel >> (s->ss_h + s->ss_v);
-            s->uvblock[1] += w4 * h4 * 64 * bytesperpixel >> (s->ss_h + s->ss_v);
-            s->eob += 4 * w4 * h4;
-            s->uveob[0] += 4 * w4 * h4 >> (s->ss_h + s->ss_v);
-            s->uveob[1] += 4 * w4 * h4 >> (s->ss_h + s->ss_v);
+            s->td[0].b++;
+            s->td[0].block += w4 * h4 * 64 * bytesperpixel;
+            s->td[0].uvblock[0] += w4 * h4 * 64 * bytesperpixel >> (s->ss_h + s->ss_v);
+            s->td[0].uvblock[1] += w4 * h4 * 64 * bytesperpixel >> (s->ss_h + s->ss_v);
+            s->td[0].eob += 4 * w4 * h4;
+            s->td[0].uveob[0] += 4 * w4 * h4 >> (s->ss_h + s->ss_v);
+            s->td[0].uveob[1] += 4 * w4 * h4 >> (s->ss_h + s->ss_v);
 
             return;
         }
@@ -1365,32 +1361,32 @@ void ff_vp9_decode_block(AVCodecContext *avctx, int row, int col,
     emu[1] = ((col + w4) * 8 >> s->ss_h) * bytesperpixel > f->linesize[1] ||
              (row + h4) > s->rows;
     if (emu[0]) {
-        s->dst[0] = s->tmp_y;
-        s->y_stride = 128;
+        td->dst[0] = td->tmp_y;
+        td->y_stride = 128;
     } else {
-        s->dst[0] = f->data[0] + yoff;
-        s->y_stride = f->linesize[0];
+        td->dst[0] = f->data[0] + yoff;
+        td->y_stride = f->linesize[0];
     }
     if (emu[1]) {
-        s->dst[1] = s->tmp_uv[0];
-        s->dst[2] = s->tmp_uv[1];
-        s->uv_stride = 128;
+        td->dst[1] = td->tmp_uv[0];
+        td->dst[2] = td->tmp_uv[1];
+        td->uv_stride = 128;
     } else {
-        s->dst[1] = f->data[1] + uvoff;
-        s->dst[2] = f->data[2] + uvoff;
-        s->uv_stride = f->linesize[1];
+        td->dst[1] = f->data[1] + uvoff;
+        td->dst[2] = f->data[2] + uvoff;
+        td->uv_stride = f->linesize[1];
     }
     if (b->intra) {
         if (s->s.h.bpp > 8) {
-            ff_vp9_intra_recon_16bpp(avctx, yoff, uvoff);
+            ff_vp9_intra_recon_16bpp(td, yoff, uvoff);
         } else {
-            ff_vp9_intra_recon_8bpp(avctx, yoff, uvoff);
+            ff_vp9_intra_recon_8bpp(td, yoff, uvoff);
         }
     } else {
         if (s->s.h.bpp > 8) {
-            ff_vp9_inter_recon_16bpp(avctx);
+            ff_vp9_inter_recon_16bpp(td);
         } else {
-            ff_vp9_inter_recon_8bpp(avctx);
+            ff_vp9_inter_recon_8bpp(td);
         }
     }
     if (emu[0]) {
@@ -1402,7 +1398,7 @@ void ff_vp9_decode_block(AVCodecContext *avctx, int row, int col,
             av_assert2(n <= 4);
             if (w & bw) {
                 s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o * bytesperpixel, f->linesize[0],
-                                         s->tmp_y + o * bytesperpixel, 128, h, 0, 0);
+                                         td->tmp_y + o * bytesperpixel, 128, h, 0, 0);
                 o += bw;
             }
         }
@@ -1417,9 +1413,9 @@ void ff_vp9_decode_block(AVCodecContext *avctx, int row, int col,
             av_assert2(n <= 4);
             if (w & bw) {
                 s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o * bytesperpixel, f->linesize[1],
-                                         s->tmp_uv[0] + o * bytesperpixel, 128, h, 0, 0);
+                                         td->tmp_uv[0] + o * bytesperpixel, 128, h, 0, 0);
                 s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o * bytesperpixel, f->linesize[2],
-                                         s->tmp_uv[1] + o * bytesperpixel, 128, h, 0, 0);
+                                         td->tmp_uv[1] + o * bytesperpixel, 128, h, 0, 0);
                 o += bw;
             }
         }
@@ -1430,7 +1426,7 @@ void ff_vp9_decode_block(AVCodecContext *avctx, int row, int col,
         (lvl = s->s.h.segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]
                                                       [b->mode[3] != ZEROMV]) > 0) {
         int x_end = FFMIN(s->cols - col, w4), y_end = FFMIN(s->rows - row, h4);
-        int skip_inter = !b->intra && b->skip, col7 = s->col7, row7 = s->row7;
+        int skip_inter = !b->intra && b->skip, col7 = td->col7, row7 = td->row7;
 
         setctx_2d(&lflvl->level[row7 * 8 + col7], w4, h4, 8, lvl);
         mask_edges(lflvl->mask[0], 0, 0, row7, col7, x_end, y_end, 0, 0, b->tx, skip_inter);
@@ -1439,29 +1435,15 @@ void ff_vp9_decode_block(AVCodecContext *avctx, int row, int col,
                        s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,
                        s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,
                        b->uvtx, skip_inter);
-
-        if (!s->filter_lut.lim_lut[lvl]) {
-            int sharp = s->s.h.filter.sharpness;
-            int limit = lvl;
-
-            if (sharp > 0) {
-                limit >>= (sharp + 3) >> 2;
-                limit = FFMIN(limit, 9 - sharp);
-            }
-            limit = FFMAX(limit, 1);
-
-            s->filter_lut.lim_lut[lvl] = limit;
-            s->filter_lut.mblim_lut[lvl] = 2 * (lvl + 2) + limit;
-        }
     }
 
     if (s->pass == 2) {
-        s->b++;
-        s->block += w4 * h4 * 64 * bytesperpixel;
-        s->uvblock[0] += w4 * h4 * 64 * bytesperpixel >> (s->ss_v + s->ss_h);
-        s->uvblock[1] += w4 * h4 * 64 * bytesperpixel >> (s->ss_v + s->ss_h);
-        s->eob += 4 * w4 * h4;
-        s->uveob[0] += 4 * w4 * h4 >> (s->ss_v + s->ss_h);
-        s->uveob[1] += 4 * w4 * h4 >> (s->ss_v + s->ss_h);
+        s->td[0].b++;
+        s->td[0].block += w4 * h4 * 64 * bytesperpixel;
+        s->td[0].uvblock[0] += w4 * h4 * 64 * bytesperpixel >> (s->ss_v + s->ss_h);
+        s->td[0].uvblock[1] += w4 * h4 * 64 * bytesperpixel >> (s->ss_v + s->ss_h);
+        s->td[0].eob += 4 * w4 * h4;
+        s->td[0].uveob[0] += 4 * w4 * h4 >> (s->ss_v + s->ss_h);
+        s->td[0].uveob[1] += 4 * w4 * h4 >> (s->ss_v + s->ss_h);
     }
 }
diff --git a/libavcodec/vp9dec.h b/libavcodec/vp9dec.h
index 4002b3a..15e0122 100644
--- a/libavcodec/vp9dec.h
+++ b/libavcodec/vp9dec.h
@@ -26,8 +26,10 @@
 
 #include <stddef.h>
 #include <stdint.h>
+#include <stdatomic.h>
 
 #include "libavutil/buffer.h"
+#include "libavutil/thread.h"
 #include "libavutil/internal.h"
 
 #include "vp9.h"
@@ -84,20 +86,21 @@ typedef struct VP9Block {
     enum BlockPartition bp;
 } VP9Block;
 
+typedef struct VP9TileData VP9TileData;
+
 typedef struct VP9Context {
+    VP9TileData *td;
     VP9SharedContext s;
 
     VP9DSPContext dsp;
     VideoDSPContext vdsp;
     GetBitContext gb;
     VP56RangeCoder c;
-    VP56RangeCoder *c_b;
-    unsigned c_b_size;
-    VP9Block *b_base, *b;
-    int pass;
-    int row, row7, col, col7;
-    uint8_t *dst[3];
-    ptrdiff_t y_stride, uv_stride;
+    int pass, active_tile_cols;
+
+    pthread_mutex_t progress_mutex;
+    pthread_cond_t progress_cond;
+    atomic_int *entries;
 
     uint8_t ss_h, ss_v;
     uint8_t last_bpp, bpp_index, bytesperpixel;
@@ -115,7 +118,6 @@ typedef struct VP9Context {
         uint8_t lim_lut[64];
         uint8_t mblim_lut[64];
     } filter_lut;
-    unsigned tile_row_start, tile_row_end, tile_col_start, tile_col_end;
     struct {
         ProbContext p;
         uint8_t coef[4][2][2][6][6][3];
@@ -124,6 +126,44 @@ typedef struct VP9Context {
         ProbContext p;
         uint8_t coef[4][2][2][6][6][11];
     } prob;
+
+    // contextual (above) cache
+    uint8_t *above_partition_ctx;
+    uint8_t *above_mode_ctx;
+    // FIXME maybe merge some of the below in a flags field?
+    uint8_t *above_y_nnz_ctx;
+    uint8_t *above_uv_nnz_ctx[2];
+    uint8_t *above_skip_ctx; // 1bit
+    uint8_t *above_txfm_ctx; // 2bit
+    uint8_t *above_segpred_ctx; // 1bit
+    uint8_t *above_intra_ctx; // 1bit
+    uint8_t *above_comp_ctx; // 1bit
+    uint8_t *above_ref_ctx; // 2bit
+    uint8_t *above_filter_ctx;
+    VP56mv (*above_mv_ctx)[2];
+
+    // whole-frame cache
+    uint8_t *intra_pred_data[3];
+    VP9Filter *lflvl;
+
+    // block reconstruction intermediates
+    int block_alloc_using_2pass;
+    uint16_t mvscale[3][2];
+    uint8_t mvstep[3][2];
+} VP9Context;
+
+struct VP9TileData {
+    //VP9Context should be const, but because of the threading API(generates
+    //a lot of warnings) it's not.
+    VP9Context *s;
+    VP56RangeCoder *c_b;
+    VP56RangeCoder *c;
+    int row, row7, col, col7;
+    uint8_t *dst[3];
+    ptrdiff_t y_stride, uv_stride;
+    VP9Block *b_base, *b;
+    unsigned tile_col_start;
+
     struct {
         unsigned y_mode[4][10];
         unsigned uv_mode[10][10];
@@ -153,7 +193,10 @@ typedef struct VP9Context {
         unsigned eob[4][2][2][6][6][2];
     } counts;
 
-    // contextual (left/above) cache
+    // whole-frame cache
+    DECLARE_ALIGNED(32, uint8_t, edge_emu_buffer)[135 * 144 * 2];
+
+    // contextual (left) cache
     DECLARE_ALIGNED(16, uint8_t, left_y_nnz_ctx)[16];
     DECLARE_ALIGNED(16, uint8_t, left_mode_ctx)[16];
     DECLARE_ALIGNED(16, VP56mv, left_mv_ctx)[16][2];
@@ -166,52 +209,30 @@ typedef struct VP9Context {
     DECLARE_ALIGNED(8, uint8_t, left_comp_ctx)[8];
     DECLARE_ALIGNED(8, uint8_t, left_ref_ctx)[8];
     DECLARE_ALIGNED(8, uint8_t, left_filter_ctx)[8];
-    uint8_t *above_partition_ctx;
-    uint8_t *above_mode_ctx;
-    // FIXME maybe merge some of the below in a flags field?
-    uint8_t *above_y_nnz_ctx;
-    uint8_t *above_uv_nnz_ctx[2];
-    uint8_t *above_skip_ctx; // 1bit
-    uint8_t *above_txfm_ctx; // 2bit
-    uint8_t *above_segpred_ctx; // 1bit
-    uint8_t *above_intra_ctx; // 1bit
-    uint8_t *above_comp_ctx; // 1bit
-    uint8_t *above_ref_ctx; // 2bit
-    uint8_t *above_filter_ctx;
-    VP56mv (*above_mv_ctx)[2];
-
-    // whole-frame cache
-    uint8_t *intra_pred_data[3];
-    VP9Filter *lflvl;
-    DECLARE_ALIGNED(32, uint8_t, edge_emu_buffer)[135 * 144 * 2];
-
     // block reconstruction intermediates
-    int block_alloc_using_2pass;
-    int16_t *block_base, *block, *uvblock_base[2], *uvblock[2];
-    uint8_t *eob_base, *uveob_base[2], *eob, *uveob[2];
-    struct { int x, y; } min_mv, max_mv;
     DECLARE_ALIGNED(32, uint8_t, tmp_y)[64 * 64 * 2];
     DECLARE_ALIGNED(32, uint8_t, tmp_uv)[2][64 * 64 * 2];
-    uint16_t mvscale[3][2];
-    uint8_t mvstep[3][2];
-} VP9Context;
+    struct { int x, y; } min_mv, max_mv;
+    int16_t *block_base, *block, *uvblock_base[2], *uvblock[2];
+    uint8_t *eob_base, *uveob_base[2], *eob, *uveob[2];
+};
 
-void ff_vp9_fill_mv(VP9Context *s, VP56mv *mv, int mode, int sb);
+void ff_vp9_fill_mv(VP9TileData *td, VP56mv *mv, int mode, int sb);
 
 void ff_vp9_adapt_probs(VP9Context *s);
 
-void ff_vp9_decode_block(AVCodecContext *ctx, int row, int col,
+void ff_vp9_decode_block(VP9TileData *td, int row, int col,
                          VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,
                          enum BlockLevel bl, enum BlockPartition bp);
 
 void ff_vp9_loopfilter_sb(AVCodecContext *avctx, VP9Filter *lflvl,
                           int row, int col, ptrdiff_t yoff, ptrdiff_t uvoff);
 
-void ff_vp9_intra_recon_8bpp(AVCodecContext *avctx,
+void ff_vp9_intra_recon_8bpp(VP9TileData *td,
                              ptrdiff_t y_off, ptrdiff_t uv_off);
-void ff_vp9_intra_recon_16bpp(AVCodecContext *avctx,
+void ff_vp9_intra_recon_16bpp(VP9TileData *td,
                               ptrdiff_t y_off, ptrdiff_t uv_off);
-void ff_vp9_inter_recon_8bpp(AVCodecContext *avctx);
-void ff_vp9_inter_recon_16bpp(AVCodecContext *avctx);
+void ff_vp9_inter_recon_8bpp(VP9TileData *td);
+void ff_vp9_inter_recon_16bpp(VP9TileData *td);
 
 #endif /* AVCODEC_VP9DEC_H */
diff --git a/libavcodec/vp9mvs.c b/libavcodec/vp9mvs.c
index e323bac..88db1c3 100644
--- a/libavcodec/vp9mvs.c
+++ b/libavcodec/vp9mvs.c
@@ -28,13 +28,13 @@
 #include "vp9dec.h"
 
 static av_always_inline void clamp_mv(VP56mv *dst, const VP56mv *src,
-                                      VP9Context *s)
+                                      VP9TileData *td)
 {
-    dst->x = av_clip(src->x, s->min_mv.x, s->max_mv.x);
-    dst->y = av_clip(src->y, s->min_mv.y, s->max_mv.y);
+    dst->x = av_clip(src->x, td->min_mv.x, td->max_mv.x);
+    dst->y = av_clip(src->y, td->min_mv.y, td->max_mv.y);
 }
 
-static void find_ref_mvs(VP9Context *s,
+static void find_ref_mvs(VP9TileData *td,
                          VP56mv *pmv, int ref, int z, int idx, int sb)
 {
     static const int8_t mv_ref_blk_off[N_BS_SIZES][8][2] = {
@@ -65,8 +65,9 @@ static void find_ref_mvs(VP9Context *s,
         [BS_4x4]   = { {  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },
                        { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 } },
     };
-    VP9Block *b = s->b;
-    int row = s->row, col = s->col, row7 = s->row7;
+    VP9Context *s = td->s;
+    VP9Block *b = td->b;
+    int row = td->row, col = td->col, row7 = td->row7;
     const int8_t (*p)[2] = mv_ref_blk_off[b->bs];
 #define INVALID_MV 0x80008000U
     uint32_t mem = INVALID_MV, mem_sub8x8 = INVALID_MV;
@@ -103,7 +104,7 @@ static void find_ref_mvs(VP9Context *s,
             av_assert2(idx == 1);                                      \
             av_assert2(mem != INVALID_MV);                             \
             if (mem_sub8x8 == INVALID_MV) {                            \
-                clamp_mv(&tmp, &mv, s);                                \
+                clamp_mv(&tmp, &mv, td);                               \
                 m = AV_RN32A(&tmp);                                    \
                 if (m != mem) {                                        \
                     AV_WN32A(pmv, m);                                  \
@@ -111,7 +112,7 @@ static void find_ref_mvs(VP9Context *s,
                 }                                                      \
                 mem_sub8x8 = AV_RN32A(&mv);                            \
             } else if (mem_sub8x8 != AV_RN32A(&mv)) {                  \
-                clamp_mv(&tmp, &mv, s);                                \
+                clamp_mv(&tmp, &mv, td);                               \
                 m = AV_RN32A(&tmp);                                    \
                 if (m != mem) {                                        \
                     AV_WN32A(pmv, m);                                  \
@@ -124,12 +125,12 @@ static void find_ref_mvs(VP9Context *s,
         } else {                                                       \
             uint32_t m = AV_RN32A(&mv);                                \
             if (!idx) {                                                \
-                clamp_mv(pmv, &mv, s);                                 \
+                clamp_mv(pmv, &mv, td);                                \
                 return;                                                \
             } else if (mem == INVALID_MV) {                            \
                 mem = m;                                               \
             } else if (m != mem) {                                     \
-                clamp_mv(pmv, &mv, s);                                 \
+                clamp_mv(pmv, &mv, td);                                \
                 return;                                                \
             }                                                          \
         }                                                              \
@@ -142,12 +143,12 @@ static void find_ref_mvs(VP9Context *s,
             else if (mv->ref[1] == ref)
                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][1]);
         }
-        if (col > s->tile_col_start) {
+        if (col > td->tile_col_start) {
             VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[row * s->sb_cols * 8 + col - 1];
             if (mv->ref[0] == ref)
-                RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][0]);
+                RETURN_MV(td->left_mv_ctx[2 * row7 + (sb >> 1)][0]);
             else if (mv->ref[1] == ref)
-                RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][1]);
+                RETURN_MV(td->left_mv_ctx[2 * row7 + (sb >> 1)][1]);
         }
         i = 2;
     } else {
@@ -158,7 +159,7 @@ static void find_ref_mvs(VP9Context *s,
     for (; i < 8; i++) {
         int c = p[i][0] + col, r = p[i][1] + row;
 
-        if (c >= s->tile_col_start && c < s->cols &&
+        if (c >= td->tile_col_start && c < s->cols &&
             r >= 0 && r < s->rows) {
             VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];
 
@@ -195,7 +196,7 @@ static void find_ref_mvs(VP9Context *s,
     for (i = 0; i < 8; i++) {
         int c = p[i][0] + col, r = p[i][1] + row;
 
-        if (c >= s->tile_col_start && c < s->cols && r >= 0 && r < s->rows) {
+        if (c >= td->tile_col_start && c < s->cols && r >= 0 && r < s->rows) {
             VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];
 
             if (mv->ref[0] != ref && mv->ref[0] >= 0)
@@ -226,69 +227,71 @@ static void find_ref_mvs(VP9Context *s,
     }
 
     AV_ZERO32(pmv);
-    clamp_mv(pmv, pmv, s);
+    clamp_mv(pmv, pmv, td);
 #undef INVALID_MV
 #undef RETURN_MV
 #undef RETURN_SCALE_MV
 }
 
-static av_always_inline int read_mv_component(VP9Context *s, int idx, int hp)
+static av_always_inline int read_mv_component(VP9TileData *td, int idx, int hp)
 {
-    int bit, sign = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].sign);
-    int n, c = vp8_rac_get_tree(&s->c, ff_vp9_mv_class_tree,
+    VP9Context *s = td->s;
+    int bit, sign = vp56_rac_get_prob(td->c, s->prob.p.mv_comp[idx].sign);
+    int n, c = vp8_rac_get_tree(td->c, ff_vp9_mv_class_tree,
                                 s->prob.p.mv_comp[idx].classes);
 
-    s->counts.mv_comp[idx].sign[sign]++;
-    s->counts.mv_comp[idx].classes[c]++;
+    td->counts.mv_comp[idx].sign[sign]++;
+    td->counts.mv_comp[idx].classes[c]++;
     if (c) {
         int m;
 
         for (n = 0, m = 0; m < c; m++) {
-            bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].bits[m]);
+            bit = vp56_rac_get_prob(td->c, s->prob.p.mv_comp[idx].bits[m]);
             n |= bit << m;
-            s->counts.mv_comp[idx].bits[m][bit]++;
+            td->counts.mv_comp[idx].bits[m][bit]++;
         }
         n <<= 3;
-        bit = vp8_rac_get_tree(&s->c, ff_vp9_mv_fp_tree,
+        bit = vp8_rac_get_tree(td->c, ff_vp9_mv_fp_tree,
                                s->prob.p.mv_comp[idx].fp);
         n  |= bit << 1;
-        s->counts.mv_comp[idx].fp[bit]++;
+        td->counts.mv_comp[idx].fp[bit]++;
         if (hp) {
-            bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].hp);
-            s->counts.mv_comp[idx].hp[bit]++;
+            bit = vp56_rac_get_prob(td->c, s->prob.p.mv_comp[idx].hp);
+            td->counts.mv_comp[idx].hp[bit]++;
             n |= bit;
         } else {
             n |= 1;
             // bug in libvpx - we count for bw entropy purposes even if the
             // bit wasn't coded
-            s->counts.mv_comp[idx].hp[1]++;
+            td->counts.mv_comp[idx].hp[1]++;
         }
         n += 8 << c;
     } else {
-        n = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0);
-        s->counts.mv_comp[idx].class0[n]++;
-        bit = vp8_rac_get_tree(&s->c, ff_vp9_mv_fp_tree,
+        n = vp56_rac_get_prob(td->c, s->prob.p.mv_comp[idx].class0);
+        td->counts.mv_comp[idx].class0[n]++;
+        bit = vp8_rac_get_tree(td->c, ff_vp9_mv_fp_tree,
                                s->prob.p.mv_comp[idx].class0_fp[n]);
-        s->counts.mv_comp[idx].class0_fp[n][bit]++;
+        td->counts.mv_comp[idx].class0_fp[n][bit]++;
         n = (n << 3) | (bit << 1);
         if (hp) {
-            bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0_hp);
-            s->counts.mv_comp[idx].class0_hp[bit]++;
+            bit = vp56_rac_get_prob(td->c, s->prob.p.mv_comp[idx].class0_hp);
+            td->counts.mv_comp[idx].class0_hp[bit]++;
             n |= bit;
         } else {
             n |= 1;
             // bug in libvpx - we count for bw entropy purposes even if the
             // bit wasn't coded
-            s->counts.mv_comp[idx].class0_hp[1]++;
+            td->counts.mv_comp[idx].class0_hp[1]++;
         }
     }
 
     return sign ? -(n + 1) : (n + 1);
 }
 
-void ff_vp9_fill_mv(VP9Context *s, VP56mv *mv, int mode, int sb)
+void ff_vp9_fill_mv(VP9TileData *td, VP56mv *mv, int mode, int sb)
 {
-    VP9Block *b = s->b;
+    VP9Context *s = td->s;
+    VP9Block *b = td->b;
 
     if (mode == ZEROMV) {
         AV_ZERO64(mv);
@@ -296,7 +299,7 @@ void ff_vp9_fill_mv(VP9Context *s, VP56mv *mv, int mode, int sb)
         int hp;
 
         // FIXME cache this value and reuse for other subblocks
-        find_ref_mvs(s, &mv[0], b->ref[0], 0, mode == NEARMV,
+        find_ref_mvs(td, &mv[0], b->ref[0], 0, mode == NEARMV,
                      mode == NEWMV ? -1 : sb);
         // FIXME maybe move this code into find_ref_mvs()
         if ((mode == NEWMV || sb == -1) &&
@@ -316,19 +319,19 @@ void ff_vp9_fill_mv(VP9Context *s, VP56mv *mv, int mode, int sb)
             }
         }
         if (mode == NEWMV) {
-            enum MVJoint j = vp8_rac_get_tree(&s->c, ff_vp9_mv_joint_tree,
+            enum MVJoint j = vp8_rac_get_tree(td->c, ff_vp9_mv_joint_tree,
                                               s->prob.p.mv_joint);
 
-            s->counts.mv_joint[j]++;
+            td->counts.mv_joint[j]++;
             if (j >= MV_JOINT_V)
-                mv[0].y += read_mv_component(s, 0, hp);
+                mv[0].y += read_mv_component(td, 0, hp);
             if (j & 1)
-                mv[0].x += read_mv_component(s, 1, hp);
+                mv[0].x += read_mv_component(td, 1, hp);
         }
 
         if (b->comp) {
             // FIXME cache this value and reuse for other subblocks
-            find_ref_mvs(s, &mv[1], b->ref[1], 1, mode == NEARMV,
+            find_ref_mvs(td, &mv[1], b->ref[1], 1, mode == NEARMV,
                          mode == NEWMV ? -1 : sb);
             if ((mode == NEWMV || sb == -1) &&
                 !(hp = s->s.h.highprecisionmvs &&
@@ -347,14 +350,14 @@ void ff_vp9_fill_mv(VP9Context *s, VP56mv *mv, int mode, int sb)
                 }
             }
             if (mode == NEWMV) {
-                enum MVJoint j = vp8_rac_get_tree(&s->c, ff_vp9_mv_joint_tree,
+                enum MVJoint j = vp8_rac_get_tree(td->c, ff_vp9_mv_joint_tree,
                                                   s->prob.p.mv_joint);
 
-                s->counts.mv_joint[j]++;
+                td->counts.mv_joint[j]++;
                 if (j >= MV_JOINT_V)
-                    mv[1].y += read_mv_component(s, 0, hp);
+                    mv[1].y += read_mv_component(td, 0, hp);
                 if (j & 1)
-                    mv[1].x += read_mv_component(s, 1, hp);
+                    mv[1].x += read_mv_component(td, 1, hp);
             }
         }
     }
diff --git a/libavcodec/vp9prob.c b/libavcodec/vp9prob.c
index cde909c..fb295b4 100644
--- a/libavcodec/vp9prob.c
+++ b/libavcodec/vp9prob.c
@@ -56,8 +56,8 @@ void ff_vp9_adapt_probs(VP9Context *s)
                 for (l = 0; l < 6; l++)
                     for (m = 0; m < 6; m++) {
                         uint8_t *pp = s->prob_ctx[s->s.h.framectxid].coef[i][j][k][l][m];
-                        unsigned *e = s->counts.eob[i][j][k][l][m];
-                        unsigned *c = s->counts.coef[i][j][k][l][m];
+                        unsigned *e = s->td[0].counts.eob[i][j][k][l][m];
+                        unsigned *c = s->td[0].counts.coef[i][j][k][l][m];
 
                         if (l == 0 && m >= 3) // dc only has 3 pt
                             break;
@@ -77,32 +77,32 @@ void ff_vp9_adapt_probs(VP9Context *s)
 
     // skip flag
     for (i = 0; i < 3; i++)
-        adapt_prob(&p->skip[i], s->counts.skip[i][0],
-                   s->counts.skip[i][1], 20, 128);
+        adapt_prob(&p->skip[i], s->td[0].counts.skip[i][0],
+                   s->td[0].counts.skip[i][1], 20, 128);
 
     // intra/inter flag
     for (i = 0; i < 4; i++)
-        adapt_prob(&p->intra[i], s->counts.intra[i][0],
-                   s->counts.intra[i][1], 20, 128);
+        adapt_prob(&p->intra[i], s->td[0].counts.intra[i][0],
+                   s->td[0].counts.intra[i][1], 20, 128);
 
     // comppred flag
     if (s->s.h.comppredmode == PRED_SWITCHABLE) {
         for (i = 0; i < 5; i++)
-            adapt_prob(&p->comp[i], s->counts.comp[i][0],
-                       s->counts.comp[i][1], 20, 128);
+            adapt_prob(&p->comp[i], s->td[0].counts.comp[i][0],
+                       s->td[0].counts.comp[i][1], 20, 128);
     }
 
     // reference frames
     if (s->s.h.comppredmode != PRED_SINGLEREF) {
         for (i = 0; i < 5; i++)
-            adapt_prob(&p->comp_ref[i], s->counts.comp_ref[i][0],
-                       s->counts.comp_ref[i][1], 20, 128);
+            adapt_prob(&p->comp_ref[i], s->td[0].counts.comp_ref[i][0],
+                       s->td[0].counts.comp_ref[i][1], 20, 128);
     }
 
     if (s->s.h.comppredmode != PRED_COMPREF) {
         for (i = 0; i < 5; i++) {
             uint8_t *pp = p->single_ref[i];
-            unsigned (*c)[2] = s->counts.single_ref[i];
+            unsigned (*c)[2] = s->td[0].counts.single_ref[i];
 
             adapt_prob(&pp[0], c[0][0], c[0][1], 20, 128);
             adapt_prob(&pp[1], c[1][0], c[1][1], 20, 128);
@@ -113,7 +113,7 @@ void ff_vp9_adapt_probs(VP9Context *s)
     for (i = 0; i < 4; i++)
         for (j = 0; j < 4; j++) {
             uint8_t *pp = p->partition[i][j];
-            unsigned *c = s->counts.partition[i][j];
+            unsigned *c = s->td[0].counts.partition[i][j];
 
             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);
             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);
@@ -123,10 +123,10 @@ void ff_vp9_adapt_probs(VP9Context *s)
     // tx size
     if (s->s.h.txfmmode == TX_SWITCHABLE) {
         for (i = 0; i < 2; i++) {
-            unsigned *c16 = s->counts.tx16p[i], *c32 = s->counts.tx32p[i];
+            unsigned *c16 = s->td[0].counts.tx16p[i], *c32 = s->td[0].counts.tx32p[i];
 
-            adapt_prob(&p->tx8p[i], s->counts.tx8p[i][0],
-                       s->counts.tx8p[i][1], 20, 128);
+            adapt_prob(&p->tx8p[i], s->td[0].counts.tx8p[i][0],
+                       s->td[0].counts.tx8p[i][1], 20, 128);
             adapt_prob(&p->tx16p[i][0], c16[0], c16[1] + c16[2], 20, 128);
             adapt_prob(&p->tx16p[i][1], c16[1], c16[2], 20, 128);
             adapt_prob(&p->tx32p[i][0], c32[0], c32[1] + c32[2] + c32[3], 20, 128);
@@ -139,7 +139,7 @@ void ff_vp9_adapt_probs(VP9Context *s)
     if (s->s.h.filtermode == FILTER_SWITCHABLE) {
         for (i = 0; i < 4; i++) {
             uint8_t *pp = p->filter[i];
-            unsigned *c = s->counts.filter[i];
+            unsigned *c = s->td[0].counts.filter[i];
 
             adapt_prob(&pp[0], c[0], c[1] + c[2], 20, 128);
             adapt_prob(&pp[1], c[1], c[2], 20, 128);
@@ -149,7 +149,7 @@ void ff_vp9_adapt_probs(VP9Context *s)
     // inter modes
     for (i = 0; i < 7; i++) {
         uint8_t *pp = p->mv_mode[i];
-        unsigned *c = s->counts.mv_mode[i];
+        unsigned *c = s->td[0].counts.mv_mode[i];
 
         adapt_prob(&pp[0], c[2], c[1] + c[0] + c[3], 20, 128);
         adapt_prob(&pp[1], c[0], c[1] + c[3], 20, 128);
@@ -159,7 +159,7 @@ void ff_vp9_adapt_probs(VP9Context *s)
     // mv joints
     {
         uint8_t *pp = p->mv_joint;
-        unsigned *c = s->counts.mv_joint;
+        unsigned *c = s->td[0].counts.mv_joint;
 
         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);
         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);
@@ -171,11 +171,11 @@ void ff_vp9_adapt_probs(VP9Context *s)
         uint8_t *pp;
         unsigned *c, (*c2)[2], sum;
 
-        adapt_prob(&p->mv_comp[i].sign, s->counts.mv_comp[i].sign[0],
-                   s->counts.mv_comp[i].sign[1], 20, 128);
+        adapt_prob(&p->mv_comp[i].sign, s->td[0].counts.mv_comp[i].sign[0],
+                   s->td[0].counts.mv_comp[i].sign[1], 20, 128);
 
         pp  = p->mv_comp[i].classes;
-        c   = s->counts.mv_comp[i].classes;
+        c   = s->td[0].counts.mv_comp[i].classes;
         sum = c[1] + c[2] + c[3] + c[4] + c[5] +
               c[6] + c[7] + c[8] + c[9] + c[10];
         adapt_prob(&pp[0], c[0], sum, 20, 128);
@@ -193,39 +193,39 @@ void ff_vp9_adapt_probs(VP9Context *s)
         adapt_prob(&pp[8], c[7], c[8], 20, 128);
         adapt_prob(&pp[9], c[9], c[10], 20, 128);
 
-        adapt_prob(&p->mv_comp[i].class0, s->counts.mv_comp[i].class0[0],
-                   s->counts.mv_comp[i].class0[1], 20, 128);
+        adapt_prob(&p->mv_comp[i].class0, s->td[0].counts.mv_comp[i].class0[0],
+                   s->td[0].counts.mv_comp[i].class0[1], 20, 128);
         pp = p->mv_comp[i].bits;
-        c2 = s->counts.mv_comp[i].bits;
+        c2 = s->td[0].counts.mv_comp[i].bits;
         for (j = 0; j < 10; j++)
             adapt_prob(&pp[j], c2[j][0], c2[j][1], 20, 128);
 
         for (j = 0; j < 2; j++) {
             pp = p->mv_comp[i].class0_fp[j];
-            c  = s->counts.mv_comp[i].class0_fp[j];
+            c  = s->td[0].counts.mv_comp[i].class0_fp[j];
             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);
             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);
             adapt_prob(&pp[2], c[2], c[3], 20, 128);
         }
         pp = p->mv_comp[i].fp;
-        c  = s->counts.mv_comp[i].fp;
+        c  = s->td[0].counts.mv_comp[i].fp;
         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);
         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);
         adapt_prob(&pp[2], c[2], c[3], 20, 128);
 
         if (s->s.h.highprecisionmvs) {
             adapt_prob(&p->mv_comp[i].class0_hp,
-                       s->counts.mv_comp[i].class0_hp[0],
-                       s->counts.mv_comp[i].class0_hp[1], 20, 128);
-            adapt_prob(&p->mv_comp[i].hp, s->counts.mv_comp[i].hp[0],
-                       s->counts.mv_comp[i].hp[1], 20, 128);
+                       s->td[0].counts.mv_comp[i].class0_hp[0],
+                       s->td[0].counts.mv_comp[i].class0_hp[1], 20, 128);
+            adapt_prob(&p->mv_comp[i].hp, s->td[0].counts.mv_comp[i].hp[0],
+                       s->td[0].counts.mv_comp[i].hp[1], 20, 128);
         }
     }
 
     // y intra modes
     for (i = 0; i < 4; i++) {
         uint8_t *pp = p->y_mode[i];
-        unsigned *c = s->counts.y_mode[i], sum, s2;
+        unsigned *c = s->td[0].counts.y_mode[i], sum, s2;
 
         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];
         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);
@@ -250,7 +250,7 @@ void ff_vp9_adapt_probs(VP9Context *s)
     // uv intra modes
     for (i = 0; i < 10; i++) {
         uint8_t *pp = p->uv_mode[i];
-        unsigned *c = s->counts.uv_mode[i], sum, s2;
+        unsigned *c = s->td[0].counts.uv_mode[i], sum, s2;
 
         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];
         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);
diff --git a/libavcodec/vp9recon.c b/libavcodec/vp9recon.c
index afdb513..49bb04e 100644
--- a/libavcodec/vp9recon.c
+++ b/libavcodec/vp9recon.c
@@ -29,15 +29,16 @@
 #include "vp9data.h"
 #include "vp9dec.h"
 
-static av_always_inline int check_intra_mode(VP9Context *s, int mode, uint8_t **a,
+static av_always_inline int check_intra_mode(VP9TileData *td, int mode, uint8_t **a,
                                              uint8_t *dst_edge, ptrdiff_t stride_edge,
                                              uint8_t *dst_inner, ptrdiff_t stride_inner,
                                              uint8_t *l, int col, int x, int w,
                                              int row, int y, enum TxfmMode tx,
                                              int p, int ss_h, int ss_v, int bytesperpixel)
 {
+    VP9Context *s = td->s;
     int have_top = row > 0 || y > 0;
-    int have_left = col > s->tile_col_start || x > 0;
+    int have_left = col > td->tile_col_start || x > 0;
     int have_right = x < w - 1;
     int bpp = s->s.h.bpp;
     static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {
@@ -214,19 +215,19 @@ static av_always_inline int check_intra_mode(VP9Context *s, int mode, uint8_t **
     return mode;
 }
 
-static av_always_inline void intra_recon(AVCodecContext *avctx, ptrdiff_t y_off,
+static av_always_inline void intra_recon(VP9TileData *td, ptrdiff_t y_off,
                                          ptrdiff_t uv_off, int bytesperpixel)
 {
-    VP9Context *s = avctx->priv_data;
-    VP9Block *b = s->b;
-    int row = s->row, col = s->col;
+    VP9Context *s = td->s;
+    VP9Block *b = td->b;
+    int row = td->row, col = td->col;
     int w4 = ff_vp9_bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;
     int h4 = ff_vp9_bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);
     int end_x = FFMIN(2 * (s->cols - col), w4);
     int end_y = FFMIN(2 * (s->rows - row), h4);
     int tx = 4 * s->s.h.lossless + b->tx, uvtx = b->uvtx + 4 * s->s.h.lossless;
     int uvstep1d = 1 << b->uvtx, p;
-    uint8_t *dst = s->dst[0], *dst_r = s->s.frames[CUR_FRAME].tf.f->data[0] + y_off;
+    uint8_t *dst = td->dst[0], *dst_r = s->s.frames[CUR_FRAME].tf.f->data[0] + y_off;
     LOCAL_ALIGNED_32(uint8_t, a_buf, [96]);
     LOCAL_ALIGNED_32(uint8_t, l, [64]);
 
@@ -238,19 +239,19 @@ static av_always_inline void intra_recon(AVCodecContext *avctx, ptrdiff_t y_off,
                                y * 2 + x : 0];
             uint8_t *a = &a_buf[32];
             enum TxfmType txtp = ff_vp9_intra_txfm_type[mode];
-            int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];
+            int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&td->eob[n]) : td->eob[n];
 
-            mode = check_intra_mode(s, mode, &a, ptr_r,
+            mode = check_intra_mode(td, mode, &a, ptr_r,
                                     s->s.frames[CUR_FRAME].tf.f->linesize[0],
-                                    ptr, s->y_stride, l,
+                                    ptr, td->y_stride, l,
                                     col, x, w4, row, y, b->tx, 0, 0, 0, bytesperpixel);
-            s->dsp.intra_pred[b->tx][mode](ptr, s->y_stride, l, a);
+            s->dsp.intra_pred[b->tx][mode](ptr, td->y_stride, l, a);
             if (eob)
-                s->dsp.itxfm_add[tx][txtp](ptr, s->y_stride,
-                                           s->block + 16 * n * bytesperpixel, eob);
+                s->dsp.itxfm_add[tx][txtp](ptr, td->y_stride,
+                                           td->block + 16 * n * bytesperpixel, eob);
         }
         dst_r += 4 * step1d * s->s.frames[CUR_FRAME].tf.f->linesize[0];
-        dst   += 4 * step1d * s->y_stride;
+        dst   += 4 * step1d * td->y_stride;
     }
 
     // U/V
@@ -259,7 +260,7 @@ static av_always_inline void intra_recon(AVCodecContext *avctx, ptrdiff_t y_off,
     end_y >>= s->ss_v;
     step = 1 << (b->uvtx * 2);
     for (p = 0; p < 2; p++) {
-        dst   = s->dst[1 + p];
+        dst   = td->dst[1 + p];
         dst_r = s->s.frames[CUR_FRAME].tf.f->data[1 + p] + uv_off;
         for (n = 0, y = 0; y < end_y; y += uvstep1d) {
             uint8_t *ptr = dst, *ptr_r = dst_r;
@@ -267,40 +268,41 @@ static av_always_inline void intra_recon(AVCodecContext *avctx, ptrdiff_t y_off,
                                    ptr_r += 4 * uvstep1d * bytesperpixel, n += step) {
                 int mode = b->uvmode;
                 uint8_t *a = &a_buf[32];
-                int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];
+                int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&td->uveob[p][n]) : td->uveob[p][n];
 
-                mode = check_intra_mode(s, mode, &a, ptr_r,
+                mode = check_intra_mode(td, mode, &a, ptr_r,
                                         s->s.frames[CUR_FRAME].tf.f->linesize[1],
-                                        ptr, s->uv_stride, l, col, x, w4, row, y,
+                                        ptr, td->uv_stride, l, col, x, w4, row, y,
                                         b->uvtx, p + 1, s->ss_h, s->ss_v, bytesperpixel);
-                s->dsp.intra_pred[b->uvtx][mode](ptr, s->uv_stride, l, a);
+                s->dsp.intra_pred[b->uvtx][mode](ptr, td->uv_stride, l, a);
                 if (eob)
-                    s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,
-                                                    s->uvblock[p] + 16 * n * bytesperpixel, eob);
+                    s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, td->uv_stride,
+                                                    td->uvblock[p] + 16 * n * bytesperpixel, eob);
             }
             dst_r += 4 * uvstep1d * s->s.frames[CUR_FRAME].tf.f->linesize[1];
-            dst   += 4 * uvstep1d * s->uv_stride;
+            dst   += 4 * uvstep1d * td->uv_stride;
         }
     }
 }
 
-void ff_vp9_intra_recon_8bpp(AVCodecContext *avctx, ptrdiff_t y_off, ptrdiff_t uv_off)
+void ff_vp9_intra_recon_8bpp(VP9TileData *td, ptrdiff_t y_off, ptrdiff_t uv_off)
 {
-    intra_recon(avctx, y_off, uv_off, 1);
+    intra_recon(td, y_off, uv_off, 1);
 }
 
-void ff_vp9_intra_recon_16bpp(AVCodecContext *avctx, ptrdiff_t y_off, ptrdiff_t uv_off)
+void ff_vp9_intra_recon_16bpp(VP9TileData *td, ptrdiff_t y_off, ptrdiff_t uv_off)
 {
-    intra_recon(avctx, y_off, uv_off, 2);
+    intra_recon(td, y_off, uv_off, 2);
 }
 
-static av_always_inline void mc_luma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2],
+static av_always_inline void mc_luma_unscaled(VP9TileData *td, vp9_mc_func (*mc)[2],
                                               uint8_t *dst, ptrdiff_t dst_stride,
                                               const uint8_t *ref, ptrdiff_t ref_stride,
                                               ThreadFrame *ref_frame,
                                               ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,
                                               int bw, int bh, int w, int h, int bytesperpixel)
 {
+    VP9Context *s = td->s;
     int mx = mv->x, my = mv->y, th;
 
     y += my >> 3;
@@ -318,18 +320,18 @@ static av_always_inline void mc_luma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2
     // (!!my * 5) than horizontally (!!mx * 4).
     if (x < !!mx * 3 || y < !!my * 3 ||
         x + !!mx * 4 > w - bw || y + !!my * 5 > h - bh) {
-        s->vdsp.emulated_edge_mc(s->edge_emu_buffer,
+        s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
                                  ref - !!my * 3 * ref_stride - !!mx * 3 * bytesperpixel,
                                  160, ref_stride,
                                  bw + !!mx * 7, bh + !!my * 7,
                                  x - !!mx * 3, y - !!my * 3, w, h);
-        ref = s->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;
+        ref = td->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;
         ref_stride = 160;
     }
     mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);
 }
 
-static av_always_inline void mc_chroma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2],
+static av_always_inline void mc_chroma_unscaled(VP9TileData *td, vp9_mc_func (*mc)[2],
                                                 uint8_t *dst_u, uint8_t *dst_v,
                                                 ptrdiff_t dst_stride,
                                                 const uint8_t *ref_u, ptrdiff_t src_stride_u,
@@ -338,6 +340,7 @@ static av_always_inline void mc_chroma_unscaled(VP9Context *s, vp9_mc_func (*mc)
                                                 ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,
                                                 int bw, int bh, int w, int h, int bytesperpixel)
 {
+    VP9Context *s = td->s;
     int mx = mv->x * (1 << !s->ss_h), my = mv->y * (1 << !s->ss_v), th;
 
     y += my >> 4;
@@ -356,20 +359,20 @@ static av_always_inline void mc_chroma_unscaled(VP9Context *s, vp9_mc_func (*mc)
     // (!!my * 5) than horizontally (!!mx * 4).
     if (x < !!mx * 3 || y < !!my * 3 ||
         x + !!mx * 4 > w - bw || y + !!my * 5 > h - bh) {
-        s->vdsp.emulated_edge_mc(s->edge_emu_buffer,
+        s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
                                  ref_u - !!my * 3 * src_stride_u - !!mx * 3 * bytesperpixel,
                                  160, src_stride_u,
                                  bw + !!mx * 7, bh + !!my * 7,
                                  x - !!mx * 3, y - !!my * 3, w, h);
-        ref_u = s->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;
+        ref_u = td->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;
         mc[!!mx][!!my](dst_u, dst_stride, ref_u, 160, bh, mx, my);
 
-        s->vdsp.emulated_edge_mc(s->edge_emu_buffer,
+        s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
                                  ref_v - !!my * 3 * src_stride_v - !!mx * 3 * bytesperpixel,
                                  160, src_stride_v,
                                  bw + !!mx * 7, bh + !!my * 7,
                                  x - !!mx * 3, y - !!my * 3, w, h);
-        ref_v = s->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;
+        ref_v = td->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;
         mc[!!mx][!!my](dst_v, dst_stride, ref_v, 160, bh, mx, my);
     } else {
         mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);
@@ -377,13 +380,13 @@ static av_always_inline void mc_chroma_unscaled(VP9Context *s, vp9_mc_func (*mc)
     }
 }
 
-#define mc_luma_dir(s, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, \
+#define mc_luma_dir(td, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, \
                     px, py, pw, ph, bw, bh, w, h, i) \
-    mc_luma_unscaled(s, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \
+    mc_luma_unscaled(td, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \
                      mv, bw, bh, w, h, bytesperpixel)
-#define mc_chroma_dir(s, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
+#define mc_chroma_dir(td, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
                       row, col, mv, px, py, pw, ph, bw, bh, w, h, i) \
-    mc_chroma_unscaled(s, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
+    mc_chroma_unscaled(td, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
                        row, col, mv, bw, bh, w, h, bytesperpixel)
 #define SCALED 0
 #define FN(x) x##_8bpp
@@ -400,7 +403,7 @@ static av_always_inline void mc_chroma_unscaled(VP9Context *s, vp9_mc_func (*mc)
 #undef BYTES_PER_PIXEL
 #undef SCALED
 
-static av_always_inline void mc_luma_scaled(VP9Context *s, vp9_scaled_mc_func smc,
+static av_always_inline void mc_luma_scaled(VP9TileData *td, vp9_scaled_mc_func smc,
                                             vp9_mc_func (*mc)[2],
                                             uint8_t *dst, ptrdiff_t dst_stride,
                                             const uint8_t *ref, ptrdiff_t ref_stride,
@@ -410,9 +413,10 @@ static av_always_inline void mc_luma_scaled(VP9Context *s, vp9_scaled_mc_func sm
                                             int bw, int bh, int w, int h, int bytesperpixel,
                                             const uint16_t *scale, const uint8_t *step)
 {
+    VP9Context *s = td->s;
     if (s->s.frames[CUR_FRAME].tf.f->width == ref_frame->f->width &&
         s->s.frames[CUR_FRAME].tf.f->height == ref_frame->f->height) {
-        mc_luma_unscaled(s, mc, dst, dst_stride, ref, ref_stride, ref_frame,
+        mc_luma_unscaled(td, mc, dst, dst_stride, ref, ref_stride, ref_frame,
                          y, x, in_mv, bw, bh, w, h, bytesperpixel);
     } else {
 #define scale_mv(n, dim) (((int64_t)(n) * scale[dim]) >> 14)
@@ -445,19 +449,19 @@ static av_always_inline void mc_luma_scaled(VP9Context *s, vp9_scaled_mc_func sm
     // needed, so switch to emulated edge one pixel sooner vertically
     // (y + 5 >= h - refbh_m1) than horizontally (x + 4 >= w - refbw_m1).
     if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 5 >= h - refbh_m1) {
-        s->vdsp.emulated_edge_mc(s->edge_emu_buffer,
+        s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
                                  ref - 3 * ref_stride - 3 * bytesperpixel,
                                  288, ref_stride,
                                  refbw_m1 + 8, refbh_m1 + 8,
                                  x - 3, y - 3, w, h);
-        ref = s->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;
+        ref = td->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;
         ref_stride = 288;
     }
     smc(dst, dst_stride, ref, ref_stride, bh, mx, my, step[0], step[1]);
     }
 }
 
-static av_always_inline void mc_chroma_scaled(VP9Context *s, vp9_scaled_mc_func smc,
+static av_always_inline void mc_chroma_scaled(VP9TileData *td, vp9_scaled_mc_func smc,
                                               vp9_mc_func (*mc)[2],
                                               uint8_t *dst_u, uint8_t *dst_v,
                                               ptrdiff_t dst_stride,
@@ -469,9 +473,10 @@ static av_always_inline void mc_chroma_scaled(VP9Context *s, vp9_scaled_mc_func
                                               int bw, int bh, int w, int h, int bytesperpixel,
                                               const uint16_t *scale, const uint8_t *step)
 {
+    VP9Context *s = td->s;
     if (s->s.frames[CUR_FRAME].tf.f->width == ref_frame->f->width &&
         s->s.frames[CUR_FRAME].tf.f->height == ref_frame->f->height) {
-        mc_chroma_unscaled(s, mc, dst_u, dst_v, dst_stride, ref_u, src_stride_u,
+        mc_chroma_unscaled(td, mc, dst_u, dst_v, dst_stride, ref_u, src_stride_u,
                            ref_v, src_stride_v, ref_frame,
                            y, x, in_mv, bw, bh, w, h, bytesperpixel);
     } else {
@@ -514,20 +519,20 @@ static av_always_inline void mc_chroma_scaled(VP9Context *s, vp9_scaled_mc_func
     // needed, so switch to emulated edge one pixel sooner vertically
     // (y + 5 >= h - refbh_m1) than horizontally (x + 4 >= w - refbw_m1).
     if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 5 >= h - refbh_m1) {
-        s->vdsp.emulated_edge_mc(s->edge_emu_buffer,
+        s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
                                  ref_u - 3 * src_stride_u - 3 * bytesperpixel,
                                  288, src_stride_u,
                                  refbw_m1 + 8, refbh_m1 + 8,
                                  x - 3, y - 3, w, h);
-        ref_u = s->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;
+        ref_u = td->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;
         smc(dst_u, dst_stride, ref_u, 288, bh, mx, my, step[0], step[1]);
 
-        s->vdsp.emulated_edge_mc(s->edge_emu_buffer,
+        s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
                                  ref_v - 3 * src_stride_v - 3 * bytesperpixel,
                                  288, src_stride_v,
                                  refbw_m1 + 8, refbh_m1 + 8,
                                  x - 3, y - 3, w, h);
-        ref_v = s->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;
+        ref_v = td->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;
         smc(dst_v, dst_stride, ref_v, 288, bh, mx, my, step[0], step[1]);
     } else {
         smc(dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my, step[0], step[1]);
@@ -536,14 +541,14 @@ static av_always_inline void mc_chroma_scaled(VP9Context *s, vp9_scaled_mc_func
     }
 }
 
-#define mc_luma_dir(s, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, \
+#define mc_luma_dir(td, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, \
                     px, py, pw, ph, bw, bh, w, h, i) \
-    mc_luma_scaled(s, s->dsp.s##mc, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \
+    mc_luma_scaled(td, s->dsp.s##mc, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \
                    mv, px, py, pw, ph, bw, bh, w, h, bytesperpixel, \
                    s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])
-#define mc_chroma_dir(s, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
+#define mc_chroma_dir(td, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
                       row, col, mv, px, py, pw, ph, bw, bh, w, h, i) \
-    mc_chroma_scaled(s, s->dsp.s##mc, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
+    mc_chroma_scaled(td, s->dsp.s##mc, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
                      row, col, mv, px, py, pw, ph, bw, bh, w, h, bytesperpixel, \
                      s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])
 #define SCALED 1
@@ -561,23 +566,23 @@ static av_always_inline void mc_chroma_scaled(VP9Context *s, vp9_scaled_mc_func
 #undef BYTES_PER_PIXEL
 #undef SCALED
 
-static av_always_inline void inter_recon(AVCodecContext *avctx, int bytesperpixel)
+static av_always_inline void inter_recon(VP9TileData *td, int bytesperpixel)
 {
-    VP9Context *s = avctx->priv_data;
-    VP9Block *b = s->b;
-    int row = s->row, col = s->col;
+    VP9Context *s = td->s;
+    VP9Block *b = td->b;
+    int row = td->row, col = td->col;
 
     if (s->mvscale[b->ref[0]][0] || (b->comp && s->mvscale[b->ref[1]][0])) {
         if (bytesperpixel == 1) {
-            inter_pred_scaled_8bpp(avctx);
+            inter_pred_scaled_8bpp(td);
         } else {
-            inter_pred_scaled_16bpp(avctx);
+            inter_pred_scaled_16bpp(td);
         }
     } else {
         if (bytesperpixel == 1) {
-            inter_pred_8bpp(avctx);
+            inter_pred_8bpp(td);
         } else {
-            inter_pred_16bpp(avctx);
+            inter_pred_16bpp(td);
         }
     }
 
@@ -590,20 +595,20 @@ static av_always_inline void inter_recon(AVCodecContext *avctx, int bytesperpixe
         int end_y = FFMIN(2 * (s->rows - row), h4);
         int tx = 4 * s->s.h.lossless + b->tx, uvtx = b->uvtx + 4 * s->s.h.lossless;
         int uvstep1d = 1 << b->uvtx, p;
-        uint8_t *dst = s->dst[0];
+        uint8_t *dst = td->dst[0];
 
         // y itxfm add
         for (n = 0, y = 0; y < end_y; y += step1d) {
             uint8_t *ptr = dst;
             for (x = 0; x < end_x; x += step1d,
                  ptr += 4 * step1d * bytesperpixel, n += step) {
-                int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];
+                int eob = b->tx > TX_8X8 ? AV_RN16A(&td->eob[n]) : td->eob[n];
 
                 if (eob)
-                    s->dsp.itxfm_add[tx][DCT_DCT](ptr, s->y_stride,
-                                                  s->block + 16 * n * bytesperpixel, eob);
+                    s->dsp.itxfm_add[tx][DCT_DCT](ptr, td->y_stride,
+                                                  td->block + 16 * n * bytesperpixel, eob);
             }
-            dst += 4 * s->y_stride * step1d;
+            dst += 4 * td->y_stride * step1d;
         }
 
         // uv itxfm add
@@ -611,29 +616,29 @@ static av_always_inline void inter_recon(AVCodecContext *avctx, int bytesperpixe
         end_y >>= s->ss_v;
         step = 1 << (b->uvtx * 2);
         for (p = 0; p < 2; p++) {
-            dst = s->dst[p + 1];
+            dst = td->dst[p + 1];
             for (n = 0, y = 0; y < end_y; y += uvstep1d) {
                 uint8_t *ptr = dst;
                 for (x = 0; x < end_x; x += uvstep1d,
                      ptr += 4 * uvstep1d * bytesperpixel, n += step) {
-                    int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];
+                    int eob = b->uvtx > TX_8X8 ? AV_RN16A(&td->uveob[p][n]) : td->uveob[p][n];
 
                     if (eob)
-                        s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,
-                                                        s->uvblock[p] + 16 * n * bytesperpixel, eob);
+                        s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, td->uv_stride,
+                                                        td->uvblock[p] + 16 * n * bytesperpixel, eob);
                 }
-                dst += 4 * uvstep1d * s->uv_stride;
+                dst += 4 * uvstep1d * td->uv_stride;
             }
         }
     }
 }
 
-void ff_vp9_inter_recon_8bpp(AVCodecContext *avctx)
+void ff_vp9_inter_recon_8bpp(VP9TileData *td)
 {
-    inter_recon(avctx, 1);
+    inter_recon(td, 1);
 }
 
-void ff_vp9_inter_recon_16bpp(AVCodecContext *avctx)
+void ff_vp9_inter_recon_16bpp(VP9TileData *td)
 {
-    inter_recon(avctx, 2);
+    inter_recon(td, 2);
 }