avcodec/vp9: Add tile threading support

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

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 */

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);