Merge "Add high-precision sub-pixel search as a speed feature" into nextgenv2

configure

@@ -284,7 +284,6 @@ EXPERIMENT_LIST="
     ext_partition
     ext_tile
     obmc
-    affine_motion
 "
 CONFIG_LIST="
     dependency_tracking

vp10/encoder/encoder.h

@@ -286,12 +286,10 @@ typedef struct IMAGE_STAT {
   double worst;
 } ImageStat;
 
-#if CONFIG_AFFINE_MOTION
 typedef struct {
   int ref_count;
   YV12_BUFFER_CONFIG buf;
 } EncRefCntBuffer;
-#endif
 
 typedef struct VP10_COMP {
   QUANTS quants;
@@ -311,11 +309,9 @@ typedef struct VP10_COMP {
   YV12_BUFFER_CONFIG *unscaled_last_source;
   YV12_BUFFER_CONFIG scaled_last_source;
 
-#if CONFIG_AFFINE_MOTION
   // Up-sampled reference buffers
   EncRefCntBuffer upsampled_ref_bufs[MAX_REF_FRAMES];
   int upsampled_ref_idx[MAX_REF_FRAMES];
-#endif
 
   TileDataEnc *tile_data;
   int allocated_tiles;  // Keep track of memory allocated for tiles.
@@ -702,7 +698,6 @@ void vp10_new_framerate(VP10_COMP *cpi, double framerate);
 
 #define LAYER_IDS_TO_IDX(sl, tl, num_tl) ((sl) * (num_tl) + (tl))
 
-#if CONFIG_AFFINE_MOTION
 // Update up-sampled reference frame index.
 static INLINE void uref_cnt_fb(EncRefCntBuffer *ubufs, int *uidx,
                                int new_uidx) {
@@ -714,7 +709,6 @@ static INLINE void uref_cnt_fb(EncRefCntBuffer *ubufs, int *uidx,
   *uidx = new_uidx;
   ubufs[new_uidx].ref_count++;
 }
-#endif
 
 #ifdef __cplusplus
 }  // extern "C"

vp10/encoder/mbgraph.c

@@ -64,11 +64,7 @@ static unsigned int do_16x16_motion_iteration(VP10_COMP *cpi,
         &v_fn_ptr, 0, mv_sf->subpel_iters_per_step,
         cond_cost_list(cpi, cost_list),
         NULL, NULL,
-#if CONFIG_AFFINE_MOTION
         &distortion, &sse, NULL, 0, 0, 0);
-#else
-        &distortion, &sse, NULL, 0, 0);
-#endif
   }
 
 #if CONFIG_EXT_INTER

vp10/encoder/mcomp.c

@@ -210,7 +210,6 @@ static INLINE const uint8_t *pre(const uint8_t *buf, int stride, int r, int c) {
 
 #define CHECK_BETTER0(v, r, c) CHECK_BETTER(v, r, c)
 
-#if CONFIG_AFFINE_MOTION
 static INLINE const uint8_t *upre(const uint8_t *buf, int stride,
                                   int r, int c) {
   return &buf[(r) * stride + (c)];
@@ -232,7 +231,6 @@ static INLINE const uint8_t *upre(const uint8_t *buf, int stride,
   } else {                                                             \
     v = INT_MAX;                                                       \
   }
-#endif
 
 #define FIRST_LEVEL_CHECKS                              \
   {                                                     \
@@ -438,11 +436,7 @@ int vp10_find_best_sub_pixel_tree_pruned_evenmore(
     int *distortion,
     unsigned int *sse1,
     const uint8_t *second_pred,
-#if CONFIG_AFFINE_MOTION
     int w, int h, int use_upsampled_ref) {
-#else
-    int w, int h) {
-#endif
   SETUP_SUBPEL_SEARCH;
   besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp,
                                z, src_stride, y, y_stride, second_pred,
@@ -455,9 +449,7 @@ int vp10_find_best_sub_pixel_tree_pruned_evenmore(
   (void) allow_hp;
   (void) forced_stop;
   (void) hstep;
-#if CONFIG_AFFINE_MOTION
   (void) use_upsampled_ref;
-#endif
 
   if (cost_list &&
       cost_list[0] != INT_MAX && cost_list[1] != INT_MAX &&
@@ -524,16 +516,10 @@ int vp10_find_best_sub_pixel_tree_pruned_more(const MACROBLOCK *x,
                                              int *distortion,
                                              unsigned int *sse1,
                                              const uint8_t *second_pred,
-#if CONFIG_AFFINE_MOTION
                                              int w, int h,
                                              int use_upsampled_ref) {
-#else
-                                             int w, int h) {
-#endif
   SETUP_SUBPEL_SEARCH;
-#if CONFIG_AFFINE_MOTION
   (void) use_upsampled_ref;
-#endif
 
   besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp,
                                z, src_stride, y, y_stride, second_pred,
@@ -607,15 +593,9 @@ int vp10_find_best_sub_pixel_tree_pruned(const MACROBLOCK *x,
                                         int *distortion,
                                         unsigned int *sse1,
                                         const uint8_t *second_pred,
-#if CONFIG_AFFINE_MOTION
                                         int w, int h, int use_upsampled_ref) {
-#else
-                                        int w, int h) {
-#endif
   SETUP_SUBPEL_SEARCH;
-#if CONFIG_AFFINE_MOTION
   (void) use_upsampled_ref;
-#endif
 
   besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp,
                                z, src_stride, y, y_stride, second_pred,
@@ -705,9 +685,8 @@ static const MV search_step_table[12] = {
     {0, -1}, {0, 1}, {-1, 0}, {1, 0}
 };
 
-
-#if CONFIG_AFFINE_MOTION
 #if CONFIG_VP9_HIGHBITDEPTH
+// TODO(yunqing): Optimize the following 2 functions.
 static void highbd_comp_avg_upsampled_pred(uint16_t *comp_pred,
                                            const uint8_t *pred8,
                                            int width, int height,
@@ -798,7 +777,6 @@ static unsigned int upsampled_setup_center_error(
   besterr += mv_err_cost(bestmv, ref_mv, mvjcost, mvcost, error_per_bit);
   return besterr;
 }
-#endif
 
 int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
                                  MV *bestmv, const MV *ref_mv,
@@ -812,11 +790,7 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
                                  int *distortion,
                                  unsigned int *sse1,
                                  const uint8_t *second_pred,
-#if CONFIG_AFFINE_MOTION
                                  int w, int h, int use_upsampled_ref) {
-#else
-                                 int w, int h) {
-#endif
   const uint8_t *const z = x->plane[0].src.buf;
   const uint8_t *const src_address = z;
   const int src_stride = x->plane[0].src.stride;
@@ -852,7 +826,6 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
   bestmv->row *= 8;
   bestmv->col *= 8;
 
-#if CONFIG_AFFINE_MOTION
   // use_upsampled_ref can be 0 or 1
   if (use_upsampled_ref)
     besterr = upsampled_setup_center_error(xd, bestmv, ref_mv, error_per_bit,
@@ -860,7 +833,6 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
                                            second_pred, w, h, (offset << 3),
                                            mvjcost, mvcost, sse1, distortion);
   else
-#endif
     besterr = setup_center_error(xd, bestmv, ref_mv, error_per_bit, vfp,
                                  z, src_stride, y, y_stride, second_pred,
                                  w, h, offset, mvjcost, mvcost,
@@ -876,7 +848,6 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
       if (tc >= minc && tc <= maxc && tr >= minr && tr <= maxr) {
         MV this_mv = {tr, tc};
 
-#if CONFIG_AFFINE_MOTION
         if (use_upsampled_ref) {
           const uint8_t *const pre_address = y + tr * y_stride + tc;
 
@@ -884,7 +855,6 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
                                          pre_address, y_stride, second_pred,
                                          w, h, &sse);
         } else {
-#endif
           const uint8_t *const pre_address = y + (tr >> 3) * y_stride +
               (tc >> 3);
           if (second_pred == NULL)
@@ -893,9 +863,7 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
           else
             thismse = vfp->svaf(pre_address, y_stride, sp(tc), sp(tr),
                                 src_address, src_stride, &sse, second_pred);
-#if CONFIG_AFFINE_MOTION
         }
-#endif
 
         cost_array[idx] = thismse +
             mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit);
@@ -920,7 +888,6 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
     if (tc >= minc && tc <= maxc && tr >= minr && tr <= maxr) {
       MV this_mv = {tr, tc};
 
-#if CONFIG_AFFINE_MOTION
       if (use_upsampled_ref) {
         const uint8_t *const pre_address = y + tr * y_stride + tc;
 
@@ -928,7 +895,6 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
                                        pre_address, y_stride, second_pred,
                                        w, h, &sse);
       } else {
-#endif
         const uint8_t *const pre_address = y + (tr >> 3) * y_stride + (tc >> 3);
 
         if (second_pred == NULL)
@@ -937,9 +903,7 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
         else
           thismse = vfp->svaf(pre_address, y_stride, sp(tc), sp(tr),
                               src_address, src_stride, &sse, second_pred);
-#if CONFIG_AFFINE_MOTION
       }
-#endif
 
       cost_array[4] = thismse +
           mv_err_cost(&this_mv, ref_mv, mvjcost, mvcost, error_per_bit);
@@ -963,15 +927,11 @@ int vp10_find_best_sub_pixel_tree(const MACROBLOCK *x,
     }
 
     if (iters_per_step > 1 && best_idx != -1) {
-#if CONFIG_AFFINE_MOTION
       if (use_upsampled_ref) {
         SECOND_LEVEL_CHECKS_BEST(1);
       } else {
-#endif
         SECOND_LEVEL_CHECKS_BEST(0);
-#if CONFIG_AFFINE_MOTION
       }
-#endif
     }
 
     tr = br;

vp10/encoder/mcomp.h

@@ -116,11 +116,7 @@ typedef int (fractional_mv_step_fp) (
     int *mvjcost, int *mvcost[2],
     int *distortion, unsigned int *sse1,
     const uint8_t *second_pred,
-#if CONFIG_AFFINE_MOTION
     int w, int h, int use_upsampled_ref);
-#else
-    int w, int h);
-#endif
 
 extern fractional_mv_step_fp vp10_find_best_sub_pixel_tree;
 extern fractional_mv_step_fp vp10_find_best_sub_pixel_tree_pruned;

vp10/encoder/rdopt.h

@@ -103,13 +103,16 @@ void vp10_build_prediction_by_left_preds(VP10_COMP *cpi,
                                          int tmp_stride[MAX_MB_PLANE]);
 #endif  // CONFIG_OBMC
 
-#if CONFIG_AFFINE_MOTION
 static INLINE const YV12_BUFFER_CONFIG *get_upsampled_ref(VP10_COMP *cpi,
                                                           const int ref) {
   // Use up-sampled reference frames.
   int ref_idx = 0;
   if (ref == LAST_FRAME)
+#if CONFIG_EXT_REFS
+    ref_idx = cpi->lst_fb_idxes[ref - LAST_FRAME];
+#else
     ref_idx = cpi->lst_fb_idx;
+#endif
   else if (ref == GOLDEN_FRAME)
     ref_idx = cpi->gld_fb_idx;
   else if (ref == ALTREF_FRAME)
@@ -117,7 +120,6 @@ static INLINE const YV12_BUFFER_CONFIG *get_upsampled_ref(VP10_COMP *cpi,
 
   return &cpi->upsampled_ref_bufs[cpi->upsampled_ref_idx[ref_idx]].buf;
 }
-#endif
 
 #ifdef __cplusplus
 }  // extern "C"

vp10/encoder/speed_features.c

@@ -177,6 +177,7 @@ static void set_good_speed_feature(VP10_COMP *cpi, VP10_COMMON *cm,
     sf->comp_inter_joint_search_thresh = BLOCK_SIZES;
     sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX;
     sf->allow_partition_search_skip = 1;
+    sf->use_upsampled_references = 0;
 #if CONFIG_EXT_TX
     sf->tx_type_search = PRUNE_TWO;
 #endif
@@ -279,6 +280,7 @@ static void set_rt_speed_feature(VP10_COMP *cpi, SPEED_FEATURES *sf,
   sf->use_fast_coef_costing = 1;
   sf->allow_exhaustive_searches = 0;
   sf->exhaustive_searches_thresh = INT_MAX;
+  sf->use_upsampled_references = 0;
 
   // Use transform domain distortion computation
   // Note var-tx expt always uses pixel domain distortion.
@@ -495,6 +497,11 @@ void vp10_set_speed_features_framesize_independent(VP10_COMP *cpi) {
   sf->disable_filter_search_var_thresh = 0;
   sf->adaptive_interp_filter_search = 0;
   sf->allow_partition_search_skip = 0;
+#if CONFIG_EXT_REFS
+  sf->use_upsampled_references = 0;
+#else
+  sf->use_upsampled_references = 1;
+#endif
 
   for (i = 0; i < TX_SIZES; i++) {
     sf->intra_y_mode_mask[i] = INTRA_ALL;

vp10/encoder/speed_features.h

@@ -479,6 +479,9 @@ typedef struct SPEED_FEATURES {
   // Fast approximation of vp10_model_rd_from_var_lapndz
   int simple_model_rd_from_var;
 
+  // Do sub-pixel search in up-sampled reference frames
+  int use_upsampled_references;
+
   // Whether to compute distortion in the image domain (slower but
   // more accurate), or in the transform domain (faster but less acurate).
   int use_transform_domain_distortion;

vp10/encoder/temporal_filter.c

@@ -320,11 +320,7 @@ static int temporal_filter_find_matching_mb_c(VP10_COMP *cpi,
                                          0, mv_sf->subpel_iters_per_step,
                                          cond_cost_list(cpi, cost_list),
                                          NULL, NULL,
-#if CONFIG_AFFINE_MOTION
                                          &distortion, &sse, NULL, 0, 0, 0);
-#else
-                                         &distortion, &sse, NULL, 0, 0);
-#endif
 
   // Restore input state
   x->plane[0].src = src;

vpx_dsp/variance.c

@@ -267,7 +267,6 @@ void vpx_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred,
   }
 }
 
-#if CONFIG_AFFINE_MOTION
 // Get pred block from up-sampled reference.
 void vpx_upsampled_pred_c(uint8_t *comp_pred,
                           int width, int height,
@@ -300,7 +299,6 @@ void vpx_comp_avg_upsampled_pred_c(uint8_t *comp_pred, const uint8_t *pred,
       ref += stride;
     }
 }
-#endif
 
 #if CONFIG_VP9_HIGHBITDEPTH
 static void highbd_variance64(const uint8_t *a8, int  a_stride,

vpx_dsp/vpx_dsp_rtcd_defs.pl

@@ -1233,12 +1233,10 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
 #
 # ...
 #
-if (vpx_config("CONFIG_AFFINE_MOTION") eq "yes") {
-  add_proto qw/void vpx_upsampled_pred/, "uint8_t *comp_pred, int width, int height, const uint8_t *ref, int ref_stride";
-    specialize qw/vpx_upsampled_pred sse2/;
-  add_proto qw/void vpx_comp_avg_upsampled_pred/, "uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride";
-    specialize qw/vpx_comp_avg_upsampled_pred sse2/;
-}
+add_proto qw/void vpx_upsampled_pred/, "uint8_t *comp_pred, int width, int height, const uint8_t *ref, int ref_stride";
+specialize qw/vpx_upsampled_pred sse2/;
+add_proto qw/void vpx_comp_avg_upsampled_pred/, "uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride";
+specialize qw/vpx_comp_avg_upsampled_pred sse2/;
 
 #
 # ...

vpx_dsp/x86/variance_sse2.c

@@ -476,7 +476,6 @@ FNS(ssse3, ssse3);
 #undef FN
 #endif  // CONFIG_USE_X86INC
 
-#if CONFIG_AFFINE_MOTION
 void vpx_upsampled_pred_sse2(uint8_t *comp_pred,
                              int width, int height,
                              const uint8_t *ref,  int ref_stride) {
@@ -703,4 +702,3 @@ void vpx_comp_avg_upsampled_pred_sse2(uint8_t *comp_pred, const uint8_t *pred,
       }
     }
 }
-#endif