Restore previous motion search bit-error scale.

The bit to error transformation got doubled as a result of going from
8-bit to 9-bit costs (change d13385c).

Use defines to derive the scale numbers and comment some of the fields.

derf: -0.023 BDRATE
hevcmr: +0.067 BDRATE
stdhd: +0.098 BDRATE
(These are substantially smaller than than the original gains from 8 to
9 bit costing.)

Change-Id: I6a2b3b029b2f1415e4f90a05709b2333ec0eea9b

vp9/encoder/vp9_block.h

@@ -65,8 +65,14 @@ struct macroblock {
   int skip_optimize;
   int q_index;
 
+  // The equivalent error at the current rdmult of one whole bit (not one
+  // bitcost unit).
   int errorperbit;
+  // The equivalend SAD error of one (whole) bit at the current quantizer
+  // for large blocks.
   int sadperbit16;
+  // The equivalend SAD error of one (whole) bit at the current quantizer
+  // for sub-8x8 blocks.
   int sadperbit4;
   int rddiv;
   int rdmult;

vp9/encoder/vp9_mcomp.c

@@ -80,27 +80,29 @@ int vp9_mv_bit_cost(const MV *mv, const MV *ref,
   return ROUND_POWER_OF_TWO(mv_cost(&diff, mvjcost, mvcost) * weight, 7);
 }
 
-static int mv_err_cost(const MV *mv, const MV *ref,
-                       const int *mvjcost, int *mvcost[2],
-                       int error_per_bit) {
+#define PIXEL_TRANSFORM_ERROR_SCALE 4
+static int mv_err_cost(const MV *mv, const MV *ref, const int *mvjcost,
+                       int *mvcost[2], int error_per_bit) {
   if (mvcost) {
-    const MV diff = { mv->row - ref->row,
-                      mv->col - ref->col };
-    // TODO(aconverse): See if this shift needs to be tied to
-    // VP9_PROB_COST_SHIFT.
-    return ROUND_POWER_OF_TWO((unsigned)mv_cost(&diff, mvjcost, mvcost) *
-                                  error_per_bit, 13);
+    const MV diff = {mv->row - ref->row, mv->col - ref->col};
+    // This product sits at a 32-bit ceiling right now and any additional
+    // accuracy in either bit cost or error cost will cause it to overflow.
+    return ROUND_POWER_OF_TWO(
+        (unsigned)mv_cost(&diff, mvjcost, mvcost) * error_per_bit,
+        RDDIV_BITS + VP9_PROB_COST_SHIFT - RD_EPB_SHIFT +
+            PIXEL_TRANSFORM_ERROR_SCALE);
   }
   return 0;
 }
 
 static int mvsad_err_cost(const MACROBLOCK *x, const MV *mv, const MV *ref,
-                          int error_per_bit) {
+                          int sad_per_bit) {
   const MV diff = { mv->row - ref->row,
                     mv->col - ref->col };
-  // TODO(aconverse): See if this shift needs to be tied to VP9_PROB_COST_SHIFT.
-  return ROUND_POWER_OF_TWO((unsigned)mv_cost(&diff, x->nmvjointsadcost,
-                                      x->nmvsadcost) * error_per_bit, 8);
+  return ROUND_POWER_OF_TWO(
+      (unsigned)mv_cost(&diff, x->nmvjointsadcost, x->nmvsadcost) *
+          sad_per_bit,
+      VP9_PROB_COST_SHIFT);
 }
 
 void vp9_init_dsmotion_compensation(search_site_config *cfg, int stride) {
@@ -152,12 +154,13 @@ void vp9_init3smotion_compensation(search_site_config *cfg, int stride) {
  * could reduce the area.
  */
 
-/* estimated cost of a motion vector (r,c) */
+/* Estimated (square) error cost of a motion vector (r,c). The 14 scale comes
+ * from the same math as in mv_err_cost(). */
 #define MVC(r, c)                                              \
     (mvcost ?                                                  \
      ((unsigned)(mvjcost[((r) != rr) * 2 + ((c) != rc)] +      \
        mvcost[0][((r) - rr)] + mvcost[1][((c) - rc)]) *        \
-      error_per_bit + 4096) >> 13 : 0)
+      error_per_bit + 8192) >> 14 : 0)
 
 
 // convert motion vector component to offset for sv[a]f calc

vp9/encoder/vp9_quantize.c

@@ -342,8 +342,7 @@ void vp9_init_plane_quantizers(VP9_COMP *cpi, MACROBLOCK *x) {
   x->skip_block = segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP);
   x->q_index = qindex;
 
-  x->errorperbit = rdmult >> 6;
-  x->errorperbit += (x->errorperbit == 0);
+  set_error_per_bit(x, rdmult);
 
   vp9_initialize_me_consts(cpi, x, x->q_index);
 }

vp9/encoder/vp9_rd.c

@@ -41,7 +41,6 @@
 #include "vp9/encoder/vp9_tokenize.h"
 
 #define RD_THRESH_POW      1.25
-#define RD_MULT_EPB_RATIO  64
 
 // Factor to weigh the rate for switchable interp filters.
 #define SWITCHABLE_INTERP_RATE_FACTOR 1
@@ -279,8 +278,7 @@ void vp9_initialize_rd_consts(VP9_COMP *cpi) {
   rd->RDDIV = RDDIV_BITS;  // In bits (to multiply D by 128).
   rd->RDMULT = vp9_compute_rd_mult(cpi, cm->base_qindex + cm->y_dc_delta_q);
 
-  x->errorperbit = rd->RDMULT / RD_MULT_EPB_RATIO;
-  x->errorperbit += (x->errorperbit == 0);
+  set_error_per_bit(x, rd->RDMULT);
 
   x->select_tx_size = (cpi->sf.tx_size_search_method == USE_LARGESTALL &&
                        cm->frame_type != KEY_FRAME) ? 0 : 1;

vp9/encoder/vp9_rd.h

@@ -24,6 +24,7 @@ extern "C" {
 #endif
 
 #define RDDIV_BITS          7
+#define RD_EPB_SHIFT        6
 
 #define RDCOST(RM, DM, R, D) \
   (ROUND_POWER_OF_TWO(((int64_t)R) * (RM), VP9_PROB_COST_SHIFT) + (D << DM))
@@ -168,6 +169,11 @@ static INLINE int rd_less_than_thresh(int64_t best_rd, int thresh,
     return best_rd < ((int64_t)thresh * thresh_fact >> 5) || thresh == INT_MAX;
 }
 
+static INLINE void set_error_per_bit(MACROBLOCK *x, int rdmult) {
+  x->errorperbit = rdmult >> RD_EPB_SHIFT;
+  x->errorperbit += (x->errorperbit == 0);
+}
+
 void vp9_mv_pred(struct VP9_COMP *cpi, MACROBLOCK *x,
                  uint8_t *ref_y_buffer, int ref_y_stride,
                  int ref_frame, BLOCK_SIZE block_size);

vp9/encoder/x86/vp9_diamond_search_sad_avx.c

@@ -47,12 +47,12 @@ static INLINE int mv_cost(const int_mv mv,
 }
 
 static int mvsad_err_cost(const MACROBLOCK *x, const int_mv mv, const MV *ref,
-                          int error_per_bit) {
+                          int sad_per_bit) {
   const int_mv diff = pack_int_mv(mv.as_mv.row - ref->row,
                                   mv.as_mv.col - ref->col);
   return ROUND_POWER_OF_TWO((unsigned)mv_cost(diff, x->nmvjointsadcost,
                                               x->nmvsadcost) *
-                                              error_per_bit, 8);
+                                              sad_per_bit, VP9_PROB_COST_SHIFT);
 }
 
 /*****************************************************************************