From e9c44a76a2e6d89abbeb455cc54aff4d919a5178 Mon Sep 17 00:00:00 2001
From: Jingning Han <jingning@google.com>
Date: Thu, 9 Jun 2016 20:44:00 -0700
Subject: [PATCH] Refactor trellis optimization process

This commit refactors the trellis coefficient optimization process.
It saves multiplications used to generate the final dequantized
coefficients. It also removes two memset operations on quantized
and dequantized coefficient sets.

The trellis coefficient optimization is on average running over
10% faster.

Change-Id: If3aa26d2a706c3012bf2b7ac059bf1825250e81f
---
 vp10/encoder/encodemb.c | 47 +++++++++++++++++++++++++++++------------
 1 file changed, 33 insertions(+), 14 deletions(-)

diff --git a/vp10/encoder/encodemb.c b/vp10/encoder/encodemb.c
index bd9dc559e..7c5f77e33 100644
--- a/vp10/encoder/encodemb.c
+++ b/vp10/encoder/encodemb.c
@@ -58,6 +58,7 @@ typedef struct vp10_token_state {
   int           next;
   int16_t       token;
   tran_low_t    qc;
+  tran_low_t    dqc;
 } vp10_token_state;
 
 // These numbers are empirically obtained.
@@ -99,17 +100,18 @@ int vp10_optimize_b(MACROBLOCK *mb, int plane, int block,
   const int eob = p->eobs[block];
   const PLANE_TYPE type = pd->plane_type;
   const int default_eob = 16 << (tx_size << 1);
-  const int16_t *dequant_ptr = pd->dequant;
-  const uint8_t *const band_translate = get_band_translate(tx_size);
+  const int16_t* const dequant_ptr = pd->dequant;
+  const uint8_t* const band_translate = get_band_translate(tx_size);
   TX_TYPE tx_type = get_tx_type(type, xd, block, tx_size);
-  const scan_order *const so =
+  const scan_order* const so =
       get_scan(tx_size, tx_type, is_inter_block(&xd->mi[0]->mbmi));
-  const int16_t *const scan = so->scan;
-  const int16_t *const nb = so->neighbors;
+  const int16_t* const scan = so->scan;
+  const int16_t* const nb = so->neighbors;
 #if CONFIG_NEW_QUANT
   const dequant_val_type_nuq *dequant_val = pd->dequant_val_nuq;
 #endif  // CONFIG_NEW_QUANT
-  int shift = get_tx_scale(xd, tx_type, tx_size);
+  const int shift = get_tx_scale(xd, tx_type, tx_size);
+  const int dq_step[2] = { dequant_ptr[0] >> shift, dequant_ptr[1] >> shift };
   int next = eob, sz = 0;
   const int64_t rdmult = (mb->rdmult * plane_rd_mult[ref][type]) >> 1;
   const int64_t rddiv = mb->rddiv;
@@ -179,6 +181,7 @@ int vp10_optimize_b(MACROBLOCK *mb, int plane, int block,
       tokens[i][0].next = next;
       tokens[i][0].token = t0;
       tokens[i][0].qc = x;
+      tokens[i][0].dqc = dqcoeff[rc];
       best_index[i][0] = best;
 
       /* Evaluate the second possibility for this state. */
@@ -271,6 +274,24 @@ int vp10_optimize_b(MACROBLOCK *mb, int plane, int block,
       tokens[i][1].next = next;
       tokens[i][1].token = best ? t1 : t0;
       tokens[i][1].qc = x;
+
+      if (x) {
+        tran_low_t offset = dq_step[rc != 0];
+        // The 32x32 transform coefficient uses half quantization step size.
+        // Account for the rounding difference in the dequantized coefficeint
+        // value when the quantization index is dropped from an even number
+        // to an odd number.
+        if (shift & x)
+          offset += (dequant_ptr[rc != 0] & 0x01);
+
+        if (sz == 0)
+          tokens[i][1].dqc = dqcoeff[rc] - offset;
+        else
+          tokens[i][1].dqc = dqcoeff[rc] + offset;
+      } else {
+        tokens[i][1].dqc = 0;
+      }
+
       best_index[i][1] = best;
       /* Finally, make this the new head of the trellis. */
       next = i;
@@ -309,24 +330,22 @@ int vp10_optimize_b(MACROBLOCK *mb, int plane, int block,
   rate1 += mb->token_costs[tx_size][type][ref][band][0][ctx][t1];
   UPDATE_RD_COST();
   best = rd_cost1 < rd_cost0;
+
   final_eob = -1;
-  memset(qcoeff, 0, sizeof(*qcoeff) * (16 << (tx_size * 2)));
-  memset(dqcoeff, 0, sizeof(*dqcoeff) * (16 << (tx_size * 2)));
+
   for (i = next; i < eob; i = next) {
     const int x = tokens[i][best].qc;
     const int rc = scan[i];
-    if (x) {
-      final_eob = i;
-    }
+
+    if (x) final_eob = i;
     qcoeff[rc] = x;
+    dqcoeff[rc] = tokens[i][best].dqc;
+
 #if CONFIG_NEW_QUANT
     dqcoeff[rc] = dequant_abscoeff_nuq(abs(x), dequant_ptr[rc != 0],
                                        dequant_val[band_translate[i]]);
     if (shift) dqcoeff[rc] = ROUND_POWER_OF_TWO(dqcoeff[rc], shift);
     if (x < 0) dqcoeff[rc] = -dqcoeff[rc];
-#else
-    dqcoeff[rc] = (abs(x * dequant_ptr[rc != 0]) >> shift);
-    if (x < 0) dqcoeff[rc] = -dqcoeff[rc];
 #endif  // CONFIG_NEW_QUANT
 
     next = tokens[i][best].next;