Merge "fast auto-determined filtering strength"

src/enc/config.c

@@ -33,7 +33,7 @@ int WebPConfigInitInternal(WebPConfig* config,
   config->target_PSNR = 0.;
   config->method = 4;
   config->sns_strength = 50;
-  config->filter_strength = 60;   // rather high filtering, helps w/ gradients.
+  config->filter_strength = 60;   // mid-filtering
   config->filter_sharpness = 0;
   config->filter_type = 1;        // default: strong (so U/V is filtered too)
   config->partitions = 0;

src/enc/filter.c

@@ -11,12 +11,58 @@
 //
 // Author: somnath@google.com (Somnath Banerjee)
 
+#include <assert.h>
 #include "./vp8enci.h"
 
 #if defined(__cplusplus) || defined(c_plusplus)
 extern "C" {
 #endif
 
+// this table gives, for a given sharpness, the filtering strength to be
+// used (at least) in order to filter a given edge step delta.
+#define MAX_DELTA_SIZE 64
+static const uint8_t kLevelsFromDelta[8][MAX_DELTA_SIZE] = {
+  { 0,   1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15,
+    16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
+    32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47,
+    48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63 },
+  { 0,  1,  2,  3,  5,  6,  7,  8,  9, 11, 12, 13, 14, 15, 17, 18,
+    20, 21, 23, 24, 26, 27, 29, 30, 32, 33, 35, 36, 38, 39, 41, 42,
+    44, 45, 47, 48, 50, 51, 53, 54, 56, 57, 59, 60, 62, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  3,  5,  6,  7,  8,  9, 11, 12, 13, 14, 16, 17, 19,
+    20, 22, 23, 25, 26, 28, 29, 31, 32, 34, 35, 37, 38, 40, 41, 43,
+    44, 46, 47, 49, 50, 52, 53, 55, 56, 58, 59, 61, 62, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  3,  5,  6,  7,  8,  9, 11, 12, 13, 15, 16, 18, 19,
+    21, 22, 24, 25, 27, 28, 30, 31, 33, 34, 36, 37, 39, 40, 42, 43,
+    45, 46, 48, 49, 51, 52, 54, 55, 57, 58, 60, 61, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  3,  5,  6,  7,  8,  9, 11, 12, 14, 15, 17, 18, 20,
+    21, 23, 24, 26, 27, 29, 30, 32, 33, 35, 36, 38, 39, 41, 42, 44,
+    45, 47, 48, 50, 51, 53, 54, 56, 57, 59, 60, 62, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  4,  5,  7,  8,  9, 11, 12, 13, 15, 16, 17, 19, 20,
+    22, 23, 25, 26, 28, 29, 31, 32, 34, 35, 37, 38, 40, 41, 43, 44,
+    46, 47, 49, 50, 52, 53, 55, 56, 58, 59, 61, 62, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  4,  5,  7,  8,  9, 11, 12, 13, 15, 16, 18, 19, 21,
+    22, 24, 25, 27, 28, 30, 31, 33, 34, 36, 37, 39, 40, 42, 43, 45,
+    46, 48, 49, 51, 52, 54, 55, 57, 58, 60, 61, 63, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 },
+  {  0,  1,  2,  4,  5,  7,  8,  9, 11, 12, 14, 15, 17, 18, 20, 21,
+    23, 24, 26, 27, 29, 30, 32, 33, 35, 36, 38, 39, 41, 42, 44, 45,
+    47, 48, 50, 51, 53, 54, 56, 57, 59, 60, 62, 63, 63, 63, 63, 63,
+    63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63, 63 }
+};
+
+int VP8FilterStrengthFromDelta(int sharpness, int delta) {
+  const int pos = (delta < MAX_DELTA_SIZE) ? delta : MAX_DELTA_SIZE - 1;
+  assert(sharpness >= 0 && sharpness <= 7);
+  return kLevelsFromDelta[sharpness][pos];
+}
+
+// -----------------------------------------------------------------------------
 // NOTE: clip1, tables and InitTables are repeated entries of dsp.c
 static uint8_t abs0[255 + 255 + 1];     // abs(i)
 static uint8_t abs1[255 + 255 + 1];     // abs(i)>>1
@@ -353,12 +399,13 @@ void VP8InitFilter(VP8EncIterator* const it) {
 
 void VP8StoreFilterStats(VP8EncIterator* const it) {
   int d;
+  VP8Encoder* const enc = it->enc_;
   const int s = it->mb_->segment_;
-  const int level0 = it->enc_->dqm_[s].fstrength_;  // TODO: ref_lf_delta[]
+  const int level0 = enc->dqm_[s].fstrength_;  // TODO: ref_lf_delta[]
 
   // explore +/-quant range of values around level0
-  const int delta_min = -it->enc_->dqm_[s].quant_;
+  const int delta_min = -enc->dqm_[s].quant_;
-  const int delta_max = it->enc_->dqm_[s].quant_;
+  const int delta_max = enc->dqm_[s].quant_;
   const int step_size = (delta_max - delta_min >= 4) ? 4 : 1;
 
   if (it->lf_stats_ == NULL) return;
@@ -385,9 +432,9 @@ void VP8StoreFilterStats(VP8EncIterator* const it) {
 }
 
 void VP8AdjustFilterStrength(VP8EncIterator* const it) {
+  VP8Encoder* const enc = it->enc_;
   if (it->lf_stats_ != NULL) {
     int s;
-    VP8Encoder* const enc = it->enc_;
     for (s = 0; s < NUM_MB_SEGMENTS; s++) {
       int i, best_level = 0;
       // Improvement over filter level 0 should be at least 1e-5 (relatively)
@@ -401,9 +448,28 @@ void VP8AdjustFilterStrength(VP8EncIterator* const it) {
       }
       enc->dqm_[s].fstrength_ = best_level;
     }
+  } else if (enc->config_->filter_strength > 0) {
+    int max_level = 0;
+    int s;
+    for (s = 0; s < NUM_MB_SEGMENTS; s++) {
+      VP8SegmentInfo* const dqm = &enc->dqm_[s];
+      // this '>> 3' accounts for some inverse WHT scaling
+      const int delta = (dqm->max_edge_ * dqm->y2_.q_[1]) >> 3;
+      const int level =
+          VP8FilterStrengthFromDelta(enc->filter_hdr_.sharpness_, delta);
+      if (level > dqm->fstrength_) {
+        dqm->fstrength_ = level;
+      }
+      if (max_level < dqm->fstrength_) {
+        max_level = dqm->fstrength_;
+      }
+    }
+    enc->filter_hdr_.level_ = max_level;
   }
 }
 
+// -----------------------------------------------------------------------------
+
 #if defined(__cplusplus) || defined(c_plusplus)
 }    // extern "C"
 #endif

src/enc/quant.c

@@ -13,6 +13,7 @@
 
 #include <assert.h>
 #include <math.h>
+#include <stdlib.h>  // for abs()
 
 #include "./vp8enci.h"
 #include "./cost.h"
@@ -181,17 +182,17 @@ static void SetupMatrices(VP8Encoder* enc) {
     q16 = ExpandMatrix(&m->y2_, 1);
     quv = ExpandMatrix(&m->uv_, 2);
 
-    // TODO: Switch to kLambda*[] tables?
+    m->lambda_i4_          = (3 * q4 * q4) >> 7;
-    {
+    m->lambda_i16_         = (3 * q16 * q16);
-      m->lambda_i4_  = (3 * q4 * q4) >> 7;
+    m->lambda_uv_          = (3 * quv * quv) >> 6;
-      m->lambda_i16_ = (3 * q16 * q16);
+    m->lambda_mode_        = (1 * q4 * q4) >> 7;
-      m->lambda_uv_  = (3 * quv * quv) >> 6;
+    m->lambda_trellis_i4_  = (7 * q4 * q4) >> 3;
-      m->lambda_mode_    = (1 * q4 * q4) >> 7;
+    m->lambda_trellis_i16_ = (q16 * q16) >> 2;
-      m->lambda_trellis_i4_  = (7 * q4 * q4) >> 3;
+    m->lambda_trellis_uv_  = (quv *quv) << 1;
-      m->lambda_trellis_i16_ = (q16 * q16) >> 2;
+    m->tlambda_            = (tlambda_scale * q4) >> 5;
-      m->lambda_trellis_uv_  = (quv *quv) << 1;
+
-      m->tlambda_            = (tlambda_scale * q4) >> 5;
+    m->min_disto_ = 10 * m->y1_.q_[0];   // quantization-aware min disto
-    }
+    m->max_edge_  = 0;
   }
 }
 
@@ -200,16 +201,21 @@ static void SetupMatrices(VP8Encoder* enc) {
 
 // Very small filter-strength values have close to no visual effect. So we can
 // save a little decoding-CPU by turning filtering off for these.
-#define FSTRENGTH_CUTOFF 3
+#define FSTRENGTH_CUTOFF 2
 
 static void SetupFilterStrength(VP8Encoder* const enc) {
   int i;
-  const int level0 = enc->config_->filter_strength;
+  // level0 is in [0..500]. Using '-f 50' as filter_strength is mid-filtering.
+  const int level0 = 5 * enc->config_->filter_strength;
   for (i = 0; i < NUM_MB_SEGMENTS; ++i) {
-    // Segments with lower quantizer will be less filtered. TODO: tune (wrt SNS)
+    VP8SegmentInfo* const m = &enc->dqm_[i];
-    const int level = level0 * 256 * enc->dqm_[i].quant_ / 128;
+    // We focus on the quantization of AC coeffs.
-    const int f = level / (256 + enc->dqm_[i].beta_);
+    const int qstep = kAcTable[clip(m->quant_, 0, 127)] >> 2;
-    enc->dqm_[i].fstrength_ = (f < FSTRENGTH_CUTOFF) ? 0 : (f > 63) ? 63 : f;
+    const int base_strength =
+        VP8FilterStrengthFromDelta(enc->filter_hdr_.sharpness_, qstep);
+    // Segments with lower complexity ('beta') will be less filtered.
+    const int f = base_strength * level0 / (256 + m->beta_);
+    m->fstrength_ = (f < FSTRENGTH_CUTOFF) ? 0 : (f > 63) ? 63 : f;
   }
   // We record the initial strength (mainly for the case of 1-segment only).
   enc->filter_hdr_.level_ = enc->dqm_[0].fstrength_;
@@ -648,10 +654,10 @@ static int ReconstructIntra16(VP8EncIterator* const it,
                               VP8ModeScore* const rd,
                               uint8_t* const yuv_out,
                               int mode) {
-  const VP8Encoder* const enc = it->enc_;
+  VP8Encoder* const enc = it->enc_;
   const uint8_t* const ref = it->yuv_p_ + VP8I16ModeOffsets[mode];
   const uint8_t* const src = it->yuv_in_ + Y_OFF;
-  const VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
+  VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
   int nz = 0;
   int n;
   int16_t tmp[16][16], dc_tmp[16];
@@ -757,6 +763,17 @@ static int ReconstructUV(VP8EncIterator* const it, VP8ModeScore* const rd,
 // RD-opt decision. Reconstruct each modes, evalue distortion and bit-cost.
 // Pick the mode is lower RD-cost = Rate + lamba * Distortion.
 
+static void StoreMaxDelta(VP8SegmentInfo* const dqm, const int16_t DCs[16]) {
+  // We look at the first three AC coefficients to determine what is the average
+  // delta between each sub-4x4 block.
+  const int v0 = abs(DCs[1]);
+  const int v1 = abs(DCs[4]);
+  const int v2 = abs(DCs[5]);
+  int max_v = (v0 > v1) ? v1 : v0;
+  max_v = (v2 > max_v) ? v2 : max_v;
+  if (max_v > dqm->max_edge_) dqm->max_edge_ = max_v;
+}
+
 static void SwapPtr(uint8_t** a, uint8_t** b) {
   uint8_t* const tmp = *a;
   *a = *b;
@@ -768,8 +785,8 @@ static void SwapOut(VP8EncIterator* const it) {
 }
 
 static void PickBestIntra16(VP8EncIterator* const it, VP8ModeScore* const rd) {
-  const VP8Encoder* const enc = it->enc_;
+  VP8Encoder* const enc = it->enc_;
-  const VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
+  VP8SegmentInfo* const dqm = &enc->dqm_[it->mb_->segment_];
   const int lambda = dqm->lambda_i16_;
   const int tlambda = dqm->tlambda_;
   const uint8_t* const src = it->yuv_in_ + Y_OFF;
@@ -804,6 +821,13 @@ static void PickBestIntra16(VP8EncIterator* const it, VP8ModeScore* const rd) {
   }
   SetRDScore(dqm->lambda_mode_, rd);   // finalize score for mode decision.
   VP8SetIntra16Mode(it, rd->mode_i16);
+
+  // we have a blocky macroblock (only DCs are non-zero) with fairly high
+  // distortion, record max delta so we can later adjust the minimal filtering
+  // strength needed to smooth these blocks out.
+  if ((rd->nz & 0xffff) == 0 && rd->D > dqm->min_disto_) {
+    StoreMaxDelta(dqm, rd->y_dc_levels);
+  }
 }
 
 //------------------------------------------------------------------------------

src/enc/vp8enci.h

@@ -248,6 +248,8 @@ typedef struct {
   int beta_;       // filter-susceptibility, range [0,255].
   int quant_;      // final segment quantizer.
   int fstrength_;  // final in-loop filtering strength
+  int max_edge_;   // max edge delta (for filtering strength)
+  int min_disto_;  // minimum distortion required to trigger filtering record
   // reactivities
   int lambda_i16_, lambda_i4_, lambda_uv_;
   int lambda_mode_, lambda_trellis_, tlambda_;
@@ -555,6 +557,10 @@ void VP8InitFilter(VP8EncIterator* const it);
 void VP8StoreFilterStats(VP8EncIterator* const it);
 void VP8AdjustFilterStrength(VP8EncIterator* const it);
 
+// returns the approximate filtering strength needed to smooth a edge
+// step of 'delta', given a sharpness parameter 'sharpness'.
+int VP8FilterStrengthFromDelta(int sharpness, int delta);
+
 //------------------------------------------------------------------------------
 
 #if defined(__cplusplus) || defined(c_plusplus)