shithub: libvpx

Info  •  Files  •  Log

Avoid division by zero for rate q_step model

Change-Id: Ic5709b79131a3969fcb2a0eb3f53994f788b5cc9

--- a/vp9/encoder/vp9_encoder.c

+++ b/vp9/encoder/vp9_encoder.c

@@ -4228,8 +4228,9 @@

   // The model predicts bits as follows.

   // target_bits = bias - ratio * log2(q_step)

   // Given the target_bits, we compute the q_step as follows.

-  const double q_step =

-      pow(2.0, (rq_model->bias - target_bits) / rq_model->ratio);

+  double q_step;

+  assert(rq_model->ratio > 0);

+  q_step = pow(2.0, (rq_model->bias - target_bits) / rq_model->ratio);

   // TODO(angiebird): Make this function support highbitdepth.

   return vp9_convert_q_to_qindex(q_step, VPX_BITS_8);

 }

@@ -4248,7 +4249,9 @@

 }

 static double get_bits_percent_diff(int target_bits, int actual_bits) {

-  double diff = abs(target_bits - actual_bits) * 1. / target_bits;

+  double diff;

+  target_bits = VPXMAX(target_bits, 1);

+  diff = abs(target_bits - actual_bits) * 1. / target_bits;

   return diff * 100;

 }

@@ -4295,17 +4298,32 @@

 static void rq_model_update(const RATE_QINDEX_HISTORY *rq_history,

                             int target_bits, RATE_QSTEP_MODEL *rq_model) {

   const int recode_count = rq_history->recode_count;

+  const double delta = 0.00001;

   if (recode_count >= 2) {

-    // Fit the ratio and bias of rq_model based on last two recode histories.

-    const double s1 = vp9_convert_qindex_to_q(

-        rq_history->q_index_history[recode_count - 2], VPX_BITS_8);

-    const double s2 = vp9_convert_qindex_to_q(

-        rq_history->q_index_history[recode_count - 1], VPX_BITS_8);

-    const double r1 = rq_history->rate_history[recode_count - 2];

-    const double r2 = rq_history->rate_history[recode_count - 1];

-    rq_model->ratio = (r2 - r1) / (log2(s1) - log2(s2));

-    rq_model->bias = r1 + (rq_model->ratio) * log2(s1);

-    rq_model->ready = 1;

+    const int q_index1 = rq_history->q_index_history[recode_count - 2];

+    const int q_index2 = rq_history->q_index_history[recode_count - 1];

+    const int r1 = rq_history->rate_history[recode_count - 2];

+    const int r2 = rq_history->rate_history[recode_count - 1];

+    int valid = 0;

+    // lower q_index should yield higher bit rate

+    if (q_index1 < q_index2) {

+      valid = r1 > r2;

+    } else if (q_index1 > q_index2) {

+      valid = r1 < r2;

+    }

+    // Only update the model when the q_index and rate behave normally.

+    if (valid) {

+      // Fit the ratio and bias of rq_model based on last two recode histories.

+      const double s1 = vp9_convert_qindex_to_q(q_index1, VPX_BITS_8);

+      const double s2 = vp9_convert_qindex_to_q(q_index2, VPX_BITS_8);

+      if (fabs(log2(s1) - log2(s2)) > delta) {

+        rq_model->ratio = (r2 - r1) / (log2(s1) - log2(s2));

+        rq_model->bias = r1 + (rq_model->ratio) * log2(s1);

+        if (rq_model->ratio > delta && rq_model->bias > delta) {

+          rq_model->ready = 1;

+        }

+      }

+    }

   } else if (recode_count == 1) {

     if (rq_model->ready) {

       // Update the ratio only when the initial model exists and we only have

@@ -4312,9 +4330,11 @@

       // one recode history.

       const int prev_q = rq_history->q_index_history[recode_count - 1];

       const double prev_q_step = vp9_convert_qindex_to_q(prev_q, VPX_BITS_8);

-      const int actual_bits = rq_history->rate_history[recode_count - 1];

-      rq_model->ratio =

-          rq_model->ratio + (target_bits - actual_bits) / log2(prev_q_step);

+      if (fabs(log2(prev_q_step)) > delta) {

+        const int actual_bits = rq_history->rate_history[recode_count - 1];

+        rq_model->ratio =

+            rq_model->ratio + (target_bits - actual_bits) / log2(prev_q_step);

+      }

     }

   }

 }