Simplify default temporal layers.

webrtc/modules/video_coding/codecs/vp8/default_temporal_layers.cc

Index: webrtc/modules/video_coding/codecs/vp8/default_temporal_layers.cc
diff --git a/webrtc/modules/video_coding/codecs/vp8/default_temporal_layers.cc b/webrtc/modules/video_coding/codecs/vp8/default_temporal_layers.cc
index 4f04743fc5d0359d62ff02117d9fe620e26fc309..c73a7b2ed04ec6fa6fe646db3ddf4d806c786584 100644
--- a/webrtc/modules/video_coding/codecs/vp8/default_temporal_layers.cc
+++ b/webrtc/modules/video_coding/codecs/vp8/default_temporal_layers.cc
@@ -26,26 +26,133 @@
 
 namespace webrtc {
 
+TemporalReferences::TemporalReferences(TemporalBufferUsage last,
+                                       TemporalBufferUsage golden,
+                                       TemporalBufferUsage arf)
+    : TemporalReferences(last, golden, arf, false, false) {}
+
+TemporalReferences::TemporalReferences(TemporalBufferUsage last,
+                                       TemporalBufferUsage golden,
+                                       TemporalBufferUsage arf,
+                                       int extra_flags)
+    : TemporalReferences(last,
+                         golden,
+                         arf,
+                         (extra_flags & kLayerSync) != 0,
+                         (extra_flags & kFreezeEntropy) != 0) {}
+
+TemporalReferences::TemporalReferences(TemporalBufferUsage last,
+                                       TemporalBufferUsage golden,
+                                       TemporalBufferUsage arf,
+                                       bool layer_sync,
+                                       bool freeze_entropy)
+    : reference_last((last & kReference) != 0),
+      update_last((last & kUpdate) != 0),
+      reference_golden((golden & kReference) != 0),
+      update_golden((golden & kUpdate) != 0),
+      reference_arf((arf & kReference) != 0),
+      update_arf((arf & kUpdate) != 0),
+      layer_sync(layer_sync),
+      freeze_entropy(freeze_entropy) {}
+
+namespace {

[brandtr, 2017/03/21 12:20:09]

nit: New line between l58 and l59. (Or remove space on l137.)

[pbos-webrtc, 2017/03/21 17:37:46]

Done.

+std::vector<unsigned int> GetTemporalIds(size_t num_layers) {
+  switch (num_layers) {
+    case 1:
+      return {0};
+    case 2:
+      return {0, 1};
+    case 3:
+      return {0, 2, 1, 2};
+    case 4:

[brandtr, 2017/03/21 12:20:09]

Might be good to add a diagram showing these IDs visually? Something like
https://cs.chromium.org/chromium/src/third_party/webrtc/modules/video_coding/....

Could be combined with the next comment.

[sprang_webrtc, 2017/03/21 13:50:38]

Do we ever use 4 temporal layers in practice?

[pbos-webrtc, 2017/03/21 17:37:46]

Done.

[pbos-webrtc, 2017/03/21 17:37:46]

No, but they are supported in RTP, hardware encoders can support them and if we
expose temporal layers to javascript we might want to support up to four layers.

+      return {0, 3, 2, 3, 1, 3, 2, 3};
+    default:
+      RTC_NOTREACHED();
+      break;
+  }
+  RTC_NOTREACHED();
+  return {0};
+}
+
+std::vector<TemporalReferences> GetTemporalPattern(
+    size_t num_layers) {
+  switch (num_layers) {
+    case 1:
+      return {TemporalReferences(kReferenceAndUpdate, kReference, kReference)};

[brandtr, 2017/03/21 12:20:09]

It would be great with some simple ASCII art here, showing the reference
dependency graphs for the different layer setups. Then it would also be fairly
easy to understand what reference buffers need to be updated for the different
layers. (The latter information is described by the now deleted comments here:
https://cs.chromium.org/chromium/src/third_party/webrtc/modules/video_coding/...)

[pbos-webrtc, 2017/03/21 17:37:46]

Since this contains all references drawing this is really hard. I put in some
descriptive comments instead, PTAL.

[brandtr, 2017/03/22 08:44:42]

Much easier to read through the lists now, thanks.

+    case 2:
+      return {TemporalReferences(kReferenceAndUpdate, kUpdate, kReference),
+              TemporalReferences(kReference, kUpdate, kReference, kLayerSync),
+              TemporalReferences(kReferenceAndUpdate, kNone, kReference),
+              TemporalReferences(kReference, kReferenceAndUpdate, kReference),
+              TemporalReferences(kReferenceAndUpdate, kNone, kReference),
+              TemporalReferences(kReference, kReferenceAndUpdate, kReference),
+              TemporalReferences(kReferenceAndUpdate, kNone, kReference),
+              TemporalReferences(kReference, kReference, kReference,
+                                 kFreezeEntropy)};
+    case 3:
+      return {TemporalReferences(kReferenceAndUpdate, kUpdate, kReference),
+              TemporalReferences(kReference, kNone, kReference,
+                                 kLayerSync | kFreezeEntropy),
+              TemporalReferences(kReference, kUpdate, kReference, kLayerSync),
+              TemporalReferences(kReference, kReference, kReference,
+                                 kFreezeEntropy),
+              TemporalReferences(kReferenceAndUpdate, kNone, kReference),
+              TemporalReferences(kReference, kReference, kReference,
+                                 kFreezeEntropy),
+              TemporalReferences(kReference, kReferenceAndUpdate, kReference),
+              TemporalReferences(kReference, kReference, kReference,
+                                 kFreezeEntropy)};
+    case 4:
+      return {TemporalReferences(kReferenceAndUpdate, kNone, kNone),
+              TemporalReferences(kReference, kReference, kReference,
+                                 kLayerSync | kFreezeEntropy),
+              TemporalReferences(kReference, kNone, kUpdate, kLayerSync),
+              TemporalReferences(kReference, kReference, kReference,
+                                 kLayerSync | kFreezeEntropy),
+              TemporalReferences(kReference, kUpdate, kNone, kLayerSync),
+              TemporalReferences(kReference, kReference, kReference,
+                                 kLayerSync | kFreezeEntropy),
+              TemporalReferences(kReference, kReference, kReferenceAndUpdate),
+              TemporalReferences(kReference, kReference, kReference,
+                                 kLayerSync | kFreezeEntropy),
+              TemporalReferences(kReferenceAndUpdate, kNone, kNone),
+              TemporalReferences(kReference, kReference, kReference,
+                                 kLayerSync | kFreezeEntropy),
+              TemporalReferences(kReference, kReference, kReferenceAndUpdate),
+              TemporalReferences(kReference, kReference, kReference,
+                                 kLayerSync | kFreezeEntropy),
+              TemporalReferences(kReference, kReferenceAndUpdate, kNone),
+              TemporalReferences(kReference, kReference, kReference,
+                                 kLayerSync | kFreezeEntropy),
+              TemporalReferences(kReference, kReference, kReferenceAndUpdate),
+              TemporalReferences(kReference, kReference, kReference,
+                                 kLayerSync | kFreezeEntropy)};
+    default:
+      RTC_NOTREACHED();
+      break;
+  }
+  RTC_NOTREACHED();
+  return {TemporalReferences(kNone, kNone, kNone)};
+}
+
+}  // namespace
+
 DefaultTemporalLayers::DefaultTemporalLayers(int number_of_temporal_layers,
                                              uint8_t initial_tl0_pic_idx)
-    : number_of_temporal_layers_(number_of_temporal_layers),
-      temporal_ids_length_(0),
-      temporal_pattern_length_(0),
+    : num_layers_(std::max(1, number_of_temporal_layers)),
+      temporal_ids_(GetTemporalIds(num_layers_)),
+      temporal_pattern_(GetTemporalPattern(num_layers_)),
       tl0_pic_idx_(initial_tl0_pic_idx),
       pattern_idx_(255),
       timestamp_(0),
       last_base_layer_sync_(false) {
   RTC_CHECK_GE(kMaxTemporalStreams, number_of_temporal_layers);
   RTC_CHECK_GE(number_of_temporal_layers, 0);
-  memset(temporal_ids_, 0, sizeof(temporal_ids_));
-  memset(temporal_pattern_, 0, sizeof(temporal_pattern_));
+  RTC_CHECK_LE(number_of_temporal_layers, 4);
 }
 
 int DefaultTemporalLayers::CurrentLayerId() const {
-  assert(temporal_ids_length_ > 0);
-  int index = pattern_idx_ % temporal_ids_length_;
-  assert(index >= 0);
-  return temporal_ids_[index];
+  return temporal_ids_[pattern_idx_ % temporal_ids_.size()];
 }
 
 std::vector<uint32_t> DefaultTemporalLayers::OnRatesUpdated(
@@ -53,17 +160,16 @@ std::vector<uint32_t> DefaultTemporalLayers::OnRatesUpdated(
     int max_bitrate_kbps,
     int framerate) {
   std::vector<uint32_t> bitrates;
-  const int num_layers = std::max(1, number_of_temporal_layers_);
-  for (int i = 0; i < num_layers; ++i) {
+  for (size_t i = 0; i < num_layers_; ++i) {
     float layer_bitrate =
-        bitrate_kbps * kVp8LayerRateAlloction[num_layers - 1][i];
+        bitrate_kbps * kVp8LayerRateAlloction[num_layers_ - 1][i];
     bitrates.push_back(static_cast<uint32_t>(layer_bitrate + 0.5));
   }
   new_bitrates_kbps_ = rtc::Optional<std::vector<uint32_t>>(bitrates);
 
   // Allocation table is of aggregates, transform to individual rates.
   uint32_t sum = 0;
-  for (int i = 0; i < num_layers; ++i) {
+  for (size_t i = 0; i < num_layers_; ++i) {
     uint32_t layer_bitrate = bitrates[i];
     RTC_DCHECK_LE(sum, bitrates[i]);
     bitrates[i] -= sum;
@@ -83,197 +189,44 @@ bool DefaultTemporalLayers::UpdateConfiguration(vpx_codec_enc_cfg_t* cfg) {
   if (!new_bitrates_kbps_)
     return false;
 
-  switch (number_of_temporal_layers_) {
-    case 0:
-      FALLTHROUGH();
-    case 1:
-      temporal_ids_length_ = 1;
-      temporal_ids_[0] = 0;
-      cfg->ts_number_layers = number_of_temporal_layers_;
-      cfg->ts_periodicity = temporal_ids_length_;
-      cfg->ts_target_bitrate[0] = (*new_bitrates_kbps_)[0];
-      cfg->ts_rate_decimator[0] = 1;
-      memcpy(cfg->ts_layer_id, temporal_ids_,
-             sizeof(unsigned int) * temporal_ids_length_);
-      temporal_pattern_length_ = 1;
-      temporal_pattern_[0] = kTemporalUpdateLastRefAll;
-      break;
-    case 2:
-      temporal_ids_length_ = 2;
-      temporal_ids_[0] = 0;
-      temporal_ids_[1] = 1;
-      cfg->ts_number_layers = number_of_temporal_layers_;
-      cfg->ts_periodicity = temporal_ids_length_;
-      // Split stream 60% 40%.
-      // Bitrate API for VP8 is the agregated bitrate for all lower layers.
-      cfg->ts_target_bitrate[0] = (*new_bitrates_kbps_)[0];
-      cfg->ts_target_bitrate[1] = (*new_bitrates_kbps_)[1];
-      cfg->ts_rate_decimator[0] = 2;
-      cfg->ts_rate_decimator[1] = 1;
-      memcpy(cfg->ts_layer_id, temporal_ids_,
-             sizeof(unsigned int) * temporal_ids_length_);
-      temporal_pattern_length_ = 8;
-      temporal_pattern_[0] = kTemporalUpdateLastAndGoldenRefAltRef;
-      temporal_pattern_[1] = kTemporalUpdateGoldenWithoutDependencyRefAltRef;
-      temporal_pattern_[2] = kTemporalUpdateLastRefAltRef;
-      temporal_pattern_[3] = kTemporalUpdateGoldenRefAltRef;
-      temporal_pattern_[4] = kTemporalUpdateLastRefAltRef;
-      temporal_pattern_[5] = kTemporalUpdateGoldenRefAltRef;
-      temporal_pattern_[6] = kTemporalUpdateLastRefAltRef;
-      temporal_pattern_[7] = kTemporalUpdateNone;
-      break;
-    case 3:
-      temporal_ids_length_ = 4;
-      temporal_ids_[0] = 0;
-      temporal_ids_[1] = 2;
-      temporal_ids_[2] = 1;
-      temporal_ids_[3] = 2;
-      cfg->ts_number_layers = number_of_temporal_layers_;
-      cfg->ts_periodicity = temporal_ids_length_;
-      // Split stream 40% 20% 40%.
-      // Bitrate API for VP8 is the agregated bitrate for all lower layers.
-      cfg->ts_target_bitrate[0] = (*new_bitrates_kbps_)[0];
-      cfg->ts_target_bitrate[1] = (*new_bitrates_kbps_)[1];
-      cfg->ts_target_bitrate[2] = (*new_bitrates_kbps_)[2];
-      cfg->ts_rate_decimator[0] = 4;
-      cfg->ts_rate_decimator[1] = 2;
-      cfg->ts_rate_decimator[2] = 1;
-      memcpy(cfg->ts_layer_id, temporal_ids_,
-             sizeof(unsigned int) * temporal_ids_length_);
-      temporal_pattern_length_ = 8;
-      temporal_pattern_[0] = kTemporalUpdateLastAndGoldenRefAltRef;
-      temporal_pattern_[1] = kTemporalUpdateNoneNoRefGoldenRefAltRef;
-      temporal_pattern_[2] = kTemporalUpdateGoldenWithoutDependencyRefAltRef;
-      temporal_pattern_[3] = kTemporalUpdateNone;
-      temporal_pattern_[4] = kTemporalUpdateLastRefAltRef;
-      temporal_pattern_[5] = kTemporalUpdateNone;
-      temporal_pattern_[6] = kTemporalUpdateGoldenRefAltRef;
-      temporal_pattern_[7] = kTemporalUpdateNone;
-      break;
-    case 4:
-      temporal_ids_length_ = 8;
-      temporal_ids_[0] = 0;
-      temporal_ids_[1] = 3;
-      temporal_ids_[2] = 2;
-      temporal_ids_[3] = 3;
-      temporal_ids_[4] = 1;
-      temporal_ids_[5] = 3;
-      temporal_ids_[6] = 2;
-      temporal_ids_[7] = 3;
-      // Split stream 25% 15% 20% 40%.
-      // Bitrate API for VP8 is the agregated bitrate for all lower layers.
-      cfg->ts_number_layers = 4;
-      cfg->ts_periodicity = temporal_ids_length_;
-      cfg->ts_target_bitrate[0] = (*new_bitrates_kbps_)[0];
-      cfg->ts_target_bitrate[1] = (*new_bitrates_kbps_)[1];
-      cfg->ts_target_bitrate[2] = (*new_bitrates_kbps_)[2];
-      cfg->ts_target_bitrate[3] = (*new_bitrates_kbps_)[3];
-      cfg->ts_rate_decimator[0] = 8;
-      cfg->ts_rate_decimator[1] = 4;
-      cfg->ts_rate_decimator[2] = 2;
-      cfg->ts_rate_decimator[3] = 1;
-      memcpy(cfg->ts_layer_id, temporal_ids_,
-             sizeof(unsigned int) * temporal_ids_length_);
-      temporal_pattern_length_ = 16;
-      temporal_pattern_[0] = kTemporalUpdateLast;
-      temporal_pattern_[1] = kTemporalUpdateNone;
-      temporal_pattern_[2] = kTemporalUpdateAltrefWithoutDependency;
-      temporal_pattern_[3] = kTemporalUpdateNone;
-      temporal_pattern_[4] = kTemporalUpdateGoldenWithoutDependency;
-      temporal_pattern_[5] = kTemporalUpdateNone;
-      temporal_pattern_[6] = kTemporalUpdateAltref;
-      temporal_pattern_[7] = kTemporalUpdateNone;
-      temporal_pattern_[8] = kTemporalUpdateLast;
-      temporal_pattern_[9] = kTemporalUpdateNone;
-      temporal_pattern_[10] = kTemporalUpdateAltref;
-      temporal_pattern_[11] = kTemporalUpdateNone;
-      temporal_pattern_[12] = kTemporalUpdateGolden;
-      temporal_pattern_[13] = kTemporalUpdateNone;
-      temporal_pattern_[14] = kTemporalUpdateAltref;
-      temporal_pattern_[15] = kTemporalUpdateNone;
-      break;
-    default:
-      RTC_NOTREACHED();
-      return false;
+  for (size_t i = 0; i < num_layers_; ++i) {
+    cfg->ts_target_bitrate[i] = (*new_bitrates_kbps_)[i];
+    // ..., 4, 2, 1
+    cfg->ts_rate_decimator[i] = 1 << (num_layers_ - i - 1);
   }
 
+  cfg->ts_number_layers = num_layers_;
+  cfg->ts_periodicity = temporal_ids_.size();
+  memcpy(cfg->ts_layer_id, &temporal_ids_[0],
+         sizeof(unsigned int) * temporal_ids_.size());
+
   new_bitrates_kbps_ = rtc::Optional<std::vector<uint32_t>>();
 
   return true;
 }
 
 int DefaultTemporalLayers::EncodeFlags(uint32_t timestamp) {
-  assert(number_of_temporal_layers_ > 0);
-  assert(kMaxTemporalPattern >= temporal_pattern_length_);
-  assert(0 < temporal_pattern_length_);
+  assert(num_layers_ > 0);
+  assert(0 < temporal_pattern_.size());
   int flags = 0;
-  int patternIdx = ++pattern_idx_ % temporal_pattern_length_;
-  assert(kMaxTemporalPattern >= patternIdx);
-  switch (temporal_pattern_[patternIdx]) {
-    case kTemporalUpdateLast:
-      flags |= VP8_EFLAG_NO_UPD_GF;
-      flags |= VP8_EFLAG_NO_UPD_ARF;
-      flags |= VP8_EFLAG_NO_REF_GF;
-      flags |= VP8_EFLAG_NO_REF_ARF;
-      break;
-    case kTemporalUpdateGoldenWithoutDependency:
-      flags |= VP8_EFLAG_NO_REF_GF;
-      // Deliberately no break here.
-      FALLTHROUGH();
-    case kTemporalUpdateGolden:
-      flags |= VP8_EFLAG_NO_REF_ARF;
-      flags |= VP8_EFLAG_NO_UPD_ARF;
-      flags |= VP8_EFLAG_NO_UPD_LAST;
-      break;
-    case kTemporalUpdateAltrefWithoutDependency:
-      flags |= VP8_EFLAG_NO_REF_ARF;
-      flags |= VP8_EFLAG_NO_REF_GF;
-      // Deliberately no break here.
-      FALLTHROUGH();
-    case kTemporalUpdateAltref:
-      flags |= VP8_EFLAG_NO_UPD_GF;
-      flags |= VP8_EFLAG_NO_UPD_LAST;
-      break;
-    case kTemporalUpdateNoneNoRefAltref:
-      flags |= VP8_EFLAG_NO_REF_ARF;
-      // Deliberately no break here.
-      FALLTHROUGH();
-    case kTemporalUpdateNone:
-      flags |= VP8_EFLAG_NO_UPD_GF;
-      flags |= VP8_EFLAG_NO_UPD_ARF;
-      flags |= VP8_EFLAG_NO_UPD_LAST;
-      flags |= VP8_EFLAG_NO_UPD_ENTROPY;
-      break;
-    case kTemporalUpdateNoneNoRefGoldenRefAltRef:
-      flags |= VP8_EFLAG_NO_REF_GF;
-      flags |= VP8_EFLAG_NO_UPD_GF;
-      flags |= VP8_EFLAG_NO_UPD_ARF;
-      flags |= VP8_EFLAG_NO_UPD_LAST;
-      flags |= VP8_EFLAG_NO_UPD_ENTROPY;
-      break;
-    case kTemporalUpdateGoldenWithoutDependencyRefAltRef:
-      flags |= VP8_EFLAG_NO_REF_GF;
-      flags |= VP8_EFLAG_NO_UPD_ARF;
-      flags |= VP8_EFLAG_NO_UPD_LAST;
-      break;
-    case kTemporalUpdateLastRefAltRef:
-      flags |= VP8_EFLAG_NO_UPD_GF;
-      flags |= VP8_EFLAG_NO_UPD_ARF;
-      flags |= VP8_EFLAG_NO_REF_GF;
-      break;
-    case kTemporalUpdateGoldenRefAltRef:
-      flags |= VP8_EFLAG_NO_UPD_ARF;
-      flags |= VP8_EFLAG_NO_UPD_LAST;
-      break;
-    case kTemporalUpdateLastAndGoldenRefAltRef:
-      flags |= VP8_EFLAG_NO_UPD_ARF;
-      flags |= VP8_EFLAG_NO_REF_GF;
-      break;
-    case kTemporalUpdateLastRefAll:
-      flags |= VP8_EFLAG_NO_UPD_ARF;
-      flags |= VP8_EFLAG_NO_UPD_GF;
-      break;
-  }
+  int patternIdx = ++pattern_idx_ % temporal_pattern_.size();

[sprang_webrtc, 2017/03/21 13:50:38]

Why not just update pattern_idx_ itself?

[pbos-webrtc, 2017/03/21 17:37:46]

pattern_idx_ is used for both temporal_pattern_.size() and temporal_ids_.size()
which are different lengths.

+  const TemporalReferences& references = temporal_pattern_[patternIdx];
+
+  if (!references.reference_last)
+    flags |= VP8_EFLAG_NO_REF_LAST;
+  if (!references.update_last)
+    flags |= VP8_EFLAG_NO_UPD_LAST;
+  if (!references.reference_golden)
+    flags |= VP8_EFLAG_NO_REF_GF;
+  if (!references.update_golden)
+    flags |= VP8_EFLAG_NO_UPD_GF;
+  if (!references.reference_arf)
+    flags |= VP8_EFLAG_NO_REF_ARF;
+  if (!references.update_arf)
+    flags |= VP8_EFLAG_NO_UPD_ARF;
+  if (references.freeze_entropy)
+    flags |= VP8_EFLAG_NO_UPD_ENTROPY;
+
   return flags;
 }
 
@@ -281,10 +234,9 @@ void DefaultTemporalLayers::PopulateCodecSpecific(
     bool base_layer_sync,
     CodecSpecificInfoVP8* vp8_info,
     uint32_t timestamp) {
-  assert(number_of_temporal_layers_ > 0);
-  assert(0 < temporal_ids_length_);
+  assert(num_layers_ > 0);

[sprang_webrtc, 2017/03/21 13:50:38]

nit: RTC_[D]CHECK_GT

[pbos-webrtc, 2017/03/21 17:37:46]

Done.

 
-  if (number_of_temporal_layers_ == 1) {
+  if (num_layers_ == 1) {
     vp8_info->temporalIdx = kNoTemporalIdx;
     vp8_info->layerSync = false;
     vp8_info->tl0PicIdx = kNoTl0PicIdx;
@@ -295,19 +247,9 @@ void DefaultTemporalLayers::PopulateCodecSpecific(
     } else {
       vp8_info->temporalIdx = CurrentLayerId();
       TemporalReferences temporal_reference =
-          temporal_pattern_[pattern_idx_ % temporal_pattern_length_];
-
-      if (temporal_reference == kTemporalUpdateAltrefWithoutDependency ||
-          temporal_reference == kTemporalUpdateGoldenWithoutDependency ||
-          temporal_reference ==
-              kTemporalUpdateGoldenWithoutDependencyRefAltRef ||
-          temporal_reference == kTemporalUpdateNoneNoRefGoldenRefAltRef ||
-          (temporal_reference == kTemporalUpdateNone &&
-           number_of_temporal_layers_ == 4)) {
-        vp8_info->layerSync = true;
-      } else {
-        vp8_info->layerSync = false;
-      }
+          temporal_pattern_[pattern_idx_ % temporal_pattern_.size()];
+
+      vp8_info->layerSync = temporal_reference.layer_sync;
     }
     if (last_base_layer_sync_ && vp8_info->temporalIdx != 0) {
       // Regardless of pattern the frame after a base layer sync will always