test/fuzzers/vp9_encoder_references_fuzzer.cc

/*
 *  Copyright (c) 2021 The WebRTC project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include <stdint.h>

#include "absl/algorithm/container.h"
#include "absl/base/macros.h"
#include "absl/container/inlined_vector.h"
#include "api/array_view.h"
#include "api/transport/webrtc_key_value_config.h"
#include "api/video/video_frame.h"
#include "api/video_codecs/video_codec.h"
#include "api/video_codecs/video_encoder.h"
#include "modules/video_coding/codecs/interface/mock_libvpx_interface.h"
#include "modules/video_coding/codecs/vp9/libvpx_vp9_encoder.h"
#include "modules/video_coding/frame_dependencies_calculator.h"
#include "rtc_base/numerics/safe_compare.h"
#include "test/fuzzers/fuzz_data_helper.h"
#include "test/gmock.h"

// Fuzzer simulates various svc configurations and libvpx encoder dropping
// layer frames.
// Validates vp9 encoder wrapper produces consistent frame references.
namespace webrtc {
namespace {

using test::FuzzDataHelper;
using ::testing::NiceMock;

constexpr int kBitrateEnabledBps = 100'000;

class FrameValidator : public EncodedImageCallback {
 public:
  ~FrameValidator() override = default;

  Result OnEncodedImage(const EncodedImage& encoded_image,
                        const CodecSpecificInfo* codec_specific_info) override {
    RTC_CHECK(codec_specific_info);
    RTC_CHECK_EQ(codec_specific_info->codecType, kVideoCodecVP9);
    if (codec_specific_info->codecSpecific.VP9.first_frame_in_picture) {
      ++picture_id_;
    }
    int64_t frame_id = frame_id_++;
    LayerFrame& layer_frame = frames_[frame_id % kMaxFrameHistorySize];
    layer_frame.picture_id = picture_id_;
    layer_frame.spatial_id = encoded_image.SpatialIndex().value_or(0);
    layer_frame.frame_id = frame_id;
    layer_frame.temporal_id =
        codec_specific_info->codecSpecific.VP9.temporal_idx;
    if (layer_frame.temporal_id == kNoTemporalIdx) {
      layer_frame.temporal_id = 0;
    }
    layer_frame.vp9_non_ref_for_inter_layer_pred =
        codec_specific_info->codecSpecific.VP9.non_ref_for_inter_layer_pred;
    CheckVp9References(layer_frame, codec_specific_info->codecSpecific.VP9);

    if (codec_specific_info->generic_frame_info.has_value()) {
      absl::InlinedVector<int64_t, 5> frame_dependencies =
          dependencies_calculator_.FromBuffersUsage(
              frame_id,
              codec_specific_info->generic_frame_info->encoder_buffers);

      CheckGenericReferences(frame_dependencies,
                             *codec_specific_info->generic_frame_info);
      CheckGenericAndCodecSpecificReferencesAreConsistent(
          frame_dependencies, *codec_specific_info, layer_frame);
    }

    return Result(Result::OK);
  }

 private:
  // With 4 spatial layers and patterns up to 8 pictures, it should be enough to
  // keep the last 32 frames to validate dependencies.
  static constexpr size_t kMaxFrameHistorySize = 32;
  struct LayerFrame {
    int64_t frame_id;
    int64_t picture_id;
    int spatial_id;
    int temporal_id;
    bool vp9_non_ref_for_inter_layer_pred;
  };

  void CheckVp9References(const LayerFrame& layer_frame,
                          const CodecSpecificInfoVP9& vp9_info) {
    if (layer_frame.frame_id == 0) {
      RTC_CHECK(!vp9_info.inter_layer_predicted);
    } else {
      const LayerFrame& previous_frame = Frame(layer_frame.frame_id - 1);
      if (vp9_info.inter_layer_predicted) {
        RTC_CHECK(!previous_frame.vp9_non_ref_for_inter_layer_pred);
        RTC_CHECK_EQ(layer_frame.picture_id, previous_frame.picture_id);
      }
      if (previous_frame.picture_id == layer_frame.picture_id) {
        RTC_CHECK_GT(layer_frame.spatial_id, previous_frame.spatial_id);
        // The check below would fail for temporal shift structures. Remove it
        // or move it to !flexible_mode section when vp9 encoder starts
        // supporting such structures.
        RTC_CHECK_EQ(layer_frame.temporal_id, previous_frame.temporal_id);
      }
    }
    if (!vp9_info.flexible_mode) {
      if (vp9_info.gof.num_frames_in_gof > 0) {
        gof_.CopyGofInfoVP9(vp9_info.gof);
      }
      RTC_CHECK_EQ(gof_.temporal_idx[vp9_info.gof_idx],
                   layer_frame.temporal_id);
    }
  }

  void CheckGenericReferences(rtc::ArrayView<const int64_t> frame_dependencies,
                              const GenericFrameInfo& generic_info) const {
    for (int64_t dependency_frame_id : frame_dependencies) {
      RTC_CHECK_GE(dependency_frame_id, 0);
      const LayerFrame& dependency = Frame(dependency_frame_id);
      RTC_CHECK_GE(generic_info.spatial_id, dependency.spatial_id);
      RTC_CHECK_GE(generic_info.temporal_id, dependency.temporal_id);
    }
  }

  void CheckGenericAndCodecSpecificReferencesAreConsistent(
      rtc::ArrayView<const int64_t> frame_dependencies,
      const CodecSpecificInfo& info,
      const LayerFrame& layer_frame) const {
    const CodecSpecificInfoVP9& vp9_info = info.codecSpecific.VP9;
    const GenericFrameInfo& generic_info = *info.generic_frame_info;

    RTC_CHECK_EQ(generic_info.spatial_id, layer_frame.spatial_id);
    RTC_CHECK_EQ(generic_info.temporal_id, layer_frame.temporal_id);
    auto picture_id_diffs =
        rtc::MakeArrayView(vp9_info.p_diff, vp9_info.num_ref_pics);
    RTC_CHECK_EQ(
        frame_dependencies.size(),
        picture_id_diffs.size() + (vp9_info.inter_layer_predicted ? 1 : 0));
    for (int64_t dependency_frame_id : frame_dependencies) {
      RTC_CHECK_GE(dependency_frame_id, 0);
      const LayerFrame& dependency = Frame(dependency_frame_id);
      if (dependency.spatial_id != layer_frame.spatial_id) {
        RTC_CHECK(vp9_info.inter_layer_predicted);
        RTC_CHECK_EQ(layer_frame.picture_id, dependency.picture_id);
        RTC_CHECK_GT(layer_frame.spatial_id, dependency.spatial_id);
      } else {
        RTC_CHECK(vp9_info.inter_pic_predicted);
        RTC_CHECK_EQ(layer_frame.spatial_id, dependency.spatial_id);
        RTC_CHECK(absl::c_linear_search(
            picture_id_diffs, layer_frame.picture_id - dependency.picture_id));
      }
    }
  }

  const LayerFrame& Frame(int64_t frame_id) const {
    auto& frame = frames_[frame_id % kMaxFrameHistorySize];
    RTC_CHECK_EQ(frame.frame_id, frame_id);
    return frame;
  }

  GofInfoVP9 gof_;
  int64_t frame_id_ = 0;
  int64_t picture_id_ = 1;
  FrameDependenciesCalculator dependencies_calculator_;
  LayerFrame frames_[kMaxFrameHistorySize];
};

class FieldTrials : public WebRtcKeyValueConfig {
 public:
  explicit FieldTrials(FuzzDataHelper& config)
      : flags_(config.ReadOrDefaultValue<uint8_t>(0)) {}

  ~FieldTrials() override = default;
  std::string Lookup(absl::string_view key) const override {
    static constexpr absl::string_view kBinaryFieldTrials[] = {
        "WebRTC-Vp9DependencyDescriptor",
        "WebRTC-Vp9ExternalRefCtrl",
        "WebRTC-Vp9IssueKeyFrameOnLayerDeactivation",
    };
    for (size_t i = 0; i < ABSL_ARRAYSIZE(kBinaryFieldTrials); ++i) {
      if (key == kBinaryFieldTrials[i]) {
        return (flags_ & (1u << i)) ? "Enabled" : "Disabled";
      }
    }

    // Ignore following field trials.
    if (key == "WebRTC-CongestionWindow" ||
        key == "WebRTC-UseBaseHeavyVP8TL3RateAllocation" ||
        key == "WebRTC-SimulcastUpswitchHysteresisPercent" ||
        key == "WebRTC-SimulcastScreenshareUpswitchHysteresisPercent" ||
        key == "WebRTC-VideoRateControl" ||
        key == "WebRTC-VP9-PerformanceFlags" ||
        key == "WebRTC-VP9VariableFramerateScreenshare" ||
        key == "WebRTC-VP9QualityScaler") {
      return "";
    }
    // Crash when using unexpected field trial to decide if it should be fuzzed
    // or have a constant value.
    RTC_CHECK(false) << "Unfuzzed field trial " << key << "\n";
  }

 private:
  const uint8_t flags_;
};

VideoCodec CodecSettings(FuzzDataHelper& rng) {
  uint16_t config = rng.ReadOrDefaultValue<uint16_t>(0);
  // Test up to to 4 spatial and 4 temporal layers.
  int num_spatial_layers = 1 + (config & 0b11);
  int num_temporal_layers = 1 + ((config >> 2) & 0b11);

  VideoCodec codec_settings = {};
  codec_settings.codecType = kVideoCodecVP9;
  codec_settings.maxFramerate = 30;
  codec_settings.width = 320 << (num_spatial_layers - 1);
  codec_settings.height = 180 << (num_spatial_layers - 1);
  if (num_spatial_layers > 1) {
    for (int sid = 0; sid < num_spatial_layers; ++sid) {
      SpatialLayer& spatial_layer = codec_settings.spatialLayers[sid];
      codec_settings.width = 320 << sid;
      codec_settings.height = 180 << sid;
      spatial_layer.width = codec_settings.width;
      spatial_layer.height = codec_settings.height;
      spatial_layer.targetBitrate = kBitrateEnabledBps * num_temporal_layers;
      spatial_layer.maxFramerate = codec_settings.maxFramerate;
      spatial_layer.numberOfTemporalLayers = num_temporal_layers;
    }
  }
  codec_settings.VP9()->numberOfSpatialLayers = num_spatial_layers;
  codec_settings.VP9()->numberOfTemporalLayers = num_temporal_layers;
  int inter_layer_pred = (config >> 4) & 0b11;
  // There are only 3 valid values.
  codec_settings.VP9()->interLayerPred = static_cast<InterLayerPredMode>(
      inter_layer_pred < 3 ? inter_layer_pred : 0);
  codec_settings.VP9()->flexibleMode = (config & (1u << 6)) != 0;
  codec_settings.VP9()->frameDroppingOn = (config & (1u << 7)) != 0;
  codec_settings.mode = VideoCodecMode::kRealtimeVideo;
  return codec_settings;
}

VideoEncoder::Settings EncoderSettings() {
  return VideoEncoder::Settings(VideoEncoder::Capabilities(false),
                                /*number_of_cores=*/1,
                                /*max_payload_size=*/0);
}

bool IsSupported(int num_spatial_layers,
                 int num_temporal_layers,
                 const VideoBitrateAllocation& allocation) {
  // VP9 encoder doesn't support certain configurations.
  // BitrateAllocator shouldn't produce them.
  if (allocation.get_sum_bps() == 0) {
    // Ignore allocation that turns off all the layers.
    // In such a case it is up to upper layer code not to call Encode.
    return false;
  }

  for (int tid = 0; tid < num_temporal_layers; ++tid) {
    int min_enabled_spatial_id = -1;
    int max_enabled_spatial_id = -1;
    int num_enabled_spatial_layers = 0;
    for (int sid = 0; sid < num_spatial_layers; ++sid) {
      if (allocation.GetBitrate(sid, tid) > 0) {
        if (min_enabled_spatial_id == -1) {
          min_enabled_spatial_id = sid;
        }
        max_enabled_spatial_id = sid;
        ++num_enabled_spatial_layers;
      }
    }
    if (num_enabled_spatial_layers == 0) {
      // Each temporal layer should be enabled because skipping a full frame is
      // not supported in non-flexible mode.
      return false;
    }
    if (max_enabled_spatial_id - min_enabled_spatial_id + 1 !=
        num_enabled_spatial_layers) {
      // To avoid odd spatial dependencies, there should be no gaps in active
      // spatial layers.
      return false;
    }
  }

  return true;
}

struct LibvpxState {
  LibvpxState() {
    pkt.kind = VPX_CODEC_CX_FRAME_PKT;
    pkt.data.frame.buf = pkt_buffer;
    pkt.data.frame.sz = ABSL_ARRAYSIZE(pkt_buffer);
    layer_id.spatial_layer_id = -1;
  }

  uint8_t pkt_buffer[1000] = {};
  vpx_codec_enc_cfg_t config = {};
  vpx_codec_priv_output_cx_pkt_cb_pair_t callback = {};
  vpx_image_t img = {};
  vpx_svc_ref_frame_config_t ref_config = {};
  vpx_svc_layer_id_t layer_id = {};
  vpx_svc_frame_drop_t frame_drop = {};
  vpx_codec_cx_pkt pkt = {};
};

class StubLibvpx : public NiceMock<MockLibvpxInterface> {
 public:
  explicit StubLibvpx(LibvpxState* state) : state_(state) { RTC_CHECK(state_); }

  vpx_codec_err_t codec_enc_config_default(vpx_codec_iface_t* iface,
                                           vpx_codec_enc_cfg_t* cfg,
                                           unsigned int usage) const override {
    state_->config = *cfg;
    return VPX_CODEC_OK;
  }

  vpx_codec_err_t codec_enc_init(vpx_codec_ctx_t* ctx,
                                 vpx_codec_iface_t* iface,
                                 const vpx_codec_enc_cfg_t* cfg,
                                 vpx_codec_flags_t flags) const override {
    RTC_CHECK(ctx);
    ctx->err = VPX_CODEC_OK;
    return VPX_CODEC_OK;
  }

  vpx_image_t* img_wrap(vpx_image_t* img,
                        vpx_img_fmt_t fmt,
                        unsigned int d_w,
                        unsigned int d_h,
                        unsigned int stride_align,
                        unsigned char* img_data) const override {
    state_->img.fmt = fmt;
    state_->img.d_w = d_w;
    state_->img.d_h = d_h;
    return &state_->img;
  }

  vpx_codec_err_t codec_encode(vpx_codec_ctx_t* ctx,
                               const vpx_image_t* img,
                               vpx_codec_pts_t pts,
                               uint64_t duration,
                               vpx_enc_frame_flags_t flags,
                               uint64_t deadline) const override {
    if (flags & VPX_EFLAG_FORCE_KF) {
      state_->pkt.data.frame.flags = VPX_FRAME_IS_KEY;
    } else {
      state_->pkt.data.frame.flags = 0;
    }
    state_->pkt.data.frame.duration = duration;
    return VPX_CODEC_OK;
  }

  vpx_codec_err_t codec_control(vpx_codec_ctx_t* ctx,
                                vp8e_enc_control_id ctrl_id,
                                void* param) const override {
    if (ctrl_id == VP9E_REGISTER_CX_CALLBACK) {
      state_->callback =
          *reinterpret_cast<vpx_codec_priv_output_cx_pkt_cb_pair_t*>(param);
    }
    return VPX_CODEC_OK;
  }

  vpx_codec_err_t codec_control(
      vpx_codec_ctx_t* ctx,
      vp8e_enc_control_id ctrl_id,
      vpx_svc_ref_frame_config_t* param) const override {
    switch (ctrl_id) {
      case VP9E_SET_SVC_REF_FRAME_CONFIG:
        state_->ref_config = *param;
        break;
      case VP9E_GET_SVC_REF_FRAME_CONFIG:
        *param = state_->ref_config;
        break;
      default:
        break;
    }
    return VPX_CODEC_OK;
  }

  vpx_codec_err_t codec_control(vpx_codec_ctx_t* ctx,
                                vp8e_enc_control_id ctrl_id,
                                vpx_svc_layer_id_t* param) const override {
    switch (ctrl_id) {
      case VP9E_SET_SVC_LAYER_ID:
        state_->layer_id = *param;
        break;
      case VP9E_GET_SVC_LAYER_ID:
        *param = state_->layer_id;
        break;
      default:
        break;
    }
    return VPX_CODEC_OK;
  }

  vpx_codec_err_t codec_control(vpx_codec_ctx_t* ctx,
                                vp8e_enc_control_id ctrl_id,
                                vpx_svc_frame_drop_t* param) const override {
    if (ctrl_id == VP9E_SET_SVC_FRAME_DROP_LAYER) {
      state_->frame_drop = *param;
    }
    return VPX_CODEC_OK;
  }

  vpx_codec_err_t codec_enc_config_set(
      vpx_codec_ctx_t* ctx,
      const vpx_codec_enc_cfg_t* cfg) const override {
    state_->config = *cfg;
    return VPX_CODEC_OK;
  }

 private:
  LibvpxState* const state_;
};

enum Actions {
  kEncode,
  kSetRates,
};

// When a layer frame is marked for drop, drops all layer frames from that
// pictures with larger spatial ids.
constexpr bool DropAbove(uint8_t layers_mask, int sid) {
  uint8_t full_mask = (uint8_t{1} << (sid + 1)) - 1;
  return (layers_mask & full_mask) != full_mask;
}
// inline unittests
static_assert(DropAbove(0b1011, /*sid=*/0) == false, "");
static_assert(DropAbove(0b1011, /*sid=*/1) == false, "");
static_assert(DropAbove(0b1011, /*sid=*/2) == true, "");
static_assert(DropAbove(0b1011, /*sid=*/3) == true, "");

// When a layer frame is marked for drop, drops all layer frames from that
// pictures with smaller spatial ids.
constexpr bool DropBelow(uint8_t layers_mask, int sid, int num_layers) {
  return (layers_mask >> sid) != (1 << (num_layers - sid)) - 1;
}
// inline unittests
static_assert(DropBelow(0b1101, /*sid=*/0, 4) == true, "");
static_assert(DropBelow(0b1101, /*sid=*/1, 4) == true, "");
static_assert(DropBelow(0b1101, /*sid=*/2, 4) == false, "");
static_assert(DropBelow(0b1101, /*sid=*/3, 4) == false, "");

}  // namespace

void FuzzOneInput(const uint8_t* data, size_t size) {
  FuzzDataHelper helper(rtc::MakeArrayView(data, size));

  FrameValidator validator;
  FieldTrials field_trials(helper);
  // Setup call callbacks for the fake
  LibvpxState state;

  // Initialize encoder
  LibvpxVp9Encoder encoder(cricket::VideoCodec(),
                           std::make_unique<StubLibvpx>(&state), field_trials);
  VideoCodec codec = CodecSettings(helper);
  if (encoder.InitEncode(&codec, EncoderSettings()) != WEBRTC_VIDEO_CODEC_OK) {
    return;
  }
  RTC_CHECK_EQ(encoder.RegisterEncodeCompleteCallback(&validator),
               WEBRTC_VIDEO_CODEC_OK);
  {
    // Enable all the layers initially. Encoder doesn't support producing
    // frames when no layers are enabled.
    LibvpxVp9Encoder::RateControlParameters parameters;
    parameters.framerate_fps = 30.0;
    for (int sid = 0; sid < codec.VP9()->numberOfSpatialLayers; ++sid) {
      for (int tid = 0; tid < codec.VP9()->numberOfTemporalLayers; ++tid) {
        parameters.bitrate.SetBitrate(sid, tid, kBitrateEnabledBps);
      }
    }
    encoder.SetRates(parameters);
  }

  std::vector<VideoFrameType> frame_types(1);
  VideoFrame fake_image = VideoFrame::Builder()
                              .set_video_frame_buffer(I420Buffer::Create(
                                  int{codec.width}, int{codec.height}))
                              .build();

  // Start producing frames at random.
  while (helper.CanReadBytes(1)) {
    uint8_t action = helper.Read<uint8_t>();
    switch (action & 0b11) {
      case kEncode: {
        // bitmask of the action: SSSS-K00, where
        // four S bit indicate which spatial layers should be produced,
        // K bit indicates if frame should be a key frame.
        frame_types[0] = (action & 0b100) ? VideoFrameType::kVideoFrameKey
                                          : VideoFrameType::kVideoFrameDelta;
        encoder.Encode(fake_image, &frame_types);
        uint8_t encode_spatial_layers = (action >> 4);
        for (size_t sid = 0; sid < state.config.ss_number_layers; ++sid) {
          if (state.config.ss_target_bitrate[sid] == 0) {
            // Don't encode disabled spatial layers.
            continue;
          }
          bool drop = true;
          switch (state.frame_drop.framedrop_mode) {
            case FULL_SUPERFRAME_DROP:
              drop = encode_spatial_layers == 0;
              break;
            case LAYER_DROP:
              drop = (encode_spatial_layers & (1 << sid)) == 0;
              break;
            case CONSTRAINED_LAYER_DROP:
              drop = DropBelow(encode_spatial_layers, sid,
                               state.config.ss_number_layers);
              break;
            case CONSTRAINED_FROM_ABOVE_DROP:
              drop = DropAbove(encode_spatial_layers, sid);
              break;
          }
          if (!drop) {
            state.layer_id.spatial_layer_id = sid;
            state.callback.output_cx_pkt(&state.pkt, state.callback.user_priv);
          }
        }
      } break;
      case kSetRates: {
        // bitmask of the action: (S2)(S1)(S0)01,
        // where Sx is number of temporal layers to enable for spatial layer x
        // In pariculat Sx = 0 indicates spatial layer x should be disabled.
        LibvpxVp9Encoder::RateControlParameters parameters;
        parameters.framerate_fps = 30.0;
        for (int sid = 0; sid < codec.VP9()->numberOfSpatialLayers; ++sid) {
          int temporal_layers = (action >> ((1 + sid) * 2)) & 0b11;
          for (int tid = 0; tid < temporal_layers; ++tid) {
            parameters.bitrate.SetBitrate(sid, tid, kBitrateEnabledBps);
          }
        }
        if (IsSupported(codec.VP9()->numberOfSpatialLayers,
                        codec.VP9()->numberOfTemporalLayers,
                        parameters.bitrate)) {
          encoder.SetRates(parameters);
        }
      } break;
      default:
        // Unspecificed values are noop.
        break;
    }
  }
}
}  // namespace webrtc