Add developer CLI flags for tuning encoder internals

Currently adds the following:
- `--deblock-strength`
- `--deblock-sharpness`
- `--ip-qidx-ratio`
- `--temporal-rdo-strength`

These flags will be available only if the `devel` feature flag
is active, as they are intended to be used to ease development
and not to be tweaked by end users. rav1e continues to subscribe
to the philosophy that end users should not need to be cargo culting
their command lines to get optimal results, and we would prefer to
tune the encoder internals so that users do not have to.

Author: shssoichiro

.github/workflows/rav1e.yml

@@ -12,7 +12,6 @@ on:
 
 jobs:
   rustfmt-clippy:
-
     runs-on: ubuntu-22.04
 
     steps:
@@ -193,7 +192,7 @@ jobs:
       - name: Check extra features
         if: matrix.toolchain == 'stable' && matrix.conf == 'check-extra-feats'
         run: |
-          cargo check --features=check_asm,capi,dump_lookahead_data,serialize,bench --all-targets
+          cargo check --features=check_asm,capi,dump_lookahead_data,serialize,bench,devel --all-targets
       - name: Check extra features
         if: matrix.toolchain == 'stable' && matrix.conf == 'check-unstable-feats'
         run: |

Cargo.toml

@@ -24,6 +24,9 @@ default-run = "rav1e"
 
 [features]
 unstable = []
+# Exposes extra flags for tuning compiler internals.
+# Intended to be used by developers to find ideal internal settings.
+devel = []
 channel-api = ["crossbeam"]
 decode_test = ["aom-sys"]
 decode_test_dav1d = ["dav1d-sys"]

src/api/config/encoder.rs

@@ -108,6 +108,38 @@ pub struct EncoderConfig {
 
   /// Settings which affect the encoding speed vs. quality trade-off.
   pub speed_settings: SpeedSettings,
+
+  /// Advanced settings which are intended for use by developers.
+  /// Non-developers should use the default values.
+  pub advanced_flags: AdvancedTuning,
+}
+
+/// Advanced settings that are intended for use by developers
+/// for tuning compiler internals.
+#[derive(Debug, Clone, Copy, Serialize, Deserialize)]
+pub struct AdvancedTuning {
+  /// Controls the strength of the deblock filter, as a multiplier to the default.
+  pub deblock_strength: f32,
+  /// Controls the sharpness of the deblock filter. Accepts a value from 0-7.
+  pub deblock_sharpness: u8,
+  /// Controls the ratio between intra frame and inter frame quantizers, as a multiplier.
+  /// Default is 1.0. Higher values create a higher quantizer difference, while lower values
+  /// create a lower quantizer difference. A value of 0.0 would mean that I and P quantizers
+  /// are the same.
+  pub ip_qidx_ratio: f32,
+  /// Controls the strength of temporal RDO, as a multiplier to the default.
+  pub temporal_rdo_strength: f32,
+}
+
+impl Default for AdvancedTuning {
+  fn default() -> Self {
+    Self {
+      deblock_strength: 1.0,
+      deblock_sharpness: 0,
+      ip_qidx_ratio: 1.0,
+      temporal_rdo_strength: 1.0,
+    }
+  }
 }
 
 /// Default preset for `EncoderConfig`: it is a balance between quality and
@@ -163,6 +195,7 @@ impl EncoderConfig {
       tile_rows: 0,
       tiles: 0,
       speed_settings: SpeedSettings::from_preset(speed),
+      advanced_flags: Default::default(),
     }
   }
 

src/api/test.rs

@@ -2164,6 +2164,7 @@ fn log_q_exp_overflow() {
       },
       ..Default::default()
     },
+    advanced_flags: Default::default(),
   };
   let config = Config::new().with_encoder_config(enc).with_threads(1);
 
@@ -2240,6 +2241,7 @@ fn guess_frame_subtypes_assert() {
       },
       ..Default::default()
     },
+    advanced_flags: Default::default(),
   };
   let config = Config::new().with_encoder_config(enc).with_threads(1);
 

src/bin/common.rs

@@ -242,10 +242,34 @@ pub struct CliOptions {
   #[clap(long, short, value_parser, help_heading = "DEBUGGING")]
   pub reconstruction: Option<PathBuf>,
 
+  /// Controls the strength of the deblock filter, as a multiplier to the default.
+  #[clap(long, value_parser = positive_float, default_value_t=1.0f32, help_heading = "ADVANCED")]
+  pub deblock_strength: f32,
+  /// Controls the sharpness of the deblock filter. Accepts a value from 0-7.
+  #[clap(long, value_parser = clap::value_parser!(u8).range(0..=7), default_value_t=0, help_heading = "ADVANCED")]
+  pub deblock_sharpness: u8,
+  /// Controls the ratio between intra frame and inter frame quantizers, as a multiplier.
+  /// Higher values create a higher quantizer difference, while lower values
+  /// create a lower quantizer difference. A value of 0.0 would mean that I and P quantizers
+  /// are the same.
+  #[clap(long, value_parser = positive_float, default_value_t=1.0f32, help_heading = "ADVANCED")]
+  pub ip_qidx_ratio: f32,
+  /// Controls the strength of temporal RDO, as a multiplier to the default.
+  #[clap(long, value_parser = positive_float, default_value_t=1.0f32, help_heading = "ADVANCED")]
+  pub temporal_rdo_strength: f32,
+
   #[clap(subcommand)]
   pub command: Option<Commands>,
 }
 
+fn positive_float(input: &str) -> Result<f32, String> {
+  let value = input.parse::<f32>().map_err(|e| e.to_string())?;
+  if value < 0.0 {
+    return Err("Value must not be negative".to_string());
+  }
+  Ok(value)
+}
+
 fn get_version() -> &'static str {
   static VERSION_STR: Lazy<String> = Lazy::new(|| {
     format!(
@@ -299,7 +323,7 @@ pub enum Commands {
     #[clap(long, short, value_parser)]
     save_config: Option<PathBuf>,
     /// Load the encoder configuration from a toml file
-    #[clap(long, short, value_parser, conflicts_with = "save-config")]
+    #[clap(long, short, value_parser, conflicts_with = "save_config")]
     load_config: Option<PathBuf>,
   },
 }
@@ -484,6 +508,16 @@ pub fn parse_cli() -> Result<ParsedCliOptions, CliError> {
   })
 }
 
+#[cfg(feature = "devel")]
+const fn parse_advanced_flags(cli: &CliOptions) -> AdvancedTuning {
+  AdvancedTuning {
+    deblock_strength: cli.deblock_strength,
+    deblock_sharpness: cli.deblock_sharpness,
+    ip_qidx_ratio: cli.ip_qidx_ratio,
+    temporal_rdo_strength: cli.temporal_rdo_strength,
+  }
+}
+
 fn parse_config(matches: &CliOptions) -> Result<EncoderConfig, CliError> {
   let maybe_quantizer = matches.quantizer;
   let maybe_bitrate = matches.bitrate;
@@ -674,5 +708,10 @@ fn parse_config(matches: &CliOptions) -> Result<EncoderConfig, CliError> {
     cfg.speed_settings.scene_detection_mode = SceneDetectionSpeed::None;
   }
 
+  #[cfg(feature = "devel")]
+  {
+    cfg.advanced_flags = parse_advanced_flags(matches);
+  }
+
   Ok(cfg)
 }

src/encoder.rs

@@ -501,7 +501,7 @@ impl<T: Pixel> FrameState<T> {
       cdfs: CDFContext::new(0),
       context_update_tile_id: 0,
       max_tile_size_bytes: 0,
-      deblock: Default::default(),
+      deblock: DeblockState::new(&fi.config),
       segmentation: Default::default(),
       restoration: rs,
       frame_me_stats: FrameMEStats::new_arc_array(fi.w_in_b, fi.h_in_b),
@@ -543,6 +543,19 @@ pub struct DeblockState {
   pub block_delta_multi: bool,
 }
 
+impl DeblockState {
+  pub fn new(config: &EncoderConfig) -> Self {
+    let mut state = DeblockState { ..Default::default() };
+    for level in &mut state.levels {
+      *level = ((*level as f32) * config.advanced_flags.deblock_strength)
+        .min(MAX_LOOP_FILTER as f32)
+        .round() as u8;
+    }
+    state.sharpness = config.advanced_flags.deblock_sharpness;
+    state
+  }
+}
+
 impl Default for DeblockState {
   fn default() -> Self {
     DeblockState {

src/fuzzing.rs

@@ -257,6 +257,7 @@ impl Arbitrary for ArbitraryEncoder {
       switch_frame_interval: u.int_in_range(0..=3)?,
       tune: *u.choose(&[Tune::Psnr, Tune::Psychovisual])?,
       film_grain_params: None,
+      advanced_flags: Default::default(),
     };
 
     let frame_count =

src/rate.rs

@@ -703,11 +703,12 @@ impl RCState {
 
   pub(crate) fn select_first_pass_qi(
     &self, bit_depth: usize, fti: usize, chroma_sampling: ChromaSampling,
+    ft_ratio: f64,
   ) -> QuantizerParameters {
     // Adjust the quantizer for the frame type, result is Q57:
     let log_q = ((self.pass1_log_base_q + (1i64 << 11)) >> 12)
       * (MQP_Q12[fti] as i64)
-      + DQP_Q57[fti];
+      + (DQP_Q57[fti] as f64 * ft_ratio) as i64;
     QuantizerParameters::new_from_log_q(
       self.pass1_log_base_q,
       log_q,
@@ -723,14 +724,20 @@ impl RCState {
     &self, ctx: &ContextInner<T>, output_frameno: u64, fti: usize,
     maybe_prev_log_base_q: Option<i64>, log_isqrt_mean_scale: i64,
   ) -> QuantizerParameters {
+    let ft_ratio = ctx.config.advanced_flags.ip_qidx_ratio as f64;
+
     // Is rate control active?
     if self.target_bitrate <= 0 {
       // Rate control is not active.
       // Derive quantizer directly from frame type.
       let bit_depth = ctx.config.bit_depth;
       let chroma_sampling = ctx.config.chroma_sampling;
-      let (log_base_q, log_q) =
-        Self::calc_flat_quantizer(ctx.config.quantizer as u8, bit_depth, fti);
+      let (log_base_q, log_q) = Self::calc_flat_quantizer(
+        ctx.config.quantizer as u8,
+        bit_depth,
+        fti,
+        ft_ratio,
+      );
       QuantizerParameters::new_from_log_q(
         log_base_q,
         log_q,
@@ -752,6 +759,7 @@ impl RCState {
             ctx.config.bit_depth,
             fti,
             ctx.config.chroma_sampling,
+            ft_ratio,
           );
         }
         // Second pass of 2-pass mode: we know exactly how much of each frame
@@ -925,7 +933,7 @@ impl RCState {
           // Modulate base quantizer by frame type.
           let log_q = ((log_base_q + (1i64 << 11)) >> 12)
             * (MQP_Q12[ftj] as i64)
-            + DQP_Q57[ftj];
+            + (DQP_Q57[ftj] as f64 * ft_ratio) as i64;
           // All the fields here are Q57 except for the exponent, which is
           //  Q6.
           bits += (nframes[ftj] as i64)
@@ -959,7 +967,7 @@ impl RCState {
       // Modulate base quantizer by frame type.
       let mut log_q = ((log_base_q + (1i64 << 11)) >> 12)
         * (MQP_Q12[fti] as i64)
-        + DQP_Q57[fti];
+        + (DQP_Q57[fti] as f64 * ft_ratio) as i64;
       // The above allocation looks only at the total rate we'll accumulate
       //  in the next reservoir_frame_delay frames.
       // However, we could overflow the bit reservoir on the very next
@@ -1019,14 +1027,22 @@ impl RCState {
       }
 
       if let Some(qi_max) = self.maybe_ac_qi_max {
-        let (max_log_base_q, max_log_q) =
-          Self::calc_flat_quantizer(qi_max, ctx.config.bit_depth, fti);
+        let (max_log_base_q, max_log_q) = Self::calc_flat_quantizer(
+          qi_max,
+          ctx.config.bit_depth,
+          fti,
+          ft_ratio,
+        );
         log_base_q = cmp::min(log_base_q, max_log_base_q);
         log_q = cmp::min(log_q, max_log_q);
       }
       if self.ac_qi_min > 0 {
-        let (min_log_base_q, min_log_q) =
-          Self::calc_flat_quantizer(self.ac_qi_min, ctx.config.bit_depth, fti);
+        let (min_log_base_q, min_log_q) = Self::calc_flat_quantizer(
+          self.ac_qi_min,
+          ctx.config.bit_depth,
+          fti,
+          ft_ratio,
+        );
         log_base_q = cmp::max(log_base_q, min_log_base_q);
         log_q = cmp::max(log_q, min_log_q);
       }
@@ -1044,7 +1060,7 @@ impl RCState {
   // Computes a quantizer directly from the frame type and base quantizer index,
   // without consideration for rate control.
   fn calc_flat_quantizer(
-    base_qi: u8, bit_depth: usize, fti: usize,
+    base_qi: u8, bit_depth: usize, fti: usize, ft_ratio: f64,
   ) -> (i64, i64) {
     // TODO: Rename "quantizer" something that indicates it is a quantizer
     //  index, and move it somewhere more sensible (or choose a better way to
@@ -1063,7 +1079,7 @@ impl RCState {
     let log_base_q = (log_ac_q + log_dc_q + 1) >> 1;
     // Adjust the quantizer for the frame type, result is Q57:
     let log_q = ((log_base_q + (1i64 << 11)) >> 12) * (MQP_Q12[fti] as i64)
-      + DQP_Q57[fti];
+      + (DQP_Q57[fti] as f64 * ft_ratio) as i64;
     (log_base_q, log_q)
   }
 

src/rdo.rs

@@ -455,6 +455,7 @@ pub fn distortion_scale<T: Pixel>(
 
   let coded_data = fi.coded_frame_data.as_ref().unwrap();
   coded_data.distortion_scales[y * coded_data.w_in_imp_b + x]
+    .strength_adjusted(fi.config.advanced_flags.temporal_rdo_strength as f64)
 }
 
 /// # Panics
@@ -504,6 +505,7 @@ pub fn spatiotemporal_scale<T: Pixel>(
     .sum::<u64>();
   }
   DistortionScale(((sum + (den >> 1)) / den) as u32)
+    .strength_adjusted(fi.config.advanced_flags.temporal_rdo_strength as f64)
 }
 
 pub fn distortion_scale_for(
@@ -617,6 +619,22 @@ impl DistortionScale {
   pub const fn mul_u64(self, dist: u64) -> u64 {
     (self.0 as u64 * dist + (1 << Self::SHIFT >> 1)) >> Self::SHIFT
   }
+
+  #[inline]
+  #[cfg(feature = "devel")]
+  pub fn strength_adjusted(self, strength: f64) -> Self {
+    let diff = 1.0 - f64::from(self);
+    let add = diff * strength;
+    DistortionScale::from((1.0 + add).max(0.0))
+  }
+
+  #[inline(always)]
+  #[cfg(not(feature = "devel"))]
+  pub fn strength_adjusted(self, _strength: f64) -> Self {
+    // If we aren't using a devel build, just return self
+    // so we do not add any performance cost.
+    self
+  }
 }
 
 impl std::ops::Mul for DistortionScale {