| 1 | use crate::api::internal::InterConfig; |
| 2 | use crate::config::EncoderConfig; |
| 3 | use crate::context::{BlockOffset, FrameBlocks, TileBlockOffset}; |
| 4 | use crate::cpu_features::CpuFeatureLevel; |
| 5 | use crate::dist::get_satd; |
| 6 | use crate::encoder::{ |
| 7 | FrameInvariants, FrameState, Sequence, IMPORTANCE_BLOCK_SIZE, |
| 8 | }; |
| 9 | use crate::frame::{AsRegion, PlaneOffset}; |
| 10 | use crate::me::{estimate_tile_motion, RefMEStats}; |
| 11 | use crate::partition::{get_intra_edges, BlockSize}; |
| 12 | use crate::predict::{IntraParam, PredictionMode}; |
| 13 | use crate::tiling::{Area, PlaneRegion, TileRect}; |
| 14 | use crate::transform::TxSize; |
| 15 | use crate::util::Aligned; |
| 16 | use crate::Pixel; |
| 17 | use rayon::iter::*; |
| 18 | use std::sync::Arc; |
| 19 | use v_frame::frame::Frame; |
| 20 | use v_frame::pixel::CastFromPrimitive; |
| 21 | use v_frame::plane::Plane; |
| 22 | |
| 23 | pub(crate) const IMP_BLOCK_MV_UNITS_PER_PIXEL: i64 = 8; |
| 24 | pub(crate) const IMP_BLOCK_SIZE_IN_MV_UNITS: i64 = |
| 25 | IMPORTANCE_BLOCK_SIZE as i64 * IMP_BLOCK_MV_UNITS_PER_PIXEL; |
| 26 | pub(crate) const IMP_BLOCK_AREA_IN_MV_UNITS: i64 = |
| 27 | IMP_BLOCK_SIZE_IN_MV_UNITS * IMP_BLOCK_SIZE_IN_MV_UNITS; |
| 28 | |
| 29 | #[profiling::function ] |
| 30 | pub(crate) fn estimate_intra_costs<T: Pixel>( |
| 31 | temp_plane: &mut Plane<T>, frame: &Frame<T>, bit_depth: usize, |
| 32 | cpu_feature_level: CpuFeatureLevel, |
| 33 | ) -> Box<[u32]> { |
| 34 | let plane = &frame.planes[0]; |
| 35 | let plane_after_prediction = temp_plane; |
| 36 | |
| 37 | let bsize = BlockSize::from_width_and_height( |
| 38 | IMPORTANCE_BLOCK_SIZE, |
| 39 | IMPORTANCE_BLOCK_SIZE, |
| 40 | ); |
| 41 | let tx_size = bsize.tx_size(); |
| 42 | |
| 43 | let h_in_imp_b = plane.cfg.height / IMPORTANCE_BLOCK_SIZE; |
| 44 | let w_in_imp_b = plane.cfg.width / IMPORTANCE_BLOCK_SIZE; |
| 45 | let mut intra_costs = Vec::with_capacity(h_in_imp_b * w_in_imp_b); |
| 46 | |
| 47 | for y in 0..h_in_imp_b { |
| 48 | for x in 0..w_in_imp_b { |
| 49 | let plane_org = plane.region(Area::Rect { |
| 50 | x: (x * IMPORTANCE_BLOCK_SIZE) as isize, |
| 51 | y: (y * IMPORTANCE_BLOCK_SIZE) as isize, |
| 52 | width: IMPORTANCE_BLOCK_SIZE, |
| 53 | height: IMPORTANCE_BLOCK_SIZE, |
| 54 | }); |
| 55 | |
| 56 | // TODO: other intra prediction modes. |
| 57 | let mut edge_buf = Aligned::uninit_array(); |
| 58 | let edge_buf = get_intra_edges( |
| 59 | &mut edge_buf, |
| 60 | &plane.as_region(), |
| 61 | TileBlockOffset(BlockOffset { x, y }), |
| 62 | 0, |
| 63 | 0, |
| 64 | bsize, |
| 65 | PlaneOffset { |
| 66 | x: (x * IMPORTANCE_BLOCK_SIZE) as isize, |
| 67 | y: (y * IMPORTANCE_BLOCK_SIZE) as isize, |
| 68 | }, |
| 69 | TxSize::TX_8X8, |
| 70 | bit_depth, |
| 71 | Some(PredictionMode::DC_PRED), |
| 72 | false, |
| 73 | IntraParam::None, |
| 74 | ); |
| 75 | |
| 76 | let mut plane_after_prediction_region = plane_after_prediction |
| 77 | .region_mut(Area::Rect { |
| 78 | x: (x * IMPORTANCE_BLOCK_SIZE) as isize, |
| 79 | y: (y * IMPORTANCE_BLOCK_SIZE) as isize, |
| 80 | width: IMPORTANCE_BLOCK_SIZE, |
| 81 | height: IMPORTANCE_BLOCK_SIZE, |
| 82 | }); |
| 83 | |
| 84 | PredictionMode::DC_PRED.predict_intra( |
| 85 | TileRect { |
| 86 | x: x * IMPORTANCE_BLOCK_SIZE, |
| 87 | y: y * IMPORTANCE_BLOCK_SIZE, |
| 88 | width: IMPORTANCE_BLOCK_SIZE, |
| 89 | height: IMPORTANCE_BLOCK_SIZE, |
| 90 | }, |
| 91 | &mut plane_after_prediction_region, |
| 92 | tx_size, |
| 93 | bit_depth, |
| 94 | &[], // Not used by DC_PRED |
| 95 | IntraParam::None, |
| 96 | None, // Not used by DC_PRED |
| 97 | &edge_buf, |
| 98 | cpu_feature_level, |
| 99 | ); |
| 100 | |
| 101 | let plane_after_prediction_region = |
| 102 | plane_after_prediction.region(Area::Rect { |
| 103 | x: (x * IMPORTANCE_BLOCK_SIZE) as isize, |
| 104 | y: (y * IMPORTANCE_BLOCK_SIZE) as isize, |
| 105 | width: IMPORTANCE_BLOCK_SIZE, |
| 106 | height: IMPORTANCE_BLOCK_SIZE, |
| 107 | }); |
| 108 | |
| 109 | let intra_cost = get_satd( |
| 110 | &plane_org, |
| 111 | &plane_after_prediction_region, |
| 112 | bsize.width(), |
| 113 | bsize.height(), |
| 114 | bit_depth, |
| 115 | cpu_feature_level, |
| 116 | ); |
| 117 | |
| 118 | intra_costs.push(intra_cost); |
| 119 | } |
| 120 | } |
| 121 | |
| 122 | intra_costs.into_boxed_slice() |
| 123 | } |
| 124 | |
| 125 | #[profiling::function ] |
| 126 | pub(crate) fn estimate_importance_block_difference<T: Pixel>( |
| 127 | frame: Arc<Frame<T>>, ref_frame: Arc<Frame<T>>, |
| 128 | ) -> f64 { |
| 129 | let plane_org = &frame.planes[0]; |
| 130 | let plane_ref = &ref_frame.planes[0]; |
| 131 | let h_in_imp_b = plane_org.cfg.height / IMPORTANCE_BLOCK_SIZE; |
| 132 | let w_in_imp_b = plane_org.cfg.width / IMPORTANCE_BLOCK_SIZE; |
| 133 | |
| 134 | let mut imp_block_costs = 0; |
| 135 | |
| 136 | (0..h_in_imp_b).for_each(|y| { |
| 137 | (0..w_in_imp_b).for_each(|x| { |
| 138 | // Coordinates of the top-left corner of the reference block, in MV |
| 139 | // units. |
| 140 | let region_org = plane_org.region(Area::Rect { |
| 141 | x: (x * IMPORTANCE_BLOCK_SIZE) as isize, |
| 142 | y: (y * IMPORTANCE_BLOCK_SIZE) as isize, |
| 143 | width: IMPORTANCE_BLOCK_SIZE, |
| 144 | height: IMPORTANCE_BLOCK_SIZE, |
| 145 | }); |
| 146 | |
| 147 | let region_ref = plane_ref.region(Area::Rect { |
| 148 | x: (x * IMPORTANCE_BLOCK_SIZE) as isize, |
| 149 | y: (y * IMPORTANCE_BLOCK_SIZE) as isize, |
| 150 | width: IMPORTANCE_BLOCK_SIZE, |
| 151 | height: IMPORTANCE_BLOCK_SIZE, |
| 152 | }); |
| 153 | |
| 154 | let sum_8x8_block = |region: &PlaneRegion<T>| { |
| 155 | region |
| 156 | .rows_iter() |
| 157 | .map(|row| { |
| 158 | // 16-bit precision is sufficient for an 8px row, as IMPORTANCE_BLOCK_SIZE * (2^12 - 1) < 2^16 - 1, |
| 159 | // so overflow is not possible |
| 160 | row.iter().map(|pixel| u16::cast_from(*pixel)).sum::<u16>() as i64 |
| 161 | }) |
| 162 | .sum::<i64>() |
| 163 | }; |
| 164 | |
| 165 | let histogram_org_sum = sum_8x8_block(®ion_org); |
| 166 | let histogram_ref_sum = sum_8x8_block(®ion_ref); |
| 167 | |
| 168 | let count = (IMPORTANCE_BLOCK_SIZE * IMPORTANCE_BLOCK_SIZE) as i64; |
| 169 | |
| 170 | let mean = (((histogram_org_sum + count / 2) / count) |
| 171 | - ((histogram_ref_sum + count / 2) / count)) |
| 172 | .abs(); |
| 173 | |
| 174 | imp_block_costs += mean as u64; |
| 175 | }); |
| 176 | }); |
| 177 | |
| 178 | imp_block_costs as f64 / (w_in_imp_b * h_in_imp_b) as f64 |
| 179 | } |
| 180 | |
| 181 | #[profiling::function ] |
| 182 | pub(crate) fn estimate_inter_costs<T: Pixel>( |
| 183 | frame: Arc<Frame<T>>, ref_frame: Arc<Frame<T>>, bit_depth: usize, |
| 184 | mut config: EncoderConfig, sequence: Arc<Sequence>, buffer: RefMEStats, |
| 185 | ) -> f64 { |
| 186 | config.low_latency = true; |
| 187 | config.speed_settings.multiref = false; |
| 188 | let inter_cfg = InterConfig::new(&config); |
| 189 | let last_fi = FrameInvariants::new_key_frame( |
| 190 | Arc::new(config), |
| 191 | sequence, |
| 192 | 0, |
| 193 | Box::new([]), |
| 194 | ); |
| 195 | let mut fi = FrameInvariants::new_inter_frame( |
| 196 | &last_fi, |
| 197 | &inter_cfg, |
| 198 | 0, |
| 199 | 1, |
| 200 | 2, |
| 201 | false, |
| 202 | Box::new([]), |
| 203 | ) |
| 204 | .unwrap(); |
| 205 | |
| 206 | // Compute the motion vectors. |
| 207 | let mut fs = FrameState::new_with_frame_and_me_stats_and_rec( |
| 208 | &fi, |
| 209 | Arc::clone(&frame), |
| 210 | buffer, |
| 211 | // We do not use this field, so we can avoid the expensive allocation |
| 212 | Arc::new(Frame { |
| 213 | planes: [ |
| 214 | Plane::new(0, 0, 0, 0, 0, 0), |
| 215 | Plane::new(0, 0, 0, 0, 0, 0), |
| 216 | Plane::new(0, 0, 0, 0, 0, 0), |
| 217 | ], |
| 218 | }), |
| 219 | ); |
| 220 | compute_motion_vectors(&mut fi, &mut fs, &inter_cfg); |
| 221 | |
| 222 | // Estimate inter costs |
| 223 | let plane_org = &frame.planes[0]; |
| 224 | let plane_ref = &ref_frame.planes[0]; |
| 225 | let h_in_imp_b = plane_org.cfg.height / IMPORTANCE_BLOCK_SIZE; |
| 226 | let w_in_imp_b = plane_org.cfg.width / IMPORTANCE_BLOCK_SIZE; |
| 227 | let stats = &fs.frame_me_stats.read().expect("poisoned lock" )[0]; |
| 228 | let bsize = BlockSize::from_width_and_height( |
| 229 | IMPORTANCE_BLOCK_SIZE, |
| 230 | IMPORTANCE_BLOCK_SIZE, |
| 231 | ); |
| 232 | |
| 233 | let mut inter_costs = 0; |
| 234 | (0..h_in_imp_b).for_each(|y| { |
| 235 | (0..w_in_imp_b).for_each(|x| { |
| 236 | let mv = stats[y * 2][x * 2].mv; |
| 237 | |
| 238 | // Coordinates of the top-left corner of the reference block, in MV |
| 239 | // units. |
| 240 | let reference_x = x as i64 * IMP_BLOCK_SIZE_IN_MV_UNITS + mv.col as i64; |
| 241 | let reference_y = y as i64 * IMP_BLOCK_SIZE_IN_MV_UNITS + mv.row as i64; |
| 242 | |
| 243 | let region_org = plane_org.region(Area::Rect { |
| 244 | x: (x * IMPORTANCE_BLOCK_SIZE) as isize, |
| 245 | y: (y * IMPORTANCE_BLOCK_SIZE) as isize, |
| 246 | width: IMPORTANCE_BLOCK_SIZE, |
| 247 | height: IMPORTANCE_BLOCK_SIZE, |
| 248 | }); |
| 249 | |
| 250 | let region_ref = plane_ref.region(Area::Rect { |
| 251 | x: reference_x as isize / IMP_BLOCK_MV_UNITS_PER_PIXEL as isize, |
| 252 | y: reference_y as isize / IMP_BLOCK_MV_UNITS_PER_PIXEL as isize, |
| 253 | width: IMPORTANCE_BLOCK_SIZE, |
| 254 | height: IMPORTANCE_BLOCK_SIZE, |
| 255 | }); |
| 256 | |
| 257 | inter_costs += get_satd( |
| 258 | ®ion_org, |
| 259 | ®ion_ref, |
| 260 | bsize.width(), |
| 261 | bsize.height(), |
| 262 | bit_depth, |
| 263 | fi.cpu_feature_level, |
| 264 | ) as u64; |
| 265 | }); |
| 266 | }); |
| 267 | inter_costs as f64 / (w_in_imp_b * h_in_imp_b) as f64 |
| 268 | } |
| 269 | |
| 270 | #[profiling::function ] |
| 271 | pub(crate) fn compute_motion_vectors<T: Pixel>( |
| 272 | fi: &mut FrameInvariants<T>, fs: &mut FrameState<T>, inter_cfg: &InterConfig, |
| 273 | ) { |
| 274 | let mut blocks: FrameBlocks = FrameBlocks::new(cols:fi.w_in_b, rows:fi.h_in_b); |
| 275 | fi.sequence |
| 276 | .tiling |
| 277 | .tile_iter_mut(fs, &mut blocks) |
| 278 | .collect::<Vec<_>>() |
| 279 | .into_par_iter() |
| 280 | .for_each(|mut ctx: TileContextMut<'_, T>| { |
| 281 | let ts: &mut TileStateMut<'_, T> = &mut ctx.ts; |
| 282 | estimate_tile_motion(fi, ts, inter_cfg); |
| 283 | }); |
| 284 | } |
| 285 | |