me.rs 29.8 KB
Newer Older
1 2 3 4 5 6 7 8 9
// Copyright (c) 2017-2018, The rav1e contributors. All rights reserved
//
// This source code is subject to the terms of the BSD 2 Clause License and
// the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
// was not distributed with this source code in the LICENSE file, you can
// obtain it at www.aomedia.org/license/software. If the Alliance for Open
// Media Patent License 1.0 was not distributed with this source code in the
// PATENTS file, you can obtain it at www.aomedia.org/license/patent.

10
#[cfg(all(target_arch = "x86_64", feature = "nasm"))]
Raphaël Zumer's avatar
Raphaël Zumer committed
11
pub use self::nasm::get_sad;
12
#[cfg(any(not(target_arch = "x86_64"), not(feature = "nasm")))]
Raphaël Zumer's avatar
Raphaël Zumer committed
13
pub use self::native::get_sad;
Raphaël Zumer's avatar
Raphaël Zumer committed
14
use crate::context::{BlockOffset, BLOCK_TO_PLANE_SHIFT, MI_SIZE};
15
use crate::encoder::ReferenceFrame;
Raphaël Zumer's avatar
Raphaël Zumer committed
16 17 18 19
use crate::FrameInvariants;
use crate::FrameState;
use crate::partition::*;
use crate::plane::*;
20
use crate::util::Pixel;
21 22

use std::sync::Arc;
23

24
#[cfg(all(target_arch = "x86_64", feature = "nasm"))]
25
mod nasm {
Raphaël Zumer's avatar
Raphaël Zumer committed
26 27
  use crate::plane::*;
  use crate::util::*;
28
  use std::mem;
Luca Barbato's avatar
Luca Barbato committed
29

Raphaël Zumer's avatar
Raphaël Zumer committed
30 31
  use libc;

Luca Barbato's avatar
Luca Barbato committed
32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61
  extern {
    fn rav1e_sad_4x4_hbd_ssse3(
      src: *const u16, src_stride: libc::ptrdiff_t, dst: *const u16,
      dst_stride: libc::ptrdiff_t
    ) -> u32;

    fn rav1e_sad_8x8_hbd10_ssse3(
      src: *const u16, src_stride: libc::ptrdiff_t, dst: *const u16,
      dst_stride: libc::ptrdiff_t
    ) -> u32;

    fn rav1e_sad_16x16_hbd_ssse3(
      src: *const u16, src_stride: libc::ptrdiff_t, dst: *const u16,
      dst_stride: libc::ptrdiff_t
    ) -> u32;

    fn rav1e_sad_32x32_hbd10_ssse3(
      src: *const u16, src_stride: libc::ptrdiff_t, dst: *const u16,
      dst_stride: libc::ptrdiff_t
    ) -> u32;

    fn rav1e_sad_64x64_hbd10_ssse3(
      src: *const u16, src_stride: libc::ptrdiff_t, dst: *const u16,
      dst_stride: libc::ptrdiff_t
    ) -> u32;

    fn rav1e_sad_128x128_hbd10_ssse3(
      src: *const u16, src_stride: libc::ptrdiff_t, dst: *const u16,
      dst_stride: libc::ptrdiff_t
    ) -> u32;
62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91

    fn rav1e_sad4x4_sse2(
      src: *const u8, src_stride: libc::ptrdiff_t, dst: *const u8,
      dst_stride: libc::ptrdiff_t
    ) -> u32;

    fn rav1e_sad8x8_sse2(
      src: *const u8, src_stride: libc::ptrdiff_t, dst: *const u8,
      dst_stride: libc::ptrdiff_t
    ) -> u32;

    fn rav1e_sad16x16_sse2(
      src: *const u8, src_stride: libc::ptrdiff_t, dst: *const u8,
      dst_stride: libc::ptrdiff_t
    ) -> u32;

    fn rav1e_sad32x32_sse2(
      src: *const u8, src_stride: libc::ptrdiff_t, dst: *const u8,
      dst_stride: libc::ptrdiff_t
    ) -> u32;

    fn rav1e_sad64x64_sse2(
      src: *const u8, src_stride: libc::ptrdiff_t, dst: *const u8,
      dst_stride: libc::ptrdiff_t
    ) -> u32;

    fn rav1e_sad128x128_sse2(
      src: *const u8, src_stride: libc::ptrdiff_t, dst: *const u8,
      dst_stride: libc::ptrdiff_t
    ) -> u32;
Luca Barbato's avatar
Luca Barbato committed
92
  }
Kyle Siefring's avatar
Kyle Siefring committed
93

Luca Barbato's avatar
Luca Barbato committed
94
  #[target_feature(enable = "ssse3")]
95
  unsafe fn sad_hbd_ssse3(
96
    plane_org: &PlaneSlice<'_, u16>, plane_ref: &PlaneSlice<'_, u16>, blk_h: usize,
Luca Barbato's avatar
Luca Barbato committed
97 98 99
    blk_w: usize, bit_depth: usize
  ) -> u32 {
    let mut sum = 0 as u32;
100 101
    let org_stride = (plane_org.plane.cfg.stride * 2) as libc::ptrdiff_t;
    let ref_stride = (plane_ref.plane.cfg.stride * 2) as libc::ptrdiff_t;
Luca Barbato's avatar
Luca Barbato committed
102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117
    assert!(blk_h >= 4 && blk_w >= 4);
    let step_size =
      blk_h.min(blk_w).min(if bit_depth <= 10 { 128 } else { 4 });
    let func = match step_size.ilog() {
      3 => rav1e_sad_4x4_hbd_ssse3,
      4 => rav1e_sad_8x8_hbd10_ssse3,
      5 => rav1e_sad_16x16_hbd_ssse3,
      6 => rav1e_sad_32x32_hbd10_ssse3,
      7 => rav1e_sad_64x64_hbd10_ssse3,
      8 => rav1e_sad_128x128_hbd10_ssse3,
      _ => rav1e_sad_128x128_hbd10_ssse3
    };
    for r in (0..blk_h).step_by(step_size) {
      for c in (0..blk_w).step_by(step_size) {
        let org_slice = plane_org.subslice(c, r);
        let ref_slice = plane_ref.subslice(c, r);
Romain Vimont's avatar
Romain Vimont committed
118 119
        let org_ptr = org_slice.as_ptr();
        let ref_ptr = ref_slice.as_ptr();
120 121 122
        // FIXME for now, T == u16
        let org_ptr = org_ptr as *const u16;
        let ref_ptr = ref_ptr as *const u16;
Luca Barbato's avatar
Luca Barbato committed
123 124
        sum += func(org_ptr, org_stride, ref_ptr, ref_stride);
      }
Kyle Siefring's avatar
Kyle Siefring committed
125
    }
126
    sum
Kyle Siefring's avatar
Kyle Siefring committed
127 128
  }

129
  #[target_feature(enable = "sse2")]
130 131
  unsafe fn sad_sse2(
    plane_org: &PlaneSlice<'_, u8>, plane_ref: &PlaneSlice<'_, u8>, blk_h: usize,
132 133 134 135
    blk_w: usize
  ) -> u32 {
    // FIXME unaligned blocks coming from hres/qres ME search
    let ptr_align_log2 = (plane_org.as_ptr() as usize).trailing_zeros() as usize;
136 137
    // The largest unaligned-safe function is for 8x8
    let ptr_align = 1 << ptr_align_log2.max(3);
138
    let mut sum = 0 as u32;
139 140
    let org_stride = plane_org.plane.cfg.stride as libc::ptrdiff_t;
    let ref_stride = plane_ref.plane.cfg.stride as libc::ptrdiff_t;
141
    assert!(blk_h >= 4 && blk_w >= 4);
142
    let step_size = blk_h.min(blk_w).min(ptr_align);
143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166
    let func = match step_size.ilog() {
      3 => rav1e_sad4x4_sse2,
      4 => rav1e_sad8x8_sse2,
      5 => rav1e_sad16x16_sse2,
      6 => rav1e_sad32x32_sse2,
      7 => rav1e_sad64x64_sse2,
      8 => rav1e_sad128x128_sse2,
      _ => rav1e_sad128x128_sse2
    };
    for r in (0..blk_h).step_by(step_size) {
      for c in (0..blk_w).step_by(step_size) {
        let org_slice = plane_org.subslice(c, r);
        let ref_slice = plane_ref.subslice(c, r);
        let org_ptr = org_slice.as_ptr();
        let ref_ptr = ref_slice.as_ptr();
        // FIXME for now, T == u8
        let org_ptr = org_ptr as *const u8;
        let ref_ptr = ref_ptr as *const u8;
        sum += func(org_ptr, org_stride, ref_ptr, ref_stride);
      }
    }
    sum
  }

Luca Barbato's avatar
Luca Barbato committed
167
  #[inline(always)]
168 169
  pub fn get_sad<T: Pixel>(
    plane_org: &PlaneSlice<'_, T>, plane_ref: &PlaneSlice<'_, T>, blk_h: usize,
Luca Barbato's avatar
Luca Barbato committed
170 171
    blk_w: usize, bit_depth: usize
  ) -> u32 {
172
    #[cfg(all(target_arch = "x86_64", feature = "nasm"))]
Luca Barbato's avatar
Luca Barbato committed
173
    {
174
      if mem::size_of::<T>() == 2 && is_x86_feature_detected!("ssse3") && blk_h >= 4 && blk_w >= 4 {
Luca Barbato's avatar
Luca Barbato committed
175
        return unsafe {
176 177
          let plane_org = &*(plane_org as *const _ as *const PlaneSlice<'_, u16>);
          let plane_ref = &*(plane_ref as *const _ as *const PlaneSlice<'_, u16>);
178
          sad_hbd_ssse3(plane_org, plane_ref, blk_h, blk_w, bit_depth)
Luca Barbato's avatar
Luca Barbato committed
179 180
        };
      }
181 182
      if mem::size_of::<T>() == 1 && is_x86_feature_detected!("sse2") && blk_h >= 4 && blk_w >= 4 {
        return unsafe {
183 184
          let plane_org = &*(plane_org as *const _ as *const PlaneSlice<'_, u8>);
          let plane_ref = &*(plane_ref as *const _ as *const PlaneSlice<'_, u8>);
185 186 187
          sad_sse2(plane_org, plane_ref, blk_h, blk_w)
        };
      }
Kyle Siefring's avatar
Kyle Siefring committed
188
    }
Luca Barbato's avatar
Luca Barbato committed
189
    super::native::get_sad(plane_org, plane_ref, blk_h, blk_w, bit_depth)
Kyle Siefring's avatar
Kyle Siefring committed
190
  }
191 192 193
}

mod native {
Raphaël Zumer's avatar
Raphaël Zumer committed
194
  use crate::plane::*;
195
  use crate::util::*;
196

Luca Barbato's avatar
Luca Barbato committed
197
  #[inline(always)]
198 199
  pub fn get_sad<T: Pixel>(
    plane_org: &PlaneSlice<'_, T>, plane_ref: &PlaneSlice<'_, T>, blk_h: usize,
Luca Barbato's avatar
Luca Barbato committed
200 201 202
    blk_w: usize, _bit_depth: usize
  ) -> u32 {
    let mut sum = 0 as u32;
203

Luca Barbato's avatar
Luca Barbato committed
204 205
    let org_iter = plane_org.iter_width(blk_w);
    let ref_iter = plane_ref.iter_width(blk_w);
206

Luca Barbato's avatar
Luca Barbato committed
207
    for (slice_org, slice_ref) in org_iter.take(blk_h).zip(ref_iter) {
208 209 210
      sum += slice_org
        .iter()
        .zip(slice_ref)
211
        .map(|(&a, &b)| (i32::cast_from(a) - i32::cast_from(b)).abs() as u32)
212
        .sum::<u32>();
Luca Barbato's avatar
Luca Barbato committed
213
    }
214

Luca Barbato's avatar
Luca Barbato committed
215 216
    sum
  }
217 218
}

219
fn get_mv_range(
220
  w_in_b: usize, h_in_b: usize, bo: &BlockOffset, blk_w: usize, blk_h: usize
221
) -> (isize, isize, isize, isize) {
222 223 224
  let border_w = 128 + blk_w as isize * 8;
  let border_h = 128 + blk_h as isize * 8;
  let mvx_min = -(bo.x as isize) * (8 * MI_SIZE) as isize - border_w;
225
  let mvx_max = (w_in_b - bo.x - blk_w / MI_SIZE) as isize * (8 * MI_SIZE) as isize + border_w;
226
  let mvy_min = -(bo.y as isize) * (8 * MI_SIZE) as isize - border_h;
227
  let mvy_max = (h_in_b - bo.y - blk_h / MI_SIZE) as isize * (8 * MI_SIZE) as isize + border_h;
228 229 230 231

  (mvx_min, mvx_max, mvy_min, mvy_max)
}

232 233
pub fn get_subset_predictors<T: Pixel>(
  fi: &FrameInvariants<T>, bo: &BlockOffset, cmv: MotionVector,
234
  frame_mvs: &[MotionVector], frame_ref_opt: &Option<Arc<ReferenceFrame<T>>>,
235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255
  ref_slot: usize
) -> (Vec<MotionVector>) {
  let mut predictors = Vec::new();

  // EPZS subset A and B predictors.

  if bo.x > 0 {
    let left = frame_mvs[bo.y * fi.w_in_b + bo.x - 1];
    predictors.push(left);
  }
  if bo.y > 0 {
    let top = frame_mvs[(bo.y - 1) * fi.w_in_b + bo.x];
    predictors.push(top);

    if bo.x < fi.w_in_b - 1 {
      let top_right = frame_mvs[(bo.y - 1) * fi.w_in_b + bo.x + 1];
      predictors.push(top_right);
    }
  }

  if predictors.len() > 0 {
Vladimir Kazakov's avatar
Vladimir Kazakov committed
256
    let mut median_mv = MotionVector::default();
257 258 259 260 261 262 263 264
    for mv in predictors.iter() {
      median_mv = median_mv + *mv;
    }
    median_mv = median_mv / (predictors.len() as i16);

    predictors.push(median_mv.quantize_to_fullpel());
  }

Vladimir Kazakov's avatar
Vladimir Kazakov committed
265
  predictors.push(MotionVector::default());
266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298

  // Coarse motion estimation.

  predictors.push(cmv.quantize_to_fullpel());

  // EPZS subset C predictors.

  if let Some(ref frame_ref) = frame_ref_opt {
    let prev_frame_mvs = &frame_ref.frame_mvs[ref_slot];

    if bo.x > 0 {
      let left = prev_frame_mvs[bo.y * fi.w_in_b + bo.x - 1];
      predictors.push(left);
    }
    if bo.y > 0 {
      let top = prev_frame_mvs[(bo.y - 1) * fi.w_in_b + bo.x];
      predictors.push(top);
    }
    if bo.x < fi.w_in_b - 1 {
      let right = prev_frame_mvs[bo.y * fi.w_in_b + bo.x + 1];
      predictors.push(right);
    }
    if bo.y < fi.h_in_b - 1 {
      let bottom = prev_frame_mvs[(bo.y + 1) * fi.w_in_b + bo.x];
      predictors.push(bottom);
    }

    predictors.push(prev_frame_mvs[bo.y * fi.w_in_b + bo.x]);
  }

  predictors
}

299
pub trait MotionEstimation {
300 301 302 303 304 305 306
  fn full_pixel_me<T: Pixel>(
    fi: &FrameInvariants<T>, fs: &FrameState<T>, rec: &Arc<ReferenceFrame<T>>, po: &PlaneOffset,
    bo: &BlockOffset, lambda: u32,
    ref_slot: usize, cmv: MotionVector, pmv: [MotionVector; 2],
    mvx_min: isize, mvx_max: isize, mvy_min: isize, mvy_max: isize,
    blk_w: usize, blk_h: usize, best_mv: &mut MotionVector,
    lowest_cost: &mut u64, ref_frame: usize
307
  );
308

309 310 311 312 313 314 315 316 317
  fn sub_pixel_me<T: Pixel>(
    fi: &FrameInvariants<T>, fs: &FrameState<T>, rec: &Arc<ReferenceFrame<T>>, po: &PlaneOffset,
    bo: &BlockOffset, lambda: u32, pmv: [MotionVector; 2],
    mvx_min: isize, mvx_max: isize, mvy_min: isize, mvy_max: isize,
    blk_w: usize, blk_h: usize, best_mv: &mut MotionVector,
    lowest_cost: &mut u64, ref_frame: usize,
    tmp_plane: Plane<T>, bsize: BlockSize
  );

318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342
  fn motion_estimation<T: Pixel> (
    fi: &FrameInvariants<T>, fs: &FrameState<T>, bsize: BlockSize,
    bo: &BlockOffset, ref_frame: usize, cmv: MotionVector,
    pmv: [MotionVector; 2], ref_slot: usize
  ) -> MotionVector {
    match fi.rec_buffer.frames[fi.ref_frames[ref_frame - LAST_FRAME] as usize]
    {
      Some(ref rec) => {
        let po = PlaneOffset {
          x: (bo.x as isize) << BLOCK_TO_PLANE_SHIFT,
          y: (bo.y as isize) << BLOCK_TO_PLANE_SHIFT
        };
        let blk_w = bsize.width();
        let blk_h = bsize.height();
        let (mvx_min, mvx_max, mvy_min, mvy_max) =
          get_mv_range(fi.w_in_b, fi.h_in_b, bo, blk_w, blk_h);

        // 0.5 is a fudge factor
        let lambda = (fi.me_lambda * 256.0 * 0.5) as u32;

        // Full-pixel motion estimation

        let mut lowest_cost = std::u64::MAX;
        let mut best_mv = MotionVector::default();

343 344 345
        Self::full_pixel_me(fi, fs, rec, &po, bo, lambda, ref_slot, cmv, pmv,
                           mvx_min, mvx_max, mvy_min, mvy_max, blk_w, blk_h,
                           &mut best_mv, &mut lowest_cost, ref_frame);
346

347 348 349 350 351
        let tmp_plane = Plane::new(blk_w, blk_h, 0, 0, 0, 0);
        Self::sub_pixel_me(fi, fs, rec, &po, bo, lambda, pmv,
                           mvx_min, mvx_max, mvy_min, mvy_max, blk_w, blk_h,
                           &mut best_mv, &mut lowest_cost, ref_frame,
                           tmp_plane, bsize);
352 353

        best_mv
Frank Bossen's avatar
Frank Bossen committed
354 355
      }

356
      None => MotionVector::default()
357
    }
358 359
  }
}
360

361 362 363
pub struct DiamondSearch {}
pub struct FullSearch {}

364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397
impl MotionEstimation for DiamondSearch {
  fn full_pixel_me<T: Pixel>(
    fi: &FrameInvariants<T>, fs: &FrameState<T>, rec: &Arc<ReferenceFrame<T>>,
    po: &PlaneOffset, bo: &BlockOffset, lambda: u32, ref_slot: usize,
    cmv: MotionVector, pmv: [MotionVector; 2], mvx_min: isize, mvx_max: isize,
    mvy_min: isize, mvy_max: isize, blk_w: usize, blk_h: usize,
    best_mv: &mut MotionVector, lowest_cost: &mut u64, ref_frame: usize
  ) {
    let frame_mvs = &fs.frame_mvs[ref_slot];
    let frame_ref = &fi.rec_buffer.frames[fi.ref_frames[0] as usize];
    let predictors =
      get_subset_predictors(fi, bo, cmv, frame_mvs, frame_ref, ref_slot);

    diamond_me_search(
      fi,
      &po,
      &fs.input.planes[0],
      &rec.frame.planes[0],
      &predictors,
      fi.sequence.bit_depth,
      pmv,
      lambda,
      mvx_min,
      mvx_max,
      mvy_min,
      mvy_max,
      blk_w,
      blk_h,
      best_mv,
      lowest_cost,
      &mut None,
      ref_frame
    );
  }
398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429

  fn sub_pixel_me<T: Pixel>(
    fi: &FrameInvariants<T>, fs: &FrameState<T>, rec: &Arc<ReferenceFrame<T>>,
    po: &PlaneOffset, _bo: &BlockOffset, lambda: u32,
    pmv: [MotionVector; 2], mvx_min: isize, mvx_max: isize,
    mvy_min: isize, mvy_max: isize, blk_w: usize, blk_h: usize,
    best_mv: &mut MotionVector, lowest_cost: &mut u64, ref_frame: usize,
    tmp_plane: Plane<T>, _bsize: BlockSize
  )
  {
    let predictors = vec![*best_mv];
    diamond_me_search(
      fi,
      &po,
      &fs.input.planes[0],
      &rec.frame.planes[0],
      &predictors,
      fi.sequence.bit_depth,
      pmv,
      lambda,
      mvx_min,
      mvx_max,
      mvy_min,
      mvy_max,
      blk_w,
      blk_h,
      best_mv,
      lowest_cost,
      &mut Some(tmp_plane),
      ref_frame
    );
  }
430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468
}

impl MotionEstimation for FullSearch {
  fn full_pixel_me<T: Pixel>(
    fi: &FrameInvariants<T>, fs: &FrameState<T>, rec: &Arc<ReferenceFrame<T>>,
    po: &PlaneOffset, _bo: &BlockOffset, lambda: u32, _ref_slot: usize,
    cmv: MotionVector, pmv: [MotionVector; 2], mvx_min: isize, mvx_max: isize,
    mvy_min: isize, mvy_max: isize, blk_w: usize, blk_h: usize,
    best_mv: &mut MotionVector, lowest_cost: &mut u64, _ref_frame: usize
  ) {
    let range = 16;
    let x_lo = po.x
      + ((-range + (cmv.col / 8) as isize).max(mvx_min / 8).min(mvx_max / 8));
    let x_hi = po.x
      + ((range + (cmv.col / 8) as isize).max(mvx_min / 8).min(mvx_max / 8));
    let y_lo = po.y
      + ((-range + (cmv.row / 8) as isize).max(mvy_min / 8).min(mvy_max / 8));
    let y_hi = po.y
      + ((range + (cmv.row / 8) as isize).max(mvy_min / 8).min(mvy_max / 8));

    full_search(
      x_lo,
      x_hi,
      y_lo,
      y_hi,
      blk_h,
      blk_w,
      &fs.input.planes[0],
      &rec.frame.planes[0],
      best_mv,
      lowest_cost,
      &po,
      2,
      fi.sequence.bit_depth,
      lambda,
      pmv,
      fi.allow_high_precision_mv
    );
  }
469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495

  fn sub_pixel_me<T: Pixel>(
    fi: &FrameInvariants<T>, fs: &FrameState<T>, _rec: &Arc<ReferenceFrame<T>>,
    po: &PlaneOffset, _bo: &BlockOffset, lambda: u32,
    pmv: [MotionVector; 2], mvx_min: isize, mvx_max: isize,
    mvy_min: isize, mvy_max: isize, _blk_w: usize, _blk_h: usize,
    best_mv: &mut MotionVector, lowest_cost: &mut u64, ref_frame: usize,
    mut tmp_plane: Plane<T>, bsize: BlockSize
  )
  {
    telescopic_subpel_search(
      fi,
      fs,
      bsize,
      &po,
      lambda,
      ref_frame,
      pmv,
      mvx_min,
      mvx_max,
      mvy_min,
      mvy_max,
      &mut tmp_plane,
      best_mv,
      lowest_cost
    );
  }
496
}
497

498 499 500
fn get_best_predictor<T: Pixel>(
  fi: &FrameInvariants<T>,
  po: &PlaneOffset, p_org: &Plane<T>, p_ref: &Plane<T>,
501
  predictors: &[MotionVector],
502
  bit_depth: usize, pmv: [MotionVector; 2], lambda: u32,
503 504
  mvx_min: isize, mvx_max: isize, mvy_min: isize, mvy_max: isize,
  blk_w: usize, blk_h: usize,
505 506
  center_mv: &mut MotionVector, center_mv_cost: &mut u64,
  tmp_plane_opt: &mut Option<Plane<T>>, ref_frame: usize) {
Vladimir Kazakov's avatar
Vladimir Kazakov committed
507
  *center_mv = MotionVector::default();
508 509 510 511 512 513
  *center_mv_cost = std::u64::MAX;

  for &init_mv in predictors.iter() {
    let cost = get_mv_rd_cost(
      fi, po, p_org, p_ref, bit_depth,
      pmv, lambda, mvx_min, mvx_max, mvy_min, mvy_max,
514
      blk_w, blk_h, init_mv, tmp_plane_opt, ref_frame);
515 516 517 518 519 520 521 522

    if cost < *center_mv_cost {
      *center_mv = init_mv;
      *center_mv_cost = cost;
    }
  }
}

523 524 525
fn diamond_me_search<T: Pixel>(
  fi: &FrameInvariants<T>,
  po: &PlaneOffset, p_org: &Plane<T>, p_ref: &Plane<T>,
526
  predictors: &[MotionVector],
527
  bit_depth: usize, pmv: [MotionVector; 2], lambda: u32,
528 529
  mvx_min: isize, mvx_max: isize, mvy_min: isize, mvy_max: isize,
  blk_w: usize, blk_h: usize,
530 531
  center_mv: &mut MotionVector, center_mv_cost: &mut u64,
  tmp_plane_opt: &mut Option<Plane<T>>, ref_frame: usize)
532 533
{
  let diamond_pattern = [(1i16, 0i16), (0, 1), (-1, 0), (0, -1)];
534 535 536 537 538 539 540 541 542
  let (mut diamond_radius, diamond_radius_end) = {
    if tmp_plane_opt.is_some() {
      // Sub-pixel motion estimation
      (4i16, if fi.allow_high_precision_mv {1i16} else {2i16})
    } else {
      // Full pixel motion estimation
      (16i16, 8i16)
    }
  };
543 544 545 546

  get_best_predictor(
    fi, po, p_org, p_ref, &predictors,
    bit_depth, pmv, lambda, mvx_min, mvx_max, mvy_min, mvy_max,
547 548
    blk_w, blk_h, center_mv, center_mv_cost,
    tmp_plane_opt, ref_frame);
549 550 551

  loop {
    let mut best_diamond_rd_cost = std::u64::MAX;
Vladimir Kazakov's avatar
Vladimir Kazakov committed
552
    let mut best_diamond_mv = MotionVector::default();
553 554 555 556 557 558 559 560 561 562 563

    for p in diamond_pattern.iter() {

        let cand_mv = MotionVector {
          row: center_mv.row + diamond_radius * p.0,
          col: center_mv.col + diamond_radius * p.1
        };

        let rd_cost = get_mv_rd_cost(
          fi, &po, p_org, p_ref, bit_depth,
          pmv, lambda, mvx_min, mvx_max, mvy_min, mvy_max,
564
          blk_w, blk_h, cand_mv, tmp_plane_opt, ref_frame);
565 566 567 568 569 570 571 572

        if rd_cost < best_diamond_rd_cost {
          best_diamond_rd_cost = rd_cost;
          best_diamond_mv = cand_mv;
        }
    }

    if *center_mv_cost <= best_diamond_rd_cost {
573
      if diamond_radius == diamond_radius_end {
574 575 576 577 578 579 580 581 582 583 584 585 586 587
        break;
      } else {
        diamond_radius /= 2;
      }
    }
    else {
      *center_mv = best_diamond_mv;
      *center_mv_cost = best_diamond_rd_cost;
    }
  }

  assert!(*center_mv_cost < std::u64::MAX);
}

588 589 590
fn get_mv_rd_cost<T: Pixel>(
  fi: &FrameInvariants<T>,
  po: &PlaneOffset, p_org: &Plane<T>, p_ref: &Plane<T>, bit_depth: usize,
591
  pmv: [MotionVector; 2], lambda: u32,
592 593
  mvx_min: isize, mvx_max: isize, mvy_min: isize, mvy_max: isize,
  blk_w: usize, blk_h: usize,
594 595
  cand_mv: MotionVector, tmp_plane_opt: &mut Option<Plane<T>>,
  ref_frame: usize) -> u64
596 597 598 599 600 601 602 603 604 605
{
  if (cand_mv.col as isize) < mvx_min || (cand_mv.col as isize) > mvx_max {
    return std::u64::MAX;
  }
  if (cand_mv.row as isize) < mvy_min || (cand_mv.row as isize) > mvy_max {
    return std::u64::MAX;
  }

  let plane_org = p_org.slice(po);

606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642
  if let Some(ref mut tmp_plane) = tmp_plane_opt {
    let mut tmp_slice = &mut tmp_plane.mut_slice(&PlaneOffset { x: 0, y: 0 });
    PredictionMode::NEWMV.predict_inter(
      fi,
      0,
      &po,
      &mut tmp_slice,
      blk_w,
      blk_h,
      [ref_frame, NONE_FRAME],
      [cand_mv, MotionVector { row: 0, col: 0 }]
    );
    let plane_ref = tmp_plane.slice(&PlaneOffset { x: 0, y: 0 });
    compute_mv_rd_cost(
      fi, pmv, lambda, bit_depth, blk_w, blk_h, cand_mv,
      &plane_org, &plane_ref
    )
  } else {
    // Full pixel motion vector
    let plane_ref = p_ref.slice(&PlaneOffset {
      x: po.x + (cand_mv.col / 8) as isize,
      y: po.y + (cand_mv.row / 8) as isize
    });
    compute_mv_rd_cost(
      fi, pmv, lambda, bit_depth, blk_w, blk_h, cand_mv,
      &plane_org, &plane_ref
    )
  }
}

fn compute_mv_rd_cost<T: Pixel>(
  fi: &FrameInvariants<T>,
  pmv: [MotionVector; 2], lambda: u32,
  bit_depth: usize, blk_w: usize, blk_h: usize, cand_mv: MotionVector,
  plane_org: &PlaneSlice<T>, plane_ref: &PlaneSlice<T>
) -> u64
{
643 644 645 646 647 648 649 650 651
  let sad = get_sad(&plane_org, &plane_ref, blk_h, blk_w, bit_depth);

  let rate1 = get_mv_rate(cand_mv, pmv[0], fi.allow_high_precision_mv);
  let rate2 = get_mv_rate(cand_mv, pmv[1], fi.allow_high_precision_mv);
  let rate = rate1.min(rate2 + 1);

  256 * sad as u64 + rate as u64 * lambda as u64
}

652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723
fn telescopic_subpel_search<T: Pixel>(
  fi: &FrameInvariants<T>, fs: &FrameState<T>, bsize: BlockSize, po: &PlaneOffset,
  lambda: u32, ref_frame: usize, pmv: [MotionVector; 2],
  mvx_min: isize, mvx_max: isize, mvy_min: isize, mvy_max: isize,
  tmp_plane: &mut Plane<T>, best_mv: &mut MotionVector, lowest_cost: &mut u64
) {
  let blk_w = bsize.width();
  let blk_h = bsize.height();

  let mode = PredictionMode::NEWMV;

  let mut steps = vec![8, 4, 2];
  if fi.allow_high_precision_mv {
    steps.push(1);
  }

  for step in steps {
    let center_mv_h = *best_mv;
    for i in 0..3 {
      for j in 0..3 {
        // Skip the center point that was already tested
        if i == 1 && j == 1 {
          continue;
        }

        let cand_mv = MotionVector {
          row: center_mv_h.row + step * (i as i16 - 1),
          col: center_mv_h.col + step * (j as i16 - 1)
        };

        if (cand_mv.col as isize) < mvx_min || (cand_mv.col as isize) > mvx_max {
          continue;
        }
        if (cand_mv.row as isize) < mvy_min || (cand_mv.row as isize) > mvy_max {
          continue;
        }

        {
          let tmp_slice =
            &mut tmp_plane.mut_slice(&PlaneOffset { x: 0, y: 0 });

          mode.predict_inter(
            fi,
            0,
            &po,
            tmp_slice,
            blk_w,
            blk_h,
            [ref_frame, NONE_FRAME],
            [cand_mv, MotionVector { row: 0, col: 0 }]
          );
        }

        let plane_org = fs.input.planes[0].slice(&po);
        let plane_ref = tmp_plane.slice(&PlaneOffset { x: 0, y: 0 });

        let sad = get_sad(&plane_org, &plane_ref, blk_h, blk_w, fi.sequence.bit_depth);

        let rate1 = get_mv_rate(cand_mv, pmv[0], fi.allow_high_precision_mv);
        let rate2 = get_mv_rate(cand_mv, pmv[1], fi.allow_high_precision_mv);
        let rate = rate1.min(rate2 + 1);
        let cost = 256 * sad as u64 + rate as u64 * lambda as u64;

        if cost < *lowest_cost {
          *lowest_cost = cost;
          *best_mv = cand_mv;
        }
      }
    }
  }
}

724
fn full_search<T: Pixel>(
Kyle Siefring's avatar
Kyle Siefring committed
725
  x_lo: isize, x_hi: isize, y_lo: isize, y_hi: isize, blk_h: usize,
726
  blk_w: usize, p_org: &Plane<T>, p_ref: &Plane<T>, best_mv: &mut MotionVector,
727
  lowest_cost: &mut u64, po: &PlaneOffset, step: usize, bit_depth: usize,
728
  lambda: u32, pmv: [MotionVector; 2], allow_high_precision_mv: bool
Kyle Siefring's avatar
Kyle Siefring committed
729
) {
Luca Barbato's avatar
Luca Barbato committed
730 731 732 733 734
    let search_range_y = (y_lo..=y_hi).step_by(step);
    let search_range_x = (x_lo..=x_hi).step_by(step);
    let search_area = search_range_y.flat_map(|y| { search_range_x.clone().map(move |x| (y, x)) });

    let (cost, mv) = search_area.map(|(y, x)| {
735 736 737
      let plane_org = p_org.slice(po);
      let plane_ref = p_ref.slice(&PlaneOffset { x, y });

Kyle Siefring's avatar
Kyle Siefring committed
738
      let sad = get_sad(&plane_org, &plane_ref, blk_h, blk_w, bit_depth);
739

Frank Bossen's avatar
Frank Bossen committed
740 741 742 743 744 745 746 747
      let mv = MotionVector {
        row: 8 * (y as i16 - po.y as i16),
        col: 8 * (x as i16 - po.x as i16)
      };

      let rate1 = get_mv_rate(mv, pmv[0], allow_high_precision_mv);
      let rate2 = get_mv_rate(mv, pmv[1], allow_high_precision_mv);
      let rate = rate1.min(rate2 + 1);
748
      let cost = 256 * sad as u64 + rate as u64 * lambda as u64;
Frank Bossen's avatar
Frank Bossen committed
749

Luca Barbato's avatar
Luca Barbato committed
750 751 752 753 754
      (cost, mv)
  }).min_by_key(|(c, _)| *c).unwrap();

    *lowest_cost = cost;
    *best_mv = mv;
755 756 757
}

// Adjust block offset such that entire block lies within frame boundaries
758
fn adjust_bo<T: Pixel>(bo: &BlockOffset, fi: &FrameInvariants<T>, blk_w: usize, blk_h: usize) -> BlockOffset {
759 760 761 762 763 764
  BlockOffset {
    x: (bo.x as isize).min(fi.w_in_b as isize - blk_w as isize / 4).max(0) as usize,
    y: (bo.y as isize).min(fi.h_in_b as isize - blk_h as isize / 4).max(0) as usize
  }
}

765
#[inline(always)]
Frank Bossen's avatar
Frank Bossen committed
766
fn get_mv_rate(a: MotionVector, b: MotionVector, allow_high_precision_mv: bool) -> u32 {
767
  #[inline(always)]
Frank Bossen's avatar
Frank Bossen committed
768 769 770 771 772 773 774 775 776 777 778 779
  fn diff_to_rate(diff: i16, allow_high_precision_mv: bool) -> u32 {
    let d = if allow_high_precision_mv { diff } else { diff >> 1 };
    if d == 0 {
      0
    } else {
      2 * (16 - d.abs().leading_zeros())
    }
  }

  diff_to_rate(a.row - b.row, allow_high_precision_mv) + diff_to_rate(a.col - b.col, allow_high_precision_mv)
}

780 781
pub fn estimate_motion_ss4<T: Pixel>(
  fi: &FrameInvariants<T>, fs: &FrameState<T>, bsize: BlockSize, ref_idx: usize,
782
  bo: &BlockOffset
783 784 785 786 787 788 789 790 791
) -> Option<MotionVector> {
  if let Some(ref rec) = fi.rec_buffer.frames[ref_idx] {
    let blk_w = bsize.width();
    let blk_h = bsize.height();
    let bo_adj = adjust_bo(bo, fi, blk_w, blk_h);
    let po = PlaneOffset {
      x: (bo_adj.x as isize) << BLOCK_TO_PLANE_SHIFT >> 2,
      y: (bo_adj.y as isize) << BLOCK_TO_PLANE_SHIFT >> 2
    };
792

793 794
    let range_x = 192 * fi.me_range_scale as isize;
    let range_y = 64 * fi.me_range_scale as isize;
795
    let (mvx_min, mvx_max, mvy_min, mvy_max) = get_mv_range(fi.w_in_b, fi.h_in_b, &bo_adj, blk_w, blk_h);
796 797 798 799
    let x_lo = po.x + (((-range_x).max(mvx_min / 8)) >> 2);
    let x_hi = po.x + (((range_x).min(mvx_max / 8)) >> 2);
    let y_lo = po.y + (((-range_y).max(mvy_min / 8)) >> 2);
    let y_hi = po.y + (((range_y).min(mvy_max / 8)) >> 2);
800

801
    let mut lowest_cost = std::u64::MAX;
Vladimir Kazakov's avatar
Vladimir Kazakov committed
802
    let mut best_mv = MotionVector::default();
803

Frank Bossen's avatar
Frank Bossen committed
804
    // Divide by 16 to account for subsampling, 0.125 is a fudge factor
805
    let lambda = (fi.me_lambda * 256.0 / 16.0 * 0.125) as u32;
Frank Bossen's avatar
Frank Bossen committed
806

807
    full_search(
Kyle Siefring's avatar
Kyle Siefring committed
808 809 810 811 812 813 814 815 816
      x_lo,
      x_hi,
      y_lo,
      y_hi,
      blk_h >> 2,
      blk_w >> 2,
      &fs.input_qres,
      &rec.input_qres,
      &mut best_mv,
Frank Bossen's avatar
Frank Bossen committed
817
      &mut lowest_cost,
Kyle Siefring's avatar
Kyle Siefring committed
818 819
      &po,
      1,
820
      fi.sequence.bit_depth,
Frank Bossen's avatar
Frank Bossen committed
821
      lambda,
Vladimir Kazakov's avatar
Vladimir Kazakov committed
822
      [MotionVector::default(); 2],
Frank Bossen's avatar
Frank Bossen committed
823
      fi.allow_high_precision_mv
824 825 826 827 828 829 830 831
    );

    Some(MotionVector { row: best_mv.row * 4, col: best_mv.col * 4 })
  } else {
    None
  }
}

832 833
pub fn estimate_motion_ss2<T: Pixel>(
  fi: &FrameInvariants<T>, fs: &FrameState<T>, bsize: BlockSize, ref_idx: usize,
834
  bo: &BlockOffset, pmvs: &[Option<MotionVector>; 3]
835 836 837 838 839 840 841 842 843 844
) -> Option<MotionVector> {
  if let Some(ref rec) = fi.rec_buffer.frames[ref_idx] {
    let blk_w = bsize.width();
    let blk_h = bsize.height();
    let bo_adj = adjust_bo(bo, fi, blk_w, blk_h);
    let po = PlaneOffset {
      x: (bo_adj.x as isize) << BLOCK_TO_PLANE_SHIFT >> 1,
      y: (bo_adj.y as isize) << BLOCK_TO_PLANE_SHIFT >> 1
    };
    let range = 16;
845
    let (mvx_min, mvx_max, mvy_min, mvy_max) = get_mv_range(fi.w_in_b, fi.h_in_b, &bo_adj, blk_w, blk_h);
846

847
    let mut lowest_cost = std::u64::MAX;
Vladimir Kazakov's avatar
Vladimir Kazakov committed
848
    let mut best_mv = MotionVector::default();
849

Frank Bossen's avatar
Frank Bossen committed
850
    // Divide by 4 to account for subsampling, 0.125 is a fudge factor
851
    let lambda = (fi.me_lambda * 256.0 / 4.0 * 0.125) as u32;
Frank Bossen's avatar
Frank Bossen committed
852

853 854
    for omv in pmvs.iter() {
      if let Some(pmv) = omv {
855 856 857 858
        let x_lo = po.x + (((pmv.col as isize / 8 - range).max(mvx_min / 8).min(mvx_max / 8)) >> 1);
        let x_hi = po.x + (((pmv.col as isize / 8 + range).max(mvx_min / 8).min(mvx_max / 8)) >> 1);
        let y_lo = po.y + (((pmv.row as isize / 8 - range).max(mvy_min / 8).min(mvy_max / 8)) >> 1);
        let y_hi = po.y + (((pmv.row as isize / 8 + range).max(mvy_min / 8).min(mvy_max / 8)) >> 1);
859 860

        full_search(
Kyle Siefring's avatar
Kyle Siefring committed
861 862 863 864 865 866 867 868 869
          x_lo,
          x_hi,
          y_lo,
          y_hi,
          blk_h >> 1,
          blk_w >> 1,
          &fs.input_hres,
          &rec.input_hres,
          &mut best_mv,
Frank Bossen's avatar
Frank Bossen committed
870
          &mut lowest_cost,
Kyle Siefring's avatar
Kyle Siefring committed
871 872
          &po,
          1,
873
          fi.sequence.bit_depth,
Frank Bossen's avatar
Frank Bossen committed
874
          lambda,
Vladimir Kazakov's avatar
Vladimir Kazakov committed
875
          [MotionVector::default(); 2],
Frank Bossen's avatar
Frank Bossen committed
876
          fi.allow_high_precision_mv
877 878 879 880 881 882 883 884 885 886
        );
      }
    }

    Some(MotionVector { row: best_mv.row * 2, col: best_mv.col * 2 })
  } else {
    None
  }
}

887 888 889
#[cfg(test)]
pub mod test {
  use super::*;
Raphaël Zumer's avatar
Raphaël Zumer committed
890 891
  use crate::partition::BlockSize;
  use crate::partition::BlockSize::*;
892 893

  // Generate plane data for get_sad_same()
894
  fn setup_sad<T: Pixel>() -> (Plane<T>, Plane<T>) {
895 896
    let mut input_plane = Plane::new(640, 480, 0, 0, 128 + 8, 128 + 8);
    let mut rec_plane = input_plane.clone();
897 898
    // Make the test pattern robust to data alignment
    let xpad_off = (input_plane.cfg.xorigin - input_plane.cfg.xpad) as i32 - 8i32;
Luca Barbato's avatar
Luca Barbato committed
899

900
    for (i, row) in input_plane.data.chunks_mut(input_plane.cfg.stride).enumerate() {
901
      for (j, pixel) in row.into_iter().enumerate() {
902
        let val = (j + i) as i32 - xpad_off & 255i32;
Luca Barbato's avatar
Luca Barbato committed
903
        assert!(val >= u8::min_value().into() &&
904
            val <= u8::max_value().into());
905
        *pixel = T::cast_from(val);
906 907 908 909
      }
    }

    for (i, row) in rec_plane.data.chunks_mut(rec_plane.cfg.stride).enumerate() {
910
      for (j, pixel) in row.into_iter().enumerate() {
911
        let val = j as i32 - i as i32 - xpad_off & 255i32;
Luca Barbato's avatar
Luca Barbato committed
912
        assert!(val >= u8::min_value().into() &&
913
            val <= u8::max_value().into());
914
        *pixel = T::cast_from(val);
915 916 917 918 919 920 921
      }
    }

    (input_plane, rec_plane)
  }

  // Regression and validation test for SAD computation
922
  fn get_sad_same_inner<T: Pixel>() {
923
    let blocks: Vec<(BlockSize, u32)> = vec![
924
      (BLOCK_4X4, 1912),
925 926
      (BLOCK_4X8, 4296),
      (BLOCK_8X4, 3496),
927
      (BLOCK_8X8, 7824),
928 929
      (BLOCK_8X16, 16592),
      (BLOCK_16X8, 14416),
930
      (BLOCK_16X16, 31136),
931 932
      (BLOCK_16X32, 60064),
      (BLOCK_32X16, 59552),
933
      (BLOCK_32X32, 120128),
934 935
      (BLOCK_32X64, 186688),
      (BLOCK_64X32, 250176),
936
      (BLOCK_64X64, 438912),
937 938
      (BLOCK_64X128, 654272),
      (BLOCK_128X64, 1016768),
939
      (BLOCK_128X128, 1689792),
940 941 942 943 944 945
      (BLOCK_4X16, 8680),
      (BLOCK_16X4, 6664),
      (BLOCK_8X32, 31056),
      (BLOCK_32X8, 27600),
      (BLOCK_16X64, 93344),
      (BLOCK_64X16, 116384),
946 947
    ];

Kyle Siefring's avatar
Kyle Siefring committed
948
    let bit_depth: usize = 8;
949
    let (input_plane, rec_plane) = setup_sad::<T>();
950 951

    for block in blocks {
Kyle Siefring's avatar
Kyle Siefring committed
952 953
      let bsw = block.0.width();
      let bsh = block.0.height();
954
      let po = PlaneOffset { x: 32, y: 40 };
955

Kyle Siefring's avatar
Kyle Siefring committed
956 957
      let mut input_slice = input_plane.slice(&po);
      let mut rec_slice = rec_plane.slice(&po);
958

Kyle Siefring's avatar
Kyle Siefring committed
959 960
      assert_eq!(
        block.1,
961
        get_sad(&mut input_slice, &mut rec_slice, bsh, bsw, bit_depth)
Kyle Siefring's avatar
Kyle Siefring committed
962
      );
963 964
    }
  }
965 966 967 968 969 970 971 972 973 974

  #[test]
  fn get_sad_same_u8() {
    get_sad_same_inner::<u8>();
  }

  #[test]
  fn get_sad_same_u16() {
    get_sad_same_inner::<u16>();
  }
975
}