From 83fcc030563ed13373d8fb559de0993905682458 Mon Sep 17 00:00:00 2001
From: fbossen <frank@bossentech.com>
Date: Thu, 23 Aug 2018 19:12:20 -0230
Subject: [PATCH] Add basic support for nonzero motion vectors (#485)
* Add motion vector paramater to inter prediction function
Motion vector is read from added data field in Block structure
* Add reference frame and motion vector fields/parameters
Reference frame and motion vector are added fields to the
RDOOutput structure and added parameters to the block
encoding function
* Fix inter prediction function for nonzero MVs
* Add call to encode new motion vector
* Import default cdfs for MV coding from libaom
* Fix handling of MV precision in MV coding function
* Add coding of DRL mode
* Add motion vector stack
* Add new MV counter to correctly determine context
* Use motion vectors in MV stack for prediction
* Sort MV stack according to weights
Also update weights after near search
* Fix log2() function
* Compute correct context for DRL flag
* Store MVs with 1/8 pel precision and fix MV generation
* Add some basic motion estimation
Motion estimation is full search (+/-16 pel range in 2-pel increments) and
is done independently for each block
* Fix chroma motion compensation for small blocks
Chroma motion compensation sometimes happens on 2x2 basis
---
src/context.rs | 239 +++++++++++++++++++++++++++++++----------------
src/encoder.rs | 90 ++++++++++++++++--
src/lib.rs | 1 +
src/me.rs | 72 ++++++++++++++
src/partition.rs | 37 +++++---
src/predict.rs | 1 +
src/rdo.rs | 19 +++-
7 files changed, 358 insertions(+), 101 deletions(-)
create mode 100644 src/me.rs
diff --git a/src/context.rs b/src/context.rs
index 9c70e162..061c8d25 100755
--- a/src/context.rs
+++ b/src/context.rs
@@ -64,6 +64,9 @@ const TX_SETS: usize = 9;
const TX_SETS_INTRA: usize = 3;
const TX_SETS_INTER: usize = 4;
+const MAX_REF_MV_STACK_SIZE: usize = 8;
+pub const REF_CAT_LEVEL: u32 = 640;
+
// Number of transform types in each set type
static num_tx_set: [usize; TX_SETS] =
[1, 2, 5, 7, 7, 10, 12, 16, 16];
@@ -377,6 +380,7 @@ const PLANE_TYPES: usize = 2;
const REF_TYPES: usize = 2;
const SKIP_CONTEXTS: usize = 3;
const INTRA_INTER_CONTEXTS: usize = 4;
+const DRL_MODE_CONTEXTS: usize = 3;
// Level Map
const TXB_SKIP_CONTEXTS: usize = 13;
@@ -758,6 +762,27 @@ pub fn uv_intra_mode_to_tx_type_context(pred: PredictionMode) -> TxType {
intra_mode_to_tx_type_context[uv2y[pred as usize] as usize]
}
+#[derive(Clone,Copy)]
+#[repr(C)]
+pub struct NMVComponent {
+ classes_cdf: [u16; MV_CLASSES + 1],
+ class0_fp_cdf: [[u16; MV_FP_SIZE + 1]; CLASS0_SIZE],
+ fp_cdf: [u16; MV_FP_SIZE + 1],
+ sign_cdf: [u16; 2 + 1],
+ class0_hp_cdf: [u16; 2 + 1],
+ hp_cdf: [u16; 2 + 1],
+ class0_cdf: [u16; CLASS0_SIZE + 1],
+ bits_cdf: [[u16; 2 + 1]; MV_OFFSET_BITS],
+}
+
+#[derive(Clone,Copy)]
+#[repr(C)]
+pub struct NMVContext {
+ joints_cdf: [u16; MV_JOINTS + 1],
+ comps: [NMVComponent; 2],
+}
+
+
extern "C" {
static default_partition_cdf:
[[u16; EXT_PARTITION_TYPES + 1]; PARTITION_CONTEXTS];
@@ -804,6 +829,9 @@ extern "C" {
static av1_default_coeff_lps_multi_cdfs: [[[[[u16; BR_CDF_SIZE + 1];
LEVEL_CONTEXTS]; PLANE_TYPES];
TxSize::TX_SIZES]; 4];
+
+ static default_nmv_context: NMVContext;
+ static default_drl_cdf:[[u16; 2 + 1]; DRL_MODE_CONTEXTS];
}
#[repr(C)]
@@ -814,6 +842,13 @@ pub struct SCAN_ORDER {
pub neighbors: &'static [u16; ((32 * 32) + 1) * 2]
}
+#[derive(Clone)]
+pub struct CandidateMV {
+ pub this_mv: MotionVector,
+ pub comp_mv: MotionVector,
+ pub weight: u32
+}
+
#[derive(Clone)]
pub struct CDFContext {
partition_cdf: [[u16; EXT_PARTITION_TYPES + 1]; PARTITION_CONTEXTS],
@@ -831,6 +866,8 @@ pub struct CDFContext {
angle_delta_cdf: [[u16; 2 * MAX_ANGLE_DELTA + 1 + 1]; DIRECTIONAL_MODES],
filter_intra_cdfs: [[u16; 3]; BlockSize::BLOCK_SIZES_ALL],
single_ref_cdfs: [[[u16; 2 + 1]; SINGLE_REFS - 1]; REF_CONTEXTS],
+ drl_cdfs: [[u16; 2 + 1]; DRL_MODE_CONTEXTS],
+ nmv_context: NMVContext,
// lv_map
txb_skip_cdf: [[[u16; 3]; TXB_SKIP_CONTEXTS]; TxSize::TX_SIZES],
@@ -877,6 +914,8 @@ impl CDFContext {
angle_delta_cdf: default_angle_delta_cdf,
filter_intra_cdfs: default_filter_intra_cdfs,
single_ref_cdfs: default_single_ref_cdf,
+ drl_cdfs: default_drl_cdf,
+ nmv_context: default_nmv_context,
// lv_map
txb_skip_cdf: av1_default_txb_skip_cdfs[qctx],
@@ -1030,7 +1069,7 @@ mod test {
const SUPERBLOCK_TO_PLANE_SHIFT: usize = MAX_SB_SIZE_LOG2;
const SUPERBLOCK_TO_BLOCK_SHIFT: usize = MAX_MIB_SIZE_LOG2;
-const BLOCK_TO_PLANE_SHIFT: usize = MI_SIZE_LOG2;
+pub const BLOCK_TO_PLANE_SHIFT: usize = MI_SIZE_LOG2;
pub const LOCAL_BLOCK_MASK: usize = (1 << SUPERBLOCK_TO_BLOCK_SHIFT) - 1;
/// Absolute offset in superblocks inside a plane, where a superblock is defined
@@ -1110,6 +1149,7 @@ pub struct Block {
pub partition: PartitionType,
pub skip: bool,
pub ref_frames: [usize; 2],
+ pub mv: [MotionVector; 2],
pub neighbors_ref_counts: [usize; TOTAL_REFS_PER_FRAME],
pub cdef_index: u8,
pub n4_w: usize, /* block width in the unit of mode_info */
@@ -1126,6 +1166,7 @@ impl Block {
partition: PartitionType::PARTITION_NONE,
skip: false,
ref_frames: [INTRA_FRAME; 2],
+ mv: [ MotionVector { row:0, col: 0 }; 2],
neighbors_ref_counts: [0; TOTAL_REFS_PER_FRAME],
cdef_index: 0,
n4_w: BLOCK_64X64.width_mi(),
@@ -1415,6 +1456,17 @@ impl BlockContext {
}
}
+ pub fn set_motion_vector(&mut self, bo: &BlockOffset, bsize: BlockSize, mv: MotionVector) {
+ let bw = bsize.width_mi();
+ let bh = bsize.height_mi();
+
+ for y in 0..bh {
+ for x in 0..bw {
+ self.blocks[bo.y + y as usize][bo.x + x as usize].mv[0] = mv;
+ }
+ }
+ }
+
pub fn set_cdef(&mut self, sbo: &SuperBlockOffset, cdef_index: u8) {
let bo = sbo.block_offset(0, 0);
// Checkme: Is 16 still the right block unit for 128x128 superblocks?
@@ -1874,31 +1926,48 @@ impl ContextWriter {
cmp::max(col_offset, -(mi_col as isize))
}
- fn add_ref_mv_candidate(&self, blk: &Block) -> bool {
+ fn find_matching_mv(&self, blk: &Block, mv_stack: &mut Vec<CandidateMV>, weight: u32) -> bool {
+ for mut mv_cand in mv_stack {
+ if blk.mv[0].row == mv_cand.this_mv.row && blk.mv[0].col == mv_cand.this_mv.col {
+ mv_cand.weight += weight;
+ return true;
+ }
+ }
+ false
+ }
+
+ fn add_ref_mv_candidate(&self, ref_frame: usize, blk: &Block, mv_stack: &mut Vec<CandidateMV>,
+ weight: u32, newmv_count: &mut usize) -> bool {
if !blk.is_inter() { /* For intrabc */
return false;
}
-/*
- let index = 0;
+ if blk.ref_frames[0] == ref_frame {
+ let found_match = self.find_matching_mv(blk, mv_stack, weight);
- if rf[1] == NONE_FRAME {
- for i in 0..2 {
- if cand.ref_frame[i] == rf[0] {
+ if !found_match && mv_stack.len() < MAX_REF_MV_STACK_SIZE {
+ let mv_cand = CandidateMV {
+ this_mv: blk.mv[0],
+ comp_mv: blk.mv[1],
+ weight: weight
+ };
- }
+ mv_stack.push(mv_cand);
}
- } else {
- if cand.ref_frame[0] == rf[0] && cand.ref_frame[1] == rf[1] {
+ if blk.mode == PredictionMode::NEWMV {
+ *newmv_count += 1;
}
+
+ true
+ } else {
+ false
}
-*/
- true
}
fn scan_row_mbmi(&mut self, bo: &BlockOffset, row_offset: isize, max_row_offs: isize,
- processed_rows: &mut isize) -> bool {
+ processed_rows: &mut isize, ref_frame: usize,
+ mv_stack: &mut Vec<CandidateMV>, newmv_count: &mut usize) -> bool {
let bc = &self.bc;
let target_n4_w = bc.at(bo).n4_w;
@@ -1907,7 +1976,6 @@ impl ContextWriter {
let n4_w_8 = BlockSize::MI_SIZE_WIDE[BLOCK_8X8 as usize];
let n4_w_16 = BlockSize::MI_SIZE_WIDE[BLOCK_16X16 as usize];
let mut col_offset = 0;
- //let shift = 0;
if row_offset.abs() > 1 {
col_offset = 1;
@@ -1932,14 +2000,17 @@ impl ContextWriter {
len = cmp::max(len, n4_w_8);
}
- //let mut weight = 2;
+ let mut weight = 2 as u32;
if target_n4_w >= n4_w_8 && target_n4_w <= n4_w {
let inc = cmp::min(-max_row_offs + row_offset + 1, cand.n4_h as isize);
- //weight = cmp::max(weight, inc << shift);
+ assert!(inc >= 0);
+ weight = cmp::max(weight, inc as u32);
*processed_rows = (inc as isize) - row_offset - 1;
}
- if self.add_ref_mv_candidate(cand) { found_match = true; }
+ if self.add_ref_mv_candidate(ref_frame, cand, mv_stack, len as u32 * weight, newmv_count) {
+ found_match = true;
+ }
i += len;
}
@@ -1948,7 +2019,8 @@ impl ContextWriter {
}
fn scan_col_mbmi(&mut self, bo: &BlockOffset, col_offset: isize, max_col_offs: isize,
- processed_cols: &mut isize) -> bool {
+ processed_cols: &mut isize, ref_frame: usize,
+ mv_stack: &mut Vec<CandidateMV>, newmv_count: &mut usize) -> bool {
let bc = &self.bc;
let target_n4_h = bc.at(bo).n4_h;
@@ -1957,7 +2029,6 @@ impl ContextWriter {
let n4_h_8 = BlockSize::MI_SIZE_HIGH[BLOCK_8X8 as usize];
let n4_h_16 = BlockSize::MI_SIZE_HIGH[BLOCK_16X16 as usize];
let mut row_offset = 0;
- //let shift = 0;
if col_offset.abs() > 1 {
row_offset = 1;
@@ -1981,14 +2052,17 @@ impl ContextWriter {
len = cmp::max(len, n4_h_8);
}
- //let mut weight = 2;
+ let mut weight = 2 as u32;
if target_n4_h >= n4_h_8 && target_n4_h <= n4_h {
let inc = cmp::min(-max_col_offs + col_offset + 1, cand.n4_w as isize);
- //weight = cmp::max(weight, inc << shift);
+ assert!(inc >= 0);
+ weight = cmp::max(weight, inc as u32);
*processed_cols = (inc as isize) - col_offset - 1;
}
- if self.add_ref_mv_candidate(cand) { found_match = true; }
+ if self.add_ref_mv_candidate(ref_frame, cand, mv_stack, len as u32 * weight, newmv_count) {
+ found_match = true;
+ }
i += len;
}
@@ -1996,16 +2070,24 @@ impl ContextWriter {
found_match
}
- fn scan_blk_mbmi(&mut self, bo: &BlockOffset) -> bool {
+ fn scan_blk_mbmi(&mut self, bo: &BlockOffset, ref_frame: usize,
+ mv_stack: &mut Vec<CandidateMV>, newmv_count: &mut usize) -> bool {
if bo.x >= self.bc.cols || bo.y >= self.bc.rows {
return false;
}
+ let weight = 2 * BLOCK_8X8.width_mi() as u32;
/* Always assume its within a tile, probably wrong */
- self.add_ref_mv_candidate(self.bc.at(bo))
+ self.add_ref_mv_candidate(ref_frame, self.bc.at(bo), mv_stack, weight, newmv_count)
+ }
+
+ fn add_offset(&mut self, mv_stack: &mut Vec<CandidateMV>) {
+ for mut cand_mv in mv_stack {
+ cand_mv.weight += REF_CAT_LEVEL;
+ }
}
- fn setup_mvref_list(&mut self, bo: &BlockOffset) -> usize {
+ fn setup_mvref_list(&mut self, bo: &BlockOffset, ref_frame: usize, mv_stack: &mut Vec<CandidateMV>) -> usize {
let (_rf, _rf_num) = self.get_mvref_ref_frames(INTRA_FRAME);
let mut max_row_offs = 0 as isize;
@@ -2045,28 +2127,33 @@ impl ContextWriter {
let mut row_match = false;
let mut col_match = false;
- let newmv_count = 0;
+ let mut newmv_count: usize = 0;
if max_row_offs.abs() >= 1 {
- let found_match = self.scan_row_mbmi(bo, -1, max_row_offs, &mut processed_rows);
+ let found_match = self.scan_row_mbmi(bo, -1, max_row_offs, &mut processed_rows, ref_frame, mv_stack,
+ &mut newmv_count);
row_match |= found_match;
}
if max_col_offs.abs() >= 1 {
- let found_match = self.scan_col_mbmi(bo, -1, max_col_offs, &mut processed_cols);
+ let found_match = self.scan_col_mbmi(bo, -1, max_col_offs, &mut processed_cols, ref_frame, mv_stack,
+ &mut newmv_count);
col_match |= found_match;
}
if self.has_tr(bo) {
let n4_w = self.bc.at(bo).n4_w;
- let found_match = self.scan_blk_mbmi(&bo.with_offset(n4_w as isize, -1));
+ let found_match = self.scan_blk_mbmi(&bo.with_offset(n4_w as isize, -1), ref_frame, mv_stack,
+ &mut newmv_count);
row_match |= found_match;
}
let nearest_match = if row_match { 1 } else { 0 } + if col_match { 1 } else { 0 };
- /* TODO: set ref_mv_stack weights to REF_CAT_LEVEL for this ref frame */
+ self.add_offset(mv_stack);
/* Scan the second outer area. */
- let found_match = self.scan_blk_mbmi(&bo.with_offset(-1, -1));
+ let mut far_newmv_count: usize = 0; // won't be used
+
+ let found_match = self.scan_blk_mbmi(&bo.with_offset(-1, -1), ref_frame, mv_stack, &mut far_newmv_count);
row_match |= found_match;
for idx in 2..MVREF_ROW_COLS+1 {
@@ -2074,12 +2161,14 @@ impl ContextWriter {
let col_offset = -2 * idx as isize + 1 + col_adj as isize;
if row_offset.abs() <= max_row_offs.abs() && row_offset.abs() > processed_rows {
- let found_match = self.scan_row_mbmi(bo, row_offset, max_row_offs, &mut processed_rows);
+ let found_match = self.scan_row_mbmi(bo, row_offset, max_row_offs, &mut processed_rows, ref_frame, mv_stack,
+ &mut far_newmv_count);
row_match |= found_match;
}
if col_offset.abs() <= max_col_offs.abs() && col_offset.abs() > processed_cols {
- let found_match = self.scan_col_mbmi(bo, col_offset, max_col_offs, &mut processed_cols);
+ let found_match = self.scan_col_mbmi(bo, col_offset, max_col_offs, &mut processed_cols, ref_frame, mv_stack,
+ &mut far_newmv_count);
col_match |= found_match;
}
}
@@ -2096,12 +2185,26 @@ impl ContextWriter {
/* TODO: Find nearest match and assign nearest and near mvs */
+ // Sort MV stack according to weight
+ if mv_stack.len() > 1 {
+ let mut i: usize = 1;
+ while i < mv_stack.len() {
+ let mut j = i;
+ while j > 0 && mv_stack[j - 1].weight < mv_stack[j].weight {
+ mv_stack.swap(j, j - 1);
+ j = j - 1;
+ }
+ i = i + 1;
+ }
+ }
+
/* TODO: Handle single reference frame extension */
mode_context
}
- pub fn find_mvrefs(&mut self, bo: &BlockOffset, ref_frame: usize) -> usize {
+ pub fn find_mvrefs(&mut self, bo: &BlockOffset, ref_frame: usize,
+ mv_stack: &mut Vec<CandidateMV>) -> usize {
if ref_frame < REF_FRAMES {
if ref_frame != INTRA_FRAME {
/* TODO: convert global mv to an mv here */
@@ -2116,7 +2219,7 @@ impl ContextWriter {
/* TODO: Set the zeromv ref to 0 */
}
- let mode_context = self.setup_mvref_list(bo);
+ let mode_context = self.setup_mvref_list(bo, ref_frame, mv_stack);
mode_context
}
@@ -2296,6 +2399,26 @@ impl ContextWriter {
}
}
+ pub fn write_drl_mode(&mut self, w: &mut dyn Writer, drl_mode: bool, ctx: usize) {
+ symbol_with_update!(self, w, drl_mode as u32, &mut self.fc.drl_cdfs[ctx]);
+ }
+
+ pub fn write_mv(&mut self, w: &mut dyn Writer,
+ mv: &MotionVector, ref_mv: &MotionVector,
+ mv_precision: MvSubpelPrecision) {
+ let diff = MotionVector { row: mv.row - ref_mv.row, col: mv.col - ref_mv.col };
+ let j: MvJointType = av1_get_mv_joint(&diff);
+
+ w.symbol_with_update(j as u32, &mut self.fc.nmv_context.joints_cdf);
+
+ if mv_joint_vertical(j) {
+ encode_mv_component(w, diff.row as i32, &mut self.fc.nmv_context.comps[0], mv_precision);
+ }
+ if mv_joint_horizontal(j) {
+ encode_mv_component(w, diff.col as i32, &mut self.fc.nmv_context.comps[1], mv_precision);
+ }
+ }
+
pub fn write_tx_type(
&mut self, w: &mut dyn Writer, tx_size: TxSize, tx_type: TxType, y_mode: PredictionMode,
is_inter: bool, use_reduced_tx_set: bool
@@ -2915,29 +3038,8 @@ const MV_UPP: i32 = (1 << MV_IN_USE_BITS);
const MV_LOW: i32 = (-(1 << MV_IN_USE_BITS));
-pub struct nmv_component {
- classes_cdf: [u16; MV_CLASSES + 1],
- class0_fp_cdf: [[u16; MV_FP_SIZE + 1]; CLASS0_SIZE],
- fp_cdf: [u16; MV_FP_SIZE + 1],
- sign_cdf: [u16; 2 + 1],
- class0_hp_cdf: [u16; 2 + 1],
- hp_cdf: [u16; 2 + 1],
- class0_cdf: [u16; CLASS0_SIZE + 1],
- bits_cdf: [[u16; 2 + 1]; MV_OFFSET_BITS],
-}
-
-pub struct nmv_context {
- joints_cdf: [u16; MV_JOINTS + 1],
- comps: [nmv_component; 2],
-}
-
-pub struct MV {
- row : i16,
- col : i16,
-}
-
#[inline(always)]
-pub fn av1_get_mv_joint(mv: &MV) -> MvJointType {
+pub fn av1_get_mv_joint(mv: &MotionVector) -> MvJointType {
if mv.row == 0 {
if mv.col == 0 { MvJointType::MV_JOINT_ZERO } else { MvJointType::MV_JOINT_HNZVZ }
} else {
@@ -2961,7 +3063,7 @@ pub fn mv_class_base(mv_class: usize) -> u32 {
#[inline(always)]
// If n != 0, returns the floor of log base 2 of n. If n == 0, returns 0.
pub fn log_in_base_2(n: u32) -> u8 {
- 32 - n.leading_zeros() as u8
+ 31 - cmp::min(31, n.leading_zeros() as u8)
}
#[inline(always)]
pub fn get_mv_class(z: u32, offset: &mut u32) -> usize {
@@ -2974,7 +3076,7 @@ pub fn get_mv_class(z: u32, offset: &mut u32) -> usize {
}
pub fn encode_mv_component(w: &mut Writer, comp: i32,
- mvcomp: &mut nmv_component, precision: MvSubpelPrecision) {
+ mvcomp: &mut NMVComponent, precision: MvSubpelPrecision) {
assert!(comp != 0);
let mut offset: u32 = 0;
let sign: u32 = if comp < 0 { 1 } else { 0 };
@@ -3016,22 +3118,3 @@ pub fn encode_mv_component(w: &mut Writer, comp: i32,
}
}
-pub fn av1_encode_mv(w: &mut Writer,
- mv: &MV, ref_mv: &MV,
- mvctx: &mut nmv_context, mut usehp: MvSubpelPrecision) {
- let diff = MV { row: mv.row - ref_mv.row, col: mv.col - ref_mv.col };
- let j: MvJointType = av1_get_mv_joint(&diff);
-
- // TODO: pass fi.force_integer_mv to this function
- if false /*fi.force_integer_mv*/ {
- usehp = MvSubpelPrecision::MV_SUBPEL_NONE;
- }
- w.symbol_with_update(j as u32, &mut mvctx.joints_cdf);
-
- if mv_joint_vertical(j) {
- encode_mv_component(w, diff.row as i32, &mut mvctx.comps[0], usehp);
- }
- if mv_joint_horizontal(j) {
- encode_mv_component(w, diff.col as i32, &mut mvctx.comps[1], usehp);
- }
-}
diff --git a/src/encoder.rs b/src/encoder.rs
index f142bd71..f6c248a7 100644
--- a/src/encoder.rs
+++ b/src/encoder.rs
@@ -340,7 +340,7 @@ impl FrameInvariants {
showable_frame: true,
error_resilient: true,
intra_only: false,
- allow_high_precision_mv: true,
+ allow_high_precision_mv: false,
frame_type: FrameType::KEY,
show_existing_frame: false,
use_reduced_tx_set,
@@ -1198,6 +1198,7 @@ pub fn encode_block_a(seq: &Sequence,
pub fn encode_block_b(fi: &FrameInvariants, fs: &mut FrameState,
cw: &mut ContextWriter, w: &mut dyn Writer,
luma_mode: PredictionMode, chroma_mode: PredictionMode,
+ ref_frame: usize, mv: MotionVector,
bsize: BlockSize, bo: &BlockOffset, skip: bool, bit_depth: usize) {
let is_inter = !luma_mode.is_intra();
if is_inter { assert!(luma_mode == chroma_mode); };
@@ -1208,15 +1209,46 @@ pub fn encode_block_b(fi: &FrameInvariants, fs: &mut FrameState,
if fi.frame_type == FrameType::INTER {
cw.write_is_inter(w, bo, is_inter);
if is_inter {
- let ref_frame = LAST_FRAME;
cw.fill_neighbours_ref_counts(bo);
cw.bc.set_ref_frame(bo, bsize, ref_frame);
+ cw.bc.set_motion_vector(bo, bsize, mv);
cw.write_ref_frames(w, bo);
- let mode_context = cw.find_mvrefs(bo, ref_frame);
+
+ let mut mv_stack = Vec::new();
+ let mode_context = cw.find_mvrefs(bo, ref_frame, &mut mv_stack);
//let mode_context = if bo.x == 0 && bo.y == 0 { 0 } else if bo.x ==0 || bo.y == 0 { 51 } else { 85 };
// NOTE: Until rav1e supports other inter modes than GLOBALMV
- assert!(luma_mode == PredictionMode::GLOBALMV);
cw.write_inter_mode(w, luma_mode, mode_context);
+
+ if luma_mode == PredictionMode::NEWMV || luma_mode == PredictionMode::NEW_NEWMV {
+ let ref_mv_idx = 0;
+ let num_mv_found = mv_stack.len();
+ for idx in 0..2 {
+ if num_mv_found > idx + 1 {
+ let drl_mode = ref_mv_idx > idx;
+ let ctx: usize = (mv_stack[idx].weight < REF_CAT_LEVEL) as usize
+ + (mv_stack[idx + 1].weight < REF_CAT_LEVEL) as usize;
+
+ cw.write_drl_mode(w, drl_mode, ctx);
+ if !drl_mode { break; }
+ }
+ }
+
+ let ref_mv = if num_mv_found > 0 {
+ mv_stack[ref_mv_idx].this_mv
+ } else {
+ MotionVector{ row: 0, col: 0 }
+ };
+
+ let mv_precision = if fi.force_integer_mv != 0 {
+ MvSubpelPrecision::MV_SUBPEL_NONE
+ } else if fi.allow_high_precision_mv {
+ MvSubpelPrecision::MV_SUBPEL_HIGH_PRECISION
+ } else {
+ MvSubpelPrecision::MV_SUBPEL_LOW_PRECISION
+ };
+ cw.write_mv(w, &mv, &ref_mv, mv_precision);
+ }
} else {
cw.write_intra_mode(w, bsize, luma_mode);
}
@@ -1267,6 +1299,9 @@ pub fn encode_block_b(fi: &FrameInvariants, fs: &mut FrameState,
};
if is_inter {
+ {
+ let ref_frame = cw.bc.at(bo).ref_frames[0];
+ let mv = &cw.bc.at(bo).mv[0];
// Inter mode prediction can take place once for a whole partition,
// instead of each tx-block.
let num_planes = 1 + if has_chroma(bo, bsize, xdec, ydec) { 2 } else { 0 };
@@ -1278,11 +1313,38 @@ pub fn encode_block_b(fi: &FrameInvariants, fs: &mut FrameState,
let rec = &mut fs.rec.planes[p];
- luma_mode.predict_inter(fi, p, &po, &mut rec.mut_slice(&po), plane_bsize);
+ // TODO: make more generic to handle 2xN and Nx2 MC
+ if p > 0 && bsize == BlockSize::BLOCK_4X4 {
+ let mv0 = &cw.bc.at(&bo.with_offset(-1,-1)).mv[0];
+ let mv1 = &cw.bc.at(&bo.with_offset(0,-1)).mv[0];
+ let po1 = PlaneOffset { x: po.x+2, y: po.y };
+ let mv2 = &cw.bc.at(&bo.with_offset(-1,0)).mv[0];
+ let po2 = PlaneOffset { x: po.x, y: po.y+2 };
+ let po3 = PlaneOffset { x: po.x+2, y: po.y+2 };
+ let some_use_intra = cw.bc.at(&bo.with_offset(-1,-1)).mode.is_intra()
+ || cw.bc.at(&bo.with_offset(0,-1)).mode.is_intra()
+ || cw.bc.at(&bo.with_offset(-1,0)).mode.is_intra();
+
+ if some_use_intra {
+ luma_mode.predict_inter(fi, p, &po, &mut rec.mut_slice(&po), plane_bsize.width(),
+ plane_bsize.height(), ref_frame, mv);
+ } else {
+ luma_mode.predict_inter(fi, p, &po, &mut rec.mut_slice(&po), 2, 2, ref_frame, mv0);
+ luma_mode.predict_inter(fi, p, &po1, &mut rec.mut_slice(&po1), 2, 2, ref_frame, mv1);
+ luma_mode.predict_inter(fi, p, &po2, &mut rec.mut_slice(&po2), 2, 2, ref_frame, mv2);
+ luma_mode.predict_inter(fi, p, &po3, &mut rec.mut_slice(&po3), 2, 2, ref_frame, mv);
+ }
+ }
+ else
+ {
+ luma_mode.predict_inter(fi, p, &po, &mut rec.mut_slice(&po), plane_bsize.width(),
+ plane_bsize.height(), ref_frame, mv);
+ }
}
- write_tx_tree(fi, fs, cw, w, luma_mode, bo, bsize, tx_size, tx_type, skip, bit_depth); // i.e. var-tx if inter mode
+ }
+ write_tx_tree(fi, fs, cw, w, luma_mode, bo, bsize, tx_size, tx_type, skip, bit_depth); // i.e. var-tx if inter mode
} else {
- write_tx_blocks(fi, fs, cw, w, luma_mode, chroma_mode, bo, bsize, tx_size, tx_type, skip, bit_depth);
+ write_tx_blocks(fi, fs, cw, w, luma_mode, chroma_mode, bo, bsize, tx_size, tx_type, skip, bit_depth);
}
}
@@ -1437,6 +1499,8 @@ fn encode_partition_bottomup(seq: &Sequence, fi: &FrameInvariants, fs: &mut Fram
bo: bo.clone(),
pred_mode_luma: PredictionMode::DC_PRED,
pred_mode_chroma: PredictionMode::DC_PRED,
+ ref_frame: INTRA_FRAME,
+ mv: MotionVector { row: 0, col: 0},
skip: false
}; // Best decision that is not PARTITION_SPLIT
@@ -1457,6 +1521,8 @@ fn encode_partition_bottomup(seq: &Sequence, fi: &FrameInvariants, fs: &mut Fram
}
let mode_decision = rdo_mode_decision(seq, fi, fs, cw, bsize, bo).part_modes[0].clone();
let (mode_luma, mode_chroma) = (mode_decision.pred_mode_luma, mode_decision.pred_mode_chroma);
+ let ref_frame = mode_decision.ref_frame;
+ let mv = mode_decision.mv;
let skip = mode_decision.skip;
let mut cdef_coded = cw.bc.cdef_coded;
rd_cost = mode_decision.rd_cost;
@@ -1464,7 +1530,7 @@ fn encode_partition_bottomup(seq: &Sequence, fi: &FrameInvariants, fs: &mut Fram
cdef_coded = encode_block_a(seq, cw, if cdef_coded {w_post_cdef} else {w_pre_cdef},
bsize, bo, skip);
encode_block_b(fi, fs, cw, if cdef_coded {w_post_cdef} else {w_pre_cdef},
- mode_luma, mode_chroma, bsize, bo, skip, seq.bit_depth);
+ mode_luma, mode_chroma, ref_frame, mv, bsize, bo, skip, seq.bit_depth);
best_decision = mode_decision;
}
@@ -1509,12 +1575,14 @@ fn encode_partition_bottomup(seq: &Sequence, fi: &FrameInvariants, fs: &mut Fram
// FIXME: redundant block re-encode
let (mode_luma, mode_chroma) = (best_decision.pred_mode_luma, best_decision.pred_mode_chroma);
+ let ref_frame = best_decision.ref_frame;
+ let mv = best_decision.mv;
let skip = best_decision.skip;
let mut cdef_coded = cw.bc.cdef_coded;
cdef_coded = encode_block_a(seq, cw, if cdef_coded {w_post_cdef} else {w_pre_cdef},
bsize, bo, skip);
encode_block_b(fi, fs, cw, if cdef_coded {w_post_cdef} else {w_pre_cdef},
- mode_luma, mode_chroma, bsize, bo, skip, seq.bit_depth);
+ mode_luma, mode_chroma, ref_frame, mv, bsize, bo, skip, seq.bit_depth);
}
}
@@ -1586,13 +1654,15 @@ fn encode_partition_topdown(seq: &Sequence, fi: &FrameInvariants, fs: &mut Frame
let (mode_luma, mode_chroma) = (part_decision.pred_mode_luma, part_decision.pred_mode_chroma);
let skip = part_decision.skip;
+ let ref_frame = part_decision.ref_frame;
+ let mv = part_decision.mv;
let mut cdef_coded = cw.bc.cdef_coded;
// FIXME: every final block that has gone through the RDO decision process is encoded twice
cdef_coded = encode_block_a(seq, cw, if cdef_coded {w_post_cdef} else {w_pre_cdef},
bsize, bo, skip);
encode_block_b(fi, fs, cw, if cdef_coded {w_post_cdef} else {w_pre_cdef},
- mode_luma, mode_chroma, bsize, bo, skip, seq.bit_depth);
+ mode_luma, mode_chroma, ref_frame, mv, bsize, bo, skip, seq.bit_depth);
},
PartitionType::PARTITION_SPLIT => {
if rdo_output.part_modes.len() >= 4 {
diff --git a/src/lib.rs b/src/lib.rs
index 62b33d02..6e16d369 100755
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -29,6 +29,7 @@ pub mod rdo;
pub mod util;
pub mod cdef;
pub mod encoder;
+pub mod me;
pub use encoder::*;
diff --git a/src/me.rs b/src/me.rs
new file mode 100644
index 00000000..68173aca
--- /dev/null
+++ b/src/me.rs
@@ -0,0 +1,72 @@
+// 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.
+
+use std::cmp;
+use FrameInvariants;
+use FrameState;
+use partition::BlockSize;
+use context::BlockOffset;
+use partition::MotionVector;
+use partition::LAST_FRAME;
+use plane::PlaneOffset;
+use context::BLOCK_TO_PLANE_SHIFT;
+
+pub fn motion_estimation(fi: &FrameInvariants, fs: &mut FrameState, bsize: BlockSize,
+ bo: &BlockOffset, ref_frame: usize) -> MotionVector {
+
+ match fi.rec_buffer.frames[fi.ref_frames[ref_frame - LAST_FRAME]] {
+ Some(ref rec) => {
+ let po = PlaneOffset { x: bo.x << BLOCK_TO_PLANE_SHIFT, y: bo.y << BLOCK_TO_PLANE_SHIFT };
+ let range = 16 as usize;
+ let blk_w = bsize.width();
+ let blk_h = bsize.height();
+ let x_lo = cmp::max(0, po.x as isize - range as isize) as usize;
+ let x_hi = cmp::min(fs.input.planes[0].cfg.width - blk_w, po.x + range);
+ let y_lo = cmp::max(0, po.y as isize - range as isize) as usize;
+ let y_hi = cmp::min(fs.input.planes[0].cfg.height - blk_h, po.y + range);
+
+ let sorg = fs.input.planes[0].slice(&po);
+ let slice_org = sorg.as_slice();
+ let stride_org = fs.input.planes[0].cfg.stride;
+ let stride_ref = rec.planes[0].cfg.stride;
+
+ let mut lowest_sad = 128*128*4096 as usize;
+ let mut best_mv = MotionVector { row: 0, col: 0 };
+
+ for y in (y_lo..y_hi).step_by(2) {
+ for x in (x_lo..x_hi).step_by(2) {
+
+ let mut sad = 0;
+ let sref = rec.planes[0].slice(&PlaneOffset { x: x, y: y });
+ let slice_ref = sref.as_slice();
+
+ for r in 0..blk_h {
+ for c in 0..blk_w {
+ let org_index = r * stride_org + c;
+ let ref_index = r * stride_ref + c;
+ let a = slice_org[org_index];
+ let b = slice_ref[ref_index];
+ let delta = b as isize - a as isize;
+ sad += delta.abs() as usize;
+ }
+ }
+
+ if sad < lowest_sad {
+ lowest_sad = sad;
+ best_mv = MotionVector { row: 8*(y as i16 - po.y as i16), col: 8*(x as i16 - po.x as i16) }
+ }
+
+ }
+ }
+ best_mv
+ }
+
+ None => MotionVector { row: 0, col : 0 }
+ }
+}
diff --git a/src/partition.rs b/src/partition.rs
index c2059ca8..27d174ce 100755
--- a/src/partition.rs
+++ b/src/partition.rs
@@ -10,6 +10,7 @@
#![allow(non_camel_case_types)]
#![allow(dead_code)]
+use std::cmp;
use self::BlockSize::*;
use self::TxSize::*;
use encoder::FrameInvariants;
@@ -348,6 +349,12 @@ pub enum PredictionMode {
NEW_NEWMV
}
+#[derive(Copy, Clone)]
+pub struct MotionVector {
+ pub row: i16,
+ pub col: i16,
+}
+
pub const NEWMV_MODE_CONTEXTS: usize = 7;
pub const GLOBALMV_MODE_CONTEXTS: usize = 2;
pub const REFMV_MODE_CONTEXTS: usize = 9;
@@ -529,24 +536,30 @@ impl PredictionMode {
}
pub fn predict_inter<'a>(self, fi: &FrameInvariants, p: usize, po: &PlaneOffset,
- dst: &'a mut PlaneMutSlice<'a>, plane_size: BlockSize) {
+ dst: &'a mut PlaneMutSlice<'a>, width: usize, height: usize,
+ ref_frame: usize, mv: &MotionVector) {
assert!(!self.is_intra());
- assert!(self == PredictionMode::GLOBALMV); // Other modes not implemented
-
- let ref_frame_idx = LAST_FRAME;
+ assert!(ref_frame == LAST_FRAME);
- match fi.rec_buffer.frames[fi.ref_frames[ref_frame_idx - LAST_FRAME]] {
+ match fi.rec_buffer.frames[fi.ref_frames[ref_frame - LAST_FRAME]] {
Some(ref rec) => {
- let ref_stride = rec.planes[p].cfg.stride;
- let src = rec.planes[p].slice(po);
- let ref_slice = src.as_slice();
+ let rec_cfg = &rec.planes[p].cfg;
+ let shift_row = 3 + rec_cfg.ydec;
+ let shift_col = 3 + rec_cfg.xdec;
+ let row_offset = mv.row as i32 >> shift_row;
+ let col_offset = mv.col as i32 >> shift_col;
+ let ref_width = rec_cfg.width;
+ let ref_height = rec_cfg.height;
+
let stride = dst.plane.cfg.stride;
let slice = dst.as_mut_slice();
- for r in 0..plane_size.height() {
- for c in 0..plane_size.width() {
- let input_index = r * ref_stride + c;
+
+ for r in 0..height {
+ for c in 0..width {
+ let rs = cmp::min(ref_height as i32 - 1, cmp::max(0, po.y as i32 + row_offset + r as i32)) as usize;
+ let cs = cmp::min(ref_width as i32 - 1, cmp::max(0, po.x as i32 + col_offset + c as i32)) as usize;
let output_index = r * stride + c;
- slice[output_index] = ref_slice[input_index];
+ slice[output_index] = rec.planes[p].p(cs, rs);
}
}
},
diff --git a/src/predict.rs b/src/predict.rs
index 723b9c78..7ae0211e 100755
--- a/src/predict.rs
+++ b/src/predict.rs
@@ -38,6 +38,7 @@ pub static RAV1E_INTRA_MODES_MINIMAL: &'static [PredictionMode] = &[
pub static RAV1E_INTER_MODES: &'static [PredictionMode] = &[
PredictionMode::GLOBALMV,
+ PredictionMode::NEWMV,
];
// Weights are quadratic from '1' to '1 / block_size', scaled by 2^sm_weight_log2_scale.
diff --git a/src/rdo.rs b/src/rdo.rs
index a7d82ddf..e0fbb005 100755
--- a/src/rdo.rs
+++ b/src/rdo.rs
@@ -12,6 +12,7 @@
#![cfg_attr(feature = "cargo-clippy", allow(cast_lossless))]
use context::*;
+use me::*;
use ec::OD_BITRES;
use ec::Writer;
use ec::WriterCounter;
@@ -48,6 +49,8 @@ pub struct RDOPartitionOutput {
pub bo: BlockOffset,
pub pred_mode_luma: PredictionMode,
pub pred_mode_chroma: PredictionMode,
+ pub ref_frame: usize,
+ pub mv: MotionVector,
pub skip: bool
}
@@ -189,6 +192,8 @@ pub fn rdo_mode_decision(
let mut best_mode_chroma = PredictionMode::DC_PRED;
let mut best_skip = false;
let mut best_rd = std::f64::MAX;
+ let mut best_ref_frame = INTRA_FRAME;
+ let mut best_mv = MotionVector { row: 0, col: 0 };
// Get block luma and chroma dimensions
let w = bsize.width();
@@ -223,6 +228,13 @@ pub fn rdo_mode_decision(
mode_set_chroma.push(PredictionMode::DC_PRED);
}
+ let ref_frame = if luma_mode.is_intra() { INTRA_FRAME } else { LAST_FRAME };
+ let mv = if luma_mode != PredictionMode::NEWMV {
+ MotionVector { row: 0, col: 0 }
+ } else {
+ motion_estimation(fi, fs, bsize, bo, ref_frame)
+ };
+
// Find the best chroma prediction mode for the current luma prediction mode
for &chroma_mode in &mode_set_chroma {
for &skip in &[false, true] {
@@ -232,8 +244,9 @@ pub fn rdo_mode_decision(
let mut wr: &mut dyn Writer = &mut WriterCounter::new();
let tell = wr.tell_frac();
+
encode_block_a(seq, cw, wr, bsize, bo, skip);
- encode_block_b(fi, fs, cw, wr, luma_mode, chroma_mode, bsize, bo, skip, seq.bit_depth);
+ encode_block_b(fi, fs, cw, wr, luma_mode, chroma_mode, ref_frame, mv, bsize, bo, skip, seq.bit_depth);
let cost = wr.tell_frac() - tell;
let rd = compute_rd_cost(
@@ -251,6 +264,8 @@ pub fn rdo_mode_decision(
best_rd = rd;
best_mode_luma = luma_mode;
best_mode_chroma = chroma_mode;
+ best_ref_frame = ref_frame;
+ best_mv = mv;
best_skip = skip;
}
@@ -270,6 +285,8 @@ pub fn rdo_mode_decision(
bo: bo.clone(),
pred_mode_luma: best_mode_luma,
pred_mode_chroma: best_mode_chroma,
+ ref_frame: best_ref_frame,
+ mv: best_mv,
rd_cost: best_rd,
skip: best_skip
}]
--
GitLab