+use nihav_core::codecs::{DecoderResult, DecoderError};
use nihav_core::frame::*;
-use nihav_codec_support::codecs::MV;
-use super::{CurrentMBInfo, I4X4_SCAN};
+use nihav_codec_support::codecs::{MV, ZERO_MV};
+use super::{CurrentMBInfo, I4X4_SCAN, Shareable};
+use super::dispatch::{ThreadDispatcher, FrameDecodingStatus};
use super::dsp::*;
-use super::pic_ref::FrameRefs;
+use super::pic_ref::SliceRefs;
use super::slice::{SliceHeader, WeightInfo, DEF_WEIGHT_INFO};
use super::types::*;
}
}
-fn do_p_mc(frm: &mut NASimpleVideoFrame<u8>, xpos: usize, ypos: usize, w: usize, h: usize, mv: MV, ref_pic: Option<NAVideoBufferRef<u8>>, weight: &WeightInfo) {
+fn do_p_mc(frm: &mut NASimpleVideoFrame<u8>, xpos: usize, ypos: usize, w: usize, h: usize, mv: MV, ref_pic: Option<NAVideoBufferRef<u8>>, weight: &WeightInfo, mc_dsp: &mut H264MC) {
if let Some(buf) = ref_pic {
if !weight.is_weighted() {
- do_mc(frm, buf, xpos, ypos, w, h, mv);
+ mc_dsp.do_mc(frm, buf, xpos, ypos, w, h, mv);
} else {
- let mut ytmp = [0; 16 * 16];
- let mut utmp = [0; 16 * 16];
- let mut vtmp = [0; 16 * 16];
- mc_blocks(&mut ytmp, &mut utmp, &mut vtmp, buf, xpos, ypos, w, h, mv);
+ let mut tmp = McBlock::new();
+ mc_dsp.mc_blocks(&mut tmp, buf, xpos, ypos, w, h, mv);
let yoff = frm.offset[0] + xpos + ypos * frm.stride[0];
let yw = if weight.luma_weighted {
} else {
[1, 0, 0]
};
- put_block_weighted(&mut frm.data[yoff..], frm.stride[0], &ytmp, w, h, yw);
+ let wmode = match w {
+ 2 => 0,
+ 4 => 1,
+ 8 => 2,
+ _ => 3,
+ };
+ (mc_dsp.put_block_weighted[wmode])(&mut frm.data[yoff..], frm.stride[0], &tmp.y, h, yw);
for chroma in 0..2 {
let cstride = frm.stride[chroma + 1];
} else {
[1, 0, 0]
};
- let csrc = if chroma == 0 { &utmp } else { &vtmp };
- put_block_weighted(&mut frm.data[coff..], cstride, csrc, w / 2, h / 2, cw);
+ let csrc = if chroma == 0 { &tmp.u } else { &tmp.v };
+ (mc_dsp.put_block_weighted[wmode - 1])(&mut frm.data[coff..], cstride, csrc, h / 2, cw);
}
}
} else {
- gray_block(frm, xpos, ypos, w, h);
+ mc_dsp.gray_block(frm, xpos, ypos, w, h);
}
}
-fn do_b_mc(frm: &mut NASimpleVideoFrame<u8>, mode: BMode, xpos: usize, ypos: usize, w: usize, h: usize, mv0: MV, ref_pic0: Option<NAVideoBufferRef<u8>>, weight0: &WeightInfo, mv1: MV, ref_pic1: Option<NAVideoBufferRef<u8>>, weight1: &WeightInfo, avg_buf: &mut NAVideoBufferRef<u8>) {
+#[allow(clippy::match_like_matches_macro)]
+fn do_b_mc(frm: &mut NASimpleVideoFrame<u8>, mode: BMode, xpos: usize, ypos: usize, w: usize, h: usize, mv0: MV, ref_pic0: Option<NAVideoBufferRef<u8>>, weight0: &WeightInfo, mv1: MV, ref_pic1: Option<NAVideoBufferRef<u8>>, weight1: &WeightInfo, mc_dsp: &mut H264MC) {
let do_weight = match (mode, weight0.is_weighted(), weight1.is_weighted()) {
(BMode::L0, true, _) => true,
(BMode::L1, _, true) => true,
match mode {
BMode::L0 => {
if let Some(buf) = ref_pic0 {
- do_mc(frm, buf, xpos, ypos, w, h, mv0);
+ mc_dsp.do_mc(frm, buf, xpos, ypos, w, h, mv0);
} else {
- gray_block(frm, xpos, ypos, w, h);
+ mc_dsp.gray_block(frm, xpos, ypos, w, h);
}
},
BMode::L1 => {
if let Some(buf) = ref_pic1 {
- do_mc(frm, buf, xpos, ypos, w, h, mv1);
+ mc_dsp.do_mc(frm, buf, xpos, ypos, w, h, mv1);
} else {
- gray_block(frm, xpos, ypos, w, h);
+ mc_dsp.gray_block(frm, xpos, ypos, w, h);
}
},
BMode::Bi => {
match (ref_pic0, ref_pic1) {
(Some(buf0), Some(buf1)) => {
- do_mc(frm, buf0, xpos, ypos, w, h, mv0);
- do_mc_avg(frm, buf1, xpos, ypos, w, h, mv1, avg_buf);
+ mc_dsp.do_mc(frm, buf0, xpos, ypos, w, h, mv0);
+ mc_dsp.do_mc_avg(frm, buf1, xpos, ypos, w, h, mv1);
},
(Some(buf0), None) => {
- do_mc(frm, buf0, xpos, ypos, w, h, mv0);
+ mc_dsp.do_mc(frm, buf0, xpos, ypos, w, h, mv0);
},
(None, Some(buf1)) => {
- do_mc(frm, buf1, xpos, ypos, w, h, mv1);
+ mc_dsp.do_mc(frm, buf1, xpos, ypos, w, h, mv1);
},
(None, None) => {
- gray_block(frm, xpos, ypos, w, h);
+ mc_dsp.gray_block(frm, xpos, ypos, w, h);
},
};
},
};
} else {
- let mut ytmp0 = [0x80; 16 * 16];
- let mut utmp0 = [0x80; 16 * 16];
- let mut vtmp0 = [0x80; 16 * 16];
- let mut ytmp1 = [0x80; 16 * 16];
- let mut utmp1 = [0x80; 16 * 16];
- let mut vtmp1 = [0x80; 16 * 16];
+ let mut tmp0 = McBlock::new();
+ let mut tmp1 = McBlock::new();
match (mode, ref_pic0, ref_pic1) {
(BMode::L0, Some(buf), _) | (BMode::L1, _, Some(buf)) => {
let (mv, weight) = if mode == BMode::L0 { (mv0, weight0) } else { (mv1, weight1) };
- mc_blocks(&mut ytmp0, &mut utmp0, &mut vtmp0, buf, xpos, ypos, w, h, mv);
+ mc_dsp.mc_blocks(&mut tmp0, buf, xpos, ypos, w, h, mv);
let yoff = frm.offset[0] + xpos + ypos * frm.stride[0];
let yw = if weight.luma_weighted {
} else {
[1, 0, 0]
};
- put_block_weighted(&mut frm.data[yoff..], frm.stride[0], &ytmp0, w, h, yw);
+ let wmode = match w {
+ 2 => 0,
+ 4 => 1,
+ 8 => 2,
+ _ => 3,
+ };
+ (mc_dsp.put_block_weighted[wmode])(&mut frm.data[yoff..], frm.stride[0], &tmp0.y, h, yw);
for chroma in 0..2 {
let cstride = frm.stride[chroma + 1];
} else {
[1, 0, 0]
};
- let csrc = if chroma == 0 { &utmp0 } else { &vtmp0 };
- put_block_weighted(&mut frm.data[coff..], cstride, csrc, w / 2, h / 2, cw);
+ let csrc = if chroma == 0 { &tmp0.u } else { &tmp0.v };
+ (mc_dsp.put_block_weighted[wmode - 1])(&mut frm.data[coff..], cstride, csrc, h / 2, cw);
}
},
(BMode::Bi, Some(buf0), Some(buf1)) => { // do both and avg
- mc_blocks(&mut ytmp0, &mut utmp0, &mut vtmp0, buf0, xpos, ypos, w, h, mv0);
- mc_blocks(&mut ytmp1, &mut utmp1, &mut vtmp1, buf1, xpos, ypos, w, h, mv1);
+ mc_dsp.mc_blocks(&mut tmp0, buf0, xpos, ypos, w, h, mv0);
+ mc_dsp.mc_blocks(&mut tmp1, buf1, xpos, ypos, w, h, mv1);
let yoff = frm.offset[0] + xpos + ypos * frm.stride[0];
let yw = match (weight0.luma_weighted, weight1.luma_weighted) {
(false, true) => [1 << weight1.luma_shift, 0, weight1.luma_weight, weight1.luma_offset, weight1.luma_shift as i8],
(false, false) => [1, 0, 1, 0, 0],
};
- put_block_weighted2(&mut frm.data[yoff..], frm.stride[0], &ytmp0, &ytmp1, w, h, yw);
+ let wmode = match w {
+ 2 => 0,
+ 4 => 1,
+ 8 => 2,
+ _ => 3,
+ };
+ (mc_dsp.put_block_weighted2[wmode])(&mut frm.data[yoff..], frm.stride[0], &tmp0.y, &tmp1.y, h, yw);
for chroma in 0..2 {
let cstride = frm.stride[chroma + 1];
(false, true) => [1 << weight1.luma_shift, 0, cw1, co1, weight1.luma_shift as i8],
(false, false) => [1, 0, 1, 0, 0],
};
- let csrc0 = if chroma == 0 { &utmp0 } else { &vtmp0 };
- let csrc1 = if chroma == 0 { &utmp1 } else { &vtmp1 };
- put_block_weighted2(&mut frm.data[coff..], cstride, csrc0, csrc1, w / 2, h / 2, cw);
+ let csrc0 = if chroma == 0 { &tmp0.u } else { &tmp0.v };
+ let csrc1 = if chroma == 0 { &tmp1.u } else { &tmp1.v };
+ (mc_dsp.put_block_weighted2[wmode - 1])(&mut frm.data[coff..], cstride, csrc0, csrc1, h / 2, cw);
}
},
_ => {
- gray_block(frm, xpos, ypos, w, h);
+ mc_dsp.gray_block(frm, xpos, ypos, w, h);
},
};
}
}
-fn get_weights(slice_hdr: &SliceHeader, frame_refs: &FrameRefs, mode: BMode, weight_mode: u8, ref_l0: PicRef, ref_l1: PicRef) -> (WeightInfo, WeightInfo) {
+fn get_weights(slice_hdr: &SliceHeader, frame_refs: &SliceRefs, mode: BMode, weight_mode: u8, ref_l0: PicRef, ref_l1: PicRef) -> (WeightInfo, WeightInfo) {
let idx_l0 = ref_l0.index();
let idx_l1 = ref_l1.index();
if mode != BMode::Bi || weight_mode != 2 {
}
}
-pub fn recon_mb(frm: &mut NASimpleVideoFrame<u8>, slice_hdr: &SliceHeader, mb_info: &CurrentMBInfo, sstate: &mut SliceState, frame_refs: &FrameRefs, avg_buf: &mut NAVideoBufferRef<u8>, weight_mode: u8) {
+pub fn recon_mb(frm: &mut NASimpleVideoFrame<u8>, slice_hdr: &SliceHeader, mb_info: &CurrentMBInfo, sstate: &mut SliceState, frame_refs: &SliceRefs, mc_dsp: &mut H264MC, weight_mode: u8) {
let xpos = sstate.mb_x * 16;
let ypos = sstate.mb_y * 16;
match mb_info.mb_type {
MBType::Intra16x16(_, _, _) => {
- pred_intra(frm, &sstate, &mb_info);
+ pred_intra(frm, sstate, mb_info);
},
MBType::Intra4x4 | MBType::Intra8x8 => {
- pred_intra(frm, &sstate, &mb_info);
+ pred_intra(frm, sstate, mb_info);
},
MBType::PCM => {},
MBType::PSkip => {
let mv = sstate.get_cur_blk4(0).mv[0];
let rpic = frame_refs.select_ref_pic(0, 0);
let weight = &slice_hdr.get_weight(0, 0);
- do_p_mc(frm, xpos, ypos, 16, 16, mv, rpic, weight);
+ do_p_mc(frm, xpos, ypos, 16, 16, mv, rpic, weight, mc_dsp);
},
MBType::P16x16 => {
let mv = sstate.get_cur_blk4(0).mv[0];
let rpic = frame_refs.select_ref_pic(0, mb_info.ref_l0[0].index());
let weight = &slice_hdr.get_weight(0, mb_info.ref_l0[0].index());
- do_p_mc(frm, xpos, ypos, 16, 16, mv, rpic, weight);
+ do_p_mc(frm, xpos, ypos, 16, 16, mv, rpic, weight, mc_dsp);
},
MBType::P16x8 | MBType::P8x16 => {
let (bw, bh, bx, by) = if mb_info.mb_type == MBType::P16x8 {
let mv = sstate.get_cur_blk4(0).mv[0];
let rpic = frame_refs.select_ref_pic(0, mb_info.ref_l0[0].index());
let weight = &slice_hdr.get_weight(0, mb_info.ref_l0[0].index());
- do_p_mc(frm, xpos, ypos, bw, bh, mv, rpic, weight);
+ do_p_mc(frm, xpos, ypos, bw, bh, mv, rpic, weight, mc_dsp);
let mv = sstate.get_cur_blk4(bx / 4 + by).mv[0];
let rpic = frame_refs.select_ref_pic(0, mb_info.ref_l0[1].index());
let weight = &slice_hdr.get_weight(0, mb_info.ref_l0[1].index());
- do_p_mc(frm, xpos + bx, ypos + by, bw, bh, mv, rpic, weight);
+ do_p_mc(frm, xpos + bx, ypos + by, bw, bh, mv, rpic, weight, mc_dsp);
},
MBType::P8x8 | MBType::P8x8Ref0 => {
for part in 0..4 {
match mb_info.sub_mb_type[part] {
SubMBType::P8x8 => {
- do_p_mc(frm, xpos + bx, ypos + by, 8, 8, mv, rpic, weight);
+ do_p_mc(frm, xpos + bx, ypos + by, 8, 8, mv, rpic, weight, mc_dsp);
},
SubMBType::P8x4 => {
- do_p_mc(frm, xpos + bx, ypos + by, 8, 4, mv, rpic.clone(), weight);
+ do_p_mc(frm, xpos + bx, ypos + by, 8, 4, mv, rpic.clone(), weight, mc_dsp);
let mv = sstate.get_cur_blk4(bx / 4 + by + 4).mv[0];
- do_p_mc(frm, xpos + bx, ypos + by + 4, 8, 4, mv, rpic, weight);
+ do_p_mc(frm, xpos + bx, ypos + by + 4, 8, 4, mv, rpic, weight, mc_dsp);
},
SubMBType::P4x8 => {
- do_p_mc(frm, xpos + bx, ypos + by, 4, 8, mv, rpic.clone(), weight);
+ do_p_mc(frm, xpos + bx, ypos + by, 4, 8, mv, rpic.clone(), weight, mc_dsp);
let mv = sstate.get_cur_blk4(bx / 4 + by + 1).mv[0];
- do_p_mc(frm, xpos + bx + 4, ypos + by, 4, 8, mv, rpic, weight);
+ do_p_mc(frm, xpos + bx + 4, ypos + by, 4, 8, mv, rpic, weight, mc_dsp);
},
SubMBType::P4x4 => {
for sb_no in 0..4 {
let sypos = ypos + by + (sb_no & 2) * 2;
let sblk_no = (bx / 4 + (sb_no & 1)) + ((by / 4) + (sb_no >> 1)) * 4;
let mv = sstate.get_cur_blk4(sblk_no).mv[0];
- do_p_mc(frm, sxpos, sypos, 4, 4, mv, rpic.clone(), weight);
+ do_p_mc(frm, sxpos, sypos, 4, 4, mv, rpic.clone(), weight, mc_dsp);
}
},
_ => unreachable!(),
let mv1 = sstate.get_cur_blk4(0).mv[1];
let rpic1 = frame_refs.select_ref_pic(1, mb_info.ref_l1[0].index());
let (weight0, weight1) = get_weights(slice_hdr, frame_refs, mode, weight_mode, mb_info.ref_l0[0], mb_info.ref_l1[0]);
- do_b_mc(frm, mode, xpos, ypos, 16, 16, mv0, rpic0, &weight0, mv1, rpic1, &weight1, avg_buf);
+ do_b_mc(frm, mode, xpos, ypos, 16, 16, mv0, rpic0, &weight0, mv1, rpic1, &weight1, mc_dsp);
},
MBType::B16x8(mode0, mode1) | MBType::B8x16(mode0, mode1) => {
let (pw, ph) = mb_info.mb_type.size();
let mv1 = sstate.get_cur_blk4(blk).mv[1];
let rpic1 = frame_refs.select_ref_pic(1, mb_info.ref_l1[part].index());
let (weight0, weight1) = get_weights(slice_hdr, frame_refs, modes[part], weight_mode, mb_info.ref_l0[part], mb_info.ref_l1[part]);
- do_b_mc(frm, modes[part], xpos + bx, ypos + by, pw, ph, mv0, rpic0, &weight0, mv1, rpic1, &weight1, avg_buf);
+ do_b_mc(frm, modes[part], xpos + bx, ypos + by, pw, ph, mv0, rpic0, &weight0, mv1, rpic1, &weight1, mc_dsp);
bx += px;
by += py;
}
let rpic0 = frame_refs.select_ref_pic(0, ref_idx[0].index());
let rpic1 = frame_refs.select_ref_pic(1, ref_idx[1].index());
let (weight0, weight1) = get_weights(slice_hdr, frame_refs, BMode::Bi, weight_mode, ref_idx[0], ref_idx[1]);
- do_b_mc(frm, BMode::Bi, xpos, ypos, 16, 16, mv[0], rpic0, &weight0, mv[1], rpic1, &weight1, avg_buf);
+ do_b_mc(frm, BMode::Bi, xpos, ypos, 16, 16, mv[0], rpic0, &weight0, mv[1], rpic1, &weight1, mc_dsp);
} else {
for blk4 in 0..16 {
let mv = sstate.get_cur_blk4(blk4).mv;
let rpic0 = frame_refs.select_ref_pic(0, ref_idx[0].index());
let rpic1 = frame_refs.select_ref_pic(1, ref_idx[1].index());
let (weight0, weight1) = get_weights(slice_hdr, frame_refs, BMode::Bi, weight_mode, ref_idx[0], ref_idx[1]);
- do_b_mc(frm, BMode::Bi, xpos + (blk4 & 3) * 4, ypos + (blk4 >> 2) * 4, 4, 4, mv[0], rpic0, &weight0, mv[1], rpic1, &weight1, avg_buf);
+ do_b_mc(frm, BMode::Bi, xpos + (blk4 & 3) * 4, ypos + (blk4 >> 2) * 4, 4, 4, mv[0], rpic0, &weight0, mv[1], rpic1, &weight1, mc_dsp);
+ }
+ }
+ sstate.apply_to_blk8(|blk8| { blk8.ref_idx[0].set_direct(); blk8.ref_idx[1].set_direct(); });
+ },
+ MBType::B8x8 => {
+ for part in 0..4 {
+ let ridx = sstate.get_cur_blk8(part).ref_idx;
+ let rpic0 = frame_refs.select_ref_pic(0, ridx[0].index());
+ let rpic1 = frame_refs.select_ref_pic(1, ridx[1].index());
+ let subtype = mb_info.sub_mb_type[part];
+ let blk8 = (part & 1) * 2 + (part & 2) * 4;
+ let mut bx = (part & 1) * 8;
+ let mut by = (part & 2) * 4;
+ match subtype {
+ SubMBType::Direct8x8 => {
+ for blk in 0..4 {
+ let mv = sstate.get_cur_blk4(bx / 4 + (by / 4) * 4).mv;
+ let ref_idx = sstate.get_cur_blk8(bx / 8 + (by / 8) * 2).ref_idx;
+ let rpic0 = frame_refs.select_ref_pic(0, ref_idx[0].index());
+ let rpic1 = frame_refs.select_ref_pic(1, ref_idx[1].index());
+ let (weight0, weight1) = get_weights(slice_hdr, frame_refs, BMode::Bi, weight_mode, ref_idx[0], ref_idx[1]);
+ do_b_mc(frm, BMode::Bi, xpos + bx, ypos + by, 4, 4, mv[0], rpic0, &weight0, mv[1], rpic1, &weight1, mc_dsp);
+ bx += 4;
+ if blk == 1 {
+ bx -= 8;
+ by += 4;
+ }
+ }
+ sstate.get_cur_blk8(part).ref_idx[0].set_direct();
+ sstate.get_cur_blk8(part).ref_idx[1].set_direct();
+ },
+ SubMBType::B8x8(mode) => {
+ let mv = sstate.get_cur_blk4(blk8).mv;
+ let (weight0, weight1) = get_weights(slice_hdr, frame_refs, mode, weight_mode, ridx[0], ridx[1]);
+ do_b_mc(frm, mode, xpos + bx, ypos + by, 8, 8, mv[0], rpic0, &weight0, mv[1], rpic1, &weight1, mc_dsp);
+ },
+ SubMBType::B8x4(mode) | SubMBType::B4x8(mode) => {
+ let (weight0, weight1) = get_weights(slice_hdr, frame_refs, mode, weight_mode, ridx[0], ridx[1]);
+ let (pw, ph) = subtype.size();
+ let mv = sstate.get_cur_blk4(blk8).mv;
+ do_b_mc(frm, mode, xpos + bx, ypos + by, pw, ph, mv[0], rpic0.clone(), &weight0, mv[1], rpic1.clone(), &weight1, mc_dsp);
+ let addr2 = blk8 + (pw & 4) / 4 + (ph & 4);
+ let mv = sstate.get_cur_blk4(addr2).mv;
+ do_b_mc(frm, mode, xpos + bx + (pw & 4), ypos + by + (ph & 4), pw, ph, mv[0], rpic0, &weight0, mv[1], rpic1, &weight1, mc_dsp);
+ },
+ SubMBType::B4x4(mode) => {
+ let (weight0, weight1) = get_weights(slice_hdr, frame_refs, mode, weight_mode, ridx[0], ridx[1]);
+ for i in 0..4 {
+ let addr2 = blk8 + (i & 1) + (i & 2) * 2;
+ let mv = sstate.get_cur_blk4(addr2).mv;
+ do_b_mc(frm, mode, xpos + bx, ypos + by, 4, 4, mv[0], rpic0.clone(), &weight0, mv[1], rpic1.clone(), &weight1, mc_dsp);
+ bx += 4;
+ if i == 1 {
+ bx -= 8;
+ by += 4;
+ }
+ }
+ },
+ _ => unreachable!(),
+ };
+ }
+ },
+ };
+ if !mb_info.mb_type.is_skip() {
+ if mb_info.mb_type != MBType::Intra4x4 && mb_info.mb_type != MBType::Intra8x8 {
+ add_luma(frm, sstate, mb_info);
+ }
+ add_chroma(frm, sstate, mb_info);
+ }
+}
+
+pub fn wait_for_mb(disp: &Shareable<ThreadDispatcher>, sstate: &SliceState, xpos: usize, ypos: usize, mv: MV, ref_id: u32) -> DecoderResult<()> {
+ let xpos = xpos as isize + ((mv.x >> 2) as isize) + 4;
+ let ypos = ypos as isize + ((mv.y >> 2) as isize) + 4;
+ let dst_mb_x = ((xpos.max(0) as usize) / 16).min(sstate.mb_w - 1);
+ let dst_mb_y = ((ypos.max(0) as usize) / 16).min(sstate.mb_h - 1);
+ let expected_mb = dst_mb_x + dst_mb_y * sstate.mb_w;
+ loop {
+ if let Ok(ds) = disp.read() {
+ match ds.check_pos(ref_id, expected_mb) {
+ FrameDecodingStatus::Ok => return Ok(()),
+ FrameDecodingStatus::NotReady => {},
+ _ => return Err(DecoderError::MissingReference),
+ };
+ }
+ std::thread::yield_now();
+ }
+}
+
+fn wait_b_mc(disp: &Shareable<ThreadDispatcher>, sstate: &SliceState, frame_refs: &SliceRefs, mv: [MV; 2], ref_idx: [PicRef; 2], xpos: usize, ypos: usize, w: usize, h: usize) -> DecoderResult<()> {
+ if let Some(ref_id) = frame_refs.get_ref_id(0, ref_idx[0].index()) {
+ wait_for_mb(disp, sstate, xpos + w, ypos + h, mv[0], ref_id)?;
+ }
+ if let Some(ref_id) = frame_refs.get_ref_id(1, ref_idx[1].index()) {
+ wait_for_mb(disp, sstate, xpos + w, ypos + h, mv[1], ref_id)?;
+ }
+ Ok(())
+}
+
+pub fn recon_mb_mt(frm: &mut NASimpleVideoFrame<u8>, slice_hdr: &SliceHeader, mb_info: &CurrentMBInfo, sstate: &mut SliceState, frame_refs: &SliceRefs, mc_dsp: &mut H264MC, weight_mode: u8, disp: &Shareable<ThreadDispatcher>) -> DecoderResult<()> {
+ let xpos = sstate.mb_x * 16;
+ let ypos = sstate.mb_y * 16;
+
+ match mb_info.mb_type {
+ MBType::Intra16x16(_, _, _) => {
+ pred_intra(frm, sstate, mb_info);
+ },
+ MBType::Intra4x4 | MBType::Intra8x8 => {
+ pred_intra(frm, sstate, mb_info);
+ },
+ MBType::PCM => {},
+ MBType::PSkip => {
+ let mv = sstate.get_cur_blk4(0).mv[0];
+ if let Some(ref_id) = frame_refs.get_ref_id(0, 0) {
+ wait_for_mb(disp, sstate, xpos + 16, ypos + 16, mv, ref_id)?;
+ }
+ let rpic = frame_refs.select_ref_pic(0, 0);
+ let weight = &slice_hdr.get_weight(0, 0);
+ do_p_mc(frm, xpos, ypos, 16, 16, mv, rpic, weight, mc_dsp);
+ },
+ MBType::P16x16 => {
+ let mv = sstate.get_cur_blk4(0).mv[0];
+ if let Some(ref_id) = frame_refs.get_ref_id(0, mb_info.ref_l0[0].index()) {
+ wait_for_mb(disp, sstate, xpos + 16, ypos + 16, mv, ref_id)?;
+ }
+ let rpic = frame_refs.select_ref_pic(0, mb_info.ref_l0[0].index());
+ let weight = &slice_hdr.get_weight(0, mb_info.ref_l0[0].index());
+ do_p_mc(frm, xpos, ypos, 16, 16, mv, rpic, weight, mc_dsp);
+ },
+ MBType::P16x8 | MBType::P8x16 => {
+ let (bw, bh, bx, by) = if mb_info.mb_type == MBType::P16x8 {
+ (16, 8, 0, 8)
+ } else {
+ (8, 16, 8, 0)
+ };
+ let mv = sstate.get_cur_blk4(0).mv[0];
+ if let Some(ref_id) = frame_refs.get_ref_id(0, mb_info.ref_l0[0].index()) {
+ wait_for_mb(disp, sstate, xpos + bw, ypos + bh, mv, ref_id)?;
+ }
+ let rpic = frame_refs.select_ref_pic(0, mb_info.ref_l0[0].index());
+ let weight = &slice_hdr.get_weight(0, mb_info.ref_l0[0].index());
+ do_p_mc(frm, xpos, ypos, bw, bh, mv, rpic, weight, mc_dsp);
+ let mv = sstate.get_cur_blk4(bx / 4 + by).mv[0];
+ if let Some(ref_id) = frame_refs.get_ref_id(0, mb_info.ref_l0[1].index()) {
+ wait_for_mb(disp, sstate, xpos + 16, ypos + 16, mv, ref_id)?;
+ }
+ let rpic = frame_refs.select_ref_pic(0, mb_info.ref_l0[1].index());
+ let weight = &slice_hdr.get_weight(0, mb_info.ref_l0[1].index());
+ do_p_mc(frm, xpos + bx, ypos + by, bw, bh, mv, rpic, weight, mc_dsp);
+ },
+ MBType::P8x8 | MBType::P8x8Ref0 => {
+ for part in 0..4 {
+ let bx = (part & 1) * 8;
+ let by = (part & 2) * 4;
+ let mv = sstate.get_cur_blk4(bx / 4 + by).mv[0];
+ let rpic = frame_refs.select_ref_pic(0, mb_info.ref_l0[part].index());
+ let weight = &slice_hdr.get_weight(0, mb_info.ref_l0[part].index());
+
+ match mb_info.sub_mb_type[part] {
+ SubMBType::P8x8 => {
+ if let Some(ref_id) = frame_refs.get_ref_id(0, mb_info.ref_l0[part].index()) {
+ wait_for_mb(disp, sstate, xpos + bx + 8, ypos + by + 8, mv, ref_id)?;
+ }
+ do_p_mc(frm, xpos + bx, ypos + by, 8, 8, mv, rpic, weight, mc_dsp);
+ },
+ SubMBType::P8x4 => {
+ if let Some(ref_id) = frame_refs.get_ref_id(0, mb_info.ref_l0[part].index()) {
+ wait_for_mb(disp, sstate, xpos + bx + 8, ypos + by + 4, mv, ref_id)?;
+ }
+ do_p_mc(frm, xpos + bx, ypos + by, 8, 4, mv, rpic.clone(), weight, mc_dsp);
+ let mv = sstate.get_cur_blk4(bx / 4 + by + 4).mv[0];
+ if let Some(ref_id) = frame_refs.get_ref_id(0, mb_info.ref_l0[part].index()) {
+ wait_for_mb(disp, sstate, xpos + bx + 8, ypos + by + 8, mv, ref_id)?;
+ }
+ do_p_mc(frm, xpos + bx, ypos + by + 4, 8, 4, mv, rpic, weight, mc_dsp);
+ },
+ SubMBType::P4x8 => {
+ if let Some(ref_id) = frame_refs.get_ref_id(0, mb_info.ref_l0[part].index()) {
+ wait_for_mb(disp, sstate, xpos + bx + 4, ypos + by + 8, mv, ref_id)?;
+ }
+ do_p_mc(frm, xpos + bx, ypos + by, 4, 8, mv, rpic.clone(), weight, mc_dsp);
+ let mv = sstate.get_cur_blk4(bx / 4 + by + 1).mv[0];
+ if let Some(ref_id) = frame_refs.get_ref_id(0, mb_info.ref_l0[part].index()) {
+ wait_for_mb(disp, sstate, xpos + bx + 8, ypos + by + 8, mv, ref_id)?;
+ }
+ do_p_mc(frm, xpos + bx + 4, ypos + by, 4, 8, mv, rpic, weight, mc_dsp);
+ },
+ SubMBType::P4x4 => {
+ for sb_no in 0..4 {
+ let sxpos = xpos + bx + (sb_no & 1) * 4;
+ let sypos = ypos + by + (sb_no & 2) * 2;
+ let sblk_no = (bx / 4 + (sb_no & 1)) + ((by / 4) + (sb_no >> 1)) * 4;
+ let mv = sstate.get_cur_blk4(sblk_no).mv[0];
+ if let Some(ref_id) = frame_refs.get_ref_id(0, mb_info.ref_l0[part].index()) {
+ wait_for_mb(disp, sstate, sxpos + 4, sypos + 4, mv, ref_id)?;
+ }
+ do_p_mc(frm, sxpos, sypos, 4, 4, mv, rpic.clone(), weight, mc_dsp);
+ }
+ },
+ _ => unreachable!(),
+ };
+ }
+ },
+ MBType::B16x16(mode) => {
+ let mv0 = sstate.get_cur_blk4(0).mv[0];
+ let rpic0 = frame_refs.select_ref_pic(0, mb_info.ref_l0[0].index());
+ let mv1 = sstate.get_cur_blk4(0).mv[1];
+ let rpic1 = frame_refs.select_ref_pic(1, mb_info.ref_l1[0].index());
+ let (weight0, weight1) = get_weights(slice_hdr, frame_refs, mode, weight_mode, mb_info.ref_l0[0], mb_info.ref_l1[0]);
+ wait_b_mc(disp, sstate, frame_refs, [mv0, mv1], [mb_info.ref_l0[0], mb_info.ref_l1[0]], xpos, ypos, 16, 16)?;
+ do_b_mc(frm, mode, xpos, ypos, 16, 16, mv0, rpic0, &weight0, mv1, rpic1, &weight1, mc_dsp);
+ },
+ MBType::B16x8(mode0, mode1) | MBType::B8x16(mode0, mode1) => {
+ let (pw, ph) = mb_info.mb_type.size();
+ let (px, py) = (pw & 8, ph & 8);
+ let modes = [mode0, mode1];
+ let (mut bx, mut by) = (0, 0);
+ for part in 0..2 {
+ let blk = if part == 0 { 0 } else { (px / 4) + py };
+ let mv0 = sstate.get_cur_blk4(blk).mv[0];
+ let rpic0 = frame_refs.select_ref_pic(0, mb_info.ref_l0[part].index());
+ let mv1 = sstate.get_cur_blk4(blk).mv[1];
+ let rpic1 = frame_refs.select_ref_pic(1, mb_info.ref_l1[part].index());
+ let (weight0, weight1) = get_weights(slice_hdr, frame_refs, modes[part], weight_mode, mb_info.ref_l0[part], mb_info.ref_l1[part]);
+ wait_b_mc(disp, sstate, frame_refs, [mv0, mv1], [mb_info.ref_l0[part], mb_info.ref_l1[part]], xpos + bx, ypos + by, pw, ph)?;
+ do_b_mc(frm, modes[part], xpos + bx, ypos + by, pw, ph, mv0, rpic0, &weight0, mv1, rpic1, &weight1, mc_dsp);
+ bx += px;
+ by += py;
+ }
+ },
+ MBType::Direct | MBType::BSkip => {
+ if let Some(ref_id) = frame_refs.get_ref_id(1, mb_info.ref_l1[0].index()) {
+ wait_for_mb(disp, sstate, xpos, ypos, ZERO_MV, ref_id)?;
+ }
+ let colo_mb_type = frame_refs.get_colocated_info(sstate.mb_x, sstate.mb_y).0.mb_type;
+ let is_16x16 = colo_mb_type.is_16x16_ref();
+
+ if is_16x16 {
+ let mv = sstate.get_cur_blk4(0).mv;
+ let ref_idx = sstate.get_cur_blk8(0).ref_idx;
+ let rpic0 = frame_refs.select_ref_pic(0, ref_idx[0].index());
+ let rpic1 = frame_refs.select_ref_pic(1, ref_idx[1].index());
+ if let Some(ref_id) = frame_refs.get_ref_id(0, mb_info.ref_l0[0].index()) {
+ wait_for_mb(disp, sstate, xpos + 16, ypos + 16, mv[0], ref_id)?;
+ }
+ if let Some(ref_id) = frame_refs.get_ref_id(1, mb_info.ref_l1[0].index()) {
+ wait_for_mb(disp, sstate, xpos + 16, ypos + 16, mv[1], ref_id)?;
+ }
+ let (weight0, weight1) = get_weights(slice_hdr, frame_refs, BMode::Bi, weight_mode, ref_idx[0], ref_idx[1]);
+ wait_b_mc(disp, sstate, frame_refs, mv, ref_idx, xpos, ypos, 16, 16)?;
+ do_b_mc(frm, BMode::Bi, xpos, ypos, 16, 16, mv[0], rpic0, &weight0, mv[1], rpic1, &weight1, mc_dsp);
+ } else {
+ for blk4 in 0..16 {
+ let mv = sstate.get_cur_blk4(blk4).mv;
+ let ref_idx = sstate.get_cur_blk8(blk4_to_blk8(blk4)).ref_idx;
+ if let Some(ref_id) = frame_refs.get_ref_id(0, ref_idx[0].index()) {
+ wait_for_mb(disp, sstate, xpos + 16, ypos + 16, mv[0], ref_id)?;
+ }
+ if let Some(ref_id) = frame_refs.get_ref_id(1, ref_idx[1].index()) {
+ wait_for_mb(disp, sstate, xpos + 16, ypos + 16, mv[1], ref_id)?;
+ }
+ let rpic0 = frame_refs.select_ref_pic(0, ref_idx[0].index());
+ let rpic1 = frame_refs.select_ref_pic(1, ref_idx[1].index());
+ let (weight0, weight1) = get_weights(slice_hdr, frame_refs, BMode::Bi, weight_mode, ref_idx[0], ref_idx[1]);
+ wait_b_mc(disp, sstate, frame_refs, mv, ref_idx, xpos + (blk4 & 3) * 4, ypos + (blk4 >> 2) * 4, 4, 4)?;
+ do_b_mc(frm, BMode::Bi, xpos + (blk4 & 3) * 4, ypos + (blk4 >> 2) * 4, 4, 4, mv[0], rpic0, &weight0, mv[1], rpic1, &weight1, mc_dsp);
}
}
sstate.apply_to_blk8(|blk8| { blk8.ref_idx[0].set_direct(); blk8.ref_idx[1].set_direct(); });
let rpic0 = frame_refs.select_ref_pic(0, ref_idx[0].index());
let rpic1 = frame_refs.select_ref_pic(1, ref_idx[1].index());
let (weight0, weight1) = get_weights(slice_hdr, frame_refs, BMode::Bi, weight_mode, ref_idx[0], ref_idx[1]);
- do_b_mc(frm, BMode::Bi, xpos + bx, ypos + by, 4, 4, mv[0], rpic0, &weight0, mv[1], rpic1, &weight1, avg_buf);
+ wait_b_mc(disp, sstate, frame_refs, mv, ref_idx, xpos + bx, ypos + by, 4, 4)?;
+ do_b_mc(frm, BMode::Bi, xpos + bx, ypos + by, 4, 4, mv[0], rpic0, &weight0, mv[1], rpic1, &weight1, mc_dsp);
bx += 4;
if blk == 1 {
bx -= 8;
SubMBType::B8x8(mode) => {
let mv = sstate.get_cur_blk4(blk8).mv;
let (weight0, weight1) = get_weights(slice_hdr, frame_refs, mode, weight_mode, ridx[0], ridx[1]);
- do_b_mc(frm, mode, xpos + bx, ypos + by, 8, 8, mv[0], rpic0, &weight0, mv[1], rpic1, &weight1, avg_buf);
+ wait_b_mc(disp, sstate, frame_refs, mv, ridx, xpos + bx, ypos + by, 8, 8)?;
+ do_b_mc(frm, mode, xpos + bx, ypos + by, 8, 8, mv[0], rpic0, &weight0, mv[1], rpic1, &weight1, mc_dsp);
},
SubMBType::B8x4(mode) | SubMBType::B4x8(mode) => {
let (weight0, weight1) = get_weights(slice_hdr, frame_refs, mode, weight_mode, ridx[0], ridx[1]);
let (pw, ph) = subtype.size();
let mv = sstate.get_cur_blk4(blk8).mv;
- do_b_mc(frm, mode, xpos + bx, ypos + by, pw, ph, mv[0], rpic0.clone(), &weight0, mv[1], rpic1.clone(), &weight1, avg_buf);
+ wait_b_mc(disp, sstate, frame_refs, mv, ridx, xpos + bx, ypos + by, pw, ph)?;
+ do_b_mc(frm, mode, xpos + bx, ypos + by, pw, ph, mv[0], rpic0.clone(), &weight0, mv[1], rpic1.clone(), &weight1, mc_dsp);
let addr2 = blk8 + (pw & 4) / 4 + (ph & 4);
let mv = sstate.get_cur_blk4(addr2).mv;
- do_b_mc(frm, mode, xpos + bx + (pw & 4), ypos + by + (ph & 4), pw, ph, mv[0], rpic0, &weight0, mv[1], rpic1, &weight1, avg_buf);
+ wait_b_mc(disp, sstate, frame_refs, mv, ridx, xpos + bx + (pw & 4), ypos + by + (ph & 4), pw, ph)?;
+ do_b_mc(frm, mode, xpos + bx + (pw & 4), ypos + by + (ph & 4), pw, ph, mv[0], rpic0, &weight0, mv[1], rpic1, &weight1, mc_dsp);
},
SubMBType::B4x4(mode) => {
let (weight0, weight1) = get_weights(slice_hdr, frame_refs, mode, weight_mode, ridx[0], ridx[1]);
for i in 0..4 {
let addr2 = blk8 + (i & 1) + (i & 2) * 2;
let mv = sstate.get_cur_blk4(addr2).mv;
- do_b_mc(frm, mode, xpos + bx, ypos + by, 4, 4, mv[0], rpic0.clone(), &weight0, mv[1], rpic1.clone(), &weight1, avg_buf);
+ wait_b_mc(disp, sstate, frame_refs, mv, ridx, xpos + bx, ypos + by, 4, 4)?;
+ do_b_mc(frm, mode, xpos + bx, ypos + by, 4, 4, mv[0], rpic0.clone(), &weight0, mv[1], rpic1.clone(), &weight1, mc_dsp);
bx += 4;
if i == 1 {
bx -= 8;
};
if !mb_info.mb_type.is_skip() {
if mb_info.mb_type != MBType::Intra4x4 && mb_info.mb_type != MBType::Intra8x8 {
- add_luma(frm, &sstate, &mb_info);
+ add_luma(frm, sstate, mb_info);
}
- add_chroma(frm, &sstate, &mb_info);
+ add_chroma(frm, sstate, mb_info);
}
+ Ok(())
}
projects / nihav.git / blobdiff