use super::{CurrentMBInfo, I4X4_SCAN, Shareable};
use super::dispatch::{ThreadDispatcher, FrameDecodingStatus};
use super::dsp::*;
-use super::pic_ref::SliceRefs;
+use super::pic_ref::SimplifiedSliceRefs;
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, mc_dsp: &mut H264MC) {
+fn do_p_mc(frm: &mut NASimpleVideoFrame<u8>, xpos: usize, ypos: usize, w: usize, h: usize, mv: MV, ref_pic: Option<&SimpleFrame>, weight: &WeightInfo, mc_dsp: &mut H264MC) {
if let Some(buf) = ref_pic {
if !weight.is_weighted() {
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_dsp.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 {
8 => 2,
_ => 3,
};
- (mc_dsp.put_block_weighted[wmode])(&mut frm.data[yoff..], frm.stride[0], &ytmp, h, yw);
+ (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 };
+ 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);
}
}
}
#[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) {
+fn do_b_mc(frm: &mut NASimpleVideoFrame<u8>, mode: BMode, xpos: usize, ypos: usize, w: usize, h: usize, mv0: MV, ref_pic0: Option<&SimpleFrame>, weight0: &WeightInfo, mv1: MV, ref_pic1: Option<&SimpleFrame>, 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,
},
};
} 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_dsp.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 {
8 => 2,
_ => 3,
};
- (mc_dsp.put_block_weighted[wmode])(&mut frm.data[yoff..], frm.stride[0], &ytmp0, h, yw);
+ (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 };
+ 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_dsp.mc_blocks(&mut ytmp0, &mut utmp0, &mut vtmp0, buf0, xpos, ypos, w, h, mv0);
- mc_dsp.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) {
8 => 2,
_ => 3,
};
- (mc_dsp.put_block_weighted2[wmode])(&mut frm.data[yoff..], frm.stride[0], &ytmp0, &ytmp1, h, yw);
+ (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 };
+ 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);
}
},
}
}
-fn get_weights(slice_hdr: &SliceHeader, frame_refs: &SliceRefs, mode: BMode, weight_mode: u8, ref_l0: PicRef, ref_l1: PicRef) -> (WeightInfo, WeightInfo) {
+fn do_b_mc_4x4bi(frm: &mut NASimpleVideoFrame<u8>, xpos: usize, ypos: usize, mv: &[MV; 2], ref_pic0: Option<&SimpleFrame>, weight0: &WeightInfo, ref_pic1: Option<&SimpleFrame>, weight1: &WeightInfo, mc_dsp: &mut H264MC) {
+ if !weight0.is_weighted() || !weight1.is_weighted() {
+ match (ref_pic0, ref_pic1) {
+ (Some(buf0), Some(buf1)) => {
+ mc_dsp.do_mc(frm, buf0, xpos, ypos, 4, 4, mv[0]);
+ mc_dsp.do_mc_avg(frm, buf1, xpos, ypos, 4, 4, mv[1]);
+ },
+ (Some(buf0), None) => {
+ mc_dsp.do_mc(frm, buf0, xpos, ypos, 4, 4, mv[0]);
+ },
+ (None, Some(buf1)) => {
+ mc_dsp.do_mc(frm, buf1, xpos, ypos, 4, 4, mv[1]);
+ },
+ (None, None) => {
+ mc_dsp.gray_block(frm, xpos, ypos, 4, 4);
+ },
+ };
+ } else {
+ let mut tmp0 = McBlock::new();
+ let mut tmp1 = McBlock::new();
+ match (ref_pic0, ref_pic1) {
+ (Some(buf0), Some(buf1)) => { // do both and avg
+ mc_dsp.mc_blocks(&mut tmp0, buf0, xpos, ypos, 4, 4, mv[0]);
+ mc_dsp.mc_blocks(&mut tmp1, buf1, xpos, ypos, 4, 4, mv[1]);
+
+ let yoff = frm.offset[0] + xpos + ypos * frm.stride[0];
+ let yw = match (weight0.luma_weighted, weight1.luma_weighted) {
+ (true, true) => [weight0.luma_weight, weight0.luma_offset, weight1.luma_weight, weight1.luma_offset, weight0.luma_shift as i8],
+ (true, false) => [weight0.luma_weight, weight0.luma_offset, 1 << weight0.luma_shift, 0, weight0.luma_shift as i8],
+ (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],
+ };
+ (mc_dsp.put_block_weighted2[1])(&mut frm.data[yoff..], frm.stride[0], &tmp0.y, &tmp1.y, 4, yw);
+
+ for chroma in 0..2 {
+ let cstride = frm.stride[chroma + 1];
+ let coff = frm.offset[chroma + 1] + xpos / 2 + ypos / 2 * cstride;
+ let cw0 = weight0.chroma_weight[chroma];
+ let co0 = weight0.chroma_offset[chroma];
+ let cw1 = weight1.chroma_weight[chroma];
+ let co1 = weight1.chroma_offset[chroma];
+ let cw = match (weight0.chroma_weighted, weight1.chroma_weighted) {
+ (true, true) => [cw0, co0, cw1, co1, weight0.luma_shift as i8],
+ (true, false) => [cw0, co0, 1 << weight0.luma_shift, 0, weight0.luma_shift as i8],
+ (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 { &tmp0.u } else { &tmp0.v };
+ let csrc1 = if chroma == 0 { &tmp1.u } else { &tmp1.v };
+ (mc_dsp.put_block_weighted2[0])(&mut frm.data[coff..], cstride, csrc0, csrc1, 2, cw);
+ }
+ },
+ _ => {
+ mc_dsp.gray_block(frm, xpos, ypos, 4, 4);
+ },
+ };
+ }
+}
+
+fn get_weights(slice_hdr: &SliceHeader, frame_refs: &SimplifiedSliceRefs, 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 {
let r0_poc = pic0.full_id as u16;
let r1_poc = pic1.full_id as u16;
let cur_id = frame_refs.cur_id as u16;
- if (r0_poc == r1_poc) || pic0.long_term.is_some() || pic1.long_term.is_some() {
+ if (r0_poc == r1_poc) || pic0.long_term || pic1.long_term {
return (DEF_WEIGHT_INFO, DEF_WEIGHT_INFO);
}
}
}
-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) {
+pub fn recon_mb(frm: &mut NASimpleVideoFrame<u8>, slice_hdr: &SliceHeader, mb_info: &CurrentMBInfo, sstate: &mut SliceState, frame_refs: &SimplifiedSliceRefs, mc_dsp: &mut H264MC, weight_mode: u8) {
let xpos = sstate.mb_x * 16;
let ypos = sstate.mb_y * 16;
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, mc_dsp);
+ do_p_mc(frm, xpos + bx, ypos + by, 8, 4, mv, rpic, 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, mc_dsp);
},
SubMBType::P4x8 => {
- do_p_mc(frm, xpos + bx, ypos + by, 4, 8, mv, rpic.clone(), weight, mc_dsp);
+ do_p_mc(frm, xpos + bx, ypos + by, 4, 8, mv, rpic, 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, mc_dsp);
},
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, mc_dsp);
+ do_p_mc(frm, sxpos, sypos, 4, 4, mv, rpic, weight, mc_dsp);
}
},
_ => unreachable!(),
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;
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, mc_dsp);
+ let mv = &sstate.get_cur_blk4(blk4).mv;
+ do_b_mc_4x4bi(frm, xpos + (blk4 & 3) * 4, ypos + (blk4 >> 2) * 4, mv, rpic0, &weight0, rpic1, &weight1, mc_dsp);
}
}
sstate.apply_to_blk8(|blk8| { blk8.ref_idx[0].set_direct(); blk8.ref_idx[1].set_direct(); });
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);
+ let mv = &sstate.get_cur_blk4(bx / 4 + (by / 4) * 4).mv;
+ do_b_mc_4x4bi(frm, xpos + bx, ypos + by, mv, rpic0, &weight0, rpic1, &weight1, mc_dsp);
bx += 4;
if blk == 1 {
bx -= 8;
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);
+ do_b_mc(frm, mode, xpos + bx, ypos + by, pw, ph, mv[0], rpic0, &weight0, mv[1], rpic1, &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);
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);
+ do_b_mc(frm, mode, xpos + bx, ypos + by, 4, 4, mv[0], rpic0, &weight0, mv[1], rpic1, &weight1, mc_dsp);
bx += 4;
if i == 1 {
bx -= 8;
}
}
-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<()> {
+fn wait_b_mc(disp: &Shareable<ThreadDispatcher>, sstate: &SliceState, frame_refs: &SimplifiedSliceRefs, 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)?;
}
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<()> {
+pub fn recon_mb_mt(frm: &mut NASimpleVideoFrame<u8>, slice_hdr: &SliceHeader, mb_info: &CurrentMBInfo, sstate: &mut SliceState, frame_refs: &SimplifiedSliceRefs, mc_dsp: &mut H264MC, weight_mode: u8, disp: &Shareable<ThreadDispatcher>) -> DecoderResult<()> {
let xpos = sstate.mb_x * 16;
let ypos = sstate.mb_y * 16;
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);
+ do_p_mc(frm, xpos + bx, ypos + by, 8, 4, mv, rpic, 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)?;
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);
+ do_p_mc(frm, xpos + bx, ypos + by, 4, 8, mv, rpic, 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)?;
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);
+ do_p_mc(frm, sxpos, sypos, 4, 4, mv, rpic, weight, mc_dsp);
}
},
_ => unreachable!(),
let (pw, ph) = subtype.size();
let mv = sstate.get_cur_blk4(blk8).mv;
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);
+ do_b_mc(frm, mode, xpos + bx, ypos + by, pw, ph, mv[0], rpic0, &weight0, mv[1], rpic1, &weight1, mc_dsp);
let addr2 = blk8 + (pw & 4) / 4 + (ph & 4);
let mv = sstate.get_cur_blk4(addr2).mv;
wait_b_mc(disp, sstate, frame_refs, mv, ridx, xpos + bx + (pw & 4), ypos + by + (ph & 4), pw, ph)?;
let addr2 = blk8 + (i & 1) + (i & 2) * 2;
let mv = sstate.get_cur_blk4(addr2).mv;
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);
+ do_b_mc(frm, mode, xpos + bx, ypos + by, 4, 4, mv[0], rpic0, &weight0, mv[1], rpic1, &weight1, mc_dsp);
bx += 4;
if i == 1 {
bx -= 8;
projects / nihav.git / blobdiff