X-Git-Url: https://git.nihav.org/?p=nihav.git;a=blobdiff_plain;f=nihav-itu%2Fsrc%2Fcodecs%2Fh264%2Fmb_recon.rs;fp=nihav-itu%2Fsrc%2Fcodecs%2Fh264%2Fmb_recon.rs;h=0a63edf5d28b9b4f7cd307f9b947a459bef6b205;hp=eb32792d6d8369ef65864766a48ae45ab5f2197c;hb=11d7aef2ac58fe8d3341ffdcec7eaf9ee4e266f1;hpb=93839abd64a4c725e0194a790ffd743dd8c19916 diff --git a/nihav-itu/src/codecs/h264/mb_recon.rs b/nihav-itu/src/codecs/h264/mb_recon.rs index eb32792..0a63edf 100644 --- a/nihav-itu/src/codecs/h264/mb_recon.rs +++ b/nihav-itu/src/codecs/h264/mb_recon.rs @@ -1,6 +1,8 @@ +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::SliceRefs; use super::slice::{SliceHeader, WeightInfo, DEF_WEIGHT_INFO}; @@ -606,3 +608,275 @@ pub fn recon_mb(frm: &mut NASimpleVideoFrame, slice_hdr: &SliceHeader, mb_in add_chroma(frm, sstate, mb_info); } } + +pub fn wait_for_mb(disp: &Shareable, 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, 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, slice_hdr: &SliceHeader, mb_info: &CurrentMBInfo, sstate: &mut SliceState, frame_refs: &SliceRefs, mc_dsp: &mut H264MC, weight_mode: u8, disp: &Shareable) -> 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(); }); + }, + 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]); + 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; + 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]); + 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; + 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; + 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; + 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; + 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); + } + Ok(()) +}