X-Git-Url: https://git.nihav.org/?a=blobdiff_plain;f=nihav-itu%2Fsrc%2Fcodecs%2Fh264%2Fmb_recon.rs;h=5d8250396fc36f44d8630ed2b7dbca7fdbf24725;hb=cc818ce8b5379d93fb5059cc50f7ff02db871dce;hp=0a63edf5d28b9b4f7cd307f9b947a459bef6b205;hpb=11d7aef2ac58fe8d3341ffdcec7eaf9ee4e266f1;p=nihav.git diff --git a/nihav-itu/src/codecs/h264/mb_recon.rs b/nihav-itu/src/codecs/h264/mb_recon.rs index 0a63edf..5d82503 100644 --- a/nihav-itu/src/codecs/h264/mb_recon.rs +++ b/nihav-itu/src/codecs/h264/mb_recon.rs @@ -210,10 +210,8 @@ fn do_p_mc(frm: &mut NASimpleVideoFrame, xpos: usize, ypos: usize, w: usize, 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 { @@ -227,7 +225,7 @@ fn do_p_mc(frm: &mut NASimpleVideoFrame, xpos: usize, ypos: usize, w: usize, 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]; @@ -237,7 +235,7 @@ fn do_p_mc(frm: &mut NASimpleVideoFrame, xpos: usize, ypos: usize, w: usize, } 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); } } @@ -289,16 +287,12 @@ fn do_b_mc(frm: &mut NASimpleVideoFrame, mode: BMode, xpos: usize, ypos: usi }, }; } 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 { @@ -312,7 +306,7 @@ fn do_b_mc(frm: &mut NASimpleVideoFrame, mode: BMode, xpos: usize, ypos: usi 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]; @@ -322,13 +316,13 @@ fn do_b_mc(frm: &mut NASimpleVideoFrame, mode: BMode, xpos: usize, ypos: usi } 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) { @@ -343,7 +337,7 @@ fn do_b_mc(frm: &mut NASimpleVideoFrame, mode: BMode, xpos: usize, ypos: usi 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]; @@ -358,8 +352,8 @@ fn do_b_mc(frm: &mut NASimpleVideoFrame, mode: BMode, xpos: usize, ypos: usi (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); } }, @@ -370,6 +364,65 @@ fn do_b_mc(frm: &mut NASimpleVideoFrame, mode: BMode, xpos: usize, ypos: usi } } +fn do_b_mc_4x4bi(frm: &mut NASimpleVideoFrame, xpos: usize, ypos: usize, mv: [MV; 2], ref_pic0: Option>, weight0: &WeightInfo, ref_pic1: Option>, 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: &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(); @@ -537,7 +590,7 @@ pub fn recon_mb(frm: &mut NASimpleVideoFrame, slice_hdr: &SliceHeader, mb_in 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); + 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(); }); @@ -559,7 +612,7 @@ pub fn recon_mb(frm: &mut NASimpleVideoFrame, slice_hdr: &SliceHeader, mb_in 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); + do_b_mc_4x4bi(frm, xpos + bx, ypos + by, mv, rpic0, &weight0, rpic1, &weight1, mc_dsp); bx += 4; if blk == 1 { bx -= 8;