diff options
Diffstat (limited to 'libvpx/vp9/encoder/vp9_encodemb.c')
| -rw-r--r-- | libvpx/vp9/encoder/vp9_encodemb.c | 123 |
1 files changed, 68 insertions, 55 deletions
diff --git a/libvpx/vp9/encoder/vp9_encodemb.c b/libvpx/vp9/encoder/vp9_encodemb.c index 75ed8ea..a85ddee 100644 --- a/libvpx/vp9/encoder/vp9_encodemb.c +++ b/libvpx/vp9/encoder/vp9_encodemb.c @@ -136,14 +136,13 @@ static void optimize_b(MACROBLOCK *mb, const int16_t *scan, *nb; const int mul = 1 + (tx_size == TX_32X32); uint8_t token_cache[1024]; - const int ib = txfrm_block_to_raster_block(plane_bsize, tx_size, block); const int16_t *dequant_ptr = pd->dequant; const uint8_t *const band_translate = get_band_translate(tx_size); assert((!type && !plane) || (type && plane)); dqcoeff_ptr = BLOCK_OFFSET(pd->dqcoeff, block); qcoeff_ptr = BLOCK_OFFSET(pd->qcoeff, block); - get_scan(xd, tx_size, type, ib, &scan, &nb); + get_scan(xd, tx_size, type, block, &scan, &nb); assert(eob <= default_eob); /* Now set up a Viterbi trellis to evaluate alternative roundings. */ @@ -179,7 +178,7 @@ static void optimize_b(MACROBLOCK *mb, t0 = (vp9_dct_value_tokens_ptr + x)->token; /* Consider both possible successor states. */ if (next < default_eob) { - band = get_coef_band(band_translate, i + 1); + band = band_translate[i + 1]; pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache); rate0 += mb->token_costs[tx_size][type][ref][band][0][pt] @@ -230,7 +229,7 @@ static void optimize_b(MACROBLOCK *mb, t0 = t1 = (vp9_dct_value_tokens_ptr + x)->token; } if (next < default_eob) { - band = get_coef_band(band_translate, i + 1); + band = band_translate[i + 1]; if (t0 != DCT_EOB_TOKEN) { pt = trellis_get_coeff_context(scan, nb, i, t0, token_cache); rate0 += mb->token_costs[tx_size][type][ref][band][!x][pt] @@ -264,7 +263,7 @@ static void optimize_b(MACROBLOCK *mb, /* There's no choice to make for a zero coefficient, so we don't * add a new trellis node, but we do need to update the costs. */ - band = get_coef_band(band_translate, i + 1); + band = band_translate[i + 1]; t0 = tokens[next][0].token; t1 = tokens[next][1].token; /* Update the cost of each path if we're past the EOB token. */ @@ -284,7 +283,7 @@ static void optimize_b(MACROBLOCK *mb, } /* Now pick the best path through the whole trellis. */ - band = get_coef_band(band_translate, i + 1); + band = band_translate[i + 1]; pt = combine_entropy_contexts(*a, *l); rate0 = tokens[next][0].rate; rate1 = tokens[next][1].rate; @@ -420,28 +419,30 @@ static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize, MACROBLOCKD *const xd = &x->e_mbd; struct optimize_ctx *const ctx = args->ctx; struct macroblockd_plane *const pd = &xd->plane[plane]; - const int raster_block = txfrm_block_to_raster_block(plane_bsize, tx_size, - block); - int16_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block); - uint8_t *const dst = raster_block_offset_uint8(plane_bsize, raster_block, - pd->dst.buf, pd->dst.stride); + int i, j; + uint8_t *dst; + txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j); + dst = &pd->dst.buf[4 * j * pd->dst.stride + 4 * i]; // TODO(jingning): per transformed block zero forcing only enabled for // luma component. will integrate chroma components as well. if (x->zcoeff_blk[tx_size][block] && plane == 0) { - int i, j; pd->eobs[block] = 0; - txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j); ctx->ta[plane][i] = 0; ctx->tl[plane][j] = 0; return; } - vp9_xform_quant(plane, block, plane_bsize, tx_size, arg); + if (!x->skip_recode) + vp9_xform_quant(plane, block, plane_bsize, tx_size, arg); - if (x->optimize) + if (x->optimize && (!x->skip_recode || !x->skip_optimize)) { vp9_optimize_b(plane, block, plane_bsize, tx_size, x, ctx); + } else { + ctx->ta[plane][i] = pd->eobs[block] > 0; + ctx->tl[plane][j] = pd->eobs[block] > 0; + } if (x->skip_encode || pd->eobs[block] == 0) return; @@ -505,9 +506,10 @@ void vp9_encode_sb(MACROBLOCK *x, BLOCK_SIZE bsize) { struct optimize_ctx ctx; struct encode_b_args arg = {x, &ctx}; - vp9_subtract_sb(x, bsize); + if (!x->skip_recode) + vp9_subtract_sb(x, bsize); - if (x->optimize) { + if (x->optimize && (!x->skip_recode || !x->skip_optimize)) { int i; for (i = 0; i < MAX_MB_PLANE; ++i) optimize_init_b(i, bsize, &arg); @@ -552,19 +554,22 @@ void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize, xoff = 32 * (block & twmask); yoff = 32 * (block >> twl); dst = pd->dst.buf + yoff * pd->dst.stride + xoff; - src = p->src.buf + yoff * p->src.stride + xoff; - src_diff = p->src_diff + 4 * bw * yoff + xoff; vp9_predict_intra_block(xd, block, bwl, TX_32X32, mode, dst, pd->dst.stride, dst, pd->dst.stride); - vp9_subtract_block(32, 32, src_diff, bw * 4, - src, p->src.stride, dst, pd->dst.stride); - if (x->use_lp32x32fdct) - vp9_fdct32x32_rd(src_diff, coeff, bw * 4); - else - vp9_fdct32x32(src_diff, coeff, bw * 4); - vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round, - p->quant, p->quant_shift, qcoeff, dqcoeff, - pd->dequant, p->zbin_extra, eob, scan, iscan); + + if (!x->skip_recode) { + src = p->src.buf + yoff * p->src.stride + xoff; + src_diff = p->src_diff + 4 * bw * yoff + xoff; + vp9_subtract_block(32, 32, src_diff, bw * 4, + src, p->src.stride, dst, pd->dst.stride); + if (x->use_lp32x32fdct) + vp9_fdct32x32_rd(src_diff, coeff, bw * 4); + else + vp9_fdct32x32(src_diff, coeff, bw * 4); + vp9_quantize_b_32x32(coeff, 1024, x->skip_block, p->zbin, p->round, + p->quant, p->quant_shift, qcoeff, dqcoeff, + pd->dequant, p->zbin_extra, eob, scan, iscan); + } if (!x->skip_encode && *eob) vp9_idct32x32_add(dqcoeff, dst, pd->dst.stride, *eob); break; @@ -577,16 +582,18 @@ void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize, xoff = 16 * (block & twmask); yoff = 16 * (block >> twl); dst = pd->dst.buf + yoff * pd->dst.stride + xoff; - src = p->src.buf + yoff * p->src.stride + xoff; - src_diff = p->src_diff + 4 * bw * yoff + xoff; vp9_predict_intra_block(xd, block, bwl, TX_16X16, mode, dst, pd->dst.stride, dst, pd->dst.stride); - vp9_subtract_block(16, 16, src_diff, bw * 4, - src, p->src.stride, dst, pd->dst.stride); - vp9_fht16x16(tx_type, src_diff, coeff, bw * 4); - vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round, - p->quant, p->quant_shift, qcoeff, dqcoeff, - pd->dequant, p->zbin_extra, eob, scan, iscan); + if (!x->skip_recode) { + src = p->src.buf + yoff * p->src.stride + xoff; + src_diff = p->src_diff + 4 * bw * yoff + xoff; + vp9_subtract_block(16, 16, src_diff, bw * 4, + src, p->src.stride, dst, pd->dst.stride); + vp9_fht16x16(tx_type, src_diff, coeff, bw * 4); + vp9_quantize_b(coeff, 256, x->skip_block, p->zbin, p->round, + p->quant, p->quant_shift, qcoeff, dqcoeff, + pd->dequant, p->zbin_extra, eob, scan, iscan); + } if (!x->skip_encode && *eob) vp9_iht16x16_add(tx_type, dqcoeff, dst, pd->dst.stride, *eob); break; @@ -599,16 +606,18 @@ void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize, xoff = 8 * (block & twmask); yoff = 8 * (block >> twl); dst = pd->dst.buf + yoff * pd->dst.stride + xoff; - src = p->src.buf + yoff * p->src.stride + xoff; - src_diff = p->src_diff + 4 * bw * yoff + xoff; vp9_predict_intra_block(xd, block, bwl, TX_8X8, mode, dst, pd->dst.stride, dst, pd->dst.stride); - vp9_subtract_block(8, 8, src_diff, bw * 4, - src, p->src.stride, dst, pd->dst.stride); - vp9_fht8x8(tx_type, src_diff, coeff, bw * 4); - vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round, p->quant, - p->quant_shift, qcoeff, dqcoeff, - pd->dequant, p->zbin_extra, eob, scan, iscan); + if (!x->skip_recode) { + src = p->src.buf + yoff * p->src.stride + xoff; + src_diff = p->src_diff + 4 * bw * yoff + xoff; + vp9_subtract_block(8, 8, src_diff, bw * 4, + src, p->src.stride, dst, pd->dst.stride); + vp9_fht8x8(tx_type, src_diff, coeff, bw * 4); + vp9_quantize_b(coeff, 64, x->skip_block, p->zbin, p->round, p->quant, + p->quant_shift, qcoeff, dqcoeff, + pd->dequant, p->zbin_extra, eob, scan, iscan); + } if (!x->skip_encode && *eob) vp9_iht8x8_add(tx_type, dqcoeff, dst, pd->dst.stride, *eob); break; @@ -624,19 +633,23 @@ void vp9_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize, xoff = 4 * (block & twmask); yoff = 4 * (block >> twl); dst = pd->dst.buf + yoff * pd->dst.stride + xoff; - src = p->src.buf + yoff * p->src.stride + xoff; - src_diff = p->src_diff + 4 * bw * yoff + xoff; vp9_predict_intra_block(xd, block, bwl, TX_4X4, mode, dst, pd->dst.stride, dst, pd->dst.stride); - vp9_subtract_block(4, 4, src_diff, bw * 4, - src, p->src.stride, dst, pd->dst.stride); - if (tx_type != DCT_DCT) - vp9_short_fht4x4(src_diff, coeff, bw * 4, tx_type); - else - x->fwd_txm4x4(src_diff, coeff, bw * 4); - vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round, p->quant, - p->quant_shift, qcoeff, dqcoeff, - pd->dequant, p->zbin_extra, eob, scan, iscan); + + if (!x->skip_recode) { + src = p->src.buf + yoff * p->src.stride + xoff; + src_diff = p->src_diff + 4 * bw * yoff + xoff; + vp9_subtract_block(4, 4, src_diff, bw * 4, + src, p->src.stride, dst, pd->dst.stride); + if (tx_type != DCT_DCT) + vp9_short_fht4x4(src_diff, coeff, bw * 4, tx_type); + else + x->fwd_txm4x4(src_diff, coeff, bw * 4); + vp9_quantize_b(coeff, 16, x->skip_block, p->zbin, p->round, p->quant, + p->quant_shift, qcoeff, dqcoeff, + pd->dequant, p->zbin_extra, eob, scan, iscan); + } + if (!x->skip_encode && *eob) { if (tx_type == DCT_DCT) // this is like vp9_short_idct4x4 but has a special case around eob<=1 |