From 8ba92a0bed9f05b79851b364c4fafd97ac88e824 Mon Sep 17 00:00:00 2001 From: Yaowu Xu Date: Fri, 17 May 2013 12:50:40 -0700 Subject: [PATCH] changes intra coding to be based on txfm block This commit changed the encoding and decoding of intra blocks to be based on transform block. In each prediction block, the intra coding iterates thorough each transform block based on raster scan order. This commit also fixed a bug in D135 prediction code. TODO next: The RD mode/txfm_size selection should take this into account when computing RD values. Change-Id: I6d1be2faa4c4948a52e830b6a9a84a6b2b6850f6 --- vp9/common/vp9_blockd.h | 2 ++ vp9/common/vp9_reconintra.c | 14 +++----- vp9/common/vp9_reconintra.h | 2 +- vp9/decoder/vp9_decodemv.c | 9 ++--- vp9/decoder/vp9_decodframe.c | 82 ++++++++++++++++++++++++++++++++++++++++--- vp9/decoder/vp9_onyxd_if.c | 10 ++++++ vp9/encoder/vp9_bitstream.c | 8 +++-- vp9/encoder/vp9_encodeframe.c | 47 +++++++++++++------------ vp9/encoder/vp9_encodeintra.h | 7 ++-- vp9/encoder/vp9_encodemb.c | 30 +++++++++++----- vp9/encoder/vp9_onyx_if.c | 9 +++++ 11 files changed, 166 insertions(+), 54 deletions(-) diff --git a/vp9/common/vp9_blockd.h b/vp9/common/vp9_blockd.h index 71d1dd9..f1722fd 100644 --- a/vp9/common/vp9_blockd.h +++ b/vp9/common/vp9_blockd.h @@ -192,6 +192,7 @@ static INLINE int b_width_log2(BLOCK_SIZE_TYPE sb_type) { case BLOCK_SIZE_SB64X32: case BLOCK_SIZE_SB64X64: return 4; default: assert(0); + return -1; } } @@ -215,6 +216,7 @@ static INLINE int b_height_log2(BLOCK_SIZE_TYPE sb_type) { case BLOCK_SIZE_SB32X64: case BLOCK_SIZE_SB64X64: return 4; default: assert(0); + return -1; } } diff --git a/vp9/common/vp9_reconintra.c b/vp9/common/vp9_reconintra.c index 1588f33..dd60a76 100644 --- a/vp9/common/vp9_reconintra.c +++ b/vp9/common/vp9_reconintra.c @@ -138,14 +138,10 @@ static void d135_predictor(uint8_t *ypred_ptr, int y_stride, ypred_ptr[y_stride] = ROUND_POWER_OF_TWO(yabove_row[-1] + yleft_col[0] * 2 + yleft_col[1], 2); - for (r = 2; r < bh - 1; ++r) + for (r = 2; r < bh; ++r) ypred_ptr[r * y_stride] = ROUND_POWER_OF_TWO(yleft_col[r - 2] + yleft_col[r - 1] * 2 + - yleft_col[r + 1], 2); - - ypred_ptr[(bh - 1) * y_stride] = ROUND_POWER_OF_TWO(yleft_col[bh - 2] + - yleft_col[bh - 1] * 3, - 2); + yleft_col[r], 2); ypred_ptr += y_stride; for (r = 1; r < bh; ++r) { @@ -396,11 +392,11 @@ void vp9_build_intra_predictors_sbuv_s(MACROBLOCKD *xd, void vp9_predict_intra_block(MACROBLOCKD *xd, int block_idx, - BLOCK_SIZE_TYPE bsize, + int bwl_in, TX_SIZE tx_size, int mode, uint8_t *predictor, int pre_stride) { - const int bwl = b_width_log2(bsize) - tx_size; + const int bwl = bwl_in - tx_size; const int wmask = (1 << bwl) - 1; const int have_top = (block_idx >> bwl) || xd->up_available; @@ -424,6 +420,6 @@ void vp9_intra4x4_predict(MACROBLOCKD *xd, BLOCK_SIZE_TYPE bsize, int mode, uint8_t *predictor, int pre_stride) { - vp9_predict_intra_block(xd, block_idx, bsize, TX_4X4, + vp9_predict_intra_block(xd, block_idx, b_width_log2(bsize), TX_4X4, mode, predictor, pre_stride); } diff --git a/vp9/common/vp9_reconintra.h b/vp9/common/vp9_reconintra.h index faecd6b..b88761b 100644 --- a/vp9/common/vp9_reconintra.h +++ b/vp9/common/vp9_reconintra.h @@ -23,7 +23,7 @@ B_PREDICTION_MODE vp9_find_bpred_context(MACROBLOCKD *xd, int block, void vp9_predict_intra_block(MACROBLOCKD *xd, int block_idx, - BLOCK_SIZE_TYPE bsize, + int bwl_in, TX_SIZE tx_size, int mode, uint8_t *predictor, int pre_stride); diff --git a/vp9/decoder/vp9_decodemv.c b/vp9/decoder/vp9_decodemv.c index ac66fb6..1ddd21c 100644 --- a/vp9/decoder/vp9_decodemv.c +++ b/vp9/decoder/vp9_decodemv.c @@ -154,11 +154,11 @@ static void kfread_modes(VP9D_COMP *pbi, MODE_INFO *m, m->mbmi.uv_mode = read_uv_mode(r, cm->kf_uv_mode_prob[m->mbmi.mode]); if (cm->txfm_mode == TX_MODE_SELECT && - !m->mbmi.mb_skip_coeff && + !(m->mbmi.mb_skip_coeff && m->mbmi.ref_frame != INTRA_FRAME) #if CONFIG_AB4X4 - m->mbmi.sb_type >= BLOCK_SIZE_SB8X8 + && m->mbmi.sb_type >= BLOCK_SIZE_SB8X8 #else - m->mbmi.mode != I4X4_PRED + && m->mbmi.mode != I4X4_PRED #endif ) { const int allow_16x16 = m->mbmi.sb_type >= BLOCK_SIZE_MB16X16; @@ -880,7 +880,8 @@ static void read_mb_modes_mv(VP9D_COMP *pbi, MODE_INFO *mi, MB_MODE_INFO *mbmi, if (cm->txfm_mode == TX_MODE_SELECT && mbmi->mb_skip_coeff == 0 && bsize >= BLOCK_SIZE_SB8X8) { #else - if (cm->txfm_mode == TX_MODE_SELECT && mbmi->mb_skip_coeff == 0 && + if (cm->txfm_mode == TX_MODE_SELECT && + (mbmi->mb_skip_coeff == 0 || mbmi->ref_frame == INTRA_FRAME) && ((mbmi->ref_frame == INTRA_FRAME && mbmi->mode != I4X4_PRED) || (mbmi->ref_frame != INTRA_FRAME && mbmi->mode != SPLITMV))) { #endif diff --git a/vp9/decoder/vp9_decodframe.c b/vp9/decoder/vp9_decodframe.c index 449e426..eea7658 100644 --- a/vp9/decoder/vp9_decodframe.c +++ b/vp9/decoder/vp9_decodframe.c @@ -239,6 +239,59 @@ static void decode_block(int plane, int block, BLOCK_SIZE_TYPE bsize, } } +static void decode_block_intra(int plane, int block, BLOCK_SIZE_TYPE bsize, + int ss_txfrm_size, void *arg) { + MACROBLOCKD* const xd = arg; + int16_t* const qcoeff = BLOCK_OFFSET(xd->plane[plane].qcoeff, block, 16); + const int stride = xd->plane[plane].dst.stride; + const int raster_block = txfrm_block_to_raster_block(xd, bsize, plane, + block, ss_txfrm_size); + uint8_t* const dst = raster_block_offset_uint8(xd, bsize, plane, + raster_block, + xd->plane[plane].dst.buf, + stride); + const TX_SIZE tx_size = (TX_SIZE)(ss_txfrm_size / 2); + TX_TYPE tx_type; + int mode, b_mode; + int plane_b_size; + int tx_ib = raster_block >> tx_size; + mode = plane == 0? xd->mode_info_context->mbmi.mode: + xd->mode_info_context->mbmi.uv_mode; + + if (bsize <= BLOCK_SIZE_SB8X8 && mode == I4X4_PRED && plane == 0) + b_mode = xd->mode_info_context->bmi[raster_block].as_mode.first; + else + b_mode = mode; + + plane_b_size = b_width_log2(bsize) - xd->plane[plane].subsampling_x; + vp9_predict_intra_block(xd, tx_ib, plane_b_size, tx_size, + b_mode, dst, xd->plane[plane].dst.stride); + + switch (ss_txfrm_size / 2) { + case TX_4X4: + tx_type = plane == 0 ? get_tx_type_4x4(xd, raster_block) : DCT_DCT; + if (tx_type == DCT_DCT) + xd->itxm_add(qcoeff, dst, stride, xd->plane[plane].eobs[block]); + else + vp9_iht_add_c(tx_type, qcoeff, dst, stride, + xd->plane[plane].eobs[block]); + break; + case TX_8X8: + tx_type = plane == 0 ? get_tx_type_8x8(xd, raster_block) : DCT_DCT; + vp9_iht_add_8x8_c(tx_type, qcoeff, dst, stride, + xd->plane[plane].eobs[block]); + break; + case TX_16X16: + tx_type = plane == 0 ? get_tx_type_16x16(xd, raster_block) : DCT_DCT; + vp9_iht_add_16x16_c(tx_type, qcoeff, dst, stride, + xd->plane[plane].eobs[block]); + break; + case TX_32X32: + vp9_idct_add_32x32(qcoeff, dst, stride, xd->plane[plane].eobs[block]); + break; + } +} + static void decode_atom_intra(VP9D_COMP *pbi, MACROBLOCKD *xd, vp9_reader *r, BLOCK_SIZE_TYPE bsize) { @@ -295,6 +348,26 @@ static void decode_atom(VP9D_COMP *pbi, MACROBLOCKD *xd, foreach_transformed_block(xd, bsize, decode_block, xd); } +static void decode_sb_intra(VP9D_COMP *pbi, MACROBLOCKD *xd, + int mi_row, int mi_col, + vp9_reader *r, BLOCK_SIZE_TYPE bsize) { + MB_MODE_INFO *const mbmi = &xd->mode_info_context->mbmi; + if (mbmi->mb_skip_coeff) { + vp9_reset_sb_tokens_context(xd, bsize); + } else { + // re-initialize macroblock dequantizer before detokenization + if (xd->segmentation_enabled) + mb_init_dequantizer(&pbi->common, xd); + + if (!vp9_reader_has_error(r)) { + vp9_decode_tokens(pbi, xd, r, bsize); + } + } + + foreach_transformed_block(xd, bsize, decode_block_intra, xd); +} + + static void decode_sb(VP9D_COMP *pbi, MACROBLOCKD *xd, int mi_row, int mi_col, vp9_reader *r, BLOCK_SIZE_TYPE bsize) { const int bwl = mi_width_log2(bsize), bhl = mi_height_log2(bsize); @@ -422,12 +495,13 @@ static void decode_modes_b(VP9D_COMP *pbi, int mi_row, int mi_col, vp9_decode_mb_mode_mv(pbi, xd, mi_row, mi_col, r); set_refs(pbi, mi_row, mi_col); + if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME) + decode_sb_intra(pbi, xd, mi_row, mi_col, r, bsize); #if CONFIG_AB4X4 - if (bsize < BLOCK_SIZE_SB8X8) + else if (bsize < BLOCK_SIZE_SB8X8) #else - if (bsize == BLOCK_SIZE_SB8X8 && - (xd->mode_info_context->mbmi.mode == SPLITMV || - xd->mode_info_context->mbmi.mode == I4X4_PRED)) + else if (bsize == BLOCK_SIZE_SB8X8 && + xd->mode_info_context->mbmi.mode == SPLITMV) #endif decode_atom(pbi, xd, mi_row, mi_col, r, BLOCK_SIZE_SB8X8); else diff --git a/vp9/decoder/vp9_onyxd_if.c b/vp9/decoder/vp9_onyxd_if.c index 9761bd6..5d5a543 100644 --- a/vp9/decoder/vp9_onyxd_if.c +++ b/vp9/decoder/vp9_onyxd_if.c @@ -352,6 +352,16 @@ int vp9_receive_compressed_data(VP9D_PTR ptr, vp9_loop_filter_frame(cm, &pbi->mb, cm->filter_level, 0, cm->dering_enabled); } + +#if WRITE_RECON_BUFFER == 2 + if (cm->show_frame) + write_dx_frame_to_file(cm->frame_to_show, + cm->current_video_frame + 2000); + else + write_dx_frame_to_file(cm->frame_to_show, + cm->current_video_frame + 3000); +#endif + vp9_extend_frame_borders(cm->frame_to_show, cm->subsampling_x, cm->subsampling_y); } diff --git a/vp9/encoder/vp9_bitstream.c b/vp9/encoder/vp9_bitstream.c index 8894333..9f32680 100644 --- a/vp9/encoder/vp9_bitstream.c +++ b/vp9/encoder/vp9_bitstream.c @@ -871,8 +871,9 @@ static void pack_inter_mode_mvs(VP9_COMP *cpi, MODE_INFO *m, if (((rf == INTRA_FRAME && mode != I4X4_PRED) || (rf != INTRA_FRAME && mode != SPLITMV)) && pc->txfm_mode == TX_MODE_SELECT && - !(skip_coeff || vp9_segfeature_active(xd, segment_id, - SEG_LVL_SKIP))) + !(rf != INTRA_FRAME && + (skip_coeff || vp9_segfeature_active(xd, segment_id, + SEG_LVL_SKIP)))) #endif { TX_SIZE sz = mi->txfm_size; @@ -941,7 +942,8 @@ static void write_mb_modes_kf(const VP9_COMP *cpi, !(skip_coeff || vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP))) { #else if (ym != I4X4_PRED && c->txfm_mode == TX_MODE_SELECT && - !(skip_coeff || vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP))) { + !(m->mbmi.ref_frame != INTRA_FRAME && (skip_coeff || + vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP)))) { #endif TX_SIZE sz = m->mbmi.txfm_size; // FIXME(rbultje) code ternary symbol once all experiments are merged diff --git a/vp9/encoder/vp9_encodeframe.c b/vp9/encoder/vp9_encodeframe.c index 277f92c..73eec56 100644 --- a/vp9/encoder/vp9_encodeframe.c +++ b/vp9/encoder/vp9_encodeframe.c @@ -1703,19 +1703,14 @@ static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, #if CONFIG_AB4X4 if (mbmi->ref_frame == INTRA_FRAME && bsize < BLOCK_SIZE_SB8X8) { -#else - if (mbmi->mode == I4X4_PRED) { - assert(bsize == BLOCK_SIZE_SB8X8 && mbmi->txfm_size == TX_4X4); -#endif - vp9_encode_intra4x4mby(cm, x, BLOCK_SIZE_SB8X8); + vp9_encode_intra_block_y(cm, x, BLOCK_SIZE_SB8X8); vp9_build_intra_predictors_sbuv_s(xd, BLOCK_SIZE_SB8X8); vp9_encode_sbuv(cm, x, BLOCK_SIZE_SB8X8); - - if (output_enabled) - sum_intra_stats(cpi, x); - } else if (mbmi->ref_frame == INTRA_FRAME) { - vp9_build_intra_predictors_sby_s(xd, bsize); - vp9_build_intra_predictors_sbuv_s(xd, bsize); +#else + if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME) { + vp9_encode_intra_block_y(cm, x, bsize); + vp9_encode_intra_block_uv(cm, x, bsize); +#endif if (output_enabled) sum_intra_stats(cpi, x); } else { @@ -1741,10 +1736,10 @@ static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, if (mbmi->ref_frame == INTRA_FRAME && bsize < BLOCK_SIZE_SB8X8) { #else - if (mbmi->mode == I4X4_PRED) { - assert(bsize == BLOCK_SIZE_SB8X8); + if (xd->mode_info_context->mbmi.ref_frame == INTRA_FRAME) { #endif - vp9_tokenize_sb(cpi, xd, t, !output_enabled, BLOCK_SIZE_SB8X8); + vp9_tokenize_sb(cpi, xd, t, !output_enabled, + (bsize < BLOCK_SIZE_SB8X8) ? BLOCK_SIZE_SB8X8 : bsize); } else if (!x->skip) { vp9_encode_sb(cm, x, (bsize < BLOCK_SIZE_SB8X8) ? BLOCK_SIZE_SB8X8 : bsize); vp9_tokenize_sb(cpi, xd, t, !output_enabled, @@ -1771,8 +1766,9 @@ static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, if (output_enabled) { if (cm->txfm_mode == TX_MODE_SELECT && + (mbmi->ref_frame == INTRA_FRAME || !(mbmi->mb_skip_coeff || - vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP))) { + vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP)))) { if (bsize >= BLOCK_SIZE_SB32X32) { cpi->txfm_count_32x32p[mbmi->txfm_size]++; } else if (bsize >= BLOCK_SIZE_MB16X16) { @@ -1783,18 +1779,23 @@ static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, } else { int x, y; TX_SIZE sz = (cm->txfm_mode == TX_MODE_SELECT) ? TX_32X32 : cm->txfm_mode; - - if (sz == TX_32X32 && bsize < BLOCK_SIZE_SB32X32) - sz = TX_16X16; - if (sz == TX_16X16 && bsize < BLOCK_SIZE_MB16X16) - sz = TX_8X8; + // The new intra coding scheme requires no change of transform size + if (mi->mbmi.ref_frame != INTRA_FRAME) { + if (sz == TX_32X32 && bsize < BLOCK_SIZE_SB32X32) + sz = TX_16X16; + if (sz == TX_16X16 && bsize < BLOCK_SIZE_MB16X16) + sz = TX_8X8; #if CONFIG_AB4X4 - if (sz == TX_8X8 && bsize < BLOCK_SIZE_SB8X8) + if (sz == TX_8X8 && bsize < BLOCK_SIZE_SB8X8) #else - if (sz == TX_8X8 && (mbmi->mode == SPLITMV || - mbmi->mode == I4X4_PRED)) + if (sz == TX_8X8 && mbmi->mode == SPLITMV) #endif + sz = TX_4X4; + } else if (mbmi->mode != I4X4_PRED) { + sz = mbmi->txfm_size; + } else { sz = TX_4X4; + } for (y = 0; y < bh; y++) { for (x = 0; x < bw; x++) { diff --git a/vp9/encoder/vp9_encodeintra.h b/vp9/encoder/vp9_encodeintra.h index 22a046e..7da164c 100644 --- a/vp9/encoder/vp9_encodeintra.h +++ b/vp9/encoder/vp9_encodeintra.h @@ -16,6 +16,9 @@ int vp9_encode_intra(VP9_COMP *cpi, MACROBLOCK *x, int use_16x16_pred); void vp9_encode_intra16x16mby(VP9_COMMON *const cm, MACROBLOCK *x); void vp9_encode_intra16x16mbuv(VP9_COMMON *const cm, MACROBLOCK *x); -void vp9_encode_intra4x4mby(VP9_COMMON *const cm, MACROBLOCK *mb, - BLOCK_SIZE_TYPE bs); +void vp9_encode_intra_block_y(VP9_COMMON *const cm, MACROBLOCK *mb, + BLOCK_SIZE_TYPE bs); +void vp9_encode_intra_block_uv(VP9_COMMON *const cm, MACROBLOCK *mb, + BLOCK_SIZE_TYPE bs); + #endif // VP9_ENCODER_VP9_ENCODEINTRA_H_ diff --git a/vp9/encoder/vp9_encodemb.c b/vp9/encoder/vp9_encodemb.c index d9cd091..a4991f2 100644 --- a/vp9/encoder/vp9_encodemb.c +++ b/vp9/encoder/vp9_encodemb.c @@ -628,16 +628,23 @@ static void encode_block_intra(int plane, int block, BLOCK_SIZE_TYPE bsize, const int txfm_b_size = 4 << tx_size; int ib = raster_block; + int tx_ib = ib >> tx_size; + int plane_b_size; TX_TYPE tx_type; + int mode, b_mode; - if (tx_size <= TX_16X16) - tx_type = txfm_map(xd->mode_info_context->bmi[ib].as_mode.first); + mode = plane == 0? xd->mode_info_context->mbmi.mode: + xd->mode_info_context->mbmi.uv_mode; + if (bsize <= BLOCK_SIZE_SB8X8 && mode == I4X4_PRED && plane == 0) + b_mode = xd->mode_info_context->bmi[ib].as_mode.first; else - tx_type = DCT_DCT; + b_mode = mode; - vp9_predict_intra_block(&x->e_mbd, ib, bsize, tx_size, - xd->mode_info_context->bmi[ib].as_mode.first, + assert(b_mode >= B_DC_PRED && b_mode <= B_TM_PRED); + + plane_b_size = b_width_log2(bsize) - xd->plane[plane].subsampling_x; + vp9_predict_intra_block(xd, tx_ib, plane_b_size, tx_size, b_mode, dst, xd->plane[plane].dst.stride); vp9_subtract_block(txfm_b_size, txfm_b_size, src_diff, bw, @@ -650,7 +657,6 @@ static void encode_block_intra(int plane, int block, BLOCK_SIZE_TYPE bsize, if (x->optimize) vp9_optimize_b(plane, block, bsize, ss_txfrm_size, args->cm, x, args->ctx); */ - switch (ss_txfrm_size / 2) { case TX_32X32: vp9_short_idct32x32_add(BLOCK_OFFSET(xd->plane[plane].dqcoeff, @@ -695,8 +701,8 @@ static void encode_block_intra(int plane, int block, BLOCK_SIZE_TYPE bsize, } } -void vp9_encode_intra4x4mby(VP9_COMMON *const cm, MACROBLOCK *x, - BLOCK_SIZE_TYPE bsize) { +void vp9_encode_intra_block_y(VP9_COMMON *const cm, MACROBLOCK *x, + BLOCK_SIZE_TYPE bsize) { MACROBLOCKD* const xd = &x->e_mbd; struct optimize_ctx ctx; struct encode_b_args arg = {cm, x, &ctx}; @@ -704,4 +710,12 @@ void vp9_encode_intra4x4mby(VP9_COMMON *const cm, MACROBLOCK *x, foreach_transformed_block_in_plane(xd, bsize, 0, encode_block_intra, &arg); } +void vp9_encode_intra_block_uv(VP9_COMMON *const cm, MACROBLOCK *x, + BLOCK_SIZE_TYPE bsize) { + MACROBLOCKD* const xd = &x->e_mbd; + struct optimize_ctx ctx; + struct encode_b_args arg = {cm, x, &ctx}; + + foreach_transformed_block_uv(xd, bsize, encode_block_intra, &arg); +} diff --git a/vp9/encoder/vp9_onyx_if.c b/vp9/encoder/vp9_onyx_if.c index a8e1a17..e3c7930 100644 --- a/vp9/encoder/vp9_onyx_if.c +++ b/vp9/encoder/vp9_onyx_if.c @@ -3170,6 +3170,15 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi, // Pick the loop filter level for the frame. loopfilter_frame(cpi, cm); +#if WRITE_RECON_BUFFER + if (cm->show_frame) + write_cx_frame_to_file(cm->frame_to_show, + cm->current_video_frame + 2000); + else + write_cx_frame_to_file(cm->frame_to_show, + cm->current_video_frame + 3000); +#endif + // build the bitstream cpi->dummy_packing = 0; vp9_pack_bitstream(cpi, dest, size); -- 2.7.4