diff options
author | Minghai Shang <minghai@google.com> | 2014-10-14 16:25:03 -0700 |
---|---|---|
committer | Minghai Shang <minghai@google.com> | 2014-10-14 16:26:39 -0700 |
commit | c113457af9880b8e15a36cdaabfd414d1c245693 (patch) | |
tree | 367ded1111a9d288fd1bf39f3c80f1dbb88e2818 /vp9/encoder | |
parent | 2040bb58fbec7d06d5bdb1f6628bb058d3132ebf (diff) | |
download | libvpx-c113457af9880b8e15a36cdaabfd414d1c245693.tar libvpx-c113457af9880b8e15a36cdaabfd414d1c245693.tar.gz libvpx-c113457af9880b8e15a36cdaabfd414d1c245693.tar.bz2 libvpx-c113457af9880b8e15a36cdaabfd414d1c245693.zip |
[spatial svc]Another workaround to avoid using prev_mi
We encode a empty invisible frame in front of the base layer frame to
avoid using prev_mi. Since there's a restriction for reference frame
scaling factor, we have to make it smaller and smaller gradually until
its size is 16x16.
Change-Id: I60b680314e33a60b4093cafc296465ee18169c19
Diffstat (limited to 'vp9/encoder')
-rw-r--r-- | vp9/encoder/vp9_bitstream.c | 9 | ||||
-rw-r--r-- | vp9/encoder/vp9_encoder.c | 58 | ||||
-rw-r--r-- | vp9/encoder/vp9_firstpass.c | 3 | ||||
-rw-r--r-- | vp9/encoder/vp9_svc_layercontext.c | 63 | ||||
-rw-r--r-- | vp9/encoder/vp9_svc_layercontext.h | 10 | ||||
-rw-r--r-- | vp9/encoder/vp9_temporal_filter.c | 3 |
6 files changed, 121 insertions, 25 deletions
diff --git a/vp9/encoder/vp9_bitstream.c b/vp9/encoder/vp9_bitstream.c index 694cac76f..3954fe6a7 100644 --- a/vp9/encoder/vp9_bitstream.c +++ b/vp9/encoder/vp9_bitstream.c @@ -1013,7 +1013,11 @@ static void write_frame_size_with_refs(VP9_COMP *cpi, ((cpi->svc.number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR) || (cpi->svc.number_spatial_layers > 1 && - cpi->svc.layer_context[cpi->svc.spatial_layer_id].is_key_frame))) { + cpi->svc.layer_context[cpi->svc.spatial_layer_id].is_key_frame) || + (is_two_pass_svc(cpi) && + cpi->svc.encode_empty_frame_state == ENCODING && + cpi->svc.layer_context[0].frames_from_key_frame < + cpi->svc.number_temporal_layers + 1))) { found = 0; } vp9_wb_write_bit(wb, found); @@ -1105,8 +1109,7 @@ static void write_uncompressed_header(VP9_COMP *cpi, // will change to show_frame flag to 0, then add an one byte frame with // show_existing_frame flag which tells the decoder which frame we want to // show. - if (!cm->show_frame || - (is_two_pass_svc(cpi) && cm->error_resilient_mode == 0)) + if (!cm->show_frame) vp9_wb_write_bit(wb, cm->intra_only); if (!cm->error_resilient_mode) diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c index dea93062b..0928c0bb7 100644 --- a/vp9/encoder/vp9_encoder.c +++ b/vp9/encoder/vp9_encoder.c @@ -225,6 +225,9 @@ static void dealloc_compressor_data(VP9_COMP *cpi) { } vpx_memset(&cpi->svc.scaled_frames[0], 0, MAX_LAG_BUFFERS * sizeof(cpi->svc.scaled_frames[0])); + + vp9_free_frame_buffer(&cpi->svc.empty_frame.img); + vpx_memset(&cpi->svc.empty_frame, 0, sizeof(cpi->svc.empty_frame)); } static void save_coding_context(VP9_COMP *cpi) { @@ -2635,7 +2638,10 @@ void set_frame_size(VP9_COMP *cpi) { // For two pass encodes analyse the first pass stats and determine // the bit allocation and other parameters for this frame / group of frames. - if ((oxcf->pass == 2) && (!cpi->use_svc || is_two_pass_svc(cpi))) { + if ((oxcf->pass == 2) && + (!cpi->use_svc || + (is_two_pass_svc(cpi) && + cpi->svc.encode_empty_frame_state != ENCODING))) { vp9_rc_get_second_pass_params(cpi); } @@ -3134,7 +3140,9 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi, } } if (is_two_pass_svc(cpi) && cm->error_resilient_mode == 0) { + // Use the last frame context for the empty frame. cm->frame_context_idx = + (cpi->svc.encode_empty_frame_state == ENCODING) ? FRAME_CONTEXTS - 1 : cpi->svc.spatial_layer_id * cpi->svc.number_temporal_layers + cpi->svc.temporal_layer_id; @@ -3269,7 +3277,9 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi, cpi->ref_frame_flags = get_ref_frame_flags(cpi); cm->last_frame_type = cm->frame_type; - vp9_rc_postencode_update(cpi, *size); + + if (!(is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING)) + vp9_rc_postencode_update(cpi, *size); #if 0 output_frame_level_debug_stats(cpi); @@ -3293,12 +3303,8 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi, cm->last_height = cm->height; // reset to normal state now that we are done. - if (!cm->show_existing_frame) { - if (is_two_pass_svc(cpi) && cm->error_resilient_mode == 0) - cm->last_show_frame = 0; - else - cm->last_show_frame = cm->show_frame; - } + if (!cm->show_existing_frame) + cm->last_show_frame = cm->show_frame; if (cm->show_frame) { vp9_swap_mi_and_prev_mi(cm); @@ -3335,7 +3341,9 @@ static void Pass2Encode(VP9_COMP *cpi, size_t *size, uint8_t *dest, unsigned int *frame_flags) { cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED; encode_frame_to_data_rate(cpi, size, dest, frame_flags); - vp9_twopass_postencode_update(cpi); + + if (!(is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING)) + vp9_twopass_postencode_update(cpi); } static void check_initial_width(VP9_COMP *cpi, @@ -3513,6 +3521,9 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags, if (is_two_pass_svc(cpi)) { #if CONFIG_SPATIAL_SVC vp9_svc_start_frame(cpi); + // Use a small empty frame instead of a real frame + if (cpi->svc.encode_empty_frame_state == ENCODING) + source = &cpi->svc.empty_frame; #endif if (oxcf->pass == 2) vp9_restore_layer_context(cpi); @@ -3531,6 +3542,11 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags, // Should we encode an arf frame. arf_src_index = get_arf_src_index(cpi); + + // Skip alt frame if we encode the empty frame + if (is_two_pass_svc(cpi) && source != NULL) + arf_src_index = 0; + if (arf_src_index) { assert(arf_src_index <= rc->frames_to_key); @@ -3818,10 +3834,18 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags, #endif - if (is_two_pass_svc(cpi) && cm->show_frame) { - ++cpi->svc.spatial_layer_to_encode; - if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers) - cpi->svc.spatial_layer_to_encode = 0; + if (is_two_pass_svc(cpi)) { + if (cpi->svc.encode_empty_frame_state == ENCODING) + cpi->svc.encode_empty_frame_state = ENCODED; + + if (cm->show_frame) { + ++cpi->svc.spatial_layer_to_encode; + if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers) + cpi->svc.spatial_layer_to_encode = 0; + + // May need the empty frame after an visible frame. + cpi->svc.encode_empty_frame_state = NEED_TO_ENCODE; + } } return 0; } @@ -3912,10 +3936,6 @@ int vp9_set_size_literal(VP9_COMP *cpi, unsigned int width, if (width) { cm->width = width; - if (cm->width * 5 < cpi->initial_width) { - cm->width = cpi->initial_width / 5 + 1; - printf("Warning: Desired width too small, changed to %d\n", cm->width); - } if (cm->width > cpi->initial_width) { cm->width = cpi->initial_width; printf("Warning: Desired width too large, changed to %d\n", cm->width); @@ -3924,10 +3944,6 @@ int vp9_set_size_literal(VP9_COMP *cpi, unsigned int width, if (height) { cm->height = height; - if (cm->height * 5 < cpi->initial_height) { - cm->height = cpi->initial_height / 5 + 1; - printf("Warning: Desired height too small, changed to %d\n", cm->height); - } if (cm->height > cpi->initial_height) { cm->height = cpi->initial_height; printf("Warning: Desired height too large, changed to %d\n", cm->height); diff --git a/vp9/encoder/vp9_firstpass.c b/vp9/encoder/vp9_firstpass.c index 9e204d7e9..760ec320c 100644 --- a/vp9/encoder/vp9_firstpass.c +++ b/vp9/encoder/vp9_firstpass.c @@ -2406,6 +2406,9 @@ void vp9_rc_get_second_pass_params(VP9_COMP *cpi) { cpi->ref_frame_flags &= (~VP9_LAST_FLAG & ~VP9_GOLD_FLAG & ~VP9_ALT_FLAG); lc->frames_from_key_frame = 0; + // Reset the empty frame resolution since we have a key frame. + cpi->svc.empty_frame_width = cm->width; + cpi->svc.empty_frame_height = cm->height; } } else { cm->frame_type = INTER_FRAME; diff --git a/vp9/encoder/vp9_svc_layercontext.c b/vp9/encoder/vp9_svc_layercontext.c index eed681c96..1573557d4 100644 --- a/vp9/encoder/vp9_svc_layercontext.c +++ b/vp9/encoder/vp9_svc_layercontext.c @@ -14,6 +14,8 @@ #include "vp9/encoder/vp9_svc_layercontext.h" #include "vp9/encoder/vp9_extend.h" +#define SMALL_FRAME_FB_IDX 7 + void vp9_init_layer_context(VP9_COMP *const cpi) { SVC *const svc = &cpi->svc; const VP9EncoderConfig *const oxcf = &cpi->oxcf; @@ -28,6 +30,25 @@ void vp9_init_layer_context(VP9_COMP *const cpi) { layer_end = svc->number_temporal_layers; } else { layer_end = svc->number_spatial_layers; + + if (cpi->oxcf.error_resilient_mode == 0 && cpi->oxcf.pass == 2) { + if (vp9_realloc_frame_buffer(&cpi->svc.empty_frame.img, + cpi->common.width, cpi->common.height, + cpi->common.subsampling_x, + cpi->common.subsampling_y, +#if CONFIG_VP9_HIGHBITDEPTH + cpi->common.use_highbitdepth, +#endif + VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL)) + vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR, + "Failed to allocate empty frame for multiple frame " + "contexts"); + + vpx_memset(cpi->svc.empty_frame.img.buffer_alloc, 0x80, + cpi->svc.empty_frame.img.buffer_alloc_sz); + cpi->svc.empty_frame_width = cpi->common.width; + cpi->svc.empty_frame_height = cpi->common.height; + } } for (layer = 0; layer < layer_end; ++layer) { @@ -310,6 +331,47 @@ int vp9_svc_start_frame(VP9_COMP *const cpi) { get_layer_resolution(cpi->oxcf.width, cpi->oxcf.height, lc->scaling_factor_num, lc->scaling_factor_den, &width, &height); + + // Workaround for multiple frame contexts. In some frames we can't use prev_mi + // since its previous frame could be changed during decoding time. The idea is + // we put a empty invisible frame in front of them, then we will not use + // prev_mi when encoding these frames. + if (cpi->oxcf.error_resilient_mode == 0 && cpi->oxcf.pass == 2 && + cpi->svc.encode_empty_frame_state == NEED_TO_ENCODE) { + if ((cpi->svc.number_temporal_layers > 1 && + cpi->svc.temporal_layer_id < cpi->svc.number_temporal_layers - 1) || + (cpi->svc.number_spatial_layers > 1 && + cpi->svc.spatial_layer_id == 0)) { + struct lookahead_entry *buf = vp9_lookahead_peek(cpi->lookahead, 0); + + if (buf != NULL) { + cpi->svc.empty_frame.ts_start = buf->ts_start; + cpi->svc.empty_frame.ts_end = buf->ts_end; + cpi->svc.encode_empty_frame_state = ENCODING; + cpi->common.show_frame = 0; + cpi->ref_frame_flags = 0; + cpi->common.frame_type = INTER_FRAME; + cpi->lst_fb_idx = + cpi->gld_fb_idx = cpi->alt_fb_idx = SMALL_FRAME_FB_IDX; + + // Gradually make the empty frame smaller to save bits. Make it half of + // its previous size because of the scaling factor restriction. + cpi->svc.empty_frame_width >>= 1; + cpi->svc.empty_frame_width = (cpi->svc.empty_frame_width + 1) & ~1; + if (cpi->svc.empty_frame_width < 16) + cpi->svc.empty_frame_width = 16; + + cpi->svc.empty_frame_height >>= 1; + cpi->svc.empty_frame_height = (cpi->svc.empty_frame_height + 1) & ~1; + if (cpi->svc.empty_frame_height < 16) + cpi->svc.empty_frame_height = 16; + + width = cpi->svc.empty_frame_width; + height = cpi->svc.empty_frame_height; + } + } + } + if (vp9_set_size_literal(cpi, width, height) != 0) return VPX_CODEC_INVALID_PARAM; @@ -317,7 +379,6 @@ int vp9_svc_start_frame(VP9_COMP *const cpi) { cpi->oxcf.best_allowed_q = vp9_quantizer_to_qindex(lc->min_q); vp9_change_config(cpi, &cpi->oxcf); - vp9_set_high_precision_mv(cpi, 1); cpi->alt_ref_source = get_layer_context(cpi)->alt_ref_source; diff --git a/vp9/encoder/vp9_svc_layercontext.h b/vp9/encoder/vp9_svc_layercontext.h index 47a5456b6..e9645ce9f 100644 --- a/vp9/encoder/vp9_svc_layercontext.h +++ b/vp9/encoder/vp9_svc_layercontext.h @@ -50,6 +50,16 @@ typedef struct { int spatial_layer_to_encode; + // Workaround for multiple frame contexts + enum { + ENCODED = 0, + ENCODING, + NEED_TO_ENCODE + }encode_empty_frame_state; + struct lookahead_entry empty_frame; + int empty_frame_width; + int empty_frame_height; + // Store scaled source frames to be used for temporal filter to generate // a alt ref frame. YV12_BUFFER_CONFIG scaled_frames[MAX_LAG_BUFFERS]; diff --git a/vp9/encoder/vp9_temporal_filter.c b/vp9/encoder/vp9_temporal_filter.c index 9ae81e761..5599227ce 100644 --- a/vp9/encoder/vp9_temporal_filter.c +++ b/vp9/encoder/vp9_temporal_filter.c @@ -719,6 +719,9 @@ void vp9_temporal_filter(VP9_COMP *cpi, int distance) { ++frame_used; } } + cm->mi = cm->mip + cm->mi_stride + 1; + cpi->mb.e_mbd.mi = cm->mi; + cpi->mb.e_mbd.mi[0].src_mi = &cpi->mb.e_mbd.mi[0]; } else { // ARF is produced at the native frame size and resized when coded. #if CONFIG_VP9_HIGHBITDEPTH |