generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge "Move frame re-sizing into the recode loop"

Browse Source
This commit is contained in:
Paul Wilkins 2014-10-24 05:50:39 -07:00 committed by Gerrit Code Review
commit 3758650c98
3 changed files with 327 additions and 259 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

473
vp9/encoder/vp9_encoder.c
View File

@ -2409,30 +2409,37 @@ void vp9_scale_references(VP9_COMP *cpi) {
const VP9_REFFRAME ref_mask[3] = {VP9_LAST_FLAG, VP9_GOLD_FLAG, VP9_ALT_FLAG};
for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
const int idx = cm->ref_frame_map[get_ref_frame_idx(cpi, ref_frame)];
const YV12_BUFFER_CONFIG *const ref = &cm->frame_bufs[idx].buf;
// Need to convert from VP9_REFFRAME to index into ref_mask (subtract 1).
if ((cpi->ref_frame_flags & ref_mask[ref_frame - 1]) &&
(ref->y_crop_width != cm->width || ref->y_crop_height != cm->height)) {
const int new_fb = get_free_fb(cm);
vp9_realloc_frame_buffer(&cm->frame_bufs[new_fb].buf,
cm->width, cm->height,
cm->subsampling_x, cm->subsampling_y,
if (cpi->ref_frame_flags & ref_mask[ref_frame - 1]) {
const int idx = cm->ref_frame_map[get_ref_frame_idx(cpi, ref_frame)];
const YV12_BUFFER_CONFIG *const ref = &cm->frame_bufs[idx].buf;
#if CONFIG_VP9_HIGHBITDEPTH
cm->use_highbitdepth,
#endif // CONFIG_VP9_HIGHBITDEPTH
VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL);
#if CONFIG_VP9_HIGHBITDEPTH
scale_and_extend_frame(ref, &cm->frame_bufs[new_fb].buf,
(int)cm->bit_depth);
if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) {
const int new_fb = get_free_fb(cm);
vp9_realloc_frame_buffer(&cm->frame_bufs[new_fb].buf,
cm->width, cm->height,
cm->subsampling_x, cm->subsampling_y,
cm->use_highbitdepth,
VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL);
scale_and_extend_frame(ref, &cm->frame_bufs[new_fb].buf,
(int)cm->bit_depth);
#else
scale_and_extend_frame(ref, &cm->frame_bufs[new_fb].buf);
if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) {
const int new_fb = get_free_fb(cm);
vp9_realloc_frame_buffer(&cm->frame_bufs[new_fb].buf,
cm->width, cm->height,
cm->subsampling_x, cm->subsampling_y,
VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL);
scale_and_extend_frame(ref, &cm->frame_bufs[new_fb].buf);
#endif // CONFIG_VP9_HIGHBITDEPTH
cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
} else {
cpi->scaled_ref_idx[ref_frame - 1] = idx;
++cm->frame_bufs[idx].ref_count;
}
} else {
cpi->scaled_ref_idx[ref_frame - 1] = idx;
cm->frame_bufs[idx].ref_count++;
cpi->scaled_ref_idx[ref_frame - 1] = INVALID_REF_BUFFER_IDX;
}
}
}
@ -2440,9 +2447,13 @@ void vp9_scale_references(VP9_COMP *cpi) {
static void release_scaled_references(VP9_COMP *cpi) {
VP9_COMMON *cm = &cpi->common;
int i;
for (i = 0; i < 3; i++)
cm->frame_bufs[cpi->scaled_ref_idx[i]].ref_count--;
for (i = 0; i < MAX_REF_FRAMES; ++i) {
const int idx = cpi->scaled_ref_idx[i];
RefCntBuffer *const buf =
idx != INVALID_REF_BUFFER_IDX ? &cm->frame_bufs[idx] : NULL;
if (buf != NULL)
--buf->ref_count;
}
}
static void full_to_model_count(unsigned int *model_count,
@ -2527,10 +2538,181 @@ static void output_frame_level_debug_stats(VP9_COMP *cpi) {
}
#endif
static void encode_without_recode_loop(VP9_COMP *cpi,
int q) {
static void set_mv_search_params(VP9_COMP *cpi) {
const VP9_COMMON *const cm = &cpi->common;
const unsigned int max_mv_def = MIN(cm->width, cm->height);
// Default based on max resolution.
cpi->mv_step_param = vp9_init_search_range(max_mv_def);
if (cpi->sf.mv.auto_mv_step_size) {
if (frame_is_intra_only(cm)) {
// Initialize max_mv_magnitude for use in the first INTER frame
// after a key/intra-only frame.
cpi->max_mv_magnitude = max_mv_def;
} else {
if (cm->show_frame) {
// Allow mv_steps to correspond to twice the max mv magnitude found
// in the previous frame, capped by the default max_mv_magnitude based
// on resolution.
cpi->mv_step_param =
vp9_init_search_range(MIN(max_mv_def, 2 * cpi->max_mv_magnitude));
}
cpi->max_mv_magnitude = 0;
}
}
}
static void set_size_dependent_vars(VP9_COMP *cpi, int *q,
int *bottom_index, int *top_index) {
VP9_COMMON *const cm = &cpi->common;
const VP9EncoderConfig *const oxcf = &cpi->oxcf;
// Setup variables that depend on the dimensions of the frame.
set_mv_search_params(cpi);
// Configure experimental use of segmentation for enhanced coding of
// static regions if indicated.
// Only allowed in the second pass of a two pass encode, as it requires
// lagged coding, and if the relevant speed feature flag is set.
if (oxcf->pass == 2 && cpi->sf.static_segmentation)
configure_static_seg_features(cpi);
#if CONFIG_VP9_POSTPROC
if (oxcf->noise_sensitivity > 0) {
int l = 0;
switch (oxcf->noise_sensitivity) {
case 1:
l = 20;
break;
case 2:
l = 40;
break;
case 3:
l = 60;
break;
case 4:
case 5:
l = 100;
break;
case 6:
l = 150;
break;
}
vp9_denoise(cpi->Source, cpi->Source, l);
}
#endif // CONFIG_VP9_POSTPROC
vp9_set_speed_features(cpi);
vp9_set_rd_speed_thresholds(cpi);
vp9_set_rd_speed_thresholds_sub8x8(cpi);
// Decide q and q bounds.
*q = vp9_rc_pick_q_and_bounds(cpi, bottom_index, top_index);
if (!frame_is_intra_only(cm)) {
cm->interp_filter = cpi->sf.default_interp_filter;
vp9_set_high_precision_mv(cpi, (*q) < HIGH_PRECISION_MV_QTHRESH);
}
}
static void init_motion_estimation(VP9_COMP *cpi) {
int y_stride = cpi->scaled_source.y_stride;
if (cpi->sf.mv.search_method == NSTEP) {
vp9_init3smotion_compensation(&cpi->ss_cfg, y_stride);
} else if (cpi->sf.mv.search_method == DIAMOND) {
vp9_init_dsmotion_compensation(&cpi->ss_cfg, y_stride);
}
}
extern void vbr_rate_correction(VP9_COMP *cpi,
int * this_frame_target,
const int64_t vbr_bits_off_target);
void set_frame_size(VP9_COMP *cpi) {
int ref_frame;
VP9_COMMON *const cm = &cpi->common;
const RATE_CONTROL *const rc = &cpi->rc;
const VP9EncoderConfig *const oxcf = &cpi->oxcf;
MACROBLOCKD *const xd = &cpi->mb.e_mbd;
if ((oxcf->pass == 2) &&
(!cpi->use_svc ||
(is_two_pass_svc(cpi) &&
cpi->svc.encode_empty_frame_state != ENCODING))) {
int target_rate = rc->base_frame_target;
if (oxcf->rc_mode == VPX_VBR)
vbr_rate_correction(cpi, &target_rate, rc->vbr_bits_off_target);
vp9_rc_set_frame_target(cpi, target_rate);
}
if (oxcf->pass == 2 &&
cm->current_video_frame == 0 &&
oxcf->allow_spatial_resampling &&
oxcf->rc_mode == VPX_VBR) {
// Internal scaling is triggered on the first frame.
vp9_set_size_literal(cpi, oxcf->scaled_frame_width,
oxcf->scaled_frame_height);
}
// Reset the frame pointers to the current frame size.
vp9_realloc_frame_buffer(get_frame_new_buffer(cm),
cm->width, cm->height,
cm->subsampling_x, cm->subsampling_y,
#if CONFIG_VP9_HIGHBITDEPTH
cm->use_highbitdepth,
#endif
VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL);
alloc_util_frame_buffers(cpi);
init_motion_estimation(cpi);
for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
const int idx = cm->ref_frame_map[get_ref_frame_idx(cpi, ref_frame)];
YV12_BUFFER_CONFIG *const buf = &cm->frame_bufs[idx].buf;
RefBuffer *const ref_buf = &cm->frame_refs[ref_frame - 1];
ref_buf->buf = buf;
ref_buf->idx = idx;
#if CONFIG_VP9_HIGHBITDEPTH
vp9_setup_scale_factors_for_frame(&ref_buf->sf,
buf->y_crop_width, buf->y_crop_height,
cm->width, cm->height,
(buf->flags & YV12_FLAG_HIGHBITDEPTH) ?
1 : 0);
#else
vp9_setup_scale_factors_for_frame(&ref_buf->sf,
buf->y_crop_width, buf->y_crop_height,
cm->width, cm->height);
#endif // CONFIG_VP9_HIGHBITDEPTH
if (vp9_is_scaled(&ref_buf->sf))
vp9_extend_frame_borders(buf);
}
set_ref_ptrs(cm, xd, LAST_FRAME, LAST_FRAME);
}
static void encode_without_recode_loop(VP9_COMP *cpi) {
int q;
int bottom_index, top_index; // Dummy.
VP9_COMMON *const cm = &cpi->common;
vp9_clear_system_state();
set_frame_size(cpi);
cpi->Source = vp9_scale_if_required(cm, cpi->un_scaled_source,
&cpi->scaled_source);
if (cpi->unscaled_last_source != NULL)
cpi->Last_Source = vp9_scale_if_required(cm, cpi->unscaled_last_source,
&cpi->scaled_last_source);
vp9_scale_references(cpi);
set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
vp9_set_quantizer(cm, q);
setup_frame(cpi);
// Variance adaptive and in frame q adjustment experiments are mutually
@ -2553,28 +2735,45 @@ static void encode_without_recode_loop(VP9_COMP *cpi,
static void encode_with_recode_loop(VP9_COMP *cpi,
size_t *size,
uint8_t *dest,
int q,
int bottom_index,
int top_index) {
uint8_t *dest) {
VP9_COMMON *const cm = &cpi->common;
RATE_CONTROL *const rc = &cpi->rc;
int q;
int q_low, q_high;
int bottom_index, top_index;
int loop_count = 0;
int loop = 0;
int overshoot_seen = 0;
int undershoot_seen = 0;
int q_low = bottom_index, q_high = top_index;
int frame_over_shoot_limit;
int frame_under_shoot_limit;
// Decide frame size bounds
vp9_rc_compute_frame_size_bounds(cpi, rc->this_frame_target,
&frame_under_shoot_limit,
&frame_over_shoot_limit);
do {
vp9_clear_system_state();
if (loop_count == 0) {
set_frame_size(cpi);
// Decide frame size bounds
vp9_rc_compute_frame_size_bounds(cpi, rc->this_frame_target,
&frame_under_shoot_limit,
&frame_over_shoot_limit);
cpi->Source = vp9_scale_if_required(cm, cpi->un_scaled_source,
&cpi->scaled_source);
if (cpi->unscaled_last_source != NULL)
cpi->Last_Source = vp9_scale_if_required(cm, cpi->unscaled_last_source,
&cpi->scaled_last_source);
vp9_scale_references(cpi);
set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
q_low = bottom_index;
q_high = top_index;
}
vp9_set_quantizer(cm, q);
if (loop_count == 0)
@ -2816,25 +3015,6 @@ YV12_BUFFER_CONFIG *vp9_scale_if_required(VP9_COMMON *cm,
}
}
static int is_skippable_frame(const VP9_COMP *cpi) {
// If the current frame does not have non-zero motion vector detected in the
// first pass, and so do its previous and forward frames, then this frame
// can be skipped for partition check, and the partition size is assigned
// according to the variance
const SVC *const svc = &cpi->svc;
const TWO_PASS *const twopass = is_two_pass_svc(cpi) ?
&svc->layer_context[svc->spatial_layer_id].twopass : &cpi->twopass;
return (!frame_is_intra_only(&cpi->common) &&
twopass->stats_in - 2 > twopass->stats_in_start &&
twopass->stats_in < twopass->stats_in_end &&
(twopass->stats_in - 1)->pcnt_inter - (twopass->stats_in - 1)->pcnt_motion
== 1 &&
(twopass->stats_in - 2)->pcnt_inter - (twopass->stats_in - 2)->pcnt_motion
== 1 &&
twopass->stats_in->pcnt_inter - twopass->stats_in->pcnt_motion == 1);
}
static void set_arf_sign_bias(VP9_COMP *cpi) {
VP9_COMMON *const cm = &cpi->common;
int arf_sign_bias;
@ -2851,31 +3031,6 @@ static void set_arf_sign_bias(VP9_COMP *cpi) {
cm->ref_frame_sign_bias[ALTREF_FRAME] = arf_sign_bias;
}
static void set_mv_search_params(VP9_COMP *cpi) {
const VP9_COMMON *const cm = &cpi->common;
const unsigned int max_mv_def = MIN(cm->width, cm->height);
// Default based on max resolution.
cpi->mv_step_param = vp9_init_search_range(max_mv_def);
if (cpi->sf.mv.auto_mv_step_size) {
if (frame_is_intra_only(cm)) {
// Initialize max_mv_magnitude for use in the first INTER frame
// after a key/intra-only frame.
cpi->max_mv_magnitude = max_mv_def;
} else {
if (cm->show_frame)
// Allow mv_steps to correspond to twice the max mv magnitude found
// in the previous frame, capped by the default max_mv_magnitude based
// on resolution.
cpi->mv_step_param =
vp9_init_search_range(MIN(max_mv_def, 2 * cpi->max_mv_magnitude));
cpi->max_mv_magnitude = 0;
}
}
}
int setup_interp_filter_search_mask(VP9_COMP *cpi) {
INTERP_FILTER ifilter;
int ref_total[MAX_REF_FRAMES] = {0};
@ -2910,21 +3065,9 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi,
const VP9EncoderConfig *const oxcf = &cpi->oxcf;
struct segmentation *const seg = &cm->seg;
TX_SIZE t;
int q;
int top_index;
int bottom_index;
set_ext_overrides(cpi);
cpi->Source = vp9_scale_if_required(cm, cpi->un_scaled_source,
&cpi->scaled_source);
if (cpi->unscaled_last_source != NULL)
cpi->Last_Source = vp9_scale_if_required(cm, cpi->unscaled_last_source,
&cpi->scaled_last_source);
vp9_scale_references(cpi);
vp9_clear_system_state();
// Enable or disable mode based tweaking of the zbin.
@ -2939,14 +3082,11 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi,
// Set default state for segment based loop filter update flags.
cm->lf.mode_ref_delta_update = 0;
set_mv_search_params(cpi);
if (cpi->oxcf.pass == 2 &&
cpi->sf.adaptive_interp_filter_search)
cpi->sf.interp_filter_search_mask =
setup_interp_filter_search_mask(cpi);
// Set various flags etc to special state if it is a key frame.
if (frame_is_intra_only(cm)) {
// Reset the loop filter deltas and segmentation map.
@ -3006,20 +3146,6 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi,
}
}
// Configure experimental use of segmentation for enhanced coding of
// static regions if indicated.
// Only allowed in second pass of two pass (as requires lagged coding)
// and if the relevant speed feature flag is set.
if (oxcf->pass == 2 && cpi->sf.static_segmentation)
configure_static_seg_features(cpi);
// Check if the current frame is skippable for the partition search in the
// second pass according to the first pass stats
if (cpi->sf.allow_partition_search_skip && oxcf->pass == 2 &&
(!cpi->use_svc || is_two_pass_svc(cpi))) {
cpi->partition_search_skippable_frame = is_skippable_frame(cpi);
}
// For 1 pass CBR, check if we are dropping this frame.
// Never drop on key frame.
if (oxcf->pass == 0 &&
@ -3034,31 +3160,6 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi,
vp9_clear_system_state();
#if CONFIG_VP9_POSTPROC
if (oxcf->noise_sensitivity > 0) {
int l = 0;
switch (oxcf->noise_sensitivity) {
case 1:
l = 20;
break;
case 2:
l = 40;
break;
case 3:
l = 60;
break;
case 4:
case 5:
l = 100;
break;
case 6:
l = 150;
break;
}
vp9_denoise(cpi->Source, cpi->Source, l);
}
#endif
#if CONFIG_INTERNAL_STATS
{
int i;
@ -3067,24 +3168,10 @@ static void encode_frame_to_data_rate(VP9_COMP *cpi,
}
#endif
vp9_set_speed_features(cpi);
vp9_set_rd_speed_thresholds(cpi);
vp9_set_rd_speed_thresholds_sub8x8(cpi);
// Decide q and q bounds.
q = vp9_rc_pick_q_and_bounds(cpi, &bottom_index, &top_index);
if (!frame_is_intra_only(cm)) {
cm->interp_filter = cpi->sf.default_interp_filter;
/* TODO: Decide this more intelligently */
vp9_set_high_precision_mv(cpi, q < HIGH_PRECISION_MV_QTHRESH);
}
if (cpi->sf.recode_loop == DISALLOW_RECODE) {
encode_without_recode_loop(cpi, q);
encode_without_recode_loop(cpi);
} else {
encode_with_recode_loop(cpi, size, dest, q, bottom_index, top_index);
encode_with_recode_loop(cpi, size, dest);
}
#if CONFIG_VP9_TEMPORAL_DENOISING
@ -3227,16 +3314,6 @@ static void Pass2Encode(VP9_COMP *cpi, size_t *size,
vp9_twopass_postencode_update(cpi);
}
static void init_motion_estimation(VP9_COMP *cpi) {
int y_stride = cpi->scaled_source.y_stride;
if (cpi->sf.mv.search_method == NSTEP) {
vp9_init3smotion_compensation(&cpi->ss_cfg, y_stride);
} else if (cpi->sf.mv.search_method == DIAMOND) {
vp9_init_dsmotion_compensation(&cpi->ss_cfg, y_stride);
}
}
static void check_initial_width(VP9_COMP *cpi,
#if CONFIG_VP9_HIGHBITDEPTH
int use_highbitdepth,
@ -3255,10 +3332,11 @@ static void check_initial_width(VP9_COMP *cpi,
alloc_ref_frame_buffers(cpi);
alloc_util_frame_buffers(cpi);
init_motion_estimation(cpi);
init_motion_estimation(cpi); // TODO(agrange) This can be removed.
cpi->initial_width = cm->width;
cpi->initial_height = cm->height;
cpi->initial_mbs = cm->MBs;
}
}
@ -3400,14 +3478,13 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
int64_t *time_stamp, int64_t *time_end, int flush) {
const VP9EncoderConfig *const oxcf = &cpi->oxcf;
VP9_COMMON *const cm = &cpi->common;
MACROBLOCKD *const xd = &cpi->mb.e_mbd;
RATE_CONTROL *const rc = &cpi->rc;
struct vpx_usec_timer cmptimer;
YV12_BUFFER_CONFIG *force_src_buffer = NULL;
struct lookahead_entry *last_source = NULL;
struct lookahead_entry *source = NULL;
MV_REFERENCE_FRAME ref_frame;
int arf_src_index;
int i;
if (is_two_pass_svc(cpi)) {
#if CONFIG_SPATIAL_SVC
@ -3546,24 +3623,11 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
vp9_restore_layer_context(cpi);
}
// start with a 0 size frame
*size = 0;
/* find a free buffer for the new frame, releasing the reference previously
* held.
*/
// Find a free buffer for the new frame, releasing the reference previously
// held.
cm->frame_bufs[cm->new_fb_idx].ref_count--;
cm->new_fb_idx = get_free_fb(cm);
// 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) &&
cpi->svc.encode_empty_frame_state != ENCODING))) {
vp9_rc_get_second_pass_params(cpi);
}
if (!cpi->use_svc && cpi->multi_arf_allowed) {
if (cm->frame_type == KEY_FRAME) {
init_buffer_indices(cpi);
@ -3573,56 +3637,27 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
}
}
// Start with a 0 size frame.
*size = 0;
cpi->frame_flags = *frame_flags;
if (oxcf->pass == 2 &&
cm->current_video_frame == 0 &&
oxcf->allow_spatial_resampling &&
oxcf->rc_mode == VPX_VBR) {
// Internal scaling is triggered on the first frame.
vp9_set_size_literal(cpi, oxcf->scaled_frame_width,
oxcf->scaled_frame_height);
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);
} else {
set_frame_size(cpi);
}
// Reset the frame pointers to the current frame size
vp9_realloc_frame_buffer(get_frame_new_buffer(cm),
cm->width, cm->height,
cm->subsampling_x, cm->subsampling_y,
#if CONFIG_VP9_HIGHBITDEPTH
cm->use_highbitdepth,
#endif
VP9_ENC_BORDER_IN_PIXELS, NULL, NULL, NULL);
alloc_util_frame_buffers(cpi);
init_motion_estimation(cpi);
for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
const int idx = cm->ref_frame_map[get_ref_frame_idx(cpi, ref_frame)];
YV12_BUFFER_CONFIG *const buf = &cm->frame_bufs[idx].buf;
RefBuffer *const ref_buf = &cm->frame_refs[ref_frame - 1];
ref_buf->buf = buf;
ref_buf->idx = idx;
#if CONFIG_VP9_HIGHBITDEPTH
vp9_setup_scale_factors_for_frame(&ref_buf->sf,
buf->y_crop_width, buf->y_crop_height,
cm->width, cm->height,
(buf->flags & YV12_FLAG_HIGHBITDEPTH) ?
1 : 0);
#else
vp9_setup_scale_factors_for_frame(&ref_buf->sf,
buf->y_crop_width, buf->y_crop_height,
cm->width, cm->height);
#endif // CONFIG_VP9_HIGHBITDEPTH
if (vp9_is_scaled(&ref_buf->sf))
vp9_extend_frame_borders(buf);
}
set_ref_ptrs(cm, xd, LAST_FRAME, LAST_FRAME);
if (oxcf->aq_mode == VARIANCE_AQ) {
vp9_vaq_init();
}
for (i = 0; i < MAX_REF_FRAMES; ++i)
cpi->scaled_ref_idx[i] = INVALID_REF_BUFFER_IDX;
if (oxcf->pass == 1 &&
(!cpi->use_svc || is_two_pass_svc(cpi))) {
const int lossless = is_lossless_requested(oxcf);
@ -3651,7 +3686,7 @@ int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
if (cm->refresh_frame_context)
cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
// Frame was dropped, release scaled references.
// No frame encoded, or frame was dropped, release scaled references.
if (*size == 0) {
release_scaled_references(cpi);
}

7
vp9/encoder/vp9_encoder.h
View File

@ -44,6 +44,7 @@ extern "C" {
#endif
#define DEFAULT_GF_INTERVAL 10
#define INVALID_REF_BUFFER_IDX -1 // Marks an invalid reference buffer id.
typedef struct {
int nmvjointcost[MV_JOINTS];
@ -241,7 +242,7 @@ typedef struct VP9_COMP {
// For a still frame, this flag is set to 1 to skip partition search.
int partition_search_skippable_frame;
int scaled_ref_idx[3];
int scaled_ref_idx[MAX_REF_FRAMES];
int lst_fb_idx;
int gld_fb_idx;
int alt_fb_idx;
@ -374,6 +375,10 @@ typedef struct VP9_COMP {
int initial_width;
int initial_height;
int initial_mbs; // Number of MBs in the full-size frame; to be used to
// normalize the firstpass stats. This will differ from the
// number of MBs in the current frame when the frame is
// scaled.
int use_svc;

106
vp9/encoder/vp9_firstpass.c
View File

@ -138,7 +138,7 @@ static void output_fpmb_stats(uint8_t *this_frame_mb_stats, VP9_COMMON *cm,
struct vpx_codec_cx_pkt pkt;
pkt.kind = VPX_CODEC_FPMB_STATS_PKT;
pkt.data.firstpass_mb_stats.buf = this_frame_mb_stats;
pkt.data.firstpass_mb_stats.sz = cm->MBs * sizeof(uint8_t);
pkt.data.firstpass_mb_stats.sz = cm->initial_mbs * sizeof(uint8_t);
vpx_codec_pkt_list_add(pktlist, &pkt);
}
#endif
@ -483,7 +483,7 @@ void vp9_first_pass(VP9_COMP *cpi, const struct lookahead_entry *source) {
#if CONFIG_FP_MB_STATS
if (cpi->use_fp_mb_stats) {
vp9_zero_array(cpi->twopass.frame_mb_stats_buf, cm->MBs);
vp9_zero_array(cpi->twopass.frame_mb_stats_buf, cm->initial_mbs);
}
#endif
@ -934,12 +934,14 @@ void vp9_first_pass(VP9_COMP *cpi, const struct lookahead_entry *source) {
vp9_clear_system_state();
{
FIRSTPASS_STATS fps;
// The minimum error here insures some bit alocation to frames even
// The minimum error here insures some bit allocation to frames even
// in static regions. The allocation per MB declines for larger formats
// where the typical "real" energy per MB also falls.
// Initial estimate here uses sqrt(mbs) to define the min_err, where the
// number of mbs is propotional to image area.
const double min_err = 200 * sqrt(cm->MBs);
// number of mbs is proportional to the image area.
const int num_mbs =
cpi->oxcf.allow_spatial_resampling ? cpi->initial_mbs : cpi->common.MBs;
const double min_err = 200 * sqrt(num_mbs);
fps.frame = cm->current_video_frame;
fps.spatial_layer_id = cpi->svc.spatial_layer_id;
@ -947,9 +949,9 @@ void vp9_first_pass(VP9_COMP *cpi, const struct lookahead_entry *source) {
fps.sr_coded_error = (double)(sr_coded_error >> 8) + min_err;
fps.intra_error = (double)(intra_error >> 8) + min_err;
fps.count = 1.0;
fps.pcnt_inter = (double)intercount / cm->MBs;
fps.pcnt_second_ref = (double)second_ref_count / cm->MBs;
fps.pcnt_neutral = (double)neutral_count / cm->MBs;
fps.pcnt_inter = (double)intercount / num_mbs;
fps.pcnt_second_ref = (double)second_ref_count / num_mbs;
fps.pcnt_neutral = (double)neutral_count / num_mbs;
if (mvcount > 0) {
fps.MVr = (double)sum_mvr / mvcount;
@ -960,7 +962,7 @@ void vp9_first_pass(VP9_COMP *cpi, const struct lookahead_entry *source) {
fps.MVcv = ((double)sum_mvcs - (fps.MVc * fps.MVc / mvcount)) / mvcount;
fps.mv_in_out_count = (double)sum_in_vectors / (mvcount * 2);
fps.new_mv_count = new_mv_count;
fps.pcnt_motion = (double)mvcount / cm->MBs;
fps.pcnt_motion = (double)mvcount / num_mbs;
} else {
fps.MVr = 0.0;
fps.mvr_abs = 0.0;
@ -1074,7 +1076,8 @@ static int get_twopass_worst_quality(const VP9_COMP *cpi,
if (section_target_bandwidth <= 0) {
return rc->worst_quality; // Highest value allowed
} else {
const int num_mbs = cpi->common.MBs;
const int num_mbs =
cpi->oxcf.allow_spatial_resampling ? cpi->initial_mbs : cpi->common.MBs;
const double section_err = stats->coded_error / stats->count;
const double err_per_mb = section_err / num_mbs;
const double speed_term = 1.0 + 0.04 * oxcf->speed;
@ -1188,9 +1191,12 @@ void vp9_init_second_pass(VP9_COMP *cpi) {
#define LOW_SR_DIFF_TRHESH 0.1
#define SR_DIFF_MAX 128.0
static double get_sr_decay_rate(const VP9_COMMON *cm,
static double get_sr_decay_rate(const VP9_COMP *cpi,
const FIRSTPASS_STATS *frame) {
double sr_diff = (frame->sr_coded_error - frame->coded_error) / cm->MBs;
const int num_mbs =
cpi->oxcf.allow_spatial_resampling ? cpi->initial_mbs : cpi->common.MBs;
double sr_diff =
(frame->sr_coded_error - frame->coded_error) / num_mbs;
double sr_decay = 1.0;
const double motion_amplitude_factor =
frame->pcnt_motion * ((frame->mvc_abs + frame->mvr_abs) / 2);
@ -1207,19 +1213,19 @@ static double get_sr_decay_rate(const VP9_COMMON *cm,
// This function gives an estimate of how badly we believe the prediction
// quality is decaying from frame to frame.
static double get_zero_motion_factor(const VP9_COMMON *cm,
static double get_zero_motion_factor(const VP9_COMP *cpi,
const FIRSTPASS_STATS *frame) {
const double zero_motion_pct = frame->pcnt_inter -
frame->pcnt_motion;
double sr_decay = get_sr_decay_rate(cm, frame);
double sr_decay = get_sr_decay_rate(cpi, frame);
return MIN(sr_decay, zero_motion_pct);
}
#define ZM_POWER_FACTOR 0.75
static double get_prediction_decay_rate(const VP9_COMMON *cm,
static double get_prediction_decay_rate(const VP9_COMP *cpi,
const FIRSTPASS_STATS *next_frame) {
const double sr_decay_rate = get_sr_decay_rate(cm, next_frame);
const double sr_decay_rate = get_sr_decay_rate(cpi, next_frame);
const double zero_motion_factor =
(0.95 * pow((next_frame->pcnt_inter - next_frame->pcnt_motion),
ZM_POWER_FACTOR));
@ -1314,9 +1320,11 @@ static double calc_frame_boost(VP9_COMP *cpi,
vp9_convert_qindex_to_q(cpi->rc.avg_frame_qindex[INTER_FRAME],
cpi->common.bit_depth);
const double boost_correction = MIN((0.5 + (lq * 0.015)), 1.5);
const int num_mbs =
cpi->oxcf.allow_spatial_resampling ? cpi->initial_mbs : cpi->common.MBs;
// Underlying boost factor is based on inter error ratio.
frame_boost = (BASELINE_ERR_PER_MB * cpi->common.MBs) /
frame_boost = (BASELINE_ERR_PER_MB * num_mbs) /
DOUBLE_DIVIDE_CHECK(this_frame->coded_error);
frame_boost = frame_boost * BOOST_FACTOR * boost_correction;
@ -1365,7 +1373,7 @@ static int calc_arf_boost(VP9_COMP *cpi, int offset,
// Accumulate the effect of prediction quality decay.
if (!flash_detected) {
decay_accumulator *= get_prediction_decay_rate(&cpi->common, this_frame);
decay_accumulator *= get_prediction_decay_rate(cpi, this_frame);
decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
? MIN_DECAY_FACTOR : decay_accumulator;
}
@ -1404,7 +1412,7 @@ static int calc_arf_boost(VP9_COMP *cpi, int offset,
// Cumulative effect of prediction quality decay.
if (!flash_detected) {
decay_accumulator *= get_prediction_decay_rate(&cpi->common, this_frame);
decay_accumulator *= get_prediction_decay_rate(cpi, this_frame);
decay_accumulator = decay_accumulator < MIN_DECAY_FACTOR
? MIN_DECAY_FACTOR : decay_accumulator;
}
@ -1723,7 +1731,8 @@ static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
gf_group_err -= gf_first_frame_err;
// Motion breakout threshold for loop below depends on image size.
mv_ratio_accumulator_thresh = (cpi->common.width + cpi->common.height) / 4.0;
mv_ratio_accumulator_thresh =
(cpi->common.height + cpi->common.width) / 4.0;
// Set a maximum and minimum interval for the GF group.
// If the image appears almost completely static we can extend beyond this.
@ -1775,14 +1784,14 @@ static void define_gf_group(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
// Accumulate the effect of prediction quality decay.
if (!flash_detected) {
last_loop_decay_rate = loop_decay_rate;
loop_decay_rate = get_prediction_decay_rate(&cpi->common, &next_frame);
loop_decay_rate = get_prediction_decay_rate(cpi, &next_frame);
decay_accumulator = decay_accumulator * loop_decay_rate;
// Monitor for static sections.
zero_motion_accumulator =
MIN(zero_motion_accumulator,
get_zero_motion_factor(&cpi->common, &next_frame));
get_zero_motion_factor(cpi, &next_frame));
// Break clause to detect very still sections after motion. For example,
// a static image after a fade or other transition.
@ -2048,8 +2057,7 @@ static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
break;
// How fast is the prediction quality decaying?
loop_decay_rate = get_prediction_decay_rate(&cpi->common,
twopass->stats_in);
loop_decay_rate = get_prediction_decay_rate(cpi, twopass->stats_in);
// We want to know something about the recent past... rather than
// as used elsewhere where we are concerned with decay in prediction
@ -2160,7 +2168,7 @@ static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
// Monitor for static sections.
zero_motion_accumulator =
MIN(zero_motion_accumulator,
get_zero_motion_factor(&cpi->common, &next_frame));
get_zero_motion_factor(cpi, &next_frame));
// Not all frames in the group are necessarily used in calculating boost.
if ((i <= rc->max_gf_interval) ||
@ -2171,7 +2179,7 @@ static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame) {
// How fast is prediction quality decaying.
if (!detect_flash(twopass, 0)) {
const double loop_decay_rate =
get_prediction_decay_rate(&cpi->common, &next_frame);
get_prediction_decay_rate(cpi, &next_frame);
decay_accumulator *= loop_decay_rate;
decay_accumulator = MAX(decay_accumulator, MIN_DECAY_FACTOR);
av_decay_accumulator += decay_accumulator;
@ -2295,6 +2303,24 @@ void configure_buffer_updates(VP9_COMP *cpi) {
}
}
int is_skippable_frame(const VP9_COMP *cpi) {
// If the current frame does not have non-zero motion vector detected in the
// first pass, and so do its previous and forward frames, then this frame
// can be skipped for partition check, and the partition size is assigned
// according to the variance
const SVC *const svc = &cpi->svc;
const TWO_PASS *const twopass = is_two_pass_svc(cpi) ?
&svc->layer_context[svc->spatial_layer_id].twopass : &cpi->twopass;
return (!frame_is_intra_only(&cpi->common) &&
twopass->stats_in - 2 > twopass->stats_in_start &&
twopass->stats_in < twopass->stats_in_end &&
(twopass->stats_in - 1)->pcnt_inter - (twopass->stats_in - 1)->pcnt_motion
== 1 &&
(twopass->stats_in - 2)->pcnt_inter - (twopass->stats_in - 2)->pcnt_motion
== 1 &&
twopass->stats_in->pcnt_inter - twopass->stats_in->pcnt_motion == 1);
}
void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
VP9_COMMON *const cm = &cpi->common;
@ -2329,11 +2355,6 @@ void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
target_rate = vp9_rc_clamp_pframe_target_size(cpi, target_rate);
rc->base_frame_target = target_rate;
// Correction to rate target based on prior over or under shoot.
if (cpi->oxcf.rc_mode == VPX_VBR)
vbr_rate_correction(cpi, &target_rate, rc->vbr_bits_off_target);
vp9_rc_set_frame_target(cpi, target_rate);
cm->frame_type = INTER_FRAME;
if (lc != NULL) {
@ -2347,6 +2368,13 @@ void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
}
}
// Do the firstpass stats indicate that this frame is skippable for the
// partition search?
if (cpi->sf.allow_partition_search_skip &&
cpi->oxcf.pass == 2 && (!cpi->use_svc || is_two_pass_svc(cpi))) {
cpi->partition_search_skippable_frame = is_skippable_frame(cpi);
}
return;
}
@ -2377,8 +2405,7 @@ void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
this_frame_copy = this_frame;
// Keyframe and section processing.
if (rc->frames_to_key == 0 ||
(cpi->frame_flags & FRAMEFLAGS_KEY)) {
if (rc->frames_to_key == 0 || (cpi->frame_flags & FRAMEFLAGS_KEY)) {
// Define next KF group and assign bits to it.
find_next_key_frame(cpi, &this_frame_copy);
} else {
@ -2431,6 +2458,13 @@ void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
configure_buffer_updates(cpi);
// Do the firstpass stats indicate that this frame is skippable for the
// partition search?
if (cpi->sf.allow_partition_search_skip && cpi->oxcf.pass == 2 &&
(!cpi->use_svc || is_two_pass_svc(cpi))) {
cpi->partition_search_skippable_frame = is_skippable_frame(cpi);
}
target_rate = gf_group->bit_allocation[gf_group->index];
if (cpi->common.frame_type == KEY_FRAME)
target_rate = vp9_rc_clamp_iframe_target_size(cpi, target_rate);
@ -2439,12 +2473,6 @@ void vp9_rc_get_second_pass_params(VP9_COMP *cpi) {
rc->base_frame_target = target_rate;
// Correction to rate target based on prior over or under shoot.
if (cpi->oxcf.rc_mode == VPX_VBR)
vbr_rate_correction(cpi, &target_rate, rc->vbr_bits_off_target);
vp9_rc_set_frame_target(cpi, target_rate);
// Update the total stats remaining structure.
subtract_stats(&twopass->total_left_stats, &this_frame);
}