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

Merge "Fixed 8K two pass encoder crash."

Browse Source
This commit is contained in:
Paul Wilkins 2016-05-11 16:25:25 +00:00 committed by Gerrit Code Review
commit 5fd142e763
2 changed files with 39 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
vp10/encoder/firstpass.c
View File

@ -45,7 +45,6 @@
#define BOOST_BREAKOUT 12.5 #define BOOST_BREAKOUT 12.5
#define BOOST_FACTOR 12.5 #define BOOST_FACTOR 12.5
#define ERR_DIVISOR 128.0
#define FACTOR_PT_LOW 0.70 #define FACTOR_PT_LOW 0.70
#define FACTOR_PT_HIGH 0.90 #define FACTOR_PT_HIGH 0.90
#define FIRST_PASS_Q 10.0 #define FIRST_PASS_Q 10.0
@ -231,6 +230,13 @@ static void subtract_stats(FIRSTPASS_STATS *section,
section->duration -= frame->duration; section->duration -= frame->duration;
} }
// Calculate the linear size relative to a baseline of 1080P
#define BASE_SIZE 2073600.0 // 1920x1080
static double get_linear_size_factor(const VP10_COMP *cpi) {
const double this_area = cpi->initial_width * cpi->initial_height;
return pow(this_area / BASE_SIZE, 0.5);
}
// Calculate an active area of the image that discounts formatting // Calculate an active area of the image that discounts formatting
// bars and partially discounts other 0 energy areas. // bars and partially discounts other 0 energy areas.
#define MIN_ACTIVE_AREA 0.5 #define MIN_ACTIVE_AREA 0.5
@ -1103,11 +1109,7 @@ static double calc_correction_factor(double err_per_mb,
return fclamp(pow(error_term, power_term), 0.05, 5.0); return fclamp(pow(error_term, power_term), 0.05, 5.0);
} }
// Larger image formats are expected to be a little harder to code relatively #define ERR_DIVISOR 100.0
// given the same prediction error score. This in part at least relates to the
// increased size and hence coding cost of motion vectors.
#define EDIV_SIZE_FACTOR 800
static int get_twopass_worst_quality(const VP10_COMP *cpi, static int get_twopass_worst_quality(const VP10_COMP *cpi,
const double section_err, const double section_err,
double inactive_zone, double inactive_zone,
@ -1126,12 +1128,22 @@ static int get_twopass_worst_quality(const VP10_COMP *cpi,
const int active_mbs = VPXMAX(1, num_mbs - (int)(num_mbs * inactive_zone)); const int active_mbs = VPXMAX(1, num_mbs - (int)(num_mbs * inactive_zone));
const double av_err_per_mb = section_err / active_mbs; const double av_err_per_mb = section_err / active_mbs;
const double speed_term = 1.0 + 0.04 * oxcf->speed; const double speed_term = 1.0 + 0.04 * oxcf->speed;
const double ediv_size_correction = (double)num_mbs / EDIV_SIZE_FACTOR; double ediv_size_correction;
const int target_norm_bits_per_mb = ((uint64_t)section_target_bandwidth << const int target_norm_bits_per_mb = ((uint64_t)section_target_bandwidth <<
BPER_MB_NORMBITS) / active_mbs; BPER_MB_NORMBITS) / active_mbs;
int q; int q;
// Larger image formats are expected to be a little harder to code
// relatively given the same prediction error score. This in part at
// least relates to the increased size and hence coding overheads of
// motion vectors. Some account of this is made through adjustment of
// the error divisor.
ediv_size_correction =
VPXMAX(0.2, VPXMIN(5.0, get_linear_size_factor(cpi)));
if (ediv_size_correction < 1.0)
ediv_size_correction = -(1.0 / ediv_size_correction);
ediv_size_correction *= 4.0;
// Try and pick a max Q that will be high enough to encode the // Try and pick a max Q that will be high enough to encode the
// content at the given rate. // content at the given rate.
for (q = rc->best_quality; q < rc->worst_quality; ++q) { for (q = rc->best_quality; q < rc->worst_quality; ++q) {

27
vp9/encoder/vp9_firstpass.c
View File

@ -45,7 +45,6 @@
#define BOOST_BREAKOUT 12.5 #define BOOST_BREAKOUT 12.5
#define BOOST_FACTOR 12.5 #define BOOST_FACTOR 12.5
#define ERR_DIVISOR 128.0
#define FACTOR_PT_LOW 0.70 #define FACTOR_PT_LOW 0.70
#define FACTOR_PT_HIGH 0.90 #define FACTOR_PT_HIGH 0.90
#define FIRST_PASS_Q 10.0 #define FIRST_PASS_Q 10.0
@ -237,6 +236,13 @@ static void subtract_stats(FIRSTPASS_STATS *section,
section->duration -= frame->duration; section->duration -= frame->duration;
} }
// Calculate the linear size relative to a baseline of 1080P
#define BASE_SIZE 2073600.0 // 1920x1080
static double get_linear_size_factor(const VP9_COMP *cpi) {
const double this_area = cpi->initial_width * cpi->initial_height;
return pow(this_area / BASE_SIZE, 0.5);
}
// Calculate an active area of the image that discounts formatting // Calculate an active area of the image that discounts formatting
// bars and partially discounts other 0 energy areas. // bars and partially discounts other 0 energy areas.
#define MIN_ACTIVE_AREA 0.5 #define MIN_ACTIVE_AREA 0.5
@ -1241,11 +1247,7 @@ static double calc_correction_factor(double err_per_mb,
return fclamp(pow(error_term, power_term), 0.05, 5.0); return fclamp(pow(error_term, power_term), 0.05, 5.0);
} }
// Larger image formats are expected to be a little harder to code relatively #define ERR_DIVISOR 100.0
// given the same prediction error score. This in part at least relates to the
// increased size and hence coding cost of motion vectors.
#define EDIV_SIZE_FACTOR 800
static int get_twopass_worst_quality(const VP9_COMP *cpi, static int get_twopass_worst_quality(const VP9_COMP *cpi,
const double section_err, const double section_err,
double inactive_zone, double inactive_zone,
@ -1267,16 +1269,25 @@ static int get_twopass_worst_quality(const VP9_COMP *cpi,
const int active_mbs = VPXMAX(1, num_mbs - (int)(num_mbs * inactive_zone)); const int active_mbs = VPXMAX(1, num_mbs - (int)(num_mbs * inactive_zone));
const double av_err_per_mb = section_err / active_mbs; const double av_err_per_mb = section_err / active_mbs;
const double speed_term = 1.0 + 0.04 * oxcf->speed; const double speed_term = 1.0 + 0.04 * oxcf->speed;
const double ediv_size_correction = (double)num_mbs / EDIV_SIZE_FACTOR; double ediv_size_correction;
const int target_norm_bits_per_mb = ((uint64_t)target_rate << const int target_norm_bits_per_mb = ((uint64_t)target_rate <<
BPER_MB_NORMBITS) / active_mbs; BPER_MB_NORMBITS) / active_mbs;
int q; int q;
int is_svc_upper_layer = 0; int is_svc_upper_layer = 0;
if (is_two_pass_svc(cpi) && cpi->svc.spatial_layer_id > 0) if (is_two_pass_svc(cpi) && cpi->svc.spatial_layer_id > 0)
is_svc_upper_layer = 1; is_svc_upper_layer = 1;
// Larger image formats are expected to be a little harder to code
// relatively given the same prediction error score. This in part at
// least relates to the increased size and hence coding overheads of
// motion vectors. Some account of this is made through adjustment of
// the error divisor.
ediv_size_correction =
VPXMAX(0.2, VPXMIN(5.0, get_linear_size_factor(cpi)));
if (ediv_size_correction < 1.0)
ediv_size_correction = -(1.0 / ediv_size_correction);
ediv_size_correction *= 4.0;
// Try and pick a max Q that will be high enough to encode the // Try and pick a max Q that will be high enough to encode the
// content at the given rate. // content at the given rate.