generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity
b3ade15a26
vpx/vp8/encoder/rdopt.c

3075 lines
100 KiB
C
Raw Normal View History

Initial WebM release
2010-05-18 17:58:33 +02:00
/*
Use WebM in copyright notice for consistency Changes 'The VP8 project' to 'The WebM project', for consistency with other webmproject.org repositories. Fixes issue #97. Change-Id: I37c13ed5fbdb9d334ceef71c6350e9febed9bbba
2010-09-09 14:16:39 +02:00
* Copyright (c) 2010 The WebM project authors. All Rights Reserved.
Initial WebM release
2010-05-18 17:58:33 +02:00
*
cosmetics: trim trailing whitespace When the license headers were updated, they accidentally contained trailing whitespace, so unfortunately we have to touch all the files again. Change-Id: I236c05fade06589e417179c0444cb39b09e4200d
2010-06-18 18:39:21 +02:00
* Use of this source code is governed by a BSD-style license
LICENSE: update with latest text Change-Id: Ieebea089095d9073b3a94932791099f614ce120c
2010-06-04 22:19:40 +02:00
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
cosmetics: trim trailing whitespace When the license headers were updated, they accidentally contained trailing whitespace, so unfortunately we have to touch all the files again. Change-Id: I236c05fade06589e417179c0444cb39b09e4200d
2010-06-18 18:39:21 +02:00
* in the file PATENTS. All contributing project authors may
LICENSE: update with latest text Change-Id: Ieebea089095d9073b3a94932791099f614ce120c
2010-06-04 22:19:40 +02:00
* be found in the AUTHORS file in the root of the source tree.
Initial WebM release
2010-05-18 17:58:33 +02:00
*/
#include <stdio.h>
#include <math.h>
#include <limits.h>
#include <assert.h>
Fix relative include paths Allow compiling without adding vp8/{common,encoder,decoder} to the include paths. Change-Id: Ifeb5dac351cdfadcd659736f5158b315a0030b6c
2011-02-10 20:41:38 +01:00
#include "vp8/common/pragmas.h"
Initial WebM release
2010-05-18 17:58:33 +02:00
#include "tokenize.h"
#include "treewriter.h"
#include "onyx_int.h"
#include "modecosts.h"
#include "encodeintra.h"
Fix relative include paths Allow compiling without adding vp8/{common,encoder,decoder} to the include paths. Change-Id: Ifeb5dac351cdfadcd659736f5158b315a0030b6c
2011-02-10 20:41:38 +01:00
#include "vp8/common/entropymode.h"
#include "vp8/common/reconinter.h"
#include "vp8/common/reconintra.h"
#include "vp8/common/reconintra4x4.h"
#include "vp8/common/findnearmv.h"
Initial WebM release
2010-05-18 17:58:33 +02:00
#include "encodemb.h"
#include "quantize.h"
Fix relative include paths Allow compiling without adding vp8/{common,encoder,decoder} to the include paths. Change-Id: Ifeb5dac351cdfadcd659736f5158b315a0030b6c
2011-02-10 20:41:38 +01:00
#include "vp8/common/idct.h"
#include "vp8/common/g_common.h"
Initial WebM release
2010-05-18 17:58:33 +02:00
#include "variance.h"
#include "mcomp.h"
Define RDCOST only once Clean up the code. Change-Id: I7db048efa4d972b528d553a7921bc45979621129
2011-04-07 22:57:25 +02:00
#include "rdopt.h"
Removal of Qindex LUTS. One of the problems arising when tweaking or adjusting the quantizer tables is that there are a lot of look up tables that depend on the QINDEX. Any adjustment to the link between QINDEX and real quantizer therefore tends to break aspects of for example the rate control. In this check in I have replaced several of the look up tables with functions that approximate the same results as the old Q luts but use a formulaic approach based on real Q values rather than QIndex. This should hopefully make it easier to experiment with changes to the Q tables without always having to go through and hand optimize a set of look up tables. Once things stabilize we may choose to re-instate luts for the sake of performance. Patch 2: Addressed Ronald's comments. vp8_init_me_luts() Added so luts only initialized once. Change-Id: Ic80db2212d2fd01e08e8cb5c7dca1fda1102be57
2011-11-21 16:45:10 +01:00
#include "ratectrl.h"
Initial WebM release
2010-05-18 17:58:33 +02:00
#include "vpx_mem/vpx_mem.h"
#include "dct.h"
Fix relative include paths Allow compiling without adding vp8/{common,encoder,decoder} to the include paths. Change-Id: Ifeb5dac351cdfadcd659736f5158b315a0030b6c
2011-02-10 20:41:38 +01:00
#include "vp8/common/systemdependent.h"
Initial WebM release
2010-05-18 17:58:33 +02:00
Segment Features. Removal of #ifdefs Removal of configure #ifdefs so that segment features always available. Removal of code supporting old segment feature method. Still a good deal of tidying up to do. Change-Id: I397855f086f8c09ab1fae0a5f65d9e06d2e3e39f
2011-11-03 13:50:09 +01:00
//#if CONFIG_SEGFEATURES
Further segment feature extensions. This quite large check in includes the following: Merge in some code from Ronald (mbgraph.c) that scans a Gf/arf group. This is used as a basis for a simple segmentation for the normal frames in a gf/arf group. This code also uses satd functions from Yaowu. Adds functionality for coding the latest possible position of an EOB for blocks in the segment. (Currently 0-15 only, hence just for 4x4 dct). Where the EOB position is 0 this acts like "skip" and the normal coding of skip at the per mb level is disabled. Added functions (seg_common.c) for setting and reading segment feature elements. These may want to be optimized away at some point but while the mecahnism is in a state of flux they provide a single location for making changes and keep things a bit cleaner. This is still proof of concept code. Currently the tested feature set:- Quantizer, Loop Filter level, Reference frame, Prediction Mode, EOB end stop. TBD:- Add functions for setting and reading the feature data with range and validity checking. Handling of signed and unsigned feature data. At the moment all is assumed to be signed and a sign bit is coded but many cannot be negative. Correct handling of EOB feature with intra coded blocks. Testing/trapping of legal/illegal ref frame and mode combinations. Transform size switch plus merge and test with 8c8 DCT work Merge and test with Sumans Segmenation coding optimizations Change-Id: Iee12e83661c7abbd1e0ce6810915eb4ec35e2d8e
2011-10-05 12:26:00 +02:00
#include "vp8/common/seg_common.h"
Initial WebM release
2010-05-18 17:58:33 +02:00
#if CONFIG_RUNTIME_CPU_DETECT
#define IF_RTCD(x) (x)
#else
#define IF_RTCD(x) NULL
#endif
Added vp8_update_zbin_extra vp8cx_mb_init_quantizer was being called for every mode checked in vp8_rd_pick_inter_mode. zbin_extra is the only value that really needs to be recalculated. This calculation is disabled when using the fast quantizer for mode selection. This gave a small performance boost (~.5% to 1%). Note: This needs to be verified with segmentation_enabled. Change-Id: I62716a870b3c82b4a998bdf95130ff0b02106f1e
2010-12-29 20:30:57 +01:00
extern void vp8cx_mb_init_quantizer(VP8_COMP *cpi, MACROBLOCK *x);
extern void vp8_update_zbin_extra(VP8_COMP *cpi, MACROBLOCK *x);
Initial WebM release
2010-05-18 17:58:33 +02:00
#define MAXF(a,b) (((a) > (b)) ? (a) : (b))
Increase static linkage, remove unused functions A large number of functions were defined with external linkage, even though they were only used from within one file. This patch changes their linkage to static and removes the vp8_ prefix from their names, which should make it more obvious to the reader that the function is contained within the current translation unit. Functions that were not referenced were removed. These symbols were identified by: $ nm -A libvpx.a | sort -k3 | uniq -c -f2 | grep ' [A-Z] ' \ | sort | grep '^ *1 ' Change-Id: I59609f58ab65312012c047036ae1e0634f795779
2011-03-17 22:07:59 +01:00
static const int auto_speed_thresh[17] =
Initial WebM release
2010-05-18 17:58:33 +02:00
{
1000,
200,
150,
130,
150,
125,
120,
115,
115,
115,
115,
115,
115,
115,
115,
115,
105
};
const MB_PREDICTION_MODE vp8_mode_order[MAX_MODES] =
{
ZEROMV,
DC_PRED,
NEARESTMV,
NEARMV,
ZEROMV,
NEARESTMV,
ZEROMV,
NEARESTMV,
NEARMV,
NEARMV,
V_PRED,
H_PRED,
TM_PRED,
NEWMV,
NEWMV,
NEWMV,
SPLITMV,
SPLITMV,
SPLITMV,
B_PRED,
enabled 8x8 intra prediction modes on inter frames This commit enabled the usage of 8x8 intra prediction modes on inter frames. There are a few TODO items related to this: 1)baseline entropy need be calibrated; 2)cost of UV need to be done more properly rather than using decision only relying on Y; 3)Threshold for allowing picking 8x8 intra prediction should be lowered to lower than the B_PRED. Even with all the TODOs, tests showed consistent gain on derf set ~0.1% (PSNR:0.08% and SSIM:0.14%). It is assumed that 8x8 intra prediction will help more on large resolution clips, especially with above TODOs addressed. Change-Id: I398ada49dfc32575cfab962a569c2885111ae3ba
2011-12-01 01:25:00 +01:00
I8X8_PRED,
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
/* dual prediction modes */
ZEROMV,
NEARESTMV,
NEARMV,
ZEROMV,
NEARESTMV,
NEARMV,
ZEROMV,
NEARESTMV,
NEARMV,
NEWMV,
NEWMV,
NEWMV,
#endif /* CONFIG_DUALPRED */
Initial WebM release
2010-05-18 17:58:33 +02:00
};
const MV_REFERENCE_FRAME vp8_ref_frame_order[MAX_MODES] =
{
LAST_FRAME,
INTRA_FRAME,
LAST_FRAME,
LAST_FRAME,
GOLDEN_FRAME,
GOLDEN_FRAME,
ALTREF_FRAME,
ALTREF_FRAME,
GOLDEN_FRAME,
ALTREF_FRAME,
INTRA_FRAME,
INTRA_FRAME,
INTRA_FRAME,
LAST_FRAME,
GOLDEN_FRAME,
ALTREF_FRAME,
LAST_FRAME,
GOLDEN_FRAME,
ALTREF_FRAME,
INTRA_FRAME,
enabled 8x8 intra prediction modes on inter frames This commit enabled the usage of 8x8 intra prediction modes on inter frames. There are a few TODO items related to this: 1)baseline entropy need be calibrated; 2)cost of UV need to be done more properly rather than using decision only relying on Y; 3)Threshold for allowing picking 8x8 intra prediction should be lowered to lower than the B_PRED. Even with all the TODOs, tests showed consistent gain on derf set ~0.1% (PSNR:0.08% and SSIM:0.14%). It is assumed that 8x8 intra prediction will help more on large resolution clips, especially with above TODOs addressed. Change-Id: I398ada49dfc32575cfab962a569c2885111ae3ba
2011-12-01 01:25:00 +01:00
INTRA_FRAME,
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
/* dual prediction modes */
LAST_FRAME,
LAST_FRAME,
LAST_FRAME,
ALTREF_FRAME,
ALTREF_FRAME,
ALTREF_FRAME,
GOLDEN_FRAME,
GOLDEN_FRAME,
GOLDEN_FRAME,
LAST_FRAME,
ALTREF_FRAME,
GOLDEN_FRAME,
#endif /* CONFIG_DUALPRED */
};
#if CONFIG_DUALPRED
const MV_REFERENCE_FRAME vp8_second_ref_frame_order[MAX_MODES] =
{
0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0,
/* dual prediction modes */
GOLDEN_FRAME,
GOLDEN_FRAME,
GOLDEN_FRAME,
LAST_FRAME,
LAST_FRAME,
LAST_FRAME,
ALTREF_FRAME,
ALTREF_FRAME,
ALTREF_FRAME,
GOLDEN_FRAME,
LAST_FRAME,
ALTREF_FRAME,
Initial WebM release
2010-05-18 17:58:33 +02:00
};
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#endif /* CONFIG_DUALPRED */
Initial WebM release
2010-05-18 17:58:33 +02:00
static void fill_token_costs(
Use MAX_ENTROPY_TOKENS and ENTROPY_NODES more consistently There were many instances in the code of vp8_coef_tokens and vp8_coef_tokens-1, which was a preprocessor macro despite the naming convention. Replace these with MAX_ENTROPY_TOKENS and ENTROPY_NODES, respectively. Change-Id: I72c4f6c7634c94e1fa066cd511471e5592c748da
2011-06-28 23:03:47 +02:00
unsigned int c [BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [MAX_ENTROPY_TOKENS],
const vp8_prob p [BLOCK_TYPES] [COEF_BANDS] [PREV_COEF_CONTEXTS] [ENTROPY_NODES]
Initial WebM release
2010-05-18 17:58:33 +02:00
)
{
int i, j, k;
for (i = 0; i < BLOCK_TYPES; i++)
for (j = 0; j < COEF_BANDS; j++)
for (k = 0; k < PREV_COEF_CONTEXTS; k++)
vp8_cost_tokens((int *)(c [i][j][k]), p [i][j][k], vp8_coef_tree);
}
Fix RD bug.
2010-06-07 18:34:46 +02:00
static int rd_iifactor [ 32 ] = { 4, 4, 3, 2, 1, 0, 0, 0,
Initial WebM release
2010-05-18 17:58:33 +02:00
0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0,
};
experiment extending the quantizer range Prior to this change, VP8 min quantizer is 4, which caps the highest quality around 51DB. This experimental change extends the min quantizer to 1, removes the cap and allows the highest quality to be around ~73DB, consistent with the fdct/idct round trip error. To test this change, at configure time use options: --enable-experimental --enable-extend_qrange The following is a brief log of changes in each of the patch sets patch set 1: In this commit, the quantization/dequantization constants are kept unchanged, instead scaling factor 4 is rolled into fdct/idct. Fixed Q0 encoding tests on mobile: Before: 9560.567kbps Overall PSNR:50.255DB VPXSSIM:98.288 Now: 18035.774kbps Overall PSNR:73.022DB VPXSSIM:99.991 patch set 2: regenerated dc/ac quantizer lookup tables based on the scaling factor rolled in the fdct/idct. Also slightly extended the range towards the high quantizer end. patch set 3: slightly tweaked the quantizer tables and generated bits_per_mb table based on Paul's suggestions. patch set 4: fix a typo in idct, re-calculated tables relating active max Q to active min Q patch set 5: added rdmult lookup table based on Q patch set 6: fix rdmult scale: dct coefficient has scaled up by 4 patch set 7: make transform coefficients to be within 16bits patch set 8: normalize 2nd order quantizers patch set 9: fix mis-spellings patch set 10: change the configure script and macros to allow experimental code to be enabled at configure time with --enable-extend_qrange patch set 11: rebase for merge Change-Id: Ib50641ddd44aba2a52ed890222c309faa31cc59c
2010-12-02 00:50:14 +01:00
// 3* dc_qlookup[Q]*dc_qlookup[Q];
Further resolution of QIndex LUTS; This commit resolves further QIndex look up tables to facilitate experimentation with the quantizer range. In some cases rather than remove the look up tables completely I have created functions that are called once to populate them using a formulaic approach base on the actual quantizer. The use of these functions based on best fit of data from the original tables does affect the results on some clips but across the derf test set the effect was broadly neutral. Change-Id: I8baa61c97ce87dc09a6340d56fdeb681b9345793
2011-11-23 12:32:20 +01:00
static int rdmult_lut[QINDEX_RANGE];
adjusted sad_per_bit to correlate with quantizer Re-calibrated sad_per_bit16 and sad_per_bit4 tables to linearly correlated to quantizer values, these two variables are used in motion search for costing motion vectors. This change has an small positive effect on compression. Change-Id: Ic9b5ea6fb8d5078ef663ba4899db019cc51f4166
2010-12-24 04:59:12 +01:00
/* values are now correlated to quantizer */
Removal of Qindex LUTS. One of the problems arising when tweaking or adjusting the quantizer tables is that there are a lot of look up tables that depend on the QINDEX. Any adjustment to the link between QINDEX and real quantizer therefore tends to break aspects of for example the rate control. In this check in I have replaced several of the look up tables with functions that approximate the same results as the old Q luts but use a formulaic approach based on real Q values rather than QIndex. This should hopefully make it easier to experiment with changes to the Q tables without always having to go through and hand optimize a set of look up tables. Once things stabilize we may choose to re-instate luts for the sake of performance. Patch 2: Addressed Ronald's comments. vp8_init_me_luts() Added so luts only initialized once. Change-Id: Ic80db2212d2fd01e08e8cb5c7dca1fda1102be57
2011-11-21 16:45:10 +01:00
static int sad_per_bit16lut[QINDEX_RANGE];
static int sad_per_bit4lut[QINDEX_RANGE];
void vp8_init_me_luts()
Initial WebM release
2010-05-18 17:58:33 +02:00
{
Removal of Qindex LUTS. One of the problems arising when tweaking or adjusting the quantizer tables is that there are a lot of look up tables that depend on the QINDEX. Any adjustment to the link between QINDEX and real quantizer therefore tends to break aspects of for example the rate control. In this check in I have replaced several of the look up tables with functions that approximate the same results as the old Q luts but use a formulaic approach based on real Q values rather than QIndex. This should hopefully make it easier to experiment with changes to the Q tables without always having to go through and hand optimize a set of look up tables. Once things stabilize we may choose to re-instate luts for the sake of performance. Patch 2: Addressed Ronald's comments. vp8_init_me_luts() Added so luts only initialized once. Change-Id: Ic80db2212d2fd01e08e8cb5c7dca1fda1102be57
2011-11-21 16:45:10 +01:00
int i;
// Initialize the sad lut tables using a formulaic calculation for now
// This is to make it easier to resolve the impact of experimental changes
// to the quantizer tables.
for ( i = 0; i < QINDEX_RANGE; i++ )
{
sad_per_bit16lut[i] =
(int)((0.0418*vp8_convert_qindex_to_q(i)) + 2.4107);
sad_per_bit4lut[i] = (int)((0.063*vp8_convert_qindex_to_q(i)) + 2.742);
}
}
Initial WebM release
2010-05-18 17:58:33 +02:00
Further resolution of QIndex LUTS; This commit resolves further QIndex look up tables to facilitate experimentation with the quantizer range. In some cases rather than remove the look up tables completely I have created functions that are called once to populate them using a formulaic approach base on the actual quantizer. The use of these functions based on best fit of data from the original tables does affect the results on some clips but across the derf test set the effect was broadly neutral. Change-Id: I8baa61c97ce87dc09a6340d56fdeb681b9345793
2011-11-23 12:32:20 +01:00
int compute_rd_mult( int qindex )
{
int q;
q = vp8_dc_quant(qindex,0);
#if CONFIG_EXTEND_QRANGE
return (3 * q * q) >> 4;
#else
return (3 * q * q);
#endif
}
Initial WebM release
2010-05-18 17:58:33 +02:00
void vp8cx_initialize_me_consts(VP8_COMP *cpi, int QIndex)
{
adjust sad per bit constants While investigating the effect of DC values on SAD and SSE in motion estimation, a side finding indicates the two table of constants need be adjusted. The adjustment was done by multiplying old constants by 90% with rounding. Also absorb the 1/2 scaling constant into the two tables. Refer to change Ifa285c3e for background of the 1/2 factor. Cif set test showed a very small gain on all metric. Change-Id: I04333527a823371175dd46cb04a817e5b9a8b752
2011-05-20 08:12:40 +02:00
cpi->mb.sadperbit16 = sad_per_bit16lut[QIndex];
cpi->mb.sadperbit4 = sad_per_bit4lut[QIndex];
Initial WebM release
2010-05-18 17:58:33 +02:00
}
experiment extending the quantizer range Prior to this change, VP8 min quantizer is 4, which caps the highest quality around 51DB. This experimental change extends the min quantizer to 1, removes the cap and allows the highest quality to be around ~73DB, consistent with the fdct/idct round trip error. To test this change, at configure time use options: --enable-experimental --enable-extend_qrange The following is a brief log of changes in each of the patch sets patch set 1: In this commit, the quantization/dequantization constants are kept unchanged, instead scaling factor 4 is rolled into fdct/idct. Fixed Q0 encoding tests on mobile: Before: 9560.567kbps Overall PSNR:50.255DB VPXSSIM:98.288 Now: 18035.774kbps Overall PSNR:73.022DB VPXSSIM:99.991 patch set 2: regenerated dc/ac quantizer lookup tables based on the scaling factor rolled in the fdct/idct. Also slightly extended the range towards the high quantizer end. patch set 3: slightly tweaked the quantizer tables and generated bits_per_mb table based on Paul's suggestions. patch set 4: fix a typo in idct, re-calculated tables relating active max Q to active min Q patch set 5: added rdmult lookup table based on Q patch set 6: fix rdmult scale: dct coefficient has scaled up by 4 patch set 7: make transform coefficients to be within 16bits patch set 8: normalize 2nd order quantizers patch set 9: fix mis-spellings patch set 10: change the configure script and macros to allow experimental code to be enabled at configure time with --enable-extend_qrange patch set 11: rebase for merge Change-Id: Ib50641ddd44aba2a52ed890222c309faa31cc59c
2010-12-02 00:50:14 +01:00
void vp8_initialize_rd_consts(VP8_COMP *cpi, int QIndex)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
int q;
int i;
vp8_clear_system_state(); //__asm emms;
Further adjustment of RD behaviour with Q and Zbin. Following conversations with Tim T (Derf) I ran a large number of tests comparing the existing polynomial expression with a simpler ^2 variant. Though the polynomial was sometimes a little better at the extremes of Q it was possible to get close for most clips and even a little better on some. This code also changes the way the RD multiplier is calculated when the ZBIN is extended to use a variant of the same ^2 expression. I hope that this simpler expression will be easier to tune further as we expand our test set and consider adjustments based on content. Change-Id: I73b2564346e74d1332c33e2c1964ae093437456c
2010-06-29 13:15:54 +02:00
// Further tests required to see if optimum is different
// for key frames, golden frames and arf frames.
// if (cpi->common.refresh_golden_frame ||
// cpi->common.refresh_alt_ref_frame)
experiment extending the quantizer range Prior to this change, VP8 min quantizer is 4, which caps the highest quality around 51DB. This experimental change extends the min quantizer to 1, removes the cap and allows the highest quality to be around ~73DB, consistent with the fdct/idct round trip error. To test this change, at configure time use options: --enable-experimental --enable-extend_qrange The following is a brief log of changes in each of the patch sets patch set 1: In this commit, the quantization/dequantization constants are kept unchanged, instead scaling factor 4 is rolled into fdct/idct. Fixed Q0 encoding tests on mobile: Before: 9560.567kbps Overall PSNR:50.255DB VPXSSIM:98.288 Now: 18035.774kbps Overall PSNR:73.022DB VPXSSIM:99.991 patch set 2: regenerated dc/ac quantizer lookup tables based on the scaling factor rolled in the fdct/idct. Also slightly extended the range towards the high quantizer end. patch set 3: slightly tweaked the quantizer tables and generated bits_per_mb table based on Paul's suggestions. patch set 4: fix a typo in idct, re-calculated tables relating active max Q to active min Q patch set 5: added rdmult lookup table based on Q patch set 6: fix rdmult scale: dct coefficient has scaled up by 4 patch set 7: make transform coefficients to be within 16bits patch set 8: normalize 2nd order quantizers patch set 9: fix mis-spellings patch set 10: change the configure script and macros to allow experimental code to be enabled at configure time with --enable-extend_qrange patch set 11: rebase for merge Change-Id: Ib50641ddd44aba2a52ed890222c309faa31cc59c
2010-12-02 00:50:14 +01:00
QIndex=(QIndex<0)? 0 : ((QIndex>127)?127 : QIndex);
Further resolution of QIndex LUTS; This commit resolves further QIndex look up tables to facilitate experimentation with the quantizer range. In some cases rather than remove the look up tables completely I have created functions that are called once to populate them using a formulaic approach base on the actual quantizer. The use of these functions based on best fit of data from the original tables does affect the results on some clips but across the derf test set the effect was broadly neutral. Change-Id: I8baa61c97ce87dc09a6340d56fdeb681b9345793
2011-11-23 12:32:20 +01:00
cpi->RDMULT = compute_rd_mult(QIndex);
Initial WebM release
2010-05-18 17:58:33 +02:00
Further adjustment of RD behaviour with Q and Zbin. Following conversations with Tim T (Derf) I ran a large number of tests comparing the existing polynomial expression with a simpler ^2 variant. Though the polynomial was sometimes a little better at the extremes of Q it was possible to get close for most clips and even a little better on some. This code also changes the way the RD multiplier is calculated when the ZBIN is extended to use a variant of the same ^2 expression. I hope that this simpler expression will be easier to tune further as we expand our test set and consider adjustments based on content. Change-Id: I73b2564346e74d1332c33e2c1964ae093437456c
2010-06-29 13:15:54 +02:00
// Extend rate multiplier along side quantizer zbin increases
if (cpi->zbin_over_quant > 0)
{
double oq_factor;
// Experimental code using the same basic equation as used for Q above
// The units of cpi->zbin_over_quant are 1/128 of Q bin size
oq_factor = 1.0 + ((double)0.0015625 * cpi->zbin_over_quant);
experiment extending the quantizer range Prior to this change, VP8 min quantizer is 4, which caps the highest quality around 51DB. This experimental change extends the min quantizer to 1, removes the cap and allows the highest quality to be around ~73DB, consistent with the fdct/idct round trip error. To test this change, at configure time use options: --enable-experimental --enable-extend_qrange The following is a brief log of changes in each of the patch sets patch set 1: In this commit, the quantization/dequantization constants are kept unchanged, instead scaling factor 4 is rolled into fdct/idct. Fixed Q0 encoding tests on mobile: Before: 9560.567kbps Overall PSNR:50.255DB VPXSSIM:98.288 Now: 18035.774kbps Overall PSNR:73.022DB VPXSSIM:99.991 patch set 2: regenerated dc/ac quantizer lookup tables based on the scaling factor rolled in the fdct/idct. Also slightly extended the range towards the high quantizer end. patch set 3: slightly tweaked the quantizer tables and generated bits_per_mb table based on Paul's suggestions. patch set 4: fix a typo in idct, re-calculated tables relating active max Q to active min Q patch set 5: added rdmult lookup table based on Q patch set 6: fix rdmult scale: dct coefficient has scaled up by 4 patch set 7: make transform coefficients to be within 16bits patch set 8: normalize 2nd order quantizers patch set 9: fix mis-spellings patch set 10: change the configure script and macros to allow experimental code to be enabled at configure time with --enable-extend_qrange patch set 11: rebase for merge Change-Id: Ib50641ddd44aba2a52ed890222c309faa31cc59c
2010-12-02 00:50:14 +01:00
cpi->RDMULT = (int)((double)cpi->RDMULT * oq_factor * oq_factor);
Further adjustment of RD behaviour with Q and Zbin. Following conversations with Tim T (Derf) I ran a large number of tests comparing the existing polynomial expression with a simpler ^2 variant. Though the polynomial was sometimes a little better at the extremes of Q it was possible to get close for most clips and even a little better on some. This code also changes the way the RD multiplier is calculated when the ZBIN is extended to use a variant of the same ^2 expression. I hope that this simpler expression will be easier to tune further as we expand our test set and consider adjustments based on content. Change-Id: I73b2564346e74d1332c33e2c1964ae093437456c
2010-06-29 13:15:54 +02:00
}
Initial WebM release
2010-05-18 17:58:33 +02:00
Fix RD bug.
2010-06-07 18:34:46 +02:00
if (cpi->pass == 2 && (cpi->common.frame_type != KEY_FRAME))
Initial WebM release
2010-05-18 17:58:33 +02:00
{
cleanup: collect twopass variables This patch collects the twopass specific memebers of VP8_COMP into a dedicated struct. This is a first step towards isolating the two pass rate control and aids readability by decorating these variables with the 'twopass.' namespace. This makes it clear to the reader in what contexts the variable will be valid, and is a hint that a section of code might be a good candidate to move to firstpass.c in later refactoring. There likely will be other rate control modes that need their own specific data as well. This notation is probably overly verbose in firstpass.c, so an alternative would be to access this struct through a pointer like 'rc->' instead of 'cpi->firstpass.' in that file. Feel free to make a review comment to that effect if you prefer. Change-Id: I0ab8254647cb4b493a77c16b5d236d0d4a94ca4d
2011-05-19 23:16:39 +02:00
if (cpi->twopass.next_iiratio > 31)
Fix RD bug.
2010-06-07 18:34:46 +02:00
cpi->RDMULT += (cpi->RDMULT * rd_iifactor[31]) >> 4;
Initial WebM release
2010-05-18 17:58:33 +02:00
else
cleanup: collect twopass variables This patch collects the twopass specific memebers of VP8_COMP into a dedicated struct. This is a first step towards isolating the two pass rate control and aids readability by decorating these variables with the 'twopass.' namespace. This makes it clear to the reader in what contexts the variable will be valid, and is a hint that a section of code might be a good candidate to move to firstpass.c in later refactoring. There likely will be other rate control modes that need their own specific data as well. This notation is probably overly verbose in firstpass.c, so an alternative would be to access this struct through a pointer like 'rc->' instead of 'cpi->firstpass.' in that file. Feel free to make a review comment to that effect if you prefer. Change-Id: I0ab8254647cb4b493a77c16b5d236d0d4a94ca4d
2011-05-19 23:16:39 +02:00
cpi->RDMULT +=
(cpi->RDMULT * rd_iifactor[cpi->twopass.next_iiratio]) >> 4;
Initial WebM release
2010-05-18 17:58:33 +02:00
}
experiment extending the quantizer range Prior to this change, VP8 min quantizer is 4, which caps the highest quality around 51DB. This experimental change extends the min quantizer to 1, removes the cap and allows the highest quality to be around ~73DB, consistent with the fdct/idct round trip error. To test this change, at configure time use options: --enable-experimental --enable-extend_qrange The following is a brief log of changes in each of the patch sets patch set 1: In this commit, the quantization/dequantization constants are kept unchanged, instead scaling factor 4 is rolled into fdct/idct. Fixed Q0 encoding tests on mobile: Before: 9560.567kbps Overall PSNR:50.255DB VPXSSIM:98.288 Now: 18035.774kbps Overall PSNR:73.022DB VPXSSIM:99.991 patch set 2: regenerated dc/ac quantizer lookup tables based on the scaling factor rolled in the fdct/idct. Also slightly extended the range towards the high quantizer end. patch set 3: slightly tweaked the quantizer tables and generated bits_per_mb table based on Paul's suggestions. patch set 4: fix a typo in idct, re-calculated tables relating active max Q to active min Q patch set 5: added rdmult lookup table based on Q patch set 6: fix rdmult scale: dct coefficient has scaled up by 4 patch set 7: make transform coefficients to be within 16bits patch set 8: normalize 2nd order quantizers patch set 9: fix mis-spellings patch set 10: change the configure script and macros to allow experimental code to be enabled at configure time with --enable-extend_qrange patch set 11: rebase for merge Change-Id: Ib50641ddd44aba2a52ed890222c309faa31cc59c
2010-12-02 00:50:14 +01:00
if (cpi->RDMULT < 7)
cpi->RDMULT = 7;
Further work on extended Q range. Fixed some further QIndex related issues and replaced some tables (eg zbin and rounding) Also Added function (currently disabled by default) to populate the main AC and DC quantizer tables. Using the original AC range the resulting computed DC values give behavior broadly comparable on the DERF set. That is not to say that the equations will hold good over a more extended range. The purpose of this code is to make it easier to experiment with further alterations to the Q range and distribution of Q values plus the relative weights given to AC and DC. The function find_fp_qindex() ensures that changes to the Q tables are reflected in the value passed in to the first pass code. Slight experimental adjustment to static segment Q offset. Change-Id: I36186267d55dfc2a3d565d0cff7218ef300d1cd5
2011-12-02 15:57:21 +01:00
adjusting the calculation of errorperbit RDMULT/RDDIV defines a bit worth of distortion in term of sum squared difference. This has also been used as errorperbit in subpixel motion search, where the distortions computed as variance of the difference. The variance of differences is different from sum squared differences by amount of DC squared. Typically, for inter predicted MBs, this difference averages around 10% between the two distortion, so this patch introduces a 110% constant in deriving errorperbit from RDMULT/RDDIV. Test on CIF set shows small but positive gain on overall PSNR (.03%) and SSIM (.07%), overall impact on average PSNR is 0. Change-Id: I95425f922d037b4d96083064a10c7cdd4948ee62
2011-06-21 01:30:26 +02:00
cpi->mb.errorperbit = (cpi->RDMULT / 110);
cap the best quantizer for 2nd order DC This commit also removes artificial RDMULT cap for low quantizers. The intention is to address some abnormal behavior of mode selections at the low quantizer end, where many macroblocks were coded with SPLITMV with all partitions using same motion vector including (0,0). This change improves the compression quality substantially for high quality encodings in both PSNR and SSIM terms. Overall effect on mid/low rate range is also positive for all metrics, but smaller in magnitude. Change-Id: I864b29c4bd9ff610d2545fa94a19cc7e80c02667
2011-01-26 07:24:22 +01:00
cpi->mb.errorperbit += (cpi->mb.errorperbit==0);
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_set_speed_features(cpi);
experiment extending the quantizer range Prior to this change, VP8 min quantizer is 4, which caps the highest quality around 51DB. This experimental change extends the min quantizer to 1, removes the cap and allows the highest quality to be around ~73DB, consistent with the fdct/idct round trip error. To test this change, at configure time use options: --enable-experimental --enable-extend_qrange The following is a brief log of changes in each of the patch sets patch set 1: In this commit, the quantization/dequantization constants are kept unchanged, instead scaling factor 4 is rolled into fdct/idct. Fixed Q0 encoding tests on mobile: Before: 9560.567kbps Overall PSNR:50.255DB VPXSSIM:98.288 Now: 18035.774kbps Overall PSNR:73.022DB VPXSSIM:99.991 patch set 2: regenerated dc/ac quantizer lookup tables based on the scaling factor rolled in the fdct/idct. Also slightly extended the range towards the high quantizer end. patch set 3: slightly tweaked the quantizer tables and generated bits_per_mb table based on Paul's suggestions. patch set 4: fix a typo in idct, re-calculated tables relating active max Q to active min Q patch set 5: added rdmult lookup table based on Q patch set 6: fix rdmult scale: dct coefficient has scaled up by 4 patch set 7: make transform coefficients to be within 16bits patch set 8: normalize 2nd order quantizers patch set 9: fix mis-spellings patch set 10: change the configure script and macros to allow experimental code to be enabled at configure time with --enable-extend_qrange patch set 11: rebase for merge Change-Id: Ib50641ddd44aba2a52ed890222c309faa31cc59c
2010-12-02 00:50:14 +01:00
#if CONFIG_EXTEND_QRANGE
Further work on extended Q range. Fixed some further QIndex related issues and replaced some tables (eg zbin and rounding) Also Added function (currently disabled by default) to populate the main AC and DC quantizer tables. Using the original AC range the resulting computed DC values give behavior broadly comparable on the DERF set. That is not to say that the equations will hold good over a more extended range. The purpose of this code is to make it easier to experiment with further alterations to the Q range and distribution of Q values plus the relative weights given to AC and DC. The function find_fp_qindex() ensures that changes to the Q tables are reflected in the value passed in to the first pass code. Slight experimental adjustment to static segment Q offset. Change-Id: I36186267d55dfc2a3d565d0cff7218ef300d1cd5
2011-12-02 15:57:21 +01:00
q = (int)pow(vp8_dc_quant(QIndex,0)>>2, 1.25);
q = q << 2;
Some further QIndex issues with extended Q Resolved or factored out some further issues with Q index. Put in a 3rd order polynomial instead of less accurate power function as the best fit on gf and kf boost adjustment. Added avg_q value to use instead of ni_av_qi. Compute segment delta Q values based on avg_q. Fixed bug in adjust_maxq_qrange(). The extended range Q on the derf set, using standard data rates (which do not extend high enough to get big benefits) still show a shortfall of between 0.5 and 1% though so there would appear to be further issues that need to be tracked down. Change-Id: Icfd49b9f401906ba487ef1bef7d397048295d959
2011-12-06 15:48:52 +01:00
cpi->RDMULT = cpi->RDMULT << 4;
Further work on extended Q range. Fixed some further QIndex related issues and replaced some tables (eg zbin and rounding) Also Added function (currently disabled by default) to populate the main AC and DC quantizer tables. Using the original AC range the resulting computed DC values give behavior broadly comparable on the DERF set. That is not to say that the equations will hold good over a more extended range. The purpose of this code is to make it easier to experiment with further alterations to the Q range and distribution of Q values plus the relative weights given to AC and DC. The function find_fp_qindex() ensures that changes to the Q tables are reflected in the value passed in to the first pass code. Slight experimental adjustment to static segment Q offset. Change-Id: I36186267d55dfc2a3d565d0cff7218ef300d1cd5
2011-12-02 15:57:21 +01:00
#else
q = (int)pow(vp8_dc_quant(QIndex,0), 1.25);
experiment extending the quantizer range Prior to this change, VP8 min quantizer is 4, which caps the highest quality around 51DB. This experimental change extends the min quantizer to 1, removes the cap and allows the highest quality to be around ~73DB, consistent with the fdct/idct round trip error. To test this change, at configure time use options: --enable-experimental --enable-extend_qrange The following is a brief log of changes in each of the patch sets patch set 1: In this commit, the quantization/dequantization constants are kept unchanged, instead scaling factor 4 is rolled into fdct/idct. Fixed Q0 encoding tests on mobile: Before: 9560.567kbps Overall PSNR:50.255DB VPXSSIM:98.288 Now: 18035.774kbps Overall PSNR:73.022DB VPXSSIM:99.991 patch set 2: regenerated dc/ac quantizer lookup tables based on the scaling factor rolled in the fdct/idct. Also slightly extended the range towards the high quantizer end. patch set 3: slightly tweaked the quantizer tables and generated bits_per_mb table based on Paul's suggestions. patch set 4: fix a typo in idct, re-calculated tables relating active max Q to active min Q patch set 5: added rdmult lookup table based on Q patch set 6: fix rdmult scale: dct coefficient has scaled up by 4 patch set 7: make transform coefficients to be within 16bits patch set 8: normalize 2nd order quantizers patch set 9: fix mis-spellings patch set 10: change the configure script and macros to allow experimental code to be enabled at configure time with --enable-extend_qrange patch set 11: rebase for merge Change-Id: Ib50641ddd44aba2a52ed890222c309faa31cc59c
2010-12-02 00:50:14 +01:00
#endif
Further work on extended Q range. Fixed some further QIndex related issues and replaced some tables (eg zbin and rounding) Also Added function (currently disabled by default) to populate the main AC and DC quantizer tables. Using the original AC range the resulting computed DC values give behavior broadly comparable on the DERF set. That is not to say that the equations will hold good over a more extended range. The purpose of this code is to make it easier to experiment with further alterations to the Q range and distribution of Q values plus the relative weights given to AC and DC. The function find_fp_qindex() ensures that changes to the Q tables are reflected in the value passed in to the first pass code. Slight experimental adjustment to static segment Q offset. Change-Id: I36186267d55dfc2a3d565d0cff7218ef300d1cd5
2011-12-02 15:57:21 +01:00
if (q < 8)
q = 8;
experiment extending the quantizer range Prior to this change, VP8 min quantizer is 4, which caps the highest quality around 51DB. This experimental change extends the min quantizer to 1, removes the cap and allows the highest quality to be around ~73DB, consistent with the fdct/idct round trip error. To test this change, at configure time use options: --enable-experimental --enable-extend_qrange The following is a brief log of changes in each of the patch sets patch set 1: In this commit, the quantization/dequantization constants are kept unchanged, instead scaling factor 4 is rolled into fdct/idct. Fixed Q0 encoding tests on mobile: Before: 9560.567kbps Overall PSNR:50.255DB VPXSSIM:98.288 Now: 18035.774kbps Overall PSNR:73.022DB VPXSSIM:99.991 patch set 2: regenerated dc/ac quantizer lookup tables based on the scaling factor rolled in the fdct/idct. Also slightly extended the range towards the high quantizer end. patch set 3: slightly tweaked the quantizer tables and generated bits_per_mb table based on Paul's suggestions. patch set 4: fix a typo in idct, re-calculated tables relating active max Q to active min Q patch set 5: added rdmult lookup table based on Q patch set 6: fix rdmult scale: dct coefficient has scaled up by 4 patch set 7: make transform coefficients to be within 16bits patch set 8: normalize 2nd order quantizers patch set 9: fix mis-spellings patch set 10: change the configure script and macros to allow experimental code to be enabled at configure time with --enable-extend_qrange patch set 11: rebase for merge Change-Id: Ib50641ddd44aba2a52ed890222c309faa31cc59c
2010-12-02 00:50:14 +01:00
Initial WebM release
2010-05-18 17:58:33 +02:00
if (cpi->RDMULT > 1000)
{
cpi->RDDIV = 1;
cpi->RDMULT /= 100;
for (i = 0; i < MAX_MODES; i++)
{
if (cpi->sf.thresh_mult[i] < INT_MAX)
{
cpi->rd_threshes[i] = cpi->sf.thresh_mult[i] * q / 100;
}
else
{
cpi->rd_threshes[i] = INT_MAX;
}
cpi->rd_baseline_thresh[i] = cpi->rd_threshes[i];
}
}
else
{
cpi->RDDIV = 100;
for (i = 0; i < MAX_MODES; i++)
{
if (cpi->sf.thresh_mult[i] < (INT_MAX / q))
{
cpi->rd_threshes[i] = cpi->sf.thresh_mult[i] * q;
}
else
{
cpi->rd_threshes[i] = INT_MAX;
}
cpi->rd_baseline_thresh[i] = cpi->rd_threshes[i];
}
}
fill_token_costs(
cpi->mb.token_costs,
(const vp8_prob( *)[8][3][11]) cpi->common.fc.coef_probs
);
enable selecting&transmitting to for intra mode entropy This commit added a 3 bit index to the bitstream, the index is used to look into the intra mode coding entropy context table. The commit uses the mode stats to calculate the cost of transmitting modes using 8 possible entropy distributions, and selects the distribution that provides the lowest cost to do the actual mode coding. Initial test show this provides additional .2%~.3% gain over quantizer adaptive intra mode coding. So the adaptive intra mode coding provides a total of .5%(psnr) to .6% gain(ssim) combined for all-key-encoding To build and test, configure with --enable-experimental --enable-qimode Change-Id: I7c41cd8bfb352bc1fe7c5da1848a58faea5ed74a
2011-08-31 21:01:58 +02:00
#if CONFIG_QIMODE
//rough estimate for costing
cpi->common.kf_ymode_probs_index = cpi->common.base_qindex>>4;
#endif
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_init_mode_costs(cpi);
}
void vp8_auto_select_speed(VP8_COMP *cpi)
{
int milliseconds_for_compress = (int)(1000000 / cpi->oxcf.frame_rate);
milliseconds_for_compress = milliseconds_for_compress * (16 - cpi->oxcf.cpu_used) / 16;
#if 0
if (0)
{
FILE *f;
f = fopen("speed.stt", "a");
fprintf(f, " %8ld %10ld %10ld %10ld\n",
cpi->common.current_video_frame, cpi->Speed, milliseconds_for_compress, cpi->avg_pick_mode_time);
fclose(f);
}
#endif
/*
// this is done during parameter valid check
clean up unused variable warnings Change-Id: I9467d7a50eac32d8e8f3a2f26db818e47c93c94b
2011-05-09 17:16:31 +02:00
if( cpi->oxcf.cpu_used > 16)
cpi->oxcf.cpu_used = 16;
if( cpi->oxcf.cpu_used < -16)
cpi->oxcf.cpu_used = -16;
Initial WebM release
2010-05-18 17:58:33 +02:00
*/
if (cpi->avg_pick_mode_time < milliseconds_for_compress && (cpi->avg_encode_time - cpi->avg_pick_mode_time) < milliseconds_for_compress)
{
if (cpi->avg_pick_mode_time == 0)
{
cpi->Speed = 4;
}
else
{
if (milliseconds_for_compress * 100 < cpi->avg_encode_time * 95)
{
cpi->Speed += 2;
cpi->avg_pick_mode_time = 0;
cpi->avg_encode_time = 0;
if (cpi->Speed > 16)
{
cpi->Speed = 16;
}
}
Increase static linkage, remove unused functions A large number of functions were defined with external linkage, even though they were only used from within one file. This patch changes their linkage to static and removes the vp8_ prefix from their names, which should make it more obvious to the reader that the function is contained within the current translation unit. Functions that were not referenced were removed. These symbols were identified by: $ nm -A libvpx.a | sort -k3 | uniq -c -f2 | grep ' [A-Z] ' \ | sort | grep '^ *1 ' Change-Id: I59609f58ab65312012c047036ae1e0634f795779
2011-03-17 22:07:59 +01:00
if (milliseconds_for_compress * 100 > cpi->avg_encode_time * auto_speed_thresh[cpi->Speed])
Initial WebM release
2010-05-18 17:58:33 +02:00
{
cpi->Speed -= 1;
cpi->avg_pick_mode_time = 0;
cpi->avg_encode_time = 0;
// In real-time mode, cpi->speed is in [4, 16].
if (cpi->Speed < 4) //if ( cpi->Speed < 0 )
{
cpi->Speed = 4; //cpi->Speed = 0;
}
}
}
}
else
{
cpi->Speed += 4;
if (cpi->Speed > 16)
cpi->Speed = 16;
cpi->avg_pick_mode_time = 0;
cpi->avg_encode_time = 0;
}
}
int vp8_block_error_c(short *coeff, short *dqcoeff)
{
int i;
int error = 0;
for (i = 0; i < 16; i++)
{
int this_diff = coeff[i] - dqcoeff[i];
error += this_diff * this_diff;
}
return error;
}
int vp8_mbblock_error_c(MACROBLOCK *mb, int dc)
{
BLOCK *be;
BLOCKD *bd;
int i, j;
int berror, error = 0;
for (i = 0; i < 16; i++)
{
be = &mb->block[i];
bd = &mb->e_mbd.block[i];
berror = 0;
for (j = dc; j < 16; j++)
{
int this_diff = be->coeff[j] - bd->dqcoeff[j];
berror += this_diff * this_diff;
}
error += berror;
}
return error;
}
int vp8_mbuverror_c(MACROBLOCK *mb)
{
BLOCK *be;
BLOCKD *bd;
int i;
int error = 0;
for (i = 16; i < 24; i++)
{
be = &mb->block[i];
bd = &mb->e_mbd.block[i];
error += vp8_block_error_c(be->coeff, bd->dqcoeff);
}
return error;
}
int VP8_UVSSE(MACROBLOCK *x, const vp8_variance_rtcd_vtable_t *rtcd)
{
unsigned char *uptr, *vptr;
unsigned char *upred_ptr = (*(x->block[16].base_src) + x->block[16].src);
unsigned char *vpred_ptr = (*(x->block[20].base_src) + x->block[20].src);
int uv_stride = x->block[16].src_stride;
unsigned int sse1 = 0;
unsigned int sse2 = 0;
Removed bmi copy to/from BLOCKD for SPLITMV and B_PRED modes. Modified code to use the bmi found in mode_info_context instead of BLOCKD. On the decode side, the uvmvs are calculated only when required, instead of every macroblock. This is WIP. (bmi should eventually be removed from BLOCKD) Small performance gains noticed for RT encodes and decodes.(VGA) Change-Id: I2ed7f0fd5ca733655df684aa82da575c77a973e7
2011-08-24 20:42:26 +02:00
int mv_row = x->e_mbd.mode_info_context->mbmi.mv.as_mv.row;
int mv_col = x->e_mbd.mode_info_context->mbmi.mv.as_mv.col;
Initial WebM release
2010-05-18 17:58:33 +02:00
int offset;
int pre_stride = x->e_mbd.block[16].pre_stride;
Removed bmi copy to/from BLOCKD for SPLITMV and B_PRED modes. Modified code to use the bmi found in mode_info_context instead of BLOCKD. On the decode side, the uvmvs are calculated only when required, instead of every macroblock. This is WIP. (bmi should eventually be removed from BLOCKD) Small performance gains noticed for RT encodes and decodes.(VGA) Change-Id: I2ed7f0fd5ca733655df684aa82da575c77a973e7
2011-08-24 20:42:26 +02:00
if (mv_row < 0)
mv_row -= 1;
else
mv_row += 1;
if (mv_col < 0)
mv_col -= 1;
else
mv_col += 1;
mv_row /= 2;
mv_col /= 2;
Initial WebM release
2010-05-18 17:58:33 +02:00
offset = (mv_row >> 3) * pre_stride + (mv_col >> 3);
uptr = x->e_mbd.pre.u_buffer + offset;
vptr = x->e_mbd.pre.v_buffer + offset;
if ((mv_row | mv_col) & 7)
{
use get8x8var directly for non-subpixel motion case in VP8_UVSSE VP8_UVSSE mistakenly used subpixvar8x8 to calculate SSE for non-subpixl motion cases. Change-Id: I4a5398bb9ef39c211039f6af4540546d4972e6a9
2011-05-22 06:51:21 +02:00
VARIANCE_INVOKE(rtcd, subpixvar8x8)(uptr, pre_stride,
mv_col & 7, mv_row & 7, upred_ptr, uv_stride, &sse2);
VARIANCE_INVOKE(rtcd, subpixvar8x8)(vptr, pre_stride,
mv_col & 7, mv_row & 7, vpred_ptr, uv_stride, &sse1);
Initial WebM release
2010-05-18 17:58:33 +02:00
sse2 += sse1;
}
else
{
Use var8x8 instead of get8x8var in VP8_UVSSE 'sum' returned by get8x8var is not used and var8x8 has optimizations for more platforms. Change-Id: I4a907fb1a05f285669fb0b95dc71d42182c980f6
2011-05-25 11:54:34 +02:00
VARIANCE_INVOKE(rtcd, var8x8)(uptr, pre_stride,
upred_ptr, uv_stride, &sse2);
VARIANCE_INVOKE(rtcd, var8x8)(vptr, pre_stride,
vpred_ptr, uv_stride, &sse1);
Initial WebM release
2010-05-18 17:58:33 +02:00
sse2 += sse1;
}
return sse2;
}
static int cost_coeffs(MACROBLOCK *mb, BLOCKD *b, int type, ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l)
{
int c = !type; /* start at coef 0, unless Y with Y2 */
int eob = b->eob;
int pt ; /* surrounding block/prev coef predictor */
int cost = 0;
short *qcoeff_ptr = b->qcoeff;
VP8_COMBINEENTROPYCONTEXTS(pt, *a, *l);
# define QC( I) ( qcoeff_ptr [vp8_default_zig_zag1d[I]] )
for (; c < eob; c++)
{
int v = QC(c);
int t = vp8_dct_value_tokens_ptr[v].Token;
cost += mb->token_costs [type] [vp8_coef_bands[c]] [pt] [t];
cost += vp8_dct_value_cost_ptr[v];
pt = vp8_prev_token_class[t];
}
# undef QC
if (c < 16)
cost += mb->token_costs [type] [vp8_coef_bands[c]] [pt] [DCT_EOB_TOKEN];
pt = (c != !type); // is eob first coefficient;
*a = *l = pt;
return cost;
}
Removed unnecessary checks. macro_block_yrd and vp8_rdcost_mby are not called for SPLITMV. Change-Id: I2224d3c8725df526d48426447482768d543752f1
2010-11-17 19:15:13 +01:00
static int vp8_rdcost_mby(MACROBLOCK *mb)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
int cost = 0;
int b;
MACROBLOCKD *x = &mb->e_mbd;
Changed above and left context data layout The main reason for the change was to reduce cycles in the token decoder. (~1.5% gain for 32 bit) This layout should be more cache friendly. As a result of this change, the encoder had to be updated. Change-Id: Id5e804169d8889da0378b3a519ac04dabd28c837 Note: dixie uses a similar layout
2010-08-31 16:49:57 +02:00
ENTROPY_CONTEXT_PLANES t_above, t_left;
ENTROPY_CONTEXT *ta;
ENTROPY_CONTEXT *tl;
vpx_memcpy(&t_above, mb->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES));
vpx_memcpy(&t_left, mb->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES));
Initial WebM release
2010-05-18 17:58:33 +02:00
Changed above and left context data layout The main reason for the change was to reduce cycles in the token decoder. (~1.5% gain for 32 bit) This layout should be more cache friendly. As a result of this change, the encoder had to be updated. Change-Id: Id5e804169d8889da0378b3a519ac04dabd28c837 Note: dixie uses a similar layout
2010-08-31 16:49:57 +02:00
ta = (ENTROPY_CONTEXT *)&t_above;
tl = (ENTROPY_CONTEXT *)&t_left;
Initial WebM release
2010-05-18 17:58:33 +02:00
for (b = 0; b < 16; b++)
Removed vp8_block2type and used defines instead. Change-Id: Idb56e0295d004793f406dfd2d8d8c546aad62e03
2011-02-23 22:37:08 +01:00
cost += cost_coeffs(mb, x->block + b, PLANE_TYPE_Y_NO_DC,
Changed above and left context data layout The main reason for the change was to reduce cycles in the token decoder. (~1.5% gain for 32 bit) This layout should be more cache friendly. As a result of this change, the encoder had to be updated. Change-Id: Id5e804169d8889da0378b3a519ac04dabd28c837 Note: dixie uses a similar layout
2010-08-31 16:49:57 +02:00
ta + vp8_block2above[b], tl + vp8_block2left[b]);
Initial WebM release
2010-05-18 17:58:33 +02:00
Removed vp8_block2type and used defines instead. Change-Id: Idb56e0295d004793f406dfd2d8d8c546aad62e03
2011-02-23 22:37:08 +01:00
cost += cost_coeffs(mb, x->block + 24, PLANE_TYPE_Y2,
Removed unnecessary checks. macro_block_yrd and vp8_rdcost_mby are not called for SPLITMV. Change-Id: I2224d3c8725df526d48426447482768d543752f1
2010-11-17 19:15:13 +01:00
ta + vp8_block2above[24], tl + vp8_block2left[24]);
Initial WebM release
2010-05-18 17:58:33 +02:00
return cost;
}
Rationalize vp8_rd_pick_intra16x16mby_mode() Use the function macro_block_yrd() to calculate error and distortion in keeping with what is done for inter frames. The old code was using a variance metric for once case and an SSE function for measuring distortion in the other case. The function vp8_encode_intra16x16mbyrd() is no longer used. Change-Id: Ic228cb00a78ff637f4365b43f58fbe5a9273d36f
2011-01-26 19:46:34 +01:00
static void macro_block_yrd( MACROBLOCK *mb,
int *Rate,
int *Distortion,
const vp8_encodemb_rtcd_vtable_t *rtcd)
{
int b;
MACROBLOCKD *const x = &mb->e_mbd;
BLOCK *const mb_y2 = mb->block + 24;
BLOCKD *const x_y2 = x->block + 24;
short *Y2DCPtr = mb_y2->src_diff;
BLOCK *beptr;
int d;
Copy macroblock data to a buffer before encoding it I got this idea from Pascal (Thanks). Before encoding a macroblock, copy it to a 16x16 buffer, and then read source data from there instead. This will help keep the source data in cache, and help with the performance. Change-Id: Id05f4cb601299150511d59dcba0ae62c49b5b757
2011-06-23 19:54:02 +02:00
ENCODEMB_INVOKE(rtcd, submby)( mb->src_diff, *(mb->block[0].base_src),
mb->e_mbd.predictor, mb->block[0].src_stride );
Rationalize vp8_rd_pick_intra16x16mby_mode() Use the function macro_block_yrd() to calculate error and distortion in keeping with what is done for inter frames. The old code was using a variance metric for once case and an SSE function for measuring distortion in the other case. The function vp8_encode_intra16x16mbyrd() is no longer used. Change-Id: Ic228cb00a78ff637f4365b43f58fbe5a9273d36f
2011-01-26 19:46:34 +01:00
// Fdct and building the 2nd order block
for (beptr = mb->block; beptr < mb->block + 16; beptr += 2)
{
mb->vp8_short_fdct8x4(beptr->src_diff, beptr->coeff, 32);
*Y2DCPtr++ = beptr->coeff[0];
*Y2DCPtr++ = beptr->coeff[16];
}
// 2nd order fdct
mb->short_walsh4x4(mb_y2->src_diff, mb_y2->coeff, 8);
// Quantization
for (b = 0; b < 16; b++)
{
mb->quantize_b(&mb->block[b], &mb->e_mbd.block[b]);
}
// DC predication and Quantization of 2nd Order block
mb->quantize_b(mb_y2, x_y2);
// Distortion
d = ENCODEMB_INVOKE(rtcd, mberr)(mb, 1) << 2;
fix build issues for experimental branch experimental branch build was broken from some merge artifacts, this commit fixes those issues to enable the experimental branch to build. Change-Id: Ic52b2d2f1d1b80abb7ecaa4c0927bcf887ac0c2a
2011-06-23 00:07:04 +02:00
#if CONFIG_EXTEND_QRANGE
d += ENCODEMB_INVOKE(rtcd, berr)(mb_y2->coeff, x_y2->dqcoeff)<<2;
#else
Rationalize vp8_rd_pick_intra16x16mby_mode() Use the function macro_block_yrd() to calculate error and distortion in keeping with what is done for inter frames. The old code was using a variance metric for once case and an SSE function for measuring distortion in the other case. The function vp8_encode_intra16x16mbyrd() is no longer used. Change-Id: Ic228cb00a78ff637f4365b43f58fbe5a9273d36f
2011-01-26 19:46:34 +01:00
d += ENCODEMB_INVOKE(rtcd, berr)(mb_y2->coeff, x_y2->dqcoeff);
fix build issues for experimental branch experimental branch build was broken from some merge artifacts, this commit fixes those issues to enable the experimental branch to build. Change-Id: Ic52b2d2f1d1b80abb7ecaa4c0927bcf887ac0c2a
2011-06-23 00:07:04 +02:00
#endif
Rationalize vp8_rd_pick_intra16x16mby_mode() Use the function macro_block_yrd() to calculate error and distortion in keeping with what is done for inter frames. The old code was using a variance metric for once case and an SSE function for measuring distortion in the other case. The function vp8_encode_intra16x16mbyrd() is no longer used. Change-Id: Ic228cb00a78ff637f4365b43f58fbe5a9273d36f
2011-01-26 19:46:34 +01:00
*Distortion = (d >> 4);
// rate
*Rate = vp8_rdcost_mby(mb);
}
Initial WebM release
2010-05-18 17:58:33 +02:00
Revisited rd_pick_intra4x4block Removed unnecessary copies. No noticeable speed gains. Change-Id: I996c50c23fedd06d54ee7a3e762cbf559cc4a9d1
2011-02-24 19:31:47 +01:00
static void copy_predictor(unsigned char *dst, const unsigned char *predictor)
Improved rd_pick_intra4x4block Eliminated unnecessary calculations. Improved performance by 10% on keyframes and 1.6% overall for the test clip used. Change-Id: I87671b26af5e2cc439e81d0fee3b15c7cd2a3309
2011-02-14 19:32:58 +01:00
{
Revisited rd_pick_intra4x4block Removed unnecessary copies. No noticeable speed gains. Change-Id: I996c50c23fedd06d54ee7a3e762cbf559cc4a9d1
2011-02-24 19:31:47 +01:00
const unsigned int *p = (const unsigned int *)predictor;
unsigned int *d = (unsigned int *)dst;
d[0] = p[0];
d[4] = p[4];
d[8] = p[8];
d[12] = p[12];
Improved rd_pick_intra4x4block Eliminated unnecessary calculations. Improved performance by 10% on keyframes and 1.6% overall for the test clip used. Change-Id: I87671b26af5e2cc439e81d0fee3b15c7cd2a3309
2011-02-14 19:32:58 +01:00
}
add 8x8 intra prediction modes Patch 1 to Patch 3 is an initial implementation of 8x8 intra prediction modes, here are with the following assumptions: a. 8x8 has 4 prediction modes DC, H, V and TM b. UV 4x4 block use the same mode as corresponding 8x8 area c. i8x8 modes are enabled for key frame only for now Patch 4: d. removed debug code from previous patches Patch 5: e. added stats code to collect entropy stats and further cleaned up Patch 6: f. changed mode stats code to collect finer stats of modes Patch 7: g. normalized i8x8 modes distribution to total at 256 (8bits). Patch 8: h. fixed a bug in decoder and removed debug printf output. Patch 9: i. more cleanups to address paul's comment Patch 10: j. messy rebase/merges to bring the commit up to date. Tests on HD clips encoded with all key frame showing consistent gain on all clips and all metrics:~0.5%(psnr) and 0.6%(ssim): http://www.corp.google.com/~yaowu/no_crawl/i8x8hd_allkey_fixedq.html To build and test, configure with: --enable-experimental --enable-i8x8 Change-Id: I9813fe07ae48cab5fdb5d904bca022514ad01e7f
2011-08-05 01:30:27 +02:00
static void copy_predictor_8x8(unsigned char *dst, const unsigned char *predictor)
{
const unsigned int *p = (const unsigned int *)predictor;
unsigned int *d = (unsigned int *)dst;
d[0] = p[0];
d[1] = p[1];
d[4] = p[4];
d[5] = p[5];
d[8] = p[8];
d[9] = p[9];
d[12] = p[12];
d[13] = p[13];
d[16] = p[16];
d[17] = p[17];
d[20] = p[20];
d[21] = p[21];
d[24] = p[24];
d[25] = p[25];
d[28] = p[28];
d[29] = p[29];
}
Added early breakout for vp8_rd_pick_intra4x4mby_modes Improved performance of good quality, speed 0 (3% average) with no average quality loss. Change-Id: Ica34473f99bd74260eaebde6b132185e09e3c09d
2011-02-08 22:50:43 +01:00
static int rd_pick_intra4x4block(
Initial WebM release
2010-05-18 17:58:33 +02:00
VP8_COMP *cpi,
MACROBLOCK *x,
BLOCK *be,
BLOCKD *b,
B_PREDICTION_MODE *best_mode,
Improved rd_pick_intra4x4block Eliminated unnecessary calculations. Improved performance by 10% on keyframes and 1.6% overall for the test clip used. Change-Id: I87671b26af5e2cc439e81d0fee3b15c7cd2a3309
2011-02-14 19:32:58 +01:00
unsigned int *bmode_costs,
Initial WebM release
2010-05-18 17:58:33 +02:00
ENTROPY_CONTEXT *a,
ENTROPY_CONTEXT *l,
int *bestrate,
int *bestratey,
int *bestdistortion)
{
B_PREDICTION_MODE mode;
Improved rd_pick_intra4x4block Eliminated unnecessary calculations. Improved performance by 10% on keyframes and 1.6% overall for the test clip used. Change-Id: I87671b26af5e2cc439e81d0fee3b15c7cd2a3309
2011-02-14 19:32:58 +01:00
int best_rd = INT_MAX;
Initial WebM release
2010-05-18 17:58:33 +02:00
int rate = 0;
int distortion;
ENTROPY_CONTEXT ta = *a, tempa = *a;
ENTROPY_CONTEXT tl = *l, templ = *l;
Revisited rd_pick_intra4x4block Removed unnecessary copies. No noticeable speed gains. Change-Id: I996c50c23fedd06d54ee7a3e762cbf559cc4a9d1
2011-02-24 19:31:47 +01:00
/*
* The predictor buffer is a 2d buffer with a stride of 16. Create
* a temp buffer that meets the stride requirements, but we are only
* interested in the left 4x4 block
* */
DECLARE_ALIGNED_ARRAY(16, unsigned char, best_predictor, 16*4);
DECLARE_ALIGNED_ARRAY(16, short, best_dqcoeff, 16);
Initial WebM release
2010-05-18 17:58:33 +02:00
for (mode = B_DC_PRED; mode <= B_HU_PRED; mode++)
{
int this_rd;
int ratey;
Improved rd_pick_intra4x4block Eliminated unnecessary calculations. Improved performance by 10% on keyframes and 1.6% overall for the test clip used. Change-Id: I87671b26af5e2cc439e81d0fee3b15c7cd2a3309
2011-02-14 19:32:58 +01:00
rate = bmode_costs[mode];
SSE2/SSSE3 optimizations for build_predictors_mbuv{,_s}(). decoding before 10.425 10.432 10.423 =10.426 after: 10.405 10.416 10.398 =10.406, 0.2% faster encoding before 14.252 14.331 14.250 14.223 14.241 14.220 14.221 =14.248 after 14.095 14.090 14.085 14.095 14.064 14.081 14.089 =14.086, 1.1% faster Change-Id: I483d3d8f0deda8ad434cea76e16028380722aee2
2011-04-27 19:05:10 +02:00
RECON_INVOKE(&cpi->rtcd.common->recon, intra4x4_predict)
(b, mode, b->predictor);
Improved rd_pick_intra4x4block Eliminated unnecessary calculations. Improved performance by 10% on keyframes and 1.6% overall for the test clip used. Change-Id: I87671b26af5e2cc439e81d0fee3b15c7cd2a3309
2011-02-14 19:32:58 +01:00
ENCODEMB_INVOKE(IF_RTCD(&cpi->rtcd.encodemb), subb)(be, b, 16);
x->vp8_short_fdct4x4(be->src_diff, be->coeff, 32);
x->quantize_b(be, b);
Initial WebM release
2010-05-18 17:58:33 +02:00
tempa = ta;
templ = tl;
Removed vp8_block2type and used defines instead. Change-Id: Idb56e0295d004793f406dfd2d8d8c546aad62e03
2011-02-23 22:37:08 +01:00
ratey = cost_coeffs(x, b, PLANE_TYPE_Y_WITH_DC, &tempa, &templ);
Initial WebM release
2010-05-18 17:58:33 +02:00
rate += ratey;
distortion = ENCODEMB_INVOKE(IF_RTCD(&cpi->rtcd.encodemb), berr)(be->coeff, b->dqcoeff) >> 2;
this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
if (this_rd < best_rd)
{
*bestrate = rate;
*bestratey = ratey;
*bestdistortion = distortion;
best_rd = this_rd;
*best_mode = mode;
*a = tempa;
*l = templ;
Revisited rd_pick_intra4x4block Removed unnecessary copies. No noticeable speed gains. Change-Id: I996c50c23fedd06d54ee7a3e762cbf559cc4a9d1
2011-02-24 19:31:47 +01:00
copy_predictor(best_predictor, b->predictor);
vpx_memcpy(best_dqcoeff, b->dqcoeff, 32);
Initial WebM release
2010-05-18 17:58:33 +02:00
}
}
Removed B_MODE_INFO Declared the bmi in BLOCKD as a union instead of B_MODE_INFO. Then removed B_MODE_INFO completely. Change-Id: Ieb7469899e265892c66f7aeac87b7f2bf38e7a67
2011-06-02 19:46:41 +02:00
b->bmi.as_mode = (B_PREDICTION_MODE)(*best_mode);
Improved rd_pick_intra4x4block Eliminated unnecessary calculations. Improved performance by 10% on keyframes and 1.6% overall for the test clip used. Change-Id: I87671b26af5e2cc439e81d0fee3b15c7cd2a3309
2011-02-14 19:32:58 +01:00
Revisited rd_pick_intra4x4block Removed unnecessary copies. No noticeable speed gains. Change-Id: I996c50c23fedd06d54ee7a3e762cbf559cc4a9d1
2011-02-24 19:31:47 +01:00
IDCT_INVOKE(IF_RTCD(&cpi->rtcd.common->idct), idct16)(best_dqcoeff, b->diff, 32);
RECON_INVOKE(IF_RTCD(&cpi->rtcd.common->recon), recon)(best_predictor, b->diff, *(b->base_dst) + b->dst, b->dst_stride);
Initial WebM release
2010-05-18 17:58:33 +02:00
Added early breakout for vp8_rd_pick_intra4x4mby_modes Improved performance of good quality, speed 0 (3% average) with no average quality loss. Change-Id: Ica34473f99bd74260eaebde6b132185e09e3c09d
2011-02-08 22:50:43 +01:00
return best_rd;
Initial WebM release
2010-05-18 17:58:33 +02:00
}
Move RD intra block mode selection to rdopt.c This change is analogous to I0b67dae1f8a74902378da7bdf565e39ab832dda7, which made the move for the non-RD path. Change-Id: If63fc1b0cd1eb7f932e710f83ff24d91454f8ed1
2011-06-08 18:05:05 +02:00
static int rd_pick_intra4x4mby_modes(VP8_COMP *cpi, MACROBLOCK *mb, int *Rate,
int *rate_y, int *Distortion, int best_rd)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
MACROBLOCKD *const xd = &mb->e_mbd;
int i;
int cost = mb->mbmode_cost [xd->frame_type] [B_PRED];
int distortion = 0;
int tot_rate_y = 0;
cosmetics: consistently use [u]int64_t Removes mixed usage of (unsigned) long long and INT64. Fixes Issue #208. Change-Id: I220d3ed5ce4bb1280cd38bb3715f208ce23cf83a
2011-07-26 03:44:59 +02:00
int64_t total_rd = 0;
Changed above and left context data layout The main reason for the change was to reduce cycles in the token decoder. (~1.5% gain for 32 bit) This layout should be more cache friendly. As a result of this change, the encoder had to be updated. Change-Id: Id5e804169d8889da0378b3a519ac04dabd28c837 Note: dixie uses a similar layout
2010-08-31 16:49:57 +02:00
ENTROPY_CONTEXT_PLANES t_above, t_left;
ENTROPY_CONTEXT *ta;
ENTROPY_CONTEXT *tl;
Improved rd_pick_intra4x4block Eliminated unnecessary calculations. Improved performance by 10% on keyframes and 1.6% overall for the test clip used. Change-Id: I87671b26af5e2cc439e81d0fee3b15c7cd2a3309
2011-02-14 19:32:58 +01:00
unsigned int *bmode_costs;
Changed above and left context data layout The main reason for the change was to reduce cycles in the token decoder. (~1.5% gain for 32 bit) This layout should be more cache friendly. As a result of this change, the encoder had to be updated. Change-Id: Id5e804169d8889da0378b3a519ac04dabd28c837 Note: dixie uses a similar layout
2010-08-31 16:49:57 +02:00
vpx_memcpy(&t_above, mb->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES));
vpx_memcpy(&t_left, mb->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES));
ta = (ENTROPY_CONTEXT *)&t_above;
tl = (ENTROPY_CONTEXT *)&t_left;
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_intra_prediction_down_copy(xd);
Improved rd_pick_intra4x4block Eliminated unnecessary calculations. Improved performance by 10% on keyframes and 1.6% overall for the test clip used. Change-Id: I87671b26af5e2cc439e81d0fee3b15c7cd2a3309
2011-02-14 19:32:58 +01:00
bmode_costs = mb->inter_bmode_costs;
Initial WebM release
2010-05-18 17:58:33 +02:00
for (i = 0; i < 16; i++)
{
MODE_INFO *const mic = xd->mode_info_context;
const int mis = xd->mode_info_stride;
B_PREDICTION_MODE UNINITIALIZED_IS_SAFE(best_mode);
int UNINITIALIZED_IS_SAFE(r), UNINITIALIZED_IS_SAFE(ry), UNINITIALIZED_IS_SAFE(d);
Improved rd_pick_intra4x4block Eliminated unnecessary calculations. Improved performance by 10% on keyframes and 1.6% overall for the test clip used. Change-Id: I87671b26af5e2cc439e81d0fee3b15c7cd2a3309
2011-02-14 19:32:58 +01:00
if (mb->e_mbd.frame_type == KEY_FRAME)
{
MODE_INFO size reduction Declared the bmi in MODE_INFO as a union instead of B_MODE_INFO. This reduced the memory footprint by 518,400 bytes for 1080 resolutions. The decoder performance improved by ~4% for the clip used and the encoder showed very small improvements. (0.5%) This reduction was first mentioned to me by John K. and in a later discussion by Yaowu. This is WIP. Change-Id: I8e175fdbc46d28c35277302a04bee4540efc8d29
2011-05-24 19:24:52 +02:00
const B_PREDICTION_MODE A = above_block_mode(mic, i, mis);
const B_PREDICTION_MODE L = left_block_mode(mic, i);
Improved rd_pick_intra4x4block Eliminated unnecessary calculations. Improved performance by 10% on keyframes and 1.6% overall for the test clip used. Change-Id: I87671b26af5e2cc439e81d0fee3b15c7cd2a3309
2011-02-14 19:32:58 +01:00
bmode_costs = mb->bmode_costs[A][L];
}
Added early breakout for vp8_rd_pick_intra4x4mby_modes Improved performance of good quality, speed 0 (3% average) with no average quality loss. Change-Id: Ica34473f99bd74260eaebde6b132185e09e3c09d
2011-02-08 22:50:43 +01:00
total_rd += rd_pick_intra4x4block(
Improved rd_pick_intra4x4block Eliminated unnecessary calculations. Improved performance by 10% on keyframes and 1.6% overall for the test clip used. Change-Id: I87671b26af5e2cc439e81d0fee3b15c7cd2a3309
2011-02-14 19:32:58 +01:00
cpi, mb, mb->block + i, xd->block + i, &best_mode, bmode_costs,
Changed above and left context data layout The main reason for the change was to reduce cycles in the token decoder. (~1.5% gain for 32 bit) This layout should be more cache friendly. As a result of this change, the encoder had to be updated. Change-Id: Id5e804169d8889da0378b3a519ac04dabd28c837 Note: dixie uses a similar layout
2010-08-31 16:49:57 +02:00
ta + vp8_block2above[i],
tl + vp8_block2left[i], &r, &ry, &d);
Initial WebM release
2010-05-18 17:58:33 +02:00
cost += r;
distortion += d;
tot_rate_y += ry;
MODE_INFO size reduction Declared the bmi in MODE_INFO as a union instead of B_MODE_INFO. This reduced the memory footprint by 518,400 bytes for 1080 resolutions. The decoder performance improved by ~4% for the clip used and the encoder showed very small improvements. (0.5%) This reduction was first mentioned to me by John K. and in a later discussion by Yaowu. This is WIP. Change-Id: I8e175fdbc46d28c35277302a04bee4540efc8d29
2011-05-24 19:24:52 +02:00
mic->bmi[i].as_mode = best_mode;
Added early breakout for vp8_rd_pick_intra4x4mby_modes Improved performance of good quality, speed 0 (3% average) with no average quality loss. Change-Id: Ica34473f99bd74260eaebde6b132185e09e3c09d
2011-02-08 22:50:43 +01:00
cosmetics: consistently use [u]int64_t Removes mixed usage of (unsigned) long long and INT64. Fixes Issue #208. Change-Id: I220d3ed5ce4bb1280cd38bb3715f208ce23cf83a
2011-07-26 03:44:59 +02:00
if(total_rd >= (int64_t)best_rd)
Revisited rd_pick_intra4x4block Removed unnecessary copies. No noticeable speed gains. Change-Id: I996c50c23fedd06d54ee7a3e762cbf559cc4a9d1
2011-02-24 19:31:47 +01:00
break;
Initial WebM release
2010-05-18 17:58:33 +02:00
}
cosmetics: consistently use [u]int64_t Removes mixed usage of (unsigned) long long and INT64. Fixes Issue #208. Change-Id: I220d3ed5ce4bb1280cd38bb3715f208ce23cf83a
2011-07-26 03:44:59 +02:00
if(total_rd >= (int64_t)best_rd)
Revisited rd_pick_intra4x4block Removed unnecessary copies. No noticeable speed gains. Change-Id: I996c50c23fedd06d54ee7a3e762cbf559cc4a9d1
2011-02-24 19:31:47 +01:00
return INT_MAX;
Added early breakout for vp8_rd_pick_intra4x4mby_modes Improved performance of good quality, speed 0 (3% average) with no average quality loss. Change-Id: Ica34473f99bd74260eaebde6b132185e09e3c09d
2011-02-08 22:50:43 +01:00
Initial WebM release
2010-05-18 17:58:33 +02:00
*Rate = cost;
*rate_y += tot_rate_y;
*Distortion = distortion;
return RDCOST(mb->rdmult, mb->rddiv, cost, distortion);
}
Move RD intra block mode selection to rdopt.c This change is analogous to I0b67dae1f8a74902378da7bdf565e39ab832dda7, which made the move for the non-RD path. Change-Id: If63fc1b0cd1eb7f932e710f83ff24d91454f8ed1
2011-06-08 18:05:05 +02:00
static int rd_pick_intra16x16mby_mode(VP8_COMP *cpi,
MACROBLOCK *x,
int *Rate,
int *rate_y,
int *Distortion)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
MB_PREDICTION_MODE mode;
MB_PREDICTION_MODE UNINITIALIZED_IS_SAFE(mode_selected);
int rate, ratey;
warning: pointer targets differ in signedness vp8/encoder/rdopt.c:728: warning: pointer targets in passing argument 3 of 'macro_block_yrd' differ in signedness vp8/encoder/rdopt.c:541: note: expected 'int *' but argument is of type 'unsigned int *' distortion is signed when calling macro_block_yrd is both other cases, as well as for RDCOST Change-Id: I5e22358b7da76a116f498793253aac8099cb3461
2011-01-27 17:50:29 +01:00
int distortion;
Initial WebM release
2010-05-18 17:58:33 +02:00
int best_rd = INT_MAX;
Rationalize vp8_rd_pick_intra16x16mby_mode() Use the function macro_block_yrd() to calculate error and distortion in keeping with what is done for inter frames. The old code was using a variance metric for once case and an SSE function for measuring distortion in the other case. The function vp8_encode_intra16x16mbyrd() is no longer used. Change-Id: Ic228cb00a78ff637f4365b43f58fbe5a9273d36f
2011-01-26 19:46:34 +01:00
int this_rd;
Initial WebM release
2010-05-18 17:58:33 +02:00
//Y Search for 16x16 intra prediction mode
for (mode = DC_PRED; mode <= TM_PRED; mode++)
{
Removed unnecessary MB_MODE_INFO copies These copies occurred for each macroblock in the encoder and decoder. Thetemp MB_MODE_INFO mbmi was removed from MACROBLOCKD. As a result, a large number compile errors had to be fixed. Change-Id: I4cf0ffae3ce244f6db04a4c217d52dd256382cf3
2010-08-12 22:25:43 +02:00
x->e_mbd.mode_info_context->mbmi.mode = mode;
Initial WebM release
2010-05-18 17:58:33 +02:00
Move build_intra_predictors_mby to RTCD framework The vp8_build_intra_predictors_mby and vp8_build_intra_predictors_mby_s functions had global function pointers rather than using the RTCD framework. This can show up as a potential data race with tools such as helgrind. See https://bugzilla.mozilla.org/show_bug.cgi?id=640935 for an example. Change-Id: I29c407f828ac2bddfc039f852f138de5de888534
2011-03-11 17:35:38 +01:00
RECON_INVOKE(&cpi->common.rtcd.recon, build_intra_predictors_mby)
(&x->e_mbd);
Initial WebM release
2010-05-18 17:58:33 +02:00
Rationalize vp8_rd_pick_intra16x16mby_mode() Use the function macro_block_yrd() to calculate error and distortion in keeping with what is done for inter frames. The old code was using a variance metric for once case and an SSE function for measuring distortion in the other case. The function vp8_encode_intra16x16mbyrd() is no longer used. Change-Id: Ic228cb00a78ff637f4365b43f58fbe5a9273d36f
2011-01-26 19:46:34 +01:00
macro_block_yrd(x, &ratey, &distortion, IF_RTCD(&cpi->rtcd.encodemb));
rate = ratey + x->mbmode_cost[x->e_mbd.frame_type]
[x->e_mbd.mode_info_context->mbmi.mode];
Initial WebM release
2010-05-18 17:58:33 +02:00
this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
if (this_rd < best_rd)
{
mode_selected = mode;
best_rd = this_rd;
*Rate = rate;
*rate_y = ratey;
warning: pointer targets differ in signedness vp8/encoder/rdopt.c:728: warning: pointer targets in passing argument 3 of 'macro_block_yrd' differ in signedness vp8/encoder/rdopt.c:541: note: expected 'int *' but argument is of type 'unsigned int *' distortion is signed when calling macro_block_yrd is both other cases, as well as for RDCOST Change-Id: I5e22358b7da76a116f498793253aac8099cb3461
2011-01-27 17:50:29 +01:00
*Distortion = distortion;
Initial WebM release
2010-05-18 17:58:33 +02:00
}
}
Removed unnecessary MB_MODE_INFO copies These copies occurred for each macroblock in the encoder and decoder. Thetemp MB_MODE_INFO mbmi was removed from MACROBLOCKD. As a result, a large number compile errors had to be fixed. Change-Id: I4cf0ffae3ce244f6db04a4c217d52dd256382cf3
2010-08-12 22:25:43 +02:00
x->e_mbd.mode_info_context->mbmi.mode = mode_selected;
Initial WebM release
2010-05-18 17:58:33 +02:00
return best_rd;
}
add 8x8 intra prediction modes Patch 1 to Patch 3 is an initial implementation of 8x8 intra prediction modes, here are with the following assumptions: a. 8x8 has 4 prediction modes DC, H, V and TM b. UV 4x4 block use the same mode as corresponding 8x8 area c. i8x8 modes are enabled for key frame only for now Patch 4: d. removed debug code from previous patches Patch 5: e. added stats code to collect entropy stats and further cleaned up Patch 6: f. changed mode stats code to collect finer stats of modes Patch 7: g. normalized i8x8 modes distribution to total at 256 (8bits). Patch 8: h. fixed a bug in decoder and removed debug printf output. Patch 9: i. more cleanups to address paul's comment Patch 10: j. messy rebase/merges to bring the commit up to date. Tests on HD clips encoded with all key frame showing consistent gain on all clips and all metrics:~0.5%(psnr) and 0.6%(ssim): http://www.corp.google.com/~yaowu/no_crawl/i8x8hd_allkey_fixedq.html To build and test, configure with: --enable-experimental --enable-i8x8 Change-Id: I9813fe07ae48cab5fdb5d904bca022514ad01e7f
2011-08-05 01:30:27 +02:00
static int rd_pick_intra8x8block(
VP8_COMP *cpi,
MACROBLOCK *x,
int ib,
B_PREDICTION_MODE *best_mode,
unsigned int *mode_costs,
ENTROPY_CONTEXT *a,
ENTROPY_CONTEXT *l,
int *bestrate,
int *bestratey,
int *bestdistortion)
{
MB_PREDICTION_MODE mode;
MACROBLOCKD *xd = &x->e_mbd;
int best_rd = INT_MAX;
int rate = 0;
int distortion;
BLOCK *be=x->block + ib;
BLOCKD *b=x->e_mbd.block + ib;
ENTROPY_CONTEXT ta0, ta1, besta0, besta1;
ENTROPY_CONTEXT tl0, tl1, bestl0, bestl1;
/*
* The predictor buffer is a 2d buffer with a stride of 16. Create
* a temp buffer that meets the stride requirements, but we are only
* interested in the left 8x8 block
* */
DECLARE_ALIGNED_ARRAY(16, unsigned char, best_predictor, 16*8);
DECLARE_ALIGNED_ARRAY(16, short, best_dqcoeff, 16*4);
for (mode = DC_PRED; mode <= TM_PRED; mode++)
{
int this_rd;
int rate_t;
rate = mode_costs[mode];
RECON_INVOKE(&cpi->rtcd.common->recon, intra8x8_predict)
(b, mode, b->predictor);
vp8_subtract_4b_c(be, b, 16);
x->vp8_short_fdct8x4(be->src_diff, be->coeff, 32);
x->vp8_short_fdct8x4(be->src_diff + 64, be->coeff + 64, 32);
x->quantize_b_pair(x->block+ib, x->block+ib+1,
xd->block+ib, xd->block+ib+1);
x->quantize_b_pair(x->block+ib+4, x->block+ib+5,
xd->block+ib+4, xd->block+ib+5);
distortion = ENCODEMB_INVOKE(IF_RTCD(&cpi->rtcd.encodemb), berr)
((x->block+ib)->coeff,(xd->block+ib)->dqcoeff)>>2;
distortion += ENCODEMB_INVOKE(IF_RTCD(&cpi->rtcd.encodemb), berr)
((x->block+ib+1)->coeff,(xd->block+ib+1)->dqcoeff)>>2;
distortion += ENCODEMB_INVOKE(IF_RTCD(&cpi->rtcd.encodemb), berr)
((x->block+ib+4)->coeff,(xd->block+ib+4)->dqcoeff)>>2;
distortion += ENCODEMB_INVOKE(IF_RTCD(&cpi->rtcd.encodemb), berr)
((x->block+ib+5)->coeff,(xd->block+ib+5)->dqcoeff)>>2;
ta0 = *(a + vp8_block2above[ib]);
ta1 = *(a + vp8_block2above[ib+1]);
tl0 = *(l + vp8_block2above[ib]);
tl1 = *(l + vp8_block2above[ib+4]);
rate_t = cost_coeffs(x, xd->block+ib, PLANE_TYPE_Y_WITH_DC,
&ta0, &tl0);
rate_t += cost_coeffs(x, xd->block+ib+1, PLANE_TYPE_Y_WITH_DC,
&ta1, &tl0);
rate_t += cost_coeffs(x, xd->block+ib+4, PLANE_TYPE_Y_WITH_DC,
&ta0, &tl1);
rate_t += cost_coeffs(x, xd->block+ib+5, PLANE_TYPE_Y_WITH_DC,
&ta1, &tl1);
rate += rate_t;
this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
if (this_rd < best_rd)
{
*bestrate = rate;
*bestratey = rate_t;
*bestdistortion = distortion;
besta0 = ta0;
besta1 = ta1;
bestl0 = tl0;
bestl1 = tl1;
best_rd = this_rd;
*best_mode = mode;
copy_predictor_8x8(best_predictor, b->predictor);
vpx_memcpy(best_dqcoeff, b->dqcoeff, 64);
vpx_memcpy(best_dqcoeff+32, b->dqcoeff+64, 64);
}
}
b->bmi.as_mode = (*best_mode);
vp8_encode_intra8x8 (IF_RTCD(&cpi->rtcd), x, ib);
*(a + vp8_block2above[ib]) = besta0;
*(a + vp8_block2above[ib+1]) = besta1;
*(l + vp8_block2above[ib]) = bestl0;
*(l + vp8_block2above[ib+4]) = bestl1;
return best_rd;
}
const int vp8_i8x8_block[4]={0, 2, 8, 10};
int rd_pick_intra8x8mby_modes(VP8_COMP *cpi,
MACROBLOCK *mb,
int *Rate,
int *rate_y,
int *Distortion,
int best_rd)
{
MACROBLOCKD *const xd = &mb->e_mbd;
int i,ib;
int cost = mb->mbmode_cost [xd->frame_type] [I8X8_PRED];
int distortion = 0;
int tot_rate_y = 0;
long long total_rd = 0;
ENTROPY_CONTEXT_PLANES t_above, t_left;
ENTROPY_CONTEXT *ta;
ENTROPY_CONTEXT *tl;
unsigned int *i8x8mode_costs;
vpx_memcpy(&t_above, mb->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES));
vpx_memcpy(&t_left, mb->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES));
ta = (ENTROPY_CONTEXT *)&t_above;
tl = (ENTROPY_CONTEXT *)&t_left;
i8x8mode_costs = mb->i8x8_mode_costs;
for (i = 0; i < 4; i++)
{
MODE_INFO *const mic = xd->mode_info_context;
const int mis = xd->mode_info_stride;
B_PREDICTION_MODE UNINITIALIZED_IS_SAFE(best_mode);
int UNINITIALIZED_IS_SAFE(r), UNINITIALIZED_IS_SAFE(ry), UNINITIALIZED_IS_SAFE(d);
ib = vp8_i8x8_block[i];
total_rd += rd_pick_intra8x8block(
cpi, mb, ib, &best_mode, i8x8mode_costs,
ta, tl, &r, &ry, &d);
cost += r;
distortion += d;
tot_rate_y += ry;
mic->bmi[ib].as_mode = best_mode;
}
*Rate = cost;
*rate_y += tot_rate_y;
*Distortion = distortion;
return RDCOST(mb->rdmult, mb->rddiv, cost, distortion);
}
Initial WebM release
2010-05-18 17:58:33 +02:00
static int rd_cost_mbuv(MACROBLOCK *mb)
{
int b;
int cost = 0;
MACROBLOCKD *x = &mb->e_mbd;
Changed above and left context data layout The main reason for the change was to reduce cycles in the token decoder. (~1.5% gain for 32 bit) This layout should be more cache friendly. As a result of this change, the encoder had to be updated. Change-Id: Id5e804169d8889da0378b3a519ac04dabd28c837 Note: dixie uses a similar layout
2010-08-31 16:49:57 +02:00
ENTROPY_CONTEXT_PLANES t_above, t_left;
ENTROPY_CONTEXT *ta;
ENTROPY_CONTEXT *tl;
vpx_memcpy(&t_above, mb->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES));
vpx_memcpy(&t_left, mb->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES));
Initial WebM release
2010-05-18 17:58:33 +02:00
Changed above and left context data layout The main reason for the change was to reduce cycles in the token decoder. (~1.5% gain for 32 bit) This layout should be more cache friendly. As a result of this change, the encoder had to be updated. Change-Id: Id5e804169d8889da0378b3a519ac04dabd28c837 Note: dixie uses a similar layout
2010-08-31 16:49:57 +02:00
ta = (ENTROPY_CONTEXT *)&t_above;
tl = (ENTROPY_CONTEXT *)&t_left;
Initial WebM release
2010-05-18 17:58:33 +02:00
Removed vp8_block2type and used defines instead. Change-Id: Idb56e0295d004793f406dfd2d8d8c546aad62e03
2011-02-23 22:37:08 +01:00
for (b = 16; b < 24; b++)
cost += cost_coeffs(mb, x->block + b, PLANE_TYPE_UV,
Changed above and left context data layout The main reason for the change was to reduce cycles in the token decoder. (~1.5% gain for 32 bit) This layout should be more cache friendly. As a result of this change, the encoder had to be updated. Change-Id: Id5e804169d8889da0378b3a519ac04dabd28c837 Note: dixie uses a similar layout
2010-08-31 16:49:57 +02:00
ta + vp8_block2above[b], tl + vp8_block2left[b]);
Initial WebM release
2010-05-18 17:58:33 +02:00
return cost;
}
Removed bmi copy to/from BLOCKD for SPLITMV and B_PRED modes. Modified code to use the bmi found in mode_info_context instead of BLOCKD. On the decode side, the uvmvs are calculated only when required, instead of every macroblock. This is WIP. (bmi should eventually be removed from BLOCKD) Small performance gains noticed for RT encodes and decodes.(VGA) Change-Id: I2ed7f0fd5ca733655df684aa82da575c77a973e7
2011-08-24 20:42:26 +02:00
static int rd_inter16x16_uv(VP8_COMP *cpi, MACROBLOCK *x, int *rate,
int *distortion, int fullpixel)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
Segment signaling of TX size Initial attempt at using new segment feature signaling to indicate 4x4 or 8x8 transform. needs --enable-experimental --enable-t8x8 Note this is work in progress. Change-Id: Ib160d46a5d810307bfcbc79853ce1a65b5b870b7
2011-11-04 19:29:51 +01:00
#if CONFIG_T8X8
int tx_type = get_seg_tx_type(&x->e_mbd,
x->e_mbd.mode_info_context->mbmi.segment_id);
#endif
Removed bmi copy to/from BLOCKD for SPLITMV and B_PRED modes. Modified code to use the bmi found in mode_info_context instead of BLOCKD. On the decode side, the uvmvs are calculated only when required, instead of every macroblock. This is WIP. (bmi should eventually be removed from BLOCKD) Small performance gains noticed for RT encodes and decodes.(VGA) Change-Id: I2ed7f0fd5ca733655df684aa82da575c77a973e7
2011-08-24 20:42:26 +02:00
ENCODEMB_INVOKE(IF_RTCD(&cpi->rtcd.encodemb), submbuv)(x->src_diff,
x->src.u_buffer, x->src.v_buffer, x->e_mbd.predictor, x->src.uv_stride);
Initial WebM release
2010-05-18 17:58:33 +02:00
Merge remote branch 'internal/upstream' into HEAD Conflicts: vp8/decoder/decodframe.c vp8/encoder/encodeframe.c vp8/encoder/encodemb.c Change-Id: I6e0d1669e4409a2dfd73ba2c7038d730842d3953
2011-09-16 15:22:29 +02:00
#if CONFIG_T8X8
Segment signaling of TX size Initial attempt at using new segment feature signaling to indicate 4x4 or 8x8 transform. needs --enable-experimental --enable-t8x8 Note this is work in progress. Change-Id: Ib160d46a5d810307bfcbc79853ce1a65b5b870b7
2011-11-04 19:29:51 +01:00
if( tx_type == TX_8X8 )
Merge remote branch 'internal/upstream' into HEAD Conflicts: vp8/decoder/decodframe.c vp8/encoder/encodeframe.c vp8/encoder/encodemb.c Change-Id: I6e0d1669e4409a2dfd73ba2c7038d730842d3953
2011-09-16 15:22:29 +02:00
vp8_transform_mbuv_8x8(x);
else
#endif
Removed bmi copy to/from BLOCKD for SPLITMV and B_PRED modes. Modified code to use the bmi found in mode_info_context instead of BLOCKD. On the decode side, the uvmvs are calculated only when required, instead of every macroblock. This is WIP. (bmi should eventually be removed from BLOCKD) Small performance gains noticed for RT encodes and decodes.(VGA) Change-Id: I2ed7f0fd5ca733655df684aa82da575c77a973e7
2011-08-24 20:42:26 +02:00
vp8_transform_mbuv(x);
vp8_quantize_mbuv(x);
*rate = rd_cost_mbuv(x);
*distortion = ENCODEMB_INVOKE(&cpi->rtcd.encodemb, mbuverr)(x) / 4;
return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
}
static int rd_inter4x4_uv(VP8_COMP *cpi, MACROBLOCK *x, int *rate,
int *distortion, int fullpixel)
{
vp8_build_inter4x4_predictors_mbuv(&x->e_mbd);
ENCODEMB_INVOKE(IF_RTCD(&cpi->rtcd.encodemb), submbuv)(x->src_diff,
x->src.u_buffer, x->src.v_buffer, x->e_mbd.predictor, x->src.uv_stride);
vp8_transform_mbuv(x);
vp8_quantize_mbuv(x);
Initial WebM release
2010-05-18 17:58:33 +02:00
*rate = rd_cost_mbuv(x);
*distortion = ENCODEMB_INVOKE(&cpi->rtcd.encodemb, mbuverr)(x) / 4;
Removed cpi->target_bits_per_mb cpi->target_bits_per_mb is currently not being used, so delete it. Also removed other unused code in rdopt.c. Change-Id: I98449f9030bcd2f15451d9b7a3b9b93dd1409923
2011-01-07 15:41:13 +01:00
return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
Initial WebM release
2010-05-18 17:58:33 +02:00
}
Move RD intra block mode selection to rdopt.c This change is analogous to I0b67dae1f8a74902378da7bdf565e39ab832dda7, which made the move for the non-RD path. Change-Id: If63fc1b0cd1eb7f932e710f83ff24d91454f8ed1
2011-06-08 18:05:05 +02:00
static void rd_pick_intra_mbuv_mode(VP8_COMP *cpi, MACROBLOCK *x, int *rate, int *rate_tokenonly, int *distortion)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
MB_PREDICTION_MODE mode;
MB_PREDICTION_MODE UNINITIALIZED_IS_SAFE(mode_selected);
int best_rd = INT_MAX;
int UNINITIALIZED_IS_SAFE(d), UNINITIALIZED_IS_SAFE(r);
int rate_to;
for (mode = DC_PRED; mode <= TM_PRED; mode++)
{
int rate;
int distortion;
int this_rd;
Removed unnecessary MB_MODE_INFO copies These copies occurred for each macroblock in the encoder and decoder. Thetemp MB_MODE_INFO mbmi was removed from MACROBLOCKD. As a result, a large number compile errors had to be fixed. Change-Id: I4cf0ffae3ce244f6db04a4c217d52dd256382cf3
2010-08-12 22:25:43 +02:00
x->e_mbd.mode_info_context->mbmi.uv_mode = mode;
SSE2/SSSE3 optimizations for build_predictors_mbuv{,_s}(). decoding before 10.425 10.432 10.423 =10.426 after: 10.405 10.416 10.398 =10.406, 0.2% faster encoding before 14.252 14.331 14.250 14.223 14.241 14.220 14.221 =14.248 after 14.095 14.090 14.085 14.095 14.064 14.081 14.089 =14.086, 1.1% faster Change-Id: I483d3d8f0deda8ad434cea76e16028380722aee2
2011-04-27 19:05:10 +02:00
RECON_INVOKE(&cpi->rtcd.common->recon, build_intra_predictors_mbuv)
(&x->e_mbd);
Improved vp8_rd_pick_intra_mbuv_mode Eliminated unnecessary calculations. Very small change to performance. Change-Id: Ib7213d43c64e36955177c4d47950ff472266f822
2011-02-14 22:34:33 +01:00
ENCODEMB_INVOKE(IF_RTCD(&cpi->rtcd.encodemb), submbuv)(x->src_diff,
x->src.u_buffer, x->src.v_buffer, x->e_mbd.predictor,
x->src.uv_stride);
vp8_transform_mbuv(x);
vp8_quantize_mbuv(x);
Initial WebM release
2010-05-18 17:58:33 +02:00
rate_to = rd_cost_mbuv(x);
Removed unnecessary MB_MODE_INFO copies These copies occurred for each macroblock in the encoder and decoder. Thetemp MB_MODE_INFO mbmi was removed from MACROBLOCKD. As a result, a large number compile errors had to be fixed. Change-Id: I4cf0ffae3ce244f6db04a4c217d52dd256382cf3
2010-08-12 22:25:43 +02:00
rate = rate_to + x->intra_uv_mode_cost[x->e_mbd.frame_type][x->e_mbd.mode_info_context->mbmi.uv_mode];
Initial WebM release
2010-05-18 17:58:33 +02:00
Inconsistent distortion metric in vp8_rd_pick_intra_mbuv_mode This function was using a variance metric compared to and SSE metric in other places (eg. vp8_rd_inter_uv) Change-Id: I9109fcc5a13bca9db1d7ead500fe14999ab233eb
2011-01-28 14:13:30 +01:00
distortion = ENCODEMB_INVOKE(&cpi->rtcd.encodemb, mbuverr)(x) / 4;
Initial WebM release
2010-05-18 17:58:33 +02:00
Removed cpi->target_bits_per_mb cpi->target_bits_per_mb is currently not being used, so delete it. Also removed other unused code in rdopt.c. Change-Id: I98449f9030bcd2f15451d9b7a3b9b93dd1409923
2011-01-07 15:41:13 +01:00
this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
Initial WebM release
2010-05-18 17:58:33 +02:00
if (this_rd < best_rd)
{
best_rd = this_rd;
d = distortion;
r = rate;
*rate_tokenonly = rate_to;
mode_selected = mode;
}
}
*rate = r;
*distortion = d;
Removed unnecessary MB_MODE_INFO copies These copies occurred for each macroblock in the encoder and decoder. Thetemp MB_MODE_INFO mbmi was removed from MACROBLOCKD. As a result, a large number compile errors had to be fixed. Change-Id: I4cf0ffae3ce244f6db04a4c217d52dd256382cf3
2010-08-12 22:25:43 +02:00
x->e_mbd.mode_info_context->mbmi.uv_mode = mode_selected;
Initial WebM release
2010-05-18 17:58:33 +02:00
}
made vp8_mode_context adaptive vp8_mode_contexts[] is an entropy table used to code inter mode choices. It was a fixed constant table. This commit made the entropy context adaptive. Tests on derf set showed very good consistent gains on all metrics: avg psnr .47%, overall psnr .46% and ssim .40%. http://www.corp.google.com/~yaowu/no_crawl/newModeContext.html Change-Id: Ia62b14485c948e2b74586118619c5eb2068b43b2
2011-12-06 21:03:42 +01:00
int vp8_cost_mv_ref(VP8_COMMON *pc,
MB_PREDICTION_MODE m,
const int near_mv_ref_ct[4])
Initial WebM release
2010-05-18 17:58:33 +02:00
{
vp8_prob p [VP8_MVREFS-1];
assert(NEARESTMV <= m && m <= SPLITMV);
made vp8_mode_context adaptive vp8_mode_contexts[] is an entropy table used to code inter mode choices. It was a fixed constant table. This commit made the entropy context adaptive. Tests on derf set showed very good consistent gains on all metrics: avg psnr .47%, overall psnr .46% and ssim .40%. http://www.corp.google.com/~yaowu/no_crawl/newModeContext.html Change-Id: Ia62b14485c948e2b74586118619c5eb2068b43b2
2011-12-06 21:03:42 +01:00
vp8_mv_ref_probs(pc, p, near_mv_ref_ct);
avoid negative array subscript warnings The mv_ref and sub_mv_ref token encodings are indexed from NEARESTMV and LEFT4X4, respectively, rather than being zero-based like the other token encodings. Change-Id: I3699c3f84111209ecfb91097c4b900773e9a3ad5
2010-08-09 19:27:26 +02:00
return vp8_cost_token(vp8_mv_ref_tree, p,
vp8_mv_ref_encoding_array - NEARESTMV + m);
Initial WebM release
2010-05-18 17:58:33 +02:00
}
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
void vp8_set_mbmode_and_mvs(MACROBLOCK *x, MB_PREDICTION_MODE mb, int_mv *mv)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
Removed unnecessary MB_MODE_INFO copies These copies occurred for each macroblock in the encoder and decoder. Thetemp MB_MODE_INFO mbmi was removed from MACROBLOCKD. As a result, a large number compile errors had to be fixed. Change-Id: I4cf0ffae3ce244f6db04a4c217d52dd256382cf3
2010-08-12 22:25:43 +02:00
x->e_mbd.mode_info_context->mbmi.mode = mb;
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
x->e_mbd.mode_info_context->mbmi.mv.as_int = mv->as_int;
Initial WebM release
2010-05-18 17:58:33 +02:00
}
static int labels2mode(
MACROBLOCK *x,
int const *labelings, int which_label,
B_PREDICTION_MODE this_mode,
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
int_mv *this_mv, int_mv *best_ref_mv,
Initial WebM release
2010-05-18 17:58:33 +02:00
int *mvcost[2]
)
{
MACROBLOCKD *const xd = & x->e_mbd;
MODE_INFO *const mic = xd->mode_info_context;
const int mis = xd->mode_info_stride;
int cost = 0;
int thismvcost = 0;
/* We have to be careful retrieving previously-encoded motion vectors.
Ones from this macroblock have to be pulled from the BLOCKD array
as they have not yet made it to the bmi array in our MB_MODE_INFO. */
int i = 0;
do
{
BLOCKD *const d = xd->block + i;
const int row = i >> 2, col = i & 3;
B_PREDICTION_MODE m;
if (labelings[i] != which_label)
continue;
if (col && labelings[i] == labelings[i-1])
m = LEFT4X4;
else if (row && labelings[i] == labelings[i-4])
m = ABOVE4X4;
else
{
// the only time we should do costing for new motion vector or mode
// is when we are on a new label (jbb May 08, 2007)
switch (m = this_mode)
{
case NEW4X4 :
thismvcost = vp8_mv_bit_cost(this_mv, best_ref_mv, mvcost, 102);
break;
case LEFT4X4:
MODE_INFO size reduction Declared the bmi in MODE_INFO as a union instead of B_MODE_INFO. This reduced the memory footprint by 518,400 bytes for 1080 resolutions. The decoder performance improved by ~4% for the clip used and the encoder showed very small improvements. (0.5%) This reduction was first mentioned to me by John K. and in a later discussion by Yaowu. This is WIP. Change-Id: I8e175fdbc46d28c35277302a04bee4540efc8d29
2011-05-24 19:24:52 +02:00
this_mv->as_int = col ? d[-1].bmi.mv.as_int : left_block_mv(mic, i);
Initial WebM release
2010-05-18 17:58:33 +02:00
break;
case ABOVE4X4:
MODE_INFO size reduction Declared the bmi in MODE_INFO as a union instead of B_MODE_INFO. This reduced the memory footprint by 518,400 bytes for 1080 resolutions. The decoder performance improved by ~4% for the clip used and the encoder showed very small improvements. (0.5%) This reduction was first mentioned to me by John K. and in a later discussion by Yaowu. This is WIP. Change-Id: I8e175fdbc46d28c35277302a04bee4540efc8d29
2011-05-24 19:24:52 +02:00
this_mv->as_int = row ? d[-4].bmi.mv.as_int : above_block_mv(mic, i, mis);
Initial WebM release
2010-05-18 17:58:33 +02:00
break;
case ZERO4X4:
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
this_mv->as_int = 0;
Initial WebM release
2010-05-18 17:58:33 +02:00
break;
default:
break;
}
if (m == ABOVE4X4) // replace above with left if same
{
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
int_mv left_mv;
MODE_INFO size reduction Declared the bmi in MODE_INFO as a union instead of B_MODE_INFO. This reduced the memory footprint by 518,400 bytes for 1080 resolutions. The decoder performance improved by ~4% for the clip used and the encoder showed very small improvements. (0.5%) This reduction was first mentioned to me by John K. and in a later discussion by Yaowu. This is WIP. Change-Id: I8e175fdbc46d28c35277302a04bee4540efc8d29
2011-05-24 19:24:52 +02:00
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
left_mv.as_int = col ? d[-1].bmi.mv.as_int :
MODE_INFO size reduction Declared the bmi in MODE_INFO as a union instead of B_MODE_INFO. This reduced the memory footprint by 518,400 bytes for 1080 resolutions. The decoder performance improved by ~4% for the clip used and the encoder showed very small improvements. (0.5%) This reduction was first mentioned to me by John K. and in a later discussion by Yaowu. This is WIP. Change-Id: I8e175fdbc46d28c35277302a04bee4540efc8d29
2011-05-24 19:24:52 +02:00
left_block_mv(mic, i);
Initial WebM release
2010-05-18 17:58:33 +02:00
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
if (left_mv.as_int == this_mv->as_int)
Initial WebM release
2010-05-18 17:58:33 +02:00
m = LEFT4X4;
}
cost = x->inter_bmode_costs[ m];
}
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
d->bmi.mv.as_int = this_mv->as_int;
Initial WebM release
2010-05-18 17:58:33 +02:00
Using partition_info instead of blockd info for splitmv The partition_info struct contains info just for SPLITMV, so it should be used instead of BLOCKD. Eventually, I want to reduce the size of B_MODE_INFO struct found in BLOCKD, so this is the first step toward that goal. Also, since SPLITMV is not supported in vp8_pick_inter_mode(), the unnecessary mem copies and checks were removed. For rt encodes, this gave a slight performance improvement. Change-Id: I5585c98fa9d5acbde1c7e0f452a01d9ecc080574
2011-05-19 21:03:36 +02:00
x->partition_info->bmi[i].mode = m;
x->partition_info->bmi[i].mv.as_int = this_mv->as_int;
Initial WebM release
2010-05-18 17:58:33 +02:00
}
while (++i < 16);
cost += thismvcost ;
return cost;
}
Changed above and left context data layout The main reason for the change was to reduce cycles in the token decoder. (~1.5% gain for 32 bit) This layout should be more cache friendly. As a result of this change, the encoder had to be updated. Change-Id: Id5e804169d8889da0378b3a519ac04dabd28c837 Note: dixie uses a similar layout
2010-08-31 16:49:57 +02:00
static int rdcost_mbsegment_y(MACROBLOCK *mb, const int *labels,
int which_label, ENTROPY_CONTEXT *ta,
ENTROPY_CONTEXT *tl)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
int cost = 0;
int b;
MACROBLOCKD *x = &mb->e_mbd;
for (b = 0; b < 16; b++)
if (labels[ b] == which_label)
Removed vp8_block2type and used defines instead. Change-Id: Idb56e0295d004793f406dfd2d8d8c546aad62e03
2011-02-23 22:37:08 +01:00
cost += cost_coeffs(mb, x->block + b, PLANE_TYPE_Y_WITH_DC,
Changed above and left context data layout The main reason for the change was to reduce cycles in the token decoder. (~1.5% gain for 32 bit) This layout should be more cache friendly. As a result of this change, the encoder had to be updated. Change-Id: Id5e804169d8889da0378b3a519ac04dabd28c837 Note: dixie uses a similar layout
2010-08-31 16:49:57 +02:00
ta + vp8_block2above[b],
tl + vp8_block2left[b]);
Initial WebM release
2010-05-18 17:58:33 +02:00
return cost;
}
static unsigned int vp8_encode_inter_mb_segment(MACROBLOCK *x, int const *labels, int which_label, const vp8_encodemb_rtcd_vtable_t *rtcd)
{
int i;
unsigned int distortion = 0;
for (i = 0; i < 16; i++)
{
if (labels[i] == which_label)
{
BLOCKD *bd = &x->e_mbd.block[i];
BLOCK *be = &x->block[i];
vp8_build_inter_predictors_b(bd, 16, x->e_mbd.subpixel_predict);
ENCODEMB_INVOKE(rtcd, subb)(be, bd, 16);
Redo the forward 4x4 dct The new fdct lowers the round trip sum squared error for a 4x4 block ~0.12. or ~0.008/pixel. For reference, the old matrix multiply version has average round trip error 1.46 for a 4x4 block. Thanks to "derf" for his suggestions and references. Change-Id: I5559d1e81d333b319404ab16b336b739f87afc79
2010-06-16 21:52:18 +02:00
x->vp8_short_fdct4x4(be->src_diff, be->coeff, 32);
Initial WebM release
2010-05-18 17:58:33 +02:00
// set to 0 no way to account for 2nd order DC so discount
//be->coeff[0] = 0;
Remove duplicate and unused functions Change-Id: I944035e720ef834561a9da0d723879a4f787312c
2010-06-04 00:08:44 +02:00
x->quantize_b(be, bd);
Initial WebM release
2010-05-18 17:58:33 +02:00
distortion += ENCODEMB_INVOKE(rtcd, berr)(be->coeff, bd->dqcoeff);
}
}
return distortion;
}
fix build issues for experimental branch experimental branch build was broken from some merge artifacts, this commit fixes those issues to enable the experimental branch to build. Change-Id: Ic52b2d2f1d1b80abb7ecaa4c0927bcf887ac0c2a
2011-06-23 00:07:04 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
static const unsigned int segmentation_to_sseshift[4] = {3, 3, 2, 0};
typedef struct
Initial WebM release
2010-05-18 17:58:33 +02:00
{
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
int_mv *ref_mv;
int_mv mvp;
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
int segment_rd;
int segment_num;
int r;
int d;
int segment_yrate;
B_PREDICTION_MODE modes[16];
Changed segmentation check order In SPLITMV, the 8x8 segment will be checked first. If the 8x8 rd is better than the best, we check the other segments. Otherwise bail. Adjustments to the thresh_mult were necessary to make up for the initial quality loss. The performance improved by 20% (average) for good quality, speed 0 and speed 1, while the overall quality remained the same. Change-Id: I717aef401323c8a254fba3e9777d2a316c774cc3
2010-12-16 23:01:27 +01:00
int_mv mvs[16];
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
unsigned char eobs[16];
int mvthresh;
int *mdcounts;
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
int_mv sv_mvp[4]; // save 4 mvp from 8x8
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
int sv_istep[2]; // save 2 initial step_param for 16x8/8x16
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
} BEST_SEG_INFO;
Increase static linkage, remove unused functions A large number of functions were defined with external linkage, even though they were only used from within one file. This patch changes their linkage to static and removes the vp8_ prefix from their names, which should make it more obvious to the reader that the function is contained within the current translation unit. Functions that were not referenced were removed. These symbols were identified by: $ nm -A libvpx.a | sort -k3 | uniq -c -f2 | grep ' [A-Z] ' \ | sort | grep '^ *1 ' Change-Id: I59609f58ab65312012c047036ae1e0634f795779
2011-03-17 22:07:59 +01:00
static void rd_check_segment(VP8_COMP *cpi, MACROBLOCK *x,
BEST_SEG_INFO *bsi, unsigned int segmentation)
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
{
int i;
Initial WebM release
2010-05-18 17:58:33 +02:00
int const *labels;
int br = 0;
int bd = 0;
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
B_PREDICTION_MODE this_mode;
Initial WebM release
2010-05-18 17:58:33 +02:00
Add half-pixel variance RTCD functions NEON has optimized 16x16 half-pixel variance functions, but they were not part of the RTCD framework. Add these functions to RTCD, so that other platforms can make use of this optimization in the future and special-case ARM code can be removed. A number of functions were taking two variance functions as parameters. These functions were changed to take a single parameter, a pointer to a struct containing all the variance functions for that block size. This provides additional flexibility for calling additional variance functions (the half-pixel special case, for example) and by initializing the table for all block sizes, we don't have to construct this function pointer table for each macroblock. Change-Id: I78289ff36b2715f9a7aa04d5f6fbe3d23acdc29c
2010-10-26 21:34:16 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
int label_count;
int this_segment_rd = 0;
int label_mv_thresh;
int rate = 0;
int sbr = 0;
int sbd = 0;
int segmentyrate = 0;
vp8_variance_fn_ptr_t *v_fn_ptr;
ENTROPY_CONTEXT_PLANES t_above, t_left;
ENTROPY_CONTEXT *ta;
ENTROPY_CONTEXT *tl;
ENTROPY_CONTEXT_PLANES t_above_b, t_left_b;
ENTROPY_CONTEXT *ta_b;
ENTROPY_CONTEXT *tl_b;
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
vpx_memcpy(&t_above, x->e_mbd.above_context, sizeof(ENTROPY_CONTEXT_PLANES));
vpx_memcpy(&t_left, x->e_mbd.left_context, sizeof(ENTROPY_CONTEXT_PLANES));
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
ta = (ENTROPY_CONTEXT *)&t_above;
tl = (ENTROPY_CONTEXT *)&t_left;
ta_b = (ENTROPY_CONTEXT *)&t_above_b;
tl_b = (ENTROPY_CONTEXT *)&t_left_b;
br = 0;
bd = 0;
Fix compiler warnings Change-Id: I2a97f08cc3c7808ce5be39e910cc5147ecf03a1d
2010-06-11 20:33:49 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
v_fn_ptr = &cpi->fn_ptr[segmentation];
labels = vp8_mbsplits[segmentation];
label_count = vp8_mbsplit_count[segmentation];
Fix compiler warnings Change-Id: I2a97f08cc3c7808ce5be39e910cc5147ecf03a1d
2010-06-11 20:33:49 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
// 64 makes this threshold really big effectively
// making it so that we very rarely check mvs on
// segments. setting this to 1 would make mv thresh
// roughly equal to what it is for macroblocks
label_mv_thresh = 1 * bsi->mvthresh / label_count ;
// Segmentation method overheads
rate = vp8_cost_token(vp8_mbsplit_tree, vp8_mbsplit_probs, vp8_mbsplit_encodings + segmentation);
made vp8_mode_context adaptive vp8_mode_contexts[] is an entropy table used to code inter mode choices. It was a fixed constant table. This commit made the entropy context adaptive. Tests on derf set showed very good consistent gains on all metrics: avg psnr .47%, overall psnr .46% and ssim .40%. http://www.corp.google.com/~yaowu/no_crawl/newModeContext.html Change-Id: Ia62b14485c948e2b74586118619c5eb2068b43b2
2011-12-06 21:03:42 +01:00
rate += vp8_cost_mv_ref(&cpi->common, SPLITMV, bsi->mdcounts);
Removed cpi->target_bits_per_mb cpi->target_bits_per_mb is currently not being used, so delete it. Also removed other unused code in rdopt.c. Change-Id: I98449f9030bcd2f15451d9b7a3b9b93dd1409923
2011-01-07 15:41:13 +01:00
this_segment_rd += RDCOST(x->rdmult, x->rddiv, rate, 0);
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
br += rate;
for (i = 0; i < label_count; i++)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
int_mv mode_mv[B_MODE_COUNT];
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
int best_label_rd = INT_MAX;
B_PREDICTION_MODE mode_selected = ZERO4X4;
int bestlabelyrate = 0;
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
// search for the best motion vector on this segment
for (this_mode = LEFT4X4; this_mode <= NEW4X4 ; this_mode ++)
{
int this_rd;
int distortion;
int labelyrate;
ENTROPY_CONTEXT_PLANES t_above_s, t_left_s;
ENTROPY_CONTEXT *ta_s;
ENTROPY_CONTEXT *tl_s;
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
vpx_memcpy(&t_above_s, &t_above, sizeof(ENTROPY_CONTEXT_PLANES));
vpx_memcpy(&t_left_s, &t_left, sizeof(ENTROPY_CONTEXT_PLANES));
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
ta_s = (ENTROPY_CONTEXT *)&t_above_s;
tl_s = (ENTROPY_CONTEXT *)&t_left_s;
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
if (this_mode == NEW4X4)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
int sseshift;
Initial WebM release
2010-05-18 17:58:33 +02:00
int num00;
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
int step_param = 0;
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
int further_steps;
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
int n;
int thissme;
int bestsme = INT_MAX;
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
int_mv temp_mv;
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
BLOCK *c;
BLOCKD *e;
Changed above and left context data layout The main reason for the change was to reduce cycles in the token decoder. (~1.5% gain for 32 bit) This layout should be more cache friendly. As a result of this change, the encoder had to be updated. Change-Id: Id5e804169d8889da0378b3a519ac04dabd28c837 Note: dixie uses a similar layout
2010-08-31 16:49:57 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
// Is the best so far sufficiently good that we cant justify doing and new motion search.
if (best_label_rd < label_mv_thresh)
break;
Initial WebM release
2010-05-18 17:58:33 +02:00
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
if(cpi->compressor_speed)
{
if (segmentation == BLOCK_8X16 || segmentation == BLOCK_16X8)
{
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
bsi->mvp.as_int = bsi->sv_mvp[i].as_int;
if (i==1 && segmentation == BLOCK_16X8)
bsi->mvp.as_int = bsi->sv_mvp[2].as_int;
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
step_param = bsi->sv_istep[i];
}
// use previous block's result as next block's MV predictor.
if (segmentation == BLOCK_4X4 && i>0)
{
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
bsi->mvp.as_int = x->e_mbd.block[i-1].bmi.mv.as_int;
if (i==4 || i==8 || i==12)
bsi->mvp.as_int = x->e_mbd.block[i-4].bmi.mv.as_int;
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
step_param = 2;
}
}
further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
Initial WebM release
2010-05-18 17:58:33 +02:00
{
further clean up of errorperbit and sadperbit this commit makes the usage errorperbit and sadperbit consistent for encoding modes and passes. Removed all different magic weight factors associated with errorperbit. Now 1/2 is used for both sadperbit16 and sadperbit4, the /2 operation is merged into initializations of the 2 variables. Tests on cif set show .23%, 0.18% and 0.19% gain by avg psnr, overall psnr and ssim respectively. Change-Id: Ifa285c3e065ce0a5a77addfc9f95aabf54ee270d
2011-06-01 20:41:05 +02:00
int sadpb = x->sadperbit4;
Adjust full-pixel clamping and motion vector limit calculation Do mvp clamping in full-pixel precision instead of 1/8-pixel precision to avoid error caused by right shifting operation. Also, further fixed the motion vector limit calculation in change: b7480454706a6b15bf091e659cd6227ab373c1a6 Change-Id: Ied88a4f7ddfb0476eb9f7afc6ceeddbf209fffd7
2011-07-07 17:21:41 +02:00
int_mv mvp_full;
mvp_full.as_mv.row = bsi->mvp.as_mv.row >>3;
mvp_full.as_mv.col = bsi->mvp.as_mv.col >>3;
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
// find first label
Increase static linkage, remove unused functions A large number of functions were defined with external linkage, even though they were only used from within one file. This patch changes their linkage to static and removes the vp8_ prefix from their names, which should make it more obvious to the reader that the function is contained within the current translation unit. Functions that were not referenced were removed. These symbols were identified by: $ nm -A libvpx.a | sort -k3 | uniq -c -f2 | grep ' [A-Z] ' \ | sort | grep '^ *1 ' Change-Id: I59609f58ab65312012c047036ae1e0634f795779
2011-03-17 22:07:59 +01:00
n = vp8_mbsplit_offset[segmentation][i];
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
c = &x->block[n];
e = &x->e_mbd.block[n];
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
{
Adjust full-pixel clamping and motion vector limit calculation Do mvp clamping in full-pixel precision instead of 1/8-pixel precision to avoid error caused by right shifting operation. Also, further fixed the motion vector limit calculation in change: b7480454706a6b15bf091e659cd6227ab373c1a6 Change-Id: Ied88a4f7ddfb0476eb9f7afc6ceeddbf209fffd7
2011-07-07 17:21:41 +02:00
bestsme = cpi->diamond_search_sad(x, c, e, &mvp_full,
remove some magic weights associated with sad_per_bit sad_per_bit has been used for a number of motion vector search routines with different magic weights: 1, 1/2 and 1/4. This commit remove these magic numbers and use 1/2 for all motion search routines, also reformat a number of source code lines to within 80 column limit. Test on cif set shows overall effect is neutral on all metrics. <=0.01% Change-Id: I8a382821fa4cffc9c0acf8e8431435a03df74885
2010-07-01 03:58:54 +02:00
&mode_mv[NEW4X4], step_param,
sadpb, &num00, v_fn_ptr,
x->mvcost, bsi->ref_mv);
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
n = num00;
num00 = 0;
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
while (n < further_steps)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
n++;
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
if (num00)
num00--;
Initial WebM release
2010-05-18 17:58:33 +02:00
else
{
remove some magic weights associated with sad_per_bit sad_per_bit has been used for a number of motion vector search routines with different magic weights: 1, 1/2 and 1/4. This commit remove these magic numbers and use 1/2 for all motion search routines, also reformat a number of source code lines to within 80 column limit. Test on cif set shows overall effect is neutral on all metrics. <=0.01% Change-Id: I8a382821fa4cffc9c0acf8e8431435a03df74885
2010-07-01 03:58:54 +02:00
thissme = cpi->diamond_search_sad(x, c, e,
Adjust full-pixel clamping and motion vector limit calculation Do mvp clamping in full-pixel precision instead of 1/8-pixel precision to avoid error caused by right shifting operation. Also, further fixed the motion vector limit calculation in change: b7480454706a6b15bf091e659cd6227ab373c1a6 Change-Id: Ied88a4f7ddfb0476eb9f7afc6ceeddbf209fffd7
2011-07-07 17:21:41 +02:00
&mvp_full, &temp_mv,
remove some magic weights associated with sad_per_bit sad_per_bit has been used for a number of motion vector search routines with different magic weights: 1, 1/2 and 1/4. This commit remove these magic numbers and use 1/2 for all motion search routines, also reformat a number of source code lines to within 80 column limit. Test on cif set shows overall effect is neutral on all metrics. <=0.01% Change-Id: I8a382821fa4cffc9c0acf8e8431435a03df74885
2010-07-01 03:58:54 +02:00
step_param + n, sadpb,
&num00, v_fn_ptr,
x->mvcost, bsi->ref_mv);
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
if (thissme < bestsme)
{
bestsme = thissme;
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
mode_mv[NEW4X4].as_int = temp_mv.as_int;
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
}
Initial WebM release
2010-05-18 17:58:33 +02:00
}
}
}
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
sseshift = segmentation_to_sseshift[segmentation];
// Should we do a full search (best quality only)
if ((cpi->compressor_speed == 0) && (bestsme >> sseshift) > 4000)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
Preload reference area in sub-pixel motion search (real-time mode) This change implemented same idea in change "Preload reference area to an intermediate buffer in sub-pixel motion search." The changes were made to vp8_find_best_sub_pixel_step() and vp8_find_best_half _pixel_step() functions which are called when speed >= 5. Test result (using tulip clip): 1. On Core2 Quad machine(Linux) rt mode, speed (-5 ~ -8), encoding speed gain: 2% ~ 3% rt mode, speed (-9 ~ -11), encoding speed gain: 1% ~ 2% rt mode, speed (-12 ~ -14), no noticeable encoding speed gain 2. On Xeon machine(Linux) Test on speed (-5 ~ -14) didn't show noticeable speed change. Change-Id: I21bec2d6e7fbe541fcc0f4c0366bbdf3e2076aa2
2011-07-22 22:01:11 +02:00
/* Check if mvp_full is within the range. */
vp8_clamp_mv(&mvp_full, x->mv_col_min, x->mv_col_max, x->mv_row_min, x->mv_row_max);
Adjust full-pixel clamping and motion vector limit calculation Do mvp clamping in full-pixel precision instead of 1/8-pixel precision to avoid error caused by right shifting operation. Also, further fixed the motion vector limit calculation in change: b7480454706a6b15bf091e659cd6227ab373c1a6 Change-Id: Ied88a4f7ddfb0476eb9f7afc6ceeddbf209fffd7
2011-07-07 17:21:41 +02:00
thissme = cpi->full_search_sad(x, c, e, &mvp_full,
remove some magic weights associated with sad_per_bit sad_per_bit has been used for a number of motion vector search routines with different magic weights: 1, 1/2 and 1/4. This commit remove these magic numbers and use 1/2 for all motion search routines, also reformat a number of source code lines to within 80 column limit. Test on cif set shows overall effect is neutral on all metrics. <=0.01% Change-Id: I8a382821fa4cffc9c0acf8e8431435a03df74885
2010-07-01 03:58:54 +02:00
sadpb, 16, v_fn_ptr,
x->mvcost, bsi->ref_mv);
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
if (thissme < bestsme)
{
bestsme = thissme;
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
mode_mv[NEW4X4].as_int = e->bmi.mv.as_int;
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
}
Initial WebM release
2010-05-18 17:58:33 +02:00
else
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
{
// The full search result is actually worse so re-instate the previous best vector
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
e->bmi.mv.as_int = mode_mv[NEW4X4].as_int;
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
}
Initial WebM release
2010-05-18 17:58:33 +02:00
}
}
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
if (bestsme < INT_MAX)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
Reduce unnecessary distortion computation In vp8_pick_inter_mode(), for NEWMV mode, use the error result got from motion search as distortion. This helps performance in real- time mode. Change-Id: I398c4e46cc5381f7d874e748cf78827ef0e0860c
2011-04-14 21:53:33 +02:00
int distortion;
Use sub-pixel search's SSE in mode selection Passed SSE from sub-pixel search back to pick_inter_mode function, which is compared with the encode_breakout to see if we could skip evaluating the remaining modes. Change-Id: I4a86442834f0d1b880a19e21ea52d17d505f941d
2011-04-18 21:48:34 +02:00
unsigned int sse;
further clean up of errorperbit and sadperbit this commit makes the usage errorperbit and sadperbit consistent for encoding modes and passes. Removed all different magic weight factors associated with errorperbit. Now 1/2 is used for both sadperbit16 and sadperbit4, the /2 operation is merged into initializations of the 2 variables. Tests on cif set show .23%, 0.18% and 0.19% gain by avg psnr, overall psnr and ssim respectively. Change-Id: Ifa285c3e065ce0a5a77addfc9f95aabf54ee270d
2011-06-01 20:41:05 +02:00
cpi->find_fractional_mv_step(x, c, e, &mode_mv[NEW4X4],
bsi->ref_mv, x->errorperbit, v_fn_ptr, x->mvcost,
&distortion, &sse);
Reduce unnecessary distortion computation In vp8_pick_inter_mode(), for NEWMV mode, use the error result got from motion search as distortion. This helps performance in real- time mode. Change-Id: I398c4e46cc5381f7d874e748cf78827ef0e0860c
2011-04-14 21:53:33 +02:00
Initial WebM release
2010-05-18 17:58:33 +02:00
}
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
} /* NEW4X4 */
rate = labels2mode(x, labels, i, this_mode, &mode_mv[this_mode],
bsi->ref_mv, x->mvcost);
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
// Trap vectors that reach beyond the UMV borders
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
if (((mode_mv[this_mode].as_mv.row >> 3) < x->mv_row_min) || ((mode_mv[this_mode].as_mv.row >> 3) > x->mv_row_max) ||
((mode_mv[this_mode].as_mv.col >> 3) < x->mv_col_min) || ((mode_mv[this_mode].as_mv.col >> 3) > x->mv_col_max))
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
{
continue;
}
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
distortion = vp8_encode_inter_mb_segment(x, labels, i, IF_RTCD(&cpi->rtcd.encodemb)) / 4;
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
labelyrate = rdcost_mbsegment_y(x, labels, i, ta_s, tl_s);
rate += labelyrate;
Initial WebM release
2010-05-18 17:58:33 +02:00
Removed cpi->target_bits_per_mb cpi->target_bits_per_mb is currently not being used, so delete it. Also removed other unused code in rdopt.c. Change-Id: I98449f9030bcd2f15451d9b7a3b9b93dd1409923
2011-01-07 15:41:13 +01:00
this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
Changed above and left context data layout The main reason for the change was to reduce cycles in the token decoder. (~1.5% gain for 32 bit) This layout should be more cache friendly. As a result of this change, the encoder had to be updated. Change-Id: Id5e804169d8889da0378b3a519ac04dabd28c837 Note: dixie uses a similar layout
2010-08-31 16:49:57 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
if (this_rd < best_label_rd)
{
sbr = rate;
sbd = distortion;
bestlabelyrate = labelyrate;
mode_selected = this_mode;
best_label_rd = this_rd;
vpx_memcpy(ta_b, ta_s, sizeof(ENTROPY_CONTEXT_PLANES));
vpx_memcpy(tl_b, tl_s, sizeof(ENTROPY_CONTEXT_PLANES));
Initial WebM release
2010-05-18 17:58:33 +02:00
}
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
} /*for each 4x4 mode*/
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
vpx_memcpy(ta, ta_b, sizeof(ENTROPY_CONTEXT_PLANES));
vpx_memcpy(tl, tl_b, sizeof(ENTROPY_CONTEXT_PLANES));
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
labels2mode(x, labels, i, mode_selected, &mode_mv[mode_selected],
bsi->ref_mv, x->mvcost);
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
br += sbr;
bd += sbd;
segmentyrate += bestlabelyrate;
this_segment_rd += best_label_rd;
Initial WebM release
2010-05-18 17:58:33 +02:00
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
if (this_segment_rd >= bsi->segment_rd)
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
break;
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
} /* for each label */
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
if (this_segment_rd < bsi->segment_rd)
{
bsi->r = br;
bsi->d = bd;
bsi->segment_yrate = segmentyrate;
bsi->segment_rd = this_segment_rd;
bsi->segment_num = segmentation;
// store everything needed to come back to this!!
for (i = 0; i < 16; i++)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
BLOCKD *bd = &x->e_mbd.block[i];
Initial WebM release
2010-05-18 17:58:33 +02:00
Using partition_info instead of blockd info for splitmv The partition_info struct contains info just for SPLITMV, so it should be used instead of BLOCKD. Eventually, I want to reduce the size of B_MODE_INFO struct found in BLOCKD, so this is the first step toward that goal. Also, since SPLITMV is not supported in vp8_pick_inter_mode(), the unnecessary mem copies and checks were removed. For rt encodes, this gave a slight performance improvement. Change-Id: I5585c98fa9d5acbde1c7e0f452a01d9ecc080574
2011-05-19 21:03:36 +02:00
bsi->mvs[i].as_mv = x->partition_info->bmi[i].mv.as_mv;
bsi->modes[i] = x->partition_info->bmi[i].mode;
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
bsi->eobs[i] = bd->eob;
Initial WebM release
2010-05-18 17:58:33 +02:00
}
}
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
}
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
static __inline
void vp8_cal_step_param(int sr, int *sp)
{
int step = 0;
if (sr > MAX_FIRST_STEP) sr = MAX_FIRST_STEP;
else if (sr < 1) sr = 1;
while (sr>>=1)
step++;
*sp = MAX_MVSEARCH_STEPS - 1 - step;
}
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
static int vp8_rd_pick_best_mbsegmentation(VP8_COMP *cpi, MACROBLOCK *x,
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
int_mv *best_ref_mv, int best_rd,
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
int *mdcounts, int *returntotrate,
int *returnyrate, int *returndistortion,
int mvthresh)
{
int i;
BEST_SEG_INFO bsi;
vpx_memset(&bsi, 0, sizeof(bsi));
bsi.segment_rd = best_rd;
bsi.ref_mv = best_ref_mv;
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
bsi.mvp.as_int = best_ref_mv->as_int;
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
bsi.mvthresh = mvthresh;
bsi.mdcounts = mdcounts;
for(i = 0; i < 16; i++)
{
bsi.modes[i] = ZERO4X4;
}
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
Changed segmentation check order In SPLITMV, the 8x8 segment will be checked first. If the 8x8 rd is better than the best, we check the other segments. Otherwise bail. Adjustments to the thresh_mult were necessary to make up for the initial quality loss. The performance improved by 20% (average) for good quality, speed 0 and speed 1, while the overall quality remained the same. Change-Id: I717aef401323c8a254fba3e9777d2a316c774cc3
2010-12-16 23:01:27 +01:00
if(cpi->compressor_speed == 0)
{
/* for now, we will keep the original segmentation order
when in best quality mode */
Increase static linkage, remove unused functions A large number of functions were defined with external linkage, even though they were only used from within one file. This patch changes their linkage to static and removes the vp8_ prefix from their names, which should make it more obvious to the reader that the function is contained within the current translation unit. Functions that were not referenced were removed. These symbols were identified by: $ nm -A libvpx.a | sort -k3 | uniq -c -f2 | grep ' [A-Z] ' \ | sort | grep '^ *1 ' Change-Id: I59609f58ab65312012c047036ae1e0634f795779
2011-03-17 22:07:59 +01:00
rd_check_segment(cpi, x, &bsi, BLOCK_16X8);
rd_check_segment(cpi, x, &bsi, BLOCK_8X16);
rd_check_segment(cpi, x, &bsi, BLOCK_8X8);
rd_check_segment(cpi, x, &bsi, BLOCK_4X4);
Changed segmentation check order In SPLITMV, the 8x8 segment will be checked first. If the 8x8 rd is better than the best, we check the other segments. Otherwise bail. Adjustments to the thresh_mult were necessary to make up for the initial quality loss. The performance improved by 20% (average) for good quality, speed 0 and speed 1, while the overall quality remained the same. Change-Id: I717aef401323c8a254fba3e9777d2a316c774cc3
2010-12-16 23:01:27 +01:00
}
else
{
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
int sr;
Increase static linkage, remove unused functions A large number of functions were defined with external linkage, even though they were only used from within one file. This patch changes their linkage to static and removes the vp8_ prefix from their names, which should make it more obvious to the reader that the function is contained within the current translation unit. Functions that were not referenced were removed. These symbols were identified by: $ nm -A libvpx.a | sort -k3 | uniq -c -f2 | grep ' [A-Z] ' \ | sort | grep '^ *1 ' Change-Id: I59609f58ab65312012c047036ae1e0634f795779
2011-03-17 22:07:59 +01:00
rd_check_segment(cpi, x, &bsi, BLOCK_8X8);
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
Changed segmentation check order In SPLITMV, the 8x8 segment will be checked first. If the 8x8 rd is better than the best, we check the other segments. Otherwise bail. Adjustments to the thresh_mult were necessary to make up for the initial quality loss. The performance improved by 20% (average) for good quality, speed 0 and speed 1, while the overall quality remained the same. Change-Id: I717aef401323c8a254fba3e9777d2a316c774cc3
2010-12-16 23:01:27 +01:00
if (bsi.segment_rd < best_rd)
{
Adjust full-pixel clamping and motion vector limit calculation Do mvp clamping in full-pixel precision instead of 1/8-pixel precision to avoid error caused by right shifting operation. Also, further fixed the motion vector limit calculation in change: b7480454706a6b15bf091e659cd6227ab373c1a6 Change-Id: Ied88a4f7ddfb0476eb9f7afc6ceeddbf209fffd7
2011-07-07 17:21:41 +02:00
int col_min = (best_ref_mv->as_mv.col>>3) - MAX_FULL_PEL_VAL + ((best_ref_mv->as_mv.col & 7)?1:0);
int row_min = (best_ref_mv->as_mv.row>>3) - MAX_FULL_PEL_VAL + ((best_ref_mv->as_mv.row & 7)?1:0);
int col_max = (best_ref_mv->as_mv.col>>3) + MAX_FULL_PEL_VAL;
int row_max = (best_ref_mv->as_mv.row>>3) + MAX_FULL_PEL_VAL;
Adjust MV borders for SPLITMV mode Add limits to avoid MV going out of range. Change-Id: I8a5deb40bf393488d29f694b5a56804d578e68b5
2010-12-28 19:23:07 +01:00
int tmp_col_min = x->mv_col_min;
int tmp_col_max = x->mv_col_max;
int tmp_row_min = x->mv_row_min;
int tmp_row_max = x->mv_row_max;
/* Get intersection of UMV window and valid MV window to reduce # of checks in diamond search. */
if (x->mv_col_min < col_min )
x->mv_col_min = col_min;
if (x->mv_col_max > col_max )
x->mv_col_max = col_max;
if (x->mv_row_min < row_min )
x->mv_row_min = row_min;
if (x->mv_row_max > row_max )
x->mv_row_max = row_max;
/* Get 8x8 result */
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
bsi.sv_mvp[0].as_int = bsi.mvs[0].as_int;
bsi.sv_mvp[1].as_int = bsi.mvs[2].as_int;
bsi.sv_mvp[2].as_int = bsi.mvs[8].as_int;
bsi.sv_mvp[3].as_int = bsi.mvs[10].as_int;
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
Adjust MV borders for SPLITMV mode Add limits to avoid MV going out of range. Change-Id: I8a5deb40bf393488d29f694b5a56804d578e68b5
2010-12-28 19:23:07 +01:00
/* Use 8x8 result as 16x8/8x16's predictor MV. Adjust search range according to the closeness of 2 MV. */
/* block 8X16 */
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
{
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
sr = MAXF((abs(bsi.sv_mvp[0].as_mv.row - bsi.sv_mvp[2].as_mv.row))>>3, (abs(bsi.sv_mvp[0].as_mv.col - bsi.sv_mvp[2].as_mv.col))>>3);
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
vp8_cal_step_param(sr, &bsi.sv_istep[0]);
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
sr = MAXF((abs(bsi.sv_mvp[1].as_mv.row - bsi.sv_mvp[3].as_mv.row))>>3, (abs(bsi.sv_mvp[1].as_mv.col - bsi.sv_mvp[3].as_mv.col))>>3);
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
vp8_cal_step_param(sr, &bsi.sv_istep[1]);
Increase static linkage, remove unused functions A large number of functions were defined with external linkage, even though they were only used from within one file. This patch changes their linkage to static and removes the vp8_ prefix from their names, which should make it more obvious to the reader that the function is contained within the current translation unit. Functions that were not referenced were removed. These symbols were identified by: $ nm -A libvpx.a | sort -k3 | uniq -c -f2 | grep ' [A-Z] ' \ | sort | grep '^ *1 ' Change-Id: I59609f58ab65312012c047036ae1e0634f795779
2011-03-17 22:07:59 +01:00
rd_check_segment(cpi, x, &bsi, BLOCK_8X16);
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
}
Adjust MV borders for SPLITMV mode Add limits to avoid MV going out of range. Change-Id: I8a5deb40bf393488d29f694b5a56804d578e68b5
2010-12-28 19:23:07 +01:00
/* block 16X8 */
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
{
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
sr = MAXF((abs(bsi.sv_mvp[0].as_mv.row - bsi.sv_mvp[1].as_mv.row))>>3, (abs(bsi.sv_mvp[0].as_mv.col - bsi.sv_mvp[1].as_mv.col))>>3);
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
vp8_cal_step_param(sr, &bsi.sv_istep[0]);
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
sr = MAXF((abs(bsi.sv_mvp[2].as_mv.row - bsi.sv_mvp[3].as_mv.row))>>3, (abs(bsi.sv_mvp[2].as_mv.col - bsi.sv_mvp[3].as_mv.col))>>3);
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
vp8_cal_step_param(sr, &bsi.sv_istep[1]);
Increase static linkage, remove unused functions A large number of functions were defined with external linkage, even though they were only used from within one file. This patch changes their linkage to static and removes the vp8_ prefix from their names, which should make it more obvious to the reader that the function is contained within the current translation unit. Functions that were not referenced were removed. These symbols were identified by: $ nm -A libvpx.a | sort -k3 | uniq -c -f2 | grep ' [A-Z] ' \ | sort | grep '^ *1 ' Change-Id: I59609f58ab65312012c047036ae1e0634f795779
2011-03-17 22:07:59 +01:00
rd_check_segment(cpi, x, &bsi, BLOCK_16X8);
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
}
Adjust MV borders for SPLITMV mode Add limits to avoid MV going out of range. Change-Id: I8a5deb40bf393488d29f694b5a56804d578e68b5
2010-12-28 19:23:07 +01:00
/* If 8x8 is better than 16x8/8x16, then do 4x4 search */
Add no_skip_block4x4_search flag in SPLITMV mode Add a flag to always enable block4x4 search for speed=0 (good quality) to guarantee no quality loss for speed0. Change-Id: Ie04bbc25f7e6a33a7bfa30e05775d33148731c81
2011-01-10 23:21:55 +01:00
/* Not skip 4x4 if speed=0 (good quality) */
if (cpi->sf.no_skip_block4x4_search || bsi.segment_num == BLOCK_8X8) /* || (sv_segment_rd8x8-bsi.segment_rd) < sv_segment_rd8x8>>5) */
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
{
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
bsi.mvp.as_int = bsi.sv_mvp[0].as_int;
Increase static linkage, remove unused functions A large number of functions were defined with external linkage, even though they were only used from within one file. This patch changes their linkage to static and removes the vp8_ prefix from their names, which should make it more obvious to the reader that the function is contained within the current translation unit. Functions that were not referenced were removed. These symbols were identified by: $ nm -A libvpx.a | sort -k3 | uniq -c -f2 | grep ' [A-Z] ' \ | sort | grep '^ *1 ' Change-Id: I59609f58ab65312012c047036ae1e0634f795779
2011-03-17 22:07:59 +01:00
rd_check_segment(cpi, x, &bsi, BLOCK_4X4);
Modify motion estimation for SPLITMV mode 1. Search for block8x16/block16x8 uses block8x8's search results. 2. Check block4x4 only if block8x8 is chosen. (This hurts quality, which will be improved in another check-in.) 3. In block4x4 search, the previous block's result is used as MV predictor for next block. This change improves performance. Change-Id: I9dc089007ca08129fb6c11fe7692777ebb8647b0
2010-12-23 17:23:03 +01:00
}
Adjust MV borders for SPLITMV mode Add limits to avoid MV going out of range. Change-Id: I8a5deb40bf393488d29f694b5a56804d578e68b5
2010-12-28 19:23:07 +01:00
/* restore UMV window */
x->mv_col_min = tmp_col_min;
x->mv_col_max = tmp_col_max;
x->mv_row_min = tmp_row_min;
x->mv_row_max = tmp_row_max;
Changed segmentation check order In SPLITMV, the 8x8 segment will be checked first. If the 8x8 rd is better than the best, we check the other segments. Otherwise bail. Adjustments to the thresh_mult were necessary to make up for the initial quality loss. The performance improved by 20% (average) for good quality, speed 0 and speed 1, while the overall quality remained the same. Change-Id: I717aef401323c8a254fba3e9777d2a316c774cc3
2010-12-16 23:01:27 +01:00
}
}
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
/* set it to the best */
Initial WebM release
2010-05-18 17:58:33 +02:00
for (i = 0; i < 16; i++)
{
BLOCKD *bd = &x->e_mbd.block[i];
Removed B_MODE_INFO Declared the bmi in BLOCKD as a union instead of B_MODE_INFO. Then removed B_MODE_INFO completely. Change-Id: Ieb7469899e265892c66f7aeac87b7f2bf38e7a67
2011-06-02 19:46:41 +02:00
bd->bmi.mv.as_int = bsi.mvs[i].as_int;
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
bd->eob = bsi.eobs[i];
Initial WebM release
2010-05-18 17:58:33 +02:00
}
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
*returntotrate = bsi.r;
*returndistortion = bsi.d;
*returnyrate = bsi.segment_yrate;
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
/* save partitions */
x->e_mbd.mode_info_context->mbmi.partitioning = bsi.segment_num;
x->partition_info->count = vp8_mbsplit_count[bsi.segment_num];
Initial WebM release
2010-05-18 17:58:33 +02:00
Reduced the size of MB_MODE_INFO Moved partition_bmi and partition_count out of MB_MODE_INFO and placed into MACROBLOCK. Also reduced the size of other members of the MB_MODE_INFO struct. For 1080p, the memory was reduced by 1,209,516 bytes. The decoder performance appeared to improve by 3% for the clip used. Note: The main goal for this change is to improve the decoder performance. The encoder will be revisited at a later date for further structure cleanup. Change-Id: I4733621292ee9cc3fffa4046cb3fd4d99bd14613
2010-09-02 22:17:52 +02:00
for (i = 0; i < x->partition_info->count; i++)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
int j;
Increase static linkage, remove unused functions A large number of functions were defined with external linkage, even though they were only used from within one file. This patch changes their linkage to static and removes the vp8_ prefix from their names, which should make it more obvious to the reader that the function is contained within the current translation unit. Functions that were not referenced were removed. These symbols were identified by: $ nm -A libvpx.a | sort -k3 | uniq -c -f2 | grep ' [A-Z] ' \ | sort | grep '^ *1 ' Change-Id: I59609f58ab65312012c047036ae1e0634f795779
2011-03-17 22:07:59 +01:00
j = vp8_mbsplit_offset[bsi.segment_num][i];
Initial WebM release
2010-05-18 17:58:33 +02:00
Using partition_info instead of blockd info for splitmv The partition_info struct contains info just for SPLITMV, so it should be used instead of BLOCKD. Eventually, I want to reduce the size of B_MODE_INFO struct found in BLOCKD, so this is the first step toward that goal. Also, since SPLITMV is not supported in vp8_pick_inter_mode(), the unnecessary mem copies and checks were removed. For rt encodes, this gave a slight performance improvement. Change-Id: I5585c98fa9d5acbde1c7e0f452a01d9ecc080574
2011-05-19 21:03:36 +02:00
x->partition_info->bmi[i].mode = bsi.modes[j];
x->partition_info->bmi[i].mv.as_mv = bsi.mvs[j].as_mv;
Initial WebM release
2010-05-18 17:58:33 +02:00
}
Using partition_info instead of blockd info for splitmv The partition_info struct contains info just for SPLITMV, so it should be used instead of BLOCKD. Eventually, I want to reduce the size of B_MODE_INFO struct found in BLOCKD, so this is the first step toward that goal. Also, since SPLITMV is not supported in vp8_pick_inter_mode(), the unnecessary mem copies and checks were removed. For rt encodes, this gave a slight performance improvement. Change-Id: I5585c98fa9d5acbde1c7e0f452a01d9ecc080574
2011-05-19 21:03:36 +02:00
/*
* used to set x->e_mbd.mode_info_context->mbmi.mv.as_int
*/
x->partition_info->bmi[15].mv.as_int = bsi.mvs[15].as_int;
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
return bsi.segment_rd;
Initial WebM release
2010-05-18 17:58:33 +02:00
}
Use insertion sort instead of quick sort Insertion sort performs better for sorting small arrays. In real- time encoding (speed=-5), test on test set showed 1.7% performance gain with 0% PSNR change in average. Change-Id: Ie02eaa6fed662866a937299194c590d41b25bc3d
2011-04-27 19:40:39 +02:00
static void insertsortmv(int arr[], int len)
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
{
Use insertion sort instead of quick sort Insertion sort performs better for sorting small arrays. In real- time encoding (speed=-5), test on test set showed 1.7% performance gain with 0% PSNR change in average. Change-Id: Ie02eaa6fed662866a937299194c590d41b25bc3d
2011-04-27 19:40:39 +02:00
int i, j, k;
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
Use insertion sort instead of quick sort Insertion sort performs better for sorting small arrays. In real- time encoding (speed=-5), test on test set showed 1.7% performance gain with 0% PSNR change in average. Change-Id: Ie02eaa6fed662866a937299194c590d41b25bc3d
2011-04-27 19:40:39 +02:00
for ( i = 1 ; i <= len-1 ; i++ )
{
for ( j = 0 ; j < i ; j++ )
{
if ( arr[j] > arr[i] )
{
int temp;
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
Use insertion sort instead of quick sort Insertion sort performs better for sorting small arrays. In real- time encoding (speed=-5), test on test set showed 1.7% performance gain with 0% PSNR change in average. Change-Id: Ie02eaa6fed662866a937299194c590d41b25bc3d
2011-04-27 19:40:39 +02:00
temp = arr[i];
for ( k = i; k >j; k--)
arr[k] = arr[k - 1] ;
arr[j] = temp ;
}
}
}
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
}
Use insertion sort instead of quick sort Insertion sort performs better for sorting small arrays. In real- time encoding (speed=-5), test on test set showed 1.7% performance gain with 0% PSNR change in average. Change-Id: Ie02eaa6fed662866a937299194c590d41b25bc3d
2011-04-27 19:40:39 +02:00
static void insertsortsad(int arr[],int idx[], int len)
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
{
Use insertion sort instead of quick sort Insertion sort performs better for sorting small arrays. In real- time encoding (speed=-5), test on test set showed 1.7% performance gain with 0% PSNR change in average. Change-Id: Ie02eaa6fed662866a937299194c590d41b25bc3d
2011-04-27 19:40:39 +02:00
int i, j, k;
for ( i = 1 ; i <= len-1 ; i++ )
{
for ( j = 0 ; j < i ; j++ )
{
if ( arr[j] > arr[i] )
{
int temp, tempi;
temp = arr[i];
tempi = idx[i];
for ( k = i; k >j; k--)
{
arr[k] = arr[k - 1] ;
idx[k] = idx[k - 1];
}
arr[j] = temp ;
idx[j] = tempi;
}
}
}
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
}
//The improved MV prediction
Remove copies of same functions Reduce the code size. Change-Id: I2e1998557a3c8776e262c442fd758c25e17aff7a
2011-01-26 18:03:13 +01:00
void vp8_mv_pred
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
(
VP8_COMP *cpi,
MACROBLOCKD *xd,
const MODE_INFO *here,
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
int_mv *mvp,
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
int refframe,
int *ref_frame_sign_bias,
int *sr,
int near_sadidx[]
)
{
const MODE_INFO *above = here - xd->mode_info_stride;
const MODE_INFO *left = here - 1;
const MODE_INFO *aboveleft = above - 1;
Refine motion vector prediction for NEWMV mode Adjust checking points in motion vector prediction to better cover possible movements, and get a better prediction. Tests on test clips showed a 0.1% improvement in SSIM, and no change in PSNR and performance. Change-Id: Ifdab05d35e10faea1445c61bb73debf888c9d2f8
2011-01-25 21:54:34 +01:00
int_mv near_mvs[8];
int near_ref[8];
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
int_mv mv;
int vcnt=0;
int find=0;
int mb_offset;
Refine motion vector prediction for NEWMV mode Adjust checking points in motion vector prediction to better cover possible movements, and get a better prediction. Tests on test clips showed a 0.1% improvement in SSIM, and no change in PSNR and performance. Change-Id: Ifdab05d35e10faea1445c61bb73debf888c9d2f8
2011-01-25 21:54:34 +01:00
int mvx[8];
int mvy[8];
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
int i;
mv.as_int = 0;
if(here->mbmi.ref_frame != INTRA_FRAME)
{
Refine motion vector prediction for NEWMV mode Adjust checking points in motion vector prediction to better cover possible movements, and get a better prediction. Tests on test clips showed a 0.1% improvement in SSIM, and no change in PSNR and performance. Change-Id: Ifdab05d35e10faea1445c61bb73debf888c9d2f8
2011-01-25 21:54:34 +01:00
near_mvs[0].as_int = near_mvs[1].as_int = near_mvs[2].as_int = near_mvs[3].as_int = near_mvs[4].as_int = near_mvs[5].as_int = near_mvs[6].as_int = near_mvs[7].as_int = 0;
near_ref[0] = near_ref[1] = near_ref[2] = near_ref[3] = near_ref[4] = near_ref[5] = near_ref[6] = near_ref[7] = 0;
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
// read in 3 nearby block's MVs from current frame as prediction candidates.
if (above->mbmi.ref_frame != INTRA_FRAME)
{
near_mvs[vcnt].as_int = above->mbmi.mv.as_int;
Remove copies of same functions Reduce the code size. Change-Id: I2e1998557a3c8776e262c442fd758c25e17aff7a
2011-01-26 18:03:13 +01:00
mv_bias(ref_frame_sign_bias[above->mbmi.ref_frame], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
near_ref[vcnt] = above->mbmi.ref_frame;
}
vcnt++;
if (left->mbmi.ref_frame != INTRA_FRAME)
{
near_mvs[vcnt].as_int = left->mbmi.mv.as_int;
Remove copies of same functions Reduce the code size. Change-Id: I2e1998557a3c8776e262c442fd758c25e17aff7a
2011-01-26 18:03:13 +01:00
mv_bias(ref_frame_sign_bias[left->mbmi.ref_frame], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
near_ref[vcnt] = left->mbmi.ref_frame;
}
vcnt++;
if (aboveleft->mbmi.ref_frame != INTRA_FRAME)
{
near_mvs[vcnt].as_int = aboveleft->mbmi.mv.as_int;
Remove copies of same functions Reduce the code size. Change-Id: I2e1998557a3c8776e262c442fd758c25e17aff7a
2011-01-26 18:03:13 +01:00
mv_bias(ref_frame_sign_bias[aboveleft->mbmi.ref_frame], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
near_ref[vcnt] = aboveleft->mbmi.ref_frame;
}
vcnt++;
Refine motion vector prediction for NEWMV mode Adjust checking points in motion vector prediction to better cover possible movements, and get a better prediction. Tests on test clips showed a 0.1% improvement in SSIM, and no change in PSNR and performance. Change-Id: Ifdab05d35e10faea1445c61bb73debf888c9d2f8
2011-01-25 21:54:34 +01:00
// read in 5 nearby block's MVs from last frame.
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
if(cpi->common.last_frame_type != KEY_FRAME)
{
Refine motion vector prediction for NEWMV mode Adjust checking points in motion vector prediction to better cover possible movements, and get a better prediction. Tests on test clips showed a 0.1% improvement in SSIM, and no change in PSNR and performance. Change-Id: Ifdab05d35e10faea1445c61bb73debf888c9d2f8
2011-01-25 21:54:34 +01:00
mb_offset = (-xd->mb_to_top_edge/128 + 1) * (xd->mode_info_stride +1) + (-xd->mb_to_left_edge/128 +1) ;
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
// current in last frame
if (cpi->lf_ref_frame[mb_offset] != INTRA_FRAME)
{
near_mvs[vcnt].as_int = cpi->lfmv[mb_offset].as_int;
Remove copies of same functions Reduce the code size. Change-Id: I2e1998557a3c8776e262c442fd758c25e17aff7a
2011-01-26 18:03:13 +01:00
mv_bias(cpi->lf_ref_frame_sign_bias[mb_offset], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
near_ref[vcnt] = cpi->lf_ref_frame[mb_offset];
}
vcnt++;
// above in last frame
Refine motion vector prediction for NEWMV mode Adjust checking points in motion vector prediction to better cover possible movements, and get a better prediction. Tests on test clips showed a 0.1% improvement in SSIM, and no change in PSNR and performance. Change-Id: Ifdab05d35e10faea1445c61bb73debf888c9d2f8
2011-01-25 21:54:34 +01:00
if (cpi->lf_ref_frame[mb_offset - xd->mode_info_stride-1] != INTRA_FRAME)
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
{
Refine motion vector prediction for NEWMV mode Adjust checking points in motion vector prediction to better cover possible movements, and get a better prediction. Tests on test clips showed a 0.1% improvement in SSIM, and no change in PSNR and performance. Change-Id: Ifdab05d35e10faea1445c61bb73debf888c9d2f8
2011-01-25 21:54:34 +01:00
near_mvs[vcnt].as_int = cpi->lfmv[mb_offset - xd->mode_info_stride-1].as_int;
Remove copies of same functions Reduce the code size. Change-Id: I2e1998557a3c8776e262c442fd758c25e17aff7a
2011-01-26 18:03:13 +01:00
mv_bias(cpi->lf_ref_frame_sign_bias[mb_offset - xd->mode_info_stride-1], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
Refine motion vector prediction for NEWMV mode Adjust checking points in motion vector prediction to better cover possible movements, and get a better prediction. Tests on test clips showed a 0.1% improvement in SSIM, and no change in PSNR and performance. Change-Id: Ifdab05d35e10faea1445c61bb73debf888c9d2f8
2011-01-25 21:54:34 +01:00
near_ref[vcnt] = cpi->lf_ref_frame[mb_offset - xd->mode_info_stride-1];
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
}
vcnt++;
// left in last frame
if (cpi->lf_ref_frame[mb_offset-1] != INTRA_FRAME)
{
near_mvs[vcnt].as_int = cpi->lfmv[mb_offset -1].as_int;
Remove copies of same functions Reduce the code size. Change-Id: I2e1998557a3c8776e262c442fd758c25e17aff7a
2011-01-26 18:03:13 +01:00
mv_bias(cpi->lf_ref_frame_sign_bias[mb_offset -1], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
near_ref[vcnt] = cpi->lf_ref_frame[mb_offset - 1];
}
vcnt++;
Refine motion vector prediction for NEWMV mode Adjust checking points in motion vector prediction to better cover possible movements, and get a better prediction. Tests on test clips showed a 0.1% improvement in SSIM, and no change in PSNR and performance. Change-Id: Ifdab05d35e10faea1445c61bb73debf888c9d2f8
2011-01-25 21:54:34 +01:00
// right in last frame
if (cpi->lf_ref_frame[mb_offset +1] != INTRA_FRAME)
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
{
Refine motion vector prediction for NEWMV mode Adjust checking points in motion vector prediction to better cover possible movements, and get a better prediction. Tests on test clips showed a 0.1% improvement in SSIM, and no change in PSNR and performance. Change-Id: Ifdab05d35e10faea1445c61bb73debf888c9d2f8
2011-01-25 21:54:34 +01:00
near_mvs[vcnt].as_int = cpi->lfmv[mb_offset +1].as_int;
Remove copies of same functions Reduce the code size. Change-Id: I2e1998557a3c8776e262c442fd758c25e17aff7a
2011-01-26 18:03:13 +01:00
mv_bias(cpi->lf_ref_frame_sign_bias[mb_offset +1], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
Refine motion vector prediction for NEWMV mode Adjust checking points in motion vector prediction to better cover possible movements, and get a better prediction. Tests on test clips showed a 0.1% improvement in SSIM, and no change in PSNR and performance. Change-Id: Ifdab05d35e10faea1445c61bb73debf888c9d2f8
2011-01-25 21:54:34 +01:00
near_ref[vcnt] = cpi->lf_ref_frame[mb_offset +1];
}
vcnt++;
// below in last frame
if (cpi->lf_ref_frame[mb_offset + xd->mode_info_stride +1] != INTRA_FRAME)
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
{
Refine motion vector prediction for NEWMV mode Adjust checking points in motion vector prediction to better cover possible movements, and get a better prediction. Tests on test clips showed a 0.1% improvement in SSIM, and no change in PSNR and performance. Change-Id: Ifdab05d35e10faea1445c61bb73debf888c9d2f8
2011-01-25 21:54:34 +01:00
near_mvs[vcnt].as_int = cpi->lfmv[mb_offset + xd->mode_info_stride +1].as_int;
Remove copies of same functions Reduce the code size. Change-Id: I2e1998557a3c8776e262c442fd758c25e17aff7a
2011-01-26 18:03:13 +01:00
mv_bias(cpi->lf_ref_frame_sign_bias[mb_offset + xd->mode_info_stride +1], refframe, &near_mvs[vcnt], ref_frame_sign_bias);
Refine motion vector prediction for NEWMV mode Adjust checking points in motion vector prediction to better cover possible movements, and get a better prediction. Tests on test clips showed a 0.1% improvement in SSIM, and no change in PSNR and performance. Change-Id: Ifdab05d35e10faea1445c61bb73debf888c9d2f8
2011-01-25 21:54:34 +01:00
near_ref[vcnt] = cpi->lf_ref_frame[mb_offset + xd->mode_info_stride +1];
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
}
vcnt++;
}
for(i=0; i< vcnt; i++)
{
if(near_ref[near_sadidx[i]] != INTRA_FRAME)
{
if(here->mbmi.ref_frame == near_ref[near_sadidx[i]])
{
mv.as_int = near_mvs[near_sadidx[i]].as_int;
find = 1;
Refine motion vector prediction for NEWMV mode Adjust checking points in motion vector prediction to better cover possible movements, and get a better prediction. Tests on test clips showed a 0.1% improvement in SSIM, and no change in PSNR and performance. Change-Id: Ifdab05d35e10faea1445c61bb73debf888c9d2f8
2011-01-25 21:54:34 +01:00
if (i < 3)
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
*sr = 3;
else
*sr = 2;
break;
}
}
}
if(!find)
{
for(i=0; i<vcnt; i++)
{
mvx[i] = near_mvs[i].as_mv.row;
mvy[i] = near_mvs[i].as_mv.col;
}
Use insertion sort instead of quick sort Insertion sort performs better for sorting small arrays. In real- time encoding (speed=-5), test on test set showed 1.7% performance gain with 0% PSNR change in average. Change-Id: Ie02eaa6fed662866a937299194c590d41b25bc3d
2011-04-27 19:40:39 +02:00
insertsortmv(mvx, vcnt);
insertsortmv(mvy, vcnt);
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
mv.as_mv.row = mvx[vcnt/2];
mv.as_mv.col = mvy[vcnt/2];
find = 1;
//sr is set to 0 to allow calling function to decide the search range.
*sr = 0;
}
}
/* Set up return values */
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
mvp->as_int = mv.as_int;
vp8_clamp_mv2(mvp, xd);
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
}
Improve MV prediction in vp8_pick_inter_mode() for speed>3 Applied same method used in vp8_rd_pick_inter_mode() to improve the accuracy of MV prediction. Change-Id: Ia50ae26208b18482695601f32febd99fe89fbc17
2011-01-28 16:00:20 +01:00
void vp8_cal_sad(VP8_COMP *cpi, MACROBLOCKD *xd, MACROBLOCK *x, int recon_yoffset, int near_sadidx[])
{
int near_sad[8] = {0}; // 0-cf above, 1-cf left, 2-cf aboveleft, 3-lf current, 4-lf above, 5-lf left, 6-lf right, 7-lf below
Copy macroblock data to a buffer before encoding it I got this idea from Pascal (Thanks). Before encoding a macroblock, copy it to a 16x16 buffer, and then read source data from there instead. This will help keep the source data in cache, and help with the performance. Change-Id: Id05f4cb601299150511d59dcba0ae62c49b5b757
2011-06-23 19:54:02 +02:00
BLOCK *b = &x->block[0];
unsigned char *src_y_ptr = *(b->base_src);
Improve MV prediction in vp8_pick_inter_mode() for speed>3 Applied same method used in vp8_rd_pick_inter_mode() to improve the accuracy of MV prediction. Change-Id: Ia50ae26208b18482695601f32febd99fe89fbc17
2011-01-28 16:00:20 +01:00
//calculate sad for current frame 3 nearby MBs.
if( xd->mb_to_top_edge==0 && xd->mb_to_left_edge ==0)
{
near_sad[0] = near_sad[1] = near_sad[2] = INT_MAX;
}else if(xd->mb_to_top_edge==0)
{ //only has left MB for sad calculation.
near_sad[0] = near_sad[2] = INT_MAX;
Copy macroblock data to a buffer before encoding it I got this idea from Pascal (Thanks). Before encoding a macroblock, copy it to a 16x16 buffer, and then read source data from there instead. This will help keep the source data in cache, and help with the performance. Change-Id: Id05f4cb601299150511d59dcba0ae62c49b5b757
2011-06-23 19:54:02 +02:00
near_sad[1] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, xd->dst.y_buffer - 16,xd->dst.y_stride, 0x7fffffff);
Improve MV prediction in vp8_pick_inter_mode() for speed>3 Applied same method used in vp8_rd_pick_inter_mode() to improve the accuracy of MV prediction. Change-Id: Ia50ae26208b18482695601f32febd99fe89fbc17
2011-01-28 16:00:20 +01:00
}else if(xd->mb_to_left_edge ==0)
{ //only has left MB for sad calculation.
near_sad[1] = near_sad[2] = INT_MAX;
Copy macroblock data to a buffer before encoding it I got this idea from Pascal (Thanks). Before encoding a macroblock, copy it to a 16x16 buffer, and then read source data from there instead. This will help keep the source data in cache, and help with the performance. Change-Id: Id05f4cb601299150511d59dcba0ae62c49b5b757
2011-06-23 19:54:02 +02:00
near_sad[0] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, xd->dst.y_buffer - xd->dst.y_stride *16,xd->dst.y_stride, 0x7fffffff);
Improve MV prediction in vp8_pick_inter_mode() for speed>3 Applied same method used in vp8_rd_pick_inter_mode() to improve the accuracy of MV prediction. Change-Id: Ia50ae26208b18482695601f32febd99fe89fbc17
2011-01-28 16:00:20 +01:00
}else
{
Copy macroblock data to a buffer before encoding it I got this idea from Pascal (Thanks). Before encoding a macroblock, copy it to a 16x16 buffer, and then read source data from there instead. This will help keep the source data in cache, and help with the performance. Change-Id: Id05f4cb601299150511d59dcba0ae62c49b5b757
2011-06-23 19:54:02 +02:00
near_sad[0] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, xd->dst.y_buffer - xd->dst.y_stride *16,xd->dst.y_stride, 0x7fffffff);
near_sad[1] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, xd->dst.y_buffer - 16,xd->dst.y_stride, 0x7fffffff);
near_sad[2] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, xd->dst.y_buffer - xd->dst.y_stride *16 -16,xd->dst.y_stride, 0x7fffffff);
Improve MV prediction in vp8_pick_inter_mode() for speed>3 Applied same method used in vp8_rd_pick_inter_mode() to improve the accuracy of MV prediction. Change-Id: Ia50ae26208b18482695601f32febd99fe89fbc17
2011-01-28 16:00:20 +01:00
}
if(cpi->common.last_frame_type != KEY_FRAME)
{
//calculate sad for last frame 5 nearby MBs.
unsigned char *pre_y_buffer = cpi->common.yv12_fb[cpi->common.lst_fb_idx].y_buffer + recon_yoffset;
int pre_y_stride = cpi->common.yv12_fb[cpi->common.lst_fb_idx].y_stride;
if(xd->mb_to_top_edge==0) near_sad[4] = INT_MAX;
if(xd->mb_to_left_edge ==0) near_sad[5] = INT_MAX;
if(xd->mb_to_right_edge ==0) near_sad[6] = INT_MAX;
if(xd->mb_to_bottom_edge==0) near_sad[7] = INT_MAX;
if(near_sad[4] != INT_MAX)
Copy macroblock data to a buffer before encoding it I got this idea from Pascal (Thanks). Before encoding a macroblock, copy it to a 16x16 buffer, and then read source data from there instead. This will help keep the source data in cache, and help with the performance. Change-Id: Id05f4cb601299150511d59dcba0ae62c49b5b757
2011-06-23 19:54:02 +02:00
near_sad[4] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, pre_y_buffer - pre_y_stride *16, pre_y_stride, 0x7fffffff);
Improve MV prediction in vp8_pick_inter_mode() for speed>3 Applied same method used in vp8_rd_pick_inter_mode() to improve the accuracy of MV prediction. Change-Id: Ia50ae26208b18482695601f32febd99fe89fbc17
2011-01-28 16:00:20 +01:00
if(near_sad[5] != INT_MAX)
Copy macroblock data to a buffer before encoding it I got this idea from Pascal (Thanks). Before encoding a macroblock, copy it to a 16x16 buffer, and then read source data from there instead. This will help keep the source data in cache, and help with the performance. Change-Id: Id05f4cb601299150511d59dcba0ae62c49b5b757
2011-06-23 19:54:02 +02:00
near_sad[5] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, pre_y_buffer - 16, pre_y_stride, 0x7fffffff);
near_sad[3] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, pre_y_buffer, pre_y_stride, 0x7fffffff);
Improve MV prediction in vp8_pick_inter_mode() for speed>3 Applied same method used in vp8_rd_pick_inter_mode() to improve the accuracy of MV prediction. Change-Id: Ia50ae26208b18482695601f32febd99fe89fbc17
2011-01-28 16:00:20 +01:00
if(near_sad[6] != INT_MAX)
Copy macroblock data to a buffer before encoding it I got this idea from Pascal (Thanks). Before encoding a macroblock, copy it to a 16x16 buffer, and then read source data from there instead. This will help keep the source data in cache, and help with the performance. Change-Id: Id05f4cb601299150511d59dcba0ae62c49b5b757
2011-06-23 19:54:02 +02:00
near_sad[6] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, pre_y_buffer + 16, pre_y_stride, 0x7fffffff);
Improve MV prediction in vp8_pick_inter_mode() for speed>3 Applied same method used in vp8_rd_pick_inter_mode() to improve the accuracy of MV prediction. Change-Id: Ia50ae26208b18482695601f32febd99fe89fbc17
2011-01-28 16:00:20 +01:00
if(near_sad[7] != INT_MAX)
Copy macroblock data to a buffer before encoding it I got this idea from Pascal (Thanks). Before encoding a macroblock, copy it to a 16x16 buffer, and then read source data from there instead. This will help keep the source data in cache, and help with the performance. Change-Id: Id05f4cb601299150511d59dcba0ae62c49b5b757
2011-06-23 19:54:02 +02:00
near_sad[7] = cpi->fn_ptr[BLOCK_16X16].sdf(src_y_ptr, b->src_stride, pre_y_buffer + pre_y_stride *16, pre_y_stride, 0x7fffffff);
Improve MV prediction in vp8_pick_inter_mode() for speed>3 Applied same method used in vp8_rd_pick_inter_mode() to improve the accuracy of MV prediction. Change-Id: Ia50ae26208b18482695601f32febd99fe89fbc17
2011-01-28 16:00:20 +01:00
}
if(cpi->common.last_frame_type != KEY_FRAME)
{
Use insertion sort instead of quick sort Insertion sort performs better for sorting small arrays. In real- time encoding (speed=-5), test on test set showed 1.7% performance gain with 0% PSNR change in average. Change-Id: Ie02eaa6fed662866a937299194c590d41b25bc3d
2011-04-27 19:40:39 +02:00
insertsortsad(near_sad, near_sadidx, 8);
Improve MV prediction in vp8_pick_inter_mode() for speed>3 Applied same method used in vp8_rd_pick_inter_mode() to improve the accuracy of MV prediction. Change-Id: Ia50ae26208b18482695601f32febd99fe89fbc17
2011-01-28 16:00:20 +01:00
}else
{
Use insertion sort instead of quick sort Insertion sort performs better for sorting small arrays. In real- time encoding (speed=-5), test on test set showed 1.7% performance gain with 0% PSNR change in average. Change-Id: Ie02eaa6fed662866a937299194c590d41b25bc3d
2011-04-27 19:40:39 +02:00
insertsortsad(near_sad, near_sadidx, 3);
Improve MV prediction in vp8_pick_inter_mode() for speed>3 Applied same method used in vp8_rd_pick_inter_mode() to improve the accuracy of MV prediction. Change-Id: Ia50ae26208b18482695601f32febd99fe89fbc17
2011-01-28 16:00:20 +01:00
}
}
Using partition_info instead of blockd info for splitmv The partition_info struct contains info just for SPLITMV, so it should be used instead of BLOCKD. Eventually, I want to reduce the size of B_MODE_INFO struct found in BLOCKD, so this is the first step toward that goal. Also, since SPLITMV is not supported in vp8_pick_inter_mode(), the unnecessary mem copies and checks were removed. For rt encodes, this gave a slight performance improvement. Change-Id: I5585c98fa9d5acbde1c7e0f452a01d9ecc080574
2011-05-19 21:03:36 +02:00
static void rd_update_mvcount(VP8_COMP *cpi, MACROBLOCK *x, int_mv *best_ref_mv)
Modify MVcount in pick_inter_mode to eliminate calling of vp8_find_near_mvs Moved MVcount modification in pick_inter_mode, and eliminated calling of vp8_find_near_mvs. Change-Id: Icd47448a1dfc8fdf526f86757d0e5a7f218cb5e8
2011-05-13 16:56:45 +02:00
{
Using partition_info instead of blockd info for splitmv The partition_info struct contains info just for SPLITMV, so it should be used instead of BLOCKD. Eventually, I want to reduce the size of B_MODE_INFO struct found in BLOCKD, so this is the first step toward that goal. Also, since SPLITMV is not supported in vp8_pick_inter_mode(), the unnecessary mem copies and checks were removed. For rt encodes, this gave a slight performance improvement. Change-Id: I5585c98fa9d5acbde1c7e0f452a01d9ecc080574
2011-05-19 21:03:36 +02:00
if (x->e_mbd.mode_info_context->mbmi.mode == SPLITMV)
Modify MVcount in pick_inter_mode to eliminate calling of vp8_find_near_mvs Moved MVcount modification in pick_inter_mode, and eliminated calling of vp8_find_near_mvs. Change-Id: Icd47448a1dfc8fdf526f86757d0e5a7f218cb5e8
2011-05-13 16:56:45 +02:00
{
Using partition_info instead of blockd info for splitmv The partition_info struct contains info just for SPLITMV, so it should be used instead of BLOCKD. Eventually, I want to reduce the size of B_MODE_INFO struct found in BLOCKD, so this is the first step toward that goal. Also, since SPLITMV is not supported in vp8_pick_inter_mode(), the unnecessary mem copies and checks were removed. For rt encodes, this gave a slight performance improvement. Change-Id: I5585c98fa9d5acbde1c7e0f452a01d9ecc080574
2011-05-19 21:03:36 +02:00
int i;
for (i = 0; i < x->partition_info->count; i++)
Modify MVcount in pick_inter_mode to eliminate calling of vp8_find_near_mvs Moved MVcount modification in pick_inter_mode, and eliminated calling of vp8_find_near_mvs. Change-Id: Icd47448a1dfc8fdf526f86757d0e5a7f218cb5e8
2011-05-13 16:56:45 +02:00
{
Using partition_info instead of blockd info for splitmv The partition_info struct contains info just for SPLITMV, so it should be used instead of BLOCKD. Eventually, I want to reduce the size of B_MODE_INFO struct found in BLOCKD, so this is the first step toward that goal. Also, since SPLITMV is not supported in vp8_pick_inter_mode(), the unnecessary mem copies and checks were removed. For rt encodes, this gave a slight performance improvement. Change-Id: I5585c98fa9d5acbde1c7e0f452a01d9ecc080574
2011-05-19 21:03:36 +02:00
if (x->partition_info->bmi[i].mode == NEW4X4)
Modify MVcount in pick_inter_mode to eliminate calling of vp8_find_near_mvs Moved MVcount modification in pick_inter_mode, and eliminated calling of vp8_find_near_mvs. Change-Id: Icd47448a1dfc8fdf526f86757d0e5a7f218cb5e8
2011-05-13 16:56:45 +02:00
{
Using partition_info instead of blockd info for splitmv The partition_info struct contains info just for SPLITMV, so it should be used instead of BLOCKD. Eventually, I want to reduce the size of B_MODE_INFO struct found in BLOCKD, so this is the first step toward that goal. Also, since SPLITMV is not supported in vp8_pick_inter_mode(), the unnecessary mem copies and checks were removed. For rt encodes, this gave a slight performance improvement. Change-Id: I5585c98fa9d5acbde1c7e0f452a01d9ecc080574
2011-05-19 21:03:36 +02:00
cpi->MVcount[0][mv_max+((x->partition_info->bmi[i].mv.as_mv.row
- best_ref_mv->as_mv.row) >> 1)]++;
cpi->MVcount[1][mv_max+((x->partition_info->bmi[i].mv.as_mv.col
- best_ref_mv->as_mv.col) >> 1)]++;
Modify MVcount in pick_inter_mode to eliminate calling of vp8_find_near_mvs Moved MVcount modification in pick_inter_mode, and eliminated calling of vp8_find_near_mvs. Change-Id: Icd47448a1dfc8fdf526f86757d0e5a7f218cb5e8
2011-05-13 16:56:45 +02:00
}
}
}
Using partition_info instead of blockd info for splitmv The partition_info struct contains info just for SPLITMV, so it should be used instead of BLOCKD. Eventually, I want to reduce the size of B_MODE_INFO struct found in BLOCKD, so this is the first step toward that goal. Also, since SPLITMV is not supported in vp8_pick_inter_mode(), the unnecessary mem copies and checks were removed. For rt encodes, this gave a slight performance improvement. Change-Id: I5585c98fa9d5acbde1c7e0f452a01d9ecc080574
2011-05-19 21:03:36 +02:00
else if (x->e_mbd.mode_info_context->mbmi.mode == NEWMV)
Modify MVcount in pick_inter_mode to eliminate calling of vp8_find_near_mvs Moved MVcount modification in pick_inter_mode, and eliminated calling of vp8_find_near_mvs. Change-Id: Icd47448a1dfc8fdf526f86757d0e5a7f218cb5e8
2011-05-13 16:56:45 +02:00
{
Using partition_info instead of blockd info for splitmv The partition_info struct contains info just for SPLITMV, so it should be used instead of BLOCKD. Eventually, I want to reduce the size of B_MODE_INFO struct found in BLOCKD, so this is the first step toward that goal. Also, since SPLITMV is not supported in vp8_pick_inter_mode(), the unnecessary mem copies and checks were removed. For rt encodes, this gave a slight performance improvement. Change-Id: I5585c98fa9d5acbde1c7e0f452a01d9ecc080574
2011-05-19 21:03:36 +02:00
cpi->MVcount[0][mv_max+((x->e_mbd.mode_info_context->mbmi.mv.as_mv.row
- best_ref_mv->as_mv.row) >> 1)]++;
cpi->MVcount[1][mv_max+((x->e_mbd.mode_info_context->mbmi.mv.as_mv.col
- best_ref_mv->as_mv.col) >> 1)]++;
Modify MVcount in pick_inter_mode to eliminate calling of vp8_find_near_mvs Moved MVcount modification in pick_inter_mode, and eliminated calling of vp8_find_near_mvs. Change-Id: Icd47448a1dfc8fdf526f86757d0e5a7f218cb5e8
2011-05-13 16:56:45 +02:00
}
}
enabled 8x8 intra prediction modes on inter frames This commit enabled the usage of 8x8 intra prediction modes on inter frames. There are a few TODO items related to this: 1)baseline entropy need be calibrated; 2)cost of UV need to be done more properly rather than using decision only relying on Y; 3)Threshold for allowing picking 8x8 intra prediction should be lowered to lower than the B_PRED. Even with all the TODOs, tests showed consistent gain on derf set ~0.1% (PSNR:0.08% and SSIM:0.14%). It is assumed that 8x8 intra prediction will help more on large resolution clips, especially with above TODOs addressed. Change-Id: I398ada49dfc32575cfab962a569c2885111ae3ba
2011-12-01 01:25:00 +01:00
static void set_i8x8_block_modes(MACROBLOCK *x, int *modes)
{
int i;
MACROBLOCKD *xd = &x->e_mbd;
for(i=0;i<4;i++)
{
int ib = vp8_i8x8_block[i];
x->e_mbd.mode_info_context->bmi[ib+0].as_mode= modes[i];
x->e_mbd.mode_info_context->bmi[ib+1].as_mode= modes[i];
x->e_mbd.mode_info_context->bmi[ib+4].as_mode= modes[i];
x->e_mbd.mode_info_context->bmi[ib+5].as_mode= modes[i];
}
for (i = 0; i < 16; i++)
{
xd->block[i].bmi = xd->mode_info_context->bmi[i];
}
}
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
void vp8_rd_pick_inter_mode(VP8_COMP *cpi, MACROBLOCK *x, int recon_yoffset, int recon_uvoffset,
int *returnrate, int *returndistortion, int *returnintra,
int *best_single_rd_diff, int *best_dual_rd_diff,
int *best_hybrid_rd_diff)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
BLOCK *b = &x->block[0];
BLOCKD *d = &x->e_mbd.block[0];
MACROBLOCKD *xd = &x->e_mbd;
Segment coding of mode and reference frame. Proof of concept test code that encodes mode and reference frame data at the segment level. Decode-able bit stream but some issues not yet resolved. As it this helps a little on a couple of clips but hurts on most as the basis for segmentation is unsound. To build and test, configure with --enable-experimental --enable-segfeatures Change-Id: I22a60774f69273523fb152db8c31f4b10b07c7f4
2011-09-30 17:45:16 +02:00
VP8_COMMON *cm = &cpi->common;
Removed B_MODE_INFO Declared the bmi in BLOCKD as a union instead of B_MODE_INFO. Then removed B_MODE_INFO completely. Change-Id: Ieb7469899e265892c66f7aeac87b7f2bf38e7a67
2011-06-02 19:46:41 +02:00
union b_mode_info best_bmodes[16];
Initial WebM release
2010-05-18 17:58:33 +02:00
MB_MODE_INFO best_mbmode;
Reduced the size of MB_MODE_INFO Moved partition_bmi and partition_count out of MB_MODE_INFO and placed into MACROBLOCK. Also reduced the size of other members of the MB_MODE_INFO struct. For 1080p, the memory was reduced by 1,209,516 bytes. The decoder performance appeared to improve by 3% for the clip used. Note: The main goal for this change is to improve the decoder performance. The encoder will be revisited at a later date for further structure cleanup. Change-Id: I4733621292ee9cc3fffa4046cb3fd4d99bd14613
2010-09-02 22:17:52 +02:00
PARTITION_INFO best_partition;
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
int_mv best_ref_mv;
int_mv mode_mv[MB_MODE_COUNT];
Initial WebM release
2010-05-18 17:58:33 +02:00
MB_PREDICTION_MODE this_mode;
int num00;
int best_mode_index = 0;
enabled 8x8 intra prediction modes on inter frames This commit enabled the usage of 8x8 intra prediction modes on inter frames. There are a few TODO items related to this: 1)baseline entropy need be calibrated; 2)cost of UV need to be done more properly rather than using decision only relying on Y; 3)Threshold for allowing picking 8x8 intra prediction should be lowered to lower than the B_PRED. Even with all the TODOs, tests showed consistent gain on derf set ~0.1% (PSNR:0.08% and SSIM:0.14%). It is assumed that 8x8 intra prediction will help more on large resolution clips, especially with above TODOs addressed. Change-Id: I398ada49dfc32575cfab962a569c2885111ae3ba
2011-12-01 01:25:00 +01:00
int mode8x8[4];
Segmentation feature logic fix. Bug fix to logic in vp8_pick_inter_mode() and vp8_rd_pick_inter_mode(). The block on the use of segment features for the cm->refresh_alt_ref_frame case was just for testing and is not correct. The special case code for alt ref can be re-enabled as an else clause. Change-Id: Ic9b57cdb5f04ea7737032b8fb953d84d7717b3ce
2011-11-03 18:45:27 +01:00
unsigned char segment_id = xd->mode_info_context->mbmi.segment_id;
Initial WebM release
2010-05-18 17:58:33 +02:00
int i;
int mode_index;
int mdcounts[4];
int rate;
int distortion;
Error accumulator stats bug. The error accumulator stats values cpi->prediction_error and cpi->intra_error were being populated with rd values not distortion values. These are only "currently" used in a limited way for RT compress key frame detection. Change-Id: I2702ba1cab6e49ab8dc096ba75b6b34ab3573021
2011-04-08 15:21:36 +02:00
int best_rd = INT_MAX;
int best_intra_rd = INT_MAX;
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
int best_dual_rd = INT_MAX;
int best_single_rd = INT_MAX;
int best_hybrid_rd = INT_MAX;
#endif /* CONFIG_DUALPRED */
Initial WebM release
2010-05-18 17:58:33 +02:00
int rate2, distortion2;
int uv_intra_rate, uv_intra_distortion, uv_intra_rate_tokenonly;
int rate_y, UNINITIALIZED_IS_SAFE(rate_uv);
Adjusted breakout RD for SPLITMV vp8_rd_pick_best_mbsegmentation looks at y only. The new breakout does not include the frame cost, the prob_skip_false cost, or the uv rate. Performance improved by a few percent and the quality remained the same. Change-Id: I94ff013998ac51e8ecce7130870f7b6600758e15
2010-12-16 15:38:02 +01:00
int distortion_uv;
int best_yrd = INT_MAX;
Initial WebM release
2010-05-18 17:58:33 +02:00
//int all_rds[MAX_MODES]; // Experimental debug code.
//int all_rates[MAX_MODES];
//int all_dist[MAX_MODES];
//int intermodecost[MAX_MODES];
MB_PREDICTION_MODE uv_intra_mode;
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
int_mv mvp;
Refine motion vector prediction for NEWMV mode Adjust checking points in motion vector prediction to better cover possible movements, and get a better prediction. Tests on test clips showed a 0.1% improvement in SSIM, and no change in PSNR and performance. Change-Id: Ifdab05d35e10faea1445c61bb73debf888c9d2f8
2011-01-25 21:54:34 +01:00
int near_sadidx[8] = {0, 1, 2, 3, 4, 5, 6, 7};
Improve MV prediction accuracy to achieve performance gain Add vp8_mv_pred() to better predict starting MV for NEWMV mode in vp8_rd_pick_inter_mode(). Set different search ranges according to MV prediction accuracy, which improves encoder performance without hurting the quality. Also, as Yaowu suggested, using diamond search result as full search starting point and therefore adjusting(reducing) full search range helps the performance. Change-Id: Ie4a3c8df87e697c1f4f6e2ddb693766bba1b77b6
2010-12-03 17:26:21 +01:00
int saddone=0;
int sr=0; //search range got from mv_pred(). It uses step_param levels. (0-7)
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
int_mv frame_nearest_mv[4];
int_mv frame_near_mv[4];
int_mv frame_best_ref_mv[4];
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
int_mv mc_search_result[4];
#endif /* CONFIG_DUALPRED */
Moved ref frame calculations Moved ref frame calculations to outside of the mode_index loop. Change-Id: I06103fc7e8af88b54b84443acf6691d29b1272ac
2011-01-11 21:00:00 +01:00
int frame_mdcounts[4][4];
unsigned char *y_buffer[4];
unsigned char *u_buffer[4];
unsigned char *v_buffer[4];
vpx_memset(&best_mbmode, 0, sizeof(best_mbmode));
Removed B_MODE_INFO Declared the bmi in BLOCKD as a union instead of B_MODE_INFO. Then removed B_MODE_INFO completely. Change-Id: Ieb7469899e265892c66f7aeac87b7f2bf38e7a67
2011-06-02 19:46:41 +02:00
vpx_memset(&best_bmodes, 0, sizeof(best_bmodes));
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
for (i = 0; i < 4; i++)
{
#define INVALID_MV 0x80008000
mc_search_result[i].as_int = INVALID_MV;
}
#endif /* CONFIG_DUALPRED */
Moved ref frame calculations Moved ref frame calculations to outside of the mode_index loop. Change-Id: I06103fc7e8af88b54b84443acf6691d29b1272ac
2011-01-11 21:00:00 +01:00
if (cpi->ref_frame_flags & VP8_LAST_FLAG)
{
YV12_BUFFER_CONFIG *lst_yv12 = &cpi->common.yv12_fb[cpi->common.lst_fb_idx];
changed find_near_mvs search to include a mb from last frame This is an experiment to include a mv contribution from last frame to nearest and near mv definition. Initial test showed some small though consistent gain. latest patch slightly better result ~.13%-~.18%. TODO: the entropy used to encode the mode choice, i.e. the mv counts based conditional distribution of modes should be re-collected to reflect this change, it is expected that there is some further gain from that. Change-Id: Ief1e284a36d8aa56b49ae5b360c91419ec494fa4
2011-11-16 01:16:30 +01:00
vp8_find_near_mvs(&x->e_mbd, x->e_mbd.mode_info_context,
#if CONFIG_NEWNEAR
x->e_mbd.prev_mode_info_context,
#endif
&frame_nearest_mv[LAST_FRAME], &frame_near_mv[LAST_FRAME],
&frame_best_ref_mv[LAST_FRAME], frame_mdcounts[LAST_FRAME], LAST_FRAME, cpi->common.ref_frame_sign_bias);
Moved ref frame calculations Moved ref frame calculations to outside of the mode_index loop. Change-Id: I06103fc7e8af88b54b84443acf6691d29b1272ac
2011-01-11 21:00:00 +01:00
y_buffer[LAST_FRAME] = lst_yv12->y_buffer + recon_yoffset;
u_buffer[LAST_FRAME] = lst_yv12->u_buffer + recon_uvoffset;
v_buffer[LAST_FRAME] = lst_yv12->v_buffer + recon_uvoffset;
}
if (cpi->ref_frame_flags & VP8_GOLD_FLAG)
{
YV12_BUFFER_CONFIG *gld_yv12 = &cpi->common.yv12_fb[cpi->common.gld_fb_idx];
Initial WebM release
2010-05-18 17:58:33 +02:00
changed find_near_mvs search to include a mb from last frame This is an experiment to include a mv contribution from last frame to nearest and near mv definition. Initial test showed some small though consistent gain. latest patch slightly better result ~.13%-~.18%. TODO: the entropy used to encode the mode choice, i.e. the mv counts based conditional distribution of modes should be re-collected to reflect this change, it is expected that there is some further gain from that. Change-Id: Ief1e284a36d8aa56b49ae5b360c91419ec494fa4
2011-11-16 01:16:30 +01:00
vp8_find_near_mvs(&x->e_mbd, x->e_mbd.mode_info_context,
#if CONFIG_NEWNEAR
x->e_mbd.prev_mode_info_context,
#endif
&frame_nearest_mv[GOLDEN_FRAME], &frame_near_mv[GOLDEN_FRAME],
&frame_best_ref_mv[GOLDEN_FRAME], frame_mdcounts[GOLDEN_FRAME], GOLDEN_FRAME, cpi->common.ref_frame_sign_bias);
Fix compiler warnings Change-Id: I2a97f08cc3c7808ce5be39e910cc5147ecf03a1d
2010-06-11 20:33:49 +02:00
Moved ref frame calculations Moved ref frame calculations to outside of the mode_index loop. Change-Id: I06103fc7e8af88b54b84443acf6691d29b1272ac
2011-01-11 21:00:00 +01:00
y_buffer[GOLDEN_FRAME] = gld_yv12->y_buffer + recon_yoffset;
u_buffer[GOLDEN_FRAME] = gld_yv12->u_buffer + recon_uvoffset;
v_buffer[GOLDEN_FRAME] = gld_yv12->v_buffer + recon_uvoffset;
}
if (cpi->ref_frame_flags & VP8_ALT_FLAG)
{
YV12_BUFFER_CONFIG *alt_yv12 = &cpi->common.yv12_fb[cpi->common.alt_fb_idx];
changed find_near_mvs search to include a mb from last frame This is an experiment to include a mv contribution from last frame to nearest and near mv definition. Initial test showed some small though consistent gain. latest patch slightly better result ~.13%-~.18%. TODO: the entropy used to encode the mode choice, i.e. the mv counts based conditional distribution of modes should be re-collected to reflect this change, it is expected that there is some further gain from that. Change-Id: Ief1e284a36d8aa56b49ae5b360c91419ec494fa4
2011-11-16 01:16:30 +01:00
vp8_find_near_mvs(&x->e_mbd, x->e_mbd.mode_info_context,
#if CONFIG_NEWNEAR
x->e_mbd.prev_mode_info_context,
#endif
&frame_nearest_mv[ALTREF_FRAME], &frame_near_mv[ALTREF_FRAME],
Moved ref frame calculations Moved ref frame calculations to outside of the mode_index loop. Change-Id: I06103fc7e8af88b54b84443acf6691d29b1272ac
2011-01-11 21:00:00 +01:00
&frame_best_ref_mv[ALTREF_FRAME], frame_mdcounts[ALTREF_FRAME], ALTREF_FRAME, cpi->common.ref_frame_sign_bias);
y_buffer[ALTREF_FRAME] = alt_yv12->y_buffer + recon_yoffset;
u_buffer[ALTREF_FRAME] = alt_yv12->u_buffer + recon_uvoffset;
v_buffer[ALTREF_FRAME] = alt_yv12->v_buffer + recon_uvoffset;
}
*returnintra = INT_MAX;
Initial WebM release
2010-05-18 17:58:33 +02:00
cpi->mbs_tested_so_far++; // Count of the number of MBs tested so far this frame
x->skip = 0;
vpx_memset(mode_mv, 0, sizeof(mode_mv));
Removed unnecessary MB_MODE_INFO copies These copies occurred for each macroblock in the encoder and decoder. Thetemp MB_MODE_INFO mbmi was removed from MACROBLOCKD. As a result, a large number compile errors had to be fixed. Change-Id: I4cf0ffae3ce244f6db04a4c217d52dd256382cf3
2010-08-12 22:25:43 +02:00
x->e_mbd.mode_info_context->mbmi.ref_frame = INTRA_FRAME;
Move RD intra block mode selection to rdopt.c This change is analogous to I0b67dae1f8a74902378da7bdf565e39ab832dda7, which made the move for the non-RD path. Change-Id: If63fc1b0cd1eb7f932e710f83ff24d91454f8ed1
2011-06-08 18:05:05 +02:00
rd_pick_intra_mbuv_mode(cpi, x, &uv_intra_rate, &uv_intra_rate_tokenonly, &uv_intra_distortion);
Removed unnecessary MB_MODE_INFO copies These copies occurred for each macroblock in the encoder and decoder. Thetemp MB_MODE_INFO mbmi was removed from MACROBLOCKD. As a result, a large number compile errors had to be fixed. Change-Id: I4cf0ffae3ce244f6db04a4c217d52dd256382cf3
2010-08-12 22:25:43 +02:00
uv_intra_mode = x->e_mbd.mode_info_context->mbmi.uv_mode;
Initial WebM release
2010-05-18 17:58:33 +02:00
for (mode_index = 0; mode_index < MAX_MODES; mode_index++)
{
int this_rd = INT_MAX;
int disable_skip = 0;
Adjusted breakout RD for SPLITMV vp8_rd_pick_best_mbsegmentation looks at y only. The new breakout does not include the frame cost, the prob_skip_false cost, or the uv rate. Performance improved by a few percent and the quality remained the same. Change-Id: I94ff013998ac51e8ecce7130870f7b6600758e15
2010-12-16 15:38:02 +01:00
int other_cost = 0;
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
int dualmode_cost = 0;
int mode_excluded = 0;
#endif /* CONFIG_DUALPRED */
Initial WebM release
2010-05-18 17:58:33 +02:00
// Experimental debug code.
// Record of rd values recorded for this MB. -1 indicates not measured
//all_rds[mode_index] = -1;
//all_rates[mode_index] = -1;
//all_dist[mode_index] = -1;
cosmetics: trim trailing whitespace When the license headers were updated, they accidentally contained trailing whitespace, so unfortunately we have to touch all the files again. Change-Id: I236c05fade06589e417179c0444cb39b09e4200d
2010-06-18 18:39:21 +02:00
//intermodecost[mode_index] = -1;
Initial WebM release
2010-05-18 17:58:33 +02:00
// Test best rd so far against threshold for trying this mode.
if (best_rd <= cpi->rd_threshes[mode_index])
continue;
// These variables hold are rolling total cost and distortion for this mode
rate2 = 0;
distortion2 = 0;
this_mode = vp8_mode_order[mode_index];
Removed unnecessary MB_MODE_INFO copies These copies occurred for each macroblock in the encoder and decoder. Thetemp MB_MODE_INFO mbmi was removed from MACROBLOCKD. As a result, a large number compile errors had to be fixed. Change-Id: I4cf0ffae3ce244f6db04a4c217d52dd256382cf3
2010-08-12 22:25:43 +02:00
x->e_mbd.mode_info_context->mbmi.mode = this_mode;
x->e_mbd.mode_info_context->mbmi.uv_mode = DC_PRED;
x->e_mbd.mode_info_context->mbmi.ref_frame = vp8_ref_frame_order[mode_index];
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
x->e_mbd.mode_info_context->mbmi.second_ref_frame = vp8_second_ref_frame_order[mode_index];
#endif /* CONFIG_DUALPRED */
Initial WebM release
2010-05-18 17:58:33 +02:00
Segment Features. Removal of #ifdefs Removal of configure #ifdefs so that segment features always available. Removal of code supporting old segment feature method. Still a good deal of tidying up to do. Change-Id: I397855f086f8c09ab1fae0a5f65d9e06d2e3e39f
2011-11-03 13:50:09 +01:00
//#if CONFIG_SEGFEATURES
Segmentation feature logic fix. Bug fix to logic in vp8_pick_inter_mode() and vp8_rd_pick_inter_mode(). The block on the use of segment features for the cm->refresh_alt_ref_frame case was just for testing and is not correct. The special case code for alt ref can be re-enabled as an else clause. Change-Id: Ic9b57cdb5f04ea7737032b8fb953d84d7717b3ce
2011-11-03 18:45:27 +01:00
// If the segment reference frame feature is enabled....
// then do nothing if the current ref frame is not allowed..
if ( segfeature_active( xd, segment_id, SEG_LVL_REF_FRAME ) &&
!check_segref( xd, segment_id,
xd->mode_info_context->mbmi.ref_frame ) )
Segment coding of mode and reference frame. Proof of concept test code that encodes mode and reference frame data at the segment level. Decode-able bit stream but some issues not yet resolved. As it this helps a little on a couple of clips but hurts on most as the basis for segmentation is unsound. To build and test, configure with --enable-experimental --enable-segfeatures Change-Id: I22a60774f69273523fb152db8c31f4b10b07c7f4
2011-09-30 17:45:16 +02:00
{
Segmentation feature logic fix. Bug fix to logic in vp8_pick_inter_mode() and vp8_rd_pick_inter_mode(). The block on the use of segment features for the cm->refresh_alt_ref_frame case was just for testing and is not correct. The special case code for alt ref can be re-enabled as an else clause. Change-Id: Ic9b57cdb5f04ea7737032b8fb953d84d7717b3ce
2011-11-03 18:45:27 +01:00
continue;
}
// If the segment mode feature is enabled....
// then do nothing if the current mode is not allowed..
else if ( segfeature_active( xd, segment_id, SEG_LVL_MODE ) &&
( this_mode !=
get_segdata( xd, segment_id, SEG_LVL_MODE ) ) )
{
continue;
Segment coding of mode and reference frame. Proof of concept test code that encodes mode and reference frame data at the segment level. Decode-able bit stream but some issues not yet resolved. As it this helps a little on a couple of clips but hurts on most as the basis for segmentation is unsound. To build and test, configure with --enable-experimental --enable-segfeatures Change-Id: I22a60774f69273523fb152db8c31f4b10b07c7f4
2011-09-30 17:45:16 +02:00
}
CONFIG_T8X8 experiment.: Block the selection of 4x4 modes in key frames if 8x8 is selected. Change-Id: Ie5729ec22a999d9a1996f020bd4b941e29514992
2011-11-18 20:44:57 +01:00
#if CONFIG_T8X8
T8x8 experiment merge. For ease of testing and merging experiments I have removed in line code in encode_frame() that assigns MBs to be t8x8 or t4x4 coded segments and have moved the decision point and segment setup to the init_seg_features0 test function. Keeping everything in one place helps make sure for now that experiments using segmentation are not fighting each other. Also made sure mode selection code can't choose 4x4 modes if t8x8 is selected. Patch2: In init_seg_features() add checks for SEG_LVL_TRANSFORM active. Change-Id: Ia1767edd99b78510011d4251539f9bc325842e3a
2011-11-09 15:05:28 +01:00
// No 4x4 modes if segment flagged as 8x8
else if ( ( get_seg_tx_type( xd, segment_id ) == TX_8X8 ) &&
( (this_mode == B_PRED) || (this_mode == SPLITMV) ) )
{
continue;
}
CONFIG_T8X8 experiment.: Block the selection of 4x4 modes in key frames if 8x8 is selected. Change-Id: Ie5729ec22a999d9a1996f020bd4b941e29514992
2011-11-18 20:44:57 +01:00
#endif
Segment Features. Removal of #ifdefs Removal of configure #ifdefs so that segment features always available. Removal of code supporting old segment feature method. Still a good deal of tidying up to do. Change-Id: I397855f086f8c09ab1fae0a5f65d9e06d2e3e39f
2011-11-03 13:50:09 +01:00
//#if !CONFIG_SEGFEATURES
Segmentation feature logic fix. Bug fix to logic in vp8_pick_inter_mode() and vp8_rd_pick_inter_mode(). The block on the use of segment features for the cm->refresh_alt_ref_frame case was just for testing and is not correct. The special case code for alt ref can be re-enabled as an else clause. Change-Id: Ic9b57cdb5f04ea7737032b8fb953d84d7717b3ce
2011-11-03 18:45:27 +01:00
// Disable this drop out case if either the mode or ref frame
// segment level feature is enabled for this segment. This is to
// prevent the possibility that the we end up unable to pick any mode.
else if ( !segfeature_active( xd, segment_id, SEG_LVL_REF_FRAME ) &&
!segfeature_active( xd, segment_id, SEG_LVL_MODE ) )
Initial WebM release
2010-05-18 17:58:33 +02:00
{
Segmentation feature logic fix. Bug fix to logic in vp8_pick_inter_mode() and vp8_rd_pick_inter_mode(). The block on the use of segment features for the cm->refresh_alt_ref_frame case was just for testing and is not correct. The special case code for alt ref can be re-enabled as an else clause. Change-Id: Ic9b57cdb5f04ea7737032b8fb953d84d7717b3ce
2011-11-03 18:45:27 +01:00
// Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
// unless ARNR filtering is enabled in which case we want
// an unfiltered alternative
if (cpi->is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0))
{
if (this_mode != ZEROMV ||
x->e_mbd.mode_info_context->mbmi.ref_frame != ALTREF_FRAME)
continue;
}
}
Initial WebM release
2010-05-18 17:58:33 +02:00
Moved ref frame calculations Moved ref frame calculations to outside of the mode_index loop. Change-Id: I06103fc7e8af88b54b84443acf6691d29b1272ac
2011-01-11 21:00:00 +01:00
/* everything but intra */
if (x->e_mbd.mode_info_context->mbmi.ref_frame)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
Moved ref frame calculations Moved ref frame calculations to outside of the mode_index loop. Change-Id: I06103fc7e8af88b54b84443acf6691d29b1272ac
2011-01-11 21:00:00 +01:00
x->e_mbd.pre.y_buffer = y_buffer[x->e_mbd.mode_info_context->mbmi.ref_frame];
x->e_mbd.pre.u_buffer = u_buffer[x->e_mbd.mode_info_context->mbmi.ref_frame];
x->e_mbd.pre.v_buffer = v_buffer[x->e_mbd.mode_info_context->mbmi.ref_frame];
mode_mv[NEARESTMV] = frame_nearest_mv[x->e_mbd.mode_info_context->mbmi.ref_frame];
mode_mv[NEARMV] = frame_near_mv[x->e_mbd.mode_info_context->mbmi.ref_frame];
best_ref_mv = frame_best_ref_mv[x->e_mbd.mode_info_context->mbmi.ref_frame];
vpx_memcpy(mdcounts, frame_mdcounts[x->e_mbd.mode_info_context->mbmi.ref_frame], sizeof(mdcounts));
Initial WebM release
2010-05-18 17:58:33 +02:00
}
// Experimental code. Special case for gf and arf zeromv modes. Increase zbin size to supress noise
if (cpi->zbin_mode_boost_enabled)
{
Tuning for the more exact quantizer. Small changes to the default zero bin and rounding tables. Though the tables are currently the same for the Y1 and Y2 cases I have left them as separate tables in case we want to tune this later. There is now some adjustment of the zbin based on the prediction mode. Previously this was restricted to an adjustment for gf/arf 0,0 MV. The exact quantizer now marginal outperforms and is the default. The overall average gain is about 0.5% Change-Id: I5e4353f3d5326dde4e86823684b236a1e9ea7f47
2010-11-08 16:28:54 +01:00
if ( vp8_ref_frame_order[mode_index] == INTRA_FRAME )
Initial WebM release
2010-05-18 17:58:33 +02:00
cpi->zbin_mode_boost = 0;
Tuning for the more exact quantizer. Small changes to the default zero bin and rounding tables. Though the tables are currently the same for the Y1 and Y2 cases I have left them as separate tables in case we want to tune this later. There is now some adjustment of the zbin based on the prediction mode. Previously this was restricted to an adjustment for gf/arf 0,0 MV. The exact quantizer now marginal outperforms and is the default. The overall average gain is about 0.5% Change-Id: I5e4353f3d5326dde4e86823684b236a1e9ea7f47
2010-11-08 16:28:54 +01:00
else
{
if (vp8_mode_order[mode_index] == ZEROMV)
{
if (vp8_ref_frame_order[mode_index] != LAST_FRAME)
cpi->zbin_mode_boost = GF_ZEROMV_ZBIN_BOOST;
else
cpi->zbin_mode_boost = LF_ZEROMV_ZBIN_BOOST;
}
correct zbin boost for splitmv mode Disable zbin boost in SPLITMV mode as intended. Was incorrectly looking at vp8_ref_frame_order instead of vp8_mode_order when comparing against SPLITMV. This condition should have always been false, as SPLITMV is not in the range of valid reference frames. Change-Id: I0408cc7595eff68f00efef6d008e79f5b60d14bf
2011-03-08 02:58:37 +01:00
else if (vp8_mode_order[mode_index] == SPLITMV)
Tuning for the more exact quantizer. Small changes to the default zero bin and rounding tables. Though the tables are currently the same for the Y1 and Y2 cases I have left them as separate tables in case we want to tune this later. There is now some adjustment of the zbin based on the prediction mode. Previously this was restricted to an adjustment for gf/arf 0,0 MV. The exact quantizer now marginal outperforms and is the default. The overall average gain is about 0.5% Change-Id: I5e4353f3d5326dde4e86823684b236a1e9ea7f47
2010-11-08 16:28:54 +01:00
cpi->zbin_mode_boost = 0;
else
cpi->zbin_mode_boost = MV_ZBIN_BOOST;
}
Initial WebM release
2010-05-18 17:58:33 +02:00
Added vp8_update_zbin_extra vp8cx_mb_init_quantizer was being called for every mode checked in vp8_rd_pick_inter_mode. zbin_extra is the only value that really needs to be recalculated. This calculation is disabled when using the fast quantizer for mode selection. This gave a small performance boost (~.5% to 1%). Note: This needs to be verified with segmentation_enabled. Change-Id: I62716a870b3c82b4a998bdf95130ff0b02106f1e
2010-12-29 20:30:57 +01:00
vp8_update_zbin_extra(cpi, x);
Initial WebM release
2010-05-18 17:58:33 +02:00
}
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
if (!x->e_mbd.mode_info_context->mbmi.second_ref_frame)
#endif /* CONFIG_DUALPRED */
Initial WebM release
2010-05-18 17:58:33 +02:00
switch (this_mode)
{
case B_PRED:
Added early breakout for vp8_rd_pick_intra4x4mby_modes Improved performance of good quality, speed 0 (3% average) with no average quality loss. Change-Id: Ica34473f99bd74260eaebde6b132185e09e3c09d
2011-02-08 22:50:43 +01:00
{
int tmp_rd;
Initial WebM release
2010-05-18 17:58:33 +02:00
// Note the rate value returned here includes the cost of coding the BPRED mode : x->mbmode_cost[x->e_mbd.frame_type][BPRED];
Move RD intra block mode selection to rdopt.c This change is analogous to I0b67dae1f8a74902378da7bdf565e39ab832dda7, which made the move for the non-RD path. Change-Id: If63fc1b0cd1eb7f932e710f83ff24d91454f8ed1
2011-06-08 18:05:05 +02:00
tmp_rd = rd_pick_intra4x4mby_modes(cpi, x, &rate, &rate_y, &distortion, best_yrd);
Initial WebM release
2010-05-18 17:58:33 +02:00
rate2 += rate;
distortion2 += distortion;
Added early breakout for vp8_rd_pick_intra4x4mby_modes Improved performance of good quality, speed 0 (3% average) with no average quality loss. Change-Id: Ica34473f99bd74260eaebde6b132185e09e3c09d
2011-02-08 22:50:43 +01:00
if(tmp_rd < best_yrd)
{
rate2 += uv_intra_rate;
rate_uv = uv_intra_rate_tokenonly;
distortion2 += uv_intra_distortion;
distortion_uv = uv_intra_distortion;
}
else
{
this_rd = INT_MAX;
disable_skip = 1;
}
}
break;
enabled 8x8 intra prediction modes on inter frames This commit enabled the usage of 8x8 intra prediction modes on inter frames. There are a few TODO items related to this: 1)baseline entropy need be calibrated; 2)cost of UV need to be done more properly rather than using decision only relying on Y; 3)Threshold for allowing picking 8x8 intra prediction should be lowered to lower than the B_PRED. Even with all the TODOs, tests showed consistent gain on derf set ~0.1% (PSNR:0.08% and SSIM:0.14%). It is assumed that 8x8 intra prediction will help more on large resolution clips, especially with above TODOs addressed. Change-Id: I398ada49dfc32575cfab962a569c2885111ae3ba
2011-12-01 01:25:00 +01:00
case I8X8_PRED:
{
int tmp_rd;
tmp_rd = rd_pick_intra8x8mby_modes(cpi,
x, &rate, &rate_y, &distortion, best_yrd);
rate2 += rate;
distortion2 += distortion;
mode8x8[0]= x->e_mbd.mode_info_context->bmi[0].as_mode;
mode8x8[1]= x->e_mbd.mode_info_context->bmi[2].as_mode;
mode8x8[2]= x->e_mbd.mode_info_context->bmi[8].as_mode;
mode8x8[3]= x->e_mbd.mode_info_context->bmi[10].as_mode;
/* TODO: uv rate maybe over-estimated here since there is UV intra
mode coded in I8X8_PRED prediction */
if(tmp_rd < best_yrd)
{
rate2 += uv_intra_rate;
rate_uv = uv_intra_rate_tokenonly;
distortion2 += uv_intra_distortion;
distortion_uv = uv_intra_distortion;
}
else
{
this_rd = INT_MAX;
disable_skip = 1;
}
}
break;
Initial WebM release
2010-05-18 17:58:33 +02:00
case SPLITMV:
{
int tmp_rd;
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
int this_rd_thresh;
Initial WebM release
2010-05-18 17:58:33 +02:00
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
this_rd_thresh = (x->e_mbd.mode_info_context->mbmi.ref_frame == LAST_FRAME) ? cpi->rd_threshes[THR_NEWMV] : cpi->rd_threshes[THR_NEWA];
this_rd_thresh = (x->e_mbd.mode_info_context->mbmi.ref_frame == GOLDEN_FRAME) ? cpi->rd_threshes[THR_NEWG]: this_rd_thresh;
tmp_rd = vp8_rd_pick_best_mbsegmentation(cpi, x, &best_ref_mv,
Adjusted breakout RD for SPLITMV vp8_rd_pick_best_mbsegmentation looks at y only. The new breakout does not include the frame cost, the prob_skip_false cost, or the uv rate. Performance improved by a few percent and the quality remained the same. Change-Id: I94ff013998ac51e8ecce7130870f7b6600758e15
2010-12-16 15:38:02 +01:00
best_yrd, mdcounts,
vp8_rd_pick_best_mbsegmentation code restructure Moved the code from the segmentation loop into a function which is now called for each segment. This will allow us to change the segment order checking more easily. Change-Id: I9510d26f0acae5a73043fcca8f1984b121d3e052
2010-12-06 22:42:52 +01:00
&rate, &rate_y, &distortion, this_rd_thresh) ;
Initial WebM release
2010-05-18 17:58:33 +02:00
rate2 += rate;
distortion2 += distortion;
// If even the 'Y' rd value of split is higher than best so far then dont bother looking at UV
Adjusted breakout RD for SPLITMV vp8_rd_pick_best_mbsegmentation looks at y only. The new breakout does not include the frame cost, the prob_skip_false cost, or the uv rate. Performance improved by a few percent and the quality remained the same. Change-Id: I94ff013998ac51e8ecce7130870f7b6600758e15
2010-12-16 15:38:02 +01:00
if (tmp_rd < best_yrd)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
// Now work out UV cost and add it in
Removed bmi copy to/from BLOCKD for SPLITMV and B_PRED modes. Modified code to use the bmi found in mode_info_context instead of BLOCKD. On the decode side, the uvmvs are calculated only when required, instead of every macroblock. This is WIP. (bmi should eventually be removed from BLOCKD) Small performance gains noticed for RT encodes and decodes.(VGA) Change-Id: I2ed7f0fd5ca733655df684aa82da575c77a973e7
2011-08-24 20:42:26 +02:00
rd_inter4x4_uv(cpi, x, &rate_uv, &distortion_uv, cpi->common.full_pixel);
Adjusted breakout RD for SPLITMV vp8_rd_pick_best_mbsegmentation looks at y only. The new breakout does not include the frame cost, the prob_skip_false cost, or the uv rate. Performance improved by a few percent and the quality remained the same. Change-Id: I94ff013998ac51e8ecce7130870f7b6600758e15
2010-12-16 15:38:02 +01:00
rate2 += rate_uv;
distortion2 += distortion_uv;
Initial WebM release
2010-05-18 17:58:33 +02:00
}
else
{
this_rd = INT_MAX;
disable_skip = 1;
}
}
break;
case DC_PRED:
case V_PRED:
case H_PRED:
case TM_PRED:
Removed unnecessary MB_MODE_INFO copies These copies occurred for each macroblock in the encoder and decoder. Thetemp MB_MODE_INFO mbmi was removed from MACROBLOCKD. As a result, a large number compile errors had to be fixed. Change-Id: I4cf0ffae3ce244f6db04a4c217d52dd256382cf3
2010-08-12 22:25:43 +02:00
x->e_mbd.mode_info_context->mbmi.ref_frame = INTRA_FRAME;
Move build_intra_predictors_mby to RTCD framework The vp8_build_intra_predictors_mby and vp8_build_intra_predictors_mby_s functions had global function pointers rather than using the RTCD framework. This can show up as a potential data race with tools such as helgrind. See https://bugzilla.mozilla.org/show_bug.cgi?id=640935 for an example. Change-Id: I29c407f828ac2bddfc039f852f138de5de888534
2011-03-11 17:35:38 +01:00
RECON_INVOKE(&cpi->common.rtcd.recon, build_intra_predictors_mby)
(&x->e_mbd);
Zero out block mv when an intra mode is selected instead of each time mode is tested. Change-Id: Ief0f5586dafde54cc14d348dcecdacb182e7c1d5
2011-02-01 18:55:51 +01:00
macro_block_yrd(x, &rate_y, &distortion, IF_RTCD(&cpi->rtcd.encodemb)) ;
rate2 += rate_y;
distortion2 += distortion;
rate2 += x->mbmode_cost[x->e_mbd.frame_type][x->e_mbd.mode_info_context->mbmi.mode];
rate2 += uv_intra_rate;
rate_uv = uv_intra_rate_tokenonly;
distortion2 += uv_intra_distortion;
distortion_uv = uv_intra_distortion;
Initial WebM release
2010-05-18 17:58:33 +02:00
break;
case NEWMV:
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
{
int thissme;
int bestsme = INT_MAX;
int step_param = cpi->sf.first_step;
int further_steps;
int n;
int do_refine=1; /* If last step (1-away) of n-step search doesn't pick the center point as the best match,
we will do a final 1-away diamond refining search */
Initial WebM release
2010-05-18 17:58:33 +02:00
further clean up of errorperbit and sadperbit this commit makes the usage errorperbit and sadperbit consistent for encoding modes and passes. Removed all different magic weight factors associated with errorperbit. Now 1/2 is used for both sadperbit16 and sadperbit4, the /2 operation is merged into initializations of the 2 variables. Tests on cif set show .23%, 0.18% and 0.19% gain by avg psnr, overall psnr and ssim respectively. Change-Id: Ifa285c3e065ce0a5a77addfc9f95aabf54ee270d
2011-06-01 20:41:05 +02:00
int sadpb = x->sadperbit16;
Adjust full-pixel clamping and motion vector limit calculation Do mvp clamping in full-pixel precision instead of 1/8-pixel precision to avoid error caused by right shifting operation. Also, further fixed the motion vector limit calculation in change: b7480454706a6b15bf091e659cd6227ab373c1a6 Change-Id: Ied88a4f7ddfb0476eb9f7afc6ceeddbf209fffd7
2011-07-07 17:21:41 +02:00
int_mv mvp_full;
Initial WebM release
2010-05-18 17:58:33 +02:00
Adjust full-pixel clamping and motion vector limit calculation Do mvp clamping in full-pixel precision instead of 1/8-pixel precision to avoid error caused by right shifting operation. Also, further fixed the motion vector limit calculation in change: b7480454706a6b15bf091e659cd6227ab373c1a6 Change-Id: Ied88a4f7ddfb0476eb9f7afc6ceeddbf209fffd7
2011-07-07 17:21:41 +02:00
int col_min = (best_ref_mv.as_mv.col>>3) - MAX_FULL_PEL_VAL + ((best_ref_mv.as_mv.col & 7)?1:0);
int row_min = (best_ref_mv.as_mv.row>>3) - MAX_FULL_PEL_VAL + ((best_ref_mv.as_mv.row & 7)?1:0);
int col_max = (best_ref_mv.as_mv.col>>3) + MAX_FULL_PEL_VAL;
int row_max = (best_ref_mv.as_mv.row>>3) + MAX_FULL_PEL_VAL;
Initial WebM release
2010-05-18 17:58:33 +02:00
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
int tmp_col_min = x->mv_col_min;
int tmp_col_max = x->mv_col_max;
int tmp_row_min = x->mv_row_min;
int tmp_row_max = x->mv_row_max;
Initial WebM release
2010-05-18 17:58:33 +02:00
Minor change in pick_inter_mode() Scott suggested to move vp8_mv_pred() under "case NEWMV" to save extra checks. Change-Id: I09e69892f34a08dd425a4d81cfcc83674e344a20
2011-07-08 20:08:45 +02:00
if(!saddone)
{
vp8_cal_sad(cpi,xd,x, recon_yoffset ,&near_sadidx[0] );
saddone = 1;
}
vp8_mv_pred(cpi, &x->e_mbd, x->e_mbd.mode_info_context, &mvp,
x->e_mbd.mode_info_context->mbmi.ref_frame, cpi->common.ref_frame_sign_bias, &sr, &near_sadidx[0]);
Adjust full-pixel clamping and motion vector limit calculation Do mvp clamping in full-pixel precision instead of 1/8-pixel precision to avoid error caused by right shifting operation. Also, further fixed the motion vector limit calculation in change: b7480454706a6b15bf091e659cd6227ab373c1a6 Change-Id: Ied88a4f7ddfb0476eb9f7afc6ceeddbf209fffd7
2011-07-07 17:21:41 +02:00
mvp_full.as_mv.col = mvp.as_mv.col>>3;
mvp_full.as_mv.row = mvp.as_mv.row>>3;
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
// Get intersection of UMV window and valid MV window to reduce # of checks in diamond search.
if (x->mv_col_min < col_min )
x->mv_col_min = col_min;
if (x->mv_col_max > col_max )
x->mv_col_max = col_max;
if (x->mv_row_min < row_min )
x->mv_row_min = row_min;
if (x->mv_row_max > row_max )
x->mv_row_max = row_max;
Initial WebM release
2010-05-18 17:58:33 +02:00
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
//adjust search range according to sr from mv prediction
if(sr > step_param)
step_param = sr;
Initial WebM release
2010-05-18 17:58:33 +02:00
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
// Initial step/diamond search
{
Adjust full-pixel clamping and motion vector limit calculation Do mvp clamping in full-pixel precision instead of 1/8-pixel precision to avoid error caused by right shifting operation. Also, further fixed the motion vector limit calculation in change: b7480454706a6b15bf091e659cd6227ab373c1a6 Change-Id: Ied88a4f7ddfb0476eb9f7afc6ceeddbf209fffd7
2011-07-07 17:21:41 +02:00
bestsme = cpi->diamond_search_sad(x, b, d, &mvp_full, &d->bmi.mv,
remove some magic weights associated with sad_per_bit sad_per_bit has been used for a number of motion vector search routines with different magic weights: 1, 1/2 and 1/4. This commit remove these magic numbers and use 1/2 for all motion search routines, also reformat a number of source code lines to within 80 column limit. Test on cif set shows overall effect is neutral on all metrics. <=0.01% Change-Id: I8a382821fa4cffc9c0acf8e8431435a03df74885
2010-07-01 03:58:54 +02:00
step_param, sadpb, &num00,
&cpi->fn_ptr[BLOCK_16X16],
x->mvcost, &best_ref_mv);
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
mode_mv[NEWMV].as_int = d->bmi.mv.as_int;
Initial WebM release
2010-05-18 17:58:33 +02:00
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
// Further step/diamond searches as necessary
n = 0;
further_steps = (cpi->sf.max_step_search_steps - 1) - step_param;
Initial WebM release
2010-05-18 17:58:33 +02:00
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
n = num00;
num00 = 0;
Initial WebM release
2010-05-18 17:58:33 +02:00
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
/* If there won't be more n-step search, check to see if refining search is needed. */
if (n > further_steps)
do_refine = 0;
Initial WebM release
2010-05-18 17:58:33 +02:00
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
while (n < further_steps)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
n++;
Initial WebM release
2010-05-18 17:58:33 +02:00
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
if (num00)
num00--;
Initial WebM release
2010-05-18 17:58:33 +02:00
else
{
Adjust full-pixel clamping and motion vector limit calculation Do mvp clamping in full-pixel precision instead of 1/8-pixel precision to avoid error caused by right shifting operation. Also, further fixed the motion vector limit calculation in change: b7480454706a6b15bf091e659cd6227ab373c1a6 Change-Id: Ied88a4f7ddfb0476eb9f7afc6ceeddbf209fffd7
2011-07-07 17:21:41 +02:00
thissme = cpi->diamond_search_sad(x, b, d, &mvp_full,
remove some magic weights associated with sad_per_bit sad_per_bit has been used for a number of motion vector search routines with different magic weights: 1, 1/2 and 1/4. This commit remove these magic numbers and use 1/2 for all motion search routines, also reformat a number of source code lines to within 80 column limit. Test on cif set shows overall effect is neutral on all metrics. <=0.01% Change-Id: I8a382821fa4cffc9c0acf8e8431435a03df74885
2010-07-01 03:58:54 +02:00
&d->bmi.mv, step_param + n, sadpb, &num00,
&cpi->fn_ptr[BLOCK_16X16], x->mvcost,
&best_ref_mv);
Initial WebM release
2010-05-18 17:58:33 +02:00
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
/* check to see if refining search is needed. */
if (num00 > (further_steps-n))
do_refine = 0;
Initial WebM release
2010-05-18 17:58:33 +02:00
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
if (thissme < bestsme)
{
bestsme = thissme;
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
mode_mv[NEWMV].as_int = d->bmi.mv.as_int;
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
}
else
{
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
d->bmi.mv.as_int = mode_mv[NEWMV].as_int;
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
}
Initial WebM release
2010-05-18 17:58:33 +02:00
}
}
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
}
/* final 1-away diamond refining search */
if (do_refine == 1)
{
int search_range;
Initial WebM release
2010-05-18 17:58:33 +02:00
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
//It seems not a good way to set search_range. Need further investigation.
//search_range = MAXF(abs((mvp.row>>3) - d->bmi.mv.as_mv.row), abs((mvp.col>>3) - d->bmi.mv.as_mv.col));
search_range = 8;
Fix a bug in motion search code(2) This fix added MV range checks for NEWMV mode as suggested by Jim. To reduce unnecessary MV range checks, I tried Yaowu's suggestion. Update UMV borders in NEWMV mode to also cover MV range check. Also, in this way, every MV that is valid gets checked in diamond search function. Change-Id: I95a89ce0daf6f178c454448f13d4249f19b30f3a
2010-12-14 23:39:25 +01:00
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
//thissme = cpi->full_search_sad(x, b, d, &d->bmi.mv.as_mv, sadpb, search_range, &cpi->fn_ptr[BLOCK_16X16], x->mvcost, &best_ref_mv);
remove some magic weights associated with sad_per_bit sad_per_bit has been used for a number of motion vector search routines with different magic weights: 1, 1/2 and 1/4. This commit remove these magic numbers and use 1/2 for all motion search routines, also reformat a number of source code lines to within 80 column limit. Test on cif set shows overall effect is neutral on all metrics. <=0.01% Change-Id: I8a382821fa4cffc9c0acf8e8431435a03df74885
2010-07-01 03:58:54 +02:00
thissme = cpi->refining_search_sad(x, b, d, &d->bmi.mv, sadpb,
search_range, &cpi->fn_ptr[BLOCK_16X16],
x->mvcost, &best_ref_mv);
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
if (thissme < bestsme)
Use sub-pixel search's SSE in mode selection Passed SSE from sub-pixel search back to pick_inter_mode function, which is compared with the encode_breakout to see if we could skip evaluating the remaining modes. Change-Id: I4a86442834f0d1b880a19e21ea52d17d505f941d
2011-04-18 21:48:34 +02:00
{
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
bestsme = thissme;
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
mode_mv[NEWMV].as_int = d->bmi.mv.as_int;
Reduce unnecessary distortion computation In vp8_pick_inter_mode(), for NEWMV mode, use the error result got from motion search as distortion. This helps performance in real- time mode. Change-Id: I398c4e46cc5381f7d874e748cf78827ef0e0860c
2011-04-14 21:53:33 +02:00
}
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
else
{
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
d->bmi.mv.as_int = mode_mv[NEWMV].as_int;
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
}
}
Initial WebM release
2010-05-18 17:58:33 +02:00
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
x->mv_col_min = tmp_col_min;
x->mv_col_max = tmp_col_max;
x->mv_row_min = tmp_row_min;
x->mv_row_max = tmp_row_max;
Initial WebM release
2010-05-18 17:58:33 +02:00
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
if (bestsme < INT_MAX)
{
int dis; /* TODO: use dis in distortion calculation later. */
unsigned int sse;
further clean up of errorperbit and sadperbit this commit makes the usage errorperbit and sadperbit consistent for encoding modes and passes. Removed all different magic weight factors associated with errorperbit. Now 1/2 is used for both sadperbit16 and sadperbit4, the /2 operation is merged into initializations of the 2 variables. Tests on cif set show .23%, 0.18% and 0.19% gain by avg psnr, overall psnr and ssim respectively. Change-Id: Ifa285c3e065ce0a5a77addfc9f95aabf54ee270d
2011-06-01 20:41:05 +02:00
cpi->find_fractional_mv_step(x, b, d, &d->bmi.mv, &best_ref_mv,
x->errorperbit,
&cpi->fn_ptr[BLOCK_16X16],
x->mvcost, &dis, &sse);
Initial WebM release
2010-05-18 17:58:33 +02:00
}
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
mc_search_result[x->e_mbd.mode_info_context->mbmi.ref_frame].as_int = d->bmi.mv.as_int;
#endif /* CONFIG_DUALPRED */
Initial WebM release
2010-05-18 17:58:33 +02:00
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
mode_mv[NEWMV].as_int = d->bmi.mv.as_int;
Use diamond search to replace full search in full-pixel refining search In NEWMV mode, currently, full search is used as the refining search after n-step search. By replacing it with an iterative diamond search of radius 1 largely reduced the computation complexity, but still maintained the same encoding quality since the refining search is done for every macroblock instead of only a small precentage of macroblocks while using full search. Tests on the test set showed a 3.4% encoding speed increase with none psnr & ssim loss. Change-Id: Ife907d7eb9544d15c34f17dc6e4cfd97cb743d41
2011-05-06 18:51:31 +02:00
// Add the new motion vector cost to our rolling cost variable
rate2 += vp8_mv_bit_cost(&mode_mv[NEWMV], &best_ref_mv, x->mvcost, 96);
}
Initial WebM release
2010-05-18 17:58:33 +02:00
case NEARESTMV:
case NEARMV:
// Clip "next_nearest" so that it does not extend to far out of image
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
vp8_clamp_mv2(&mode_mv[this_mode], xd);
Initial WebM release
2010-05-18 17:58:33 +02:00
// Do not bother proceeding if the vector (from newmv,nearest or near) is 0,0 as this should then be coded using the zeromv mode.
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
if (((this_mode == NEARMV) || (this_mode == NEARESTMV)) && (mode_mv[this_mode].as_int == 0))
Initial WebM release
2010-05-18 17:58:33 +02:00
continue;
case ZEROMV:
// Trap vectors that reach beyond the UMV borders
// Note that ALL New MV, Nearest MV Near MV and Zero MV code drops through to this point
// because of the lack of break statements in the previous two cases.
Using int_mv instead of MV The compiler produces better assembly when using int_mv for assignments. The compiler shifts and ors the two 16bit values when assigning MV. Change-Id: I52ce4bc2bfbfaf3f1151204b2f21e1e0654f960f
2011-05-12 16:50:16 +02:00
if (((mode_mv[this_mode].as_mv.row >> 3) < x->mv_row_min) || ((mode_mv[this_mode].as_mv.row >> 3) > x->mv_row_max) ||
((mode_mv[this_mode].as_mv.col >> 3) < x->mv_col_min) || ((mode_mv[this_mode].as_mv.col >> 3) > x->mv_col_max))
Initial WebM release
2010-05-18 17:58:33 +02:00
continue;
vp8_set_mbmode_and_mvs(x, this_mode, &mode_mv[this_mode]);
Removed redundant checks of the mode_info_context flags Code cleanup. The build inter predictor functions are redundantly checking the mode_info_context for either INTRA_FRAME or SPLITMV. Change-Id: I4d58c3a5192a4c2cec5c24ab1caf608bf13aebfb
2011-04-19 17:42:15 +02:00
vp8_build_inter16x16_predictors_mby(&x->e_mbd);
Initial WebM release
2010-05-18 17:58:33 +02:00
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
Minor fixes: fixed issues caused by conflicts between two experiments. Change-Id: I56a9bd69493e4850c121ea057a6233c55777c2a5
2011-12-07 14:32:45 +01:00
{
MB_MODE_INFO *t = &x->e_mbd.mode_info_context[-cpi->common.mode_info_stride].mbmi;
MB_MODE_INFO *l = &x->e_mbd.mode_info_context[-1].mbmi;
int cnt = (t->second_ref_frame != INTRA_FRAME) + (l->second_ref_frame != INTRA_FRAME);
dualmode_cost = vp8_cost_bit(cpi->prob_dualpred[cnt], 0);
}
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#endif /* CONFIG_DUALPRED */
vp8e - static threshold play Realized no need for new assembly code sum is already calculated. Change-Id: Ie2d94feb4b7c1f77c5359bca29b66228e41638c9
2010-12-07 22:07:23 +01:00
if (cpi->active_map_enabled && x->active_ptr[0] == 0) {
Initial WebM release
2010-05-18 17:58:33 +02:00
x->skip = 1;
}
fix uninitialized read in encode breakout Change I3430820 performed an uninitialized read when encode_breakout == 0, since the sum and sse wouldn't be set: if(x->encode_breakout) VARIANCE_INVOKE(..., get16x16var)(..., &sum, &sse); if (cpi->active_map_enabled && x->active_ptr[0] == 0) { ... } else if (sse < x->encode_breakout) Change-Id: I915eb76d1227b4b6d1137a0dedf2c143860098a2
2010-12-09 21:59:22 +01:00
else if (x->encode_breakout)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
Use sub-pixel search's SSE in mode selection Passed SSE from sub-pixel search back to pick_inter_mode function, which is compared with the encode_breakout to see if we could skip evaluating the remaining modes. Change-Id: I4a86442834f0d1b880a19e21ea52d17d505f941d
2011-04-18 21:48:34 +02:00
unsigned int sse;
remove redundant functions The encoder defined about 4 set of similar functions to calculate sum, variance or sse or a combination of them. This commit removed one set of these functions, get8x8var and get16x16var, where calls to the later function are replaced with var16x16 by using the fact on a 16x16 MB: variance == sse - sum*sum/256 Change-Id: I803eabd1fb3ab177780a40338cbd596dffaed267
2011-06-07 01:42:58 +02:00
unsigned int var;
change the threshold of DC check for encode breakout Previously, the DC check is to make sure there is no code-able DC shift for quantizer Q0, which has been verified rather conservative. This commit changes the criteria to have two components, DC and AC, to address the conservativeness. First, it checks if all AC energy is enough to contribute a single non-zero quantized AC coefficient. Second, for DC, the decision to skip further considers two possible scenarios: 1. There is no code-able 2nd order DC coefficient at all; 2 The residue is relatively flat, but the uniform DC change is very small, i.e. less than 1/2 gray level per pixel. Comparing to previous criteria, the new criteria is about 10% to 15% faster in encoding time with a very small quality loss. (threshold ~1000 and quality range 33db-45db) It should be noted that this commit enables "automatic" static threshold for encodebreakout if a non-zero small value is passed in to encoder. Change-Id: I0f77719a1ac2c2dfddbd950d84920df374515ce3
2011-01-20 01:21:01 +01:00
int threshold = (xd->block[0].dequant[1]
* xd->block[0].dequant[1] >>4);
if(threshold < x->encode_breakout)
threshold = x->encode_breakout;
Initial WebM release
2010-05-18 17:58:33 +02:00
remove redundant functions The encoder defined about 4 set of similar functions to calculate sum, variance or sse or a combination of them. This commit removed one set of these functions, get8x8var and get16x16var, where calls to the later function are replaced with var16x16 by using the fact on a 16x16 MB: variance == sse - sum*sum/256 Change-Id: I803eabd1fb3ab177780a40338cbd596dffaed267
2011-06-07 01:42:58 +02:00
var = VARIANCE_INVOKE(&cpi->rtcd.variance, var16x16)
Copy macroblock data to a buffer before encoding it I got this idea from Pascal (Thanks). Before encoding a macroblock, copy it to a 16x16 buffer, and then read source data from there instead. This will help keep the source data in cache, and help with the performance. Change-Id: Id05f4cb601299150511d59dcba0ae62c49b5b757
2011-06-23 19:54:02 +02:00
(*(b->base_src), b->src_stride,
remove redundant functions The encoder defined about 4 set of similar functions to calculate sum, variance or sse or a combination of them. This commit removed one set of these functions, get8x8var and get16x16var, where calls to the later function are replaced with var16x16 by using the fact on a 16x16 MB: variance == sse - sum*sum/256 Change-Id: I803eabd1fb3ab177780a40338cbd596dffaed267
2011-06-07 01:42:58 +02:00
x->e_mbd.predictor, 16, &sse);
fix uninitialized read in encode breakout Change I3430820 performed an uninitialized read when encode_breakout == 0, since the sum and sse wouldn't be set: if(x->encode_breakout) VARIANCE_INVOKE(..., get16x16var)(..., &sum, &sse); if (cpi->active_map_enabled && x->active_ptr[0] == 0) { ... } else if (sse < x->encode_breakout) Change-Id: I915eb76d1227b4b6d1137a0dedf2c143860098a2
2010-12-09 21:59:22 +01:00
change the threshold of DC check for encode breakout Previously, the DC check is to make sure there is no code-able DC shift for quantizer Q0, which has been verified rather conservative. This commit changes the criteria to have two components, DC and AC, to address the conservativeness. First, it checks if all AC energy is enough to contribute a single non-zero quantized AC coefficient. Second, for DC, the decision to skip further considers two possible scenarios: 1. There is no code-able 2nd order DC coefficient at all; 2 The residue is relatively flat, but the uniform DC change is very small, i.e. less than 1/2 gray level per pixel. Comparing to previous criteria, the new criteria is about 10% to 15% faster in encoding time with a very small quality loss. (threshold ~1000 and quality range 33db-45db) It should be noted that this commit enables "automatic" static threshold for encodebreakout if a non-zero small value is passed in to encoder. Change-Id: I0f77719a1ac2c2dfddbd950d84920df374515ce3
2011-01-20 01:21:01 +01:00
if (sse < threshold)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
remove redundant functions The encoder defined about 4 set of similar functions to calculate sum, variance or sse or a combination of them. This commit removed one set of these functions, get8x8var and get16x16var, where calls to the later function are replaced with var16x16 by using the fact on a 16x16 MB: variance == sse - sum*sum/256 Change-Id: I803eabd1fb3ab177780a40338cbd596dffaed267
2011-06-07 01:42:58 +02:00
unsigned int q2dc = xd->block[24].dequant[0];
change the threshold of DC check for encode breakout Previously, the DC check is to make sure there is no code-able DC shift for quantizer Q0, which has been verified rather conservative. This commit changes the criteria to have two components, DC and AC, to address the conservativeness. First, it checks if all AC energy is enough to contribute a single non-zero quantized AC coefficient. Second, for DC, the decision to skip further considers two possible scenarios: 1. There is no code-able 2nd order DC coefficient at all; 2 The residue is relatively flat, but the uniform DC change is very small, i.e. less than 1/2 gray level per pixel. Comparing to previous criteria, the new criteria is about 10% to 15% faster in encoding time with a very small quality loss. (threshold ~1000 and quality range 33db-45db) It should be noted that this commit enables "automatic" static threshold for encodebreakout if a non-zero small value is passed in to encoder. Change-Id: I0f77719a1ac2c2dfddbd950d84920df374515ce3
2011-01-20 01:21:01 +01:00
/* If theres is no codeable 2nd order dc
or a very small uniform pixel change change */
remove redundant functions The encoder defined about 4 set of similar functions to calculate sum, variance or sse or a combination of them. This commit removed one set of these functions, get8x8var and get16x16var, where calls to the later function are replaced with var16x16 by using the fact on a 16x16 MB: variance == sse - sum*sum/256 Change-Id: I803eabd1fb3ab177780a40338cbd596dffaed267
2011-06-07 01:42:58 +02:00
if ((sse - var < q2dc * q2dc >>4) ||
(sse /2 > var && sse-var < 64))
vp8e - static threshold play Realized no need for new assembly code sum is already calculated. Change-Id: Ie2d94feb4b7c1f77c5359bca29b66228e41638c9
2010-12-07 22:07:23 +01:00
{
remove redundant functions The encoder defined about 4 set of similar functions to calculate sum, variance or sse or a combination of them. This commit removed one set of these functions, get8x8var and get16x16var, where calls to the later function are replaced with var16x16 by using the fact on a 16x16 MB: variance == sse - sum*sum/256 Change-Id: I803eabd1fb3ab177780a40338cbd596dffaed267
2011-06-07 01:42:58 +02:00
// Check u and v to make sure skip is ok
int sse2= VP8_UVSSE(x, IF_RTCD(&cpi->rtcd.variance));
change the threshold of DC check for encode breakout Previously, the DC check is to make sure there is no code-able DC shift for quantizer Q0, which has been verified rather conservative. This commit changes the criteria to have two components, DC and AC, to address the conservativeness. First, it checks if all AC energy is enough to contribute a single non-zero quantized AC coefficient. Second, for DC, the decision to skip further considers two possible scenarios: 1. There is no code-able 2nd order DC coefficient at all; 2 The residue is relatively flat, but the uniform DC change is very small, i.e. less than 1/2 gray level per pixel. Comparing to previous criteria, the new criteria is about 10% to 15% faster in encoding time with a very small quality loss. (threshold ~1000 and quality range 33db-45db) It should be noted that this commit enables "automatic" static threshold for encodebreakout if a non-zero small value is passed in to encoder. Change-Id: I0f77719a1ac2c2dfddbd950d84920df374515ce3
2011-01-20 01:21:01 +01:00
if (sse2 * 2 < threshold)
fix uninitialized read in encode breakout Change I3430820 performed an uninitialized read when encode_breakout == 0, since the sum and sse wouldn't be set: if(x->encode_breakout) VARIANCE_INVOKE(..., get16x16var)(..., &sum, &sse); if (cpi->active_map_enabled && x->active_ptr[0] == 0) { ... } else if (sse < x->encode_breakout) Change-Id: I915eb76d1227b4b6d1137a0dedf2c143860098a2
2010-12-09 21:59:22 +01:00
{
x->skip = 1;
Adjusted breakout RD for SPLITMV vp8_rd_pick_best_mbsegmentation looks at y only. The new breakout does not include the frame cost, the prob_skip_false cost, or the uv rate. Performance improved by a few percent and the quality remained the same. Change-Id: I94ff013998ac51e8ecce7130870f7b6600758e15
2010-12-16 15:38:02 +01:00
distortion2 = sse + sse2;
fix uninitialized read in encode breakout Change I3430820 performed an uninitialized read when encode_breakout == 0, since the sum and sse wouldn't be set: if(x->encode_breakout) VARIANCE_INVOKE(..., get16x16var)(..., &sum, &sse); if (cpi->active_map_enabled && x->active_ptr[0] == 0) { ... } else if (sse < x->encode_breakout) Change-Id: I915eb76d1227b4b6d1137a0dedf2c143860098a2
2010-12-09 21:59:22 +01:00
rate2 = 500;
Adjusted breakout RD for SPLITMV vp8_rd_pick_best_mbsegmentation looks at y only. The new breakout does not include the frame cost, the prob_skip_false cost, or the uv rate. Performance improved by a few percent and the quality remained the same. Change-Id: I94ff013998ac51e8ecce7130870f7b6600758e15
2010-12-16 15:38:02 +01:00
/* for best_yrd calculation */
rate_uv = 0;
distortion_uv = sse2;
fix uninitialized read in encode breakout Change I3430820 performed an uninitialized read when encode_breakout == 0, since the sum and sse wouldn't be set: if(x->encode_breakout) VARIANCE_INVOKE(..., get16x16var)(..., &sum, &sse); if (cpi->active_map_enabled && x->active_ptr[0] == 0) { ... } else if (sse < x->encode_breakout) Change-Id: I915eb76d1227b4b6d1137a0dedf2c143860098a2
2010-12-09 21:59:22 +01:00
disable_skip = 1;
Use sub-pixel search's SSE in mode selection Passed SSE from sub-pixel search back to pick_inter_mode function, which is compared with the encode_breakout to see if we could skip evaluating the remaining modes. Change-Id: I4a86442834f0d1b880a19e21ea52d17d505f941d
2011-04-18 21:48:34 +02:00
this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
fix uninitialized read in encode breakout Change I3430820 performed an uninitialized read when encode_breakout == 0, since the sum and sse wouldn't be set: if(x->encode_breakout) VARIANCE_INVOKE(..., get16x16var)(..., &sum, &sse); if (cpi->active_map_enabled && x->active_ptr[0] == 0) { ... } else if (sse < x->encode_breakout) Change-Id: I915eb76d1227b4b6d1137a0dedf2c143860098a2
2010-12-09 21:59:22 +01:00
break;
}
vp8e - static threshold play Realized no need for new assembly code sum is already calculated. Change-Id: Ie2d94feb4b7c1f77c5359bca29b66228e41638c9
2010-12-07 22:07:23 +01:00
}
Initial WebM release
2010-05-18 17:58:33 +02:00
}
}
fix uninitialized read in encode breakout Change I3430820 performed an uninitialized read when encode_breakout == 0, since the sum and sse wouldn't be set: if(x->encode_breakout) VARIANCE_INVOKE(..., get16x16var)(..., &sum, &sse); if (cpi->active_map_enabled && x->active_ptr[0] == 0) { ... } else if (sse < x->encode_breakout) Change-Id: I915eb76d1227b4b6d1137a0dedf2c143860098a2
2010-12-09 21:59:22 +01:00
Initial WebM release
2010-05-18 17:58:33 +02:00
//intermodecost[mode_index] = vp8_cost_mv_ref(this_mode, mdcounts); // Experimental debug code
// Add in the Mv/mode cost
made vp8_mode_context adaptive vp8_mode_contexts[] is an entropy table used to code inter mode choices. It was a fixed constant table. This commit made the entropy context adaptive. Tests on derf set showed very good consistent gains on all metrics: avg psnr .47%, overall psnr .46% and ssim .40%. http://www.corp.google.com/~yaowu/no_crawl/newModeContext.html Change-Id: Ia62b14485c948e2b74586118619c5eb2068b43b2
2011-12-06 21:03:42 +01:00
rate2 += vp8_cost_mv_ref(&cpi->common, this_mode, mdcounts);
Initial WebM release
2010-05-18 17:58:33 +02:00
// Y cost and distortion
Adjusted breakout RD for SPLITMV vp8_rd_pick_best_mbsegmentation looks at y only. The new breakout does not include the frame cost, the prob_skip_false cost, or the uv rate. Performance improved by a few percent and the quality remained the same. Change-Id: I94ff013998ac51e8ecce7130870f7b6600758e15
2010-12-16 15:38:02 +01:00
macro_block_yrd(x, &rate_y, &distortion, IF_RTCD(&cpi->rtcd.encodemb));
rate2 += rate_y;
Initial WebM release
2010-05-18 17:58:33 +02:00
distortion2 += distortion;
// UV cost and distortion
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
vp8_build_inter16x16_predictors_mbuv(&x->e_mbd);
Removed bmi copy to/from BLOCKD for SPLITMV and B_PRED modes. Modified code to use the bmi found in mode_info_context instead of BLOCKD. On the decode side, the uvmvs are calculated only when required, instead of every macroblock. This is WIP. (bmi should eventually be removed from BLOCKD) Small performance gains noticed for RT encodes and decodes.(VGA) Change-Id: I2ed7f0fd5ca733655df684aa82da575c77a973e7
2011-08-24 20:42:26 +02:00
rd_inter16x16_uv(cpi, x, &rate_uv, &distortion_uv, cpi->common.full_pixel);
Adjusted breakout RD for SPLITMV vp8_rd_pick_best_mbsegmentation looks at y only. The new breakout does not include the frame cost, the prob_skip_false cost, or the uv rate. Performance improved by a few percent and the quality remained the same. Change-Id: I94ff013998ac51e8ecce7130870f7b6600758e15
2010-12-16 15:38:02 +01:00
rate2 += rate_uv;
distortion2 += distortion_uv;
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
mode_excluded = cpi->common.dual_pred_mode == DUAL_PREDICTION_ONLY;
#endif /* CONFIG_DUALPRED */
Initial WebM release
2010-05-18 17:58:33 +02:00
break;
default:
break;
}
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
else /* x->e_mbd.mode_info_context->mbmi.second_ref_frame != 0 */
{
int ref1 = x->e_mbd.mode_info_context->mbmi.ref_frame;
int ref2 = x->e_mbd.mode_info_context->mbmi.second_ref_frame;
mode_excluded = cpi->common.dual_pred_mode == SINGLE_PREDICTION_ONLY;
switch (this_mode)
{
case NEWMV:
if (mc_search_result[ref1].as_int == INVALID_MV ||
mc_search_result[ref2].as_int == INVALID_MV)
continue;
x->e_mbd.mode_info_context->mbmi.mv.as_int = mc_search_result[ref1].as_int;
x->e_mbd.mode_info_context->mbmi.second_mv.as_int = mc_search_result[ref2].as_int;
rate2 += vp8_mv_bit_cost(&mc_search_result[ref1],
&frame_best_ref_mv[ref1], x->mvcost, 96);
rate2 += vp8_mv_bit_cost(&mc_search_result[ref2],
&frame_best_ref_mv[ref2], x->mvcost, 96);
break;
case ZEROMV:
x->e_mbd.mode_info_context->mbmi.mv.as_int = 0;
x->e_mbd.mode_info_context->mbmi.second_mv.as_int = 0;
break;
case NEARMV:
if (frame_near_mv[ref1].as_int == 0 || frame_near_mv[ref2].as_int == 0)
continue;
x->e_mbd.mode_info_context->mbmi.mv.as_int = frame_near_mv[ref1].as_int;
x->e_mbd.mode_info_context->mbmi.second_mv.as_int = frame_near_mv[ref2].as_int;
break;
case NEARESTMV:
if (frame_nearest_mv[ref1].as_int == 0 || frame_nearest_mv[ref2].as_int == 0)
continue;
x->e_mbd.mode_info_context->mbmi.mv.as_int = frame_nearest_mv[ref1].as_int;
x->e_mbd.mode_info_context->mbmi.second_mv.as_int = frame_nearest_mv[ref2].as_int;
break;
default:
break;
}
/* Add in the Mv/mode cost */
made vp8_mode_context adaptive vp8_mode_contexts[] is an entropy table used to code inter mode choices. It was a fixed constant table. This commit made the entropy context adaptive. Tests on derf set showed very good consistent gains on all metrics: avg psnr .47%, overall psnr .46% and ssim .40%. http://www.corp.google.com/~yaowu/no_crawl/newModeContext.html Change-Id: Ia62b14485c948e2b74586118619c5eb2068b43b2
2011-12-06 21:03:42 +01:00
rate2 += vp8_cost_mv_ref(&cpi->common,this_mode, mdcounts);
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
vp8_clamp_mv2(&x->e_mbd.mode_info_context->mbmi.mv, xd);
vp8_clamp_mv2(&x->e_mbd.mode_info_context->mbmi.second_mv, xd);
if (((x->e_mbd.mode_info_context->mbmi.mv.as_mv.row >> 3) < x->mv_row_min) ||
((x->e_mbd.mode_info_context->mbmi.mv.as_mv.row >> 3) > x->mv_row_max) ||
((x->e_mbd.mode_info_context->mbmi.mv.as_mv.col >> 3) < x->mv_col_min) ||
((x->e_mbd.mode_info_context->mbmi.mv.as_mv.col >> 3) > x->mv_col_max) ||
((x->e_mbd.mode_info_context->mbmi.second_mv.as_mv.row >> 3) < x->mv_row_min) ||
((x->e_mbd.mode_info_context->mbmi.second_mv.as_mv.row >> 3) > x->mv_row_max) ||
((x->e_mbd.mode_info_context->mbmi.second_mv.as_mv.col >> 3) < x->mv_col_min) ||
((x->e_mbd.mode_info_context->mbmi.second_mv.as_mv.col >> 3) > x->mv_col_max))
continue;
/* build first and second prediction */
vp8_build_inter16x16_predictors_mby(&x->e_mbd);
vp8_build_inter16x16_predictors_mbuv(&x->e_mbd);
/* do second round and average the results */
x->e_mbd.second_pre.y_buffer = y_buffer[ref2];
x->e_mbd.second_pre.u_buffer = u_buffer[ref2];
x->e_mbd.second_pre.v_buffer = v_buffer[ref2];
vp8_build_2nd_inter16x16_predictors_mb(&x->e_mbd, x->e_mbd.predictor,
&x->e_mbd.predictor[256],
&x->e_mbd.predictor[320], 16, 8);
/* Y cost and distortion */
macro_block_yrd(x, &rate_y, &distortion, IF_RTCD(&cpi->rtcd.encodemb));
rate2 += rate_y;
distortion2 += distortion;
/* UV cost and distortion */
rd_inter16x16_uv(cpi, x, &rate_uv, &distortion_uv, cpi->common.full_pixel);
rate2 += rate_uv;
distortion2 += distortion_uv;
/* don't bother w/ skip, we would never have come here if skip were enabled */
x->e_mbd.mode_info_context->mbmi.mode = this_mode;
/* We don't include the cost of the second reference here, because there are only
* three options: Last/Golden, ARF/Last or Golden/ARF, or in other words if you
* present them in that order, the second one is always known if the first is known */
Minor fixes: fixed issues caused by conflicts between two experiments. Change-Id: I56a9bd69493e4850c121ea057a6233c55777c2a5
2011-12-07 14:32:45 +01:00
{
MB_MODE_INFO *t = &x->e_mbd.mode_info_context[-cpi->common.mode_info_stride].mbmi;
MB_MODE_INFO *l = &x->e_mbd.mode_info_context[-1].mbmi;
int cnt = (t->second_ref_frame != INTRA_FRAME) + (l->second_ref_frame != INTRA_FRAME);
dualmode_cost = vp8_cost_bit(cpi->prob_dualpred[cnt], 1);
}
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
}
#endif /* CONFIG_DUALPRED */
Initial WebM release
2010-05-18 17:58:33 +02:00
Adjusted breakout RD for SPLITMV vp8_rd_pick_best_mbsegmentation looks at y only. The new breakout does not include the frame cost, the prob_skip_false cost, or the uv rate. Performance improved by a few percent and the quality remained the same. Change-Id: I94ff013998ac51e8ecce7130870f7b6600758e15
2010-12-16 15:38:02 +01:00
// Where skip is allowable add in the default per mb cost for the no skip case.
// where we then decide to skip we have to delete this and replace it with the
// cost of signallying a skip
if (cpi->common.mb_no_coeff_skip)
{
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
int prob_skip_cost = vp8_cost_bit(cpi->prob_skip_false, 0);
other_cost += prob_skip_cost;
rate2 += prob_skip_cost;
Adjusted breakout RD for SPLITMV vp8_rd_pick_best_mbsegmentation looks at y only. The new breakout does not include the frame cost, the prob_skip_false cost, or the uv rate. Performance improved by a few percent and the quality remained the same. Change-Id: I94ff013998ac51e8ecce7130870f7b6600758e15
2010-12-16 15:38:02 +01:00
}
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
if (cpi->common.dual_pred_mode == HYBRID_PREDICTION)
{
rate2 += dualmode_cost;
}
#endif /* CONFIG_DUALPRED */
Calc ref_frame_cost once per frame instead of every macro block. Change-Id: I2604e94c6b89e3a8457777e21c8c38406d55b165
2011-06-13 15:50:54 +02:00
/* Estimate the reference frame signaling cost and add it
* to the rolling cost variable.
*/
rate2 +=
x->e_mbd.ref_frame_cost[x->e_mbd.mode_info_context->mbmi.ref_frame];
Adjusted breakout RD for SPLITMV vp8_rd_pick_best_mbsegmentation looks at y only. The new breakout does not include the frame cost, the prob_skip_false cost, or the uv rate. Performance improved by a few percent and the quality remained the same. Change-Id: I94ff013998ac51e8ecce7130870f7b6600758e15
2010-12-16 15:38:02 +01:00
Initial WebM release
2010-05-18 17:58:33 +02:00
if (!disable_skip)
{
// Test for the condition where skip block will be activated because there are no non zero coefficients and make any necessary adjustment for rate
if (cpi->common.mb_no_coeff_skip)
{
Adjusted breakout RD for SPLITMV vp8_rd_pick_best_mbsegmentation looks at y only. The new breakout does not include the frame cost, the prob_skip_false cost, or the uv rate. Performance improved by a few percent and the quality remained the same. Change-Id: I94ff013998ac51e8ecce7130870f7b6600758e15
2010-12-16 15:38:02 +01:00
int tteob;
Initial WebM release
2010-05-18 17:58:33 +02:00
tteob = 0;
for (i = 0; i <= 24; i++)
{
tteob += x->e_mbd.block[i].eob;
}
if (tteob == 0)
{
rate2 -= (rate_y + rate_uv);
Adjusted breakout RD for SPLITMV vp8_rd_pick_best_mbsegmentation looks at y only. The new breakout does not include the frame cost, the prob_skip_false cost, or the uv rate. Performance improved by a few percent and the quality remained the same. Change-Id: I94ff013998ac51e8ecce7130870f7b6600758e15
2010-12-16 15:38:02 +01:00
//for best_yrd calculation
rate_uv = 0;
Initial WebM release
2010-05-18 17:58:33 +02:00
// Back out no skip flag costing and add in skip flag costing
if (cpi->prob_skip_false)
{
Adjusted breakout RD for SPLITMV vp8_rd_pick_best_mbsegmentation looks at y only. The new breakout does not include the frame cost, the prob_skip_false cost, or the uv rate. Performance improved by a few percent and the quality remained the same. Change-Id: I94ff013998ac51e8ecce7130870f7b6600758e15
2010-12-16 15:38:02 +01:00
int prob_skip_cost;
Initial WebM release
2010-05-18 17:58:33 +02:00
Adjusted breakout RD for SPLITMV vp8_rd_pick_best_mbsegmentation looks at y only. The new breakout does not include the frame cost, the prob_skip_false cost, or the uv rate. Performance improved by a few percent and the quality remained the same. Change-Id: I94ff013998ac51e8ecce7130870f7b6600758e15
2010-12-16 15:38:02 +01:00
prob_skip_cost = vp8_cost_bit(cpi->prob_skip_false, 1);
prob_skip_cost -= vp8_cost_bit(cpi->prob_skip_false, 0);
rate2 += prob_skip_cost;
other_cost += prob_skip_cost;
Initial WebM release
2010-05-18 17:58:33 +02:00
}
}
}
// Calculate the final RD estimate for this mode
Removed cpi->target_bits_per_mb cpi->target_bits_per_mb is currently not being used, so delete it. Also removed other unused code in rdopt.c. Change-Id: I98449f9030bcd2f15451d9b7a3b9b93dd1409923
2011-01-07 15:41:13 +01:00
this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
Initial WebM release
2010-05-18 17:58:33 +02:00
}
// Experimental debug code.
//all_rds[mode_index] = this_rd;
//all_rates[mode_index] = rate2;
//all_dist[mode_index] = distortion2;
Error accumulator stats bug. The error accumulator stats values cpi->prediction_error and cpi->intra_error were being populated with rd values not distortion values. These are only "currently" used in a limited way for RT compress key frame detection. Change-Id: I2702ba1cab6e49ab8dc096ba75b6b34ab3573021
2011-04-08 15:21:36 +02:00
// Keep record of best intra distortion
if ((x->e_mbd.mode_info_context->mbmi.ref_frame == INTRA_FRAME) &&
(this_rd < best_intra_rd) )
Initial WebM release
2010-05-18 17:58:33 +02:00
{
Error accumulator stats bug. The error accumulator stats values cpi->prediction_error and cpi->intra_error were being populated with rd values not distortion values. These are only "currently" used in a limited way for RT compress key frame detection. Change-Id: I2702ba1cab6e49ab8dc096ba75b6b34ab3573021
2011-04-08 15:21:36 +02:00
best_intra_rd = this_rd;
*returnintra = distortion2 ;
Initial WebM release
2010-05-18 17:58:33 +02:00
}
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
if (!disable_skip &&
(this_mode == SPLITMV || x->e_mbd.mode_info_context->mbmi.ref_frame == INTRA_FRAME))
{
if (this_rd < best_dual_rd)
best_dual_rd = this_rd;
if (this_rd < best_single_rd)
best_single_rd = this_rd;
if (this_rd < best_hybrid_rd)
best_hybrid_rd = this_rd;
}
#endif /* CONFIG_DUALPRED */
Initial WebM release
2010-05-18 17:58:33 +02:00
// Did this mode help.. i.i is it the new best mode
if (this_rd < best_rd || x->skip)
{
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
if (!mode_excluded)
{
#endif /* CONFIG_DUALPRED */
Initial WebM release
2010-05-18 17:58:33 +02:00
// Note index of best mode so far
best_mode_index = mode_index;
if (this_mode <= B_PRED)
{
Removed unnecessary MB_MODE_INFO copies These copies occurred for each macroblock in the encoder and decoder. Thetemp MB_MODE_INFO mbmi was removed from MACROBLOCKD. As a result, a large number compile errors had to be fixed. Change-Id: I4cf0ffae3ce244f6db04a4c217d52dd256382cf3
2010-08-12 22:25:43 +02:00
x->e_mbd.mode_info_context->mbmi.uv_mode = uv_intra_mode;
MODE_INFO size reduction Declared the bmi in MODE_INFO as a union instead of B_MODE_INFO. This reduced the memory footprint by 518,400 bytes for 1080 resolutions. The decoder performance improved by ~4% for the clip used and the encoder showed very small improvements. (0.5%) This reduction was first mentioned to me by John K. and in a later discussion by Yaowu. This is WIP. Change-Id: I8e175fdbc46d28c35277302a04bee4540efc8d29
2011-05-24 19:24:52 +02:00
/* required for left and above block mv */
x->e_mbd.mode_info_context->mbmi.mv.as_int = 0;
Initial WebM release
2010-05-18 17:58:33 +02:00
}
Calc ref_frame_cost once per frame instead of every macro block. Change-Id: I2604e94c6b89e3a8457777e21c8c38406d55b165
2011-06-13 15:50:54 +02:00
other_cost +=
x->e_mbd.ref_frame_cost[x->e_mbd.mode_info_context->mbmi.ref_frame];
Adjusted breakout RD for SPLITMV vp8_rd_pick_best_mbsegmentation looks at y only. The new breakout does not include the frame cost, the prob_skip_false cost, or the uv rate. Performance improved by a few percent and the quality remained the same. Change-Id: I94ff013998ac51e8ecce7130870f7b6600758e15
2010-12-16 15:38:02 +01:00
/* Calculate the final y RD estimate for this mode */
Removed cpi->target_bits_per_mb cpi->target_bits_per_mb is currently not being used, so delete it. Also removed other unused code in rdopt.c. Change-Id: I98449f9030bcd2f15451d9b7a3b9b93dd1409923
2011-01-07 15:41:13 +01:00
best_yrd = RDCOST(x->rdmult, x->rddiv, (rate2-rate_uv-other_cost),
(distortion2-distortion_uv));
Adjusted breakout RD for SPLITMV vp8_rd_pick_best_mbsegmentation looks at y only. The new breakout does not include the frame cost, the prob_skip_false cost, or the uv rate. Performance improved by a few percent and the quality remained the same. Change-Id: I94ff013998ac51e8ecce7130870f7b6600758e15
2010-12-16 15:38:02 +01:00
Initial WebM release
2010-05-18 17:58:33 +02:00
*returnrate = rate2;
*returndistortion = distortion2;
best_rd = this_rd;
Removed unnecessary MB_MODE_INFO copies These copies occurred for each macroblock in the encoder and decoder. Thetemp MB_MODE_INFO mbmi was removed from MACROBLOCKD. As a result, a large number compile errors had to be fixed. Change-Id: I4cf0ffae3ce244f6db04a4c217d52dd256382cf3
2010-08-12 22:25:43 +02:00
vpx_memcpy(&best_mbmode, &x->e_mbd.mode_info_context->mbmi, sizeof(MB_MODE_INFO));
Reduced the size of MB_MODE_INFO Moved partition_bmi and partition_count out of MB_MODE_INFO and placed into MACROBLOCK. Also reduced the size of other members of the MB_MODE_INFO struct. For 1080p, the memory was reduced by 1,209,516 bytes. The decoder performance appeared to improve by 3% for the clip used. Note: The main goal for this change is to improve the decoder performance. The encoder will be revisited at a later date for further structure cleanup. Change-Id: I4733621292ee9cc3fffa4046cb3fd4d99bd14613
2010-09-02 22:17:52 +02:00
vpx_memcpy(&best_partition, x->partition_info, sizeof(PARTITION_INFO));
Initial WebM release
2010-05-18 17:58:33 +02:00
enabled 8x8 intra prediction modes on inter frames This commit enabled the usage of 8x8 intra prediction modes on inter frames. There are a few TODO items related to this: 1)baseline entropy need be calibrated; 2)cost of UV need to be done more properly rather than using decision only relying on Y; 3)Threshold for allowing picking 8x8 intra prediction should be lowered to lower than the B_PRED. Even with all the TODOs, tests showed consistent gain on derf set ~0.1% (PSNR:0.08% and SSIM:0.14%). It is assumed that 8x8 intra prediction will help more on large resolution clips, especially with above TODOs addressed. Change-Id: I398ada49dfc32575cfab962a569c2885111ae3ba
2011-12-01 01:25:00 +01:00
if ((this_mode == B_PRED)
||(this_mode == I8X8_PRED)
|| (this_mode == SPLITMV))
Removed B_MODE_INFO Declared the bmi in BLOCKD as a union instead of B_MODE_INFO. Then removed B_MODE_INFO completely. Change-Id: Ieb7469899e265892c66f7aeac87b7f2bf38e7a67
2011-06-02 19:46:41 +02:00
for (i = 0; i < 16; i++)
{
best_bmodes[i] = x->e_mbd.block[i].bmi;
}
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
}
#endif /* CONFIG_DUALPRED */
Initial WebM release
2010-05-18 17:58:33 +02:00
// Testing this mode gave rise to an improvement in best error score. Lower threshold a bit for next time
cpi->rd_thresh_mult[mode_index] = (cpi->rd_thresh_mult[mode_index] >= (MIN_THRESHMULT + 2)) ? cpi->rd_thresh_mult[mode_index] - 2 : MIN_THRESHMULT;
cpi->rd_threshes[mode_index] = (cpi->rd_baseline_thresh[mode_index] >> 7) * cpi->rd_thresh_mult[mode_index];
}
// If the mode did not help improve the best error case then raise the threshold for testing that mode next time around.
else
{
cpi->rd_thresh_mult[mode_index] += 4;
if (cpi->rd_thresh_mult[mode_index] > MAX_THRESHMULT)
cpi->rd_thresh_mult[mode_index] = MAX_THRESHMULT;
cpi->rd_threshes[mode_index] = (cpi->rd_baseline_thresh[mode_index] >> 7) * cpi->rd_thresh_mult[mode_index];
}
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
/* keep record of best dual/single-only prediction */
if (!disable_skip &&
x->e_mbd.mode_info_context->mbmi.ref_frame != INTRA_FRAME &&
this_mode != SPLITMV)
{
int single_rd, hybrid_rd, single_rate, hybrid_rate;
if (cpi->common.dual_pred_mode == HYBRID_PREDICTION)
{
single_rate = rate2 - dualmode_cost;
hybrid_rate = rate2;
}
else
{
single_rate = rate2;
hybrid_rate = rate2 + dualmode_cost;
}
single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
if (x->e_mbd.mode_info_context->mbmi.second_ref_frame == INTRA_FRAME &&
single_rd < best_single_rd)
{
best_single_rd = single_rd;
if (0) printf("single rd [DMC: %d]: %d\n", dualmode_cost, single_rd);
}
else if (x->e_mbd.mode_info_context->mbmi.second_ref_frame != INTRA_FRAME &&
single_rd < best_dual_rd)
{
best_dual_rd = single_rd;
if (0) printf("dual rd [DMC: %d]: %d\n", dualmode_cost, single_rd);
}
if (hybrid_rd < best_hybrid_rd)
{
best_hybrid_rd = hybrid_rd;
if (0) printf("hybrid rd [DMC: %d]: %d\n", best_hybrid_rd, hybrid_rd);
}
}
#endif /* CONFIG_DUALPRED */
Initial WebM release
2010-05-18 17:58:33 +02:00
if (x->skip)
break;
change the threshold of DC check for encode breakout Previously, the DC check is to make sure there is no code-able DC shift for quantizer Q0, which has been verified rather conservative. This commit changes the criteria to have two components, DC and AC, to address the conservativeness. First, it checks if all AC energy is enough to contribute a single non-zero quantized AC coefficient. Second, for DC, the decision to skip further considers two possible scenarios: 1. There is no code-able 2nd order DC coefficient at all; 2 The residue is relatively flat, but the uniform DC change is very small, i.e. less than 1/2 gray level per pixel. Comparing to previous criteria, the new criteria is about 10% to 15% faster in encoding time with a very small quality loss. (threshold ~1000 and quality range 33db-45db) It should be noted that this commit enables "automatic" static threshold for encodebreakout if a non-zero small value is passed in to encoder. Change-Id: I0f77719a1ac2c2dfddbd950d84920df374515ce3
2011-01-20 01:21:01 +01:00
Initial WebM release
2010-05-18 17:58:33 +02:00
}
// Reduce the activation RD thresholds for the best choice mode
if ((cpi->rd_baseline_thresh[best_mode_index] > 0) && (cpi->rd_baseline_thresh[best_mode_index] < (INT_MAX >> 2)))
{
int best_adjustment = (cpi->rd_thresh_mult[best_mode_index] >> 2);
cpi->rd_thresh_mult[best_mode_index] = (cpi->rd_thresh_mult[best_mode_index] >= (MIN_THRESHMULT + best_adjustment)) ? cpi->rd_thresh_mult[best_mode_index] - best_adjustment : MIN_THRESHMULT;
cpi->rd_threshes[best_mode_index] = (cpi->rd_baseline_thresh[best_mode_index] >> 7) * cpi->rd_thresh_mult[best_mode_index];
// If we chose a split mode then reset the new MV thresholds as well
/*if ( vp8_mode_order[best_mode_index] == SPLITMV )
{
best_adjustment = 4; //(cpi->rd_thresh_mult[THR_NEWMV] >> 4);
cpi->rd_thresh_mult[THR_NEWMV] = (cpi->rd_thresh_mult[THR_NEWMV] >= (MIN_THRESHMULT+best_adjustment)) ? cpi->rd_thresh_mult[THR_NEWMV]-best_adjustment: MIN_THRESHMULT;
cpi->rd_threshes[THR_NEWMV] = (cpi->rd_baseline_thresh[THR_NEWMV] >> 7) * cpi->rd_thresh_mult[THR_NEWMV];
best_adjustment = 4; //(cpi->rd_thresh_mult[THR_NEWG] >> 4);
cpi->rd_thresh_mult[THR_NEWG] = (cpi->rd_thresh_mult[THR_NEWG] >= (MIN_THRESHMULT+best_adjustment)) ? cpi->rd_thresh_mult[THR_NEWG]-best_adjustment: MIN_THRESHMULT;
cpi->rd_threshes[THR_NEWG] = (cpi->rd_baseline_thresh[THR_NEWG] >> 7) * cpi->rd_thresh_mult[THR_NEWG];
best_adjustment = 4; //(cpi->rd_thresh_mult[THR_NEWA] >> 4);
cpi->rd_thresh_mult[THR_NEWA] = (cpi->rd_thresh_mult[THR_NEWA] >= (MIN_THRESHMULT+best_adjustment)) ? cpi->rd_thresh_mult[THR_NEWA]-best_adjustment: MIN_THRESHMULT;
cpi->rd_threshes[THR_NEWA] = (cpi->rd_baseline_thresh[THR_NEWA] >> 7) * cpi->rd_thresh_mult[THR_NEWA];
}*/
}
// Note how often each mode chosen as best
cpi->mode_chosen_counts[best_mode_index] ++;
Segmentation feature logic fix. Bug fix to logic in vp8_pick_inter_mode() and vp8_rd_pick_inter_mode(). The block on the use of segment features for the cm->refresh_alt_ref_frame case was just for testing and is not correct. The special case code for alt ref can be re-enabled as an else clause. Change-Id: Ic9b57cdb5f04ea7737032b8fb953d84d7717b3ce
2011-11-03 18:45:27 +01:00
// This code force Altref,0,0 and skip for the frame that overlays a
// an alrtef unless Altref is filtered. However, this is unsafe if
// segment level coding of ref frame or mode is enabled for this
// segment.
if (!segfeature_active( xd, segment_id, SEG_LVL_REF_FRAME ) &&
!segfeature_active( xd, segment_id, SEG_LVL_MODE ) &&
cpi->is_src_frame_alt_ref &&
(cpi->oxcf.arnr_max_frames == 0) &&
vp8_pick_inter_mode code cleanup Small code cleanups before attempting to reduce the size of bmi found in BLOCKD. Change-Id: Ie9c14adb53afd847716a75bcce067d0e6c04f225
2011-05-31 20:24:42 +02:00
(best_mbmode.mode != ZEROMV || best_mbmode.ref_frame != ALTREF_FRAME))
Initial WebM release
2010-05-18 17:58:33 +02:00
{
vp8_pick_inter_mode code cleanup Small code cleanups before attempting to reduce the size of bmi found in BLOCKD. Change-Id: Ie9c14adb53afd847716a75bcce067d0e6c04f225
2011-05-31 20:24:42 +02:00
x->e_mbd.mode_info_context->mbmi.mode = ZEROMV;
x->e_mbd.mode_info_context->mbmi.ref_frame = ALTREF_FRAME;
Removed unnecessary MB_MODE_INFO copies These copies occurred for each macroblock in the encoder and decoder. Thetemp MB_MODE_INFO mbmi was removed from MACROBLOCKD. As a result, a large number compile errors had to be fixed. Change-Id: I4cf0ffae3ce244f6db04a4c217d52dd256382cf3
2010-08-12 22:25:43 +02:00
x->e_mbd.mode_info_context->mbmi.mv.as_int = 0;
vp8_pick_inter_mode code cleanup Small code cleanups before attempting to reduce the size of bmi found in BLOCKD. Change-Id: Ie9c14adb53afd847716a75bcce067d0e6c04f225
2011-05-31 20:24:42 +02:00
x->e_mbd.mode_info_context->mbmi.uv_mode = DC_PRED;
x->e_mbd.mode_info_context->mbmi.mb_skip_coeff =
(cpi->common.mb_no_coeff_skip) ? 1 : 0;
x->e_mbd.mode_info_context->mbmi.partitioning = 0;
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
*best_single_rd_diff = *best_dual_rd_diff = *best_hybrid_rd_diff = 0;
#endif /* CONFIG_DUALPRED */
Bugfix for error accumulator stats Previous to commit de4e9e3, there was an early return in the alt-ref case that was inadvertantly removed when the function was refactored to return void. This patch restores the prior behavior. Change-Id: I783ffd594a4690297e2742f99526fd7ad67698b2
2011-04-11 19:05:08 +02:00
return;
Initial WebM release
2010-05-18 17:58:33 +02:00
}
vp8_pick_inter_mode code cleanup Small code cleanups before attempting to reduce the size of bmi found in BLOCKD. Change-Id: Ie9c14adb53afd847716a75bcce067d0e6c04f225
2011-05-31 20:24:42 +02:00
Initial WebM release
2010-05-18 17:58:33 +02:00
// macroblock modes
Removed unnecessary MB_MODE_INFO copies These copies occurred for each macroblock in the encoder and decoder. Thetemp MB_MODE_INFO mbmi was removed from MACROBLOCKD. As a result, a large number compile errors had to be fixed. Change-Id: I4cf0ffae3ce244f6db04a4c217d52dd256382cf3
2010-08-12 22:25:43 +02:00
vpx_memcpy(&x->e_mbd.mode_info_context->mbmi, &best_mbmode, sizeof(MB_MODE_INFO));
Initial WebM release
2010-05-18 17:58:33 +02:00
MODE_INFO size reduction Declared the bmi in MODE_INFO as a union instead of B_MODE_INFO. This reduced the memory footprint by 518,400 bytes for 1080 resolutions. The decoder performance improved by ~4% for the clip used and the encoder showed very small improvements. (0.5%) This reduction was first mentioned to me by John K. and in a later discussion by Yaowu. This is WIP. Change-Id: I8e175fdbc46d28c35277302a04bee4540efc8d29
2011-05-24 19:24:52 +02:00
if (best_mbmode.mode == B_PRED)
Initial WebM release
2010-05-18 17:58:33 +02:00
{
MODE_INFO size reduction Declared the bmi in MODE_INFO as a union instead of B_MODE_INFO. This reduced the memory footprint by 518,400 bytes for 1080 resolutions. The decoder performance improved by ~4% for the clip used and the encoder showed very small improvements. (0.5%) This reduction was first mentioned to me by John K. and in a later discussion by Yaowu. This is WIP. Change-Id: I8e175fdbc46d28c35277302a04bee4540efc8d29
2011-05-24 19:24:52 +02:00
for (i = 0; i < 16; i++)
enabled 8x8 intra prediction modes on inter frames This commit enabled the usage of 8x8 intra prediction modes on inter frames. There are a few TODO items related to this: 1)baseline entropy need be calibrated; 2)cost of UV need to be done more properly rather than using decision only relying on Y; 3)Threshold for allowing picking 8x8 intra prediction should be lowered to lower than the B_PRED. Even with all the TODOs, tests showed consistent gain on derf set ~0.1% (PSNR:0.08% and SSIM:0.14%). It is assumed that 8x8 intra prediction will help more on large resolution clips, especially with above TODOs addressed. Change-Id: I398ada49dfc32575cfab962a569c2885111ae3ba
2011-12-01 01:25:00 +01:00
{
Removed bmi copy to/from BLOCKD for SPLITMV and B_PRED modes. Modified code to use the bmi found in mode_info_context instead of BLOCKD. On the decode side, the uvmvs are calculated only when required, instead of every macroblock. This is WIP. (bmi should eventually be removed from BLOCKD) Small performance gains noticed for RT encodes and decodes.(VGA) Change-Id: I2ed7f0fd5ca733655df684aa82da575c77a973e7
2011-08-24 20:42:26 +02:00
xd->mode_info_context->bmi[i].as_mode = best_bmodes[i].as_mode;
enabled 8x8 intra prediction modes on inter frames This commit enabled the usage of 8x8 intra prediction modes on inter frames. There are a few TODO items related to this: 1)baseline entropy need be calibrated; 2)cost of UV need to be done more properly rather than using decision only relying on Y; 3)Threshold for allowing picking 8x8 intra prediction should be lowered to lower than the B_PRED. Even with all the TODOs, tests showed consistent gain on derf set ~0.1% (PSNR:0.08% and SSIM:0.14%). It is assumed that 8x8 intra prediction will help more on large resolution clips, especially with above TODOs addressed. Change-Id: I398ada49dfc32575cfab962a569c2885111ae3ba
2011-12-01 01:25:00 +01:00
xd->block[i].bmi.as_mode = xd->mode_info_context->bmi[i].as_mode;
}
}
if (best_mbmode.mode == I8X8_PRED)
{
set_i8x8_block_modes(x, mode8x8);
Initial WebM release
2010-05-18 17:58:33 +02:00
}
Using partition_info instead of blockd info for splitmv The partition_info struct contains info just for SPLITMV, so it should be used instead of BLOCKD. Eventually, I want to reduce the size of B_MODE_INFO struct found in BLOCKD, so this is the first step toward that goal. Also, since SPLITMV is not supported in vp8_pick_inter_mode(), the unnecessary mem copies and checks were removed. For rt encodes, this gave a slight performance improvement. Change-Id: I5585c98fa9d5acbde1c7e0f452a01d9ecc080574
2011-05-19 21:03:36 +02:00
if (best_mbmode.mode == SPLITMV)
{
MODE_INFO size reduction Declared the bmi in MODE_INFO as a union instead of B_MODE_INFO. This reduced the memory footprint by 518,400 bytes for 1080 resolutions. The decoder performance improved by ~4% for the clip used and the encoder showed very small improvements. (0.5%) This reduction was first mentioned to me by John K. and in a later discussion by Yaowu. This is WIP. Change-Id: I8e175fdbc46d28c35277302a04bee4540efc8d29
2011-05-24 19:24:52 +02:00
for (i = 0; i < 16; i++)
Removed bmi copy to/from BLOCKD for SPLITMV and B_PRED modes. Modified code to use the bmi found in mode_info_context instead of BLOCKD. On the decode side, the uvmvs are calculated only when required, instead of every macroblock. This is WIP. (bmi should eventually be removed from BLOCKD) Small performance gains noticed for RT encodes and decodes.(VGA) Change-Id: I2ed7f0fd5ca733655df684aa82da575c77a973e7
2011-08-24 20:42:26 +02:00
xd->mode_info_context->bmi[i].mv.as_int = best_bmodes[i].mv.as_int;
MODE_INFO size reduction Declared the bmi in MODE_INFO as a union instead of B_MODE_INFO. This reduced the memory footprint by 518,400 bytes for 1080 resolutions. The decoder performance improved by ~4% for the clip used and the encoder showed very small improvements. (0.5%) This reduction was first mentioned to me by John K. and in a later discussion by Yaowu. This is WIP. Change-Id: I8e175fdbc46d28c35277302a04bee4540efc8d29
2011-05-24 19:24:52 +02:00
Using partition_info instead of blockd info for splitmv The partition_info struct contains info just for SPLITMV, so it should be used instead of BLOCKD. Eventually, I want to reduce the size of B_MODE_INFO struct found in BLOCKD, so this is the first step toward that goal. Also, since SPLITMV is not supported in vp8_pick_inter_mode(), the unnecessary mem copies and checks were removed. For rt encodes, this gave a slight performance improvement. Change-Id: I5585c98fa9d5acbde1c7e0f452a01d9ecc080574
2011-05-19 21:03:36 +02:00
vpx_memcpy(x->partition_info, &best_partition, sizeof(PARTITION_INFO));
MODE_INFO size reduction Declared the bmi in MODE_INFO as a union instead of B_MODE_INFO. This reduced the memory footprint by 518,400 bytes for 1080 resolutions. The decoder performance improved by ~4% for the clip used and the encoder showed very small improvements. (0.5%) This reduction was first mentioned to me by John K. and in a later discussion by Yaowu. This is WIP. Change-Id: I8e175fdbc46d28c35277302a04bee4540efc8d29
2011-05-24 19:24:52 +02:00
Using partition_info instead of blockd info for splitmv The partition_info struct contains info just for SPLITMV, so it should be used instead of BLOCKD. Eventually, I want to reduce the size of B_MODE_INFO struct found in BLOCKD, so this is the first step toward that goal. Also, since SPLITMV is not supported in vp8_pick_inter_mode(), the unnecessary mem copies and checks were removed. For rt encodes, this gave a slight performance improvement. Change-Id: I5585c98fa9d5acbde1c7e0f452a01d9ecc080574
2011-05-19 21:03:36 +02:00
x->e_mbd.mode_info_context->mbmi.mv.as_int =
x->partition_info->bmi[15].mv.as_int;
}
rd_update_mvcount(cpi, x, &frame_best_ref_mv[xd->mode_info_context->mbmi.ref_frame]);
Modify MVcount in pick_inter_mode to eliminate calling of vp8_find_near_mvs Moved MVcount modification in pick_inter_mode, and eliminated calling of vp8_find_near_mvs. Change-Id: Icd47448a1dfc8fdf526f86757d0e5a7f218cb5e8
2011-05-13 16:56:45 +02:00
Dual 16x16 inter prediction. This patch introduces the concept of dual inter16x16 prediction. A 16x16 inter-predicted macroblock can use 2 references instead of 1, where both references use the same mvmode (new, near/est, zero). In the case of newmv, this means that two MVs are coded instead of one. The frame can be encoded in 3 ways: all MBs single-prediction, all MBs dual prediction, or per-MB single/dual prediction selection ("hybrid"), in which case a single bit is coded per-MB to indicate whether the MB uses single or dual inter prediction. In the future, we can (maybe?) get further gains by mixing this with Adrian's 32x32 work, per-segment dual prediction settings, or adding support for dual splitmv/8x8mv inter prediction. Gain (on derf-set, CQ mode) is ~2.8% (SSIM) or ~3.6% (glb PSNR). Most gain is at medium/high bitrates, but there's minor gains at low bitrates also. Output was confirmed to match between encoder and decoder. Note for optimization people: this patch introduces a 2nd version of 16x16/8x8 sixtap/bilin functions, which does an avg instead of a store. They may want to look and make sure this is implemented to their satisfaction so we can optimize it best in the future. Change-ID: I59dc84b07cbb3ccf073ac0f756d03d294cb19281
2011-12-06 20:53:02 +01:00
#if CONFIG_DUALPRED
*best_single_rd_diff = best_rd - best_single_rd;
*best_dual_rd_diff = best_rd - best_dual_rd;
*best_hybrid_rd_diff = best_rd - best_hybrid_rd;
#endif /* CONFIG_DUALPRED */
Initial WebM release
2010-05-18 17:58:33 +02:00
}
Move RD intra block mode selection to rdopt.c This change is analogous to I0b67dae1f8a74902378da7bdf565e39ab832dda7, which made the move for the non-RD path. Change-Id: If63fc1b0cd1eb7f932e710f83ff24d91454f8ed1
2011-06-08 18:05:05 +02:00
void vp8_rd_pick_intra_mode(VP8_COMP *cpi, MACROBLOCK *x, int *rate_)
{
CONFIG_T8X8 experiment.: Block the selection of 4x4 modes in key frames if 8x8 is selected. Change-Id: Ie5729ec22a999d9a1996f020bd4b941e29514992
2011-11-18 20:44:57 +01:00
MACROBLOCKD *xd = &x->e_mbd;
Move RD intra block mode selection to rdopt.c This change is analogous to I0b67dae1f8a74902378da7bdf565e39ab832dda7, which made the move for the non-RD path. Change-Id: If63fc1b0cd1eb7f932e710f83ff24d91454f8ed1
2011-06-08 18:05:05 +02:00
int error4x4, error16x16;
int rate4x4, rate16x16 = 0, rateuv;
int dist4x4, dist16x16, distuv;
int rate;
int rate4x4_tokenonly = 0;
int rate16x16_tokenonly = 0;
int rateuv_tokenonly = 0;
add 8x8 intra prediction modes Patch 1 to Patch 3 is an initial implementation of 8x8 intra prediction modes, here are with the following assumptions: a. 8x8 has 4 prediction modes DC, H, V and TM b. UV 4x4 block use the same mode as corresponding 8x8 area c. i8x8 modes are enabled for key frame only for now Patch 4: d. removed debug code from previous patches Patch 5: e. added stats code to collect entropy stats and further cleaned up Patch 6: f. changed mode stats code to collect finer stats of modes Patch 7: g. normalized i8x8 modes distribution to total at 256 (8bits). Patch 8: h. fixed a bug in decoder and removed debug printf output. Patch 9: i. more cleanups to address paul's comment Patch 10: j. messy rebase/merges to bring the commit up to date. Tests on HD clips encoded with all key frame showing consistent gain on all clips and all metrics:~0.5%(psnr) and 0.6%(ssim): http://www.corp.google.com/~yaowu/no_crawl/i8x8hd_allkey_fixedq.html To build and test, configure with: --enable-experimental --enable-i8x8 Change-Id: I9813fe07ae48cab5fdb5d904bca022514ad01e7f
2011-08-05 01:30:27 +02:00
int error8x8, rate8x8_tokenonly=0;
int rate8x8, dist8x8;
int mode16x16;
int mode8x8[4];
Move RD intra block mode selection to rdopt.c This change is analogous to I0b67dae1f8a74902378da7bdf565e39ab832dda7, which made the move for the non-RD path. Change-Id: If63fc1b0cd1eb7f932e710f83ff24d91454f8ed1
2011-06-08 18:05:05 +02:00
x->e_mbd.mode_info_context->mbmi.ref_frame = INTRA_FRAME;
rd_pick_intra_mbuv_mode(cpi, x, &rateuv, &rateuv_tokenonly, &distuv);
rate = rateuv;
error16x16 = rd_pick_intra16x16mby_mode(cpi, x,
&rate16x16, &rate16x16_tokenonly,
&dist16x16);
add 8x8 intra prediction modes Patch 1 to Patch 3 is an initial implementation of 8x8 intra prediction modes, here are with the following assumptions: a. 8x8 has 4 prediction modes DC, H, V and TM b. UV 4x4 block use the same mode as corresponding 8x8 area c. i8x8 modes are enabled for key frame only for now Patch 4: d. removed debug code from previous patches Patch 5: e. added stats code to collect entropy stats and further cleaned up Patch 6: f. changed mode stats code to collect finer stats of modes Patch 7: g. normalized i8x8 modes distribution to total at 256 (8bits). Patch 8: h. fixed a bug in decoder and removed debug printf output. Patch 9: i. more cleanups to address paul's comment Patch 10: j. messy rebase/merges to bring the commit up to date. Tests on HD clips encoded with all key frame showing consistent gain on all clips and all metrics:~0.5%(psnr) and 0.6%(ssim): http://www.corp.google.com/~yaowu/no_crawl/i8x8hd_allkey_fixedq.html To build and test, configure with: --enable-experimental --enable-i8x8 Change-Id: I9813fe07ae48cab5fdb5d904bca022514ad01e7f
2011-08-05 01:30:27 +02:00
mode16x16 = x->e_mbd.mode_info_context->mbmi.mode;
error8x8 = rd_pick_intra8x8mby_modes(cpi, x,
&rate8x8, &rate8x8_tokenonly,
&dist8x8, error16x16);
mode8x8[0]= x->e_mbd.mode_info_context->bmi[0].as_mode;
mode8x8[1]= x->e_mbd.mode_info_context->bmi[2].as_mode;
mode8x8[2]= x->e_mbd.mode_info_context->bmi[8].as_mode;
mode8x8[3]= x->e_mbd.mode_info_context->bmi[10].as_mode;
CONFIG_T8X8 experiment.: Block the selection of 4x4 modes in key frames if 8x8 is selected. Change-Id: Ie5729ec22a999d9a1996f020bd4b941e29514992
2011-11-18 20:44:57 +01:00
#if CONFIG_T8X8
if ( get_seg_tx_type( xd,
xd->mode_info_context->mbmi.segment_id ) == TX_4X4 )
{
error4x4 = rd_pick_intra4x4mby_modes(cpi, x,
&rate4x4, &rate4x4_tokenonly,
&dist4x4, error16x16);
}
else
{
error4x4 = INT_MAX;
}
#else
Move RD intra block mode selection to rdopt.c This change is analogous to I0b67dae1f8a74902378da7bdf565e39ab832dda7, which made the move for the non-RD path. Change-Id: If63fc1b0cd1eb7f932e710f83ff24d91454f8ed1
2011-06-08 18:05:05 +02:00
error4x4 = rd_pick_intra4x4mby_modes(cpi, x,
&rate4x4, &rate4x4_tokenonly,
&dist4x4, error16x16);
CONFIG_T8X8 experiment.: Block the selection of 4x4 modes in key frames if 8x8 is selected. Change-Id: Ie5729ec22a999d9a1996f020bd4b941e29514992
2011-11-18 20:44:57 +01:00
#endif
Move RD intra block mode selection to rdopt.c This change is analogous to I0b67dae1f8a74902378da7bdf565e39ab832dda7, which made the move for the non-RD path. Change-Id: If63fc1b0cd1eb7f932e710f83ff24d91454f8ed1
2011-06-08 18:05:05 +02:00
add 8x8 intra prediction modes Patch 1 to Patch 3 is an initial implementation of 8x8 intra prediction modes, here are with the following assumptions: a. 8x8 has 4 prediction modes DC, H, V and TM b. UV 4x4 block use the same mode as corresponding 8x8 area c. i8x8 modes are enabled for key frame only for now Patch 4: d. removed debug code from previous patches Patch 5: e. added stats code to collect entropy stats and further cleaned up Patch 6: f. changed mode stats code to collect finer stats of modes Patch 7: g. normalized i8x8 modes distribution to total at 256 (8bits). Patch 8: h. fixed a bug in decoder and removed debug printf output. Patch 9: i. more cleanups to address paul's comment Patch 10: j. messy rebase/merges to bring the commit up to date. Tests on HD clips encoded with all key frame showing consistent gain on all clips and all metrics:~0.5%(psnr) and 0.6%(ssim): http://www.corp.google.com/~yaowu/no_crawl/i8x8hd_allkey_fixedq.html To build and test, configure with: --enable-experimental --enable-i8x8 Change-Id: I9813fe07ae48cab5fdb5d904bca022514ad01e7f
2011-08-05 01:30:27 +02:00
if(error8x8> error16x16)
{
Removed #if CONFIG_I8X8 This commit removed the macro CONFIG_I8X8, which was used to indicate the 8x8 intra prediction experiment, made the change fully merged in. Change-Id: Iafa4443781ce6e83f5591c12ba615a0e92ce0ea0
2011-12-07 22:03:57 +01:00
if (error4x4 < error16x16)
{
x->e_mbd.mode_info_context->mbmi.mode = B_PRED;
rate += rate4x4;
}
else
{
x->e_mbd.mode_info_context->mbmi.mode = mode16x16;
rate += rate16x16;
add 8x8 intra prediction modes Patch 1 to Patch 3 is an initial implementation of 8x8 intra prediction modes, here are with the following assumptions: a. 8x8 has 4 prediction modes DC, H, V and TM b. UV 4x4 block use the same mode as corresponding 8x8 area c. i8x8 modes are enabled for key frame only for now Patch 4: d. removed debug code from previous patches Patch 5: e. added stats code to collect entropy stats and further cleaned up Patch 6: f. changed mode stats code to collect finer stats of modes Patch 7: g. normalized i8x8 modes distribution to total at 256 (8bits). Patch 8: h. fixed a bug in decoder and removed debug printf output. Patch 9: i. more cleanups to address paul's comment Patch 10: j. messy rebase/merges to bring the commit up to date. Tests on HD clips encoded with all key frame showing consistent gain on all clips and all metrics:~0.5%(psnr) and 0.6%(ssim): http://www.corp.google.com/~yaowu/no_crawl/i8x8hd_allkey_fixedq.html To build and test, configure with: --enable-experimental --enable-i8x8 Change-Id: I9813fe07ae48cab5fdb5d904bca022514ad01e7f
2011-08-05 01:30:27 +02:00
Removed #if CONFIG_I8X8 This commit removed the macro CONFIG_I8X8, which was used to indicate the 8x8 intra prediction experiment, made the change fully merged in. Change-Id: Iafa4443781ce6e83f5591c12ba615a0e92ce0ea0
2011-12-07 22:03:57 +01:00
}
add 8x8 intra prediction modes Patch 1 to Patch 3 is an initial implementation of 8x8 intra prediction modes, here are with the following assumptions: a. 8x8 has 4 prediction modes DC, H, V and TM b. UV 4x4 block use the same mode as corresponding 8x8 area c. i8x8 modes are enabled for key frame only for now Patch 4: d. removed debug code from previous patches Patch 5: e. added stats code to collect entropy stats and further cleaned up Patch 6: f. changed mode stats code to collect finer stats of modes Patch 7: g. normalized i8x8 modes distribution to total at 256 (8bits). Patch 8: h. fixed a bug in decoder and removed debug printf output. Patch 9: i. more cleanups to address paul's comment Patch 10: j. messy rebase/merges to bring the commit up to date. Tests on HD clips encoded with all key frame showing consistent gain on all clips and all metrics:~0.5%(psnr) and 0.6%(ssim): http://www.corp.google.com/~yaowu/no_crawl/i8x8hd_allkey_fixedq.html To build and test, configure with: --enable-experimental --enable-i8x8 Change-Id: I9813fe07ae48cab5fdb5d904bca022514ad01e7f
2011-08-05 01:30:27 +02:00
}
else
{
if (error4x4 < error8x8)
{
x->e_mbd.mode_info_context->mbmi.mode = B_PRED;
rate += rate4x4;
}
else
{
Move RD intra block mode selection to rdopt.c This change is analogous to I0b67dae1f8a74902378da7bdf565e39ab832dda7, which made the move for the non-RD path. Change-Id: If63fc1b0cd1eb7f932e710f83ff24d91454f8ed1
2011-06-08 18:05:05 +02:00
add 8x8 intra prediction modes Patch 1 to Patch 3 is an initial implementation of 8x8 intra prediction modes, here are with the following assumptions: a. 8x8 has 4 prediction modes DC, H, V and TM b. UV 4x4 block use the same mode as corresponding 8x8 area c. i8x8 modes are enabled for key frame only for now Patch 4: d. removed debug code from previous patches Patch 5: e. added stats code to collect entropy stats and further cleaned up Patch 6: f. changed mode stats code to collect finer stats of modes Patch 7: g. normalized i8x8 modes distribution to total at 256 (8bits). Patch 8: h. fixed a bug in decoder and removed debug printf output. Patch 9: i. more cleanups to address paul's comment Patch 10: j. messy rebase/merges to bring the commit up to date. Tests on HD clips encoded with all key frame showing consistent gain on all clips and all metrics:~0.5%(psnr) and 0.6%(ssim): http://www.corp.google.com/~yaowu/no_crawl/i8x8hd_allkey_fixedq.html To build and test, configure with: --enable-experimental --enable-i8x8 Change-Id: I9813fe07ae48cab5fdb5d904bca022514ad01e7f
2011-08-05 01:30:27 +02:00
x->e_mbd.mode_info_context->mbmi.mode = I8X8_PRED;
set_i8x8_block_modes(x, mode8x8);
rate += rate8x8;
}
}
Move RD intra block mode selection to rdopt.c This change is analogous to I0b67dae1f8a74902378da7bdf565e39ab832dda7, which made the move for the non-RD path. Change-Id: If63fc1b0cd1eb7f932e710f83ff24d91454f8ed1
2011-06-08 18:05:05 +02:00
*rate_ = rate;
}
Reference in New Issue Copy Permalink