]> granicus.if.org Git - libvpx/log
libvpx
13 years agoFirst pass loop bug.
Paul Wilkins [Wed, 19 Jan 2011 15:50:22 +0000 (15:50 +0000)]
First pass loop bug.

Incorrect value loop_decay_rate used in GF loop.

The intent was to test the  cumulative value decay_accumulator.

Change-Id: I62928c63eb09f4f6936a45ebd1c23784d1c9681b

13 years agoMerge "fix last frame buffer copy logic regression"
John Koleszar [Tue, 18 Jan 2011 20:54:57 +0000 (12:54 -0800)]
Merge "fix last frame buffer copy logic regression"

13 years agovp8e -removed undefined max call
Jim Bankoski [Tue, 18 Jan 2011 16:14:17 +0000 (11:14 -0500)]
vp8e -removed undefined max call

Change-Id: I42a86b0488f44115f09551fc5ad6d711fd470f0d

13 years agoMerge "Further CQ, Key frame and ARF changes"
Paul Wilkins [Tue, 18 Jan 2011 16:04:46 +0000 (08:04 -0800)]
Merge "Further CQ, Key frame and ARF changes"

13 years agoFurther CQ, Key frame and ARF changes
Paul Wilkins [Tue, 18 Jan 2011 15:19:05 +0000 (15:19 +0000)]
Further CQ, Key frame and ARF changes

This code fixes a bug in the calculation of
the minimum Q for alt ref frames.

It also allows an extended gf/arf interval for sections
of clips that completely static (or nearly so).

Change-Id: I1a21aaa16d4f0578e5f99b13bebd78d59403c73b

13 years agoFix encoder real-time only configuration.
Attila Nagy [Mon, 10 Jan 2011 09:14:10 +0000 (11:14 +0200)]
Fix encoder real-time only configuration.

Remove allocation/deallocation of stats storage.
Remove full search functions in machine specific encoder inits.
Remove last pass validation in  validate_config.

Change-Id: I7f29be69273981a4fef6e80ecdb6217c68cbad4e

13 years agoFix CQ range and experimental KF sizing changes.
Paul Wilkins [Mon, 17 Jan 2011 17:23:11 +0000 (17:23 +0000)]
Fix CQ range and experimental KF sizing changes.

The CQ level was not using the q_trans[] array to convert
to a 0-127 range as per min and maxq

Experimental change to try and match the reconstruction
error for forced key frames approximately to that of the
previous frame by means of the recode loop. Though this
may cause extra recodes and the recode behavior has not
been optimized, it can only happen on forced key frames.

Change-Id: I1f7e42d526f1b1cb556dd461eff1a692bd1b5b2f

13 years agoMerge "KF/GF Pulsing"
Paul Wilkins [Fri, 14 Jan 2011 17:20:37 +0000 (09:20 -0800)]
Merge "KF/GF Pulsing"

13 years agoMerge "Testing of modes with Alt Ref frame"
Paul Wilkins [Fri, 14 Jan 2011 15:26:37 +0000 (07:26 -0800)]
Merge "Testing of modes with Alt Ref frame"

13 years agoMerge "Experimental change to help with ARNR problem."
Paul Wilkins [Fri, 14 Jan 2011 15:26:01 +0000 (07:26 -0800)]
Merge "Experimental change to help with ARNR problem."

13 years agoTesting of modes with Alt Ref frame
Paul Wilkins [Fri, 14 Jan 2011 14:52:15 +0000 (14:52 +0000)]
Testing of modes with Alt Ref frame

Previously when a frame was being overlaid on a previously coded
alt ref frame we only checked the alt ref 0,0 mode. Where there is
a possibility that the alt ref buffer is a filtered frame we should allow
the other prediction modes as normal or at the least allow use of
the last frame buffer.

Change-Id: I4d6227223d125c96b4f3066ec6ec9484fee7768c

13 years agoARNR filter pointer update bug fix
Adrian Grange [Fri, 14 Jan 2011 15:04:39 +0000 (15:04 +0000)]
ARNR filter pointer update bug fix

In cases where the frame width is not a multiple of 16 the
ARNR filter would go wrong.

In vp8_temporal_filter_iterate_c when updating pointers
at the end of a row of MBs,  the image size was
incorrectly used rather than using Num_MBs_In_Row
times 16 (Y) or 8 (U,V).

This worked when width is multiple of 16 but failed
otherwise.

Change-Id: I008919062715bd3d17c7aa2562ab58d1cb37053a

13 years agoExperimental change to help with ARNR problem.
Paul Wilkins [Fri, 14 Jan 2011 12:07:12 +0000 (12:07 +0000)]
Experimental change to help with ARNR problem.

Allow use of other reference frames for the ARF overlay frame
when ARNR filtering is enabled

Change-Id: Icd6a9fb38977a88fbe7cc9b9c18198eb454c0273

13 years agoKF/GF Pulsing
Paul Wilkins [Fri, 14 Jan 2011 11:34:53 +0000 (11:34 +0000)]
KF/GF Pulsing

This change is designed to try and reduce pulsing effects when moving
with a complex transition like a fade, into an easy or static section in
an otherwise difficult clip in CQ mode.

The active CQ level is relaxed down to the user entered level for frames that
are generating less than the passed in minimum bandwidth.

Change-Id: Id6d8b551daad4f489c087bd742bc95418a95f3f0

13 years agoMerge "Moved ref frame calculations"
Scott LaVarnway [Thu, 13 Jan 2011 14:59:28 +0000 (06:59 -0800)]
Merge "Moved ref frame calculations"

13 years agoOne pass rate control correction.
Paul Wilkins [Thu, 13 Jan 2011 12:45:00 +0000 (12:45 +0000)]
One pass rate control correction.

Fixed discrepancy cpi->ni_frames vs cm->current_video_frame > 150.

Make one pass path explicit.

There is still scope for some odd behaviour around the transition
point at cpi->ni_frames > 150.

Change-Id: Icdee130fe6e2a832206d30e45bf65963edd7a74d

13 years agoLimit key frame quantizer for forced key frames.
Paul Wilkins [Wed, 12 Jan 2011 17:08:42 +0000 (17:08 +0000)]
Limit key frame quantizer for forced key frames.

Where a key frame occurs because of a minimum interval
selected by the user, then these forced key frames ideally need
to be more closely matched in quality to the surrounding frame.

Change-Id: Ia55b1f047e77dc7fbd78379c45869554f25b3df7

13 years agoMoved ref frame calculations
Scott LaVarnway [Tue, 11 Jan 2011 20:00:00 +0000 (15:00 -0500)]
Moved ref frame calculations

Moved ref frame calculations to outside of the
mode_index loop.

Change-Id: I06103fc7e8af88b54b84443acf6691d29b1272ac

13 years agoMerge "Add no_skip_block4x4_search flag in SPLITMV mode"
Yunqing Wang [Tue, 11 Jan 2011 16:34:24 +0000 (08:34 -0800)]
Merge "Add no_skip_block4x4_search flag in SPLITMV mode"

13 years agoMerge "use unaligned load"
Johann [Tue, 11 Jan 2011 16:25:22 +0000 (08:25 -0800)]
Merge "use unaligned load"

13 years agouse unaligned load
Johann [Tue, 11 Jan 2011 16:22:29 +0000 (11:22 -0500)]
use unaligned load

source buffer is not guaranteed to be aligned for odd size buffers

Change-Id: Id0b1fd40ba3bd6c994bcfada788feccd2b53c5a9

13 years agoAdd no_skip_block4x4_search flag in SPLITMV mode
Yunqing Wang [Mon, 10 Jan 2011 22:21:55 +0000 (17:21 -0500)]
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

13 years agoRemove unused local variables
Henrik Lundin [Tue, 11 Jan 2011 12:01:24 +0000 (13:01 +0100)]
Remove unused local variables

Removing unused local variables causing compiler warnings in
Visual Studio.

Change-Id: I0e2096303be1fdbc01428a6e57cca9796bb32c8a

13 years agoFix bug in motion search
Yunqing Wang [Mon, 10 Jan 2011 21:16:59 +0000 (16:16 -0500)]
Fix bug in motion search

The maximum possible MV in 1/8 pel units is (1<<11), which could
cause mvcost out of its range that is 1023. Change maximum
possible MV in 1/8 pel units to (1<<11)-8 will fix this problem.

Change-Id: I5788ed1de773f66658c14f225fb4ab5b1679b74b

13 years agoTwo Pass VBR change
Paul Wilkins [Mon, 10 Jan 2011 16:41:53 +0000 (16:41 +0000)]
Two Pass VBR change

Further experiment with restriction of the Q range.

This uses the average non KF/GF/ARF quantizer,  instead
of just relying on the initial value. It is not such a strong constraint
but there may be a reduced risk of rate misses.

Change-Id: I424fe782a37a2f4e18c70805e240db55bfaa25ec

13 years agoRevert BASE_ERRPERMB
Paul Wilkins [Mon, 10 Jan 2011 16:02:51 +0000 (16:02 +0000)]
Revert BASE_ERRPERMB

Constant value reverted pending more tests
on different video formats.

Change-Id: I07d11a0e0185e60724698c835416caf2e0774e61

13 years agoMerge "CQ Mode"
Paul Wilkins [Fri, 7 Jan 2011 19:05:56 +0000 (11:05 -0800)]
Merge "CQ Mode"

13 years agoCQ Mode
Paul Wilkins [Fri, 7 Jan 2011 18:29:37 +0000 (18:29 +0000)]
CQ Mode

The merge includes hooks to for CQ mode and other code
changes merged from the test branch.

CQ mode attempts to maintain a more stable quantizer within a clip
whilst also trying to adhere to a guidline maximum bitrate.

The existing target data rate parameter is used to specify the
guideline maximum bitrate.

A new parameter allows the user to specify a target CQ level.

For normal (non kf/gf/arf) frames, the quantizer will not drop BELOW the
user specified value (0-63). However, in some cases the encoder may
choose to impose a target CQ that is above that specified by the user,
if it estimates that consistent use of the target value is not compatible
with guideline maximum bitrate.

Change-Id: I2221f9eecae8cc3c431d36caf83503941b25e4c1

13 years agoMerge "Limit Q variability in two pass."
Paul Wilkins [Fri, 7 Jan 2011 17:32:29 +0000 (09:32 -0800)]
Merge "Limit Q variability in two pass."

13 years agoLimit Q variability in two pass.
Paul Wilkins [Fri, 7 Jan 2011 16:33:59 +0000 (16:33 +0000)]
Limit Q variability in two pass.

In two pass encoding each frame is given an active
Q range to work with. This change limits how much this
Q range can be altered over time from the initial estimate
made for the clip as a whole.

There is some danger this could lead to overshoot or undershoot
in some corner cases but it helps considerably in regard to
clips where either there is a glut or famine of bits in some sections,
particularly near the end of a clip.

Change-Id: I34fcd1af31d2ee3d5444f93e334645254043026e

13 years agoMerge "Disable some features for first pass."
Paul Wilkins [Fri, 7 Jan 2011 16:34:27 +0000 (08:34 -0800)]
Merge "Disable some features for first pass."

13 years agoMerge "Removed cpi->target_bits_per_mb"
Scott LaVarnway [Fri, 7 Jan 2011 14:46:45 +0000 (06:46 -0800)]
Merge "Removed cpi->target_bits_per_mb"

13 years agoRemoved cpi->target_bits_per_mb
Scott LaVarnway [Fri, 7 Jan 2011 14:41:13 +0000 (09:41 -0500)]
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

13 years agox86 sse2 temporal_filter_apply
Johann [Wed, 22 Dec 2010 16:23:51 +0000 (11:23 -0500)]
x86 sse2 temporal_filter_apply

count can be reduced to short because the max number of filtered frames
is set to 15. the max value for any frame is 32 (modifier = 16,
filter_weight = 2). 15*32 = 480 which requires 9 bits

this function goes from about 7000 us / 1000 iterations for the C code
to < 275 us / 1000 iterations for sse2 for block_size = 16 and from
about 1800 us / 1000 iters to < 100 us / 1000 iters for block_size = 8

Change-Id: I64a32607f58a2d33c39286f468b04ccd457d9e6e

13 years agofix last frame buffer copy logic regression
John Koleszar [Thu, 6 Jan 2011 18:07:39 +0000 (13:07 -0500)]
fix last frame buffer copy logic regression

Commit 0ce3901 introduced a change in the frame buffer copy logic where
the NEW frame could be copied to the ARF or GF buffer through the
copy_buffer_to_{arf,gf}==1 flags, if the LAST frame was not being
refreshed. This is not correct. The intent of the
copy_buffer_to_{arf,gf}==1 flag is to copy the LAST buffer. To copy the
NEW buffer, the refresh_{alt_ref,golden}_frame flag should be used.

The original buffer copy logic is fairly convoluted. For example:

    if (cm->refresh_last_frame)
    {
        vp8_swap_yv12_buffer(&cm->last_frame, &cm->new_frame);

        cm->frame_to_show = &cm->last_frame;
    }
    else
    {
        cm->frame_to_show = &cm->new_frame;
    }
    ...
    if (cm->copy_buffer_to_arf)
    {
        if (cm->copy_buffer_to_arf == 1)
        {
            if (cm->refresh_last_frame)
                vp8_yv12_copy_frame_ptr(&cm->new_frame, &cm->alt_ref_frame);
            else
                vp8_yv12_copy_frame_ptr(&cm->last_frame, &cm->alt_ref_frame);
        }
        else if (cm->copy_buffer_to_arf == 2)
            vp8_yv12_copy_frame_ptr(&cm->golden_frame, &cm->alt_ref_frame);
    }

Effectively, if refresh_last_frame, then new and last are swapped, so
when "new" is copied to ARF, it's equivalent to copying LAST to ARF. If
not refresh_last_frame, then LAST is copied to ARF. So LAST is copied to
ARF in both cases.

Commit 0ce3901 removed the first buffer swap but kept the
refresh_last_frame?new:last behavior, changing the sense since the first
swap wasn't done to the more readable refresh_last_frame?last:new, but
this logic is not correct when !refresh_last_frame.

This commit restores the correct behavior from v0.9.1 and prior. This
case is missing from the test vector set.

Change-Id: I8369fc13a37ae882e31a8a104da808a08bc8428f

13 years agoDisable some features for first pass.
Paul Wilkins [Thu, 6 Jan 2011 17:10:07 +0000 (17:10 +0000)]
Disable some features for first pass.

The following features don't make sense for the first
pass in its current form and have a significant impact on its
speed (up to 50%).

Slow quantizer, slow dct and trellis optimization.

Change-Id: Id9943f6765ffbd71fc0084ec7dfbc9d376fd6fcd

13 years agoAdjustment to boost calculation in two pass.
Paul Wilkins [Tue, 4 Jan 2011 17:55:49 +0000 (17:55 +0000)]
Adjustment to boost calculation in two pass.

Calculate a minimum intra value to be used in determining the
IIratio scores used in two pass, second pass.

This is to make sure sections that are low complexity" in the
intra domain are still boosted appropriately for KF/GF/ARF.

For now I have commented out the Q based adjustment of
KF boost.

Change-Id: I15deb09c5bd9b53180a2ddd3e5f575b2aba244b3

13 years agoFixed encoder crash when mult-threading is enabled.
Scott LaVarnway [Wed, 29 Dec 2010 21:28:24 +0000 (16:28 -0500)]
Fixed encoder crash when mult-threading is enabled.

Happens in real-time mode.  Will happen in good quality, speed 1.

Change-Id: I3e5b68827b1a5798d0431b088a709256d1ce2c95

13 years agoAlways update last_frame_type
Yunqing Wang [Wed, 29 Dec 2010 15:28:35 +0000 (10:28 -0500)]
Always update last_frame_type

Scott pointed out that last_frame_type only gets updated while
loopfilter exists. Since last_frame_type is also needed in
motion search now, it needs to be updated every frame.

Change-Id: I9203532fd67361588d4024628d9ddb8e391ad912

13 years agoMerge "Use the fast quantizer for inter mode selection"
Scott LaVarnway [Tue, 28 Dec 2010 19:56:11 +0000 (11:56 -0800)]
Merge "Use the fast quantizer for inter mode selection"

13 years agoUse the fast quantizer for inter mode selection
Scott LaVarnway [Tue, 28 Dec 2010 19:51:46 +0000 (14:51 -0500)]
Use the fast quantizer for inter mode selection

Use the fast quantizer for inter mode selection and the
regular quantizer for the rest of the encode for good quality,
speed 1.  Both performance and quality were improved.  The
quality gains will make up for the quality loss mentioned in
I9dc089007ca08129fb6c11fe7692777ebb8647b0.

Change-Id: Ia90bc9cf326a7c65d60d31fa32f6465ab6984d21

13 years agoAdjust MV borders for SPLITMV mode
Yunqing Wang [Tue, 28 Dec 2010 18:23:07 +0000 (13:23 -0500)]
Adjust MV borders for SPLITMV mode

Add limits to avoid MV going out of range.

Change-Id: I8a5deb40bf393488d29f694b5a56804d578e68b5

13 years agoMerge "Modify motion estimation for SPLITMV mode"
Yunqing Wang [Tue, 28 Dec 2010 16:12:26 +0000 (08:12 -0800)]
Merge "Modify motion estimation for SPLITMV mode"

13 years agoModify motion estimation for SPLITMV mode
Yunqing Wang [Thu, 23 Dec 2010 16:23:03 +0000 (11:23 -0500)]
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

13 years agoMerge "adjusted sad_per_bit to correlate with quantizer"
Yaowu Xu [Sun, 26 Dec 2010 21:45:37 +0000 (13:45 -0800)]
Merge "adjusted sad_per_bit to correlate with quantizer"

13 years agoadjusted sad_per_bit to correlate with quantizer
Yaowu Xu [Fri, 24 Dec 2010 03:59:12 +0000 (19:59 -0800)]
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

13 years agovpxenc stats_close() memleak fix
James Berry [Thu, 23 Dec 2010 19:47:56 +0000 (14:47 -0500)]
vpxenc stats_close() memleak fix

stats_close() was not freeing memory for
single pass runs.  It now takes in arg_passes
to determine when it should free memory.

Change-Id: I6623b7e30b76f9bf2e16008490f9b20484d03f31

13 years agoMerge "improve integer version of filter"
Johann [Thu, 23 Dec 2010 14:14:28 +0000 (06:14 -0800)]
Merge "improve integer version of filter"

13 years agoMerge "temporal filter naming changes"
Johann [Thu, 23 Dec 2010 14:14:20 +0000 (06:14 -0800)]
Merge "temporal filter naming changes"

13 years agoMerge "abstract apply_temporal_filter"
Johann [Thu, 23 Dec 2010 14:14:07 +0000 (06:14 -0800)]
Merge "abstract apply_temporal_filter"

13 years agoMerge "make yasm generate cv8 debug data on win32"
John Koleszar [Wed, 22 Dec 2010 19:11:08 +0000 (11:11 -0800)]
Merge "make yasm generate cv8 debug data on win32"

13 years agomake yasm generate cv8 debug data on win32
John Koleszar [Fri, 17 Dec 2010 20:06:23 +0000 (15:06 -0500)]
make yasm generate cv8 debug data on win32

Native Windows targets should use CV8 format debugging symbols, not
DWARF.

Change-Id: I9489163fcd9d749b72f6c70ecbce67a6f0790802

13 years agoimprove integer version of filter
Johann [Wed, 22 Dec 2010 16:15:56 +0000 (11:15 -0500)]
improve integer version of filter

the lookup table is based on floating point calculations (see source)

by moving the *3 before the downshift and adding the rounding bit, the
delta (LUT - integer) goes from:
______________________________________
__ 1__ 1______________________________
__ 1__ 1______________________________
____ 1______ 1________________________
____ 1 2__ 2 1________________________
______ 1 1 2__ 2__ 2__ 2 1 1__________
________ 1 1 2 2__ 1 2 3 1 2__ 2__ 2__
to:
__-1__-1______________________________
______________________________________
____-1______-1________________________
______________________________________
________-1______________-1____________
______________________________________

it's important to be able to use the integer version because the LUT
more or less precludes SIMD optimizations

Change-Id: I45a81127dc7b72a06fba951649135d9d918386c0

13 years agotemporal filter naming changes
Johann [Tue, 30 Nov 2010 15:23:43 +0000 (10:23 -0500)]
temporal filter naming changes

be more consistant with the naming pattern, especially wrt rtcd

Change-Id: I3df50686a09f1dab0a9620b5adbb8a1577b40f2f

13 years agoabstract apply_temporal_filter
Johann [Mon, 29 Nov 2010 19:21:11 +0000 (14:21 -0500)]
abstract apply_temporal_filter

allow for optimized versions of apply_temporal_filter
(now vp8_apply_temporal_filter_c)

the function was previously declared as static and appears to have been
inlined. with this change, that's no longer possible. performance takes
a small hit.

the declaration for vp8_cx_temp_filter_c was moved to onyx_if.c because
of a circular dependency. for rtcd, temporal_filter.h holds the
definition for the rtcd table, so it needs to be included by onyx_int.h.
however, onyx_int.h holds the definition for VP8_COMP which is needed
for the function prototype. blah.

Change-Id: I499c055fdc652ac4659c21c5a55fe10ceb7e95e3

13 years agoMerge "Add psnr/ssim tuning option"
Jim Bankoski [Mon, 20 Dec 2010 17:32:13 +0000 (09:32 -0800)]
Merge "Add psnr/ssim tuning option"

13 years agopropagate user private data on decode
John Koleszar [Fri, 17 Dec 2010 16:34:02 +0000 (11:34 -0500)]
propagate user private data on decode

The pointer passed in the user_priv argument to vpx_codec_decode()
should be propagated through to the corresponding output frame and
made available in the image's user_priv member. Fixes issue #252

Change-Id: I182746a6882c8549fb146b4a4fdb64f1789eb750

13 years agoMerge "Inform caller of decoder about updated references"
John Koleszar [Fri, 17 Dec 2010 15:08:21 +0000 (07:08 -0800)]
Merge "Inform caller of decoder about updated references"

13 years agoAdd psnr/ssim tuning option
John Koleszar [Fri, 17 Dec 2010 14:43:39 +0000 (09:43 -0500)]
Add psnr/ssim tuning option

Add a new encoder control, VP8E_SET_TUNING, to allow the application
to inform the encoder that the material will benefit from certain
tuning. Expose this control as the --tune option to vpxenc. The args
helper is expanded to support enumerated arguments by name or value.

Two tunings are provided by this patch, PSNR (default) and SSIM.
Activity masking is made dependent on setting --tune=ssim, as the
current implementation hurts speed (10%) and PSNR (2.7% avg,
10% peak) too much for it to be a default yet.

Change-Id: I110d969381c4805347ff5a0ffaf1a14ca1965257

13 years agoInform caller of decoder about updated references
Henrik Lundin [Tue, 14 Dec 2010 13:05:06 +0000 (14:05 +0100)]
Inform caller of decoder about updated references

Inform the caller of the decoder if a decoded frame updated last,
golden, or altref frames, required for realtime communication
proposed in document VP8 RTP payload format.

Added a new vpx_codec_control called VP8D_GET_LAST_REF_UPDATES, to be
called after vpx_codec_decode. The control will indicate which of the
reference frames that were updated by setting the 3 LSBs in the input
int (pointer).

Change-Id: Iac9db60dac414356c7ffa0b0fede88cb91e11bd7

13 years agoChanged segmentation check order
Scott LaVarnway [Thu, 16 Dec 2010 22:01:27 +0000 (17:01 -0500)]
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

13 years agoAdjusted breakout RD for SPLITMV
Scott LaVarnway [Thu, 16 Dec 2010 14:38:02 +0000 (09:38 -0500)]
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

13 years agoMerge "Fix a bug in motion search code(2)"
Yunqing Wang [Wed, 15 Dec 2010 16:10:34 +0000 (08:10 -0800)]
Merge "Fix a bug in motion search code(2)"

13 years agoFix a bug in motion search code(2)
Yunqing Wang [Tue, 14 Dec 2010 22:39:25 +0000 (17:39 -0500)]
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

13 years agoMerge "fix a bug that "optimize" flag is not set for sub-threads"
Yaowu Xu [Tue, 14 Dec 2010 21:32:04 +0000 (13:32 -0800)]
Merge "fix a bug that "optimize" flag is not set for sub-threads"

13 years agoMerge "Fix a bug in motion search code"
Yunqing Wang [Tue, 14 Dec 2010 21:25:34 +0000 (13:25 -0800)]
Merge "Fix a bug in motion search code"

13 years agoFix a bug in motion search code
Yunqing Wang [Tue, 14 Dec 2010 16:00:25 +0000 (11:00 -0500)]
Fix a bug in motion search code

The MV's range is 256. Since the new motion search uses a different
starting MV than the center ref MV, a MV range checking needs to
be done to avoid corruption.

Change-Id: I8ae0721d1bd203639e13891e2e54a2e87276f306

13 years agofix a bug that "optimize" flag is not set for sub-threads
Yaowu Xu [Tue, 14 Dec 2010 18:12:21 +0000 (10:12 -0800)]
fix a bug that "optimize" flag is not set for sub-threads

The flag for quantization optimization was not properly propagated to
mb row encoding threads.

Change-Id: Ic561599c35acd94cd5698c9b314bccd596ac2deb

13 years agoshrink TOKENEXTRA and vp8_extra_bit_struct
Johann [Tue, 14 Dec 2010 14:35:18 +0000 (09:35 -0500)]
shrink TOKENEXTRA and vp8_extra_bit_struct

Per John's previous change, shrink TOKENEXTRA from 20 to 8 bytes
original: b7b1e6fb
reverted: 41f4458a

Also drop unused field from vp8_extra_bit_struct

Update ARM ASM to deal with this change. In particular, Extra is signed
and needs to be sign-extended when loaded.

Change-Id: Ibd0ddc058432bc7bb09222d6ce4ef77e93a30b41

13 years agoRevert "Reduce size of TOKENEXTRA struct"
John Koleszar [Mon, 13 Dec 2010 22:11:48 +0000 (17:11 -0500)]
Revert "Reduce size of TOKENEXTRA struct"

This reverts commit b7b1e6fb55c6b12ccd078a20cb9855f6734931b5. Previous
fix is incomplete, breaks ARM. Itchy submit finger.

Change-Id: I939dc0d3bf4173cf951c1d152338ab6ea2184bb9

13 years agoMerge "remove unused temporal preproc code"
John Koleszar [Mon, 13 Dec 2010 21:57:59 +0000 (13:57 -0800)]
Merge "remove unused temporal preproc code"

13 years agoMerge "Reduce size of TOKENEXTRA struct"
John Koleszar [Mon, 13 Dec 2010 21:57:55 +0000 (13:57 -0800)]
Merge "Reduce size of TOKENEXTRA struct"

13 years agoremove unused temporal preproc code
John Koleszar [Mon, 13 Dec 2010 21:47:34 +0000 (16:47 -0500)]
remove unused temporal preproc code

This code is unused, as the current preproc implementation uses the
same spatial filter that postproc uses.

Change-Id: Ia06d5664917d67283f279e2480016bebed602ea7

13 years agoReduce size of TOKENEXTRA struct
John Koleszar [Mon, 13 Dec 2010 15:11:42 +0000 (10:11 -0500)]
Reduce size of TOKENEXTRA struct

Change the size of structure elements to reduce memory utilization.
Removed the 'section' member entirely, as it is set but never read.

Change-Id: Iad043830392fb4168cb3cd6075fb0eb70c7f691c

13 years agofixed vpxenc bug where ivf files would be read incorrectly
James Berry [Mon, 13 Dec 2010 18:10:58 +0000 (13:10 -0500)]
fixed vpxenc bug where ivf files would be read incorrectly

read_frame would incorrectly insert detect->buf into img
for ivf files.  detect->position now set to 4 if input file is
detected to be ivf in file_is_ivf to keep this from occuring.

Change-Id: I5e235dd3033985bc62707a35c13af5984620208e

13 years agofix a bug in multithreaded encoding with active_map enabled
Yaowu Xu [Fri, 10 Dec 2010 18:48:30 +0000 (10:48 -0800)]
fix a bug in multithreaded encoding with active_map enabled

Added the initialization of the pointer to active map. Also added the
same logic for cyclic refresh in mbrow encoding threads.

Change-Id: Ic48d0849dc706b27fba72d07dcc498075725663d

13 years agoMerge "vp8 fast quantizer sse2 optimizations for eob."
Fritz Koenig [Fri, 10 Dec 2010 17:25:04 +0000 (09:25 -0800)]
Merge "vp8 fast quantizer sse2 optimizations for eob."

13 years agovp8 fast quantizer sse2 optimizations for eob.
Fritz Koenig [Thu, 9 Dec 2010 23:00:30 +0000 (15:00 -0800)]
vp8 fast quantizer sse2 optimizations for eob.

Changed the end of block computation to use pmaxw.  Removed
additional pushing and popping of registers that was not needed.

Change-Id: I08cb9b424513cd8a2c7ad8cea53b4e2adc66ef98

13 years agofix uninitialized read in encode breakout
John Koleszar [Thu, 9 Dec 2010 20:59:22 +0000 (15:59 -0500)]
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

13 years agoCorrect q_low and q_high limits for the recode loop
Paul Wilkins [Thu, 9 Dec 2010 15:02:04 +0000 (15:02 +0000)]
Correct q_low and q_high limits for the recode loop

Corrected the initial Q range limits for the recode loop
to reflect the current allowed range for the frame.

In experimental work on constrained quality this bug was
causing unnecessary recodes.

Change-Id: I7e256fbfa681293b0223fe21ec329933d76c229f

13 years agoMerge "vp8e - static threshold play"
Yaowu Xu [Wed, 8 Dec 2010 21:08:04 +0000 (13:08 -0800)]
Merge "vp8e - static threshold play"

13 years agoMerge "vp8e - remove unnecessary variance calc"
Yaowu Xu [Wed, 8 Dec 2010 17:19:22 +0000 (09:19 -0800)]
Merge "vp8e - remove unnecessary variance calc"

13 years agovp8e - static threshold play
Jim Bankoski [Tue, 7 Dec 2010 21:07:23 +0000 (16:07 -0500)]
vp8e - static threshold play

Realized no need for new assembly code sum is already
calculated.

Change-Id: Ie2d94feb4b7c1f77c5359bca29b66228e41638c9

13 years agoMerge "vp8_rd_pick_best_mbsegmentation code restructure"
Scott LaVarnway [Tue, 7 Dec 2010 15:53:12 +0000 (07:53 -0800)]
Merge "vp8_rd_pick_best_mbsegmentation code restructure"

13 years agoMerge "adjust RDMULT for UV plane in quantization RDO"
Yaowu Xu [Tue, 7 Dec 2010 06:04:44 +0000 (22:04 -0800)]
Merge "adjust RDMULT for UV plane in quantization RDO"

13 years agoadjust RDMULT for UV plane in quantization RDO
Yaowu Xu [Mon, 6 Dec 2010 21:33:01 +0000 (13:33 -0800)]
adjust RDMULT for UV plane in quantization RDO

This patch adds a weighting factor on RDMULT for UV blocks. The change
has an overall gain about 0.5% based on ssim, between 0.1 and 0.2% by
psnr numbers.

Change-Id: I97781b077ce3bb7e34241b03268491917e8d1d72

13 years agoFix a memory leak problem in encoder
Yunqing Wang [Mon, 6 Dec 2010 22:21:37 +0000 (17:21 -0500)]
Fix a memory leak problem in encoder

Deallocating the buffers before re-allocating them.

The fix passed James Berry's test program for memory
leak check.

Change-Id: I18c3cf665412c0e313a523e3d435106c03ca438d

13 years agovp8_rd_pick_best_mbsegmentation code restructure
Scott LaVarnway [Mon, 6 Dec 2010 21:42:52 +0000 (16:42 -0500)]
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

13 years agoMerge "Improve MV prediction accuracy to achieve performance gain"
Scott LaVarnway [Mon, 6 Dec 2010 17:41:09 +0000 (09:41 -0800)]
Merge "Improve MV prediction accuracy to achieve performance gain"

13 years agoFix for manual Golden frame frequency
Patrik Westin [Tue, 16 Nov 2010 10:06:00 +0000 (11:06 +0100)]
Fix for manual Golden frame frequency

When auto_golden wasn't set it forced all frames to be a golden
frame. Now the manual configured frequency is adhered to.

Change-Id: I360acac9bc487db0d9c4d4da6ee41f70c227c539

13 years agoMerge "Change to inter_minq table."
Paul Wilkins [Sat, 4 Dec 2010 10:06:33 +0000 (02:06 -0800)]
Merge "Change to inter_minq table."

13 years agoChange to inter_minq table.
Paul Wilkins [Sat, 4 Dec 2010 10:04:12 +0000 (10:04 +0000)]
Change to inter_minq table.

The inter_minq table controls the range of quantizers available
for a particular frame in two pass relative to a max Q value.

The changes reduces the range somewhat. The effect of this
was a small increase (0.3% average) in psnr for the test set
but it should also help encode speed somewhat for higher
quality modes as it will reduce the number of iterations in the
recode loop.

The change damps the range of quantizers available locally
within a section of a clip and should therefore help keep quality
more uniform. If there is systematic overshoot or undershoot the
range can shift gradually to accommodate. However, there is
some increased risk of overshoot or undershoot against the target
bit rate in VBR mode and this risk will be more pronounced for short
clips.

The change damps the range of quantizers available locally
within a section of a clip and should therefore help keep quality
more uniform. If there is systematic overshoot or undershoot the
range can shift gradually to accommodate. However, there is
some increased risk of overshoot or undershoot against the
target bit rate in VBR mode and this risk will be more
pronounced for short clips.

Change-Id: I84465567d49ae767c6c73ff2a2aac30c895adb52

13 years agoImprove MV prediction accuracy to achieve performance gain
Yunqing Wang [Fri, 3 Dec 2010 16:26:21 +0000 (11:26 -0500)]
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

13 years agoMerge 'Add simple version of activity masking.'
John Koleszar [Fri, 3 Dec 2010 18:30:50 +0000 (13:30 -0500)]
Merge 'Add simple version of activity masking.'

Merge commit 'refs/changes/79/779/2' of
    https://review.webmproject.org/p/libvpx

Conflicts:
vp8/encoder/encodeintra.c
vp8/encoder/encodemb.c

Change-Id: Id607063fabe92d99eeb3c380e8ca670b01bfb3ef

13 years agoSet refresh_alt_ref_frame on keyframe encode.
Fritz Koenig [Thu, 18 Nov 2010 18:40:58 +0000 (10:40 -0800)]
Set refresh_alt_ref_frame on keyframe encode.

On a keyframe alt ref and golden are refreshed.  The flag was
not being set and so on the frame after a keyframe, motion
search would occur on the alt ref frame.  This is not necessary
because the alt ref frame identical to the last frame in this
scenario.

Handle corner case where a forward alt-ref frame is put
directly after a keyframe.

Change-Id: I9be4cf290d694f8cf2f9a31852014b5ccf1504d3

13 years agovp8e - remove unnecessary variance calc
Jim Bankoski [Sat, 27 Nov 2010 19:02:59 +0000 (14:02 -0500)]
vp8e - remove unnecessary variance calc

only do the variance calculation if necessary
( eg needed for breakout test)

13 years agoallow dimensions as low as 1 pixel
Pascal Massimino [Wed, 24 Nov 2010 08:22:59 +0000 (00:22 -0800)]
allow dimensions as low as 1 pixel

remove warning comment in vpxenc.c: in case of 1x1 picture,
detect_bytes will be equal to '3' and we'll fall back to
RAW_TYPE.
fix read_frame() by tracking the pre-read buffer length
in the struct detect

Change-Id: If1ed86ee5260dcdbc8f9d10da6cbb84a4cc2f151

13 years agoMerge "vpxdec: fix use of uninitialized memory for raw files"
John Koleszar [Tue, 23 Nov 2010 20:39:03 +0000 (12:39 -0800)]
Merge "vpxdec: fix use of uninitialized memory for raw files"

13 years agoMerge changes I3aed713e,I9ef7f56e,Ic18c60df
John Koleszar [Tue, 23 Nov 2010 20:38:20 +0000 (12:38 -0800)]
Merge changes I3aed713e,I9ef7f56e,Ic18c60df

* changes:
  vp8_set_maps: remove hard-coded width/height
  vp8mt_alloc_temp_buffers: make prototype return void
  Disable compile warning for ERROR macro

13 years agovpxdec: fix use of uninitialized memory for raw files
John Koleszar [Tue, 23 Nov 2010 18:40:31 +0000 (13:40 -0500)]
vpxdec: fix use of uninitialized memory for raw files

The sz member of the vpx_codec_stream_info_t structure must be
initialized when passed to vpx_codec_peek_stream_info().

Change-Id: I2d13d287d9639262b932cf44671a595fdf3c38ef