platform/upstream/libvpx.git
9 years agoRemove unnecessary dqcoeff memset.
hkuang [Thu, 11 Dec 2014 20:27:25 +0000 (12:27 -0800)]
Remove unnecessary dqcoeff memset.

dqcoeff is set to be 0 on initialization. And set back to 0 after being
used everytime.

Change-Id: I32b8e149bba40a8d707849f737a8e49a691f319c

9 years agoMerge "Refactor choose_partitioning computing scheme"
Jingning Han [Thu, 11 Dec 2014 19:14:07 +0000 (11:14 -0800)]
Merge "Refactor choose_partitioning computing scheme"

9 years agoRefactor choose_partitioning computing scheme
Jingning Han [Thu, 11 Dec 2014 17:29:36 +0000 (09:29 -0800)]
Refactor choose_partitioning computing scheme

This commit refactors the choose_partitioning function. It removes
redundant memset calls and makes the encoder to calculate
variance value per block only when it is needed. It reduces the
average runtime cost of choose_partitioning by 60%. Overall it
reduces speed -6 runtime by 2-5%.

Change-Id: I951922c50d901d0fff77a3bafc45992179bacef9

9 years agoEnable neon idct tests for intrinsics
Johann [Wed, 10 Dec 2014 23:20:22 +0000 (15:20 -0800)]
Enable neon idct tests for intrinsics

Change-Id: I45d4a22f3ecb9af172e37c95f168805e492c5493

9 years agoVP9 common for ARMv8 by using NEON intrinsics 18
James Yu [Tue, 18 Feb 2014 13:56:46 +0000 (21:56 +0800)]
VP9 common for ARMv8 by using NEON intrinsics 18

Add vp9_idct32x32_add_neon.c
- vp9_idct32x32_1024_add_neon

Change-Id: Ic598b772c28bd3487a8ead7a4598a66b25f9b00f
Signed-off-by: James Yu <james.yu@linaro.org>
9 years agoVP9 common for ARMv8 by using NEON intrinsics 14
James Yu [Fri, 7 Feb 2014 17:52:15 +0000 (01:52 +0800)]
VP9 common for ARMv8 by using NEON intrinsics 14

Add vp9_idct16x16_add_neon.c
- vp9_idct16x16_256_add_neon_pass1
- vp9_idct16x16_256_add_neon_pass2
- vp9_idct16x16_10_add_neon_pass1
- vp9_idct16x16_10_add_neon_pass2

Change-Id: I54d25b54a36f4371760f54e4036693aaea40a5de
Signed-off-by: James Yu <james.yu@linaro.org>
9 years agoVP9 common for ARMv8 by using NEON intrinsics 13
James Yu [Sat, 1 Feb 2014 06:56:06 +0000 (14:56 +0800)]
VP9 common for ARMv8 by using NEON intrinsics 13

Add vp9_idct8x8_add_neon.c
- vp9_idct8x8_64_add_neon
- vp9_idct8x8_10_add_neon

Change-Id: I6ee7b4496765aa36ed52990f2ef73e9f24459610
Signed-off-by: James Yu <james.yu@linaro.org>
9 years agoVP9 common for ARMv8 by using NEON intrinsics 12
James Yu [Sat, 1 Feb 2014 06:01:05 +0000 (14:01 +0800)]
VP9 common for ARMv8 by using NEON intrinsics 12

Add vp9_idct4x4_add_neon.c
- vp9_idct4x4_16_add_neon

Change-Id: I011a96b10f1992dbd52246019ce05bae7ca8ea4f
Signed-off-by: James Yu <james.yu@linaro.org>
9 years agoVP9 common for ARMv8 by using NEON intrinsics 11
James Yu [Fri, 31 Jan 2014 05:18:15 +0000 (13:18 +0800)]
VP9 common for ARMv8 by using NEON intrinsics 11

Add vp9_idct16x16_1_add_neon.c
- vp9_idct16x16_1_add_neon

Change-Id: I7c6524024ad4cb4e66aa38f1c887e733503c39df
Signed-off-by: James Yu <james.yu@linaro.org>
9 years agoVP9 common for ARMv8 by using NEON intrinsics 10
James Yu [Thu, 30 Jan 2014 07:26:31 +0000 (15:26 +0800)]
VP9 common for ARMv8 by using NEON intrinsics 10

Add vp9_idct32x32_1_add_neon.c
- vp9_idct32x32_1_add_neon

Change-Id: If9ffe9a857228f5c67f61dc2b428b40965816eda
Signed-off-by: James Yu <james.yu@linaro.org>
9 years agoVP9 common for ARMv8 by using NEON intrinsics 09
James Yu [Thu, 30 Jan 2014 04:26:44 +0000 (12:26 +0800)]
VP9 common for ARMv8 by using NEON intrinsics 09

Add vp9_idct8x8_1_add_neon.c
- vp9_idct8x8_1_add_neon

Change-Id: I9d23e01fa96013febbf64db6c76c6c955f14e3ff
Signed-off-by: James Yu <james.yu@linaro.org>
9 years agoVP9 common for ARMv8 by using NEON intrinsics 08
James Yu [Thu, 30 Jan 2014 03:54:35 +0000 (11:54 +0800)]
VP9 common for ARMv8 by using NEON intrinsics 08

Add vp9_idct4x4_1_add_neon.c
- vp9_idct4x4_1_add_neon

Change-Id: Ieab9af107dbd07a4f9503bc945890c90faccb8ac
Signed-off-by: James Yu <james.yu@linaro.org>
9 years agoMerge "VP9 common for ARMv8 by using NEON intrinsics 07"
Johann [Wed, 10 Dec 2014 19:40:46 +0000 (11:40 -0800)]
Merge "VP9 common for ARMv8 by using NEON intrinsics 07"

9 years agoMerge "VP9 common for ARMv8 by using NEON intrinsics 04"
Johann [Wed, 10 Dec 2014 19:40:29 +0000 (11:40 -0800)]
Merge "VP9 common for ARMv8 by using NEON intrinsics 04"

9 years agoMerge "Substantial restructuring of AQ mode 2."
Paul Wilkins [Wed, 10 Dec 2014 18:44:27 +0000 (10:44 -0800)]
Merge "Substantial restructuring of AQ mode 2."

9 years agoMerge "Use use_prev_frame_mvs flag for ref mv search branch"
Jingning Han [Wed, 10 Dec 2014 17:25:12 +0000 (09:25 -0800)]
Merge "Use use_prev_frame_mvs flag for ref mv search branch"

9 years agoMerge "Refactor update_state_rt"
Jingning Han [Wed, 10 Dec 2014 17:25:05 +0000 (09:25 -0800)]
Merge "Refactor update_state_rt"

9 years agoMerge "Make RTC coding flow support sub8x8 in key frame coding"
Jingning Han [Wed, 10 Dec 2014 17:24:56 +0000 (09:24 -0800)]
Merge "Make RTC coding flow support sub8x8 in key frame coding"

9 years agoMerge "Cosmetic naming change"
Jingning Han [Wed, 10 Dec 2014 17:05:34 +0000 (09:05 -0800)]
Merge "Cosmetic naming change"

9 years agoMerge "Take out redundant setting of mode_info from set_block_size"
Jingning Han [Wed, 10 Dec 2014 17:05:26 +0000 (09:05 -0800)]
Merge "Take out redundant setting of mode_info from set_block_size"

9 years agoMerge "Remove unused rd cost calculation from nonrd_use_partition"
Jingning Han [Wed, 10 Dec 2014 17:05:18 +0000 (09:05 -0800)]
Merge "Remove unused rd cost calculation from nonrd_use_partition"

9 years agoMerge changes I92251a8b,I5d23a685
Jim Bankoski [Wed, 10 Dec 2014 14:42:08 +0000 (06:42 -0800)]
Merge changes I92251a8b,I5d23a685

* changes:
  Adds a decode perf test that builds a new file.
  Make the decoder Cfg available to encoder tests..

9 years agoVP9 common for ARMv8 by using NEON intrinsics 07
James Yu [Wed, 29 Jan 2014 15:12:41 +0000 (23:12 +0800)]
VP9 common for ARMv8 by using NEON intrinsics 07

Add vp9_convolve8_neon.c
- vp9_convolve8_horiz_neon
- vp9_convolve8_vert_neon

Change-Id: I0bdd99ff72d275223fe211ac7243c25a5a60cf87
Signed-off-by: James Yu <james.yu@linaro.org>
9 years agoVP9 common for ARMv8 by using NEON intrinsics 04
James Yu [Sat, 25 Jan 2014 12:51:49 +0000 (20:51 +0800)]
VP9 common for ARMv8 by using NEON intrinsics 04

Add vp9_convolve8_avg_neon.c
- vp9_convolve8_avg_horiz_neon
- vp9_convolve8_avg_vert_neon

Change-Id: I617971e37b02186fec5aca181f4f9622050ea2df
Signed-off-by: James Yu <james.yu@linaro.org>
9 years agoVP9 common for ARMv8 by using NEON intrinsics 03
James Yu [Tue, 21 Jan 2014 09:23:27 +0000 (17:23 +0800)]
VP9 common for ARMv8 by using NEON intrinsics 03

Add vp9_copy_neon.c
- vp9_convolve_copy_neon

Change-Id: I291fc5423d06240876411bbceab03eae5ef585be
Signed-off-by: James Yu <james.yu@linaro.org>
9 years agoVP9 common for ARMv8 by using NEON intrinsics 02
Scott LaVarnway [Wed, 10 Sep 2014 16:49:34 +0000 (09:49 -0700)]
VP9 common for ARMv8 by using NEON intrinsics 02

Add vp9_avg_neon.c
- vp9_convolve_avg_neon

Change-Id: Id2c9d5bcfa37cff1a16417aba1656ff07bdf10fd
Signed-off-by: James Yu <james.yu@linaro.org>
9 years agoMerge "Fix clang ioc warning due to NULL src_mi pointer."
James Zern [Wed, 10 Dec 2014 02:31:46 +0000 (18:31 -0800)]
Merge "Fix clang ioc warning due to NULL src_mi pointer."

9 years agoUse use_prev_frame_mvs flag for ref mv search branch
Jingning Han [Wed, 10 Dec 2014 02:01:17 +0000 (18:01 -0800)]
Use use_prev_frame_mvs flag for ref mv search branch

Replace error_resilient flag with use_prev_frame_mvs in
vp9_pick_inter_mode reference motion vector search selection.
This effectively turns off the simplified ref mv search in the
settings of frame resizing, even if error-resilient mode is off.

Change-Id: I7fed814ee7bc0cb419a03b846e0fc2de46ba7686

9 years agoMerge "Add convolve_copy and convolve_avg to the test"
Johann [Wed, 10 Dec 2014 00:51:35 +0000 (16:51 -0800)]
Merge "Add convolve_copy and convolve_avg to the test"

9 years agoRefactor update_state_rt
Jingning Han [Tue, 9 Dec 2014 20:09:36 +0000 (12:09 -0800)]
Refactor update_state_rt

Update the frame motion vector only if previous frame motion vector
is needed for next frame reference motion vector.

Change-Id: Ica50f9d7b46ad4f815bba0d9e30f5546df29546f

9 years agoFix clang ioc warning due to NULL src_mi pointer.
hkuang [Tue, 9 Dec 2014 22:32:48 +0000 (14:32 -0800)]
Fix clang ioc warning due to NULL src_mi pointer.

The warning only happens in VP9 encoder's first pass due to src_mi
is not set up yet. But it will not fail the encoder as left_mi and
above_mi are not used in the first_pass and they will be set up again
in the second pass.

Change-Id: I12dffcd5fb1002b2b2dabb083c8726650e4b5f08

9 years agoMerge "VP9 common for ARMv8 by using NEON intrinsics 01"
Johann [Tue, 9 Dec 2014 21:41:49 +0000 (13:41 -0800)]
Merge "VP9 common for ARMv8 by using NEON intrinsics 01"

9 years agoAdd convolve_copy and convolve_avg to the test
Johann [Tue, 9 Dec 2014 20:05:15 +0000 (12:05 -0800)]
Add convolve_copy and convolve_avg to the test

Change-Id: Ic9438031282e63e627550f7e4cdeda36e43e647b

9 years agoMerge "Disable neon assembly when neon is disabled"
Johann [Tue, 9 Dec 2014 20:47:12 +0000 (12:47 -0800)]
Merge "Disable neon assembly when neon is disabled"

9 years agoAdds a decode perf test that builds a new file.
Jim Bankoski [Tue, 9 Dec 2014 20:44:45 +0000 (12:44 -0800)]
Adds a decode perf test that builds a new file.

This allows us to track decode speed for new encodes so that we catch
problems like an encode change that makes decode really slow.

Change-Id: I92251a8b1f710b241f66e1042413df1b71b76038

9 years agoVP9 common for ARMv8 by using NEON intrinsics 01
James Yu [Tue, 21 Jan 2014 01:43:29 +0000 (09:43 +0800)]
VP9 common for ARMv8 by using NEON intrinsics 01

Add vp9_loopfilter_neon.c
- vp9_lpf_horizontal_4_neon
- vp9_lpf_vertical_4_neon
- vp9_lpf_horizontal_8_neon
- vp9_lpf_vertical_8_neon

Change-Id: I97a0d7b399a431c21ee77396be3d5f5a1f7ebccb
Signed-off-by: James Yu <james.yu@linaro.org>
9 years agoMake RTC coding flow support sub8x8 in key frame coding
Jingning Han [Tue, 9 Dec 2014 19:31:45 +0000 (11:31 -0800)]
Make RTC coding flow support sub8x8 in key frame coding

This commit enables the use of sub8x8 blocks in RTC key frame
encoding. It requires the block size to be preset and will decide
the coding mode and encode the bit-stream.

Change-Id: I35aaf8ee2d4d6085432410c7963f339f85a2c19b

9 years agoDisable neon assembly when neon is disabled
Johann [Mon, 8 Dec 2014 23:13:37 +0000 (15:13 -0800)]
Disable neon assembly when neon is disabled

Change-Id: Idde266cd7287bb6bee016c90efeafa67550f94c6

9 years agoCosmetic naming change
Jingning Han [Tue, 9 Dec 2014 18:30:39 +0000 (10:30 -0800)]
Cosmetic naming change

Rename set_modeinfo_offsets as set_mode_info_offsets, to be more
consistent with naming convention.

Change-Id: I68ca1f36c4a78127d9439a50c1506a2afd07927d

9 years agoTake out redundant setting of mode_info from set_block_size
Jingning Han [Tue, 9 Dec 2014 18:24:37 +0000 (10:24 -0800)]
Take out redundant setting of mode_info from set_block_size

The later encoding process will take the top-left block's
mode_info for pre-determined block size.

Change-Id: I76a90f9ce7f3b2dbc2975b52442114e461c465b5

9 years agoMerge "Clean up the logic of handling corrupted frame."
hkuang [Tue, 9 Dec 2014 18:23:18 +0000 (10:23 -0800)]
Merge "Clean up the logic of handling corrupted frame."

9 years agoSubstantial restructuring of AQ mode 2.
Paul Wilkins [Thu, 27 Nov 2014 10:50:56 +0000 (10:50 +0000)]
Substantial restructuring of AQ mode 2.

The restructure moves the decision into the rd pick
modes loop and makes a decision based at the 16x16
block level instead of only the 64x64 level.

This gives finer granularity and better visual results
on the clips I have tested. Metrics results are worse
than the old AQ2 especially for PSNR and this mode
now falls between AQ0 and AQ1 in terms of visual
impact and metrics results.

Further tuning of this to follow.

It should be noted that if there are multiple iterations
of the recode loop the segment for a MB could change
in each loop if the previous loop causes a change in the
complexity / variance bin of the block. Also where a block
gets a delta Q this will alter the rd multiplier for this block
in subsequent recode iterations and frames where the
segmentation is applied.

Change-Id: I20256c125daa14734c16f7cc9aefab656ab808f7

9 years agoRemove unused rd cost calculation from nonrd_use_partition
Jingning Han [Tue, 9 Dec 2014 02:43:36 +0000 (18:43 -0800)]
Remove unused rd cost calculation from nonrd_use_partition

The per block rd cost calculation is not needed when partition
size is preset.

Change-Id: Ie5575248bbffb584e908aa13097f697ace6ec747

9 years agoMerge "Extend x32 check by also checking for __x86_64__."
Johann [Mon, 8 Dec 2014 22:52:31 +0000 (14:52 -0800)]
Merge "Extend x32 check by also checking for __x86_64__."

9 years agoMerge "SSSE3 Optimization for Atom processors using new instruction selection and...
Yunqing Wang [Mon, 8 Dec 2014 21:34:53 +0000 (13:34 -0800)]
Merge "SSSE3 Optimization for Atom processors using new instruction selection and ordering"

9 years agoMerge "Changes to assembler for NASM on mac."
James Zern [Mon, 8 Dec 2014 20:55:06 +0000 (12:55 -0800)]
Merge "Changes to assembler for NASM on mac."

9 years agoSSSE3 Optimization for Atom processors using new instruction selection and ordering
levytamar82 [Fri, 5 Dec 2014 18:14:33 +0000 (11:14 -0700)]
SSSE3 Optimization for Atom processors using new instruction selection and ordering

The function vp9_filter_block1d16_h8_ssse3 uses the PSHUFB instruction which has a 3 cycle latency and slows execution when done in blocks of 5 or more on Atom processors.
By replacing the PSHUFB instructions with other more efficient single cycle instructions (PUNPCKLBW + PUNPCHBW + PALIGNR) performance can be improved.
In the original code, the PSHUBF uses every byte and is consecutively copied.
This is done more efficiently by PUNPCKLBW and PUNPCHBW, using PALIGNR to concatenate the intermediate result and then shift right the next consecutive 16 bytes for the final result.

For example:
filter = 0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,8
Reg = 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15
REG1 = PUNPCKLBW Reg, Reg = 0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7
REG2 = PUNPCHBW Reg, Reg = 8,8,9,9,10,10,11,11,12,12,13,13,14,14,15,15
PALIGNR REG2, REG1, 1 = 0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,8

This optimization improved the function performance by 23% and produced a 3% user level gain on 1080p content on Atom processors.
There was no observed performance impact on Core processors (expected).

Change-Id: I3cec701158993d95ed23ff04516942b5a4a461c0

9 years agoMerge "Improve the performance by caching the left_mi and right_mi in macroblockd."
hkuang [Mon, 8 Dec 2014 18:24:17 +0000 (10:24 -0800)]
Merge "Improve the performance by caching the left_mi and right_mi in macroblockd."

9 years agoMerge "Use average mb energy from first pass in AQ2 test."
Paul Wilkins [Mon, 8 Dec 2014 17:01:39 +0000 (09:01 -0800)]
Merge "Use average mb energy from first pass in AQ2 test."

9 years agoMerge "Fix potential integer overflow."
Frank Galligan [Mon, 8 Dec 2014 05:37:39 +0000 (21:37 -0800)]
Merge "Fix potential integer overflow."

9 years agoMake the decoder Cfg available to encoder tests..
Jim Bankoski [Sun, 7 Dec 2014 19:28:51 +0000 (11:28 -0800)]
Make the decoder Cfg available to encoder tests..

Adds decoder config as a changeable parameter to unit tests, and
changes end to end test to use commonly used parameters to enable
base test of tiles encoding and frame parallel decoding.

Change-Id: I5d23a6857303b4d68b92b15c3f2f04a1bcb4c2bb

9 years agoMerge "vp9 asserts: fix compile warning"
James Zern [Sat, 6 Dec 2014 05:09:42 +0000 (21:09 -0800)]
Merge "vp9 asserts: fix compile warning"

9 years agoMerge "fix building with --disable-spatial-resampling"
James Zern [Sat, 6 Dec 2014 04:36:44 +0000 (20:36 -0800)]
Merge "fix building with --disable-spatial-resampling"

9 years agofix building with --disable-spatial-resampling
James Zern [Sat, 6 Dec 2014 00:02:42 +0000 (16:02 -0800)]
fix building with --disable-spatial-resampling

vpx_scale.c is only used by the vp8 encoder when spatial resampling is
enabled.

Change-Id: If3d3ad81e9ee6e0b59f8c040b9624ef52598fe03

9 years agoMerge "vp9: fix frame-parallel encoding"
James Zern [Sat, 6 Dec 2014 03:00:48 +0000 (19:00 -0800)]
Merge "vp9: fix frame-parallel encoding"

9 years agoMerge "Some internal-stats, vp9-highbitdepth bug fixes"
Deb Mukherjee [Sat, 6 Dec 2014 01:49:52 +0000 (17:49 -0800)]
Merge "Some internal-stats, vp9-highbitdepth bug fixes"

9 years agovp9: fix frame-parallel encoding
James Zern [Sat, 6 Dec 2014 01:44:46 +0000 (17:44 -0800)]
vp9: fix frame-parallel encoding

the flag in the header wasn't being set based on the encoder
configuration in non-intra only mode

broken since:
fbc2fbf Adding oxcf temp variable.

Change-Id: Ib4cff9901889824bc4e68d7f0f6deb1e41df2f53

9 years agoMerge "Remove redundant rdcost reset"
Jingning Han [Sat, 6 Dec 2014 01:35:07 +0000 (17:35 -0800)]
Merge "Remove redundant rdcost reset"

9 years agoMerge "Fix a motion search skip condition in vp9_pick_inter_mode"
Jingning Han [Sat, 6 Dec 2014 01:35:04 +0000 (17:35 -0800)]
Merge "Fix a motion search skip condition in vp9_pick_inter_mode"

9 years agoMerge "Remove redundant MB_MODE_INFO reset from vp9_pick_mode_inter"
Jingning Han [Sat, 6 Dec 2014 00:59:50 +0000 (16:59 -0800)]
Merge "Remove redundant MB_MODE_INFO reset from vp9_pick_mode_inter"

9 years agoImprove the performance by caching the left_mi and right_mi in macroblockd.
hkuang [Fri, 5 Dec 2014 22:39:21 +0000 (14:39 -0800)]
Improve the performance by caching the left_mi and right_mi in macroblockd.

This improve the deocde performance by ~2% on Nexus 7 2013.

Change-Id: Ie9c4ba0371a149eb7fddc687a6a291c17298d6c3

9 years agovp9 asserts: fix compile warning
James Zern [Sat, 6 Dec 2014 00:20:42 +0000 (16:20 -0800)]
vp9 asserts: fix compile warning

string literal to int within an assert

Change-Id: I76a173f96b9add5bf27c3f5ad5d72c6f30e51629

9 years agoRemove redundant rdcost reset
Jingning Han [Sat, 6 Dec 2014 00:04:22 +0000 (16:04 -0800)]
Remove redundant rdcost reset

The initial reset of this_rdc in vp9_pick_inter_mode is not needed,
since it will be re-assign when used.

Change-Id: Ic0e12d741cbab292fc214c1eabb48b129af7839b

9 years agoFix a motion search skip condition in vp9_pick_inter_mode
Jingning Han [Fri, 5 Dec 2014 23:56:39 +0000 (15:56 -0800)]
Fix a motion search skip condition in vp9_pick_inter_mode

Compare the current best mode rate-distortion cost with the skip
threshold to decide if performing motion search.

Change-Id: Ia071824f8dd3b7db485f424692a485a2da6a1a9f

9 years agoRemove redundant MB_MODE_INFO reset from vp9_pick_mode_inter
Jingning Han [Fri, 5 Dec 2014 23:50:31 +0000 (15:50 -0800)]
Remove redundant MB_MODE_INFO reset from vp9_pick_mode_inter

Change-Id: I0222f7abc61202f4a83b117bbfb042ada6304562

9 years agoMerge "Merge set_prev_mi function into encoder function."
hkuang [Fri, 5 Dec 2014 23:12:50 +0000 (15:12 -0800)]
Merge "Merge set_prev_mi function into encoder function."

9 years agoSome internal-stats, vp9-highbitdepth bug fixes
Deb Mukherjee [Fri, 5 Dec 2014 21:40:50 +0000 (13:40 -0800)]
Some internal-stats, vp9-highbitdepth bug fixes

Change-Id: I0363d98f6f6558a43276aec48f27dca37c93f5ad

9 years agoMerge "Remove redundant vp9_zero in choose_partitioning"
Jingning Han [Fri, 5 Dec 2014 19:47:57 +0000 (11:47 -0800)]
Merge "Remove redundant vp9_zero in choose_partitioning"

9 years agoMerge "Enable conditional skip path in rd_pick_intra_sby_mode"
Jingning Han [Fri, 5 Dec 2014 19:25:30 +0000 (11:25 -0800)]
Merge "Enable conditional skip path in rd_pick_intra_sby_mode"

9 years agoMerge "Use hybrid RD and non-RD coding flow for key frame coding"
Jingning Han [Fri, 5 Dec 2014 19:25:19 +0000 (11:25 -0800)]
Merge "Use hybrid RD and non-RD coding flow for key frame coding"

9 years agoRemove redundant vp9_zero in choose_partitioning
Jingning Han [Fri, 5 Dec 2014 18:38:09 +0000 (10:38 -0800)]
Remove redundant vp9_zero in choose_partitioning

It makes the overall speed -6 about 2% faster with no compression
performance change.

Change-Id: I680a967b421caa2c5a5cdb821311c4726a2df45a

9 years agoEnable conditional skip path in rd_pick_intra_sby_mode
Jingning Han [Thu, 4 Dec 2014 22:53:36 +0000 (14:53 -0800)]
Enable conditional skip path in rd_pick_intra_sby_mode

These speed-up features for key frame coding are only turned on
in the settings of hybrid non-RD and RD mode decision. It provides
about 20% speed-up to the hybrid key frame coding at the expense
of certain compression performance loss. For vidyo1, the key frame
coding statistics are changed
9838F, 35.020 dB, 61677 us -> 9920F, 34.834 dB, 47556 us

Overall rtc set compression performance is down by -0.257%.

Change-Id: I0025447fda26bb7855e982955642b5f55d71b51f

9 years agoUse hybrid RD and non-RD coding flow for key frame coding
Jingning Han [Thu, 4 Dec 2014 17:29:24 +0000 (09:29 -0800)]
Use hybrid RD and non-RD coding flow for key frame coding

When block size is below 16x16, the encoder swap from non-RD to
RD mode for key frame coding. This largely brough back the key
frame compression performance. For vidyo1 at 1000 kbps, the key
frame coding statistics are changed

9978F, 34.183 dB, 36807 us -> 9838F, 35.020 dB, 61677 us

As compared to the full RD case
7187F, 34.930 dB, 214470 us

The overall rtc set coding performance (single key frame setting)
is improved by 1.5%.

Change-Id: I78a4ecf025d7b24ec911e85be94e01da05e77878

9 years agoMerge "vp9_ethread: the tile-based multi-threaded encoder"
Yunqing Wang [Fri, 5 Dec 2014 16:23:49 +0000 (08:23 -0800)]
Merge "vp9_ethread: the tile-based multi-threaded encoder"

9 years agoFix potential integer overflow.
Frank Galligan [Fri, 5 Dec 2014 00:26:47 +0000 (16:26 -0800)]
Fix potential integer overflow.

ioc found a potential integer overflow in the rate control.

This is related to https://code.google.com/p/webm/issues/detail?id=821

Change-Id: Ib6c4acd6e964972f932fce7490592eb134f2b7ea

9 years agoMerge "Increase strength of AQ1."
Paul Wilkins [Fri, 5 Dec 2014 12:11:43 +0000 (04:11 -0800)]
Merge "Increase strength of AQ1."

9 years agoMerge "Use the RTC optimizations when in high bitdepth mode."
Debargha Mukherjee [Fri, 5 Dec 2014 03:22:27 +0000 (19:22 -0800)]
Merge "Use the RTC optimizations when in high bitdepth mode."

9 years agoMerge "vp9_reader: reorder struct members"
James Zern [Fri, 5 Dec 2014 00:08:08 +0000 (16:08 -0800)]
Merge "vp9_reader: reorder struct members"

9 years agoMerge "Corrected the renaming of CONFIG_VP9_HIGH ro CONFIG_VP9_HIGHBITDEPTH."
Debargha Mukherjee [Thu, 4 Dec 2014 23:52:35 +0000 (15:52 -0800)]
Merge "Corrected the renaming of CONFIG_VP9_HIGH ro CONFIG_VP9_HIGHBITDEPTH."

9 years agoUse the RTC optimizations when in high bitdepth mode.
Peter de Rivaz [Thu, 4 Dec 2014 10:51:10 +0000 (10:51 +0000)]
Use the RTC optimizations when in high bitdepth mode.

Change 72193 made the encoder behave differently
when configured with and without high bitdepth.
This change means the same algorithm is used for both.

Change-Id: I707a44a94afca773a9e0c2f7ebeeea83030257c5

9 years agoMerge "Loop filter test fix"
James Zern [Thu, 4 Dec 2014 23:14:20 +0000 (15:14 -0800)]
Merge "Loop filter test fix"

9 years agoClean up the logic of handling corrupted frame.
hkuang [Thu, 4 Dec 2014 23:06:31 +0000 (15:06 -0800)]
Clean up the logic of handling corrupted frame.

No more checking of corrupted reference frame as we skip
decoding any non-intra frame in case of frame corrupted.

Change-Id: I77d41bbb02fc5f61972740e2d411441eb6a17073

9 years agoMerge set_prev_mi function into encoder function.
hkuang [Thu, 4 Dec 2014 22:44:23 +0000 (14:44 -0800)]
Merge set_prev_mi function into encoder function.

Change-Id: Ifcf2efbb232ea4cabcdebbe77e0820d121e4a6da

9 years agovp9_ethread: the tile-based multi-threaded encoder
Yunqing Wang [Tue, 2 Dec 2014 23:47:41 +0000 (15:47 -0800)]
vp9_ethread: the tile-based multi-threaded encoder

Currently, VP9 supports column-tile encoding, which allows a frame
to be encoded in multiple column tiles independently. The number of
column tiles are set by encoder option "--tile-columns". This
provides a way to encode a frame in parallel.

Based on previous set of patches, this patch implemented the tile-
based multi-threaded encoder. Each thread processes one or more
tiles.

Usage:
For HD clips:
--tile-columns=2 --threads=1/2/3/4

While using 4 threads, tests showed that the encoder achieved
2.3X - 2.5X speedup at good-quality speed 3, and 2X speedup at
realtime speed 5.

Change-Id: Ied987f8f2618b1283a8643ad255e88341733c9d4

9 years agoMerge "Fixes a missing highbitdepth convolve call bug"
Deb Mukherjee [Thu, 4 Dec 2014 19:19:59 +0000 (11:19 -0800)]
Merge "Fixes a missing highbitdepth convolve call bug"

9 years agoMerge "Free motion vector array before re-allocating"
Adrian Grange [Thu, 4 Dec 2014 15:08:37 +0000 (07:08 -0800)]
Merge "Free motion vector array before re-allocating"

9 years agoCorrected the renaming of CONFIG_VP9_HIGH ro CONFIG_VP9_HIGHBITDEPTH.
Peter de Rivaz [Thu, 4 Dec 2014 11:01:46 +0000 (11:01 +0000)]
Corrected the renaming of CONFIG_VP9_HIGH ro CONFIG_VP9_HIGHBITDEPTH.

Change 71789 renamed CONFIG_VP9_HIGH to CONFIG_VP9_HIGHBITDEPTH.
However, one use of CONFIG_VP9_HIGH was missed.

Change-Id: I0ebb9c71380c6d810a25708d15471abf9533e695

9 years agoLoop filter test fix
Deb Mukherjee [Thu, 4 Dec 2014 00:26:48 +0000 (16:26 -0800)]
Loop filter test fix

Related to fixing random number range.

Change-Id: I84d004600e1f6918badf08d08da421b4e8cc4826

9 years agoMerge "Various updates to vp8."
Marco [Thu, 4 Dec 2014 02:12:47 +0000 (18:12 -0800)]
Merge "Various updates to vp8."

9 years agoMerge "acm_random: provide full range for Rand16()"
James Zern [Thu, 4 Dec 2014 02:05:36 +0000 (18:05 -0800)]
Merge "acm_random: provide full range for Rand16()"

9 years agoMerge "sse2 visual studio build fix"
Tom Finegan [Thu, 4 Dec 2014 02:05:03 +0000 (18:05 -0800)]
Merge "sse2 visual studio build fix"

9 years agoFixes a missing highbitdepth convolve call bug
Deb Mukherjee [Thu, 4 Dec 2014 01:48:50 +0000 (17:48 -0800)]
Fixes a missing highbitdepth convolve call bug

Bug was introduced in https://gerrit.chromium.org/gerrit/#/c/72122/

Change-Id: Idb500ea619a30e7bc50e22fb8ee03be5282f41db

9 years agoacm_random: provide full range for Rand16()
James Zern [Mon, 1 Dec 2014 23:10:00 +0000 (15:10 -0800)]
acm_random: provide full range for Rand16()

the gtest implementation used only returns values between 0 and 2^31-1

+ temporarily disable some tests in fdct8x8_test which misbehave with the
  new range

Change-Id: I45381076f0bea3317cc6728305890e4fd2f2facd

9 years agoMerge "acm_random: provide full range for Rand8()"
James Zern [Thu, 4 Dec 2014 00:58:42 +0000 (16:58 -0800)]
Merge "acm_random: provide full range for Rand8()"

9 years agoMerge "Use memset for initialization to 0"
Adrian Grange [Thu, 4 Dec 2014 00:50:39 +0000 (16:50 -0800)]
Merge "Use memset for initialization to 0"

9 years agosse2 visual studio build fix
Deb Mukherjee [Wed, 3 Dec 2014 23:18:44 +0000 (15:18 -0800)]
sse2 visual studio build fix

Change-Id: Id8c8c3be882bcd92afea3ccec6ebdf3f208d28ef

9 years agoFree motion vector array before re-allocating
Adrian Grange [Tue, 2 Dec 2014 00:05:22 +0000 (16:05 -0800)]
Free motion vector array before re-allocating

Change-Id: I0c39136d67e1e83020d61f86b062a04182ec9b00

9 years agoMerge "Increase delta-qp for aq=3 mode, after key frame."
Marco [Thu, 4 Dec 2014 00:03:06 +0000 (16:03 -0800)]
Merge "Increase delta-qp for aq=3 mode, after key frame."

9 years agoVarious updates to vp8.
Marco [Mon, 10 Nov 2014 21:07:05 +0000 (13:07 -0800)]
Various updates to vp8.

Change-Id: Icc7a816491897107764e4c936288e9000e6319b8

9 years agoMerge "Fix indent in source_var_based_partition_search_method"
Jingning Han [Wed, 3 Dec 2014 23:43:40 +0000 (15:43 -0800)]
Merge "Fix indent in source_var_based_partition_search_method"