platform/upstream/libvpx.git
10 years agoCorrect the mode decisions in special cases
Yunqing Wang [Thu, 4 Sep 2014 22:16:12 +0000 (15:16 -0700)]
Correct the mode decisions in special cases

The rate costs calculated for inter modes are not precise in some
cases, which causes NEWMV is chosen instead of NEARESTMV, NEARMV,
and ZEROMV. This patch added checks for these cases, and corrected
the mode decisions.

Borg tests at speed 3 showed:
1. stdhd set: 0.102% PSNR gain and 0.088% SSIM gain.
2. derf set:  0.147% PSNR gain and 0.132% SSIM gain.
No speed change.

Change-Id: I35d17684b89ad4734fb610942d707899146426db

10 years agoMerge "fix x86-darwin* build"
James Zern [Fri, 5 Sep 2014 18:58:55 +0000 (11:58 -0700)]
Merge "fix x86-darwin* build"

10 years agoMerge "vp9: skip loopfilter when the frame is corrupt"
James Zern [Fri, 5 Sep 2014 18:58:09 +0000 (11:58 -0700)]
Merge "vp9: skip loopfilter when the frame is corrupt"

10 years agoMerge "Adding temp cpi var."
Dmitry Kovalev [Fri, 5 Sep 2014 17:31:41 +0000 (10:31 -0700)]
Merge "Adding temp cpi var."

10 years agovp8 common: change 'HAVE_NEON_ASM' to 'HAVE_NEON' for compiling functions of NEON...
Jia Jia [Fri, 5 Sep 2014 05:02:22 +0000 (05:02 +0000)]
vp8 common: change 'HAVE_NEON_ASM' to 'HAVE_NEON' for compiling functions of NEON intrinsics.

Change-Id: I975e5eac16f8b623ff589f0ec072cdaff2183b04

10 years agofix x86-darwin* build
James Zern [Fri, 5 Sep 2014 06:32:13 +0000 (23:32 -0700)]
fix x86-darwin* build

vp9_variance_sse2.c contains a mix of intrinsics and references to
assembly which uses x86inc.asm; it's conditionally included as a result.

Change-Id: I254451483a65881c0b8e18e27bf0c3ddef60c4ec

10 years agobilinearpredict_neon: fix type conversion warnings
James Zern [Fri, 11 Jul 2014 07:05:57 +0000 (00:05 -0700)]
bilinearpredict_neon: fix type conversion warnings

make bifilter4_coeff[][] uint8_t, no values exceed this range and
they're loaded with vdup_n_u8().

Change-Id: I921983e9edd828d29820e40ac30a7801dbe0fb4f

10 years agoMerge "Removing sz member from vpx_codec_priv. "
Dmitry Kovalev [Fri, 5 Sep 2014 00:28:22 +0000 (17:28 -0700)]
Merge "Removing sz member from vpx_codec_priv. "

10 years agoMerge "Removing unused function prototypes."
Dmitry Kovalev [Fri, 5 Sep 2014 00:28:16 +0000 (17:28 -0700)]
Merge "Removing unused function prototypes."

10 years agoMerge "Map motion magnitude in VP9 denoiser."
JackyChen [Thu, 4 Sep 2014 23:59:53 +0000 (16:59 -0700)]
Merge "Map motion magnitude in VP9 denoiser."

10 years agoMerge "Speed up compound inter prediction mode check"
Jingning Han [Thu, 4 Sep 2014 23:09:07 +0000 (16:09 -0700)]
Merge "Speed up compound inter prediction mode check"

10 years agoMerge "Update the condition when COPY_BLOCK is chosen."
JackyChen [Thu, 4 Sep 2014 22:48:22 +0000 (15:48 -0700)]
Merge "Update the condition when COPY_BLOCK is chosen."

10 years agoMerge "Removing local set_speed_features() function."
Dmitry Kovalev [Thu, 4 Sep 2014 22:36:52 +0000 (15:36 -0700)]
Merge "Removing local set_speed_features() function."

10 years agoMap motion magnitude in VP9 denoiser.
JackyChen [Thu, 4 Sep 2014 21:44:46 +0000 (14:44 -0700)]
Map motion magnitude in VP9 denoiser.

This is to keep the same with VP8 denoiser.
If motion magnitude is small,
make denoiser more aggressive.

Change-Id: I942a6e2f2ed9aec6f0c4c1f9e5fa47066cadcc0c

10 years agoAdding temp cpi var.
Dmitry Kovalev [Thu, 4 Sep 2014 21:51:29 +0000 (14:51 -0700)]
Adding temp cpi var.

Change-Id: Ifa3c1cc2317c1bc21d1042b9662b35056d1e9ed0

10 years agoRemoving sz member from vpx_codec_priv.
Dmitry Kovalev [Thu, 4 Sep 2014 21:47:42 +0000 (14:47 -0700)]
Removing sz member from vpx_codec_priv.

Change-Id: I811526a9ee9f237604f72abe7fc677e39e0f457f

10 years agoUpdate the condition when COPY_BLOCK is chosen.
JackyChen [Thu, 4 Sep 2014 21:08:18 +0000 (14:08 -0700)]
Update the condition when COPY_BLOCK is chosen.

The change is just to keep the condition the same with VP8.

Change-Id: I9662b40996126605945dd853c0cbe8916c1ce578

10 years agoMerge "Actually resetting random generator for all variance test cases."
Dmitry Kovalev [Thu, 4 Sep 2014 21:27:04 +0000 (14:27 -0700)]
Merge "Actually resetting random generator for all variance test cases."

10 years agoRemoving unused function prototypes.
Dmitry Kovalev [Thu, 4 Sep 2014 21:26:30 +0000 (14:26 -0700)]
Removing unused function prototypes.

Change-Id: Ia5e383e2cf18052f6f1eacf8b9495ab8e4d58878

10 years agoMerge "Fix a bug in VP9 denoiser."
JackyChen [Thu, 4 Sep 2014 21:16:26 +0000 (14:16 -0700)]
Merge "Fix a bug in VP9 denoiser."

10 years agoMerge "Adding sse2 variant for vp9_mse{8x8, 8x16, 16x8}."
Dmitry Kovalev [Thu, 4 Sep 2014 20:59:37 +0000 (13:59 -0700)]
Merge "Adding sse2 variant for vp9_mse{8x8, 8x16, 16x8}."

10 years agoActually resetting random generator for all variance test cases.
Dmitry Kovalev [Thu, 4 Sep 2014 19:24:52 +0000 (12:24 -0700)]
Actually resetting random generator for all variance test cases.

Calling Reset(int) method instead of overloaded operator()(int).
Adding underscore at the end of class member name.

Change-Id: I01934e7bc056d4b594e5d05d693328febd34ac3c

10 years agoFix a bug in VP9 denoiser.
JackyChen [Thu, 4 Sep 2014 18:16:56 +0000 (11:16 -0700)]
Fix a bug in VP9 denoiser.

When the first try of denoising turns out to be too much,
we will use a softer filter by adopting an adjustment to
make the result closer to original pixel (as in VP8 denoiser).
The old code made the adjustment in the wrong direction.

Change-Id: I84e28fa9e01eef47c5a37d5a2e6d3d378a06786b

10 years agoMerge "arm: Fix building vp8_subpixelvariance_neon.c with MSVC"
James Zern [Thu, 4 Sep 2014 18:00:53 +0000 (11:00 -0700)]
Merge "arm: Fix building vp8_subpixelvariance_neon.c with MSVC"

10 years agoRevert "Revert "VP8 for ARMv8 by using NEON intrinsics 10""
Scott LaVarnway [Tue, 2 Sep 2014 12:27:28 +0000 (05:27 -0700)]
Revert "Revert "VP8 for ARMv8 by using NEON intrinsics 10""

This reverts commit 677fb5123e0ece1d6c30be9d0282e1e1f594a56d

Compiles with 4.6.

Change-Id: I7f87048911b6bc28a61741d95501fa45ee97b819

10 years agovp8 common: change 'HAVE_NEON_ASM' to 'HAVE_NEON' for compiling idct_blk_neon.c.
Jia Jia [Thu, 4 Sep 2014 09:36:18 +0000 (09:36 +0000)]
vp8 common: change 'HAVE_NEON_ASM' to 'HAVE_NEON' for compiling idct_blk_neon.c.

Change-Id: Ib89107fb824b5fe58afef6841104d5a27b2e0f2d

10 years agoarm: Fix building vp8_subpixelvariance_neon.c with MSVC
Martin Storsjo [Thu, 4 Sep 2014 06:23:08 +0000 (09:23 +0300)]
arm: Fix building vp8_subpixelvariance_neon.c with MSVC

Use the right return values - vget_low_s64 returns int64x1_t, not
a normal int64_t.

Also make __builtin_prefetch a no-op on MSVC for this file.

Change-Id: I4d2fce01d0ba106b98d3d53b137803119c2c2c08

10 years agoMerge "Consistent allocation of vpx_codec_alg_priv_t."
Dmitry Kovalev [Thu, 4 Sep 2014 02:41:28 +0000 (19:41 -0700)]
Merge "Consistent allocation of vpx_codec_alg_priv_t."

10 years agoAdding sse2 variant for vp9_mse{8x8, 8x16, 16x8}.
Dmitry Kovalev [Tue, 2 Sep 2014 19:09:14 +0000 (12:09 -0700)]
Adding sse2 variant for vp9_mse{8x8, 8x16, 16x8}.

Change-Id: I6786d25ce4f32b8d8912f2d239a45ca15b310c4b

10 years agoMerge "Replacing asm 16x16 variance calculation with intrinsics."
Dmitry Kovalev [Thu, 4 Sep 2014 01:57:33 +0000 (18:57 -0700)]
Merge "Replacing asm 16x16 variance calculation with intrinsics."

10 years agoMerge "Small cleanup: reusing existing code."
Dmitry Kovalev [Thu, 4 Sep 2014 01:57:25 +0000 (18:57 -0700)]
Merge "Small cleanup: reusing existing code."

10 years agoMerge "Add a more complex SEG_LVL_SKIP test vector."
Yaowu Xu [Thu, 4 Sep 2014 00:17:12 +0000 (17:17 -0700)]
Merge "Add a more complex SEG_LVL_SKIP test vector."

10 years agoSpeed up compound inter prediction mode check
Jingning Han [Wed, 3 Sep 2014 00:32:12 +0000 (17:32 -0700)]
Speed up compound inter prediction mode check

This commit allows the encoder to store outcomes of single reference
frame modes and compares them to decide if the inter prediction
filter, forward transform, and quantization can be skipped.

The compression performance of speed 3 is down
derf  -0.364%
stdhd -0.198%

For test sequences, the speed 3 runtime is reduced
highway CIF 100 kbps, 51976 ms -> 45033 ms, 13% speed-up
stockholm 720p 1000 kbps, 71826 ms -> 67838 ms, 5.5% speed-up
pedestrian 1080p 2000 kbps, 154924 ms -> 150702 ms, 2.6% speed-up

Change-Id: I5aa26f918d2b4b5197a2c0afa2779319f1c88e44

10 years agoAdd a more complex SEG_LVL_SKIP test vector.
Alex Converse [Wed, 3 Sep 2014 22:08:49 +0000 (15:08 -0700)]
Add a more complex SEG_LVL_SKIP test vector.

This file changes the segmentation map, deactivates the map, and
reactivates the map.

Change-Id: I1c8b8949887257b9f4c49a1a9ba9a7c74666ef2e

10 years agoMerge "Change last_partition_redo_frequency for speed 3"
Yaowu Xu [Wed, 3 Sep 2014 21:57:02 +0000 (14:57 -0700)]
Merge "Change last_partition_redo_frequency for speed 3"

10 years agoMerge "Remove redundant code"
Yaowu Xu [Wed, 3 Sep 2014 21:55:28 +0000 (14:55 -0700)]
Merge "Remove redundant code"

10 years agoMerge "Make process_denoise_mode_change function cond-compiled."
Marco [Wed, 3 Sep 2014 20:58:16 +0000 (13:58 -0700)]
Merge "Make process_denoise_mode_change function cond-compiled."

10 years agoNeon version of vp8_build_intra_predictors_mby_s() and
Scott LaVarnway [Wed, 3 Sep 2014 18:26:27 +0000 (11:26 -0700)]
Neon version of vp8_build_intra_predictors_mby_s() and

vp8_build_intra_predictors_mbuv_s().

This patch replaces the assembly version with an intrinsic
version.

On a Nexus 7, vpxenc (in realtime mode, speed -12)
reported a performance improvement of ~2.6%.

Change-Id: I9ef65bad929450c0215253fdae1c16c8b4a8f26f

10 years agoVP8 for ARMv8 by using NEON intrinsics 17
Scott LaVarnway [Tue, 26 Aug 2014 19:13:03 +0000 (12:13 -0700)]
VP8 for ARMv8 by using NEON intrinsics 17

Add vp8_subpixelvariance_neon.c
- vp8_sub_pixel_variance16x16_neon_func
- vp8_variance_halfpixvar16x16_h_neon
- vp8_variance_halfpixvar16x16_v_neon
- vp8_variance_halfpixvar16x16_hv_neon
- vp8_sub_pixel_variance8x8_neon

Change-Id: I3e5d85b2eafc26be0eef6a777789b80e4579257b
Signed-off-by: James Yu <james.yu@linaro.org>
10 years agoMerge "Revert "Revert "VP8 for ARMv8 by using NEON intrinsics 06" This reverts commit...
Johann [Wed, 3 Sep 2014 20:27:11 +0000 (13:27 -0700)]
Merge "Revert "Revert "VP8 for ARMv8 by using NEON intrinsics 06" This reverts commit 81ad047ee57ecb0e2c1ee4dcebda54a44ea54ae9. Revert "VP8 for ARMv8 by using NEON intrinsics 15" This reverts commit 727af7cebe3698b8493ba6c1360b0a6606c310fb.""

10 years agoMerge "Cleaning up vp9_variance_avx2.c."
Dmitry Kovalev [Wed, 3 Sep 2014 20:21:38 +0000 (13:21 -0700)]
Merge "Cleaning up vp9_variance_avx2.c."

10 years agoMerge "Merge two similar functions into one"
Yaowu Xu [Wed, 3 Sep 2014 20:00:02 +0000 (13:00 -0700)]
Merge "Merge two similar functions into one"

10 years agoSmall cleanup: reusing existing code.
Dmitry Kovalev [Wed, 3 Sep 2014 19:20:29 +0000 (12:20 -0700)]
Small cleanup: reusing existing code.

Change-Id: Iac4775ad98e988f2b9cf5bd0dc91ab994d0262ce

10 years agoMerge "Removing duplicated code."
Dmitry Kovalev [Wed, 3 Sep 2014 19:11:37 +0000 (12:11 -0700)]
Merge "Removing duplicated code."

10 years agoMerge "select_tx_mode(): remove special case for key frame"
Yaowu Xu [Wed, 3 Sep 2014 18:54:44 +0000 (11:54 -0700)]
Merge "select_tx_mode(): remove special case for key frame"

10 years agoMake process_denoise_mode_change function cond-compiled.
Marco [Wed, 3 Sep 2014 18:35:06 +0000 (11:35 -0700)]
Make process_denoise_mode_change function cond-compiled.

This function is called only when temporal denoising is enabled.

Change-Id: I25a036e3d628edd34410d6e6fe1f9e68a90a8e56

10 years agoMerge two similar functions into one
Yaowu Xu [Tue, 2 Sep 2014 19:42:34 +0000 (12:42 -0700)]
Merge two similar functions into one

intra_super_block_yrd() and inter_super_block_yrd() are largely same,
this commit merges them into one to reduce code duplication.

Change-Id: I64d7042a5b099345627cf55663010c185b25ec37

10 years agoMerge "Removing clear_system_state() call from update_coef_probs()."
Dmitry Kovalev [Wed, 3 Sep 2014 18:05:45 +0000 (11:05 -0700)]
Merge "Removing clear_system_state() call from update_coef_probs()."

10 years agoConsistent allocation of vpx_codec_alg_priv_t.
Dmitry Kovalev [Wed, 3 Sep 2014 01:26:53 +0000 (18:26 -0700)]
Consistent allocation of vpx_codec_alg_priv_t.

Change-Id: I5a03496de035fbcf31e4527cd25fcae4627a57a0

10 years agoMerge "[svc] Temporal svc with two pass rate control"
Minghai Shang [Wed, 3 Sep 2014 17:51:19 +0000 (10:51 -0700)]
Merge "[svc] Temporal svc with two pass rate control"

10 years agoChange last_partition_redo_frequency for speed 3
Yaowu Xu [Wed, 3 Sep 2014 01:53:10 +0000 (18:53 -0700)]
Change last_partition_redo_frequency for speed 3

From 3 to 2, which seems to be slightly positive on compression for
all test sets, also reduces encoding time by 2%-5%, varying on the
test clips.

Change-Id: If045417bd27311700c919b4a335eff0dc1130ae0

10 years agoRemove redundant code
Yaowu Xu [Wed, 3 Sep 2014 01:00:49 +0000 (18:00 -0700)]
Remove redundant code

Change-Id: I453b167f03811a3cd3592089593b3f2823f62ab3

10 years agoselect_tx_mode(): remove special case for key frame
Yaowu Xu [Tue, 2 Sep 2014 23:21:57 +0000 (16:21 -0700)]
select_tx_mode(): remove special case for key frame

This commit removes the special case for key frame, as transform size
decision is controlled by the appropriate speed feature for all lossy
coding modes: tx_size_search_method.

Change-Id: I9677171e3f2432ec23705f7c5ea8170dd4562fae

10 years agoMerge "Add a test vector for SEG_LVL_SKIP."
Yaowu Xu [Wed, 3 Sep 2014 15:12:20 +0000 (08:12 -0700)]
Merge "Add a test vector for SEG_LVL_SKIP."

10 years agoMerge "Skip comp inter mode test in RD loop with same frame bias signs"
Paul Wilkins [Wed, 3 Sep 2014 09:26:47 +0000 (02:26 -0700)]
Merge "Skip comp inter mode test in RD loop with same frame bias signs"

10 years agoSkip comp inter mode test in RD loop with same frame bias signs
Jingning Han [Wed, 3 Sep 2014 01:05:03 +0000 (18:05 -0700)]
Skip comp inter mode test in RD loop with same frame bias signs

This commit allows the encoder to skip check on compound inter
modes in the rate-distortion optimization loop, if the reference
frame bias signs are the same.

Change-Id: Ib753e6bb11cbdd338aee69dbe2b649671f75a6b0

10 years agoRemoving duplicated code.
Dmitry Kovalev [Wed, 3 Sep 2014 00:57:35 +0000 (17:57 -0700)]
Removing duplicated code.

Change-Id: I7b5c776d5e6f5ca428b87fa9411ae4012a9538ba

10 years agoMerge "Removing MMX SAD calculation code."
Dmitry Kovalev [Wed, 3 Sep 2014 00:35:59 +0000 (17:35 -0700)]
Merge "Removing MMX SAD calculation code."

10 years agoAdd a test vector for SEG_LVL_SKIP.
Alex Converse [Tue, 2 Sep 2014 23:39:21 +0000 (16:39 -0700)]
Add a test vector for SEG_LVL_SKIP.

Change-Id: Ib41acade12fe6cdd4c4a4efdb1260d5f100a3e7f

10 years agoMerge "Adds config opt for highbitdepth + misc. vpx"
Deb Mukherjee [Tue, 2 Sep 2014 22:41:27 +0000 (15:41 -0700)]
Merge "Adds config opt for highbitdepth + misc. vpx"

10 years agoRemoving MMX SAD calculation code.
Dmitry Kovalev [Tue, 2 Sep 2014 21:41:36 +0000 (14:41 -0700)]
Removing MMX SAD calculation code.

Removed functions:
* vp9_sad_16x16_mmx
* vp9_sad_8x16_mmx
* vp9_sad_16x8_mmx
* vp9_sad_8x8_mmx
* vp9_sad_4x4_mmx

Change-Id: Ic5174b93b64d65d846f0c11e72cab149e9472bc3

10 years agoAdds config opt for highbitdepth + misc. vpx
Deb Mukherjee [Tue, 26 Aug 2014 19:35:15 +0000 (12:35 -0700)]
Adds config opt for highbitdepth + misc. vpx

Adds config parameter vp9_highbitdepth, to support highbitdepth profiles.
Also includes most vpx level high bit-depth functions. However
encode/decode in the highbitdepth profiles will not work until
the rest of the code is in place.

Change-Id: I34c53b253c38873611057a6cbc89a1361b8985a6

10 years agoMerge "Switching back to #define instead of static const int."
Dmitry Kovalev [Tue, 2 Sep 2014 21:02:04 +0000 (14:02 -0700)]
Merge "Switching back to #define instead of static const int."

10 years agoReplacing asm 16x16 variance calculation with intrinsics.
Dmitry Kovalev [Tue, 2 Sep 2014 17:50:12 +0000 (10:50 -0700)]
Replacing asm 16x16 variance calculation with intrinsics.

New code is 20% faster for 64-bit and 15% faster for 32-bit. Compiled
using clang.

Change-Id: Icfea461238411001fd093561293dbfedfbf8d0bb

10 years ago[svc] Temporal svc with two pass rate control
Minghai Shang [Tue, 2 Sep 2014 19:05:14 +0000 (12:05 -0700)]
[svc] Temporal svc with two pass rate control

It's built based on current spatial svc code.
We only support one spatial two temporal layers at this time.
Change-Id: I1fdc8584354b910331e626bfae60473b3b701ba1

10 years agoSwitching back to #define instead of static const int.
Dmitry Kovalev [Tue, 2 Sep 2014 18:29:40 +0000 (11:29 -0700)]
Switching back to #define instead of static const int.

To avoid 'variable length array' warnings from gcc.

Change-Id: I426f7e93ce674a10b901e79c0c9d9df5d4e47cb6

10 years agoSkip comp inter mode tests for arf coding
Jingning Han [Tue, 2 Sep 2014 17:57:10 +0000 (10:57 -0700)]
Skip comp inter mode tests for arf coding

This commit skips the compound inter mode prediction check in the
rate-distortion optimization loop for ARF coding. It reduces the
runtime for certain test clips at speed 3, at no compression
performance change:

bus CIF 1000 kbps, 8260 ms -> 8090 ms, 1.8% speed-up
stockholm 720p 1000 kbps, 74453 ms -> 71826 ms, 2.9% speed-up

No visible speed-up for pedestrian area 1080p at 2000 kbps.

Change-Id: Ic68aa56837159b726563b784e2e3729e846465ad

10 years agoCleaning up vp9_variance_avx2.c.
Dmitry Kovalev [Wed, 28 May 2014 18:44:45 +0000 (11:44 -0700)]
Cleaning up vp9_variance_avx2.c.

Change-Id: I75eb47dd21f87015efd673dbd2aa71f4386afdf5

10 years agoMerge "Updates to adaptive/aggressive denoiser mode."
Marco [Tue, 2 Sep 2014 17:58:41 +0000 (10:58 -0700)]
Merge "Updates to adaptive/aggressive denoiser mode."

10 years agoMerge "Removing 'frames' field from VP9_COMP."
Dmitry Kovalev [Tue, 2 Sep 2014 17:01:20 +0000 (10:01 -0700)]
Merge "Removing 'frames' field from VP9_COMP."

10 years agoMerge "Adding get_frame_pkt_flags() function."
Dmitry Kovalev [Tue, 2 Sep 2014 17:00:51 +0000 (10:00 -0700)]
Merge "Adding get_frame_pkt_flags() function."

10 years agoUpdates to adaptive/aggressive denoiser mode.
Marco [Tue, 2 Sep 2014 16:45:37 +0000 (09:45 -0700)]
Updates to adaptive/aggressive denoiser mode.

Parameter changes and modification to zero_last bias.

Change-Id: I50a408d47fde049c562bbe95075194cb0f17c31b

10 years agoMerge "Revert "Revert "VP8 for ARMv8 by using NEON intrinsics 08"""
Johann [Tue, 2 Sep 2014 16:28:14 +0000 (09:28 -0700)]
Merge "Revert "Revert "VP8 for ARMv8 by using NEON intrinsics 08"""

10 years agoMerge "Removing lookup_next_frame_stats()."
Dmitry Kovalev [Tue, 2 Sep 2014 16:25:16 +0000 (09:25 -0700)]
Merge "Removing lookup_next_frame_stats()."

10 years agoMerge "Skip intra mode tests depending on inter residuals"
Jingning Han [Tue, 2 Sep 2014 15:32:52 +0000 (08:32 -0700)]
Merge "Skip intra mode tests depending on inter residuals"

10 years agoMerge "Replacing asm 8x8 variance calculation with intrinsics."
Dmitry Kovalev [Mon, 1 Sep 2014 01:39:46 +0000 (18:39 -0700)]
Merge "Replacing asm 8x8 variance calculation with intrinsics."

10 years agoRemoving dummy_packing member from VP9_COMP.
Dmitry Kovalev [Fri, 22 Aug 2014 00:28:44 +0000 (17:28 -0700)]
Removing dummy_packing member from VP9_COMP.

Change-Id: I571ce84c97087f8a1a36a10058393bfdcefbf72a

10 years agoReplacing asm 8x8 variance calculation with intrinsics.
Dmitry Kovalev [Fri, 29 Aug 2014 21:25:31 +0000 (14:25 -0700)]
Replacing asm 8x8 variance calculation with intrinsics.

New code is 10% faster for 64-bit and 25% faster for 32-bit. Compiled
using clang.

Change-Id: I8ba1544c30dd6f3ca479db806384317549650dfc

10 years agoMerge "Fix int64_t to unsigned int conversion warnings"
Jingning Han [Sat, 30 Aug 2014 00:15:46 +0000 (17:15 -0700)]
Merge "Fix int64_t to unsigned int conversion warnings"

10 years agoMerge "Extend block level sse to support multiple txfm blocks"
Jingning Han [Sat, 30 Aug 2014 00:15:30 +0000 (17:15 -0700)]
Merge "Extend block level sse to support multiple txfm blocks"

10 years agoFix int64_t to unsigned int conversion warnings
Jingning Han [Thu, 28 Aug 2014 16:21:58 +0000 (09:21 -0700)]
Fix int64_t to unsigned int conversion warnings

Use unsigned int type to store the sse in the pixel domain. The
precision is sufficient to handle sse of block size up to 64x64.
The transform domain version however needs int64_t, since there is
a transfer gain applied in the forward transformation that might
cause unsigned int overflow.

Change-Id: Ifef97c38597e426262290f35341fbb093cf0a079

10 years agoMerge "Removing variance MMX code."
Dmitry Kovalev [Fri, 29 Aug 2014 21:08:02 +0000 (14:08 -0700)]
Merge "Removing variance MMX code."

10 years agoMerge "vp9: sync workers at the start of decode_tiles_mt()"
James Zern [Fri, 29 Aug 2014 21:07:37 +0000 (14:07 -0700)]
Merge "vp9: sync workers at the start of decode_tiles_mt()"

10 years agoMerge "vp9: fix m/t loop filter invalid free"
James Zern [Fri, 29 Aug 2014 21:07:02 +0000 (14:07 -0700)]
Merge "vp9: fix m/t loop filter invalid free"

10 years agoMerge "Minor fix in vp9_encoder.h"
Yunqing Wang [Fri, 29 Aug 2014 20:44:10 +0000 (13:44 -0700)]
Merge "Minor fix in vp9_encoder.h"

10 years agoRevert "Revert "VP8 for ARMv8 by using NEON intrinsics 08""
Scott LaVarnway [Fri, 29 Aug 2014 20:29:36 +0000 (13:29 -0700)]
Revert "Revert "VP8 for ARMv8 by using NEON intrinsics 08""

This reverts commit 928ff03889dadc3f63883553b443c08e625b4885

Compiles with 4.6 now.

Change-Id: Ib455da1098bb0e0623248be07579882a425fcbd1

10 years agoMinor fix in vp9_encoder.h
Yunqing Wang [Fri, 29 Aug 2014 18:27:24 +0000 (11:27 -0700)]
Minor fix in vp9_encoder.h

Added the missing "int".

Change-Id: I7c8af3dee700837b40f010d53e1431a59370ae3a

10 years agovp9: fix m/t loop filter invalid free
James Zern [Fri, 29 Aug 2014 00:11:31 +0000 (17:11 -0700)]
vp9: fix m/t loop filter invalid free

store the number of allocated rows in VP9LfSync, the calculated values
can not be relied on when dealing with corrupt material.

Change-Id: I13b8bcec9738c299a71df726772ab7ac05511e5b

10 years agoRemoving variance MMX code.
Dmitry Kovalev [Fri, 29 Aug 2014 00:31:00 +0000 (17:31 -0700)]
Removing variance MMX code.

Removed functions:
* vp9_mse16x16_mmx
* vp9_get_mb_ss_mmx
* vp9_get4x4var_mmx
* vp9_get8x8var_mmx
* vp9_variance4x4_mmx
* vp9_variance8x8_mmx
* vp9_variance16x16_mmx
* vp9_variance16x8_mmx
* vp9_variance8x16_mmx

They all have SSE2 equivalent.

Change-Id: I3796f2477c4f59b35b4828f46a300c16e62a2615

10 years agoSkip intra mode tests depending on inter residuals
Jingning Han [Fri, 29 Aug 2014 05:58:30 +0000 (22:58 -0700)]
Skip intra mode tests depending on inter residuals

This commit allows encoder to skip intra coding mode test, when
the known inter residual is less than the source variance. It
reduces the runtime of speed 3 for test clips:
bus cif 1000 kbps: 8587 ms -> 8260 ms, 3.8% speed-up
pedestrian 1080p 2000 kbps: 161381 ms -> 155241 ms, 3.7% speed-up.

The compression performance is down by
derf   -0.36%
stdhd  -0.25%

Change-Id: I75ce1e035b4da2153cb1ac14111d1a07c05a735d

10 years agoExtend block level sse to support multiple txfm blocks
Jingning Han [Thu, 28 Aug 2014 16:09:37 +0000 (09:09 -0700)]
Extend block level sse to support multiple txfm blocks

This commit extends the sse and forward transform computation flag
to support the case 64x64 blocks where there are 4 32x32 2D-DCT
blocks.

Change-Id: I86a3e805dfaa0f3abd812f590520c71aa0e40473

10 years agoMerge "vp9: fix crash in inline loopfilter w/corrupt file"
James Zern [Fri, 29 Aug 2014 01:37:30 +0000 (18:37 -0700)]
Merge "vp9: fix crash in inline loopfilter w/corrupt file"

10 years agoMerge "vp9: fix crash in mt loopfilter w/corrupt file"
James Zern [Fri, 29 Aug 2014 01:36:31 +0000 (18:36 -0700)]
Merge "vp9: fix crash in mt loopfilter w/corrupt file"

10 years agovp9: sync workers at the start of decode_tiles_mt()
James Zern [Fri, 29 Aug 2014 00:50:38 +0000 (17:50 -0700)]
vp9: sync workers at the start of decode_tiles_mt()

prevents any problems resuming decode after decoding a corrupt frame

Change-Id: Ib7eb1b5c062aebe71074fef1ece32a32822c16be

10 years agoMerge "Implementing 4x4 variance calculation with SSE2."
Dmitry Kovalev [Fri, 29 Aug 2014 00:25:46 +0000 (17:25 -0700)]
Merge "Implementing 4x4 variance calculation with SSE2."

10 years agoImplementing 4x4 variance calculation with SSE2.
Dmitry Kovalev [Thu, 28 Aug 2014 00:39:44 +0000 (17:39 -0700)]
Implementing 4x4 variance calculation with SSE2.

New SSE2 function is three times faster than MMX one.

Change-Id: I4f387ce9f75b88379176ec7bdc62d86eb5f70fbe

10 years agoRemoving alg_priv from vpx_codec_priv struct.
Dmitry Kovalev [Thu, 21 Aug 2014 00:02:10 +0000 (17:02 -0700)]
Removing alg_priv from vpx_codec_priv struct.

In order to understand memory layout consider the declaration of the
following structs. The first one is a part of our API:

struct vpx_codec_ctx {
  // ...
  struct vpx_codec_priv *priv;
};

The second one is defined in vpx_codec_internal.h:

struct vpx_codec_priv {
  // ...
};

The following struct is defined 4 times for encoder/decoder VP8/VP9:

struct vpx_codec_alg_priv {
  struct vpx_codec_priv base;
  // ...
};

Private data allocation for the given ctx:

struct vpx_codec_ctx *ctx = <get>
struct vpx_codec_alg_priv *alg_priv = <allocate>
ctx->priv = (struct vpx_codec_priv *)alg_priv;

The cast works because vpx_codec_alg_priv has a
vpx_codec_priv instance as a first member 'base'.

Change-Id: I10d1afc8c9a7dfda50baade8c7b0296678bdb0d0

10 years agoMerge "Removing unused arnr_type from VP9EncoderConfig and vp9_extracfg."
Dmitry Kovalev [Thu, 28 Aug 2014 20:50:05 +0000 (13:50 -0700)]
Merge "Removing unused arnr_type from VP9EncoderConfig and vp9_extracfg."

10 years agoMerge "Early termination in encoding partition search"
Yunqing Wang [Thu, 28 Aug 2014 20:49:39 +0000 (13:49 -0700)]
Merge "Early termination in encoding partition search"

10 years agoMerge "vp8_multi_resolution_encoder: Huge cleanup."
Dmitry Kovalev [Thu, 28 Aug 2014 19:13:29 +0000 (12:13 -0700)]
Merge "vp8_multi_resolution_encoder: Huge cleanup."