This reverts commit f1342a7b07.
This breaks 32-bit builds:
runtime error: load of misaligned address 0xf72fdd48 for type 'const
__m128i' (vector of 2 'long long' values), which requires 16 byte
alignment
+ _mm_set1_epi64x is incompatible with some versions of visual studio
Change-Id: I6f6fc3c11403344cef78d1c432cdc9147e5c1673
This function now has an AVX intrinsics version which is about 80%
faster compared to the C implementation. This provides a 2-4% total
speed-up for encode, depending on encoding parameters. The function
utilizes 3 properties of the cost function lookup table, constructed
in 'cal_nmvjointsadcost' and 'cal_nmvsadcosts'.
For the joint cost:
- mvjointsadcost[1] == mvjointsadcost[2] == mvjointsadcost[3]
For the component costs:
- For all i: mvsadcost[0][i] == mvsadcost[1][i]
(equal per component cost)
- For all i: mvsadcost[0][i] == mvsadcost[0][-i]
(Cost function is even)
These must hold, otherwise the AVX version of the function cannot be used.
Change-Id: I184055b864c5a2dc37b2d8c5c9012eb801e9daf6
A new version of vp9_highbd_error_8bit is now available which is
optimized with AVX assembly. AVX itself does not buy us too much, but
the non-destructive 3 operand format encoding of the 128bit SSEn integer
instructions helps to eliminate move instructions. The Sandy Bridge
micro-architecture cannot eliminate move instructions in the processor
front end, so AVX will help on these machines.
Further 2 optimizations are applied:
1. The common case of computing block error on 4x4 blocks is optimized
as a special case.
2. All arithmetic is speculatively done on 32 bits only. At the end of
the loop, the code detects if overflow might have happened and if so,
the whole computation is re-executed using higher precision arithmetic.
This case however is extremely rare in real use, so we can achieve a
large net gain here.
The optimizations rely on the fact that the coefficients are in the
range [-(2^15-1), 2^15-1], and that the quantized coefficients always
have the same sign as the input coefficients (in the worst case they are
0). These are the same assumptions that the old SSE2 assembly code for
the non high bitdepth configuration relied on. The unit tests have been
updated to take this constraint into consideration when generating test
input data.
Change-Id: I57d9888a74715e7145a5d9987d67891ef68f39b7
If high bit depth configuration is enabled, but encoding in profile 0,
the code now falls back on optimized SSE2 assembler to compute the
block errors, similar to when high bit depth is not enabled.
Change-Id: I471d1494e541de61a4008f852dbc0d548856484f
Rename updated version of x86inc.asm
Use "private_prefix" instead of "program_name" and make vpx the default
prefix.
Change-Id: I4883a99b2aee8e5dc9f2c16a2e6f4b5d6e4de458
This commit moves the module inverse transform functions from vp9
to vpx_dsp folder. The hybrid transform wrapper functions stay in
the vp9 folder, since it involves codec-specific data structures.
Change-Id: Ib066367c953d3d024c73ba65157bbd70a95c9ef8
This commit replaces vp9_idct.h with txfm_common.h in many SIMD
implementation files for precise file dependency.
Change-Id: If73dd726bb16537e7494f28538b0a169810f9756
Separate the common coefficient constant into vpx_dsp/txfm_common.h.
Move the SSE2 macro definitions to vpx_dsp/x86/txfm_common_sse2.h.
This clears the use case of vp9_idct.h in vpx_dsp folder.
Change-Id: I319735a2abf42888e5080ac14cfbcde34be7b121
This commit factors the 4x4, 8x8, and 16x16 2D-DCT forward
transform operations into vpx_dsp folder.
Change-Id: I084b117b79c0925edcbcabb93f62b9f4bf8dbe7d
The SSE2 version high bit-depth forward hybrid transforms are
essentially using the C functions via cross referencing to 1-D
functions in vp9_dct.c. This commit unifies the two versions and
removes the unnecessary dependency.
Change-Id: Ib4d0702a138f8daf7d0bd97c141ee7088f293765
The following quantization functions were moved:
vp9_quantize_b
vp9_quantize_b_32x32
vp9_highbd_quantize_b
vp9_highbd_quantize_b_32x32
vp9_quantize_dc
vp9_quantize_dc_32x32
vp9_highbd_quantize_dc
vp9_highbd_quantize_dc_32x32
The purpose of doing that was to allow these functions to be shared
by multiple codecs.
Change-Id: Id8ab939f283353cdd07bd930d47db3d932a5d87f
The clamp calls with INT32_MIN and INT32_MAX have no effect at all on
int values passed in, therefore this commit removes those effectless
clamps and also adds more const intermediate results to make the code
more readable.
Change-Id: I66d8811f58bb74ec31cbec9a6c441983a662352e
This commit fixes a potential integer overflow issue in function
hadamard_16x16. It adds corresponding dynamic range comment.
Change-Id: Iec22f3be345fb920ec79178e016378e2f65b20be
The only difference between the two was that the vp9 function allowed
for every step in the bilinear filter (16 steps) while vp8 only allowed
for half of those. Since all the call sites in vp9 (<< 1) the input, it
only ever used the same steps as vp8.
This will allow moving the subpel variance to vpx_dsp with the rest of
the variance functions.
Change-Id: I6fa2509350a2dc610c46b3e15bde98a15a084b75
this file shouldn't be built directly, it is included in vp9_dct_sse2.c
to create a non-high-bitdepth and a high-bitdepth version
silences missing prototype warnings for the unused FDCT* functions
Change-Id: Ide6ff8c24ab31bdb0f833260505ae33660a1ad5b
this file shouldn't be built directly, it is included in vp9_dct_sse2.c
to create a non-high-bitdepth and a high-bitdepth version
silences missing prototype warnings for the unused FDCT32x32* functions
Change-Id: I0e38f16dae5ea1728de184ee2c89287d48675c51
this file shouldn't be built directly, it is included in vp9_dct_avx2.c
to create a non-high-bitdepth and a high-bitdepth version
silences missing prototype warnings for the unused FDCT32x32* functions
Change-Id: I4c19935c0e035b393be513bde735e9a78064a494
With the sad functions, and hopefully the variance functions soon,
moving to the vpx_dsp location, place the defines used in the
reference C code in a common location.
Change-Id: I4c8ce7778eb38a0a3ee674d2f1c488eda01cfeca
this macro was used inconsistently and only differs in behavior from
DECLARE_ALIGNED when an alignment attribute is unavailable. this macro
is used with calls to assembly, while generic c-code doesn't rely on it,
so in a c-only build without an alignment attribute the code will
function as expected.
Change-Id: Ie9d06d4028c0de17c63b3a27e6c1b0491cc4ea79
Create a new component, vpx_dsp, for code that can be shared
between codecs. Move the SAD code into the component.
This reduces the size of vpxenc/dec by 36k on x86_64 builds.
Change-Id: I73f837ddaecac6b350bf757af0cfe19c4ab9327a
Force split on 16x16 block (to 8x8) based on the minmax over the 8x8 sub-blocks.
Also increase variance threshold for 32x32, and add exit condiiton in choose_partition
(with very safe threshold) based on sad used to select reference frame.
Some visual improvement near moving boundaries.
Average gain in psnr/ssim: ~0.6%, some clips go up ~1 or 2%.
Encoding time increase (due to more 8x8 blocks) from ~1-4%, depending on clip.
Change-Id: I4759bb181251ac41517cd45e326ce2997dadb577
This commit separates Hadamard transform/quantization operations
from rate and distortion computation in block_yrd. This allows one
to skip SATD computation when all transform blocks are quantized
to zero. It also uses a new block error function that skips
repeated computation of sum of squared residuals. It reduces the
CPU cycles spent on block error calculation in block_yrd by 40%.
Change-Id: I726acb2454b44af1c3bd95385abecac209959b10
This commit allows the quantizer to compare the AC coefficients to
the quantization step size to determine if further multiplication
operations are needed. It makes the quantization process 20% faster
without coding statistics change.
Change-Id: I735aaf6a9c0874c82175bb565b20e131464db64a
This commit fixes the SSE2 version 8x8 Hadamard transform
alignment and makes it consistent with the C version.
Change-Id: I1304e5f97e0e5ef2d798fe38081609c39f5bfe74
This commit replaces the 16x16 2D-DCT transform with Hadamard
transform for RTC coding mode. It reduces the CPU cycles cost
on 16x16 transform by 5X. Overall it makes the speed -6 encoding
speed 1.5% faster without compromise on compression performance.
Change-Id: If6c993831dc4c678d841edc804ff395ed37f2a1b
This commit uses Hadamard transform based rate-distortion cost
estimate for rtc coding mode decision. It improves the compression
performance of speed -6 for many hard clips at lower bit-rates.
For example, 5.5% for jimredvga, 6.7% for mmmoving, 6.1% for
niklas720p. This will introduce extra encoding cycle costs at
this point.
Change-Id: Iaf70634fa2417a705ee29f2456175b981db3d375
Move the scaling factor outside column projection. This avoids
repeated calculation of the same scaling factor. Profiling shows
that the percentage of vp9_int_pro_col_sse2 of overall cycles
goes from 2.29% down to 1.88%.
Change-Id: I5ac4e324ab2d7f33ba2de66dd2a12e04e04dfd66
This commit enables the rtc coding mode to run integral projection
based motion search for golden reference frame. It improves the
speed -6 compression performance by 1.1% on average, 3.46% for
jimred_vga, 6.46% for tacomascmvvga, and 0.5% for vidyo clips. The
speed -6 is about 6% slower.
Change-Id: I0fe402ad2edf0149d0349ad304ab9b2abdf0c804
This commit replaces the SAD with variance as metric for the
integral projection vector match. It improves the search accuracy
in the presence of slight light change. The average speed -6
compression performance for rtc set is improved by 1.7%. No speed
changes are observed for the test clips.
Change-Id: I71c1d27e42de2aa429fb3564e6549bba1c7d6d4d
In ssse3 functions, DEFINE_ARGS macro hard codes qcoeff and dqcoeff
to r3 and r4. If skip is 1, qcoeff and dqcoeff need to be loaded
from the stack, which doesn't work because of the above definitions.
Currently, skip=1 case is not used in the encoder. This patch fixed
the issue, so it can be turned on later.
Change-Id: I998d696b1a7a85dca2b3bcee790b21c21e039147
This commit introduces a new block match motion estimation
using integral projection measurement. The 2-D block and the nearby
region is projected onto the horizontal and vertical 1-D vectors,
respectively. It then runs vector match, instead of block match,
over the two separate 1-D vectors to locate the motion compensated
reference block.
This process is run per 64x64 block to align the reference before
choosing partitioning in speed 6. The overall CPU cycle cost due
to this additional 64x64 block match (SSE2 version) takes around 2%
at low bit-rate rtc speed 6. When strong motion activities exist in
the video sequence, it substantially improves the partition
selection accuracy, thereby achieving better compression performance
and lower CPU cycles.
The experiments were tested in RTC speed -6 setting:
cloud 1080p 500 kbps
17006 b/f, 37.086 dB, 5386 ms ->
16669 b/f, 37.970 dB, 5085 ms (>0.9dB gain and 6% faster)
pedestrian_area 1080p 500 kbps
53537 b/f, 36.771 dB, 18706 ms ->
51897 b/f, 36.792 dB, 18585 ms (4% bit-rate savings)
blue_sky 1080p 500 kbps
70214 b/f, 33.600 dB, 13979 ms ->
53885 b/f, 33.645 dB, 10878 ms (30% bit-rate savings, 25% faster)
jimred 400 kbps
13380 b/f, 36.014 dB, 5723 ms ->
13377 b/f, 36.087 dB, 5831 ms (2% bit-rate savings, 2% slower)
Change-Id: Iffdb6ea5b16b77016bfa3dd3904d284168ae649c
The high bit depth build failed while building for 32bit target.
The bugs were in vp9_highbd_subpel_variance.asm and
vp9_highbd_sad4d_sse2.asm functions. This patch fixed the bugs,
and made 32bit build work.
Change-Id: Idc8e5e1b7965bb70d4afba140c6583c5d9666b75
This patch was to fix issue 924:
https://code.google.com/p/webm/issues/detail?id=924
The SECTION_RODATA macro was modified to support macho32 format.
The sub-pixel functions were modified to pass in 2 more parameters
to handle the global offsets for PIC build.
Change-Id: I3bfcd336bcae945edf300bca4ab40376a2628cd4
For key frame at speed 6: enable the non-rd mode selection in speed setting
and use the (non-rd) variance_based partition.
Adjust some logic/thresholds in variance partition selection for key frame only (no change to delta frames),
mainly to bias to selecting smaller prediction blocks, and also set max tx size of 16x16.
Loss in key frame quality (~0.6-0.7dB) compared to rd coding,
but speeds up key frame encoding by at least 6x.
Average PNSR/SSIM metrics over RTC clips go down by ~1-2% for speed 6.
Change-Id: Ie4845e0127e876337b9c105aa37e93b286193405
Also removes some spurious changes in common/vp9_blockd.h which
was introduced by a rebase issue between nextgen and master branches.
Change-Id: If359f0e9a71bca9c2ba685a87a355873536bb282
(cherry picked from commit 005d80cd05)
(cherry picked from commit 08d2f54800)
(cherry picked from commit 4230c2306c)
This commit reworks the forward transform and quantization process
for 8x8 block coding. It combines the two operations in a single
function to save a store/load stage of the original transform
coefficients. Overall the speed -6 is slightly faster (around 1%
range). The compression performance of speed -6 is improved by
3.4%.
Change-Id: Id6628daef123f3e4649248735ec2ad7423629387
vp9_quantize_fp is the quantization process used by rtc coding
mode. This commit adds a sse2 implementation of it. The
implementation is modified based on vp9_quantize_b_sse2. No speed
difference from ssse3 version.
Change-Id: I24949c5b27df160b4f35117d28858d269454e64a
Combined vp9_denoiser_8xM_sse2 and vp9_denoiser_4xM_sse2 into one
function vp9_denoiser_NxM_sse2_small and passed the bitexact testing.
Changed the name of the function vp9_denoiser_64_32_16xM_sse2 to
vp9_denoiser_NxM_sse2_big.
Change-Id: Ib22478df585994dd347ebae04202c0b701e7f451
All sad function that process above 32 consecutive elements are optimized
for AVX2:
vp9_sad64x64
vp9_sad64x32
vp9_sad32x64
vp9_sad32x32
vp9_sad32x16
vp9_sad64x64_avg
vp9_sad64x32_avg
vp9_sad32x64_avg
vp9_sad32x32_avg
vp9_sad32x16_avg
The functions that appeared as a hotspot is vp9_sad32x32 and vp9_sad64x64
vp9_sad32x32 was optimized by 68% and vp9_sad64x64 was optimized by 90%
both of them gave and overall ~2.3% user level gain
Change-Id: Iccf86b375a2b54c5fbbe685902ead0c9a561b9fd
The concept:
There's too much noise in source pixels for variance and at low bitrate
the reconstructed looks nothing like the source so we have problems
getting good partitionings with either. This skirts the issue by using
a box blur scaled down version for variance calculations. To compare
against source_var_ moved keyframe to be rd based like source_var.
Change-Id: Ie3babdbfadae324b7b5a76bea192893af27f0624
This SSE2 is based on VP8 denoiser's SSE2 code. In VP8, there are
only 16x16 blocks in denoiser, while in VP9, there are 13 different
block sizes.
By adding this SSE2 code, the improvement of encoder speed is around
20%(using C code vs using SSE2 code), vary for different clips.
The unit test for VP9 denoiser is to confirm that the SSE2 code is
bit-exact with the C code. The unit test covers all block size.
Change-Id: Ic8d8ac26db4ea40a5f146b5678a065af07eaaa3d
