This commit re-designs the quantization process for transform
coefficient blocks of size 4x4 to 16x16. It improves compression
performance for speed 7 by 3.85%. The SSSE3 version for the
new quantization process is included.
The average runtime of the 8x8 block quantization is reduced
from 285 cycles -> 255 cycles, i.e., over 10% faster.
Change-Id: I61278aa02efc70599b962d3314671db5b0446a50
This commit enables a fast path computational flow for forward
transformation. It checks the sse and variance of prediction
residuals and decides if the quantized coefficients are all
zero, dc only, or more. It then selects the corresponding coding
path in the forward transformation and quantization stage.
It is currently enabled in rtc coding mode. Will do it for rd
coding mode next.
In speed -6, the runtime for pedestrian_area 1080p at 1000 kbps
goes down from 14234 ms to 13704 ms, i.e., about 4% speed-up.
Overall coding performance for rtc set is changed by -0.18%.
Change-Id: I0452da1786d59bc8bcbe0a35fdae9f623d1d44e1
The SSSE3 implementation might find a potential overflow issue in
its second 1-D transform, if all input residual pixels are close to
255. This commit fixes the issue and re-enables the unit test on
the SSSE3 version.
Change-Id: I0520478abdab7afd3ff2842516bec951111e9b3c
Renames all x86_64 specific assembly files to consistently
end in _x86_64.asm. This will be useful for build systems to
handle these files differently.
All new 64-bit specific assembly files should use the new
naming convention.
Change-Id: I36c89584967c82ffc4088b1b5044ac15d2bb7536
This commit adjusts the forward 16x16 DCT computation steps to
simplify the register level operations. It fixes the corresponding
sse2 version accordingly.
Change-Id: I72a9c25b8ca9442fc5e113f47cd701ae55aa7f08
vp9_block_error_sse2 can only handle 16 bytes at a time but
the function requires to handle a sequence of 32 bytes at a time
so each 16 bytes is handled in a different register.
With AVX2 optimization the 32 bytes can be handled in one register instead
of two in the SSE2
The vp9_block_error was optimized by 85%.
The user level was optimized by 1.2%
Change-Id: Ia8fffe60e61eff7432a5fbd538757894f6c319fd
Includes changes that are not compatible with VS windows builds.
Amongst other things stdint.h is not supported in VS.
This reverts commit 89fbf3de50.
Change-Id: Ifa86d7df250578d1ada9b539c9ff12ed0c523cdd
This commit enables SSSE3 version full inverse 8x8 2D-DCT and
reconstruction. It makes the runtime of vp9_idct8x8_64_add down
from 256 cycles (SSE2) to 246 cycles.
Change-Id: I0600feac894d6a443a3c9d18daf34156d4e225c3
Assembly implementation of ssse3 8x8 forward 2D-DCT. The current
version is turned on only for x86_64. The average unit runtime
goes from 157 cycles down to 136 cycles, i.e., about 12.8% faster.
This translates into about 1.5% speed-up for pedestrian_area 1080p
at speed 2.
Change-Id: I0f12435857e9425ed7ce12541344dfa16837f4f4
We don't use declarations from this file. The real declarations
(differently named) are in vp9_rtcd_defs.pl, e.g. vp9_full_search_sad.
Change-Id: I73cbf064305710ba20747233cfdbe67366f069a0
2 functions were optimized for avx2 by using full 256 bit register
In order to handle 32 elements in parallel instead of only 16 in parallel:
1. vp9_sad32x32x4d
2. vp9_sad64x64x4d
The function level gain is 66% and the user level gain is ~1%.
Change-Id: I4efbb3bc7d8bc03b64b6c98f5cd5c4a9dd3212cb
* speed improvment of 30 percent achieved
* multiplies and adds remain the same
* non-arithmetic instructions minimized by hand, by:
-expanding 2 pass loop
-removing irrelivant "shuffles"
-combining last two rounding steps
* further improvments may be possible
Change-Id: Idec2c3f52910c48e6a0e0f9aefed5cae31b0b8c0
Optimizing 2 functions to process 32 elements in parallel instead of 16:
1. vp9_sub_pixel_avg_variance64x64
2. vp9_sub_pixel_avg_variance32x32
both of those function were calling vp9_sub_pixel_avg_variance16xh_ssse3
instead of calling that function, it calls vp9_sub_pixel_avg_variance32xh_avx2
that is written in avx2 and process 32 elements in parallel.
This Optimization gave 80% function level gain and 2% user level gain
Change-Id: Iea694654e1b7612dc6ed11e2626208c2179502c8
Optimizing 2 functions to process 32 elements in parallel instead of 16:
1. vp9_sub_pixel_variance64x64
2. vp9_sub_pixel_variance32x32
both of those function were calling vp9_sub_pixel_variance16xh_ssse3
instead of calling that function, it calls vp9_sub_pixel_variance32xh_avx2
that is written in avx2 and process 32 elements in parallel.
This Optimization gave 70% function level gain and 2% user level gain
Change-Id: I4f5cb386b346ff6c878a094e1c3b37e418e50bde
A bug was reported in Issue 702: "SIGILL (Illegal instruction) when
transcoding with vp9 - using FFmpeg". It was reproduced and fixed.
Change-Id: Ie32c149a89af02856084aeaf289e848a905c7700
Optimizing the variance functions: vp9_variance16x16, vp9_variance32x32,
vp9_variance64x64, vp9_variance32x16, vp9_variance64x32,
vp9_mse16x16 by migrating to AVX2
some of the functions were optimized by processing 32 elements instead of 16.
some of the functions were optimized by processing 2 loop strides of 16
elements in a single 256 bit register
This optimization gives between 2.4% - 2.7% user level performance gain
and 42% function level gain.
Change-Id: I265ae08a2b0196057a224a86450153ef3aebd85d
Modifications are done to reduce the total clock cycle.
Speedup: 1.2
Tested with: park_joy_420_720p50.y4m
Change-Id: Ia36b87e62e2f80a5fadaf5628729aedc80f38f3f
The step that sums three input samples could potentially cause the
intermediate result go beyond 16 bit limit, when operating as the
second 1-D transform. This commit fixes the issue.
Change-Id: Iaf512449ac2d25ddd8a806d760afab362c62a516
This patch fixed the issue reported in "Issue 655: remove textrel's
from 32-bit vp9 encoder". The set of vp9_subpel_variance functions
that used x86inc.asm ABI didn't build correctly for 32bit PIC. The
fix was carefully done under the situation that there was not
enough registers.
After the change, we got
$ eu-findtextrel libvpx.so
eu-findtextrel: no text relocations reported in 'libvpx.so'
Change-Id: I1b176311dedaf48eaee0a1e777588043c97cea82
For consistency with idct function names. Renames:
vp9_short_fdct4x4 -> vp9_fdct4x4
vp9_short_walsh4x4 -> vp9_fwht4x4
Change-Id: Id15497cc1270acca626447d846f0ce9199770f58
Just making fdct consistent with iht/idct/fht functions which all use
stride (# of elements) as input argument.
Change-Id: I0ba3c52513a5fdd194f1e7e2901092671398985b
Just making fdct consistent with iht/idct/fht functions which all use
stride (# of elements) as input argument.
Change-Id: Ibc944952a192e6c7b2b6a869ec2894c01da82ed1
Just making fdct consistent with iht/idct/fht functions which all use
stride (# of elements) as input argument.
Change-Id: I2d95fdcbba96aaa0ed24a80870cb38f53487a97d
Just making fdct consistent with iht/idct/fht functions which all use
stride (# of elements) as input argument.
Change-Id: Id623c5113262655fa50f7c9d6cec9a91fcb20bb4
Simplify the k_cvtlo_epi16 and k_cvthi_epi16 to only two
instructions. Then inlined them.
quoting from intel MMX_App_Compute_16bit_Vector.pdf
"The PMADDWD instruction multiplies four
pairs of 16-bit numbers and produces partial sums of the results
and can do so once per clock (with a three-clock latency)."
so I am assuming that there will be three clock overhead after the
last _mm_madd_pi16 command.
Even with the overhead the number of clocks in general should be
smaller. I am not sure though becasue I could not find information
about number of clocks required for instructions in k_cvtlo_epi16
and k_cvthi_epi16. I will run a test and compare the execution time.
Change-Id: Ieda4aa338f69ad3dd196ac6e7892da3cf1b47ea7
The 16x16 transform unit test suggested that the peak coefficient
value can reach 32639. This could cause potential overflow issue
in the SSSE3 implmentation of 16x16 block quantization. This commit
fixes this issue by replacing addition with saturated addition.
Change-Id: I6d5bb7c5faad4a927be53292324bd2728690717e
The 32x32 forward transform can potentially reach peak coefficient
value close to 32700, while the rounding factor can go upto 610.
This could cause overflow issue in the SSSE3 implementation of 32x32
quantization process.
This commit resolves this issue by replacing the addition operations
with saturated addition operations in 32x32 block quantization.
Change-Id: Id6b98996458e16c5b6241338ca113c332bef6e70
This commit fixed the potential overflow issue in the SSE2
implementation of 32x32 forward DCT. It resolved the corrupted
coded frames in the border of scenes.
Change-Id: If87eef2d46209269f74ef27e7295b6707fbf56f9
The 32x32 quantization process can potentially have the intermediate
stacks over 16-bit range, thereby causing enc/dec mismatch. This commit
fixes this overflow issue in the SSSE3 implementation, as well as the
prototype, of 32x32 quantization.
This fixes issue 607 from webm@googlecode.
Change-Id: I85635e6ca236b90c3dcfc40d449215c7b9caa806
In subpel_avg_variance functions, code similar to the following
punpkldq m2, [addr]
actually reads 8 bytes. For functions that are supposed to work on
buffers only have less 8 bytes a line, this caused valgrind error
of reading uninitialized memory.
Change-Id: I2a4c079dbdbc747829bd9e2ed85f0018ad2a3a34
Enable SSE2 implementation of high precision 32x32 forward DCT. The
intermediate stacks are of 32-bits. The run-time goes down from
32126 cycles to 13442 cycles.
Change-Id: Ib5ccafe3176c65bd6f2dbdef790bd47bbc880e56
This is in preparation for the SSE2 version of the high-precision
32x32 forward DCT which will share a lot of code with the existing
low precision version used for rate-distortion search.
Change-Id: I7084b6bdfb480b1fabb8493fb14e3f7fcc7888c0
This commit enables 16x16 ADST/DCT forward hybrid transform using SSE2
operations. It reduces the runtime from 5433 cycles to 1621 cycles, at
no compression performance loss.
Change-Id: I75fd7f1984e9e28846af459f810ff0d6ae125230
If none of the 16 coefficients that we quantize per loop iteration
are larger than the zbin, directly skip to the next round of coeffs,
rather than doing a full quantize loop that will eventually result
in 16 zeroes. This incurs a jump cost, but saves a lot of other work.
32x32 quant goes from 1349 -> 1184 cycles. The same approach yielded
no significantly positive results for smaller transforms, so is not
used there (8x8: 103 -> 101 cycles; 16x16: 302 -> 306 cycles).
Change-Id: I8fca17dc2543fc8eed1dbcd5100145e3c3a9b647
Encode time of bus (speed 0) 50 frames @ 1500kbps goes from 2min14.4 to
2min10.1, i.e. a 2.3% overall speed increase.
Change-Id: I3699580e74ec26c7d24e03681bc47ba25ee1ee87
Total encoding time for first 50 frames of bus (speed 0) @ 1500kbps
goes 2min34.8 to 2min14.4, i.e. a 10.4% overall speedup. The code is
x86-64 only, it needs some minor modifications to be 32bit compatible,
because it uses 15 xmm registers, whereas 32bit only has 8.
Change-Id: I2df53770c2e850813ffa713e1a91b45b0082b904
This commit enables SSE2 4x4 foward hybrid transform. The runtime
goes from 249 cycles down to 74 cycles. Overall around 2% speed-up
at no compression performance change.
Change-Id: Iad4d526346e05c7be896466c05500711bb763660
This commit replaces zrun_zbin_boost, a method of biasing non-zero
coefficients following runs of zero-coefficients to be rounded towards
zero, with an explicit skip-block choice in the RD loop.
The logic is basically that if individual coefficients should be rounded
towards zero (from a RD point of view), the trellis/optimize loop should
take care of it. If whole blocks should be zero (from a RD point of
view), a single RD check is much more efficient than a complete
serialization of the quantization loop.
Quality change: derf +0.5% psnr, +1.6% ssim; yt +0.6% psnr, +1.1% ssim.
SIMD for quantize will follow in a separate patch. Results for other
test sets pending.
Change-Id: Ife5fa641163ac5150ac428011e87188f1937c1f4
The aligned array in parameter list caused win32 build to report
c2719 error. This commit fixed the issue by make the parameter
type a pointer instead of an array.
Change-Id: I4ed654ce4eba2db4995d9cdc136c68e9a6acc992
Makes first 50 frames of bus @ 1500kbps encode from 3min22.7 to 3min18.2,
i.e. 2.3% faster. In addition, use the sub_pixel_avg functions to calc
the variance of the averaging predictor. This is slightly suboptimal
because the function is subpixel-position-aware, but it will (at least
for the SSE2 version) not actually use a bilinear filter for a full-pixel
position, thus leading to approximately the same performance compared to
if we implemented an actual average-aware full-pixel variance function.
That gains another 0.3 seconds (i.e. encode time goes to 3min17.4), thus
leading to a total gain of 2.7%.
Change-Id: I3f059d2b04243921868cfed2568d4fa65d7b5acd
