Jingning Han [Fri, 6 Sep 2013 23:06:21 +0000 (16:06 -0700)]
Fix overflow issue in 16x16 quantization SSSE3
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.
Deb Mukherjee [Thu, 29 Aug 2013 23:21:44 +0000 (16:21 -0700)]
Support a constant quality mode in VP9
Adds a new end-usage option for constant quality encoding in vpx. This
first version implemented for VP9, encodes all regular inter frames
using the quality specified in the --cq-level= option, while encoding
all key frames and golden/altref frames at a quality better than that.
The current performance on derfraw300 is +0.910% up from bitrate control,
but achieved without multiple recode loops per frame.
The decision for qp for each altref/golden/key frame will be improved
in subsequent patches based on better use of stats from the first pass.
Further, the qp for regular inter frames may also be varied around the
provided cq-level.
Scott LaVarnway [Fri, 6 Sep 2013 16:33:34 +0000 (12:33 -0400)]
New mode_info_context storage
mode_info_context was stored as a grid of MODE_INFO structs.
The grid now constists of a pointer to a MODE_INFO struct and
a "in the image" flag. The MODE_INFO structs are now stored
as a stream, eliminating unnecessary copies and is a little
more cache friendly.
For the test clips used, the decoder performance improved
by ~4.3% (1080p) and ~9.7% (720p).
Patch Set 2: Re-encoded clips with latest. Now ~1.7% (1080p)
and 5.9% (720p).
Jingning Han [Tue, 3 Sep 2013 18:57:34 +0000 (11:57 -0700)]
Enable 32x32 Transform unit test
This commit enabled a full functional test on 32x32 forward/inverse
transform, including round-trip error and memory overflow check. It
tests the prototype functions in C and all other implementations if
applicable.
Jingning Han [Wed, 4 Sep 2013 19:03:28 +0000 (12:03 -0700)]
Use saturated addition in SSSE3 of 32x32 quant
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.
The segment feature SEG_LVL_SKIP requires the prediction unit size
to be at least BLOCK_8X8. This commit makes the requirement to be
explicit. This is to prevent future encoder implementations from
making wrong choices.
There is another unit test that has been failing randomly on win32
build. Investigation has shown that the failure was caused by simd
register state is not reset appropriately in the fdct8x8 test. This
commit added ClearSystemState() in the teardown of this test, tests
showed it resolved the random failure issue for win32 build.
Related issue: https://code.google.com/p/webm/issues/detail?id=614
Jingning Han [Fri, 30 Aug 2013 17:57:23 +0000 (10:57 -0700)]
Fix 32x32 forward transform SSE2 version
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.
Tero Rintaluoma [Thu, 22 Aug 2013 08:29:19 +0000 (11:29 +0300)]
Fix intermediate height in convolve_c
- Intermediate height was not correct i.e. when block size is 4 and
y_step_q4 is 6. In this case intermediate height was
(4*6) >> 4 = 1 and vertical interpolation needs two source pixels
plus 7 extra pixels for taps.
- Also if the current output block is 16x16 and we are using 4x upscaling
we need only 12 rows after horizontal filtering instead of 16.
Patch Set 2: Intermediate_height updated after CL 66723
"Fix bug in convolution functions (filter selection)"
Paul Wilkins [Wed, 21 Aug 2013 11:34:14 +0000 (12:34 +0100)]
Added per pixel inter rd hit count stats
Added some code to output normalized rd hit count stats.
In effect this approximates to the average number of rd
operations/tests per pixel for the sequence.
The results are not quite accurate and I have not bothered
to account for partial SB64s at frame edges and for key frames
However they do give some idea of the number of modes /
prediction methods being tested for each pixel across the
different partition sizes. This indicates how much scope their
is for further gains either by reducing the number of partitions
examined or the modes per partition through heuristics.
Patch 3 moved place where count incremented so partial rd
tests that are aborted with INT_MAX return are also counted.
Example numbers for first 50 frames of Akiyo.
Speed 0 ~84.4 rd operations / pixel
Speed 1 ~28.8
Speed 2 ~11.9
Jingning Han [Mon, 26 Aug 2013 23:12:16 +0000 (16:12 -0700)]
Fix overflow issue in SSSE3 32x32 quantization
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.
Yaowu Xu [Thu, 29 Aug 2013 17:26:52 +0000 (10:26 -0700)]
Fixed potential overflows
The two arrays are typically initialized to INT64_MAX, if they are not
filled with valid values before the addition, the values can overflow
and lead to wrong results.
Scott LaVarnway [Thu, 29 Aug 2013 12:30:17 +0000 (08:30 -0400)]
Improved mb_lpf_horizontal_edge_w_sse2_8
This patch is a reformatted version of optimizations done by
engineers at Intel (Erik/Tamar) who have been providing
performance feedback for VP9. For the test clips used (720p, 1080p),
up to 1.2% performance improvement was seen.
Dmitry Kovalev [Wed, 28 Aug 2013 19:22:37 +0000 (12:22 -0700)]
General code cleanup.
Switching from mi_{width, height}_log2 and b_{width, height}_log2 to
num_8x8_blocks_{wide, high} and num_4x4_blocks_{wide, high}. Removing
redundant code, adding const.
Deb Mukherjee [Tue, 27 Aug 2013 22:07:50 +0000 (15:07 -0700)]
Adds a speed feature for fast 1-loop forw updates
Incorporates a speed feature for fast forward updates of
coefficients. This feature takes 3 values:
0 - use standard 2-loop version
1 - use a 1-loop version
2 - use a 1-loop version with reduced updates
Results: derfraw300 +0.007% (on speed 0) at feature value = 1
-0.160% (on speed 0) at feature value = 2
There is substantial speed up at speeds 2 and above for low
resolution sequences where the entropy updates are a big part
of the overall computations.
Yaowu Xu [Tue, 27 Aug 2013 15:39:20 +0000 (08:39 -0700)]
fixed the reading too many bytes
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.