Marco [Tue, 14 Jul 2015 17:01:45 +0000 (10:01 -0700)]
Fix to source scaling for dynamic_resize.
The fast scaling for 1 pass mode was being used only on the
first frame after resizing event (because resize_scale_num/den
is set to 1 and only changed for first frame following resize event).
Fixes part of merge regression from adding arf parameters.
From Change Ibf0c30b72074b3f71918ab278ccccc02a95a70a0
There is still an issue relating to one animated test clip with repeat
patterns where this change effectively increase the default maximum
arf interval by +1. This can be examined seperately.
Jingning Han [Mon, 13 Jul 2015 18:49:57 +0000 (11:49 -0700)]
Refactor intra block prediction and reconstruction process
Flaten the intra block decoding process. It removes the legacy
foreach_transformed_block use in the decoder. This saves cycles
spent on retrieving the transform block position.
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.
Jingning Han [Wed, 8 Jul 2015 16:15:39 +0000 (09:15 -0700)]
Refactor transform block loop for inter mode decoding
Rework the inter mode transform block decoding loop. Replace the
block index with the row and col index as the input argument. It
saves function call to compute the row and col index according to
the block index and overall block size, and many if statements
associated with the transform block position relative to the coding
block. For the test bit-stream pedestrian_area 1080p at 5 Mbps,
the decoding speed goes up from 81.13 fps to 81.92 fps.
Note that the intra coded block decoding needs more refactoring
work than the inter ones. So keep it using foreach_transforme_block
as for now.
paulwilkins [Tue, 30 Jun 2015 09:18:42 +0000 (10:18 +0100)]
Add extra resize trigger for frames above maximum allowed size.
Even if the recode loop is not enabled for the current frame type
trap the case where the projected size of a a frame is above the
maximum allowed in recode_loop_test()
paulwilkins [Wed, 24 Jun 2015 10:36:51 +0000 (11:36 +0100)]
Changes to use of rectangular partitions.
Changes to allow more use of rectangular partitions at
speeds 1 and 2 for content classed by the first pass as
animation and for blocks near the active image edge.
This has quite a big impact in quality for the animated
test sequence but also hurts encode speed for speed 2.
For other content types the impact on both speed and
quality is small.
Added some plumbing for detection of internal vertical
image edges.
paulwilkins [Wed, 24 Jun 2015 10:36:51 +0000 (11:36 +0100)]
Change speed and rd features for formatting bars.
Change speed features / behavior for split mode when there
is an internal active edge (e.g. formatting bars).
Remove some threshold constraints in rd code near the active
edge of the image.
Add some plumbing for left and right active edge detection.
Patch set 5. Limit rd pass through for sub 8x8 to internal active edges.
This takes away any speed penalty for most clips but keeps the enhanced
edge coding for the more critical case of internal image edges
Jingning Han [Tue, 7 Jul 2015 22:32:27 +0000 (15:32 -0700)]
Refactor inverse_transform_block argument list
Replace block index with transform type in the argument list. This
allows to save an extra fetch to the prediction mode. For pedestrian
area 1080p coded at 5 Mbps with single tile, the average decoding
speed goes up from 80.55 fps (before the refactoring series) to
81.13 fps.
Jingning Han [Tue, 7 Jul 2015 18:36:05 +0000 (11:36 -0700)]
Reduce dqcoeff array size in decoder
The decoding process handles detokenization and reconstruction per
transform block sequentially. There is no need to offset the dqcoeff
buffer according to the transform block index. This allows to
reduce the memory spill and improve cache performance.
Jingning Han [Mon, 6 Jul 2015 23:52:24 +0000 (16:52 -0700)]
Unify subtract function used in VP8/9
This commit replaces the vp8_ prefixed subtract function with the
common vpx_subtract_block function. It removes redundant SIMD
optimization codes and unit tests.
Adds two new vp9 parameters --min-gf-interval and --max-gf-interval
to enable testing based on frequency of alt-ref frames.
Also adds a unit-test to test enforcement of min-gf-interval.
For both these parameters the default value is 0, which indicates
they are picked by the encoder, based on resolution and framerate
considerations. If they are greater than zero, the specified
parameter is honored.
(Additional note by paulwilkins)
Note that there is a slight oddity in that key frames are also GFs and
considered part of GF only group. However they are treated as not
being part of an arf group because for arf groups the previous GF is
assumed to be the terminal or overlay frame for the previous group.
this change causes MSA/VP9SubpelVarianceTest.Ref and
MSA/VP9SubpelVarianceTest.ExtremeRef failures under
mips32r5el-msa-linux-gnu and mips64r6el-msa-linux-gnu
The vp9_lpf_vertical_16_dual function optimized for x86 32bit target. The hot code in that function was caused by the call to the transpose8x16.
The gcc generated assembly created uneeded fills and spills to the stack. By interleaving 2 loads and unpack instructions, in addition to hoisting the consumer
instruction closer to the producer instructions, we eliminated most of the fills and spills and improve the function-level performance by 17%.
credit for writing the function as well as finding the root cause goes to Erik Niemeyer (erik.a.niemeyer@intel.com)
projects / libvpx / log