We pick a mapping with quality 0 mapping to max diff 32, to quality 100
mapping to max_diff 1.
For 6k GIF image set, this improves compression by:
4% at quality 0
0.05% at quality 75
Benefits the MovingThumbnailer test videos too.
Change-Id: I6838ce864d41e1e65311d26b9b8115a12390a253
This way we can ignore some noisy pixels and get tighter frame
rectangles.
Some results:
- Correctness:
Tested that anim_diff reports all images are identical for lossless, and
similar min_psnr value for lossy and mixed modes.
Also checked output images visually to make sure there weren't any
obvious kinks.
- Compression:
A very tiny improvement for 6000 image GIF set we have (0.03%) for lossy
and mixed mode. For some of these images, frames get dropped
automatically as they have a very small diff from previous frame.
10 images from test_video_frames_png show a clear improvement in
compression though. This CL leads to 7 out of 9 lossy WebPs getting
smaller -- for one of them, this leads to a higher quality being picked
(as that’s still < 150 KB).
Change-Id: If539b9e77e1375aa15edc8f926933593a9865f1c
and also pass 'VP8Io* io' extra param to VP8DecompressAlphaRows()
This is somehow in preparation for some memory optimizations in
the 'cropping' case. For now, only the easy crop_bottom case is
optimized.
Change-Id: Ib54531ba057bf62b98422dbb6c181dda626c72c2
This avoids generating file that would trigger a decoding bug
found in 0.4.0 -> 0.4.3 libwebp versions.
This reverts commit 6ecd72f845.
Change-Id: I4667cc8f7b851ba44479e3fe2b9d844b2c56fcf4
The mode's bits were not taken into account, which is ok for most of cases.
But in case of super large image, with 'easy' content, their overhead starts
mattering a lot and we were omitting to optimize for these.
Now, these mode bits have their own lambda values associated, limiting
the jerkiness. We also limit (for -m 2 only) the individual number of bits
to something that will prevent the partition 0 overflow.
removed the I4_PENALTY constant, which was a rather crude approximation.
Replaced by some q-dependent expression.
fixes issue #289
Change-Id: I956ae2d2308c339adc4706d52722f0bb61ccf18c
If value is '2', it means the buffer is a 'slow' one, like GPU-mapped memory.
This change is backward compatible (setting is_external_memory to 2
will be a no-op in previous libraries)
dwebp: add flags to force a particular colorspace format
new flags is:
-pixel_format {RGB,RGBA,BGR,BGRA,ARGB,RGBA_4444,RGB_565,
rgbA,bgrA,Argb,rgbA_4444,YUV,YUVA}
and also,external_memory {0,1,2}
These flags are mostly for debuggging purpose, and hence are not documented.
Change-Id: Iac88ce1e10b35163dd7af57f9660f062f5d8ed5e
This is in preparation for some SSE2 code.
And generally speaking, the whole SSIM code needs some
revamp: we're not averaging the SSIM value at each pixels
but just computing the overall SSIM value once, for the whole
plane. The former might be better than the latter.
Change-Id: I935784a917f84a18ef08dc5ec9a7b528abea46a5
based on the sse2 change in:
9960c31 Remove an unnecessary transposition in TTransform.
~9-10.5% faster at the function-level, < 1% overall
Change-Id: I44413369b230b250fb0dbc51ff2f17cfeda609b7
- The result is now indeed closest among possible results for all inputs, which
was not the case for bits>4, where the mapping was not even monotonic because
GetValAndDistance was correct only if the significant part of initial fit in
a byte at most twice.
- The set of results for a larger number of bits dropped is a subset of values
for a smaller number of bits dropped. This implies that subsequent
discretizations for a smaller number of bits dropped do not change already
discretized pixels, which improves the quality (changes do not accumulate)
and compression density (values tend to repeat more often).
- Errors are more fairly distributed between upwards and downwards thanks to
bankers’ rounding, which avoids images getting darker or lighter in overall.
- Deltas between discretized values are more repetitive. This improves
compression density if delta encoding is used.
Also, the implementation is much shorter now.
Change-Id: I0a98e7d5255e91a7b9c193a156cf5405d9701f16
Pass them along to internal 'pic' object, so that progress can be reported back
and user data can also be inspected.
Change-Id: Idb5d0d4a76d07283d704a86c5892e1ad7bda09fa
SSE4.1 is slower than the SSE2 implementation and this seems to
be due to a slow _mm_loadl_epi64 implementation by gcc
(hence a bug with my gcc 4.8) and a very slow _mm_hadd_epi32. Both
got confirmed by IACA and experiments.
Change-Id: I05607f66b7ccd8f4f42e000693aea583ffd5768f
We were not updating the current_width_, which is usually
not a problem, unless we use Delta Palette with small number
of colors
-> Addressed this re-entrancy problem by checking we have
enough capacity for transform buffer.
The problem is not currently visible, until we restrict
the number of gradient used in delta-palette to less than 16.
Then the buffers have different current_width_ and the problem
surfaces.
Change-Id: Icd84b919905d7789014bb6668bfb6813c93fb36e
The transpose refactoring will help removing a transpose in a
later CL.
The horizontal add function helps removing a _mm_sad_epu8 in DC8uv
=> the latency/throughput went from 29/25 to 23/19
Change-Id: I5f3dfd4aad614eb079b1e83631e6a7cef49a3766
'implicit conversion from 'int' to 'short' changes value from 33050 to
-32486'
original patch:
https://codereview.chromium.org/1657313003/
Make libwebp build with -Wconstant-conversion from newer clangs.
After http://llvm.org/viewvc/llvm-project?rev=259271&view=rev, clang
points out that _mm_set1_epi16(33050) causes an overflow in the short
argument to _mm_set1_epi16(). Since there's no version that takes an
unsigned short, add an explicit cast to tell the compiler that this is
intentional.
No behavior change.
Change-Id: I6b4e3401b15cfbcc895f9e81b5c2dc59d43ffb9b
The code and logic is unified when computing bit entropy + Huffman cost.
Speed-wise, we gain 8% for lossless encoding.
Logic-wise, the beginning/end of the distributions are handled properly
and the compression ratio does not change much.
Change-Id: Ifa91d7d3e667c9a9a421faec4e845ecb6479a633
setting all transparent pixels to black rather than the "flatten" method.
0.3% smaller filesize on the 1000 PNGs if alpha cleanup is used (before: 18685774, after: 18622472)
Change-Id: Ib0db9e7ccde55b36e82de07855f2dbb630fe62b1
The functions containing magic constants are moved out of ./dsp .
VP8LPopulationCost got put back in ./enc
VP8LGetCombinedEntropy is now unrefined (refinement happening in ./enc)
VP8LBitsEntropy is now unrefined (refinement happening in ./enc)
VP8LHistogramEstimateBits got put back in ./enc
VP8LHistogramEstimateBitsBulk got deleted.
Change-Id: I09c4101eebbc6f174403157026fe4a23a5316beb
This implementation brings:
- an SSE implementation of packing / unpacking
- bigger buffers processed at the same time
The speedup is of 4% on lossy decoding (YUV to RGB), 0.5% on
lossy encoding (RGB to YUV was already optimized).
Change-Id: Iec677ee17f91c08614d1adab67c6df551925767f
* changes:
demux: remove GetFragment()
demux: remove dead fragment related TODO
demux, Frame: remove is_fragment_ field
demux,WebPIterator: remove fragment_num/num_fragments
demux: remove WebPDemuxSelectFragment
this hasn't been set since parsing of the experimental chunk was
removed.
+ cleanup IsValidExtendedFormat(). is_fragmented has caused immediate
failure since:
4e2589f demux: restore strict fragment flag check
Change-Id: If9ecfc19556297100a6d5de1ba2cffdcbdc6c8fd
INFO: From Compiling src/dsp/cpu.c:
src/dsp/cpu.c: In function 'x86CPUInfo':
src/dsp/cpu.c:36:3: inconsistent operand constraints in an 'asm'
With PIC and mcmodel=medium, the %rbx register must be saved and
restored which causes this problem. This was also solved in GCC-4.9 with
this patch:
https://gcc.gnu.org/ml/gcc-patches/2012-12/msg01484.html
Tested:
Builds fine with this change.
Change-Id: Icca8eea7bf5af3ef9f17f6ae2886e3430143febf
CombinedShannonEntropy takes 30% for lossless compression.
This implementation speeds up the overall process by 2 to 3 %.
Change-Id: I04a71743284c38814fd0726034d51a02b1b6ba8f
The previous priority system used a heap which was too heavy to
maintain (what was gained from insertions / deletions was lost
due to a linear that still happened on the heap for invalidation).
The new structure is a priority queue where only the head is
ordered.
Change-Id: Id13f8694885a934fe2b2f115f8f84ada061b9016
SimpleQuantize()
it's now a single function, that reconstructs the intra4x4 block during the scan
The I4_PENALTY had to be adjusted.
Overall, result is better quality-wise (esp. at q < 50), and a tad faster too.
method #0, #1 and #3+ are unchanged
Change-Id: If262aeb552397860b3dd532df8df6b1357779222
* Precision is slightly different
* also implemented in SSE2 the missing WebPUpsamplers for MODE_ARGB, MODE_Argb, MODE_RGB565, etc.
* removing yuv_tables_sse2.h saved ~8k of binary size
* the mips32/mips_dsp_r2 code is disabled for now, since it has drifted away
* the NEON code is somewhat tricky
Change-Id: Icf205faa62cf46c2825d79f3af6725dc1ec7f052
Gives 0.9% smaller (2.4% compared to before alpha cleanup) size on the 1000 PNGs dataset:
Alpha cleanup before: 18856614
Alpha cleanup after: 18685802
For reference, with no alpha cleanup: 19159992
Note: WebPCleanupTransparentArea is still also called in WebPEncode. This cleanup still helps
preprocessing in the encoder, and the cases when the prediction transform is not used.
Change-Id: I63e69f48af6ddeb9804e2e603c59dde2718c6c28
The 32-bit buffers are actually rarely 64-bit aligned.
The new solution uses memcmp and is alignment agnostic.
It is also slightly faster.
Change-Id: I863003e9ee4ee8a3eed25b7b2478cb82a0ddbb20
for more speed.
This gives a roughly a 1% speedup for low_effort. But actually this is a
preparation for the upcoming CL that changes RGB values of transparent pixels
based on prediction, which should not be done for low_effort because that would
slightly hurt its performance.
On 1000 PNGs, with quality 0, method 0:
Before:
Compression (output/input): 2.9120/3.2667 bpp, Encode rate (raw data): 36.034 MP/s
After:
Compression (output/input): 2.9120/3.2667 bpp, Encode rate (raw data): 36.428 MP/s
Change-Id: I5ed9f599bbf908a917723f3c780551ceb7fd724d
Arrays were compared 32 bits at a time, it is now done 64 bits at a time.
Overall encoding speed-up is only of 0.2% on @skal's small PNG corpus.
It is of 3% on my initial 1.3 Mp desktop screenshot image.
Change-Id: I1acb32b437397a7bf3dcffbecbcd4b06d29c05e1
The same computation was done for both values: go over two buffers,
sum them up, and take a decision on the sum at each iteration.
MIPS32 code has been disabled for now, pending a code update.
Change-Id: I997984326f7092b3dbb8cfa1e524bd8132b2ab9d
instead of per block. This prepares for a next CL that can make the
predictors alter RGB value behind transparent pixels for denser
encoding. Some predictors depend on the top-right pixel, and it must
have been already processed to know its new RGB value, so requires per
scanline instead of per block.
Running the encode speed test on 1000 PNGs 10 times with default
settings:
Before:
Compression (output/input): 2.3745/3.2667 bpp, Encode rate (raw data): 1.497 MP/s
After:
Compression (output/input): 2.3745/3.2667 bpp, Encode rate (raw data): 1.501 MP/s
Same but with quality 0, method 0 and 30 iterations:
Before:
Compression (output/input): 2.9120/3.2667 bpp, Encode rate (raw data): 36.379 MP/s
After:
Compression (output/input): 2.9120/3.2667 bpp, Encode rate (raw data): 36.462 MP/s
No effect on compressed size, this produces exactly same files. No
significant measured effect on speed. Expected faster speed from better
memory layout with scanline processing but slower speed due to needing
to get predictor mode per pixel, may compensate each other.
Change-Id: I40f766f1c1c19f87b62c1e2a1c4cd7627a2c3334
the problem was the incorporation of the extra constant 1<<16 in the kC1
constant, to emulate the addition. It's now removed and the addition is
performed explicitly.
No real speed difference observed.
cf. issue #278
Change-Id: I2c6499031571d98afff392fb5ebe21a5fa60722d
* changes:
Makefile.vc: enable WEBP_USE_THREAD for windows phone
thread: use CreateThread for windows phone
thread: use WaitForSingleObjectEx if available
thread: use InitializeCriticalSectionEx if available
thread: use native windows cond var if available
if FANCY_UPSAMPLING was not defined but io->fancy_upsampling was set,
then the call to WebPInitSamplers() was skipped -> boom.
Change-Id: Id63e2ecc09f532fbe2ec9936d9ce4b502ba8fac5
Rename the flag to exact instead of the opposite cleanup_alpha. Add the flag to
WebPConfig. Do the cleanup in the webp encoder library rather than the cwebp
binary, this will be needed for the next stage: smarter alpha cleanup for
better compression which cannot be done as a preprocessing due to depending on
predictor choices in the encoder.
Change-Id: I2fbf57f918a35f2da6186ef0b5d85e5fd0020eef
the original change triggered several internal API modifs.
This is to ensure that we're never computing pointer that can
possibly wrap around, or differences between pointers that can
overflow.
no observed speed difference
Change-Id: I9c94dda38d94fecc010305e4ad12f13b8fda5380
We now consider 3 special cases:
* htree-group has only 1 code (no bit is read from bitstream)
* htree-group has few enough literal symbols, so that all the bit
codes can fit into a look-up table of less than 64 entries
* htree-group has a trivial arb literal (not GREEN!), like before
No overall speed change.
Change-Id: I6077fa0b7e5c31a6c67aa8aca859c22cc50ee254
We now get error string instead of printing it.
The verbose option is now only used to print info and warnings.
Change-Id: I985c5acd427a9d1973068e7b7a8af5dd0d6d2585
It was needed earlier for WebPAnimEncoder API when it was using structs
like WebPConfig, but it only uses pointers to those now.
Change-Id: Ic0c144966421c678e8ef54b3fa81574bb2c9cd08
global effect is ~2% faster encoding from JPG source
and ~8% faster lossless-webp source decoding to PGM (e.g.)
Also revamped the YUVA case to first accumulate R/G/B value into 16b
temporary buffer, and then doing the UV conversion.
-> New function: WebPConvertRGBA32ToUV
Change-Id: I1d7d0c4003aa02966ad33490ce0fcdc7925cf9f5
Just for RGB24/BGR24 for now, which are the hard-to-optimize ones.
SSE2 implementation coming next.
ConvertRowToY() should go into dsp/ too, at some point.
Change-Id: Ibc705ede5cbf674deefd0d9332cd82f618bc2425
also switch to using ExtractAlpha() instead of hard-coding the loop.
The ARGBToY/UV functions are rather easy to port to SSE2 / NEON.
Change-Id: I8f1346a9ca427a36ce2d6c848369ca7964d8b3c7
use 'u' rather than the unnecessary 'l' as a suffix. this prevents a
conversion warning with some toolchains
Change-Id: I21c33ce08819b3c839c75e03a8f7f3a6041d0695
Note that ALIGN_CST is still kept different in dec/frame.c for now,
because the values is 31 there, not 15. We might re-unite these two
later.
Change-Id: Ibbee607fac4eef02f175b56f0bb0ba359fda3b87
same functionality, but better code layout.
What changed:
* don't trash the palette_[] in EncodePalette(), so it can be re-used
* split generation of image from bit-stream coding
* move all the delta-palette code to delta_palettization.c, and only have 1 entry point there WebPSearchOptimalDeltaPalette()
* minimize the number of "#ifdef WEBP_EXPERIMENTAL_FEATURES" in vp8l.c
* clarify the TransformBuffer stuff. more clean-up to come here...
This should make experimenting with delta-palettization easier and more compartimentalized.
Change-Id: Iadaa90e6c5b9dabc7791aec2530e18c973a94610
some limitations: only for RGBA output,
and if reduction factor is not too small (dst_width > src_width / 128)
20-25% faster, ~4-6% global improvement total decoding.
Change-Id: I95366ddaa4a38e0a96bed754dfe790126f7bb84a
It's better to stay with a 32b fixed-point precision overall, otherwise
the C-version on ARM gets *slower*.
Actually, gcc ARM compiler optimizes some instructions pretty
well when WEBP_RESCALER_FIX is exactly 32, even in C.
Change-Id: I0eea97f7db5947470f5af355dee098eca81e178d
New palette compresses more than 20% better with minimum quality loss.
Tested on set of wikipedia images with command line:
cwebp -delta_palettization
Change-Id: I82ec7d513136599cd70386f607f634502eb9095d
Earlier, we stored a 1x1 frame for such frames. Now, we drop every such
frame and increase the duration of its previous frame instead.
Also, modify the anim_diff tool to handle animated images that are
equivalent, but have different number of frames.
Change-Id: I2688b1771e1f5f9f6a78e48ec81b01c3cd495403
The rounding and arithmetic is not the same as previously, to prevent overflow cases for large upscale factors.
We still rely on 32b x 32b -> 64b multiplies. Raised the fixed-point precision to 32b
so that we have some nice shifts from epi64 to epi32.
Changed rescaler_t type to 'uint32_t' in order to squeeze in all the precision required.
The MIPS code has been disabled because it's now out-of-sync. Will be fixed in
a subsequent CL when the dust settles.
~30-35% faster
Change-Id: I32e4ddc00933f1b1aa3463403086199fd5dad07b
* vertical expansion now uses bilinear interpolation
* heavily assumes that the alpha plane is decoded in full, not row-by-row
* split the RescalerExportRow and RescalerImportRow methods into Shrink
and Expand variants.
* MIPS implementation of ExportRowExpand is missing.
There's room for extra speed optim and code re-org, but let's keep that for later patches.
addresses https://code.google.com/p/webp/issues/detail?id=254
Change-Id: I8f12b855342bf07dd467fe85e4fde5fd814effdb
This is designed for the simple use-case where one wants to decode all
frames one-by-one in order.
Also, use this API in anim_util library, which is in turn used by
anim_diff tool.
Change-Id: Ie8b653c04e867d40fd23321b3dd41b87689656c7
This makes the chains more efficient and a larger variety of data is tested.
0.02 % compression gain at q 100, 0.05 % at default quality. 0.8 % speedup by
callgrind.
0.16 % compression gain for lossy alpha ?!
Change-Id: I888120133352799eb14f5f602c7f40ab404bd665
this allows scaling to a particular width/height while preserving the
source aspect ratio using WebPRescalerGetScaledDimensions().
Change-Id: I77b11528753290c1e9bb942ac761c215ccfb8701
using a *tmp_plane buffer to split a/r/g/b planes up appeared to
be the easiest route, compared to copy-pasting the whole code and
making it x_stride aware...
Change-Id: I0898ef1df62bd3e1713b77187b31b5eeef3832fe
allows the values to be used in preproc checks, fixing a
-Wunreachable-code warning in 64-bit builds where VP8L_WRITER_BITS != 16
Change-Id: Ie98dff4e8ef896436557c64d5da2c5d70228a730
Slightly faster on -m 0 -q 0, particularly for small images (50 x 75
image was 0.1 % faster on callgrind measurement).
Increases compression density by 0.005 % for the 1000 images, but small
images can improve even 0.5 % (about 4 bytes, depending on the
characteristics of the palette).
Change-Id: I94f568d396ac62a054a829abeeef3eb0af6b3f94
the x_add/x_sub increments were wrong for u/v in the upscaling case.
They shouldn't be left to the caller's discretion, but set up by
WebPRescalerInit to their exact necessary values.
-> Cleaned-up WebPRescalerInit() param list.
-> added safety asserts
-> removed the mips32/mips_r2 variant of "ImportRow" which were buggy prior
Change-Id: I347c75804d835811e7025de92a0758d7929dfc09
this moves the function outside the WEBP_USE_INTRINSICS check.
there's no alternative version and it's ~54% faster at the
function level and mildly faster overall
Change-Id: Ibc648e9ee35021d48901e05aa596aa01067796a2
a total impact of 1 % on encoding speed
This allows for performance neutral removal of the binary search
in cache bits selection. This will give a small improvement in
compression density.
Change-Id: If5d4d59460fa1924ce71af977320834a47c2054a
0.21 % compression density improvement for 1000 png corpus in
lossless mode
0.50 % compression density improvement for 1000 png corpus in
lossy mode
Change-Id: I14ee8c427ae5d3e116b0ee6695fcdea3321a319d
valgrind --tool=callgrind shows a 9 % speedup: 1021201984 ticks before vs.
927917709 after
-q 0 -m 0 -lossless ~/alpi/1.png
22.040 MP/s before
24.796 MP/s after
Change-Id: Iaab928167b3e20fb0d9401c6f8317a26c5a610b4
* changes:
lossless: combine the Huffman code with extra bits
lossless: Inlining add literal
lossless: simplify HashChainFindCopy heuristics
lossless: 0.5 % compression density improvement
lossless: Add zeroes into the predicted histograms.
lossless: encoding, don't compute unnecessary histo
lossless: Remove about 25 % of the speed degradation
Faster alpha coding for webp
lossless: rle mode not to accept lengths smaller than 4.
lossless: Less code for the entropy selection
lossless: 0.37 % compression density improvement
_BitScanReverse() takes an unsigned long*
http://msdn.microsoft.com/en-us/library/fbxyd7zd.aspx
fixes:
C4057: 'function': 'unsigned long *' differs in indirection to slightly
different base types from 'uint32_t *'
fixes issue #253
Change-Id: I0101ef7be18c7ed188b35e9b17e7f71290953786
do not do length 2 matches far away
speedup for non compressible data by inserting two literals at a time
when no matches are found
Change-Id: Ia8e033071f4186bb8148bb2bf13ca37586734aa3
Increases compression density by 0.03 % for lossy.
Speeds up at least one of the lossy alpha images by 20 %.
Palette entropy 'kludge' seems to save 1-2 % on alpha images.
Change-Id: I2116b8d81593ac8173bfba54a7c833997fca0804
share the computation between different modes
3-5 % speedup for lossless alpha
1 % for lossy alpha
no change in compression density
Change-Id: I5e31413b3efcd4319121587da8320ac4f14550b2
introduced in:
"lossless: 0.37 % compression density improvement"
Uses the statistics of red and blue histograms to decide if to run
cross color correction at all.
Improves compression density by 0.02 % or so.
Change-Id: I47429557e9cdbd9fa90c584696f241b17427d73f
No significant size degradation (+0.001 %) for 1000 image corpus
Fixes the 8 ms vs 2 ms degradation from:
"lossless: 0.37 % compression density improvement"
Change-Id: Id540169a305d9d5c6213a82b46c879761b3ca608
counting the entropy expectation for five different configurations:
palette
non-predicted
non-predicted with subtract green
predicted
predicted with subtract green
and choose the strategy with the smallest expected entropy
Change-Id: Iaaf209c0d565660a54a4f9b3959067afb9951960
this should be used in preference to free() for releasing memory
returned from WebPDecode*() / WebPEncode*(). this simplifies memory
management when working through language bindings
Change-Id: I15eb538a45390efc552fda8e5c251a3fbdc13c29
After several trials at re-organizing the main loop and accumulation scheme,
this is apparently the faster variant.
removed the SSE41 version, which is no longer faster now.
For some reason, the AVX variant seems to benefit most for the change.
Change-Id: Ib11ee18dbb69596cee1a3a289af8e2b4253de7b5
this moves the function outside the WEBP_USE_INTRINSICS check.
there's no alternative version and it's ~70% faster at the
function level and 1-2% faster overall
Change-Id: I59fb4918ec86b1ac3a47cbd5d05ce62f007461cb
Changed the code (again) to process 4 pixels at a time. Loop is more
involved, but overall it's faster.
Removed the SSE4.1 implementation which is now slower than SSE2.
Change-Id: I7734e371033ad8929ace7f7e1373ba930d9bb5f1
New implementations: SubtractGreenFromBlueAndRed and TransformColor
around 1-2% faster lossless encoding.
Change-Id: I1668e36fdc316ba55b3b798b91b4a3e36ce62861
DispatchAlpha* functions are hard to speed up, compared to SSE2.
ExtractAlpha sees a ~15% speed-up though.
Change-Id: I8715c2defecbc832f469eed7e6ffd012146b52de
over a 1000 image corpus
Single photograph benchmark:
Before:
Q=20: 2.560 MP/s
Q=40: 2.593 MP/s
Q=60: 1.795 MP/s
Q=80: 1.603 MP/s
Q=99: 1.122 MP/s
After:
Q=20: 3.334 MP/s
Q=40: 2.464 MP/s
Q=60: 2.009 MP/s
Q=80: 1.871 MP/s
Q=99: 1.163 MP/s
This CL allows for some further improvements that would not be possible
otherwise.
Change-Id: I61ba154beca2266cb96469281cf96e84a4412586
