* We were re-doing most of the work in plain-C as 'left-over'.
* we were always returning has_alpha = true because of a bad mask all_0xff
These bugs were conservative and silent, in the sense that we were 'just' doing
more work than necessary.
Now, the SSE2 version is really 2x faster than the C version.
Change-Id: I6c8132a267fe3c7a3d1fa70e7a5fcd10719543fa
Split initialization of YUV444Converters[] out of Upsamplers init.
update test for NULL function pointers
Change-Id: I9603f54250f90c85a12ffbecfd6c59e9b06c47e0
vtbl4_u8 is available everywhere except iOS arm64: use vtbl2q_u8 there
with a corresponding change in the load.
Change-Id: Ib84212dda3c7875348282726c29e3b79b78b0eac
_M_IX86 will be defined in mingw builds after including windows.h. as
the gcc inline asm is first, this missing check would only have caused
an error if the code was reorganized.
Change-Id: I395679bcfc43e94d308d1ceb0c0fbf932b2c378c
this defines WORDS_BIGENDIAN, replacing uses of
__BIG_ENDIAN__/__BYTE_ORDER__ with it
+ fixes lossless BGRA output with big-endian toolchains
that do not define __BIG_ENDIAN__ (codesourcery mips gcc)
Change-Id: Ieaccd623292d235343b5e34b7a720fc251c432d7
this will remove a warning about the shift amount not being
an immediate (=constant).
Change-Id: Ie9a00fefdb9a07ec8994fb113f24234518bc878a
Also: fix the NULL sharpen argument mismatch.
Another store to load forward block was detected coming from the function
FTransform.
FTransform save the output data 4 times 8 bytes each. when this data is
later being loaded by the QuantizeBlock function in one chunk of 16 bytes
that caused a store to load forward block.
The fix was done in the FTransform function where each two consecutive 8 bytes
were merged into one 16 bytes register and saved into the memory.
This fix gives ~21% function level gain and 1.6% user level gain.
Change-Id: Idc27c307d5083f3ebe206d3ca19059e5bd465992
only 1 of <lib>_CPPFLAGS and AM_CPPFLAGS is used, with the former
getting precedence when it's defined. configure's DEFAULT_INCLUDES is
covering what's necessary given the include paths are all source
relative.
Change-Id: I7d14076acd266b28a88a3d92bcc3d7165284d5f3
this change has the side-effect of using directory names in the
include, silencing a lint warning.
Change-Id: Ib91cf63a90534e32fadfa5c2372bfdb29f854d02
if res->first = 1, coeffs[0]=0 because of quant.c:749 and line
added at quant.c:744
So, no need for the extra case.
Going forward, TrellisQuantizeBlock() should also be calling
a variant of VP8SetResidualCoeffs() to set the 'last' field.
also: fixes a warning for win64
+ slight speed-up
Change-Id: Ib24b611f7396d24aeb5b56dc74d5c39160f048f0
+ add a WEBP_HAVE_SSE2 to dsp.h
not all 32-bit toolchain configurations will have sse2 enabled by
default
Change-Id: I7c675e511581f93cf55c79f960fa7efa2df4987e
this is used to set WEBP_USE_AVX2 in files where the build flag won't be
used, i.e., dsp/enc.c, which enables VP8EncDspInitAVX2() to be called
Change-Id: I362f4ba39ca40d3e07a081292d5f743c649d9d7f
Extract loop invariant and avoid storing/loading samples
if they can be re-used. This is particularly interesting when
a transpose is involved (HFilter16i).
Change-Id: I93274620f6da220a35025ff8708ff0c9ee8c4139
The luminance needs to be pre- and post- multiplied by
the alpha value in case of rescaling, for proper averaging.
Also:
- removed util/alpha_processing and moved it to dsp/
- removed WebPInitPremultiply() which was mostly useless
and merged it with the new function WebPInitAlphaProcessing()
Change-Id: If089cefd4ec53f6880a791c476fb1c7f7c5a8e60
VP8EncDspInitAVX2 is included in sse2 builds for now, later a configure
flag should be added to avoid the stub when avx2 is unavailable/disabled
Change-Id: I6127b687c273f46f41652aaf8e3b86ae3cfb8108
* remove some sign-bit flipping
* turn some macro into inline functions
* fix some 'const' in signatures
* clarify the int8/uint8 usage
Change-Id: Ib04459ac34cb280c57579c5d79a5efd2f8d5e99d
The predictors based on Average2 are tad slower.
Following is the performance data for these predictors normalized to
number of instruction cycles (as per valgrind) per operation:
- Predictor6 & Predictor7 now takes 15 instruction cycles compared to 11
instruction cycles for the C version.
- Predictor8 & Predictor9 now takes 15 instruction cycles compared to 12
instruction cycles for the C version.
The predictors based on Average4 is faster and Average3 is tad slower:
- Predictor10 (Average4) now takes 23 instruction cycles compared to 25
instruction cycles for the C version.
- Predictor5 (Average3) now takes 20 instruction cycles compared to 18
instruction cycles for the C version.
Maybe SSE2 version of Average2 can be improved further. Otherwise, we can
remove the SSE2 version and always fallback to the C version.
Change-Id: I388b2871919985bc28faaad37c1d4beeb20ba029
* merged the two HistogramAdd/AddEval() into a single call
(with detection of special case when b==out)
* added a SSE2 variant
* harmonize the histogram type to 'uint32_t' instead
of just 'int'. This has a lot of ripples on signatures.
* 1-2% faster
Change-Id: I10299ff300f36cdbca5a560df1ae4d4df149d306
move simple loop filter defines closer to their use and LOAD* to a
location common with the intrinsics
Change-Id: Iaec506d27bbc9a01be20936e30b68a4b0e690ee3
the complex loop filter has no inline equivalent; the simple loop filter
remains conditional on USE_INTRINSICS: it's left undefined for now.
Change-Id: I4f258e10458df53a7a1819707c8f46b450e9d9d2
CollectHistogram / SSE* / QuantizeBlock have no inline equivalents,
enable them where possible and use USE_INTRINSICS to control borderline
cases: it's left undefined for now.
Change-Id: I62235bc4ddb8aa0769d1ce18a90e0d7da1e18155
using this in Load4x16 was slightly slower and didn't help mitigate any
of the remaining build issues with 4.6.x.
Change-Id: Idabfe1b528842a514d14a85f4cefeb90abe08e51
HuffmanCost and HuffmanCostCombined optimized and added
'const' to some variables from ExtraCost functions.
Change-Id: I28b2b357a06766bee78bdab294b5fc8c05ac120d
Some versions of compiler in debug build can't find
a register in class 'GR_REGS' while reloading 'asm'
Number of used registers is decreased in this fix.
Change-Id: I7d7b8172b8f37f1de4db3d8534a346d7a72c5065
This is to help further optimizations.
(like in https://gerrit.chromium.org/gerrit/#/c/69787/)
There's a small slowdown (~0.5% at -z 9 quality) due to
function pointer usage. Note that, for speed, it's important
to return VP8LStreaks by value, and not pass a pointer.
Change-Id: Id4167366765fb7fc5dff89c1fd75dee456737000
avoids:
src/dsp/enc_mips32.c: In function 'ITransformOne':
src/dsp/enc_mips32.c:123:3: can't find a register in class 'GR_REGS' while reloading 'asm'
src/dsp/enc_mips32.c:123:3: 'asm' operand has impossible constraints
Change-Id: Ic469667ee572f25e502c9873c913643cf7bbe89d
apparently faster, but we might save some load/store to/from memory
once we settle for the intrinsics-based FTransform()
(also: fixed some #ifdef USE_INTRINSICS problems)
Change-Id: I426dea299cea0c64eb21c4d81a04a960e0c263c7
Functions VP8LFastLog2Slow and VP8LFastSLog2Slow
also: replaced some "% y" by "& (y-1)" in the C-version
(since y is a power-of-two)
Change-Id: I875170384e3c333812ca42d6ce7278aecabd60f0
Verified OK, but right now they don't seem faster.
So they are disabled behind a USE_INTRINSICS flag (off for now)
Change-Id: I72a1c4fa3798f98c1e034f7ca781914c36d3392c
+ reorganize the cost-evaluation code by moving some functions
to cost.h/cost.c and exposing VP8Residual
Change-Id: Id976299b5d4484e65da8bed31b3d2eb9cb4c1f7d
slightly faster than the inline asm
in practice not much faster than the C-code in a full NEON build, but
still better overall in an Android-like one that only enables NEON for
certain files.
Change-Id: I69534016186064fd92476d5eabc0f53462d53146
* inverse transform is actually slower with intrinsics + gcc-4.6,
so is left disabled for now.
With gcc-4.8, it's a bit faster than inlined assembly.
* Sum of Square error function provide a 2-3% speed up
There's enabled by default (since there's no inlined-asm equivalent)
Change-Id: I361b3f0497bc935da4cf5b35e330e379e71f498a
+ misc cosmetics
* seems 4% slower than inlined-asm with gcc-4.6
* is a tad faster (<1%) with gcc-4.8
(disabled for now)
Change-Id: Iea6cd00053a2e9c1b1ccfdad1378be26584f1095
The nice trick is to pack 8 u + 8 v samples into a single uint8x16x_t
register, and re-use the previous (luma) functions
Change-Id: Idf50ed2d6b7137ea080d603062bc9e0c66d79f38
+ added some work-around gcc-4.6 to make it compile (except one function).
+ lots of revamping
All variants tested ok.
Speed-up is ~5-7%
Change-Id: I5ceda2ee5debfada090907fe3696889eb66269c3
vertical only currently, 2.5-3% faster
placed under USE_INTRINSICS as this change depends on the simple
loopfilter
improves the simple loopfilter slightly thanks to some reorganization
Change-Id: I6611441fa54228549b21ea74c013cb78d53c7155
When 4 pixels are left, they should be processed with SSE2.
Decoding is marginally faster (~0.4%).
Encoding speed: No observable difference.
Change-Id: I3cf21c07145a560ff795451e65e64faf148d5c3e
new file: lossless_neon.c
speedup is ~5%
gcc 4.6.3 seems to be doing some sub-optimal things here,
storing register on stack using 'vstmia' and such.
Looks similar to gcc.gnu.org/bugzilla/show_bug.cgi?id=51509
I've tried adding -fno-split-wide-types and it does help
the generated assembly. But the overall speed gets worse with
this flag. We should only compile lossless_neon.c with it -> urk.
Change-Id: I2ccc0929f5ef9dfb0105960e65c0b79b5f18d3b0
It's disable for now, because it crashes gcc-4.6.3 during compilation
with -O2 or -O3. It's been tested OK with -O1.
Code is still globally disabled with USE_INTRINSICS, though.
Change-Id: I3ca6cf83f3b9545ad8909556f700758b3cefa61c
disabled for now (but tested OK), thanks to the USE_INTRINSICS #define
We'll activate the code when we're on par with non-intrinsics
Change-Id: Idbfb9cb01f4c7c9f5131b270f8c11b70d0d485ff
expose the predictor array as function pointers instead
of each individual sub-function
+ merged Average2() into ClampedAddSubtractHalf directly
+ unified the signature as "VP8LProcessBlueAndRedFunc"
no speed diff observed
Change-Id: Ic3c45dff11884a8330a9ad38c2c8e82491c6e044
Get back some of the compression gains by extending the search space for
GetBestGreenRedToBlue. Also removed the SkipRepeatedPixels call, as it was not
helping much in yielding better compression density.
Before:
1000 files, 63530337 pixels, 1 loops => 45.0s (45.0 ms/file/iterations)
Compression (output/input): 2.463/3.268 bpp, Encode rate (raw data): 1.347 MP/s
After:
1000 files, 63530337 pixels, 1 loops => 45.9s (45.9 ms/file/iterations)
Compression (output/input): 2.461/3.268 bpp, Encode rate (raw data): 1.321 MP/s
Change-Id: I044ba9d3f5bec088305e94a7c40c053ca237fd9d
Restructure PredictorInverseTransform & ColorSpaceInverseTransform to remove
one if condition inside the main/critial loop. Also separated TransformColor &
TransformColorInverse into separate functions and avoid one 'if condition'
inside this critical method.
This change speeds up lossless decoding for Lenna image about 5% and 1000 image
corpus by 3-4%.
Change-Id: I4bd390ffa4d3bcf70ca37ef2ff2e81bedbba197d
Speedup lossless encoder by 20-25% by optimizing:
- GetBestColorTransformForTile: Use techniques like binary search and
local minima search to reduce the search space.
- VP8LFastSLog2Slow & VP8LFastLog2Slow: Adding the correction factor for
log(1 + x) and increase the threshold for calling the approximate
version of log_2 (compared to costly call to log()).
Change-Id: Ia2444c914521ac298492aafa458e617028fc2f9d
converts 2 s16 vectors to 2 u8 and store to uint8_t destination;
TransformAC3 can reuse this after a rework
Change-Id: Ia9370283ee3d9bfbc8c008fa883412100ff483d0
-> remove the 'color_transform' multiplier, use more constants, etc.
This function is particularly critical, mostly because of
GetBestColorTransformForTile().
Loop is a bit faster (maybe ~1%)
Change-Id: I90c96a3437cafb184773acef55c77e40c224388f
The WEBP_SWAP_16BIT_CSP flag needs to be honored while filling the Alpha (4 bits)
data in the destination buffer and while pre-multiplying the alpha to RGB colors.
Change-Id: I3b07307d60963db8d09c3b078888a839cefb35ba
The registers and instructions are quite different to 32bit
and the assembly code needs a rewrite.
more info: http://people.linaro.org/~rikuvoipio/aarch64-talk/
Change-Id: Id75dbc1b7bf47f43a426ba2831f25bb8fa252c4f
add TransformDC special case, and make the switch function inlined.
Recovers a few of the CPU lost during the addition of TransformAC3
(only on ARM)
Change-Id: I21c1f0c6a9cb9d1dfc1e307b4f473a2791273bd6
WHT is somewhat a special case: no sharpen[] bias, etc.
Will be useful in a later CL when precision of input is changed.
Change-Id: I851b06deb94abdfc1ef00acafb8aa731801b4299
This is in preparation for a future change where input will
be 16bit instead of 12bit
No speed diff observed.
Note that the NEON implementation was using 32bit calc already.
Change-Id: If06935db5c56a77fc9cefcb2dec617483f5f62b4
* remove the sharpening for non luma-AC coeffs
* adjust the bias a little bit to compensate for this
Using the multiply-by-reciprocal doesn't always give the same result
as the exact divide, given the QFIX fixed-point precision we use.
-> removed few now-unneeded SSE2 instructions (and checked for
bit-exactness using -noasm)
Change-Id: Ib68057cbdd69c4e589af56a01a8e7085db762c24
-> helps debanding (sky, gradients, etc.)
This dithering can only be triggered when using -preset photo
or -pre 2 (as a preprocessing). Everything is unchanged otherwise.
Note that this change is likely to make the perceived PSNR/SSIM drop
since we're altering the input internally.
Change-Id: Id8d4326245d9b828141de162c94ba381b1fa5813
use of uint8_t type was causing error like:
src/dsp/upsampling.c:223:1: internal compiler error: in vect_determine_vectorization_factor, at tree-vect-loop.c:349
with gcc 4.6.3
Change-Id: Ieb6189a1375c47fc4ff992e6c09b34a7f1f605da
The C-version gets ~7-8% slower in order to match the SSE2
output exactly. The old (now off-by-1) code is kept under
the WEBP_YUV_USE_TABLE flag for reference.
(note that calc rounding precision is slightly better ~= +0.02dB)
on ARM-neon, we somehow recover the ~4% speed that was lost by mimicking
the initial C-version (see https://gerrit.chromium.org/gerrit/#/c/41610)
Change-Id: Ia4363c5ed9b4c9edff5d932b002e57bb7814bf6f
If 'top' was meant to be NULL, then bottom and top can be
swapped. Logic is simpler.
+ fix compilation in non-FANCY_UPSAMPLING mode
Change-Id: I7c62bbb59454017f072c0945d1ff2d24d89286ff
Also created variant VP8LPrefixEncodeBits that returns the
code & extra_bits only.
There's no impact on compression density and compression speed.
Change-Id: I2cafdd3438ac9270cd72ad9d57b383cdddfdfa4c
This speeds up WebP lossless decoding by 20%. In particular, the
photographic images get 35% speedup.
Change-Id: Idb94750342a140ec05df52c07e12be4bba335adc
rather than symlink the webm/vpx terms, use the same header as libvpx to
reference in-tree files
based on the discussion in:
https://codereview.chromium.org/12771026/
Change-Id: Ia3067ecddefaa7ee01550136e00f7b3f086d4af4
* "declaration of ‘index’ shadows a global declaration [-Wshadow]"
* "signed and unsigned type in conditional expression [-Wsign-compare]"
Change-Id: I891182d919b18b6c84048486e0385027bd93b57d
Earlier such images were using roughly 9 * width * height bytes for
decoding. Now, they take 6 * width * height memory.
Change-Id: Ie4a681ca5074d96d64f30b2597fafdca648dd8f7
no precision loss observed
speed is not really faster (0.5% at max), as forward-WHT isn't called often.
also: replaced a "int << 3" (undefined by C-spec) by a "int * 8"
( supersedes https://gerrit.chromium.org/gerrit/#/c/48739/ )
Change-Id: I2d980ec2f20f4ff6be5636105ff4f1c70ffde401
Saturation was done on input coeff, not quantized one.
This saturation is not absolutely needed: output of FTransformWHT
is in range [-16320, 16321]. At quality 100, max quantization steps is 8,
so the maximal range used by QuantizeBlock() is [-2040, 2040].
But there's some extra bias (mtx->bias_[] and mtx->sharpen_[]) so
it's better to leave this saturation check for now.
addresses issue #145
Change-Id: I4b14f71cdc80c46f9eaadb2a4e8e03d396879d28
subdirectories with more than one target can have the install targets
run in parallel with make -jN. group the shared headers in one place to
produce a common install target.
Change-Id: I1f3aa338a8ee6d681de1e5d0b2c6244d2c3d5451
Reported to eventually be 4% on ARM
(see https://code.google.com/p/webp/issues/detail?id=134 for details)
We might activate it selectively later...
Output values is not bitwise the same as the LUT-based
version, but difference is only +/-1 at max.
Change-Id: I1cc790ff4459885ed2ae2e72f31c5f3740095f07
larger values are still dealt with in the .cc
~5% faster encoding
Output size is slightly different (variably), because of
different floating-point calculation ordering.
Change-Id: I6ede18b09c753997cf78aa1199a807d9ddb5d4b4
* add SSE2 variant for lossless
* speed-up TransformColor calls using specialized TransformColorBlue/Red
* Fuse the Shannon Entropy calls to compute it for X and X+Y simultaneously.
This latter changes the output size a little bit.
Change-Id: Ie5df94da78bf51a58da859c9099b56340da9ec89
This flag will make the code use no uint64, no asm, and no fancy
trick, but instead aim at being as simple and straightforward as
possible.
Main use is to help emscripten generate proper JS code.
More code needs to be simplified later.
Also: tune the BITS values to be 24 and make use of WEBP_RIGHT_JUSTIFY
Here are the typical timing for decoding a large image:
ARM7-a:
dwebp_justify_32_neon Time to decode picture: 3.280s
dwebp_justify_24_neon Time to decode picture: 2.640s
dwebp_justify_16_neon Time to decode picture: 2.723s
dwebp_justify_8_neon Time to decode picture: 2.802s
dwebp_justify_32 Time to decode picture: 4.264s
dwebp_justify_24 Time to decode picture: 3.696s
dwebp_justify_16 Time to decode picture: 3.779s
dwebp_justify_8 Time to decode picture: 3.834s
dwebp_32_neon Time to decode picture: 4.010s
dwebp_24_neon Time to decode picture: 2.725s
dwebp_16_neon Time to decode picture: 2.852s
dwebp_8_neon Time to decode picture: 2.778s
dwebp_32 Time to decode picture: 4.587s
dwebp_24 Time to decode picture: 3.800s
dwebp_16 Time to decode picture: 3.902s
dwebp_8 Time to decode picture: 3.815s
REFERENCE (HEAD) Time to decode picture: 3.818s
x86_64:
dwebp_justify_32 Time to decode picture: 0.473s
dwebp_justify_24 Time to decode picture: 0.434s
dwebp_justify_16 Time to decode picture: 0.450s
dwebp_justify_8 Time to decode picture: 0.467s
dwebp_32 Time to decode picture: 0.474s
dwebp_24 Time to decode picture: 0.468s
dwebp_16 Time to decode picture: 0.468s
dwebp_8 Time to decode picture: 0.481s
REFERENCE (HEAD) Time to decode picture: 0.436s
i386:
dwebp_justify_32 Time to decode picture: 0.723s
dwebp_justify_24 Time to decode picture: 0.618s
dwebp_justify_16 Time to decode picture: 0.626s
dwebp_justify_8 Time to decode picture: 0.651s
dwebp_32 Time to decode picture: 0.744s
dwebp_24 Time to decode picture: 0.627s
dwebp_16 Time to decode picture: 0.642s
dwebp_8 Time to decode picture: 0.642s
Change-Id: Ie56c7235733a24f94fbfc2e4351aae36ec39c225
store values to a temporary variable before calling functions that take
vector types.
removes non-standard constructs such as:
(uint8x8x2_t){{ a, b }}
fixing:
src/dsp/upsampling_neon.c:69:32: error: macro "vst2_u8" passed 3
arguments, but takes just 2
Change-Id: Ib4368e16e3a3efac18024f02be94e76243ade2dc
Fixes: https://code.google.com/p/webp/issues/detail?id=140
- along the lines of the SSE chroma upsampling.
Total speedup is ~30%.
4% speed loss on YuvToRgbXX conversion using tables instead
of 14-bit fixed precision. TODO(later): investigate, and compare
to x86.
see http://code.google.com/p/webp/issues/detail?id=134
Change-Id: Idc2261037cd13b4553ca20ecc4c4007099c37009
When the config option '--enable-libwebpdecoder' is specified, the
lean decoder library 'libwebpdecoder' will be created in addition to
libwebp. Also dwebp binary will be linked to libwebpdecoder, if this
config option is specified.
Change-Id: I9de3e149b59c9a8390fae2ba660941749640e54a
We don't need to use the exact forward transform,
since it's only a rough evaluation.
-> Removed some shifts and rounding constants.
Change-Id: I3fdf8b4fe9720473894155e1ad0345f4d1fd9a33
Contributed by Wayne Chen (datoudatou at gmail dot com)
+ some header cleanup
+ remove the NEON suffix in static functions
Change-Id: I75bf5e9b54cf5e1acc53764c6f081d61690f8e3d
(implements the backward and forward transforms in the encoder)
original patch by Wayne Chen (datoudatou at gmail dot com)
Change-Id: Ic00f3bffcdf7a924f043006728735c810ee47a57
This is mostly for experimentation!
Need to define USE_YUVj flag in the code for that.
suggested by benwreder at hotmail dot com
Change-Id: If0b8e2c1863efc08ce097de6de20f4c7efc3f7e8
fixes the 'blocky sky problem' (saturation problem: when luma was flat,
chroma noise was taking over, resulting in random segment id assigned.
When just using a common uniform segment was better).
+ side clean-up and readibility/experimentability MACRO'ization
+ added '-map 7' option
Change-Id: I35982a9e43c0fecbfdd7b05e4813e8ba8c121d71
this will avoid the "dec_neon.o has no symbol" warning
no change in binary size observed on linux.
Change-Id: Ia27ae2bc5a03d714afa7e46671fdcf4cb630784d
