also changed the token-page layout a little bit to remove
a not-needed field.
This reduces the number of malloc()/free() calls substantially
with minimal increase in memory consumption (~2%).
For the tail of large sources, the number of malloc calls goes
typically from ~10000 to ~100 (e.g.: bryce_big.jpg: 22711 -> 105)
Change-Id: Ib847f41e618ed8c303d26b76da982fbc48de45b9
MALLOC_FAIL_AT flag can be used to set-up a pre-determined failure
point during malloc calls. The counter value is retrieved using
getenv().
Example usage: export MALLOC_FAIL_AT=37 && cwebp input.png
will make 'cwebp' report a memory allocation error the 37th time
malloc() or calloc() is called.
MALLOC_MEM_LIMIT can be used similarly to prevent allocating more
than a given amount of memory. This is usually less convenient to
use than MALLOC_FAIL_AT since one has to know in advance the typical
memory size allocated.
Both these flags are meant to be used for debugging only!
Also: added a 'total_mem_allocated' to record the overall memory allocated
Change-Id: I9d408095ee7d76acba0f3a31b1276fc36478720a
Non-photo source produce far less literal reference and their
buffer is usually much smaller than the picture size if its compresses
well. Hence, use a block-base allocation (and recycling) to avoid
pre-allocating a buffer with maximal size.
This can reduce memory consumption up to 50% for non-photographic
content. Encode speed is also a little better (1-2%)
Change-Id: Icbc229e1e5a08976348e600c8906beaa26954a11
This change reduces the number of calls to WebPSafeMalloc from 200 to
100. The overall memory consumption is down 3% for Lenna image.
Change-Id: I1b351a1f61abf2634c035ef1ccb34050b7876bdd
Some tracing code is activated by PRINT_MEM_INFO flag.
For debugging only! (not thread-safe, and slow).
Change-Id: I282c623c960f97d474a35b600981b761ef89ace9
the unique instance of VP8LHashChain (1MB size corresponding to hash_to_first_index_)
is now wholy part of VP8LEncoder, instead of maintaining the pointer to VP8LHashChain
in the encoder.
Change-Id: Ib6fe52019fdd211fbbc78dc0ba731a4af0728677
We use automatic int->uint64_t promotion where applicable.
(uint64_t should be kept only for overflow checking and memory alloc).
Change-Id: I1f41b0f73e2e6380e7d65cc15c1f730696862125
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
Reduce calls to Malloc (WebPSafeMalloc/WebPSafeCalloc) for:
- Building HashChain data-structure used in creating the backward references.
- Creating Backward references for LZ77 or RLE coding.
- Creating Huffman tree for encoding the image.
For the above mentioned code-paths, allocate memory once and re-use it
subsequently.
Reduce the foorprint of VP8LHistogram struct by changing the Struct
field 'literal_' from an array of constant size to dynamically allocated
buffer based on the input parameter cache_bits.
Initialize BitWriter buffer corresponding to 16bpp (2*W*H).
There are some hard-files that are compressed at 12 bpp or more. The
realloc is costly and can be avoided for most of the WebP lossless
images by allocating some extra memory at the encoder initializaiton.
Change-Id: I1ea8cf60df727b8eb41547901f376c9a585e6095
HuffmanCost and HuffmanCostCombined optimized and added
'const' to some variables from ExtraCost functions.
Change-Id: I28b2b357a06766bee78bdab294b5fc8c05ac120d
When remapping buffer, br->eos_ was wrongly being set to true for
certain
images.
Also, refactored the end-of-stream detection as a function.
Reported in http://crbug.com/364830
Change-Id: I716ce082ef2b505fe24246b9c14912d8e97b5d84
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
.set at - Indicates that macro expansions may clobber
the assembler temporary ($at or $28) register.
Some macros may not be expanded without this
and will generate an error message if noat
is in effect.
"at" also added to the clobber list.
Change-Id: I67feebbd9f2944fc7f26c28496e49e1e2348529d
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
This change gains back 1% in compression density for method=3 and 0.5% for
method=4, at the expense of 10% slower compression speed.
Change-Id: I491aa1c726def934161d4a4377e009737fbeff82
+ 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
Tune HistogramCombineBin for hard images that are larger than 1-2 Mega
pixel and represent photographic images.
This speeds up lossless encoding on 1000 image corpus by 10-12% and compression
penalty of 0.1-0.2%.
Change-Id: Ifd03b75c503b9e886098e5fe6f86be0391ca8e81
there's still some malloc/free in the external example
This is an encoder API change because of the introduction
of WebPMemoryWriterClear() for symmetry reasons.
The MemoryWriter object should probably go in examples/ instead
of being in the main lib, though.
mux_types.h stil contain some inlined free()/malloc() that are
harder to remove (we need to put them in the libwebputils lib
and make sure link is ok). Left as a TODO for now.
Also: WebPDecodeRGB*() function are still returning a pointer
that needs to be free()'d. We should call WebPSafeFree() on
these, but it means exposing the whole mechanism. TODO(later).
Change-Id: Iad2c9060f7fa6040e3ba489c8b07f4caadfab77b
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
With -bypass_filter switched on, the lossless-compressed
data is decoded ahead of time (before being transformed and
display). Hence, the last row was called twice.
http://code.google.com/p/webp/issues/detail?id=193
Change-Id: I9e13f495f6bd6f75fa84c4a21911f14c402d4b10
(and ~2-3% on ARM)
We don't need to store cost/score for each node, but only for
the current and previous one -> simplify code and save some memory.
Also made the 'Node' structure tighter.
Change-Id: Ie3ad7d3b678992b396242f56e2ac387fe43852e6
all the functions involved return double and later these locals are used
in double calculations. fixes a vs build warning
Change-Id: Idb547104ef00b48c71c124a774ef6f2ec5f30f14
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
Optimize and re-structured VP8LGetHistoImageSymbols method, by using the bin-hash
for merging the Histograms more efficiently, instead of the randomized
heuristic of existing method HistogramCombine.
This change speeds up the Lossless encoding by 40-50% (for method=4 and Q > 50)
with 0.8% penalty in compression density. For lower method, the speed up is 25-30%,
with 0.4% penalty in the compression density.
Change-Id: If61adadb1a041b95def6405aa1fe3b83c3cb25ce
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
These are presets for lossless coding, similar to zlib.
The shortcut for lossless coding is now, e.g.:
cwebp -z 5 in.png -o out_lossless.webp
There are 10 possible values for -z parameter:
0 (fastest, lowest compression)
to 9 (slowest, best compression)
A reasonable tradeoff is -z 6, e.g.
-z 9 can be quite slow, so use with care.
This -z option is just a shortcut for some pre-defined
'-lossless -m xx -q yy' combinations.
Change-Id: I6ae716456456aea065469c916c2d5ca4d6c6cf04
(We didn't need the exact value of the max_error properly.
We can work with relative values instead of absolute)
Output is bitwise the same as before.
Change-Id: I67aeaaea5f81bfd9ca8e1158387a5083a2b6c649
Refactor code for HistogramCombine and optimize the code by calculating
the combined entropy and avoid un-necessary Histogram merges.
This speeds up lossless encoding by 1-2% and almost no impact on compression
density.
Change-Id: Iedfcf4c1f3e88077bc77fc7b8c780c4cd5d6362b
mostly by:
- storing a single rd-score instead of cost / distortion separately
- evaluating terminal cost only once
- getting some invariants out of the loops
- more consts behind fewer variables
Change-Id: I79451f3fd1143d6537200fb8b90d0ba252809f8c
incorporate non-last cost in per-level cost table
also: correct trellis-quant cost evaluation at nodes
(output a little bit different now). Method 6 is ~4% faster.
Change-Id: Ic48bd6d33f9193838216e7dc3a9f9c5508a1fbe8
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
Separate the C version from the MIPS32 version and have run-time
initialization during RescalerInit()
Change-Id: I93cfa5691c073a099fe62eda1333ad2bb749915b
Increase the initial buffer size for VP8L Bit Writer from 4bpp to 8bpp.
The resize buffer is expensive (requires realloc and copy) and this additional
memory (0.5 * W * H) doesn't add much overhead on the lossless encoder.
Change-Id: Ic1fe55cd7bc3d1afadc799e4c2c8786ec848ee66
Optimize 'VP8LCalculateEstimateForCacheSize' for lower quality ranges (Q < 50).
The entropy is generally lower for higher cache_bits, so start searching from
higher cache_bits and settle for a local minima, instead of evaluating all
values.
This speeds up the lossless encoding at lower qualities by 10-15%.
Change-Id: I33c1e958515a2549f2e6f64b1aab3f128660dcec
* simplify the endian logic
* remove the need for memset()
* write 16 or 32 at a time (likely aligned)
Makes the code a bit faster on ARM (~1%)
Change-Id: I650bc5654e8d0b0454318b7a78206b301c5f6c2c
-> 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
(instead of per-macroblock)
speed unchanged.
simplified the context-saving for incremental decoding
Change-Id: I301be581bab581ff68de14c4ffe5bc0ec63f34be
VP8GetThreadMethod() may be called with a NULL headers param; correct an
assert.
broken since:
8a2fa09 Add a second multi-thread method
Change-Id: If7b6d1b8f4ec874d343a806cee5f5e6bb6438620
This makes the segmentation overall less prone to
local-optimum or boundary effect.
(and overall, encoding is a little faster)
Change-Id: I35688098b0f43c28b5cb81c4a92e1575bb0eddb9
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
the -alpha_cleanup flag was ineffective since we switched cwebp
to using ARGB input always.
Original idea by David Eckel (dvdckl at gmail dot com)
Change-Id: I0917a8b91ce15a43199728ff4ee2a163be443bab
New API options: WebPDecoderOptions.flip and 'dwebp -flip ...'
it uses negative stride trick.
Also changed the decoder code to support user-supplied
buffers with negative stride, independently of the
WebPDecoderOptions.flip value.
Change-Id: I4dc0d06f0c87e51a3f3428be4fee2d6b5ad76053
this partially reverts
f626fe2 Detect canvas and image size mismatch in decoder.
the original change would cause calls to e.g., WebPGetInfo to fail until
a portion of the image chunk was available. With lossy+alpha this meant
waiting for the entire ALPH chunk to be received.
this change restores the original behavior -- reporting the values from
VP8X if available -- while retaining some of the added canvas/image size
checks if the image data is available
Change-Id: I6295b00a2e2d0d4d8847371756af347e4a80bc0e
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
the *quantized* level should be clipped to 2047, not the
original coeff.
(similar problem was fixed in the regular quantize function
quite some time ago)
Change-Id: I2fd2f8d94561ff0204e60535321ab41a565e8f85
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
Even at high quality setting, the U/V quantizer step is limited
to 4 which can lead to banding on gradient.
This option allows to selectively apply some randomness to
potentially flattened-out U/V blocks and attenuate the banding.
This option is off by default in 'dwebp', but set to -dither 50
by default in 'vwebp'.
Note: depending on the number of blocks selectively dithered,
we can have up to a 10% slow-down in decoding speed it seems.
Change-Id: Icc2446007f33ddacb60b3a80a9e63f2d5ad162de
RGBToU/V calls expects two extra precision bits, they were only
given one by SUM2H and SUM2H macros.
For rounding coherency, also changed SUM1 macro.
Change-Id: I05f96a46f5d4f17b830d0420eaf79b066cdf78d4
otherwise make sure that all frames are marked as a fragment. there's
still some work to do with validation if fragments are expected to cover
the entire canvas.
Change-Id: Id59e95ac01b9340ba8c6039b0c3b65484b91c42f
this avoids local-minima that look bad, even if the distortion
looks low (e.g. gradients, sky,...). Mostly visible in the q=50-80 range.
Output size is mostly unchanged.
Change-Id: I425b600ec45420db409911367cda375870bc2c63
Earlier we were only testing for bit_pos == LBITS. But this is not
sufficient,
as bit_pos can jump from < LBITS to > LBITS.
This was resulting in some bit-stream truncation errors not being
caught.
Note: Not a security bug though, as br->pos wasn't incremented in such
cases
and so we weren't reading beyond the buffer.
Change-Id: Idadcdcbc6a5713f8fac3470f907fa37a63074836
* raise U/V quantization bias to more neutral values
* also raise the non-zero AC bias for Y1/Y2 matrices
(we need all the precision we can for U/V leves, which are often empty)
This will increase quality in the higher range (q >= 90) mostly.
Files size is exacted to raise a little (5-7%). and SSIM accordingly of course.
Change-Id: I8a9ffdb6d8fb6dadb959e3fd392e66dc5aaed64e
kLevelsFromDelta[sharpness][delta] is an inverse look-up table
that tells the minimum filtering strength needed to trigger the
filtering of a step with amplitude 'delta'. We use this table
in various situations:
a) when computing the initial (/global) filtering
strength for each segment. We look at the quantization
step and deduce the proper filtering strength needed
to result this quantization noise (talking the -f option
into account).
b) during intra16 calculation, when a block ends up
very empty (only DC coeffs are non-zero, all ACs have
vanished). We'll rely on the in-loop filtering to
restore the smoothness (if the source was gradient-like
smooth. That's why we look at the distortion too before
triggering the filtering).
Step b) goes _in addition_ to a), potentially raising
the filtering strength if blockiness is likely.
Change-Id: Icaeca93ef21da195b079e6587a44d9edfc8e9efa
Earlier "f = f->next_" was executing for both inner and outer loop, thus
skipping validation of some frames.
Change-Id: Ice5cdb4ff5da78384aa0573addd3a5e5efa0b10c
-> 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
method 1 grouping: [parse + reconstruction] // [filtering + output]
method 2 grouping: [parse] // [reconstruction+filtering + output]
Depending on some heuristics (see VP8ThreadMethod()), we
can pick one of the other when -mt flag (or option.use_threads)
is selected.
Conservatively, we always use method #2 for now until the heuristic
is refined (so, timing should be the same the before this patch)
+ replace 'use_threads' by 'mt_method'
+ define MIN_WIDTH_FOR_THREADS constant
+ fix comment alignment
Change-Id: I11a756dea9070d6e21b1a9481d357a1e8aa0663e
Mostly visible for large images.
Reconstruction+filtering is now done in parallel to bitstream-parsing.
Change-Id: I4cc4483d803b255f4d97a2fcd9158b1c291dd900
Needs more memory but allows for future parallelization.
Noticeably faster on ARM, slightly faster on x86
also: remove dec->filter_row_ unnecessary field
Change-Id: I044a808839b4e000c838a477e3e8688820436d9a
