It was a bad implementation of a Lehmer random number generator
(the saturation was done wrong and mostly & was used instead of % .....).
That lead to "for" loop stuck with the same values given a specific seed,
hence wasted "for" loops (e.g. seed getting at 374988608 and modulo of 64
later leads to 0 even when updating the seed with the old formula).
As the "for" loops now always return a proper pair of histograms, their
number can greatly be reduced, hence a speedup.
Change-Id: I9f5b44d66cc96fd4824189d92276c3756c8ead5b
This code is ultra-critical for lossless decoding, especially on ARM.
The extra call VP8LIsEndOfStream() was causing unnecessary slow-down.
Now, we check for bitstream-end separately in the main loop.
Change-Id: I739b5d74cc29578e2b712ba99b544fd995ef0e0d
Currently, none are available. If WEBP_HAVE_SSE2 eventually works,
we'll have to refine this conditionals.
BUG=webp:261
Change-Id: Ibc63ee1c013f2a4169eeb85cc8b6317b6420c2ad
Previously, the stochastic method for histogram
combination could finish in a greedy way
if the number of iterations to perform so was smaller.
Except that another greedy combination was performed
afterwards ... hence wasted CPU in some cases.
Change-Id: Ic0f26873e6dc746679486b91cb35d73efee91931
The initial re-writing of this part of the code with intervals
had to be done with a complex logic (mostly intervals with a
lower and upper bound, not a constant value like now) to properly
deal with the inefficiencies of the then LZ77 algorithm.
The improvements made to LZ77 since, now allow for a simpler logic.
There were also small errors in the interval insertion logic
that lead to small inefficiencies (hence a slightly better
compression rate).
Change-Id: If079a0cafaae7be8e3f253485d9015a7177cf973
Documentation says: "if kmin == 0, then key-frame insertion is disabled;
and if kmax == 0, then all frames will be key-frames."
Reading this, you'd expect that if kmax == 0, then with any kmin <= 0
all frames will be key-frames. But actually the kmin <= 0 test is caught
first and you get the opposite (no keyframes but the first). You'd have
instead to set kmax == 0 and any value kmin > 0, which is absolutely
counter-intuitive (reversing order).
Moreover kmax == 1 has no valid kmin (kmin == 1 conflicts with the
`kmax > kmin` rule and kmin == 0 conflicts with `kmin >= kmax / 2 + 1`).
So it should be considered an exception too.
Instead I propose this new logic:
- kmax == 1 means that all frames are keyframes (you are explicitly
requesting a keyframe every 1 frame at most, i.e. all frames).
- kmax == 0 means no keyframes (you ask for a keyframe every 0 frames,
i.e. never).
This is more "logical" language-wise, and also does not involve any
conflicts about what if both kmax and kmin are 0, since now a single
property value is meaningful for the 2 exceptional cases.
Change-Id: Ia90fb963bc26904ff078d2e4ef9f74b22b13a0fd
(cherry picked from commit 2dc0bdcaee)
Compile with XCode, it appears quite slower than the C-version,
especially for arm64.
Change-Id: Ic46dba184a36be454fef674129d2f909003788fc
(cherry picked from commit 4f3e3bbd44)
Documentation says: "if kmin == 0, then key-frame insertion is disabled;
and if kmax == 0, then all frames will be key-frames."
Reading this, you'd expect that if kmax == 0, then with any kmin <= 0
all frames will be key-frames. But actually the kmin <= 0 test is caught
first and you get the opposite (no keyframes but the first). You'd have
instead to set kmax == 0 and any value kmin > 0, which is absolutely
counter-intuitive (reversing order).
Moreover kmax == 1 has no valid kmin (kmin == 1 conflicts with the
`kmax > kmin` rule and kmin == 0 conflicts with `kmin >= kmax / 2 + 1`).
So it should be considered an exception too.
Instead I propose this new logic:
- kmax == 1 means that all frames are keyframes (you are explicitly
requesting a keyframe every 1 frame at most, i.e. all frames).
- kmax == 0 means no keyframes (you ask for a keyframe every 0 frames,
i.e. never).
This is more "logical" language-wise, and also does not involve any
conflicts about what if both kmax and kmin are 0, since now a single
property value is meaningful for the 2 exceptional cases.
Change-Id: Ia90fb963bc26904ff078d2e4ef9f74b22b13a0fd
this avoids duplicates between these trees and dsp/, e.g., enc/tree.c,
dec/tree.c, making pulling the whole library source tree into one target
possible
BUG=webp:279
Change-Id: I060a614833c7c24ddd37bf641702ae6a5eef1775
We can switch at run-time between the standard GetCoeffs() critical
function, that uses a fast variant of VP8GetBit().
However, some platforms have slow instructions that make standard
VP8GetBit() slow. GetCoeffs() is the right level of branching to
switch to GetCoeffsAlt() that avoids these slow instructions in some
not-frequent cases.
Next patch will upgrade VP8GetBit() to use clz, after this one
is proved to be neutral speed-wise.
Change-Id: Ia6cef5de9de6131574d2202bbc0bea8559c9b693
vmlal_u8() is prone to overflow during the accumulation.
There was a mismatch happening at low q mostly. Because in this
case the distortion is important and the accumulated sum was
later than 16bit-unsigned.
Change-Id: I1a08a2f744bcdf0b26647e61b9ee92a0c2e28fe8
This makes the structure more generic, without the hard-coded
internal structure.
This is a borderline incompatible ABI change, even if WebPIDecoder structure
is opaque.
Change-Id: I518765c3f76fc17a136cef045a5a8aa70ed70e85
30% faster on x86, 5% faster on N5.
New generic function: WebPLog2FloorC()
This function is called as fallback for BitsLog2Floor() when there's
no clz() available.
Change-Id: Ica15c6092112e514c0e200fab89c434de48d4b19
This is meant to be used for run-time detection of slow platforms
regarding instructions like pshufb and bsr.
Adapted from libvpx patch: https://chromium-review.googlesource.com/#/c/367731
Change-Id: I2c22fbb9aae699d87a041393ba1ad5f1f21ff640
and 15% faster MultARGBRow()
by switching to formulae:
X / 255 = (X + 1 + (X >> 8)) >> 8 for any 16bit value X.
(X / 255 + .5) = (XX + (XX >> 8)) >> 8, with XX = X + 128
Change-Id: Ia4a7408aee74d7f61b58f5dff304d05546c04e81
The previous optimization was performing dichotomy on a function that
is anything in practice, hence a bit of randomness.
Also, two magic constants were used, one for an extra constant cost,
one for an extra linear cost. Both values/models were empirical.
A brute force search for the best cache size is now performed.
To have less CPU impact, a speed optimization is also made by not
inserting a value again and again.
This makes sense but it's also the most common case of when LZ77 is
useful hence an overall improvement sometimes.
Change-Id: I57de5750ad2313b2feecbcd15cd6e4feeb98e5c8
