The existing code applied a 6-tap filter with 0's on either end.
We're already paying the branch penalty to avoid computing the two
extra columns needed as input to this filter.
We might as well save time computing the filter as well.
This reduces the inner loop from 21 instructions to 16, the number
of loads per iteration from 4 to 1, and the number of multiplies
from 7 to 4.
The gain in overall decoding performance, however, is small (less
than 1%).
This change also means we now valgrind clean on ARMv6, which is
its real purpose.
The errors reported here were valgrind's fault (it does not detect
that 0 times an uninitialized value is initialized), but Julian
Seward says it would slow down valgrind considerably to make such
checks.
Speeding up libvpx rather, even by a small amount, seems a much
better idea if only to enable proper valgrind checking of the
rest of the codec.
Change-Id: Ifb376ea195e086b60f61daf1097d8910c4d8ff16
This function was accessing values below the stack pointer, which
can be corrupted by signal delivery at any time.
Change-Id: I92945b30817562eb0340f289e74c108da72aeaca
previous implementation compared each set of values to limit and then
&'d them together, requiring a compare and & for each value.
this does the accumulation first, requiring only one compare
Change-Id: Ia5e3a1a50e47699c88470b8c41964f92a0dc1323
This patch avoids compiling some debugging code in onyx_if.c. The most
significant fix is to avoid generating code for vp8_write_yuv_frame,
which is never called. Some other code was removed by the dead code
elimination performed by the compiler, and this patch does it with the
preprocessor instead. There are advantages both ways.
Change-Id: I044fd43179d2e947553f0d6f2cad5b40907ac458
reconintra_mt.c is only required for building the decoder right now.
It could definitely be used for the encoder in the future, but it
currently depends on decoder only data structures. (onyxd_int.h,
VP8D_COMP, etc). Move it from common/ to decoder/ until the
necessary changes to the common multithread code are complete.
This patch is needed to build with --disable-vp8-decoder.
Change-Id: I568c52221a2b309234d269675cba97131ce35c86
Having these symbols be available as functions rather than data is
occasionally more convenient. Implemented this way rather than a
get-codec-by-id style to avoid creating a link-time dependency
between the encoder and the decoder.
Fixes issue #169
Change-Id: I319f281277033a5e7e3ee3b092b9a87cce2f463d
The MV decoding changes in c5fb0eb introduced a bug where the
macroblock clamping state was reset for each partition, so if an
earlier partition needed clamping but a subsequent one didn't,
the MB wouldn't receive clamping. Instead, the state is only
set during splitmv decoding, never cleared.
Change-Id: I224fe258493405ee0f6a04596acdb622c475e845
Movdqu is more expensive (throughput, uops) than movq. Minimal
impact for newer big cores, but ~2.25% gain on Atom.
Change-Id: I62c80bb1cc01d8a91c350c4c7719462809a4ef7f
Use pmaxub instead of a combination of psubusb/por to
determine if any comparisons go over the limit.
Change-Id: I3f0bd7d2aabe5fee9ba6620508e2b60605abcb82
The patch related with issue #55 (5a72620) fixed some warnings, but the
fix was not optimal. It actually was a trick to confuse compiler rather
than a fix.
This patch fixes it by creating a new macro used when needed just a high
limit check for an unsigned.
Change-Id: I94b322e0f7fb07604b3b1df1f9321185f48cfcb5
the previous commit laid the groundwork by doing two sets of idcts
together. this moved that further by grouping the interesting data
(q[0], q+16[0]) together to allow using wider instructions. also
managed to drop a few instructions by recognizing that the constant
for sinpi8sqrt2 could be downshifted all the time which avoided a
dowshift as well as workarounds for a function which only accepted
signed data
looks like a modest gain for performance: at qcif, went from ~180
fps to ~183
Change-Id: I842673f3080b8239e026cc9b50346dbccbab4adf
On each MB, loopfiltering is done right after MB decoding. This
combines two loops in multi-threaded code into one, which reduces
number of synchronizations to half.
The above-row/left-col data are saved in temp buffers for
next-row/next MB decoding.
Tests on 4-core gLucid machine showed 10% decoder performance
gain with threads=4 (tulip clip). Testing on other platforms
isn't done yet.
Change-Id: Id18ea7c1e84965dabea65d4c01ca5bc056ddeac9
This patch reduces the size of the global tables maintained by the
tokenizer to 16k from 80k-96k. See issue #177.
Change-Id: If0275d5f28389af11ac83c5d929d1157cde90fbe
GET_GOT was producing a zero length call. This resulted in
pipeline flushes occuring when returing from the assembly
functions. Masked on out of order cores, but evident on
Atom cores.
Change-Id: I8c375af313e8a169c77adbaf956693c0cfeb5ccd
Modified code so that:
-When above and left contexts are same and not equal to current segment id, it needs to read a maximum of 2 segment_tree_probabilities.
- When above and left contexts are different and not equal to current segment id, it needs to read only a single segment_tree_probability.
Change-Id: Idc2cf2c4afcc6179b8162ac5a32c948ff5a9a2ba
There is no need to make sure that the lower byte of the
register is 0 because the downshift by 11 overwrites that byte.
Change-Id: I89cbf004b2ff532a2c68e0dc399c45a49cdad5a1
