The primary goal is to allow a binary to be built which supports
NEON, but can fall back to non-NEON routines, since some Android
devices do not have NEON, even if they are otherwise ARMv7 (e.g.,
Tegra).
The configure-generated flags HAVE_ARMV7, etc., are used to decide
which versions of each function to build, and when
CONFIG_RUNTIME_CPU_DETECT is enabled, the correct version is chosen
at run time.
In order for this to work, the CFLAGS must be set to something
appropriate (e.g., without -mfpu=neon for ARMv7, and with
appropriate -march and -mcpu for even earlier configurations), or
the native C code will not be able to run.
The ASFLAGS must remain set for the most advanced instruction set
required at build time, since the ARM assembler will refuse to emit
them otherwise.
I have not attempted to make any changes to configure to do this
automatically.
Doing so will probably require the addition of new configure options.
Many of the hooks for RTCD on ARM were already there, but a lot of
the code had bit-rotted, and a good deal of the ARM-specific code
is not integrated into the RTCD structs at all.
I did not try to resolve the latter, merely to add the minimal amount
of protection around them to allow RTCD to work.
Those functions that were called based on an ifdef at the calling
site were expanded to check the RTCD flags at that site, but they
should be added to an RTCD struct somewhere in the future.
The functions invoked with global function pointers still are, but
these should be moved into an RTCD struct for thread safety (I
believe every platform currently supported has atomic pointer
stores, but this is not guaranteed).
The encoder's boolhuff functions did not even have _c and armv7
suffixes, and the correct version was resolved at link time.
The token packing functions did have appropriate suffixes, but the
version was selected with a define, with no associated RTCD struct.
However, for both of these, the only armv7 instruction they actually
used was rbit, and this was completely superfluous, so I reworked
them to avoid it.
The only non-ARMv4 instruction remaining in them is clz, which is
ARMv5 (not even ARMv5TE is required).
Considering that there are no ARM-specific configs which are not at
least ARMv5TE, I did not try to detect these at runtime, and simply
enable them for ARMv5 and above.
Finally, the NEON register saving code was completely non-reentrant,
since it saved the registers to a global, static variable.
I moved the storage for this onto the stack.
A single binary built with this code was tested on an ARM11 (ARMv6)
and a Cortex A8 (ARMv7 w/NEON), for both the encoder and decoder,
and produced identical output, while using the correct accelerated
functions on each.
I did not test on any earlier processors.
Change-Id: I45cbd63a614f4554c3b325c45d46c0806f009eaa
Changes 'The VP8 project' to 'The WebM project', for consistency
with other webmproject.org repositories.
Fixes issue #97.
Change-Id: I37c13ed5fbdb9d334ceef71c6350e9febed9bbba
Moving the eob structure allows for a non-struct based
function to handle decoding an entire mb of
idct/dequant/recon data. This allows for SIMD functions
to idct/dequant/recon multiple blocks at once.
SSE2 implementation gives 3% gain on Atom.
Change-Id: I8a8f3efd546ea4e0535f517d94f347cfb737c9c2
Jeff Muizelaar posted some changes to the idct/reconstruction c code.
This is the equivalent update for the arm assembly.
This shows a good boost on v6, and a minor boost on neon.
Here are some numbers for highway in qcif, 2641 frames:
HEAD neon: ~161 fps
new neon: ~162 fps
HEAD v6: ~102 fps
new v6: ~106 fps
The following functions have been updated for armv6 and neon:
vp8_dc_only_idct_add
vp8_dequant_idct_add
vp8_dequant_dc_idct_add
Conflicts:
vp8/decoder/arm/armv6/dequantdcidct_v6.asm
vp8/decoder/arm/armv6/dequantidct_v6.asm
Resolved by removing these files. When I rewrote the functions, I also
moved the files to dequant_dc_idct_v6.asm/dequant_idct_v6.asm
Change-Id: Ie3300df824d52474eca1a5134cf22d8b7809a5d4
This moves the prediction step before the idct and combines the idct and
reconstruction steps into a single step. Combining them seems to give an
overall decoder performance improvement of about 1%.
Change-Id: I90d8b167ec70d79c7ba2ee484106a78b3d16e318
When the license headers were updated, they accidentally contained
trailing whitespace, so unfortunately we have to touch all the files
again.
Change-Id: I236c05fade06589e417179c0444cb39b09e4200d
Change bitreading functions to use a larger window which is refilled less
often.
This makes it cheap enough to do bounds checking each time the window is
refilled, which avoids the need to copy the input into a large circular
buffer.
This uses less memory and speeds up the total decode time by 1.6% on an ARM11,
2.8% on a Cortex A8, and 2.2% on x86-32, but less than 1% on x86-64.
Inlining vp8dx_bool_decoder_fill() has a big penalty on x86-32, as does moving
the refill loop to the front of vp8dx_decode_bool().
However, having the refill loop between computation of the split values and
the branch in vp8_decode_mb_tokens() is a big win on ARM (presumably due to
memory latency and code size: refilling after normalization duplicates the
code in the DECODE_AND_BRANCH_IF_ZERO and DECODE_AND_LOOP_IF_ZERO cases.
Unfortunately, refilling at the end of vp8dx_bool_decoder_fill() and at the
beginning of each decode step in vp8_decode_mb_tokens() means the latter
requires an extra refill at the end.
Platform-specific versions could avoid the problem, but would require most of
detokenize.c to be duplicated.
Change-Id: I16c782a63376f2a15b78f8086d899b987204c1c7