Replace generic with block size specific function.
Signed-off-by: Kaustubh Raste <kaustubh.raste@imgtec.com>
Reviewed-by: Manojkumar Bhosale <Manojkumar.Bhosale@imgtec.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
The initialisation should be common. For libmfx, it was previously
happening in the derivation function and this moves it out. For VAAPI,
it fixes some failures when deriving from a DRM device because this
initialisation did not run.
* commit 'e6bff23f1e11aefb16a2b5d6ee72bf7469c5a66e':
cpu: add a function for querying maximum required data alignment
Adapted to work with the arbitrary runtime cpuflag changes av_force_cpu_flags()
can generate.
Merged-by: James Almer <jamrial@gmail.com>
Replace generic with block size specific function.
Signed-off-by: Kaustubh Raste <kaustubh.raste@imgtec.com>
Reviewed-by: Manojkumar Bhosale <Manojkumar.Bhosale@imgtec.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Load the specific destination bytes instead of MSA load and pack.
Pack the data to half word before clipping.
Use immediate unsigned saturation for clip to max saving one vector register.
Signed-off-by: Kaustubh Raste <kaustubh.raste@imgtec.com>
Reviewed-by: Manojkumar Bhosale <Manojkumar.Bhosale@imgtec.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Preload data in band filter 0-8 for better pipeline parallelization.
Signed-off-by: Kaustubh Raste <kaustubh.raste@imgtec.com>
Reviewed-by: Manojkumar Bhosale <Manojkumar.Bhosale@imgtec.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Refer to "checkasm: use perf API on Linux ARM*" commit for the
rationale.
The implementation is somehow duplicated with checkasm, but so is the
current usage of AV_READ_TIME(). Until these implementations and
heuristics are made consistent, I don't see a way of sharing that code.
Note: when using libavutil/timer.h, it is now important to include
before any other include due to the _GNU_SOURCE requirement.
Load the specific destination bytes instead of MSA load and pack.
Signed-off-by: Kaustubh Raste <kaustubh.raste@imgtec.com>
Reviewed-by: Manojkumar Bhosale <Manojkumar.Bhosale@imgtec.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Improved version of VBROADCASTSS that works like the avx2 instruction.
Emulation of vpbroadcastd.
Horizontal sum HSUMPS that places the result in all elements.
Emulation of blendvps and pblendvb.
Signed-off-by: Ivan Kalvachev <ikalvachev@gmail.com>
If using the winstore compat library, a fallback LoadLibrary
function does exist, that only calls LoadPackagedLibrary though
(which doesn't work for dynamically loading d3d11 DLLs).
Therefore explicitly check the targeted API family instead.
Make this check a reusable HAVE_* component which other parts
of the libraries can check when necessary as well.
Signed-off-by: Martin Storsjö <martin@martin.st>
Merged from Libav commit 4d330da006.
Black isn't always just memset(ptr, 0, size). Limited YUV in particular
requires relatively non-obvious values, and filling a frame with
repeating 0 bytes is disallowed in some contexts. With component sizes
larger than 8 or packed YUV, this can become relatively complicated. So
having a generic function for this seems helpful.
In order to handle the complex cases in a generic way without destroying
performance, this code attempts to compute a black pixel, and then uses
that value to clear the image data quickly by using a function like
memset.
Common cases like yuv410p10 or rgba can't be handled with a simple
memset, so there is some code to fill memory with 2/4/8 byte patterns.
For the remaining cases, a generic slow fallback is used.
Signed-off-by: Anton Khirnov <anton@khirnov.net>
Merged from Libav commit 45df7adc1d.
Many image formats support embedding of ICC profiles directly in
their bitstreams. Add a new side data type to allow exposing them to
API users.
Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
Reduced msa load-store code.
Removed inline asm of GP load-store for 64 bit.
Updated variable names in GP load-store macros for naming consistency.
Corrected macro descriptions.
Signed-off-by: Kaustubh Raste <kaustubh.raste@imgtec.com>
Reviewed-by: Manojkumar Bhosale <Manojkumar.Bhosale@imgtec.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Rework it to improve performance. Now mutex is not shared by workers,
instead each worker has its own mutex and condition variable. This
reduces lock contention between workers. Also use atomic variable for
counter.
The interface also allows execute to run special function on main
thread, requested by Ronald.
Signed-off-by: Muhammad Faiz <mfcc64@gmail.com>
