texture() uses bilinear scaling; imageLoad() accesses the image directly.
The reason why texture() was used throughout Vulkan filters is that
back when they were written, they were targetting old Intel hardware,
which had a texel cache only for sampled images.
These days, GPUs have a generic cache that doesn't care what source it
gets populated with. Additionally, bypassing the sampling circuitry saves
us some performance.
Finally, all the old texture() code had an issue where unnormalized
coordinates were used, but an offset of 0.5 was not added, hence each
pixel ended up being interpolated. This fixes this.
4b77a0a681 added a new consumer of ff_scale_adjust_dimensions
which was recently modified to allow for square pixel output.
This commit extends the new option to vpp_amf, and unbreaks the building
of vf_amf_common.c
For anamorphic videos, enabling this option leads to adjustment of
output dimensions to obtain square pixels when the user requests
proportional scaling through either of the w/h expressions or
force_original_aspect_ratio.
Output SAR is always reset to 1.
Option added to scale, scale_cuda, scale_npp & scale_vaapi.
libplacebo already has a similar option with different semantics,
scale_vt and scale_vulkan don't implement force_oar, so for these
three filters, I've made minimal changes needed to not break building
or change output.
Free AVOption in drawtext private context when freeing old copy of drawtext
private context during processing of "reinit" runtime command.
Signed-off-by: Steven Zhou <steven.zhou@netint.ca>
Signed-off-by: James Almer <jamrial@gmail.com>
The ref input may have its frame rate unset, which would then lead to
SIGFPE. So fall back to the main link frame rate. If that too is unset,
default to 0.
Related to #11428
This commit adds two AMF filters: vpp_amf & sr_amf.
Both filters are using AMF hardware acceleration.
vpp_amf supports simple scaling algorithms & color conversion.
sr_amf supports advanced scaling algorithms such as FSR & can
be used for upscaling only.
In f121d95, the outlink framerate was unconditionally unset.
This breaks/bloats outputs from CFR muxers unless the user explicitly
sets a sane framerate. And the most common invocation for setpts seen in
workflows, our docs and across the web is `PTS-STARTPTS` or others of the
general form `PTS+constant` which preserves the input framerate.
Default value is false, which restores old behaviour.
Fixes#11428
Fixes: applying zero offset to null pointer
partly Fixes: verysmall.flv
Found-by: Elias Myllymäki <elias.myllymaki04p@gmail.com>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
This filter uses the AVBuffer API to allocate buffers that are never
shared at all and frees them via av_freep() (actually, it passes
pointers to AVBufferRefs to av_freep, so that only the AVBuffer
structures are freed at all (because AVBufferRef has a AVBuffer* as its
first member), not the AVBufferRef and not the underlying buffers;
and due to a wrong check the AVBuffers corresponding
to buf_org[c] with c>0 were never freed at all). This is a violation
of the AVBuffer API and causes a memleak. Fix this by avoiding the
AVBuffer API altogether.
(The FATE tests don't catch this, because they use piping to awk,
so that the error code from ffmpeg is ignored.)
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
This patch is analogous to 20f9727018:
It hides the internal part of AVFilter by adding a new internal
structure FFFilter (declared in filters.h) that has an AVFilter
as its first member; the internal part of AVFilter is moved to
this new structure.
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Fixes: crash
Found-by: Elias Myllymäki <elias.myllymaki04p@gmail.com>
Reviewed-by: Alexander Strasser <eclipse7@gmx.net>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Fixes: out of array read
Found-by: Elias Myllymäki <elias.myllymaki04p@gmail.com>
Reviewed-by: Alexander Strasser <eclipse7@gmx.net>
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
The scale_vulkan filter initializes the shader once, with the crop
region set by the original frame. However, subsequent frames may
specify a different crop region than the first frame. This change
updates the cropping to match the behavior present on the other
hardware frame scale filters.
The scale filter should also allow negative values
that respect aspect ratio, similar to other scale filters.
Signed-off-by: Koushik Dutta <koushd@gmail.com>
References https://trac.ffmpeg.org/ticket/11387
Defining anchors is needed, even when a filename with the labels is set.
The issue was identified when using yolov4-tiny model using openvino.
More information about how to reproduce the bug can be found on the trac
issue referenced by this commit.
Signed-off-by: Leandro Santiago <leandrosansilva@gmail.com>
The string type is a remnant of the old channel layout API implementation.
Reviewed-by: Nicolas George <george@nsup.org>
Signed-off-by: James Almer <jamrial@gmail.com>
A sample rate <= 0 is invalid.
Fixes an assert in ffmpeg_enc.c that assumed a valid sample rate would be set.
Fixes ticket #11385.
Signed-off-by: James Almer <jamrial@gmail.com>
Switches to av_frame_side_data_remove_by_props(), covering a number of
cases that we previously ignored. Additionally, stop stripping metadata
when merely changing colorspace or color range, since these do not
affect the actual color volume of the image data, only the encoding.
We recently introduced a public field which was a superset
of the queue context we used to have.
Switch to using it entirely.
This also allows us to get rid of the NIH function which was
valid only for video queues.
