| Age | Commit message (Collapse) | Author |
|---|
|
When a source has no colour conversion (e.g. an already-XYZ J2K DCP),
every frame normally bypasses encoding via the has_j2k() fast-path. But
has_j2k() returns false as soon as _text is set, so the first
burned-subtitle frame goes through the encode path and hits the
colour_conversion assertion in the Grok variant of
convert_to_xyz(), killing the Grok threads one by one until the queue
can no longer drain.
Handle the no-conversion case the same way the local convert_to_xyz()
overload does: the image is already XYZ12LE, so copy it straight into
the destination buffer.
|
|
There's a few things going on here:
1. Improve the regression test for 3042. Previously we made a DCP from
the reporter's _original_ prores file (before they converted it to XYZ)
and compared the result to a reference J2K file of uncertain origin.
This seems wrong because:
a) We never got confirmation from the reporter that the fix worked for
them, so any arbitrary reference is dubious.
b) It doesn't seem to reflect their actual complaint, which was that
they got a different result when making a DCP from XYZ TIFFs compared
to their "XYZ" Prores.
The new test makes a DCP from their TIFFs and "XYZ" Prores and compares
the result.
2. Revert to the old approach to "no conversion" handling. In the good
old days we did everything -> RGB48LE except XYZ12LE -> XYZ12LE, and
that's what we do again here.
3. Change the YUV->RGB conversion from Rec.601 to Rec.709 for the
"no conversion" case. This fixes the 3042 regression test.
The supposed "XYZ" Prores is yuv444p12le according to ffprobe. So I
think what we have here is actually a file that was converted to XYZ
and then back to YUV by Resolve. I experimented with using the raw
YUV values and considering them as XYZ but this was clearly wrong.
I think 3 is probably what I should have done in the first place.
|
|
For a long time we would keep XYZ12LE, if that's what we have, otherwise
ask FFmpeg to switch to RGB48LE.
Then in 1d5c211dadb9a9dc2318adce86ca9c31b367cabe I tried to fix the case
of an XYZ source mis-tagged as YUV. I changed things so that with no
colour conversion we'd always ask FFmpeg to convert to XYZ. This meant
that RGB sources with no colour conversion would get treatment by FFmpeg
due to the RGB -> XYZ switch.
Here we're going back to the more-or-less the "long time" behaviour when
there is a conversion (keep XYZ12LE but otherwise convert to RGB48).
When there's no conversion, keep RGB (to avoid the FFmpeg conversion
from RGB -> XYZ) but convert everything else to XYZ.
|
|
This was removed in df9b4676aba8b941f124b174393988cad21677e1 and
surrounding commits, but it turns out we need it after all.
|
|
sed -i "/include.*compose.hpp/d;" src/lib/*.cc src/wx/*.cc src/wx/*.h src/tools/*.cc src/lib/*.h test/*.cc
|
|
sed -i "/Plural-Forms/n;/%100/n;/scanf/n;s/%[123456789]/{}/g" src/lib/*.cc src/lib/*.h src/wx/*.cc src/tools/*.cc src/lib/po/*.po src/wx/po/*.po src/tools/po/*.po test/*.cc
sed -i "s/String::compose */fmt::format/g" src/lib/*.cc src/lib/*.h src/wx/*.cc src/tools/*.cc test/*.cc
|
|
|
|
|
|
|
|
AV_PIX_FMT_XYZ12LE.
Prepare the image for where it is going to next. I don't understand the difference
in this situation between XYZ12LE and RGB48LE (it seems they are both 16-bit-per-channel)
but there you go.
|
|
|
|
This avoids some boilerplate.
|
|
raw_convert<string> is gone, so here we use fmt::{to_string,format} instead.
Other raw_converts now use fast_float.
|
|
|
|
|
|
|
|
via his tool "grok".
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Replace it with a fixed 16KB as per #1902.
|
|
|
|
|
|
|
|
|
|
The recovery code assumes that encoding a given frame from the
same data will always give the same output; it always encodes frame 0
again to set things up properly and if that frame is a different size
to the previous run everything breaks.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Our export formats all currently use limited range but we weren't
making sure data fed to the encoders was limited range.
Should fix #1832.
|
|
|
|
Forward-ported-from: d1e9749ca290673639a49d693a8fe5c6557cc2de
|
|
|
|
various places.
|
|
network corruption (#1381).
|
|
|
|
|
|
|
|
|
|
|
|
|
