The scalar loop is replaced with masked AVX512 instructions.
For extracting the Y from UYVY, vperm2b is used instead of
various AND and packuswb.
Instead of loading the vectors with interleaved lanes as done
in AVX2 version, normal load is used. At the end of packuswb,
for U and V, an extra permute operation is done to get the
required layout.
AMD 7950x Zen 4 benchmark data:
uyvytoyuv422_c: 29105.0 ( 1.00x)
uyvytoyuv422_sse2: 3888.0 ( 7.49x)
uyvytoyuv422_avx: 3374.2 ( 8.63x)
uyvytoyuv422_avx2: 2649.8 (10.98x)
uyvytoyuv422_avx512icl: 1615.0 (18.02x)
Signed-off-by: Shreesh Adiga <16567adigashreesh@gmail.com>
Signed-off-by: James Almer <jamrial@gmail.com>
ff_w1111 and ff_bgr2(Y|UV)Offset are only used there
(and only on x86-32 since caaec2ea95).
Also make them static.
Reviewed-by: Michael Niedermayer <michael@niedermayer.cc>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Values in csp, prim, trc, etc, are irrelevant if there's no conversion needed.
Reviewed-by: Niklas Haas <ffmpeg@haasn.xyz>
Signed-off-by: James Almer <jamrial@gmail.com>
The reason for the failure is that the function yuv2rgb_1_c_template was modified
in 095f8038fa. The corresponding functional test
was added in c601bb8df5. The code on loongarch was
not updated in a timely manner, resulting in the error.
Signed-off-by: yuanhecai <yuanhecai@loongson.cn>
Reviewed-by: yinshiyou-hf@loongson.cn
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
Symbols with the sws_* prefix are exported.
Reviewed-by: Alexander Strasser <eclipse7@gmx.net>
Signed-off-by: Andreas Rheinhardt <andreas.rheinhardt@outlook.com>
Fixes: signed integer overflow: -1082982400 + -1195645138 cannot be represented in type 'int'
Fixes: 376136843/clusterfuzz-testcase-minimized-ffmpeg_SWS_fuzzer-4791844321427456
Found-by: continuous fuzzing process https://github.com/google/oss-fuzz/tree/master/projects/ffmpeg
Signed-off-by: Michael Niedermayer <michael@niedermayer.cc>
This has the downside of requiring these tables to be recomputed on every
init, but saves ~270 kB of static data.
Signed-off-by: Niklas Haas <git@haasn.dev>
The current logic uses 12-bit linear light math, which is woefully insufficient
and leads to nasty postarization artifacts. This patch simply switches the
internal logic to 16-bit precision.
This raises the memory requirement of these tables from 32 kB to 272 kB.
All relevant FATE tests updated for improved accuracy.
Fixes: #4829
Signed-off-by: Niklas Haas <git@haasn.dev>
Sponsored-by: Sovereign Tech Fund
While this one was technically supported on account of the generic options
code allowing "none" as a valid alias for 0, not having it listed here meant
it never showed up in e.g. the -h output, and is also inconsistent with other
places in the code that generally add an explicit alias with appropriate
documentation. Reduces user confusion at negligible cost.
Fixes: #9192
Signed-off-by: Niklas Haas <git@haasn.dev>
Sponsored-by: Sovereign Tech Fund
This leverages the previously introduced color management subsystem in order
to adapt between transfer functions and color spaces, as well as for HDR tone
mapping.
Take special care to handle grayscale formats without a colorspace
gracefully.
This is a lightweight wrapper around the underlying color management system,
whose job it is merely to manage the 3DLUT state and apply them to the frame
data. This is where we might add platform-specific optimizations in the future.
I also plan on adding support for more pixel formats in the future. In
particular, we could support YUV or XYZ input formats directly using only
negligible additional code in the 3DLUT setup functions. This would eliminate
the major source of slowdown, which is currently the roundtrip to RGBA64.
The underlying color mapping logic was ported as straightforwardly as possible
from libplacebo, although the API and glue code has been very heavily
refactored / rewritten. In particular, the generalization of gamut mapping
methods is replaced by a single ICC intent selection, and constants have been
hard-coded.
To minimize the amount of overall operations, this gamut mapping LUT now embeds
a direct end-to-end transformation to the output color space; something that
libplacebo does in shaders, but which is prohibitively expensive in software.
In order to preserve compatibility with dynamic tone mapping without severely
regressing performance, we add the ability to generate a pair of "split" LUTS,
one for encoding the input and output to the perceptual color space, and a
third to embed the tone mapping operation. Additionally, this intermediate
space could be used for additional subjective effect (e.g. changing
saturation or brightness).
The big downside of the new approach is that generating a static color mapping
LUT is now fairly slow, as the chromaticity lobe peaks have to be recomputed
for every single RGB value, since correlated RGB colors are not necessarily
aligned in ICh space. Generating a split 3DLUT significantly alleviates this
problem because the expensive step is done as part of the IPT input LUT, which
can share the same hue peak calculation at least for all input intensities.
This setting can be used to infuence the type of tone and gamut mapping used
internally when color space conversions are required. As discussed at VDD'24,
the default was set to relative colorimetric clipping, which is approximately
associative, surjective and idempotent. As such, it roundtrips well, although
it is strictly speaking not associative on out-of-gamut colors.
Swscale currently handles XYZ by embedding a forced conversion to
BT.709 RGB with a hardcoded matrix. This is not ideal, but to preserve the
status quo and avoid any unexpected changes in behavior, this patch merely
fixes the inferred primaries tag to match the reality.
In the future, I would like to handle XYZ properly, via direct conversion
to the target colorspace (or possibly simply by using a more fitting
RGB intermediate like SMPTE428), but for now just keep the status quo.
Logic is loosely on equivalent decisions in libplacebo. The basic idea is to try
and be a bit conservative by treating AVCOL_*_UNSPECIFIED as a no-op, unless the
other primaries set are non-standard / wide-gamut or HDR. This helps avoid
unintended or unexpected colorspace conversions, while forcing it in cases where
we are almost certain it is needed. The major departure from libplacebo semantics
is that we no default to a 1000:1 contrast ration for SDR displays, instead modelling
them as idealized devices with an infinite contrast ratio.
In either case, setting SWS_STRICT overrides this behavior in favor of always
requiring explicit colorspace metadata.
Logic ported from libplacebo's AVFrame helpers. The basic idea is to use the
provided MaxRGB/MaxSCL values to infer what the actual luminance would have
been, which HDR10+ metadata does not provide directly. It's worth pointing out
that this gives us an *upper* bound on the true maximum luminance, so any
error in the estimation cannot result in clipping.
Only add the condensed values that we actually care about. Group them into
a new struct to make it easier to discard or replace this metadata.
Define a special comparison function that does not choke on undefined/unknown
metadata.
As fixed in the previous commit, this enables semipacked range and
bit depth conversions. Previously these would go through the general
purpose path.
Signed-off-by: Niklas Haas <git@haasn.dev>
Sponsored-by: Sovereign Tech Fund
This fixes multiple bugs with semiplanar formats like NV12. Not only do these
false positive the grayscale format checks (because dst[2] in NULL), but they
also copied an incorrect number of pixels.
Fixes conversions such as nv12 -> nv12, gray8 -> nv12, nv20le -> nv20be, etc.
Fixes: #11239
Signed-off-by: Niklas Haas <git@haasn.dev>
Sponsored-by: Sovereign Tech Fund
We should at least bias towards the nearest integer, instead of always
rounding down, when not dithering. This is a bit more correct.
The FATE changes are only in the cases where sws_dither was explicitly set
to "none", which is exactly as expected.
Signed-off-by: Niklas Haas <git@haasn.dev>
Sponsored-by: Sovereign Tech Fund
These functions were divided into two special cases; one assuming that
uvalpha == 0, and the other assuming that uvalpha == 2048. This worked fine
for simple 2x chroma upscaling but broke for e.g. yuv410p, non-centered chroma,
or other special cases that involved non-aligned chroma filters.
Fix it by instead dividing this check into two cases, a uvalpha==0 fast path
and a uvalpha>0 general path. Instead of (A+B)/2 the general path now multiplies
in the true uvalpha weight.
I tried preserving the old fast path for the case of uvalpha == 2048, but this
was significantly slower in practise versus having just one general path.
However, we still need a uvalpha == 0 path for the unscaled case.
Fixes: ticket #5083
Signed-off-by: Niklas Haas <git@haasn.dev>
Sponsored-by: Sovereign Tech Fund
Since its introduction, this function has claimed to return 0 on success, yet
never actually did so (until the introduction of the new graph based API). It
always returned the number of scaled lines, and continues to do so.
To avoid confusion, but also avoid regressing possible clients that relied on
the existing semantics, simply update the documentation to reflect the actual
behavior. Remain ambiguous about the exact interpretation of the return value
on account of the unfortunate difference in behavior between the legacy and
new scaling APIs.
Signed-off-by: Niklas Haas <git@haasn.dev>
Sponsored-by: Sovereign Tech Fund
This logic was inverted, but || was not replaced by &&.
Fixes: ed5dd67562
Fixes: ticket #11353
Signed-off-by: Niklas Haas <git@haasn.dev>
Sponsored-by: Sovereign Tech Fund
In input.c and output.c and many other places, swscale follows the rule of using
15-bit intermediate if output bpc is <= 8, and 19-bit (inside int32_t)
intermediate otherwise. See e.g. the comments on hyScale() on
swscale_internal.h. These are also the coefficients that yuv2gbrpf32_full_X_c()
is using.
In contrast to this, the plane init code in slice.c (function fill_ones) is
assuming that we use 35-bit intermediates (inside 64-bit integers) for this
case, seemingly added by commit b4967fc71c with
no further justification.
This causes a mismatch whenever the implicitly initialized plane contents leak
out to the output, e.g. when converting from grayscale to RGB.
Fixes: ticket #10716
Signed-off-by: Niklas Haas <git@haasn.dev>
Sponsored-by: Sovereign Tech Fund
We only parse 12 values, so this check always failed. Regression caused by
a change to the print format.
Fixes: 59c39a79ca
Signed-off-by: Niklas Haas <git@haasn.dev>
Sponsored-by: Sovereign Tech Fund
There is an issue with the constants used in YUV to YUV range conversion,
where the upper bound is not respected when converting to mpeg range.
With this commit, the constants are calculated at runtime, depending on
the bit depth. This approach also allows us to more easily understand how
the constants are derived.
For bit depths <= 14, the number of fixed point bits has been set to 14
for all conversions, to simplify the code.
For bit depths > 14, the number of fixed points bits has been raised and
set to 18, to allow for the conversion to be accurate enough for the mpeg
range to be respected.
The convert functions now take the conversion constants (coeff and offset)
as function arguments.
For bit depths <= 14, coeff is unsigned 16-bit and offset is 32-bit.
For bit depths > 14, coeff is unsigned 32-bit and offset is 64-bit.
x86_64:
chrRangeFromJpeg8_1920_c: 2127.4 2125.0 (1.00x)
chrRangeFromJpeg16_1920_c: 2325.2 2127.2 (1.09x)
chrRangeToJpeg8_1920_c: 3166.9 3168.7 (1.00x)
chrRangeToJpeg16_1920_c: 2152.4 3164.8 (0.68x)
lumRangeFromJpeg8_1920_c: 1263.0 1302.5 (0.97x)
lumRangeFromJpeg16_1920_c: 1080.5 1299.2 (0.83x)
lumRangeToJpeg8_1920_c: 1886.8 2112.2 (0.89x)
lumRangeToJpeg16_1920_c: 1077.0 1906.5 (0.56x)
aarch64 A55:
chrRangeFromJpeg8_1920_c: 28835.2 28835.6 (1.00x)
chrRangeFromJpeg16_1920_c: 28839.8 32680.8 (0.88x)
chrRangeToJpeg8_1920_c: 23074.7 23075.4 (1.00x)
chrRangeToJpeg16_1920_c: 17318.9 24996.0 (0.69x)
lumRangeFromJpeg8_1920_c: 15389.7 15384.5 (1.00x)
lumRangeFromJpeg16_1920_c: 15388.2 17306.7 (0.89x)
lumRangeToJpeg8_1920_c: 19227.8 19226.6 (1.00x)
lumRangeToJpeg16_1920_c: 15387.0 21146.3 (0.73x)
aarch64 A76:
chrRangeFromJpeg8_1920_c: 6324.4 6268.1 (1.01x)
chrRangeFromJpeg16_1920_c: 6339.9 11521.5 (0.55x)
chrRangeToJpeg8_1920_c: 9656.0 9612.8 (1.00x)
chrRangeToJpeg16_1920_c: 6340.4 11651.8 (0.54x)
lumRangeFromJpeg8_1920_c: 4422.0 4420.8 (1.00x)
lumRangeFromJpeg16_1920_c: 4420.9 5762.0 (0.77x)
lumRangeToJpeg8_1920_c: 5949.1 5977.5 (1.00x)
lumRangeToJpeg16_1920_c: 4446.8 5946.2 (0.75x)
NOTE: all simd optimizations for range_convert have been disabled.
they will be re-enabled when they are fixed for each architecture.
NOTE2: the same issue still exists in rgb2yuv conversions, which is not
addressed in this commit.
Shreesh Adiga
Krzysztof Pyrkosz
James Almer
Andreas Rheinhardt
Michael Niedermayer
yuanhecai
Niklas Haas
Ramiro Polla