PR #25859 opened by ksu-gitreaper807
URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/25859
Patch URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/25859.patch

RGB input can still run its U/V conversion callbacks when the destination is 
grayscale: `ff_init_filters()` schedules the input chroma-conversion descriptor 
even though horizontal chroma scaling is already disabled. This patch excludes 
grayscale destinations from `need_chr_conv` and retains the existing no-chroma 
descriptor for slice metadata.

For a 65x18 RGB0→GRAY8 diagnostic case, actual U/V converter calls dropped from 
18 to zero with unchanged output. The source change is limited to 
`libswscale/slice.c`.

Validation:

- Baseline/patched active output bytes matched across 2,912 cases with native 
CPU dispatch and another 2,912 with dispatch forced to scalar, per tested 
target.
- Another 112 modern frame API cases matched per target, using the legacy 
backend with one and three threads.
- Targets: x86_64 Linux; physical ARM64 Android 13; QEMU AArch64 and big-endian 
PowerPC64. The matrix includes grayscale and color outputs, RGB0/BGR0/0RGB/0BGR 
layouts, planar inputs, endian-specific pixel formats, odd widths, scaling, 
negative strides, and sliced input.
- `fate-sws-range-convert`, `fate-sws-unscaled-threaded`, and 
`fate-sws-floatimg-cmp` passed on baseline and patched non-x86 builds.
- The configured `make -j6 fate` suite passed on the rebased x86 build (201 
tests).

Conversion-only timing on an Intel Core i7-12800H, single thread pinned to one 
CPU, native x86 dispatch, RGB0 input, BT.709/limited output range, bilinear, 
unchanged dimensions:

| Output | Resolution | Baseline ms/frame | Patched ms/frame |
|---|---|---:|---:|
| GRAY8 | 1920x1080 | 7.498 | 5.707 |
| GRAY8 | 3840x2160 | 30.813 | 24.729 |

Values are medians of seven alternating baseline/patched runs, with five warmup 
and 60 timed frames per case. The unchanged YUV420P control varied by 
approximately -2% to +5%; no YUV420P speedup is claimed. No phone or emulated 
performance results are claimed.

The detailed comparisons and timings were collected against `7afbfcbd92d6`. 
This branch is based on upstream `cc583787bbec`; there are no intervening 
`libswscale` changes.

Diagnostic sources, per-target result hashes, individual timings, and the 
configured x86 FATE log are available in the [validation 
branch](https://github.com/ksu-gitreaper807/FFmpeg/tree/validation/sws-gray-chroma/validation/gray-chroma).


From 5dea1542165c6946efcc6766088dc29000e1ac85 Mon Sep 17 00:00:00 2001
From: ksu-gitreaper807 <[email protected]>
Date: Sun, 8 Nov 2026 10:00:15 +0000
Subject: [PATCH] swscale: skip chroma conversion for grayscale output

Grayscale output already bypasses horizontal chroma scaling, but input
conversion callbacks can still cause the chroma conversion descriptor to
be scheduled. RGB input therefore computes U and V even though neither
component is used by the output.

Do not schedule the input chroma conversion for grayscale destinations.
Keep the existing no-chroma descriptor to maintain slice metadata.

For RGB0 to GRAY8 with native x86 dispatch, local conversion-only timing
showed about 24% less time at 1920x1080 and 20% less at 3840x2160 on an
Intel Core i7-12800H. Tested output is unchanged with native and scalar
dispatch on x86_64, physical ARM64 Android, and QEMU AArch64 and big-endian
PowerPC64.
---
 libswscale/slice.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/libswscale/slice.c b/libswscale/slice.c
index 44c3bd74b44..0dbd40624d9 100644
--- a/libswscale/slice.c
+++ b/libswscale/slice.c
@@ -251,7 +251,8 @@ int ff_init_filters(SwsInternal * c)
     int num_cdesc;
     int num_vdesc = isPlanarYUV(c->opts.dst_format) && 
!isGray(c->opts.dst_format) ? 2 : 1;
     int need_lum_conv = c->lumToYV12 || c->readLumPlanar || c->alpToYV12 || 
c->readAlpPlanar;
-    int need_chr_conv = c->chrToYV12 || c->readChrPlanar;
+    int need_chr_conv = !isGray(c->opts.dst_format) &&
+                        (c->chrToYV12 || c->readChrPlanar);
     int need_gamma = c->is_internal_gamma;
     int srcIdx, dstIdx;
     int dst_stride = FFALIGN(c->opts.dst_w * sizeof(int16_t) + 66, 16);
-- 
2.54.0

_______________________________________________
ffmpeg-devel mailing list -- [email protected]
To unsubscribe send an email to [email protected]

Reply via email to