PR #23024 opened by Niklas Haas (haasn)
URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/23024
Patch URL: https://code.ffmpeg.org/FFmpeg/FFmpeg/pulls/23024.patch
In an op list like:
[ u8 +XXX] SWS_OP_READ : 1 elem(s) planar >> 3
[ u8 .XXX] SWS_OP_FILTER_V : 256 -> 320 bilinear (2 taps)
[f32 .XXX] SWS_OP_SCALE : * 65535
[f32 +XXX] SWS_OP_CONVERT : f32 -> u16
[u16 zXXX] SWS_OP_SWAP_BYTES
[u16 zzzX] SWS_OP_SWIZZLE : 0003
[u16 zzz+] SWS_OP_CLEAR : {_ _ _ 65535}
[u16 XXXX] SWS_OP_WRITE : 4 elem(s) packed >> 0
The current version of the code would happily push the SWS_OP_SCALE past
the truncating conversion, leading to degenerate loss of information. (In
this case, the result was quite extreme)
Incidentally, this does not change the output of the sws_ops test because that
test does not currently include split scaling passes.
Signed-off-by: Niklas Haas <[email protected]>
>From 423b0eb1ecb2339c1d07348d998e4f114a8095dd Mon Sep 17 00:00:00 2001
From: Niklas Haas <[email protected]>
Date: Tue, 5 May 2026 20:11:25 +0200
Subject: [PATCH] swscale/ops_optimizer: don't push scale past truncating
conversions
In an op list like:
[ u8 +XXX] SWS_OP_READ : 1 elem(s) planar >> 3
[ u8 .XXX] SWS_OP_FILTER_V : 256 -> 320 bilinear (2 taps)
[f32 .XXX] SWS_OP_SCALE : * 65535
[f32 +XXX] SWS_OP_CONVERT : f32 -> u16
[u16 zXXX] SWS_OP_SWAP_BYTES
[u16 zzzX] SWS_OP_SWIZZLE : 0003
[u16 zzz+] SWS_OP_CLEAR : {_ _ _ 65535}
[u16 XXXX] SWS_OP_WRITE : 4 elem(s) packed >> 0
The current version of the code would happily push the SWS_OP_SCALE past
the truncating conversion, leading to degenerate loss of information. (In
this case, the result was quite extreme)
Incidentally, this does not change the output of the sws_ops test because that
test does not currently include split scaling passes.
Signed-off-by: Niklas Haas <[email protected]>
---
libswscale/ops_optimizer.c | 15 +++++++++++++--
1 file changed, 13 insertions(+), 2 deletions(-)
diff --git a/libswscale/ops_optimizer.c b/libswscale/ops_optimizer.c
index c7972e471f..619c6cf42b 100644
--- a/libswscale/ops_optimizer.c
+++ b/libswscale/ops_optimizer.c
@@ -337,6 +337,16 @@ static bool extract_swizzle(SwsLinearOp *op, SwsComps
prev, SwsSwizzleOp *out_sw
return true;
}
+static int op_result_is_exact(const SwsOp *op)
+{
+ for (int i = 0; i < 4; i++) {
+ if (SWS_OP_NEEDED(op, i) && !(op->comps.flags[i] & SWS_COMP_EXACT))
+ return false;
+ }
+
+ return true;
+}
+
int ff_sws_op_list_optimize(SwsOpList *ops)
{
int ret;
@@ -769,9 +779,10 @@ retry:
}
case SWS_OP_SCALE:
- /* Scaling by integer before conversion to int */
+ /* Exact integer multiplication */
if (op->scale.factor.den == 1 && next->op == SWS_OP_CONVERT &&
- ff_sws_pixel_type_is_int(next->convert.to))
+ ff_sws_pixel_type_is_int(next->convert.to) &&
+ op_result_is_exact(op))
{
op->type = next->convert.to;
FFSWAP(SwsOp, *op, *next);
--
2.52.0
_______________________________________________
ffmpeg-devel mailing list -- [email protected]
To unsubscribe send an email to [email protected]
