Fixes: out of array access
Fixes: matejsmycka/poc.mp4
Introducing commit: baad75cafa6bac298b72c177f657a2eb8e31cff1 — "aacdec_usac: add support for parsing Mpsp212 (MPEG surround)", 2025-11-17.
Found-by: Matěj Smyčka [email protected]
Signed-off-by: Michael Niedermayer [email protected]
I'll relay that Q to the author.
]]>Not all I-frames can be used for random-access, we should also
respect bit 7 of the index table entry flags which indicates random
access.
Refs: #22467
]]>What about the makeinfo checks in configure?
@Lynne wrote in #22609 (comment):
]]>Plan to push this on Thursday, barring objections.
]]>Plan to push this on Thursday, barring objections.
]]>A lot of the warnings are due to drivers, and valgrind not dealing with GPU mapped memory well. The log is as clean as all the other Vulkan code, so I think its okay.
]]>@Lynne wrote in #22609 (comment):
]]>can directly check whether ret > 13 here to simplify the code
@ramiro wrote in #22609 (comment):
]]>I'd prefer to leave them commented out to remind me and others of what has to happen.
]]>The commit should be splited into two parts:
]]>Valgrind still has a lot of uninitialised value errors. It seems to come from the output buffer after a hw conversion.
]]>It seems I forgot to comment on this, but do they need to be commented out like this? Can't they be checked for later?
]]>@haasn Any comment on this?
]]>Oh, forgot to add a note to "configure: enable Vulkan only if the SPIR-V headers are installed". Added the following:
]]>@ramiro Updated PR with your feedback. I split up the SPIR-V construction into its 3 logical sections, the same way the file is required to be laid out: headers+decorations, types+constants, descriptors.
It should be more organized, especially to those familiar with SPIR-V.
This change is no longer necessary after fmt_is_supported_by_hw.
Add ARM MVE SIMD optimization for resample_common in libswresample,
targeting Cortex-M processors with ARMv8.1-M Helium extension.
MVE and NEON are mutually exclusive (Cortex-M vs Cortex-A), so
compile-time selection is used. The assembly uses MVE tail-predicated
loops (wlstp/letp) which handle arbitrary filter lengths without
remainder loops. S32P support is added as MVE-only since vmlaldava.s32
provides native 32x32->64 bit accumulation not available in NEON.
Also rename resample.S to resample_neon.S for clarity.
Benchmarked on Cortex-M55, 200ms 2ch stereo, speedup vs C fallback:
44.1->48k 48->44.1k 8->48k 48->8k 16->48k 48->16k
S16P: 1.70x 1.38x 1.77x 2.45x 1.75x 2.04x
FLTP: 1.32x 1.19x 1.32x 2.14x 1.34x 1.76x
S32P: 1.38x 1.17x 1.37x 1.80x 1.36x 1.59x
Signed-off-by: Jianguang Ren [email protected]
]]>Previous commits have mentioned this a few times, but sure, I'll add a comment.
]]>Sure, will do.
]]>