On Wed, 26 Mar 2025, Martin Storsjö wrote:

This allows catching whether the functions write outside of
the designated rectangle, and if run with "checkasm -v", it also
prints out on which side of the rectangle the overwrite was.
---
tests/checkasm/hevc_pel.c | 157 ++++++++++++++++++++++----------------
1 file changed, 90 insertions(+), 67 deletions(-)

This patch exposes existing issues in the Loongson HEVC simd; running with "checkasm --test=hevc_pel -v" after applying these patches, shows the following errors:

LSX:
 - hevc_pel.qpel       [OK]
 - hevc_pel.qpel_uni   [OK]
 - hevc_pel.qpel_uni_w [OK]
   put_hevc_pel_bi_pixels12_8_lsx (src/tests/checkasm/hevc_pel.c:272)
dst (12x12):
 overwrite below
 - hevc_pel.qpel_bi    [FAILED]
 - hevc_pel.epel       [OK]
   put_hevc_epel_uni_v24_8_lsx (src/tests/checkasm/hevc_pel.c:433)
dst (24x24):
 overwrite below
   put_hevc_epel_uni_hv12_8_lsx (src/tests/checkasm/hevc_pel.c:433)
dst (12x12):
 overwrite below
 - hevc_pel.epel_uni   [FAILED]
 - hevc_pel.epel_uni_w [OK]
   put_hevc_epel_bi_hv6_8_lsx (src/tests/checkasm/hevc_pel.c:545)
dst (6x6):
 overwrite below
 - hevc_pel.epel_bi    [FAILED]
LASX:
 - hevc_pel.qpel_uni   [OK]
 - hevc_pel.qpel_uni_w [OK]
 - hevc_pel.epel_uni_w [OK]
 - hevc_pel.epel_bi    [OK]
checkasm: 4 of 231 tests have failed


So the functions seem to produce more output rows than requested, for "uneven" heights like 6, 12 or 24 pixels.

Can someone from Loongson please have a look at this?

We don't seem to have any Loongson instances on fate, so can I go ahead and push this, without waiting for the Loongson assembly to be fixed? We do have some on patchwork though, so I guess that'll make patchwork noisy for everybody else until it is fixed, so I guess I'll have to hold off of this patch until it is fixed.

// Martin
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".

Reply via email to