On 4/24/24 12:22, Robin Dapp wrote:
> The dynamic icounts looks sane (vs. Apr 10 snapshot) except for a
>> regression in x264 which is likely independent of the chaos going on.
>>
>>      Apr 10     |     Apr 23      |
>>   109f1b28fc94  |  6f0a646dd2fc   |
>> ----------------+-----------------+--------
>> 276,584,692,883 | 277,816,987,018 |  -0.45%
>> 913,452,236,000 | 927,291,935,180 |  -1.52%
>> 903,916,092,805 | 915,364,006,176 |  -1.27%
> x264 uses widening arithmetic so it could be the reverts.
> Can you compare the hot functions (e.g. x264_pixel_sad_16x16)

Function                                     old     new   delta
x264_pixel_sad_x4_16x8.lto_priv             5188    5288    +100
x264_pixel_sad_x4_8x16.lto_priv             5844    5924     +80
x264_pixel_sad_x3_16x8.lto_priv             3904    3980     +76
x264_pixel_sad_x4_16x16.lto_priv             834     898     +64
x264_pixel_sad_x3_8x16.lto_priv             4408    4468     +60
x264_pixel_sad_x4_8x8.lto_priv              3010    3058     +48
x264_pixel_sad_x3_8x8.lto_priv              2290    2338     +48
...
...
x264_pixel_sad_x4_4x8.lto_priv              1366    1362      -4
x264_pixel_sad_x4_4x4.lto_priv               716     712      -4
x264_pixel_sad_4x8.lto_priv                  332     328      -4
x264_pixel_sad_4x4.lto_priv                  172     168      -4
hpel_filter.lto_priv                         984     980      -4



> if anything stands out surrounding the vwadd.wv for example?

Yeah it does:  not specifically in the routine you mentioned above but
in its various brethren: see attached objdump for
x264_pixel_sad_x4_16x16 () for the 2 cases.

-Vineet