volk_32fc_s32fc_rotator2puppet_32fc fails in {a,u}_sse4_1

[marcusmueller]

found this while working on packaging: ctest fails in a container, and pretty drastically so:

offset 3 in1: 0.85403 + 0.188718j in2: -0.00175621 + 0.0954956j tolerance was: 0.001

|in2| is too small by a factor of roughly 10.

Full output

105/218 Test #105: qa_volk_32fc_s32fc_rotator2puppet_32fc ......................................................................................***Failed    0.01 sec
RUN_VOLK_TESTS: volk_32fc_s32fc_rotator2puppet_32fc(131071,1)
generic completed in 0.233065 ms
a_sse4_1 completed in 0.110075 ms
u_sse4_1 completed in 0.109965 ms
a_avx completed in 0.056566 ms
u_avx completed in 0.056075 ms
a_avx_fma completed in 0.061925 ms
u_avx_fma completed in 0.062336 ms
offset 2 in1: -0.84885 + 0.194543j  in2: -0.805366 + 0.269719j tolerance was: 0.001
offset 3 in1: 0.85403 + 0.188718j  in2: -0.00175621 + 0.0954956j tolerance was: 0.001
offset 4 in1: -0.845072 + -0.717403j  in2: -0.921176 + -0.513012j tolerance was: 0.001
offset 5 in1: 0.395931 + 0.139802j  in2: -0.00499856 + 0.000294006j tolerance was: 0.001
offset 6 in1: 0.58138 + 0.301571j  in2: 0.597208 + 0.111701j tolerance was: 0.001
offset 7 in1: 0.373094 + 0.818401j  in2: 0.000608006 + -0.00100112j tolerance was: 0.001
offset 8 in1: 0.41406 + 0.391089j  in2: 0.481277 + 0.184129j tolerance was: 0.001
offset 9 in1: 0.850985 + -0.569144j  in2: 2.67737e-05 + -0.000143105j tolerance was: 0.001
offset 10 in1: -0.537605 + -0.63551j  in2: -0.682274 + -0.271972j tolerance was: 0.001
offset 11 in1: -0.392492 + 0.269725j  in2: -7.39303e-06 + 1.97787e-07j tolerance was: 0.001
volk_32fc_s32fc_rotator2puppet_32fc: fail on arch a_sse4_1
offset 2 in1: -0.84885 + 0.194543j  in2: -0.805366 + 0.269719j tolerance was: 0.001
offset 3 in1: 0.85403 + 0.188718j  in2: -0.00175621 + 0.0954956j tolerance was: 0.001
offset 4 in1: -0.845072 + -0.717403j  in2: -0.921176 + -0.513012j tolerance was: 0.001
offset 5 in1: 0.395931 + 0.139802j  in2: -0.00499856 + 0.000294006j tolerance was: 0.001
offset 6 in1: 0.58138 + 0.301571j  in2: 0.597208 + 0.111701j tolerance was: 0.001
offset 7 in1: 0.373094 + 0.818401j  in2: 0.000608006 + -0.00100112j tolerance was: 0.001
offset 8 in1: 0.41406 + 0.391089j  in2: 0.481277 + 0.184129j tolerance was: 0.001
offset 9 in1: 0.850985 + -0.569144j  in2: 2.67737e-05 + -0.000143105j tolerance was: 0.001
offset 10 in1: -0.537605 + -0.63551j  in2: -0.682274 + -0.271972j tolerance was: 0.001
offset 11 in1: -0.392492 + 0.269725j  in2: -7.39303e-06 + 1.97787e-07j tolerance was: 0.001
volk_32fc_s32fc_rotator2puppet_32fc: fail on arch u_sse4_1
Best aligned arch: u_avx
Best unaligned arch: u_avx

[marcusmueller]

(you can find this by re-enabling the commented out %check parts in fedora's volk.spec, and running fedpkg local, at least here on my local Zen5)

[jdemel]

Thanks for the report!

The rotator is one of those notoriously failing kernels. Mostly due to differences in how we accumulate errors though. An offset of 2 should not be an issue though.

I can't reproduce this issue on my machine (older Zen 2 machine).

This issue might be a good candidate to write a (hopefully) more robust test with our new gtest based framework and go from there. Maybe that helps to discover the issue.

[marcusmueller]

I'm sorry, the error here is a factor of 10 in magnitude of an individual complex sample coming out of the puppet. The rotator is very very likely broken.

[marcusmueller]

and if it's notoriously failing, um, maybe disable protokernels or deprecate it? I mean, slow results are better than known-to-be-potentially-bad results…

[jdemel]

I'm sorry, the error here is a factor of 10 in magnitude of an individual complex sample coming out of the puppet. The rotator is very very likely broken.

Agreed. Smth in the rotator is broken.

It is a challenge to track this down though. Even more so because I can't reproduce the same error on my machine.

You mentioned fedpkg. Can you share a link to an instance where it fails?

[jdemel]

and if it's notoriously failing, um, maybe disable protokernels or deprecate it? I mean, slow results are better than known-to-be-potentially-bad results…

This was considered before but is still not the solution we'd be looking for. The use case allows for some small error. The error you reported is way off though.

Due to the structure of the algorithm, some error between implementations will always accumulate. This is expected.

The kernel should do 2 things

• not change the amplitude
• consistently rotate by phase_increment.

As long as these requirements are met sufficiently well, this is a useful thing.

[jdemel]

I implemented new tests: #795 I hope they allow for more fine-grained analysis.

[maitbot]

FYI this issue shows up when doing Debian builds with gcc-15

Debian Bug #1098070 volk: ftbfs with GCC-15
https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=1098070