Adds Neonv8 protokernel to 'volk_32f_64f_add_64f'

[dmiralles2009]

Adds Adds Neonv8 protokernel to volk_32f_64f_add_64f. Code in toolchains credited to Albin
Stigo (@ast). Execution improvement shown below

ubuntu@ubuntu:~/volk/build$ ./apps/volk_profile -d -R 32f_64f_add

RUN_VOLK_TESTS: volk_32f_64f_add_64f(131071,1987)
generic completed in 1079.18 ms
neon completed in 1056.38 ms
Best aligned arch: neon
Best unaligned arch: neon
Writing /home/ubuntu/.volk/volk_config...

[dmiralles2009]

Hello friends,
It seems like the code is currently breaking in areas not related to the changes this PR is trying to fix. I am adding this as part of a discussion initiated by @ast in issue #243 (more details there). Is the build system currently broken?. Below is the output error.

The following tests FAILED:
	 85 - qa_volk_32fc_s32f_magnitude_16i (Failed)
Errors while running CTest
The command "mkdir build && cd build && BOOST_ROOT=$BOOST_ROOT cmake ${CMAKE_ARG} ../ && make && ctest -V && cd .." exited with 8.

Thanks in advance,
Damian

[jdemel]

@dmiralles2009 could you test this with volk_profile -n -R 32fc_s32f_magnitude_16i?
Just a dry-run of this specific kernel.
Travis CI reports this same kernel as broken for your PR. But only for one build while all other configurations go through smoothly.

85: RUN_VOLK_TESTS: volk_32fc_s32f_magnitude_16i(131071,1)
85: a_avx2 completed in 0.10756 ms
85: a_sse3 completed in 0.566765 ms
85: a_sse completed in 0.594867 ms
85: generic completed in 0.490799 ms
85: u_avx2 completed in 0.131633 ms
85: offset 77213 in1: 407 in2: 406 tolerance was: 0
85: volk_32fc_s32f_magnitude_16i: fail on arch a_avx2
85: offset 77213 in1: 407 in2: 406 tolerance was: 0
85: volk_32fc_s32f_magnitude_16i: fail on arch a_sse3
85: offset 77213 in1: 407 in2: 406 tolerance was: 0
85: volk_32fc_s32f_magnitude_16i: fail on arch a_sse
85: offset 77213 in1: 407 in2: 406 tolerance was: 0
85: volk_32fc_s32f_magnitude_16i: fail on arch u_avx2
85: Best aligned arch: a_avx2
85: Best unaligned arch: u_avx2

Since this is a converter from float to int and the error says in1: 407 in2: 406 aka off-by-one error, for the time being this is one of those flaky QA tests that we'd love to fix. BTW, I can't make this test fail on my machine.

[dmiralles2009]

Hi @jdemel, that is curious because my PR does not modify that portion of the code. I also did a dry run at home and it seems to build without failing in my machine. I might look over the failing QA in the next few days. Thanks for the response :)

[hcab14]

85: RUN_VOLK_TESTS: volk_32fc_s32f_magnitude_16i(131071,1)
85: a_avx2 completed in 0.10756 ms
85: a_sse3 completed in 0.566765 ms
85: a_sse completed in 0.594867 ms
85: generic completed in 0.490799 ms
85: u_avx2 completed in 0.131633 ms
85: offset 77213 in1: 407 in2: 406 tolerance was: 0
85: volk_32fc_s32f_magnitude_16i: fail on arch a_avx2
85: offset 77213 in1: 407 in2: 406 tolerance was: 0
85: volk_32fc_s32f_magnitude_16i: fail on arch a_sse3
85: offset 77213 in1: 407 in2: 406 tolerance was: 0
85: volk_32fc_s32f_magnitude_16i: fail on arch a_sse
85: offset 77213 in1: 407 in2: 406 tolerance was: 0
85: volk_32fc_s32f_magnitude_16i: fail on arch u_avx2
85: Best aligned arch: a_avx2
85: Best unaligned arch: u_avx2

Since this is a converter from float to int and the error says in1: 407 in2: 406 aka off-by-one error, for the time being this is one of those flaky QA tests that we'd love to fix. BTW, I can't make this test fail on my machine.

#285

[jdemel]

Recently we had a discussion on -ffast-math and my understanding is that we should not use it because it breaks things.

[ast]

Yes if I remember correctly we agreed that it might break things because it allows non IEEE-754 compliant optimizations, and is therefore not a good default.

On the other hand NEONv7 is also not strictly IEEE-754 compliant and all tests passed for me with fast-math enabled.

But for now I think it's better to leave it out!

[michaelld]

Any further updates to this PR, or do y'all think it is ready for review?

[michaelld]

Please leave out the fast math flag for this specific PR. Otherwise, looks OK to me.

[dmiralles2009]

Hi all, I have been busy with work but will look at this closely during the weekend. Thanks for the replies :)

[michaelld]

LGTM

[jdemel]

LGTM now

[michaelld]

CI checks pass. Multiple positive reviews. Merging.