[ndpiitb]

I have successfully installed SU2 in my system but while running any problem it is showing a warning " WARNING: SU2 was not compiled for an AVX-capable architecture.".

Can anyone help regarding this? How can I install SU2 with AVX- capable architecture?

[pcg]

See here https://su2code.github.io/docs_v7/Bu...-optimizations

If you don’t know details use
export CXX_FLAGS="-march=native"
Before running meson.py

[CleverBoy]

I have the same error and I am curious about a thing. I have installed the mpi and run some tutorials with parallel computation. Does this error mean that simulation is not being solved by multiple cores, or something else?

[bigfootedrockmidget]

If you are worried that your parallelization is not working, then check:
1. that you get a speed-up if you use 1,2,4,8 cores
2. When you use 8 cores and mpi is not working correctly, you will not get a speed-up but su2 will be executed 8 times on the same core, and all the output is shown 8 times every iteration.


You can get an additional speed-up by making use of vectorization. Only certain parts of the code make use of it, so the amount of speed-up differs depending on your setup. You have to compile SU2 with vectorization support. If SU2 is not compiled with vectorization but you have the keyword USE_VECTORIZATION= YES in the config file, you will get this warning.

[CleverBoy]

Originally Posted by bigfootedrockmidget

If you are worried that your parallelization is not working, then check:
1. that you get a speed-up if you use 1,2,4,8 cores
2. When you use 8 cores and mpi is not working correctly, you will not get a speed-up but su2 will be executed 8 times on the same core, and all the output is shown 8 times every iteration.


You can get an additional speed-up by making use of vectorization. Only certain parts of the code make use of it, so the amount of speed-up differs depending on your setup. You have to compile SU2 with vectorization support. If SU2 is not compiled with vectorization but you have the keyword USE_VECTORIZATION= YES in the config file, you will get this warning.

Thanks for the tip, it was clearly faster (1.5 min faster) when computed with mpi. I also checked for a USE_VECTORIZATION keyword on the Turbulent_ONERAM6 tutorial (which I tried mpi) and there was no such line. Error was there anyway.

[pcg]

Some schemes are vectorized by default, hence the option has no effect.
On x86 you always get SSE vectorization (2 doubles), which is still faster, just not ideal since most x86 cpus also have AVX (4 doubles).