Cannot build avx512 binaries on Azure

[fplazaonate]

Hi,

I have written a recipe that generates binaries for different cpu instruction sets:
#50691

Sometimes, the Azure workflow is successful (https://github.com/bioconda/bioconda-recipes/runs/30100279149), sometimes it fails (https://github.com/bioconda/bioconda-recipes/pull/50691/checks?check_run_id=30107543181).
When it fails, the error log indicates that cargo was not able to build an avx512 binary (aka x86-64-v4).

From what I understand, the workflow is sometimes executed on a machine that supports avx512 instructions, and sometimes not.

How can I make sure that the Azure workflow will run successfully?
Is there any way to select a machine with a specific CPU?

[aliciaaevans]

I see what you mean. I see from the logs that when it succeeded, it was using Agent: Azure Pipelines 4 and when it failed it was using Agent: Hosted Agent. I reran the job, but it seems it is using Hosted Agent again.

I can't find any way we can control which of the agents in the pool is used. If you find anything in the Azure Pipelines docs that you think would work, we can try that. We are limited to using the Microsoft-hosted agents.

[martin-g]

Would it be possible to detect the CPU features at runtime and enable the optimisations?
Otherwise now if the build host supports AVX512 but the user's machine does not it will fail instead of falling back to AVX256 or SSE.

[fplazaonate]

@martin-g
The current recipe already implements detection of CPU features at runtime.
I have tested it and machines with or without AVX512 support and it works.

@aliciaaevans
From what I understand, there is no way for an end user to select an agent with a specific configuration.
This is something that can be changed only by administrators via the Azure portal.