Target granularity setting

[manodeep]

While working with spack, I found out that installing for different cpu-types (e.g., cascadelake, sapphirerapids) was not possible while the setting was generic (as currently done). With generic, the only (optimised) targets would be x86_64_v3/v4 and not the actual micro-arch. Setting this config value to microarchitectures allows us to create optimised binaries targeting each cpu type available on the HPC system (ie queues).

[harshula]

Hi @manodeep , Are you finding that targetting x86_64_v3/x86_64_v4 is insufficient?

[manodeep]

@harshula Not particularly. The generic x86_64_v4 is probably fine for the time being. Since we are not running on the broadwell queue, we should not use v3 (which is AVX2).

For context, x86_64_v4 is the same as -xCORE-AVX512 - which is -march=skylake-avx512. Since the cascadelake, sapphirerapids cpus provide newer instructions, -march=cascadelake -mtune=sapphirerapids would be the right flag to use (such that the generated exe runs on both cascadelake and sapphirerapids queues, but the exe is tuned for sapphirerapids). However, in my experiments with compiler flags, I have not found that much of a performance improvement; and certainly within the supposed 5-10% runtime scatter on the sapphirerapids queue on gadi.

[harshula]

Hi @manodeep , What are your thoughts on having the default granularity set to generic and if required it can be over-ridden via an environment's spack.yaml?

[manodeep]

@harshula Works for me