vec_active_tier() names the SIMD kernel tier the package resolved when its
shared library was loaded, and vec_active_lanes() gives the number of
double lanes that tier works on.
The tier is chosen from the instruction sets the CPU reports: "avx512"
(8 lanes), "avx2" (4 lanes), or "baseline" (2 lanes) where neither of
those is available.
Value
A scalar integer representing the number of active SIMD lanes.
A scalar character describing the active SIMD tier.
Forcing a lower tier
Setting the environment variable LOGMU_TIER to "baseline", "avx2" or
"avx512" before the package is loaded places a ceiling on the tier that may
be selected. It is there for comparing tiers in a benchmark, and for
reproducing a fault that only one of them shows.
The ceiling can only ever lower the tier, never raise it. A CPU that does not support the tier asked for still falls back to the best one it does support, because the processor feature checks apply exactly as they otherwise would.
An unrecognised value places no ceiling at all, and nothing is reported, so
call vec_active_tier() to confirm which tier is in use rather than assuming
the request was honoured.
The variable is read once, when the shared library is loaded. Changing it later has no effect for the rest of the session.