You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I am witnessing a strange behavior on my machine wrt arch ranking. I've narrowed it down to lib/volk_rank_archs.c:101.
The code does a popcnt on impl_deps, which contains 1 << various LV_{GENERIC,SSE4_1,AVX} constants, and picks the arch with the most 1s. However, since they are powers of two, they all get the same score.
Now, simply getting rid of popcnt solves the problem (as the best archs have a higher LV_ number), but I'm not sure that's the right thing to do
The text was updated successfully, but these errors were encountered:
n-west
added a commit
to n-west/volk
that referenced
this issue
Jul 24, 2015
You're correct about this behavior. However, it's only a problem when you don't have a volk_config (generated by volk_profile). I removed the popcnt and will push a fix with the upcoming release.
…-fma-trig-functions to next
* commit '6ec8fdb8194657d451c4bcc5f798e6ea56da51ad':
bugfix: Put FMA intrinsics in separate protokernel for sin and cos functions
I am witnessing a strange behavior on my machine wrt arch ranking. I've narrowed it down to lib/volk_rank_archs.c:101.
The code does a popcnt on impl_deps, which contains 1 << various LV_{GENERIC,SSE4_1,AVX} constants, and picks the arch with the most 1s. However, since they are powers of two, they all get the same score.
Now, simply getting rid of popcnt solves the problem (as the best archs have a higher LV_ number), but I'm not sure that's the right thing to do
The text was updated successfully, but these errors were encountered: