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
Eltwise unary reciprocal and rsqrt require Newton iteration to meet reasonable accuracy with ASIMD. Accuracy is reasonable for the SVE based code-path (Newton iteration).
The text was updated successfully, but these errors were encountered:
hfp
changed the title
Eltwise unary reciprocal and rsqrt require Newton iteration to meet accuracy with ASIMD
Eltwise unary reciprocal and rsqrt require Newton iteration on ASIMD
Jun 5, 2023
This is a low-priority issue if an issue at all. The unwritten policy for numerical accuracy can be "architectural accuracy" being "embraced" by default (even if exposing big differences betwee uArchs), i.e., LIBXSMM_ULP_PRECISION may be the only way to trigger higher precision (or Newton iterations on top of "native" precision).
Remaining work (need to keep issue open?): check if LIBXSMM_ULP_PRECISION can be made a vehicle to close the gap between architectural differences. For example, on x86 native sqrt, native rcp, and native rsqrt exist depending on type (FP32 vs F64), but even that exposes differences due to type (and/or vendor's implementation). On ARM64, precision and ISA coverage is apparently the weakest.
Eltwise unary reciprocal and rsqrt require Newton iteration to meet reasonable accuracy with ASIMD. Accuracy is reasonable for the SVE based code-path (Newton iteration).
The text was updated successfully, but these errors were encountered: